drivers/md/raid5.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * raid5.c : Multiple Devices driver for Linux
   4  *         Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
   5  *         Copyright (C) 1999, 2000 Ingo Molnar
   6  *         Copyright (C) 2002, 2003 H. Peter Anvin
   7  *
   8  * RAID-4/5/6 management functions.
   9  * Thanks to Penguin Computing for making the RAID-6 development possible
  10  * by donating a test server!
  11  */
  12
  13 /*
  14  * BITMAP UNPLUGGING:
  15  *
  16  * The sequencing for updating the bitmap reliably is a little
  17  * subtle (and I got it wrong the first time) so it deserves some
  18  * explanation.
  19  *
  20  * We group bitmap updates into batches.  Each batch has a number.
  21  * We may write out several batches at once, but that isn't very important.
  22  * conf->seq_write is the number of the last batch successfully written.
  23  * conf->seq_flush is the number of the last batch that was closed to
  24  *    new additions.
  25  * When we discover that we will need to write to any block in a stripe
  26  * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
  27  * the number of the batch it will be in. This is seq_flush+1.
  28  * When we are ready to do a write, if that batch hasn't been written yet,
  29  *   we plug the array and queue the stripe for later.
  30  * When an unplug happens, we increment bm_flush, thus closing the current
  31  *   batch.
  32  * When we notice that bm_flush > bm_write, we write out all pending updates
  33  * to the bitmap, and advance bm_write to where bm_flush was.
  34  * This may occasionally write a bit out twice, but is sure never to
  35  * miss any bits.
  36  */
  37
  38 #include <linux/blkdev.h>
  39 #include <linux/delay.h>
  40 #include <linux/kthread.h>
  41 #include <linux/raid/pq.h>
  42 #include <linux/async_tx.h>
  43 #include <linux/module.h>
  44 #include <linux/async.h>
  45 #include <linux/seq_file.h>
  46 #include <linux/cpu.h>
  47 #include <linux/slab.h>
  48 #include <linux/ratelimit.h>
  49 #include <linux/nodemask.h>
  50
  51 #include <trace/events/block.h>
  52 #include <linux/list_sort.h>
  53
  54 #include "md.h"
  55 #include "raid5.h"
  56 #include "raid0.h"
  57 #include "md-bitmap.h"
  58 #include "raid5-log.h"
  59
  60 #define UNSUPPORTED_MDDEV_FLAGS (1L << MD_FAILFAST_SUPPORTED)
  61
  62 #define cpu_to_group(cpu) cpu_to_node(cpu)
  63 #define ANY_GROUP NUMA_NO_NODE
  64
  65 #define RAID5_MAX_REQ_STRIPES 256
  66
  67 static bool devices_handle_discard_safely = false;
  68 module_param(devices_handle_discard_safely, bool, 0644);
  69 MODULE_PARM_DESC(devices_handle_discard_safely,
  70                  "Set to Y if all devices in each array reliably return zeroes on reads from discarded regions");
  71 static struct workqueue_struct *raid5_wq;
  72
  73 static void raid5_quiesce(struct mddev *mddev, int quiesce);
  74
  75 static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
  76 {
  77         int hash = (sect >> RAID5_STRIPE_SHIFT(conf)) & HASH_MASK;
  78         return &conf->stripe_hashtbl[hash];
  79 }
  80
  81 static inline int stripe_hash_locks_hash(struct r5conf *conf, sector_t sect)
  82 {
  83         return (sect >> RAID5_STRIPE_SHIFT(conf)) & STRIPE_HASH_LOCKS_MASK;
  84 }
  85
  86 static inline void lock_device_hash_lock(struct r5conf *conf, int hash)
  87         __acquires(&conf->device_lock)
  88 {
  89         spin_lock_irq(conf->hash_locks + hash);
  90         spin_lock(&conf->device_lock);
  91 }
  92
  93 static inline void unlock_device_hash_lock(struct r5conf *conf, int hash)
  94         __releases(&conf->device_lock)
  95 {
  96         spin_unlock(&conf->device_lock);
  97         spin_unlock_irq(conf->hash_locks + hash);
  98 }
  99
 100 static inline void lock_all_device_hash_locks_irq(struct r5conf *conf)
 101         __acquires(&conf->device_lock)
 102 {
 103         int i;
 104         spin_lock_irq(conf->hash_locks);
 105         for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
 106                 spin_lock_nest_lock(conf->hash_locks + i, conf->hash_locks);
 107         spin_lock(&conf->device_lock);
 108 }
 109
 110 static inline void unlock_all_device_hash_locks_irq(struct r5conf *conf)
 111         __releases(&conf->device_lock)
 112 {
 113         int i;
 114         spin_unlock(&conf->device_lock);
 115         for (i = NR_STRIPE_HASH_LOCKS - 1; i; i--)
 116                 spin_unlock(conf->hash_locks + i);
 117         spin_unlock_irq(conf->hash_locks);
 118 }
 119
 120 /* Find first data disk in a raid6 stripe */
 121 static inline int raid6_d0(struct stripe_head *sh)
 122 {
 123         if (sh->ddf_layout)
 124                 /* ddf always start from first device */
 125                 return 0;
 126         /* md starts just after Q block */
 127         if (sh->qd_idx == sh->disks - 1)
 128                 return 0;
 129         else
 130                 return sh->qd_idx + 1;
 131 }
 132 static inline int raid6_next_disk(int disk, int raid_disks)
 133 {
 134         disk++;
 135         return (disk < raid_disks) ? disk : 0;
 136 }
 137
 138 /* When walking through the disks in a raid5, starting at raid6_d0,
 139  * We need to map each disk to a 'slot', where the data disks are slot
 140  * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
 141  * is raid_disks-1.  This help does that mapping.
 142  */
 143 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
 144                              int *count, int syndrome_disks)
 145 {
 146         int slot = *count;
 147
 148         if (sh->ddf_layout)
 149                 (*count)++;
 150         if (idx == sh->pd_idx)
 151                 return syndrome_disks;
 152         if (idx == sh->qd_idx)
 153                 return syndrome_disks + 1;
 154         if (!sh->ddf_layout)
 155                 (*count)++;
 156         return slot;
 157 }
 158
 159 static void print_raid5_conf (struct r5conf *conf);
 160
 161 static int stripe_operations_active(struct stripe_head *sh)
 162 {
 163         return sh->check_state || sh->reconstruct_state ||
 164                test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
 165                test_bit(STRIPE_COMPUTE_RUN, &sh->state);
 166 }
 167
 168 static bool stripe_is_lowprio(struct stripe_head *sh)
 169 {
 170         return (test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state) ||
 171                 test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state)) &&
 172                !test_bit(STRIPE_R5C_CACHING, &sh->state);
 173 }
 174
 175 static void raid5_wakeup_stripe_thread(struct stripe_head *sh)
 176         __must_hold(&sh->raid_conf->device_lock)
 177 {
 178         struct r5conf *conf = sh->raid_conf;
 179         struct r5worker_group *group;
 180         int thread_cnt;
 181         int i, cpu = sh->cpu;
 182
 183         if (!cpu_online(cpu)) {
 184                 cpu = cpumask_any(cpu_online_mask);
 185                 sh->cpu = cpu;
 186         }
 187
 188         if (list_empty(&sh->lru)) {
 189                 struct r5worker_group *group;
 190                 group = conf->worker_groups + cpu_to_group(cpu);
 191                 if (stripe_is_lowprio(sh))
 192                         list_add_tail(&sh->lru, &group->loprio_list);
 193                 else
 194                         list_add_tail(&sh->lru, &group->handle_list);
 195                 group->stripes_cnt++;
 196                 sh->group = group;
 197         }
 198
 199         if (conf->worker_cnt_per_group == 0) {
 200                 md_wakeup_thread(conf->mddev->thread);
 201                 return;
 202         }
 203
 204         group = conf->worker_groups + cpu_to_group(sh->cpu);
 205
 206         group->workers[0].working = true;
 207         /* at least one worker should run to avoid race */
 208         queue_work_on(sh->cpu, raid5_wq, &group->workers[0].work);
 209
 210         thread_cnt = group->stripes_cnt / MAX_STRIPE_BATCH - 1;
 211         /* wakeup more workers */
 212         for (i = 1; i < conf->worker_cnt_per_group && thread_cnt > 0; i++) {
 213                 if (group->workers[i].working == false) {
 214                         group->workers[i].working = true;
 215                         queue_work_on(sh->cpu, raid5_wq,
 216                                       &group->workers[i].work);
 217                         thread_cnt--;
 218                 }
 219         }
 220 }
 221
 222 static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh,
 223                               struct list_head *temp_inactive_list)
 224         __must_hold(&conf->device_lock)
 225 {
 226         int i;
 227         int injournal = 0;      /* number of date pages with R5_InJournal */
 228
 229         BUG_ON(!list_empty(&sh->lru));
 230         BUG_ON(atomic_read(&conf->active_stripes)==0);
 231
 232         if (r5c_is_writeback(conf->log))
 233                 for (i = sh->disks; i--; )
 234                         if (test_bit(R5_InJournal, &sh->dev[i].flags))
 235                                 injournal++;
 236         /*
 237          * In the following cases, the stripe cannot be released to cached
 238          * lists. Therefore, we make the stripe write out and set
 239          * STRIPE_HANDLE:
 240          *   1. when quiesce in r5c write back;
 241          *   2. when resync is requested fot the stripe.
 242          */
 243         if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) ||
 244             (conf->quiesce && r5c_is_writeback(conf->log) &&
 245              !test_bit(STRIPE_HANDLE, &sh->state) && injournal != 0)) {
 246                 if (test_bit(STRIPE_R5C_CACHING, &sh->state))
 247                         r5c_make_stripe_write_out(sh);
 248                 set_bit(STRIPE_HANDLE, &sh->state);
 249         }
 250
 251         if (test_bit(STRIPE_HANDLE, &sh->state)) {
 252                 if (test_bit(STRIPE_DELAYED, &sh->state) &&
 253                     !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 254                         list_add_tail(&sh->lru, &conf->delayed_list);
 255                 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
 256                            sh->bm_seq - conf->seq_write > 0)
 257                         list_add_tail(&sh->lru, &conf->bitmap_list);
 258                 else {
 259                         clear_bit(STRIPE_DELAYED, &sh->state);
 260                         clear_bit(STRIPE_BIT_DELAY, &sh->state);
 261                         if (conf->worker_cnt_per_group == 0) {
 262                                 if (stripe_is_lowprio(sh))
 263                                         list_add_tail(&sh->lru,
 264                                                         &conf->loprio_list);
 265                                 else
 266                                         list_add_tail(&sh->lru,
 267                                                         &conf->handle_list);
 268                         } else {
 269                                 raid5_wakeup_stripe_thread(sh);
 270                                 return;
 271                         }
 272                 }
 273                 md_wakeup_thread(conf->mddev->thread);
 274         } else {
 275                 BUG_ON(stripe_operations_active(sh));
 276                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 277                         if (atomic_dec_return(&conf->preread_active_stripes)
 278                             < IO_THRESHOLD)
 279                                 md_wakeup_thread(conf->mddev->thread);
 280                 atomic_dec(&conf->active_stripes);
 281                 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
 282                         if (!r5c_is_writeback(conf->log))
 283                                 list_add_tail(&sh->lru, temp_inactive_list);
 284                         else {
 285                                 WARN_ON(test_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags));
 286                                 if (injournal == 0)
 287                                         list_add_tail(&sh->lru, temp_inactive_list);
 288                                 else if (injournal == conf->raid_disks - conf->max_degraded) {
 289                                         /* full stripe */
 290                                         if (!test_and_set_bit(STRIPE_R5C_FULL_STRIPE, &sh->state))
 291                                                 atomic_inc(&conf->r5c_cached_full_stripes);
 292                                         if (test_and_clear_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state))
 293                                                 atomic_dec(&conf->r5c_cached_partial_stripes);
 294                                         list_add_tail(&sh->lru, &conf->r5c_full_stripe_list);
 295                                         r5c_check_cached_full_stripe(conf);
 296                                 } else
 297                                         /*
 298                                          * STRIPE_R5C_PARTIAL_STRIPE is set in
 299                                          * r5c_try_caching_write(). No need to
 300                                          * set it again.
 301                                          */
 302                                         list_add_tail(&sh->lru, &conf->r5c_partial_stripe_list);
 303                         }
 304                 }
 305         }
 306 }
 307
 308 static void __release_stripe(struct r5conf *conf, struct stripe_head *sh,
 309                              struct list_head *temp_inactive_list)
 310         __must_hold(&conf->device_lock)
 311 {
 312         if (atomic_dec_and_test(&sh->count))
 313                 do_release_stripe(conf, sh, temp_inactive_list);
 314 }
 315
 316 /*
 317  * @hash could be NR_STRIPE_HASH_LOCKS, then we have a list of inactive_list
 318  *
 319  * Be careful: Only one task can add/delete stripes from temp_inactive_list at
 320  * given time. Adding stripes only takes device lock, while deleting stripes
 321  * only takes hash lock.
 322  */
 323 static void release_inactive_stripe_list(struct r5conf *conf,
 324                                          struct list_head *temp_inactive_list,
 325                                          int hash)
 326 {
 327         int size;
 328         bool do_wakeup = false;
 329         unsigned long flags;
 330
 331         if (hash == NR_STRIPE_HASH_LOCKS) {
 332                 size = NR_STRIPE_HASH_LOCKS;
 333                 hash = NR_STRIPE_HASH_LOCKS - 1;
 334         } else
 335                 size = 1;
 336         while (size) {
 337                 struct list_head *list = &temp_inactive_list[size - 1];
 338
 339                 /*
 340                  * We don't hold any lock here yet, raid5_get_active_stripe() might
 341                  * remove stripes from the list
 342                  */
 343                 if (!list_empty_careful(list)) {
 344                         spin_lock_irqsave(conf->hash_locks + hash, flags);
 345                         if (list_empty(conf->inactive_list + hash) &&
 346                             !list_empty(list))
 347                                 atomic_dec(&conf->empty_inactive_list_nr);
 348                         list_splice_tail_init(list, conf->inactive_list + hash);
 349                         do_wakeup = true;
 350                         spin_unlock_irqrestore(conf->hash_locks + hash, flags);
 351                 }
 352                 size--;
 353                 hash--;
 354         }
 355
 356         if (do_wakeup) {
 357                 wake_up(&conf->wait_for_stripe);
 358                 if (atomic_read(&conf->active_stripes) == 0)
 359                         wake_up(&conf->wait_for_quiescent);
 360                 if (conf->retry_read_aligned)
 361                         md_wakeup_thread(conf->mddev->thread);
 362         }
 363 }
 364
 365 static int release_stripe_list(struct r5conf *conf,
 366                                struct list_head *temp_inactive_list)
 367         __must_hold(&conf->device_lock)
 368 {
 369         struct stripe_head *sh, *t;
 370         int count = 0;
 371         struct llist_node *head;
 372
 373         head = llist_del_all(&conf->released_stripes);
 374         head = llist_reverse_order(head);
 375         llist_for_each_entry_safe(sh, t, head, release_list) {
 376                 int hash;
 377
 378                 /* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
 379                 smp_mb();
 380                 clear_bit(STRIPE_ON_RELEASE_LIST, &sh->state);
 381                 /*
 382                  * Don't worry the bit is set here, because if the bit is set
 383                  * again, the count is always > 1. This is true for
 384                  * STRIPE_ON_UNPLUG_LIST bit too.
 385                  */
 386                 hash = sh->hash_lock_index;
 387                 __release_stripe(conf, sh, &temp_inactive_list[hash]);
 388                 count++;
 389         }
 390
 391         return count;
 392 }
 393
 394 void raid5_release_stripe(struct stripe_head *sh)
 395 {
 396         struct r5conf *conf = sh->raid_conf;
 397         unsigned long flags;
 398         struct list_head list;
 399         int hash;
 400         bool wakeup;
 401
 402         /* Avoid release_list until the last reference.
 403          */
 404         if (atomic_add_unless(&sh->count, -1, 1))
 405                 return;
 406
 407         if (unlikely(!conf->mddev->thread) ||
 408                 test_and_set_bit(STRIPE_ON_RELEASE_LIST, &sh->state))
 409                 goto slow_path;
 410         wakeup = llist_add(&sh->release_list, &conf->released_stripes);
 411         if (wakeup)
 412                 md_wakeup_thread(conf->mddev->thread);
 413         return;
 414 slow_path:
 415         /* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
 416         if (atomic_dec_and_lock_irqsave(&sh->count, &conf->device_lock, flags)) {
 417                 INIT_LIST_HEAD(&list);
 418                 hash = sh->hash_lock_index;
 419                 do_release_stripe(conf, sh, &list);
 420                 spin_unlock_irqrestore(&conf->device_lock, flags);
 421                 release_inactive_stripe_list(conf, &list, hash);
 422         }
 423 }
 424
 425 static inline void remove_hash(struct stripe_head *sh)
 426 {
 427         pr_debug("remove_hash(), stripe %llu\n",
 428                 (unsigned long long)sh->sector);
 429
 430         hlist_del_init(&sh->hash);
 431 }
 432
 433 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
 434 {
 435         struct hlist_head *hp = stripe_hash(conf, sh->sector);
 436
 437         pr_debug("insert_hash(), stripe %llu\n",
 438                 (unsigned long long)sh->sector);
 439
 440         hlist_add_head(&sh->hash, hp);
 441 }
 442
 443 /* find an idle stripe, make sure it is unhashed, and return it. */
 444 static struct stripe_head *get_free_stripe(struct r5conf *conf, int hash)
 445 {
 446         struct stripe_head *sh = NULL;
 447         struct list_head *first;
 448
 449         if (list_empty(conf->inactive_list + hash))
 450                 goto out;
 451         first = (conf->inactive_list + hash)->next;
 452         sh = list_entry(first, struct stripe_head, lru);
 453         list_del_init(first);
 454         remove_hash(sh);
 455         atomic_inc(&conf->active_stripes);
 456         BUG_ON(hash != sh->hash_lock_index);
 457         if (list_empty(conf->inactive_list + hash))
 458                 atomic_inc(&conf->empty_inactive_list_nr);
 459 out:
 460         return sh;
 461 }
 462
 463 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
 464 static void free_stripe_pages(struct stripe_head *sh)
 465 {
 466         int i;
 467         struct page *p;
 468
 469         /* Have not allocate page pool */
 470         if (!sh->pages)
 471                 return;
 472
 473         for (i = 0; i < sh->nr_pages; i++) {
 474                 p = sh->pages[i];
 475                 if (p)
 476                         put_page(p);
 477                 sh->pages[i] = NULL;
 478         }
 479 }
 480
 481 static int alloc_stripe_pages(struct stripe_head *sh, gfp_t gfp)
 482 {
 483         int i;
 484         struct page *p;
 485
 486         for (i = 0; i < sh->nr_pages; i++) {
 487                 /* The page have allocated. */
 488                 if (sh->pages[i])
 489                         continue;
 490
 491                 p = alloc_page(gfp);
 492                 if (!p) {
 493                         free_stripe_pages(sh);
 494                         return -ENOMEM;
 495                 }
 496                 sh->pages[i] = p;
 497         }
 498         return 0;
 499 }
 500
 501 static int
 502 init_stripe_shared_pages(struct stripe_head *sh, struct r5conf *conf, int disks)
 503 {
 504         int nr_pages, cnt;
 505
 506         if (sh->pages)
 507                 return 0;
 508
 509         /* Each of the sh->dev[i] need one conf->stripe_size */
 510         cnt = PAGE_SIZE / conf->stripe_size;
 511         nr_pages = (disks + cnt - 1) / cnt;
 512
 513         sh->pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
 514         if (!sh->pages)
 515                 return -ENOMEM;
 516         sh->nr_pages = nr_pages;
 517         sh->stripes_per_page = cnt;
 518         return 0;
 519 }
 520 #endif
 521
 522 static void shrink_buffers(struct stripe_head *sh)
 523 {
 524         int i;
 525         int num = sh->raid_conf->pool_size;
 526
 527 #if PAGE_SIZE == DEFAULT_STRIPE_SIZE
 528         for (i = 0; i < num ; i++) {
 529                 struct page *p;
 530
 531                 WARN_ON(sh->dev[i].page != sh->dev[i].orig_page);
 532                 p = sh->dev[i].page;
 533                 if (!p)
 534                         continue;
 535                 sh->dev[i].page = NULL;
 536                 put_page(p);
 537         }
 538 #else
 539         for (i = 0; i < num; i++)
 540                 sh->dev[i].page = NULL;
 541         free_stripe_pages(sh); /* Free pages */
 542 #endif
 543 }
 544
 545 static int grow_buffers(struct stripe_head *sh, gfp_t gfp)
 546 {
 547         int i;
 548         int num = sh->raid_conf->pool_size;
 549
 550 #if PAGE_SIZE == DEFAULT_STRIPE_SIZE
 551         for (i = 0; i < num; i++) {
 552                 struct page *page;
 553
 554                 if (!(page = alloc_page(gfp))) {
 555                         return 1;
 556                 }
 557                 sh->dev[i].page = page;
 558                 sh->dev[i].orig_page = page;
 559                 sh->dev[i].offset = 0;
 560         }
 561 #else
 562         if (alloc_stripe_pages(sh, gfp))
 563                 return -ENOMEM;
 564
 565         for (i = 0; i < num; i++) {
 566                 sh->dev[i].page = raid5_get_dev_page(sh, i);
 567                 sh->dev[i].orig_page = sh->dev[i].page;
 568                 sh->dev[i].offset = raid5_get_page_offset(sh, i);
 569         }
 570 #endif
 571         return 0;
 572 }
 573
 574 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
 575                             struct stripe_head *sh);
 576
 577 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
 578 {
 579         struct r5conf *conf = sh->raid_conf;
 580         int i, seq;
 581
 582         BUG_ON(atomic_read(&sh->count) != 0);
 583         BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
 584         BUG_ON(stripe_operations_active(sh));
 585         BUG_ON(sh->batch_head);
 586
 587         pr_debug("init_stripe called, stripe %llu\n",
 588                 (unsigned long long)sector);
 589 retry:
 590         seq = read_seqcount_begin(&conf->gen_lock);
 591         sh->generation = conf->generation - previous;
 592         sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
 593         sh->sector = sector;
 594         stripe_set_idx(sector, conf, previous, sh);
 595         sh->state = 0;
 596
 597         for (i = sh->disks; i--; ) {
 598                 struct r5dev *dev = &sh->dev[i];
 599
 600                 if (dev->toread || dev->read || dev->towrite || dev->written ||
 601                     test_bit(R5_LOCKED, &dev->flags)) {
 602                         pr_err("sector=%llx i=%d %p %p %p %p %d\n",
 603                                (unsigned long long)sh->sector, i, dev->toread,
 604                                dev->read, dev->towrite, dev->written,
 605                                test_bit(R5_LOCKED, &dev->flags));
 606                         WARN_ON(1);
 607                 }
 608                 dev->flags = 0;
 609                 dev->sector = raid5_compute_blocknr(sh, i, previous);
 610         }
 611         if (read_seqcount_retry(&conf->gen_lock, seq))
 612                 goto retry;
 613         sh->overwrite_disks = 0;
 614         insert_hash(conf, sh);
 615         sh->cpu = smp_processor_id();
 616         set_bit(STRIPE_BATCH_READY, &sh->state);
 617 }
 618
 619 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
 620                                          short generation)
 621 {
 622         struct stripe_head *sh;
 623
 624         pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
 625         hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
 626                 if (sh->sector == sector && sh->generation == generation)
 627                         return sh;
 628         pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
 629         return NULL;
 630 }
 631
 632 static struct stripe_head *find_get_stripe(struct r5conf *conf,
 633                 sector_t sector, short generation, int hash)
 634 {
 635         int inc_empty_inactive_list_flag;
 636         struct stripe_head *sh;
 637
 638         sh = __find_stripe(conf, sector, generation);
 639         if (!sh)
 640                 return NULL;
 641
 642         if (atomic_inc_not_zero(&sh->count))
 643                 return sh;
 644
 645         /*
 646          * Slow path. The reference count is zero which means the stripe must
 647          * be on a list (sh->lru). Must remove the stripe from the list that
 648          * references it with the device_lock held.
 649          */
 650
 651         spin_lock(&conf->device_lock);
 652         if (!atomic_read(&sh->count)) {
 653                 if (!test_bit(STRIPE_HANDLE, &sh->state))
 654                         atomic_inc(&conf->active_stripes);
 655                 BUG_ON(list_empty(&sh->lru) &&
 656                        !test_bit(STRIPE_EXPANDING, &sh->state));
 657                 inc_empty_inactive_list_flag = 0;
 658                 if (!list_empty(conf->inactive_list + hash))
 659                         inc_empty_inactive_list_flag = 1;
 660                 list_del_init(&sh->lru);
 661                 if (list_empty(conf->inactive_list + hash) &&
 662                     inc_empty_inactive_list_flag)
 663                         atomic_inc(&conf->empty_inactive_list_nr);
 664                 if (sh->group) {
 665                         sh->group->stripes_cnt--;
 666                         sh->group = NULL;
 667                 }
 668         }
 669         atomic_inc(&sh->count);
 670         spin_unlock(&conf->device_lock);
 671
 672         return sh;
 673 }
 674
 675 /*
 676  * Need to check if array has failed when deciding whether to:
 677  *  - start an array
 678  *  - remove non-faulty devices
 679  *  - add a spare
 680  *  - allow a reshape
 681  * This determination is simple when no reshape is happening.
 682  * However if there is a reshape, we need to carefully check
 683  * both the before and after sections.
 684  * This is because some failed devices may only affect one
 685  * of the two sections, and some non-in_sync devices may
 686  * be insync in the section most affected by failed devices.
 687  *
 688  * Most calls to this function hold &conf->device_lock. Calls
 689  * in raid5_run() do not require the lock as no other threads
 690  * have been started yet.
 691  */
 692 int raid5_calc_degraded(struct r5conf *conf)
 693 {
 694         int degraded, degraded2;
 695         int i;
 696
 697         rcu_read_lock();
 698         degraded = 0;
 699         for (i = 0; i < conf->previous_raid_disks; i++) {
 700                 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
 701                 if (rdev && test_bit(Faulty, &rdev->flags))
 702                         rdev = rcu_dereference(conf->disks[i].replacement);
 703                 if (!rdev || test_bit(Faulty, &rdev->flags))
 704                         degraded++;
 705                 else if (test_bit(In_sync, &rdev->flags))
 706                         ;
 707                 else
 708                         /* not in-sync or faulty.
 709                          * If the reshape increases the number of devices,
 710                          * this is being recovered by the reshape, so
 711                          * this 'previous' section is not in_sync.
 712                          * If the number of devices is being reduced however,
 713                          * the device can only be part of the array if
 714                          * we are reverting a reshape, so this section will
 715                          * be in-sync.
 716                          */
 717                         if (conf->raid_disks >= conf->previous_raid_disks)
 718                                 degraded++;
 719         }
 720         rcu_read_unlock();
 721         if (conf->raid_disks == conf->previous_raid_disks)
 722                 return degraded;
 723         rcu_read_lock();
 724         degraded2 = 0;
 725         for (i = 0; i < conf->raid_disks; i++) {
 726                 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
 727                 if (rdev && test_bit(Faulty, &rdev->flags))
 728                         rdev = rcu_dereference(conf->disks[i].replacement);
 729                 if (!rdev || test_bit(Faulty, &rdev->flags))
 730                         degraded2++;
 731                 else if (test_bit(In_sync, &rdev->flags))
 732                         ;
 733                 else
 734                         /* not in-sync or faulty.
 735                          * If reshape increases the number of devices, this
 736                          * section has already been recovered, else it
 737                          * almost certainly hasn't.
 738                          */
 739                         if (conf->raid_disks <= conf->previous_raid_disks)
 740                                 degraded2++;
 741         }
 742         rcu_read_unlock();
 743         if (degraded2 > degraded)
 744                 return degraded2;
 745         return degraded;
 746 }
 747
 748 static bool has_failed(struct r5conf *conf)
 749 {
 750         int degraded = conf->mddev->degraded;
 751
 752         if (test_bit(MD_BROKEN, &conf->mddev->flags))
 753                 return true;
 754
 755         if (conf->mddev->reshape_position != MaxSector)
 756                 degraded = raid5_calc_degraded(conf);
 757
 758         return degraded > conf->max_degraded;
 759 }
 760
 761 enum stripe_result {
 762         STRIPE_SUCCESS = 0,
 763         STRIPE_RETRY,
 764         STRIPE_SCHEDULE_AND_RETRY,
 765         STRIPE_FAIL,
 766 };
 767
 768 struct stripe_request_ctx {
 769         /* a reference to the last stripe_head for batching */
 770         struct stripe_head *batch_last;
 771
 772         /* first sector in the request */
 773         sector_t first_sector;
 774
 775         /* last sector in the request */
 776         sector_t last_sector;
 777
 778         /*
 779          * bitmap to track stripe sectors that have been added to stripes
 780          * add one to account for unaligned requests
 781          */
 782         DECLARE_BITMAP(sectors_to_do, RAID5_MAX_REQ_STRIPES + 1);
 783
 784         /* the request had REQ_PREFLUSH, cleared after the first stripe_head */
 785         bool do_flush;
 786 };
 787
 788 /*
 789  * Block until another thread clears R5_INACTIVE_BLOCKED or
 790  * there are fewer than 3/4 the maximum number of active stripes
 791  * and there is an inactive stripe available.
 792  */
 793 static bool is_inactive_blocked(struct r5conf *conf, int hash)
 794 {
 795         if (list_empty(conf->inactive_list + hash))
 796                 return false;
 797
 798         if (!test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state))
 799                 return true;
 800
 801         return (atomic_read(&conf->active_stripes) <
 802                 (conf->max_nr_stripes * 3 / 4));
 803 }
 804
 805 struct stripe_head *raid5_get_active_stripe(struct r5conf *conf,
 806                 struct stripe_request_ctx *ctx, sector_t sector,
 807                 unsigned int flags)
 808 {
 809         struct stripe_head *sh;
 810         int hash = stripe_hash_locks_hash(conf, sector);
 811         int previous = !!(flags & R5_GAS_PREVIOUS);
 812
 813         pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
 814
 815         spin_lock_irq(conf->hash_locks + hash);
 816
 817         for (;;) {
 818                 if (!(flags & R5_GAS_NOQUIESCE) && conf->quiesce) {
 819                         /*
 820                          * Must release the reference to batch_last before
 821                          * waiting, on quiesce, otherwise the batch_last will
 822                          * hold a reference to a stripe and raid5_quiesce()
 823                          * will deadlock waiting for active_stripes to go to
 824                          * zero.
 825                          */
 826                         if (ctx && ctx->batch_last) {
 827                                 raid5_release_stripe(ctx->batch_last);
 828                                 ctx->batch_last = NULL;
 829                         }
 830
 831                         wait_event_lock_irq(conf->wait_for_quiescent,
 832                                             !conf->quiesce,
 833                                             *(conf->hash_locks + hash));
 834                 }
 835
 836                 sh = find_get_stripe(conf, sector, conf->generation - previous,
 837                                      hash);
 838                 if (sh)
 839                         break;
 840
 841                 if (!test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state)) {
 842                         sh = get_free_stripe(conf, hash);
 843                         if (sh) {
 844                                 r5c_check_stripe_cache_usage(conf);
 845                                 init_stripe(sh, sector, previous);
 846                                 atomic_inc(&sh->count);
 847                                 break;
 848                         }
 849
 850                         if (!test_bit(R5_DID_ALLOC, &conf->cache_state))
 851                                 set_bit(R5_ALLOC_MORE, &conf->cache_state);
 852                 }
 853
 854                 if (flags & R5_GAS_NOBLOCK)
 855                         break;
 856
 857                 set_bit(R5_INACTIVE_BLOCKED, &conf->cache_state);
 858                 r5l_wake_reclaim(conf->log, 0);
 859
 860                 /* release batch_last before wait to avoid risk of deadlock */
 861                 if (ctx && ctx->batch_last) {
 862                         raid5_release_stripe(ctx->batch_last);
 863                         ctx->batch_last = NULL;
 864                 }
 865
 866                 wait_event_lock_irq(conf->wait_for_stripe,
 867                                     is_inactive_blocked(conf, hash),
 868                                     *(conf->hash_locks + hash));
 869                 clear_bit(R5_INACTIVE_BLOCKED, &conf->cache_state);
 870         }
 871
 872         spin_unlock_irq(conf->hash_locks + hash);
 873         return sh;
 874 }
 875
 876 static bool is_full_stripe_write(struct stripe_head *sh)
 877 {
 878         BUG_ON(sh->overwrite_disks > (sh->disks - sh->raid_conf->max_degraded));
 879         return sh->overwrite_disks == (sh->disks - sh->raid_conf->max_degraded);
 880 }
 881
 882 static void lock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
 883                 __acquires(&sh1->stripe_lock)
 884                 __acquires(&sh2->stripe_lock)
 885 {
 886         if (sh1 > sh2) {
 887                 spin_lock_irq(&sh2->stripe_lock);
 888                 spin_lock_nested(&sh1->stripe_lock, 1);
 889         } else {
 890                 spin_lock_irq(&sh1->stripe_lock);
 891                 spin_lock_nested(&sh2->stripe_lock, 1);
 892         }
 893 }
 894
 895 static void unlock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
 896                 __releases(&sh1->stripe_lock)
 897                 __releases(&sh2->stripe_lock)
 898 {
 899         spin_unlock(&sh1->stripe_lock);
 900         spin_unlock_irq(&sh2->stripe_lock);
 901 }
 902
 903 /* Only freshly new full stripe normal write stripe can be added to a batch list */
 904 static bool stripe_can_batch(struct stripe_head *sh)
 905 {
 906         struct r5conf *conf = sh->raid_conf;
 907
 908         if (raid5_has_log(conf) || raid5_has_ppl(conf))
 909                 return false;
 910         return test_bit(STRIPE_BATCH_READY, &sh->state) &&
 911                 !test_bit(STRIPE_BITMAP_PENDING, &sh->state) &&
 912                 is_full_stripe_write(sh);
 913 }
 914
 915 /* we only do back search */
 916 static void stripe_add_to_batch_list(struct r5conf *conf,
 917                 struct stripe_head *sh, struct stripe_head *last_sh)
 918 {
 919         struct stripe_head *head;
 920         sector_t head_sector, tmp_sec;
 921         int hash;
 922         int dd_idx;
 923
 924         /* Don't cross chunks, so stripe pd_idx/qd_idx is the same */
 925         tmp_sec = sh->sector;
 926         if (!sector_div(tmp_sec, conf->chunk_sectors))
 927                 return;
 928         head_sector = sh->sector - RAID5_STRIPE_SECTORS(conf);
 929
 930         if (last_sh && head_sector == last_sh->sector) {
 931                 head = last_sh;
 932                 atomic_inc(&head->count);
 933         } else {
 934                 hash = stripe_hash_locks_hash(conf, head_sector);
 935                 spin_lock_irq(conf->hash_locks + hash);
 936                 head = find_get_stripe(conf, head_sector, conf->generation,
 937                                        hash);
 938                 spin_unlock_irq(conf->hash_locks + hash);
 939                 if (!head)
 940                         return;
 941                 if (!stripe_can_batch(head))
 942                         goto out;
 943         }
 944
 945         lock_two_stripes(head, sh);
 946         /* clear_batch_ready clear the flag */
 947         if (!stripe_can_batch(head) || !stripe_can_batch(sh))
 948                 goto unlock_out;
 949
 950         if (sh->batch_head)
 951                 goto unlock_out;
 952
 953         dd_idx = 0;
 954         while (dd_idx == sh->pd_idx || dd_idx == sh->qd_idx)
 955                 dd_idx++;
 956         if (head->dev[dd_idx].towrite->bi_opf != sh->dev[dd_idx].towrite->bi_opf ||
 957             bio_op(head->dev[dd_idx].towrite) != bio_op(sh->dev[dd_idx].towrite))
 958                 goto unlock_out;
 959
 960         if (head->batch_head) {
 961                 spin_lock(&head->batch_head->batch_lock);
 962                 /* This batch list is already running */
 963                 if (!stripe_can_batch(head)) {
 964                         spin_unlock(&head->batch_head->batch_lock);
 965                         goto unlock_out;
 966                 }
 967                 /*
 968                  * We must assign batch_head of this stripe within the
 969                  * batch_lock, otherwise clear_batch_ready of batch head
 970                  * stripe could clear BATCH_READY bit of this stripe and
 971                  * this stripe->batch_head doesn't get assigned, which
 972                  * could confuse clear_batch_ready for this stripe
 973                  */
 974                 sh->batch_head = head->batch_head;
 975
 976                 /*
 977                  * at this point, head's BATCH_READY could be cleared, but we
 978                  * can still add the stripe to batch list
 979                  */
 980                 list_add(&sh->batch_list, &head->batch_list);
 981                 spin_unlock(&head->batch_head->batch_lock);
 982         } else {
 983                 head->batch_head = head;
 984                 sh->batch_head = head->batch_head;
 985                 spin_lock(&head->batch_lock);
 986                 list_add_tail(&sh->batch_list, &head->batch_list);
 987                 spin_unlock(&head->batch_lock);
 988         }
 989
 990         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 991                 if (atomic_dec_return(&conf->preread_active_stripes)
 992                     < IO_THRESHOLD)
 993                         md_wakeup_thread(conf->mddev->thread);
 994
 995         if (test_and_clear_bit(STRIPE_BIT_DELAY, &sh->state)) {
 996                 int seq = sh->bm_seq;
 997                 if (test_bit(STRIPE_BIT_DELAY, &sh->batch_head->state) &&
 998                     sh->batch_head->bm_seq > seq)
 999                         seq = sh->batch_head->bm_seq;
1000                 set_bit(STRIPE_BIT_DELAY, &sh->batch_head->state);
1001                 sh->batch_head->bm_seq = seq;
1002         }
1003
1004         atomic_inc(&sh->count);
1005 unlock_out:
1006         unlock_two_stripes(head, sh);
1007 out:
1008         raid5_release_stripe(head);
1009 }
1010
1011 /* Determine if 'data_offset' or 'new_data_offset' should be used
1012  * in this stripe_head.
1013  */
1014 static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
1015 {
1016         sector_t progress = conf->reshape_progress;
1017         /* Need a memory barrier to make sure we see the value
1018          * of conf->generation, or ->data_offset that was set before
1019          * reshape_progress was updated.
1020          */
1021         smp_rmb();
1022         if (progress == MaxSector)
1023                 return 0;
1024         if (sh->generation == conf->generation - 1)
1025                 return 0;
1026         /* We are in a reshape, and this is a new-generation stripe,
1027          * so use new_data_offset.
1028          */
1029         return 1;
1030 }
1031
1032 static void dispatch_bio_list(struct bio_list *tmp)
1033 {
1034         struct bio *bio;
1035
1036         while ((bio = bio_list_pop(tmp)))
1037                 submit_bio_noacct(bio);
1038 }
1039
1040 static int cmp_stripe(void *priv, const struct list_head *a,
1041                       const struct list_head *b)
1042 {
1043         const struct r5pending_data *da = list_entry(a,
1044                                 struct r5pending_data, sibling);
1045         const struct r5pending_data *db = list_entry(b,
1046                                 struct r5pending_data, sibling);
1047         if (da->sector > db->sector)
1048                 return 1;
1049         if (da->sector < db->sector)
1050                 return -1;
1051         return 0;
1052 }
1053
1054 static void dispatch_defer_bios(struct r5conf *conf, int target,
1055                                 struct bio_list *list)
1056 {
1057         struct r5pending_data *data;
1058         struct list_head *first, *next = NULL;
1059         int cnt = 0;
1060
1061         if (conf->pending_data_cnt == 0)
1062                 return;
1063
1064         list_sort(NULL, &conf->pending_list, cmp_stripe);
1065
1066         first = conf->pending_list.next;
1067
1068         /* temporarily move the head */
1069         if (conf->next_pending_data)
1070                 list_move_tail(&conf->pending_list,
1071                                 &conf->next_pending_data->sibling);
1072
1073         while (!list_empty(&conf->pending_list)) {
1074                 data = list_first_entry(&conf->pending_list,
1075                         struct r5pending_data, sibling);
1076                 if (&data->sibling == first)
1077                         first = data->sibling.next;
1078                 next = data->sibling.next;
1079
1080                 bio_list_merge(list, &data->bios);
1081                 list_move(&data->sibling, &conf->free_list);
1082                 cnt++;
1083                 if (cnt >= target)
1084                         break;
1085         }
1086         conf->pending_data_cnt -= cnt;
1087         BUG_ON(conf->pending_data_cnt < 0 || cnt < target);
1088
1089         if (next != &conf->pending_list)
1090                 conf->next_pending_data = list_entry(next,
1091                                 struct r5pending_data, sibling);
1092         else
1093                 conf->next_pending_data = NULL;
1094         /* list isn't empty */
1095         if (first != &conf->pending_list)
1096                 list_move_tail(&conf->pending_list, first);
1097 }
1098
1099 static void flush_deferred_bios(struct r5conf *conf)
1100 {
1101         struct bio_list tmp = BIO_EMPTY_LIST;
1102
1103         if (conf->pending_data_cnt == 0)
1104                 return;
1105
1106         spin_lock(&conf->pending_bios_lock);
1107         dispatch_defer_bios(conf, conf->pending_data_cnt, &tmp);
1108         BUG_ON(conf->pending_data_cnt != 0);
1109         spin_unlock(&conf->pending_bios_lock);
1110
1111         dispatch_bio_list(&tmp);
1112 }
1113
1114 static void defer_issue_bios(struct r5conf *conf, sector_t sector,
1115                                 struct bio_list *bios)
1116 {
1117         struct bio_list tmp = BIO_EMPTY_LIST;
1118         struct r5pending_data *ent;
1119
1120         spin_lock(&conf->pending_bios_lock);
1121         ent = list_first_entry(&conf->free_list, struct r5pending_data,
1122                                                         sibling);
1123         list_move_tail(&ent->sibling, &conf->pending_list);
1124         ent->sector = sector;
1125         bio_list_init(&ent->bios);
1126         bio_list_merge(&ent->bios, bios);
1127         conf->pending_data_cnt++;
1128         if (conf->pending_data_cnt >= PENDING_IO_MAX)
1129                 dispatch_defer_bios(conf, PENDING_IO_ONE_FLUSH, &tmp);
1130
1131         spin_unlock(&conf->pending_bios_lock);
1132
1133         dispatch_bio_list(&tmp);
1134 }
1135
1136 static void
1137 raid5_end_read_request(struct bio *bi);
1138 static void
1139 raid5_end_write_request(struct bio *bi);
1140
1141 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
1142 {
1143         struct r5conf *conf = sh->raid_conf;
1144         int i, disks = sh->disks;
1145         struct stripe_head *head_sh = sh;
1146         struct bio_list pending_bios = BIO_EMPTY_LIST;
1147         struct r5dev *dev;
1148         bool should_defer;
1149
1150         might_sleep();
1151
1152         if (log_stripe(sh, s) == 0)
1153                 return;
1154
1155         should_defer = conf->batch_bio_dispatch && conf->group_cnt;
1156
1157         for (i = disks; i--; ) {
1158                 enum req_op op;
1159                 blk_opf_t op_flags = 0;
1160                 int replace_only = 0;
1161                 struct bio *bi, *rbi;
1162                 struct md_rdev *rdev, *rrdev = NULL;
1163
1164                 sh = head_sh;
1165                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
1166                         op = REQ_OP_WRITE;
1167                         if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
1168                                 op_flags = REQ_FUA;
1169                         if (test_bit(R5_Discard, &sh->dev[i].flags))
1170                                 op = REQ_OP_DISCARD;
1171                 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
1172                         op = REQ_OP_READ;
1173                 else if (test_and_clear_bit(R5_WantReplace,
1174                                             &sh->dev[i].flags)) {
1175                         op = REQ_OP_WRITE;
1176                         replace_only = 1;
1177                 } else
1178                         continue;
1179                 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
1180                         op_flags |= REQ_SYNC;
1181
1182 again:
1183                 dev = &sh->dev[i];
1184                 bi = &dev->req;
1185                 rbi = &dev->rreq; /* For writing to replacement */
1186
1187                 rcu_read_lock();
1188                 rrdev = rcu_dereference(conf->disks[i].replacement);
1189                 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
1190                 rdev = rcu_dereference(conf->disks[i].rdev);
1191                 if (!rdev) {
1192                         rdev = rrdev;
1193                         rrdev = NULL;
1194                 }
1195                 if (op_is_write(op)) {
1196                         if (replace_only)
1197                                 rdev = NULL;
1198                         if (rdev == rrdev)
1199                                 /* We raced and saw duplicates */
1200                                 rrdev = NULL;
1201                 } else {
1202                         if (test_bit(R5_ReadRepl, &head_sh->dev[i].flags) && rrdev)
1203                                 rdev = rrdev;
1204                         rrdev = NULL;
1205                 }
1206
1207                 if (rdev && test_bit(Faulty, &rdev->flags))
1208                         rdev = NULL;
1209                 if (rdev)
1210                         atomic_inc(&rdev->nr_pending);
1211                 if (rrdev && test_bit(Faulty, &rrdev->flags))
1212                         rrdev = NULL;
1213                 if (rrdev)
1214                         atomic_inc(&rrdev->nr_pending);
1215                 rcu_read_unlock();
1216
1217                 /* We have already checked bad blocks for reads.  Now
1218                  * need to check for writes.  We never accept write errors
1219                  * on the replacement, so we don't to check rrdev.
1220                  */
1221                 while (op_is_write(op) && rdev &&
1222                        test_bit(WriteErrorSeen, &rdev->flags)) {
1223                         sector_t first_bad;
1224                         int bad_sectors;
1225                         int bad = is_badblock(rdev, sh->sector, RAID5_STRIPE_SECTORS(conf),
1226                                               &first_bad, &bad_sectors);
1227                         if (!bad)
1228                                 break;
1229
1230                         if (bad < 0) {
1231                                 set_bit(BlockedBadBlocks, &rdev->flags);
1232                                 if (!conf->mddev->external &&
1233                                     conf->mddev->sb_flags) {
1234                                         /* It is very unlikely, but we might
1235                                          * still need to write out the
1236                                          * bad block log - better give it
1237                                          * a chance*/
1238                                         md_check_recovery(conf->mddev);
1239                                 }
1240                                 /*
1241                                  * Because md_wait_for_blocked_rdev
1242                                  * will dec nr_pending, we must
1243                                  * increment it first.
1244                                  */
1245                                 atomic_inc(&rdev->nr_pending);
1246                                 md_wait_for_blocked_rdev(rdev, conf->mddev);
1247                         } else {
1248                                 /* Acknowledged bad block - skip the write */
1249                                 rdev_dec_pending(rdev, conf->mddev);
1250                                 rdev = NULL;
1251                         }
1252                 }
1253
1254                 if (rdev) {
1255                         if (s->syncing || s->expanding || s->expanded
1256                             || s->replacing)
1257                                 md_sync_acct(rdev->bdev, RAID5_STRIPE_SECTORS(conf));
1258
1259                         set_bit(STRIPE_IO_STARTED, &sh->state);
1260
1261                         bio_init(bi, rdev->bdev, &dev->vec, 1, op | op_flags);
1262                         bi->bi_end_io = op_is_write(op)
1263                                 ? raid5_end_write_request
1264                                 : raid5_end_read_request;
1265                         bi->bi_private = sh;
1266
1267                         pr_debug("%s: for %llu schedule op %d on disc %d\n",
1268                                 __func__, (unsigned long long)sh->sector,
1269                                 bi->bi_opf, i);
1270                         atomic_inc(&sh->count);
1271                         if (sh != head_sh)
1272                                 atomic_inc(&head_sh->count);
1273                         if (use_new_offset(conf, sh))
1274                                 bi->bi_iter.bi_sector = (sh->sector
1275                                                  + rdev->new_data_offset);
1276                         else
1277                                 bi->bi_iter.bi_sector = (sh->sector
1278                                                  + rdev->data_offset);
1279                         if (test_bit(R5_ReadNoMerge, &head_sh->dev[i].flags))
1280                                 bi->bi_opf |= REQ_NOMERGE;
1281
1282                         if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1283                                 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1284
1285                         if (!op_is_write(op) &&
1286                             test_bit(R5_InJournal, &sh->dev[i].flags))
1287                                 /*
1288                                  * issuing read for a page in journal, this
1289                                  * must be preparing for prexor in rmw; read
1290                                  * the data into orig_page
1291                                  */
1292                                 sh->dev[i].vec.bv_page = sh->dev[i].orig_page;
1293                         else
1294                                 sh->dev[i].vec.bv_page = sh->dev[i].page;
1295                         bi->bi_vcnt = 1;
1296                         bi->bi_io_vec[0].bv_len = RAID5_STRIPE_SIZE(conf);
1297                         bi->bi_io_vec[0].bv_offset = sh->dev[i].offset;
1298                         bi->bi_iter.bi_size = RAID5_STRIPE_SIZE(conf);
1299                         /*
1300                          * If this is discard request, set bi_vcnt 0. We don't
1301                          * want to confuse SCSI because SCSI will replace payload
1302                          */
1303                         if (op == REQ_OP_DISCARD)
1304                                 bi->bi_vcnt = 0;
1305                         if (rrdev)
1306                                 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
1307
1308                         if (conf->mddev->gendisk)
1309                                 trace_block_bio_remap(bi,
1310                                                 disk_devt(conf->mddev->gendisk),
1311                                                 sh->dev[i].sector);
1312                         if (should_defer && op_is_write(op))
1313                                 bio_list_add(&pending_bios, bi);
1314                         else
1315                                 submit_bio_noacct(bi);
1316                 }
1317                 if (rrdev) {
1318                         if (s->syncing || s->expanding || s->expanded
1319                             || s->replacing)
1320                                 md_sync_acct(rrdev->bdev, RAID5_STRIPE_SECTORS(conf));
1321
1322                         set_bit(STRIPE_IO_STARTED, &sh->state);
1323
1324                         bio_init(rbi, rrdev->bdev, &dev->rvec, 1, op | op_flags);
1325                         BUG_ON(!op_is_write(op));
1326                         rbi->bi_end_io = raid5_end_write_request;
1327                         rbi->bi_private = sh;
1328
1329                         pr_debug("%s: for %llu schedule op %d on "
1330                                  "replacement disc %d\n",
1331                                 __func__, (unsigned long long)sh->sector,
1332                                 rbi->bi_opf, i);
1333                         atomic_inc(&sh->count);
1334                         if (sh != head_sh)
1335                                 atomic_inc(&head_sh->count);
1336                         if (use_new_offset(conf, sh))
1337                                 rbi->bi_iter.bi_sector = (sh->sector
1338                                                   + rrdev->new_data_offset);
1339                         else
1340                                 rbi->bi_iter.bi_sector = (sh->sector
1341                                                   + rrdev->data_offset);
1342                         if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1343                                 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1344                         sh->dev[i].rvec.bv_page = sh->dev[i].page;
1345                         rbi->bi_vcnt = 1;
1346                         rbi->bi_io_vec[0].bv_len = RAID5_STRIPE_SIZE(conf);
1347                         rbi->bi_io_vec[0].bv_offset = sh->dev[i].offset;
1348                         rbi->bi_iter.bi_size = RAID5_STRIPE_SIZE(conf);
1349                         /*
1350                          * If this is discard request, set bi_vcnt 0. We don't
1351                          * want to confuse SCSI because SCSI will replace payload
1352                          */
1353                         if (op == REQ_OP_DISCARD)
1354                                 rbi->bi_vcnt = 0;
1355                         if (conf->mddev->gendisk)
1356                                 trace_block_bio_remap(rbi,
1357                                                 disk_devt(conf->mddev->gendisk),
1358                                                 sh->dev[i].sector);
1359                         if (should_defer && op_is_write(op))
1360                                 bio_list_add(&pending_bios, rbi);
1361                         else
1362                                 submit_bio_noacct(rbi);
1363                 }
1364                 if (!rdev && !rrdev) {
1365                         if (op_is_write(op))
1366                                 set_bit(STRIPE_DEGRADED, &sh->state);
1367                         pr_debug("skip op %d on disc %d for sector %llu\n",
1368                                 bi->bi_opf, i, (unsigned long long)sh->sector);
1369                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1370                         set_bit(STRIPE_HANDLE, &sh->state);
1371                 }
1372
1373                 if (!head_sh->batch_head)
1374                         continue;
1375                 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1376                                       batch_list);
1377                 if (sh != head_sh)
1378                         goto again;
1379         }
1380
1381         if (should_defer && !bio_list_empty(&pending_bios))
1382                 defer_issue_bios(conf, head_sh->sector, &pending_bios);
1383 }
1384
1385 static struct dma_async_tx_descriptor *
1386 async_copy_data(int frombio, struct bio *bio, struct page **page,
1387         unsigned int poff, sector_t sector, struct dma_async_tx_descriptor *tx,
1388         struct stripe_head *sh, int no_skipcopy)
1389 {
1390         struct bio_vec bvl;
1391         struct bvec_iter iter;
1392         struct page *bio_page;
1393         int page_offset;
1394         struct async_submit_ctl submit;
1395         enum async_tx_flags flags = 0;
1396         struct r5conf *conf = sh->raid_conf;
1397
1398         if (bio->bi_iter.bi_sector >= sector)
1399                 page_offset = (signed)(bio->bi_iter.bi_sector - sector) * 512;
1400         else
1401                 page_offset = (signed)(sector - bio->bi_iter.bi_sector) * -512;
1402
1403         if (frombio)
1404                 flags |= ASYNC_TX_FENCE;
1405         init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
1406
1407         bio_for_each_segment(bvl, bio, iter) {
1408                 int len = bvl.bv_len;
1409                 int clen;
1410                 int b_offset = 0;
1411
1412                 if (page_offset < 0) {
1413                         b_offset = -page_offset;
1414                         page_offset += b_offset;
1415                         len -= b_offset;
1416                 }
1417
1418                 if (len > 0 && page_offset + len > RAID5_STRIPE_SIZE(conf))
1419                         clen = RAID5_STRIPE_SIZE(conf) - page_offset;
1420                 else
1421                         clen = len;
1422
1423                 if (clen > 0) {
1424                         b_offset += bvl.bv_offset;
1425                         bio_page = bvl.bv_page;
1426                         if (frombio) {
1427                                 if (conf->skip_copy &&
1428                                     b_offset == 0 && page_offset == 0 &&
1429                                     clen == RAID5_STRIPE_SIZE(conf) &&
1430                                     !no_skipcopy)
1431                                         *page = bio_page;
1432                                 else
1433                                         tx = async_memcpy(*page, bio_page, page_offset + poff,
1434                                                   b_offset, clen, &submit);
1435                         } else
1436                                 tx = async_memcpy(bio_page, *page, b_offset,
1437                                                   page_offset + poff, clen, &submit);
1438                 }
1439                 /* chain the operations */
1440                 submit.depend_tx = tx;
1441
1442                 if (clen < len) /* hit end of page */
1443                         break;
1444                 page_offset +=  len;
1445         }
1446
1447         return tx;
1448 }
1449
1450 static void ops_complete_biofill(void *stripe_head_ref)
1451 {
1452         struct stripe_head *sh = stripe_head_ref;
1453         int i;
1454         struct r5conf *conf = sh->raid_conf;
1455
1456         pr_debug("%s: stripe %llu\n", __func__,
1457                 (unsigned long long)sh->sector);
1458
1459         /* clear completed biofills */
1460         for (i = sh->disks; i--; ) {
1461                 struct r5dev *dev = &sh->dev[i];
1462
1463                 /* acknowledge completion of a biofill operation */
1464                 /* and check if we need to reply to a read request,
1465                  * new R5_Wantfill requests are held off until
1466                  * !STRIPE_BIOFILL_RUN
1467                  */
1468                 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
1469                         struct bio *rbi, *rbi2;
1470
1471                         BUG_ON(!dev->read);
1472                         rbi = dev->read;
1473                         dev->read = NULL;
1474                         while (rbi && rbi->bi_iter.bi_sector <
1475                                 dev->sector + RAID5_STRIPE_SECTORS(conf)) {
1476                                 rbi2 = r5_next_bio(conf, rbi, dev->sector);
1477                                 bio_endio(rbi);
1478                                 rbi = rbi2;
1479                         }
1480                 }
1481         }
1482         clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
1483
1484         set_bit(STRIPE_HANDLE, &sh->state);
1485         raid5_release_stripe(sh);
1486 }
1487
1488 static void ops_run_biofill(struct stripe_head *sh)
1489 {
1490         struct dma_async_tx_descriptor *tx = NULL;
1491         struct async_submit_ctl submit;
1492         int i;
1493         struct r5conf *conf = sh->raid_conf;
1494
1495         BUG_ON(sh->batch_head);
1496         pr_debug("%s: stripe %llu\n", __func__,
1497                 (unsigned long long)sh->sector);
1498
1499         for (i = sh->disks; i--; ) {
1500                 struct r5dev *dev = &sh->dev[i];
1501                 if (test_bit(R5_Wantfill, &dev->flags)) {
1502                         struct bio *rbi;
1503                         spin_lock_irq(&sh->stripe_lock);
1504                         dev->read = rbi = dev->toread;
1505                         dev->toread = NULL;
1506                         spin_unlock_irq(&sh->stripe_lock);
1507                         while (rbi && rbi->bi_iter.bi_sector <
1508                                 dev->sector + RAID5_STRIPE_SECTORS(conf)) {
1509                                 tx = async_copy_data(0, rbi, &dev->page,
1510                                                      dev->offset,
1511                                                      dev->sector, tx, sh, 0);
1512                                 rbi = r5_next_bio(conf, rbi, dev->sector);
1513                         }
1514                 }
1515         }
1516
1517         atomic_inc(&sh->count);
1518         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
1519         async_trigger_callback(&submit);
1520 }
1521
1522 static void mark_target_uptodate(struct stripe_head *sh, int target)
1523 {
1524         struct r5dev *tgt;
1525
1526         if (target < 0)
1527                 return;
1528
1529         tgt = &sh->dev[target];
1530         set_bit(R5_UPTODATE, &tgt->flags);
1531         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1532         clear_bit(R5_Wantcompute, &tgt->flags);
1533 }
1534
1535 static void ops_complete_compute(void *stripe_head_ref)
1536 {
1537         struct stripe_head *sh = stripe_head_ref;
1538
1539         pr_debug("%s: stripe %llu\n", __func__,
1540                 (unsigned long long)sh->sector);
1541
1542         /* mark the computed target(s) as uptodate */
1543         mark_target_uptodate(sh, sh->ops.target);
1544         mark_target_uptodate(sh, sh->ops.target2);
1545
1546         clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
1547         if (sh->check_state == check_state_compute_run)
1548                 sh->check_state = check_state_compute_result;
1549         set_bit(STRIPE_HANDLE, &sh->state);
1550         raid5_release_stripe(sh);
1551 }
1552
1553 /* return a pointer to the address conversion region of the scribble buffer */
1554 static struct page **to_addr_page(struct raid5_percpu *percpu, int i)
1555 {
1556         return percpu->scribble + i * percpu->scribble_obj_size;
1557 }
1558
1559 /* return a pointer to the address conversion region of the scribble buffer */
1560 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
1561                                  struct raid5_percpu *percpu, int i)
1562 {
1563         return (void *) (to_addr_page(percpu, i) + sh->disks + 2);
1564 }
1565
1566 /*
1567  * Return a pointer to record offset address.
1568  */
1569 static unsigned int *
1570 to_addr_offs(struct stripe_head *sh, struct raid5_percpu *percpu)
1571 {
1572         return (unsigned int *) (to_addr_conv(sh, percpu, 0) + sh->disks + 2);
1573 }
1574
1575 static struct dma_async_tx_descriptor *
1576 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
1577 {
1578         int disks = sh->disks;
1579         struct page **xor_srcs = to_addr_page(percpu, 0);
1580         unsigned int *off_srcs = to_addr_offs(sh, percpu);
1581         int target = sh->ops.target;
1582         struct r5dev *tgt = &sh->dev[target];
1583         struct page *xor_dest = tgt->page;
1584         unsigned int off_dest = tgt->offset;
1585         int count = 0;
1586         struct dma_async_tx_descriptor *tx;
1587         struct async_submit_ctl submit;
1588         int i;
1589
1590         BUG_ON(sh->batch_head);
1591
1592         pr_debug("%s: stripe %llu block: %d\n",
1593                 __func__, (unsigned long long)sh->sector, target);
1594         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1595
1596         for (i = disks; i--; ) {
1597                 if (i != target) {
1598                         off_srcs[count] = sh->dev[i].offset;
1599                         xor_srcs[count++] = sh->dev[i].page;
1600                 }
1601         }
1602
1603         atomic_inc(&sh->count);
1604
1605         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
1606                           ops_complete_compute, sh, to_addr_conv(sh, percpu, 0));
1607         if (unlikely(count == 1))
1608                 tx = async_memcpy(xor_dest, xor_srcs[0], off_dest, off_srcs[0],
1609                                 RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1610         else
1611                 tx = async_xor_offs(xor_dest, off_dest, xor_srcs, off_srcs, count,
1612                                 RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1613
1614         return tx;
1615 }
1616
1617 /* set_syndrome_sources - populate source buffers for gen_syndrome
1618  * @srcs - (struct page *) array of size sh->disks
1619  * @offs - (unsigned int) array of offset for each page
1620  * @sh - stripe_head to parse
1621  *
1622  * Populates srcs in proper layout order for the stripe and returns the
1623  * 'count' of sources to be used in a call to async_gen_syndrome.  The P
1624  * destination buffer is recorded in srcs[count] and the Q destination
1625  * is recorded in srcs[count+1]].
1626  */
1627 static int set_syndrome_sources(struct page **srcs,
1628                                 unsigned int *offs,
1629                                 struct stripe_head *sh,
1630                                 int srctype)
1631 {
1632         int disks = sh->disks;
1633         int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1634         int d0_idx = raid6_d0(sh);
1635         int count;
1636         int i;
1637
1638         for (i = 0; i < disks; i++)
1639                 srcs[i] = NULL;
1640
1641         count = 0;
1642         i = d0_idx;
1643         do {
1644                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1645                 struct r5dev *dev = &sh->dev[i];
1646
1647                 if (i == sh->qd_idx || i == sh->pd_idx ||
1648                     (srctype == SYNDROME_SRC_ALL) ||
1649                     (srctype == SYNDROME_SRC_WANT_DRAIN &&
1650                      (test_bit(R5_Wantdrain, &dev->flags) ||
1651                       test_bit(R5_InJournal, &dev->flags))) ||
1652                     (srctype == SYNDROME_SRC_WRITTEN &&
1653                      (dev->written ||
1654                       test_bit(R5_InJournal, &dev->flags)))) {
1655                         if (test_bit(R5_InJournal, &dev->flags))
1656                                 srcs[slot] = sh->dev[i].orig_page;
1657                         else
1658                                 srcs[slot] = sh->dev[i].page;
1659                         /*
1660                          * For R5_InJournal, PAGE_SIZE must be 4KB and will
1661                          * not shared page. In that case, dev[i].offset
1662                          * is 0.
1663                          */
1664                         offs[slot] = sh->dev[i].offset;
1665                 }
1666                 i = raid6_next_disk(i, disks);
1667         } while (i != d0_idx);
1668
1669         return syndrome_disks;
1670 }
1671
1672 static struct dma_async_tx_descriptor *
1673 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
1674 {
1675         int disks = sh->disks;
1676         struct page **blocks = to_addr_page(percpu, 0);
1677         unsigned int *offs = to_addr_offs(sh, percpu);
1678         int target;
1679         int qd_idx = sh->qd_idx;
1680         struct dma_async_tx_descriptor *tx;
1681         struct async_submit_ctl submit;
1682         struct r5dev *tgt;
1683         struct page *dest;
1684         unsigned int dest_off;
1685         int i;
1686         int count;
1687
1688         BUG_ON(sh->batch_head);
1689         if (sh->ops.target < 0)
1690                 target = sh->ops.target2;
1691         else if (sh->ops.target2 < 0)
1692                 target = sh->ops.target;
1693         else
1694                 /* we should only have one valid target */
1695                 BUG();
1696         BUG_ON(target < 0);
1697         pr_debug("%s: stripe %llu block: %d\n",
1698                 __func__, (unsigned long long)sh->sector, target);
1699
1700         tgt = &sh->dev[target];
1701         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1702         dest = tgt->page;
1703         dest_off = tgt->offset;
1704
1705         atomic_inc(&sh->count);
1706
1707         if (target == qd_idx) {
1708                 count = set_syndrome_sources(blocks, offs, sh, SYNDROME_SRC_ALL);
1709                 blocks[count] = NULL; /* regenerating p is not necessary */
1710                 BUG_ON(blocks[count+1] != dest); /* q should already be set */
1711                 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1712                                   ops_complete_compute, sh,
1713                                   to_addr_conv(sh, percpu, 0));
1714                 tx = async_gen_syndrome(blocks, offs, count+2,
1715                                 RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1716         } else {
1717                 /* Compute any data- or p-drive using XOR */
1718                 count = 0;
1719                 for (i = disks; i-- ; ) {
1720                         if (i == target || i == qd_idx)
1721                                 continue;
1722                         offs[count] = sh->dev[i].offset;
1723                         blocks[count++] = sh->dev[i].page;
1724                 }
1725
1726                 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1727                                   NULL, ops_complete_compute, sh,
1728                                   to_addr_conv(sh, percpu, 0));
1729                 tx = async_xor_offs(dest, dest_off, blocks, offs, count,
1730                                 RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1731         }
1732
1733         return tx;
1734 }
1735
1736 static struct dma_async_tx_descriptor *
1737 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1738 {
1739         int i, count, disks = sh->disks;
1740         int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1741         int d0_idx = raid6_d0(sh);
1742         int faila = -1, failb = -1;
1743         int target = sh->ops.target;
1744         int target2 = sh->ops.target2;
1745         struct r5dev *tgt = &sh->dev[target];
1746         struct r5dev *tgt2 = &sh->dev[target2];
1747         struct dma_async_tx_descriptor *tx;
1748         struct page **blocks = to_addr_page(percpu, 0);
1749         unsigned int *offs = to_addr_offs(sh, percpu);
1750         struct async_submit_ctl submit;
1751
1752         BUG_ON(sh->batch_head);
1753         pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1754                  __func__, (unsigned long long)sh->sector, target, target2);
1755         BUG_ON(target < 0 || target2 < 0);
1756         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1757         BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1758
1759         /* we need to open-code set_syndrome_sources to handle the
1760          * slot number conversion for 'faila' and 'failb'
1761          */
1762         for (i = 0; i < disks ; i++) {
1763                 offs[i] = 0;
1764                 blocks[i] = NULL;
1765         }
1766         count = 0;
1767         i = d0_idx;
1768         do {
1769                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1770
1771                 offs[slot] = sh->dev[i].offset;
1772                 blocks[slot] = sh->dev[i].page;
1773
1774                 if (i == target)
1775                         faila = slot;
1776                 if (i == target2)
1777                         failb = slot;
1778                 i = raid6_next_disk(i, disks);
1779         } while (i != d0_idx);
1780
1781         BUG_ON(faila == failb);
1782         if (failb < faila)
1783                 swap(faila, failb);
1784         pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1785                  __func__, (unsigned long long)sh->sector, faila, failb);
1786
1787         atomic_inc(&sh->count);
1788
1789         if (failb == syndrome_disks+1) {
1790                 /* Q disk is one of the missing disks */
1791                 if (faila == syndrome_disks) {
1792                         /* Missing P+Q, just recompute */
1793                         init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1794                                           ops_complete_compute, sh,
1795                                           to_addr_conv(sh, percpu, 0));
1796                         return async_gen_syndrome(blocks, offs, syndrome_disks+2,
1797                                                   RAID5_STRIPE_SIZE(sh->raid_conf),
1798                                                   &submit);
1799                 } else {
1800                         struct page *dest;
1801                         unsigned int dest_off;
1802                         int data_target;
1803                         int qd_idx = sh->qd_idx;
1804
1805                         /* Missing D+Q: recompute D from P, then recompute Q */
1806                         if (target == qd_idx)
1807                                 data_target = target2;
1808                         else
1809                                 data_target = target;
1810
1811                         count = 0;
1812                         for (i = disks; i-- ; ) {
1813                                 if (i == data_target || i == qd_idx)
1814                                         continue;
1815                                 offs[count] = sh->dev[i].offset;
1816                                 blocks[count++] = sh->dev[i].page;
1817                         }
1818                         dest = sh->dev[data_target].page;
1819                         dest_off = sh->dev[data_target].offset;
1820                         init_async_submit(&submit,
1821                                           ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1822                                           NULL, NULL, NULL,
1823                                           to_addr_conv(sh, percpu, 0));
1824                         tx = async_xor_offs(dest, dest_off, blocks, offs, count,
1825                                        RAID5_STRIPE_SIZE(sh->raid_conf),
1826                                        &submit);
1827
1828                         count = set_syndrome_sources(blocks, offs, sh, SYNDROME_SRC_ALL);
1829                         init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1830                                           ops_complete_compute, sh,
1831                                           to_addr_conv(sh, percpu, 0));
1832                         return async_gen_syndrome(blocks, offs, count+2,
1833                                                   RAID5_STRIPE_SIZE(sh->raid_conf),
1834                                                   &submit);
1835                 }
1836         } else {
1837                 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1838                                   ops_complete_compute, sh,
1839                                   to_addr_conv(sh, percpu, 0));
1840                 if (failb == syndrome_disks) {
1841                         /* We're missing D+P. */
1842                         return async_raid6_datap_recov(syndrome_disks+2,
1843                                                 RAID5_STRIPE_SIZE(sh->raid_conf),
1844                                                 faila,
1845                                                 blocks, offs, &submit);
1846                 } else {
1847                         /* We're missing D+D. */
1848                         return async_raid6_2data_recov(syndrome_disks+2,
1849                                                 RAID5_STRIPE_SIZE(sh->raid_conf),
1850                                                 faila, failb,
1851                                                 blocks, offs, &submit);
1852                 }
1853         }
1854 }
1855
1856 static void ops_complete_prexor(void *stripe_head_ref)
1857 {
1858         struct stripe_head *sh = stripe_head_ref;
1859
1860         pr_debug("%s: stripe %llu\n", __func__,
1861                 (unsigned long long)sh->sector);
1862
1863         if (r5c_is_writeback(sh->raid_conf->log))
1864                 /*
1865                  * raid5-cache write back uses orig_page during prexor.
1866                  * After prexor, it is time to free orig_page
1867                  */
1868                 r5c_release_extra_page(sh);
1869 }
1870
1871 static struct dma_async_tx_descriptor *
1872 ops_run_prexor5(struct stripe_head *sh, struct raid5_percpu *percpu,
1873                 struct dma_async_tx_descriptor *tx)
1874 {
1875         int disks = sh->disks;
1876         struct page **xor_srcs = to_addr_page(percpu, 0);
1877         unsigned int *off_srcs = to_addr_offs(sh, percpu);
1878         int count = 0, pd_idx = sh->pd_idx, i;
1879         struct async_submit_ctl submit;
1880
1881         /* existing parity data subtracted */
1882         unsigned int off_dest = off_srcs[count] = sh->dev[pd_idx].offset;
1883         struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1884
1885         BUG_ON(sh->batch_head);
1886         pr_debug("%s: stripe %llu\n", __func__,
1887                 (unsigned long long)sh->sector);
1888
1889         for (i = disks; i--; ) {
1890                 struct r5dev *dev = &sh->dev[i];
1891                 /* Only process blocks that are known to be uptodate */
1892                 if (test_bit(R5_InJournal, &dev->flags)) {
1893                         /*
1894                          * For this case, PAGE_SIZE must be equal to 4KB and
1895                          * page offset is zero.
1896                          */
1897                         off_srcs[count] = dev->offset;
1898                         xor_srcs[count++] = dev->orig_page;
1899                 } else if (test_bit(R5_Wantdrain, &dev->flags)) {
1900                         off_srcs[count] = dev->offset;
1901                         xor_srcs[count++] = dev->page;
1902                 }
1903         }
1904
1905         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1906                           ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1907         tx = async_xor_offs(xor_dest, off_dest, xor_srcs, off_srcs, count,
1908                         RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1909
1910         return tx;
1911 }
1912
1913 static struct dma_async_tx_descriptor *
1914 ops_run_prexor6(struct stripe_head *sh, struct raid5_percpu *percpu,
1915                 struct dma_async_tx_descriptor *tx)
1916 {
1917         struct page **blocks = to_addr_page(percpu, 0);
1918         unsigned int *offs = to_addr_offs(sh, percpu);
1919         int count;
1920         struct async_submit_ctl submit;
1921
1922         pr_debug("%s: stripe %llu\n", __func__,
1923                 (unsigned long long)sh->sector);
1924
1925         count = set_syndrome_sources(blocks, offs, sh, SYNDROME_SRC_WANT_DRAIN);
1926
1927         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_PQ_XOR_DST, tx,
1928                           ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1929         tx = async_gen_syndrome(blocks, offs, count+2,
1930                         RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1931
1932         return tx;
1933 }
1934
1935 static struct dma_async_tx_descriptor *
1936 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1937 {
1938         struct r5conf *conf = sh->raid_conf;
1939         int disks = sh->disks;
1940         int i;
1941         struct stripe_head *head_sh = sh;
1942
1943         pr_debug("%s: stripe %llu\n", __func__,
1944                 (unsigned long long)sh->sector);
1945
1946         for (i = disks; i--; ) {
1947                 struct r5dev *dev;
1948                 struct bio *chosen;
1949
1950                 sh = head_sh;
1951                 if (test_and_clear_bit(R5_Wantdrain, &head_sh->dev[i].flags)) {
1952                         struct bio *wbi;
1953
1954 again:
1955                         dev = &sh->dev[i];
1956                         /*
1957                          * clear R5_InJournal, so when rewriting a page in
1958                          * journal, it is not skipped by r5l_log_stripe()
1959                          */
1960                         clear_bit(R5_InJournal, &dev->flags);
1961                         spin_lock_irq(&sh->stripe_lock);
1962                         chosen = dev->towrite;
1963                         dev->towrite = NULL;
1964                         sh->overwrite_disks = 0;
1965                         BUG_ON(dev->written);
1966                         wbi = dev->written = chosen;
1967                         spin_unlock_irq(&sh->stripe_lock);
1968                         WARN_ON(dev->page != dev->orig_page);
1969
1970                         while (wbi && wbi->bi_iter.bi_sector <
1971                                 dev->sector + RAID5_STRIPE_SECTORS(conf)) {
1972                                 if (wbi->bi_opf & REQ_FUA)
1973                                         set_bit(R5_WantFUA, &dev->flags);
1974                                 if (wbi->bi_opf & REQ_SYNC)
1975                                         set_bit(R5_SyncIO, &dev->flags);
1976                                 if (bio_op(wbi) == REQ_OP_DISCARD)
1977                                         set_bit(R5_Discard, &dev->flags);
1978                                 else {
1979                                         tx = async_copy_data(1, wbi, &dev->page,
1980                                                              dev->offset,
1981                                                              dev->sector, tx, sh,
1982                                                              r5c_is_writeback(conf->log));
1983                                         if (dev->page != dev->orig_page &&
1984                                             !r5c_is_writeback(conf->log)) {
1985                                                 set_bit(R5_SkipCopy, &dev->flags);
1986                                                 clear_bit(R5_UPTODATE, &dev->flags);
1987                                                 clear_bit(R5_OVERWRITE, &dev->flags);
1988                                         }
1989                                 }
1990                                 wbi = r5_next_bio(conf, wbi, dev->sector);
1991                         }
1992
1993                         if (head_sh->batch_head) {
1994                                 sh = list_first_entry(&sh->batch_list,
1995                                                       struct stripe_head,
1996                                                       batch_list);
1997                                 if (sh == head_sh)
1998                                         continue;
1999                                 goto again;
2000                         }
2001                 }
2002         }
2003
2004         return tx;
2005 }
2006
2007 static void ops_complete_reconstruct(void *stripe_head_ref)
2008 {
2009         struct stripe_head *sh = stripe_head_ref;
2010         int disks = sh->disks;
2011         int pd_idx = sh->pd_idx;
2012         int qd_idx = sh->qd_idx;
2013         int i;
2014         bool fua = false, sync = false, discard = false;
2015
2016         pr_debug("%s: stripe %llu\n", __func__,
2017                 (unsigned long long)sh->sector);
2018
2019         for (i = disks; i--; ) {
2020                 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
2021                 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
2022                 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
2023         }
2024
2025         for (i = disks; i--; ) {
2026                 struct r5dev *dev = &sh->dev[i];
2027
2028                 if (dev->written || i == pd_idx || i == qd_idx) {
2029                         if (!discard && !test_bit(R5_SkipCopy, &dev->flags)) {
2030                                 set_bit(R5_UPTODATE, &dev->flags);
2031                                 if (test_bit(STRIPE_EXPAND_READY, &sh->state))
2032                                         set_bit(R5_Expanded, &dev->flags);
2033                         }
2034                         if (fua)
2035                                 set_bit(R5_WantFUA, &dev->flags);
2036                         if (sync)
2037                                 set_bit(R5_SyncIO, &dev->flags);
2038                 }
2039         }
2040
2041         if (sh->reconstruct_state == reconstruct_state_drain_run)
2042                 sh->reconstruct_state = reconstruct_state_drain_result;
2043         else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
2044                 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
2045         else {
2046                 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
2047                 sh->reconstruct_state = reconstruct_state_result;
2048         }
2049
2050         set_bit(STRIPE_HANDLE, &sh->state);
2051         raid5_release_stripe(sh);
2052 }
2053
2054 static void
2055 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
2056                      struct dma_async_tx_descriptor *tx)
2057 {
2058         int disks = sh->disks;
2059         struct page **xor_srcs;
2060         unsigned int *off_srcs;
2061         struct async_submit_ctl submit;
2062         int count, pd_idx = sh->pd_idx, i;
2063         struct page *xor_dest;
2064         unsigned int off_dest;
2065         int prexor = 0;
2066         unsigned long flags;
2067         int j = 0;
2068         struct stripe_head *head_sh = sh;
2069         int last_stripe;
2070
2071         pr_debug("%s: stripe %llu\n", __func__,
2072                 (unsigned long long)sh->sector);
2073
2074         for (i = 0; i < sh->disks; i++) {
2075                 if (pd_idx == i)
2076                         continue;
2077                 if (!test_bit(R5_Discard, &sh->dev[i].flags))
2078                         break;
2079         }
2080         if (i >= sh->disks) {
2081                 atomic_inc(&sh->count);
2082                 set_bit(R5_Discard, &sh->dev[pd_idx].flags);
2083                 ops_complete_reconstruct(sh);
2084                 return;
2085         }
2086 again:
2087         count = 0;
2088         xor_srcs = to_addr_page(percpu, j);
2089         off_srcs = to_addr_offs(sh, percpu);
2090         /* check if prexor is active which means only process blocks
2091          * that are part of a read-modify-write (written)
2092          */
2093         if (head_sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
2094                 prexor = 1;
2095                 off_dest = off_srcs[count] = sh->dev[pd_idx].offset;
2096                 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
2097                 for (i = disks; i--; ) {
2098                         struct r5dev *dev = &sh->dev[i];
2099                         if (head_sh->dev[i].written ||
2100                             test_bit(R5_InJournal, &head_sh->dev[i].flags)) {
2101                                 off_srcs[count] = dev->offset;
2102                                 xor_srcs[count++] = dev->page;
2103                         }
2104                 }
2105         } else {
2106                 xor_dest = sh->dev[pd_idx].page;
2107                 off_dest = sh->dev[pd_idx].offset;
2108                 for (i = disks; i--; ) {
2109                         struct r5dev *dev = &sh->dev[i];
2110                         if (i != pd_idx) {
2111                                 off_srcs[count] = dev->offset;
2112                                 xor_srcs[count++] = dev->page;
2113                         }
2114                 }
2115         }
2116
2117         /* 1/ if we prexor'd then the dest is reused as a source
2118          * 2/ if we did not prexor then we are redoing the parity
2119          * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
2120          * for the synchronous xor case
2121          */
2122         last_stripe = !head_sh->batch_head ||
2123                 list_first_entry(&sh->batch_list,
2124                                  struct stripe_head, batch_list) == head_sh;
2125         if (last_stripe) {
2126                 flags = ASYNC_TX_ACK |
2127                         (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
2128
2129                 atomic_inc(&head_sh->count);
2130                 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, head_sh,
2131                                   to_addr_conv(sh, percpu, j));
2132         } else {
2133                 flags = prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST;
2134                 init_async_submit(&submit, flags, tx, NULL, NULL,
2135                                   to_addr_conv(sh, percpu, j));
2136         }
2137
2138         if (unlikely(count == 1))
2139                 tx = async_memcpy(xor_dest, xor_srcs[0], off_dest, off_srcs[0],
2140                                 RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
2141         else
2142                 tx = async_xor_offs(xor_dest, off_dest, xor_srcs, off_srcs, count,
2143                                 RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
2144         if (!last_stripe) {
2145                 j++;
2146                 sh = list_first_entry(&sh->batch_list, struct stripe_head,
2147                                       batch_list);
2148                 goto again;
2149         }
2150 }
2151
2152 static void
2153 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
2154                      struct dma_async_tx_descriptor *tx)
2155 {
2156         struct async_submit_ctl submit;
2157         struct page **blocks;
2158         unsigned int *offs;
2159         int count, i, j = 0;
2160         struct stripe_head *head_sh = sh;
2161         int last_stripe;
2162         int synflags;
2163         unsigned long txflags;
2164
2165         pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
2166
2167         for (i = 0; i < sh->disks; i++) {
2168                 if (sh->pd_idx == i || sh->qd_idx == i)
2169                         continue;
2170                 if (!test_bit(R5_Discard, &sh->dev[i].flags))
2171                         break;
2172         }
2173         if (i >= sh->disks) {
2174                 atomic_inc(&sh->count);
2175                 set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
2176                 set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
2177                 ops_complete_reconstruct(sh);
2178                 return;
2179         }
2180
2181 again:
2182         blocks = to_addr_page(percpu, j);
2183         offs = to_addr_offs(sh, percpu);
2184
2185         if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
2186                 synflags = SYNDROME_SRC_WRITTEN;
2187                 txflags = ASYNC_TX_ACK | ASYNC_TX_PQ_XOR_DST;
2188         } else {
2189                 synflags = SYNDROME_SRC_ALL;
2190                 txflags = ASYNC_TX_ACK;
2191         }
2192
2193         count = set_syndrome_sources(blocks, offs, sh, synflags);
2194         last_stripe = !head_sh->batch_head ||
2195                 list_first_entry(&sh->batch_list,
2196                                  struct stripe_head, batch_list) == head_sh;
2197
2198         if (last_stripe) {
2199                 atomic_inc(&head_sh->count);
2200                 init_async_submit(&submit, txflags, tx, ops_complete_reconstruct,
2201                                   head_sh, to_addr_conv(sh, percpu, j));
2202         } else
2203                 init_async_submit(&submit, 0, tx, NULL, NULL,
2204                                   to_addr_conv(sh, percpu, j));
2205         tx = async_gen_syndrome(blocks, offs, count+2,
2206                         RAID5_STRIPE_SIZE(sh->raid_conf),  &submit);
2207         if (!last_stripe) {
2208                 j++;
2209                 sh = list_first_entry(&sh->batch_list, struct stripe_head,
2210                                       batch_list);
2211                 goto again;
2212         }
2213 }
2214
2215 static void ops_complete_check(void *stripe_head_ref)
2216 {
2217         struct stripe_head *sh = stripe_head_ref;
2218
2219         pr_debug("%s: stripe %llu\n", __func__,
2220                 (unsigned long long)sh->sector);
2221
2222         sh->check_state = check_state_check_result;
2223         set_bit(STRIPE_HANDLE, &sh->state);
2224         raid5_release_stripe(sh);
2225 }
2226
2227 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
2228 {
2229         int disks = sh->disks;
2230         int pd_idx = sh->pd_idx;
2231         int qd_idx = sh->qd_idx;
2232         struct page *xor_dest;
2233         unsigned int off_dest;
2234         struct page **xor_srcs = to_addr_page(percpu, 0);
2235         unsigned int *off_srcs = to_addr_offs(sh, percpu);
2236         struct dma_async_tx_descriptor *tx;
2237         struct async_submit_ctl submit;
2238         int count;
2239         int i;
2240
2241         pr_debug("%s: stripe %llu\n", __func__,
2242                 (unsigned long long)sh->sector);
2243
2244         BUG_ON(sh->batch_head);
2245         count = 0;
2246         xor_dest = sh->dev[pd_idx].page;
2247         off_dest = sh->dev[pd_idx].offset;
2248         off_srcs[count] = off_dest;
2249         xor_srcs[count++] = xor_dest;
2250         for (i = disks; i--; ) {
2251                 if (i == pd_idx || i == qd_idx)
2252                         continue;
2253                 off_srcs[count] = sh->dev[i].offset;
2254                 xor_srcs[count++] = sh->dev[i].page;
2255         }
2256
2257         init_async_submit(&submit, 0, NULL, NULL, NULL,
2258                           to_addr_conv(sh, percpu, 0));
2259         tx = async_xor_val_offs(xor_dest, off_dest, xor_srcs, off_srcs, count,
2260                            RAID5_STRIPE_SIZE(sh->raid_conf),
2261                            &sh->ops.zero_sum_result, &submit);
2262
2263         atomic_inc(&sh->count);
2264         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
2265         tx = async_trigger_callback(&submit);
2266 }
2267
2268 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
2269 {
2270         struct page **srcs = to_addr_page(percpu, 0);
2271         unsigned int *offs = to_addr_offs(sh, percpu);
2272         struct async_submit_ctl submit;
2273         int count;
2274
2275         pr_debug("%s: stripe %llu checkp: %d\n", __func__,
2276                 (unsigned long long)sh->sector, checkp);
2277
2278         BUG_ON(sh->batch_head);
2279         count = set_syndrome_sources(srcs, offs, sh, SYNDROME_SRC_ALL);
2280         if (!checkp)
2281                 srcs[count] = NULL;
2282
2283         atomic_inc(&sh->count);
2284         init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
2285                           sh, to_addr_conv(sh, percpu, 0));
2286         async_syndrome_val(srcs, offs, count+2,
2287                            RAID5_STRIPE_SIZE(sh->raid_conf),
2288                            &sh->ops.zero_sum_result, percpu->spare_page, 0, &submit);
2289 }
2290
2291 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
2292 {
2293         int overlap_clear = 0, i, disks = sh->disks;
2294         struct dma_async_tx_descriptor *tx = NULL;
2295         struct r5conf *conf = sh->raid_conf;
2296         int level = conf->level;
2297         struct raid5_percpu *percpu;
2298
2299         local_lock(&conf->percpu->lock);
2300         percpu = this_cpu_ptr(conf->percpu);
2301         if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
2302                 ops_run_biofill(sh);
2303                 overlap_clear++;
2304         }
2305
2306         if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
2307                 if (level < 6)
2308                         tx = ops_run_compute5(sh, percpu);
2309                 else {
2310                         if (sh->ops.target2 < 0 || sh->ops.target < 0)
2311                                 tx = ops_run_compute6_1(sh, percpu);
2312                         else
2313                                 tx = ops_run_compute6_2(sh, percpu);
2314                 }
2315                 /* terminate the chain if reconstruct is not set to be run */
2316                 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
2317                         async_tx_ack(tx);
2318         }
2319
2320         if (test_bit(STRIPE_OP_PREXOR, &ops_request)) {
2321                 if (level < 6)
2322                         tx = ops_run_prexor5(sh, percpu, tx);
2323                 else
2324                         tx = ops_run_prexor6(sh, percpu, tx);
2325         }
2326
2327         if (test_bit(STRIPE_OP_PARTIAL_PARITY, &ops_request))
2328                 tx = ops_run_partial_parity(sh, percpu, tx);
2329
2330         if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
2331                 tx = ops_run_biodrain(sh, tx);
2332                 overlap_clear++;
2333         }
2334
2335         if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
2336                 if (level < 6)
2337                         ops_run_reconstruct5(sh, percpu, tx);
2338                 else
2339                         ops_run_reconstruct6(sh, percpu, tx);
2340         }
2341
2342         if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
2343                 if (sh->check_state == check_state_run)
2344                         ops_run_check_p(sh, percpu);
2345                 else if (sh->check_state == check_state_run_q)
2346                         ops_run_check_pq(sh, percpu, 0);
2347                 else if (sh->check_state == check_state_run_pq)
2348                         ops_run_check_pq(sh, percpu, 1);
2349                 else
2350                         BUG();
2351         }
2352
2353         if (overlap_clear && !sh->batch_head) {
2354                 for (i = disks; i--; ) {
2355                         struct r5dev *dev = &sh->dev[i];
2356                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
2357                                 wake_up(&sh->raid_conf->wait_for_overlap);
2358                 }
2359         }
2360         local_unlock(&conf->percpu->lock);
2361 }
2362
2363 static void free_stripe(struct kmem_cache *sc, struct stripe_head *sh)
2364 {
2365 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
2366         kfree(sh->pages);
2367 #endif
2368         if (sh->ppl_page)
2369                 __free_page(sh->ppl_page);
2370         kmem_cache_free(sc, sh);
2371 }
2372
2373 static struct stripe_head *alloc_stripe(struct kmem_cache *sc, gfp_t gfp,
2374         int disks, struct r5conf *conf)
2375 {
2376         struct stripe_head *sh;
2377
2378         sh = kmem_cache_zalloc(sc, gfp);
2379         if (sh) {
2380                 spin_lock_init(&sh->stripe_lock);
2381                 spin_lock_init(&sh->batch_lock);
2382                 INIT_LIST_HEAD(&sh->batch_list);
2383                 INIT_LIST_HEAD(&sh->lru);
2384                 INIT_LIST_HEAD(&sh->r5c);
2385                 INIT_LIST_HEAD(&sh->log_list);
2386                 atomic_set(&sh->count, 1);
2387                 sh->raid_conf = conf;
2388                 sh->log_start = MaxSector;
2389
2390                 if (raid5_has_ppl(conf)) {
2391                         sh->ppl_page = alloc_page(gfp);
2392                         if (!sh->ppl_page) {
2393                                 free_stripe(sc, sh);
2394                                 return NULL;
2395                         }
2396                 }
2397 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
2398                 if (init_stripe_shared_pages(sh, conf, disks)) {
2399                         free_stripe(sc, sh);
2400                         return NULL;
2401                 }
2402 #endif
2403         }
2404         return sh;
2405 }
2406 static int grow_one_stripe(struct r5conf *conf, gfp_t gfp)
2407 {
2408         struct stripe_head *sh;
2409
2410         sh = alloc_stripe(conf->slab_cache, gfp, conf->pool_size, conf);
2411         if (!sh)
2412                 return 0;
2413
2414         if (grow_buffers(sh, gfp)) {
2415                 shrink_buffers(sh);
2416                 free_stripe(conf->slab_cache, sh);
2417                 return 0;
2418         }
2419         sh->hash_lock_index =
2420                 conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
2421         /* we just created an active stripe so... */
2422         atomic_inc(&conf->active_stripes);
2423
2424         raid5_release_stripe(sh);
2425         conf->max_nr_stripes++;
2426         return 1;
2427 }
2428
2429 static int grow_stripes(struct r5conf *conf, int num)
2430 {
2431         struct kmem_cache *sc;
2432         size_t namelen = sizeof(conf->cache_name[0]);
2433         int devs = max(conf->raid_disks, conf->previous_raid_disks);
2434
2435         if (conf->mddev->gendisk)
2436                 snprintf(conf->cache_name[0], namelen,
2437                         "raid%d-%s", conf->level, mdname(conf->mddev));
2438         else
2439                 snprintf(conf->cache_name[0], namelen,
2440                         "raid%d-%p", conf->level, conf->mddev);
2441         snprintf(conf->cache_name[1], namelen, "%.27s-alt", conf->cache_name[0]);
2442
2443         conf->active_name = 0;
2444         sc = kmem_cache_create(conf->cache_name[conf->active_name],
2445                                struct_size_t(struct stripe_head, dev, devs),
2446                                0, 0, NULL);
2447         if (!sc)
2448                 return 1;
2449         conf->slab_cache = sc;
2450         conf->pool_size = devs;
2451         while (num--)
2452                 if (!grow_one_stripe(conf, GFP_KERNEL))
2453                         return 1;
2454
2455         return 0;
2456 }
2457
2458 /**
2459  * scribble_alloc - allocate percpu scribble buffer for required size
2460  *                  of the scribble region
2461  * @percpu: from for_each_present_cpu() of the caller
2462  * @num: total number of disks in the array
2463  * @cnt: scribble objs count for required size of the scribble region
2464  *
2465  * The scribble buffer size must be enough to contain:
2466  * 1/ a struct page pointer for each device in the array +2
2467  * 2/ room to convert each entry in (1) to its corresponding dma
2468  *    (dma_map_page()) or page (page_address()) address.
2469  *
2470  * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2471  * calculate over all devices (not just the data blocks), using zeros in place
2472  * of the P and Q blocks.
2473  */
2474 static int scribble_alloc(struct raid5_percpu *percpu,
2475                           int num, int cnt)
2476 {
2477         size_t obj_size =
2478                 sizeof(struct page *) * (num + 2) +
2479                 sizeof(addr_conv_t) * (num + 2) +
2480                 sizeof(unsigned int) * (num + 2);
2481         void *scribble;
2482
2483         /*
2484          * If here is in raid array suspend context, it is in memalloc noio
2485          * context as well, there is no potential recursive memory reclaim
2486          * I/Os with the GFP_KERNEL flag.
2487          */
2488         scribble = kvmalloc_array(cnt, obj_size, GFP_KERNEL);
2489         if (!scribble)
2490                 return -ENOMEM;
2491
2492         kvfree(percpu->scribble);
2493
2494         percpu->scribble = scribble;
2495         percpu->scribble_obj_size = obj_size;
2496         return 0;
2497 }
2498
2499 static int resize_chunks(struct r5conf *conf, int new_disks, int new_sectors)
2500 {
2501         unsigned long cpu;
2502         int err = 0;
2503
2504         /* Never shrink. */
2505         if (conf->scribble_disks >= new_disks &&
2506             conf->scribble_sectors >= new_sectors)
2507                 return 0;
2508
2509         raid5_quiesce(conf->mddev, true);
2510         cpus_read_lock();
2511
2512         for_each_present_cpu(cpu) {
2513                 struct raid5_percpu *percpu;
2514
2515                 percpu = per_cpu_ptr(conf->percpu, cpu);
2516                 err = scribble_alloc(percpu, new_disks,
2517                                      new_sectors / RAID5_STRIPE_SECTORS(conf));
2518                 if (err)
2519                         break;
2520         }
2521
2522         cpus_read_unlock();
2523         raid5_quiesce(conf->mddev, false);
2524
2525         if (!err) {
2526                 conf->scribble_disks = new_disks;
2527                 conf->scribble_sectors = new_sectors;
2528         }
2529         return err;
2530 }
2531
2532 static int resize_stripes(struct r5conf *conf, int newsize)
2533 {
2534         /* Make all the stripes able to hold 'newsize' devices.
2535          * New slots in each stripe get 'page' set to a new page.
2536          *
2537          * This happens in stages:
2538          * 1/ create a new kmem_cache and allocate the required number of
2539          *    stripe_heads.
2540          * 2/ gather all the old stripe_heads and transfer the pages across
2541          *    to the new stripe_heads.  This will have the side effect of
2542          *    freezing the array as once all stripe_heads have been collected,
2543          *    no IO will be possible.  Old stripe heads are freed once their
2544          *    pages have been transferred over, and the old kmem_cache is
2545          *    freed when all stripes are done.
2546          * 3/ reallocate conf->disks to be suitable bigger.  If this fails,
2547          *    we simple return a failure status - no need to clean anything up.
2548          * 4/ allocate new pages for the new slots in the new stripe_heads.
2549          *    If this fails, we don't bother trying the shrink the
2550          *    stripe_heads down again, we just leave them as they are.
2551          *    As each stripe_head is processed the new one is released into
2552          *    active service.
2553          *
2554          * Once step2 is started, we cannot afford to wait for a write,
2555          * so we use GFP_NOIO allocations.
2556          */
2557         struct stripe_head *osh, *nsh;
2558         LIST_HEAD(newstripes);
2559         struct disk_info *ndisks;
2560         int err = 0;
2561         struct kmem_cache *sc;
2562         int i;
2563         int hash, cnt;
2564
2565         md_allow_write(conf->mddev);
2566
2567         /* Step 1 */
2568         sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
2569                                struct_size_t(struct stripe_head, dev, newsize),
2570                                0, 0, NULL);
2571         if (!sc)
2572                 return -ENOMEM;
2573
2574         /* Need to ensure auto-resizing doesn't interfere */
2575         mutex_lock(&conf->cache_size_mutex);
2576
2577         for (i = conf->max_nr_stripes; i; i--) {
2578                 nsh = alloc_stripe(sc, GFP_KERNEL, newsize, conf);
2579                 if (!nsh)
2580                         break;
2581
2582                 list_add(&nsh->lru, &newstripes);
2583         }
2584         if (i) {
2585                 /* didn't get enough, give up */
2586                 while (!list_empty(&newstripes)) {
2587                         nsh = list_entry(newstripes.next, struct stripe_head, lru);
2588                         list_del(&nsh->lru);
2589                         free_stripe(sc, nsh);
2590                 }
2591                 kmem_cache_destroy(sc);
2592                 mutex_unlock(&conf->cache_size_mutex);
2593                 return -ENOMEM;
2594         }
2595         /* Step 2 - Must use GFP_NOIO now.
2596          * OK, we have enough stripes, start collecting inactive
2597          * stripes and copying them over
2598          */
2599         hash = 0;
2600         cnt = 0;
2601         list_for_each_entry(nsh, &newstripes, lru) {
2602                 lock_device_hash_lock(conf, hash);
2603                 wait_event_cmd(conf->wait_for_stripe,
2604                                     !list_empty(conf->inactive_list + hash),
2605                                     unlock_device_hash_lock(conf, hash),
2606                                     lock_device_hash_lock(conf, hash));
2607                 osh = get_free_stripe(conf, hash);
2608                 unlock_device_hash_lock(conf, hash);
2609
2610 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
2611         for (i = 0; i < osh->nr_pages; i++) {
2612                 nsh->pages[i] = osh->pages[i];
2613                 osh->pages[i] = NULL;
2614         }
2615 #endif
2616                 for(i=0; i<conf->pool_size; i++) {
2617                         nsh->dev[i].page = osh->dev[i].page;
2618                         nsh->dev[i].orig_page = osh->dev[i].page;
2619                         nsh->dev[i].offset = osh->dev[i].offset;
2620                 }
2621                 nsh->hash_lock_index = hash;
2622                 free_stripe(conf->slab_cache, osh);
2623                 cnt++;
2624                 if (cnt >= conf->max_nr_stripes / NR_STRIPE_HASH_LOCKS +
2625                     !!((conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS) > hash)) {
2626                         hash++;
2627                         cnt = 0;
2628                 }
2629         }
2630         kmem_cache_destroy(conf->slab_cache);
2631
2632         /* Step 3.
2633          * At this point, we are holding all the stripes so the array
2634          * is completely stalled, so now is a good time to resize
2635          * conf->disks and the scribble region
2636          */
2637         ndisks = kcalloc(newsize, sizeof(struct disk_info), GFP_NOIO);
2638         if (ndisks) {
2639                 for (i = 0; i < conf->pool_size; i++)
2640                         ndisks[i] = conf->disks[i];
2641
2642                 for (i = conf->pool_size; i < newsize; i++) {
2643                         ndisks[i].extra_page = alloc_page(GFP_NOIO);
2644                         if (!ndisks[i].extra_page)
2645                                 err = -ENOMEM;
2646                 }
2647
2648                 if (err) {
2649                         for (i = conf->pool_size; i < newsize; i++)
2650                                 if (ndisks[i].extra_page)
2651                                         put_page(ndisks[i].extra_page);
2652                         kfree(ndisks);
2653                 } else {
2654                         kfree(conf->disks);
2655                         conf->disks = ndisks;
2656                 }
2657         } else
2658                 err = -ENOMEM;
2659
2660         conf->slab_cache = sc;
2661         conf->active_name = 1-conf->active_name;
2662
2663         /* Step 4, return new stripes to service */
2664         while(!list_empty(&newstripes)) {
2665                 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2666                 list_del_init(&nsh->lru);
2667
2668 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
2669                 for (i = 0; i < nsh->nr_pages; i++) {
2670                         if (nsh->pages[i])
2671                                 continue;
2672                         nsh->pages[i] = alloc_page(GFP_NOIO);
2673                         if (!nsh->pages[i])
2674                                 err = -ENOMEM;
2675                 }
2676
2677                 for (i = conf->raid_disks; i < newsize; i++) {
2678                         if (nsh->dev[i].page)
2679                                 continue;
2680                         nsh->dev[i].page = raid5_get_dev_page(nsh, i);
2681                         nsh->dev[i].orig_page = nsh->dev[i].page;
2682                         nsh->dev[i].offset = raid5_get_page_offset(nsh, i);
2683                 }
2684 #else
2685                 for (i=conf->raid_disks; i < newsize; i++)
2686                         if (nsh->dev[i].page == NULL) {
2687                                 struct page *p = alloc_page(GFP_NOIO);
2688                                 nsh->dev[i].page = p;
2689                                 nsh->dev[i].orig_page = p;
2690                                 nsh->dev[i].offset = 0;
2691                                 if (!p)
2692                                         err = -ENOMEM;
2693                         }
2694 #endif
2695                 raid5_release_stripe(nsh);
2696         }
2697         /* critical section pass, GFP_NOIO no longer needed */
2698
2699         if (!err)
2700                 conf->pool_size = newsize;
2701         mutex_unlock(&conf->cache_size_mutex);
2702
2703         return err;
2704 }
2705
2706 static int drop_one_stripe(struct r5conf *conf)
2707 {
2708         struct stripe_head *sh;
2709         int hash = (conf->max_nr_stripes - 1) & STRIPE_HASH_LOCKS_MASK;
2710
2711         spin_lock_irq(conf->hash_locks + hash);
2712         sh = get_free_stripe(conf, hash);
2713         spin_unlock_irq(conf->hash_locks + hash);
2714         if (!sh)
2715                 return 0;
2716         BUG_ON(atomic_read(&sh->count));
2717         shrink_buffers(sh);
2718         free_stripe(conf->slab_cache, sh);
2719         atomic_dec(&conf->active_stripes);
2720         conf->max_nr_stripes--;
2721         return 1;
2722 }
2723
2724 static void shrink_stripes(struct r5conf *conf)
2725 {
2726         while (conf->max_nr_stripes &&
2727                drop_one_stripe(conf))
2728                 ;
2729
2730         kmem_cache_destroy(conf->slab_cache);
2731         conf->slab_cache = NULL;
2732 }
2733
2734 /*
2735  * This helper wraps rcu_dereference_protected() and can be used when
2736  * it is known that the nr_pending of the rdev is elevated.
2737  */
2738 static struct md_rdev *rdev_pend_deref(struct md_rdev __rcu *rdev)
2739 {
2740         return rcu_dereference_protected(rdev,
2741                         atomic_read(&rcu_access_pointer(rdev)->nr_pending));
2742 }
2743
2744 /*
2745  * This helper wraps rcu_dereference_protected() and should be used
2746  * when it is known that the mddev_lock() is held. This is safe
2747  * seeing raid5_remove_disk() has the same lock held.
2748  */
2749 static struct md_rdev *rdev_mdlock_deref(struct mddev *mddev,
2750                                          struct md_rdev __rcu *rdev)
2751 {
2752         return rcu_dereference_protected(rdev,
2753                         lockdep_is_held(&mddev->reconfig_mutex));
2754 }
2755
2756 static void raid5_end_read_request(struct bio * bi)
2757 {
2758         struct stripe_head *sh = bi->bi_private;
2759         struct r5conf *conf = sh->raid_conf;
2760         int disks = sh->disks, i;
2761         struct md_rdev *rdev = NULL;
2762         sector_t s;
2763
2764         for (i=0 ; i<disks; i++)
2765                 if (bi == &sh->dev[i].req)
2766                         break;
2767
2768         pr_debug("end_read_request %llu/%d, count: %d, error %d.\n",
2769                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2770                 bi->bi_status);
2771         if (i == disks) {
2772                 BUG();
2773                 return;
2774         }
2775         if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2776                 /* If replacement finished while this request was outstanding,
2777                  * 'replacement' might be NULL already.
2778                  * In that case it moved down to 'rdev'.
2779                  * rdev is not removed until all requests are finished.
2780                  */
2781                 rdev = rdev_pend_deref(conf->disks[i].replacement);
2782         if (!rdev)
2783                 rdev = rdev_pend_deref(conf->disks[i].rdev);
2784
2785         if (use_new_offset(conf, sh))
2786                 s = sh->sector + rdev->new_data_offset;
2787         else
2788                 s = sh->sector + rdev->data_offset;
2789         if (!bi->bi_status) {
2790                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
2791                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2792                         /* Note that this cannot happen on a
2793                          * replacement device.  We just fail those on
2794                          * any error
2795                          */
2796                         pr_info_ratelimited(
2797                                 "md/raid:%s: read error corrected (%lu sectors at %llu on %pg)\n",
2798                                 mdname(conf->mddev), RAID5_STRIPE_SECTORS(conf),
2799                                 (unsigned long long)s,
2800                                 rdev->bdev);
2801                         atomic_add(RAID5_STRIPE_SECTORS(conf), &rdev->corrected_errors);
2802                         clear_bit(R5_ReadError, &sh->dev[i].flags);
2803                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
2804                 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2805                         clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2806
2807                 if (test_bit(R5_InJournal, &sh->dev[i].flags))
2808                         /*
2809                          * end read for a page in journal, this
2810                          * must be preparing for prexor in rmw
2811                          */
2812                         set_bit(R5_OrigPageUPTDODATE, &sh->dev[i].flags);
2813
2814                 if (atomic_read(&rdev->read_errors))
2815                         atomic_set(&rdev->read_errors, 0);
2816         } else {
2817                 int retry = 0;
2818                 int set_bad = 0;
2819
2820                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
2821                 if (!(bi->bi_status == BLK_STS_PROTECTION))
2822                         atomic_inc(&rdev->read_errors);
2823                 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2824                         pr_warn_ratelimited(
2825                                 "md/raid:%s: read error on replacement device (sector %llu on %pg).\n",
2826                                 mdname(conf->mddev),
2827                                 (unsigned long long)s,
2828                                 rdev->bdev);
2829                 else if (conf->mddev->degraded >= conf->max_degraded) {
2830                         set_bad = 1;
2831                         pr_warn_ratelimited(
2832                                 "md/raid:%s: read error not correctable (sector %llu on %pg).\n",
2833                                 mdname(conf->mddev),
2834                                 (unsigned long long)s,
2835                                 rdev->bdev);
2836                 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
2837                         /* Oh, no!!! */
2838                         set_bad = 1;
2839                         pr_warn_ratelimited(
2840                                 "md/raid:%s: read error NOT corrected!! (sector %llu on %pg).\n",
2841                                 mdname(conf->mddev),
2842                                 (unsigned long long)s,
2843                                 rdev->bdev);
2844                 } else if (atomic_read(&rdev->read_errors)
2845                          > conf->max_nr_stripes) {
2846                         if (!test_bit(Faulty, &rdev->flags)) {
2847                                 pr_warn("md/raid:%s: %d read_errors > %d stripes\n",
2848                                     mdname(conf->mddev),
2849                                     atomic_read(&rdev->read_errors),
2850                                     conf->max_nr_stripes);
2851                                 pr_warn("md/raid:%s: Too many read errors, failing device %pg.\n",
2852                                     mdname(conf->mddev), rdev->bdev);
2853                         }
2854                 } else
2855                         retry = 1;
2856                 if (set_bad && test_bit(In_sync, &rdev->flags)
2857                     && !test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2858                         retry = 1;
2859                 if (retry)
2860                         if (sh->qd_idx >= 0 && sh->pd_idx == i)
2861                                 set_bit(R5_ReadError, &sh->dev[i].flags);
2862                         else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
2863                                 set_bit(R5_ReadError, &sh->dev[i].flags);
2864                                 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2865                         } else
2866                                 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2867                 else {
2868                         clear_bit(R5_ReadError, &sh->dev[i].flags);
2869                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
2870                         if (!(set_bad
2871                               && test_bit(In_sync, &rdev->flags)
2872                               && rdev_set_badblocks(
2873                                       rdev, sh->sector, RAID5_STRIPE_SECTORS(conf), 0)))
2874                                 md_error(conf->mddev, rdev);
2875                 }
2876         }
2877         rdev_dec_pending(rdev, conf->mddev);
2878         bio_uninit(bi);
2879         clear_bit(R5_LOCKED, &sh->dev[i].flags);
2880         set_bit(STRIPE_HANDLE, &sh->state);
2881         raid5_release_stripe(sh);
2882 }
2883
2884 static void raid5_end_write_request(struct bio *bi)
2885 {
2886         struct stripe_head *sh = bi->bi_private;
2887         struct r5conf *conf = sh->raid_conf;
2888         int disks = sh->disks, i;
2889         struct md_rdev *rdev;
2890         sector_t first_bad;
2891         int bad_sectors;
2892         int replacement = 0;
2893
2894         for (i = 0 ; i < disks; i++) {
2895                 if (bi == &sh->dev[i].req) {
2896                         rdev = rdev_pend_deref(conf->disks[i].rdev);
2897                         break;
2898                 }
2899                 if (bi == &sh->dev[i].rreq) {
2900                         rdev = rdev_pend_deref(conf->disks[i].replacement);
2901                         if (rdev)
2902                                 replacement = 1;
2903                         else
2904                                 /* rdev was removed and 'replacement'
2905                                  * replaced it.  rdev is not removed
2906                                  * until all requests are finished.
2907                                  */
2908                                 rdev = rdev_pend_deref(conf->disks[i].rdev);
2909                         break;
2910                 }
2911         }
2912         pr_debug("end_write_request %llu/%d, count %d, error: %d.\n",
2913                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2914                 bi->bi_status);
2915         if (i == disks) {
2916                 BUG();
2917                 return;
2918         }
2919
2920         if (replacement) {
2921                 if (bi->bi_status)
2922                         md_error(conf->mddev, rdev);
2923                 else if (is_badblock(rdev, sh->sector,
2924                                      RAID5_STRIPE_SECTORS(conf),
2925                                      &first_bad, &bad_sectors))
2926                         set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
2927         } else {
2928                 if (bi->bi_status) {
2929                         set_bit(STRIPE_DEGRADED, &sh->state);
2930                         set_bit(WriteErrorSeen, &rdev->flags);
2931                         set_bit(R5_WriteError, &sh->dev[i].flags);
2932                         if (!test_and_set_bit(WantReplacement, &rdev->flags))
2933                                 set_bit(MD_RECOVERY_NEEDED,
2934                                         &rdev->mddev->recovery);
2935                 } else if (is_badblock(rdev, sh->sector,
2936                                        RAID5_STRIPE_SECTORS(conf),
2937                                        &first_bad, &bad_sectors)) {
2938                         set_bit(R5_MadeGood, &sh->dev[i].flags);
2939                         if (test_bit(R5_ReadError, &sh->dev[i].flags))
2940                                 /* That was a successful write so make
2941                                  * sure it looks like we already did
2942                                  * a re-write.
2943                                  */
2944                                 set_bit(R5_ReWrite, &sh->dev[i].flags);
2945                 }
2946         }
2947         rdev_dec_pending(rdev, conf->mddev);
2948
2949         if (sh->batch_head && bi->bi_status && !replacement)
2950                 set_bit(STRIPE_BATCH_ERR, &sh->batch_head->state);
2951
2952         bio_uninit(bi);
2953         if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2954                 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2955         set_bit(STRIPE_HANDLE, &sh->state);
2956
2957         if (sh->batch_head && sh != sh->batch_head)
2958                 raid5_release_stripe(sh->batch_head);
2959         raid5_release_stripe(sh);
2960 }
2961
2962 static void raid5_error(struct mddev *mddev, struct md_rdev *rdev)
2963 {
2964         struct r5conf *conf = mddev->private;
2965         unsigned long flags;
2966         pr_debug("raid456: error called\n");
2967
2968         pr_crit("md/raid:%s: Disk failure on %pg, disabling device.\n",
2969                 mdname(mddev), rdev->bdev);
2970
2971         spin_lock_irqsave(&conf->device_lock, flags);
2972         set_bit(Faulty, &rdev->flags);
2973         clear_bit(In_sync, &rdev->flags);
2974         mddev->degraded = raid5_calc_degraded(conf);
2975
2976         if (has_failed(conf)) {
2977                 set_bit(MD_BROKEN, &conf->mddev->flags);
2978                 conf->recovery_disabled = mddev->recovery_disabled;
2979
2980                 pr_crit("md/raid:%s: Cannot continue operation (%d/%d failed).\n",
2981                         mdname(mddev), mddev->degraded, conf->raid_disks);
2982         } else {
2983                 pr_crit("md/raid:%s: Operation continuing on %d devices.\n",
2984                         mdname(mddev), conf->raid_disks - mddev->degraded);
2985         }
2986
2987         spin_unlock_irqrestore(&conf->device_lock, flags);
2988         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2989
2990         set_bit(Blocked, &rdev->flags);
2991         set_mask_bits(&mddev->sb_flags, 0,
2992                       BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
2993         r5c_update_on_rdev_error(mddev, rdev);
2994 }
2995
2996 /*
2997  * Input: a 'big' sector number,
2998  * Output: index of the data and parity disk, and the sector # in them.
2999  */
3000 sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
3001                               int previous, int *dd_idx,
3002                               struct stripe_head *sh)
3003 {
3004         sector_t stripe, stripe2;
3005         sector_t chunk_number;
3006         unsigned int chunk_offset;
3007         int pd_idx, qd_idx;
3008         int ddf_layout = 0;
3009         sector_t new_sector;
3010         int algorithm = previous ? conf->prev_algo
3011                                  : conf->algorithm;
3012         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
3013                                          : conf->chunk_sectors;
3014         int raid_disks = previous ? conf->previous_raid_disks
3015                                   : conf->raid_disks;
3016         int data_disks = raid_disks - conf->max_degraded;
3017
3018         /* First compute the information on this sector */
3019
3020         /*
3021          * Compute the chunk number and the sector offset inside the chunk
3022          */
3023         chunk_offset = sector_div(r_sector, sectors_per_chunk);
3024         chunk_number = r_sector;
3025
3026         /*
3027          * Compute the stripe number
3028          */
3029         stripe = chunk_number;
3030         *dd_idx = sector_div(stripe, data_disks);
3031         stripe2 = stripe;
3032         /*
3033          * Select the parity disk based on the user selected algorithm.
3034          */
3035         pd_idx = qd_idx = -1;
3036         switch(conf->level) {
3037         case 4:
3038                 pd_idx = data_disks;
3039                 break;
3040         case 5:
3041                 switch (algorithm) {
3042                 case ALGORITHM_LEFT_ASYMMETRIC:
3043                         pd_idx = data_disks - sector_div(stripe2, raid_disks);
3044                         if (*dd_idx >= pd_idx)
3045                                 (*dd_idx)++;
3046                         break;
3047                 case ALGORITHM_RIGHT_ASYMMETRIC:
3048                         pd_idx = sector_div(stripe2, raid_disks);
3049                         if (*dd_idx >= pd_idx)
3050                                 (*dd_idx)++;
3051                         break;
3052                 case ALGORITHM_LEFT_SYMMETRIC:
3053                         pd_idx = data_disks - sector_div(stripe2, raid_disks);
3054                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
3055                         break;
3056                 case ALGORITHM_RIGHT_SYMMETRIC:
3057                         pd_idx = sector_div(stripe2, raid_disks);
3058                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
3059                         break;
3060                 case ALGORITHM_PARITY_0:
3061                         pd_idx = 0;
3062                         (*dd_idx)++;
3063                         break;
3064                 case ALGORITHM_PARITY_N:
3065                         pd_idx = data_disks;
3066                         break;
3067                 default:
3068                         BUG();
3069                 }
3070                 break;
3071         case 6:
3072
3073                 switch (algorithm) {
3074                 case ALGORITHM_LEFT_ASYMMETRIC:
3075                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
3076                         qd_idx = pd_idx + 1;
3077                         if (pd_idx == raid_disks-1) {
3078                                 (*dd_idx)++;    /* Q D D D P */
3079                                 qd_idx = 0;
3080                         } else if (*dd_idx >= pd_idx)
3081                                 (*dd_idx) += 2; /* D D P Q D */
3082                         break;
3083                 case ALGORITHM_RIGHT_ASYMMETRIC:
3084                         pd_idx = sector_div(stripe2, raid_disks);
3085                         qd_idx = pd_idx + 1;
3086                         if (pd_idx == raid_disks-1) {
3087                                 (*dd_idx)++;    /* Q D D D P */
3088                                 qd_idx = 0;
3089                         } else if (*dd_idx >= pd_idx)
3090                                 (*dd_idx) += 2; /* D D P Q D */
3091                         break;
3092                 case ALGORITHM_LEFT_SYMMETRIC:
3093                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
3094                         qd_idx = (pd_idx + 1) % raid_disks;
3095                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
3096                         break;
3097                 case ALGORITHM_RIGHT_SYMMETRIC:
3098                         pd_idx = sector_div(stripe2, raid_disks);
3099                         qd_idx = (pd_idx + 1) % raid_disks;
3100                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
3101                         break;
3102
3103                 case ALGORITHM_PARITY_0:
3104                         pd_idx = 0;
3105                         qd_idx = 1;
3106                         (*dd_idx) += 2;
3107                         break;
3108                 case ALGORITHM_PARITY_N:
3109                         pd_idx = data_disks;
3110                         qd_idx = data_disks + 1;
3111                         break;
3112
3113                 case ALGORITHM_ROTATING_ZERO_RESTART:
3114                         /* Exactly the same as RIGHT_ASYMMETRIC, but or
3115                          * of blocks for computing Q is different.
3116                          */
3117                         pd_idx = sector_div(stripe2, raid_disks);
3118                         qd_idx = pd_idx + 1;
3119                         if (pd_idx == raid_disks-1) {
3120                                 (*dd_idx)++;    /* Q D D D P */
3121                                 qd_idx = 0;
3122                         } else if (*dd_idx >= pd_idx)
3123                                 (*dd_idx) += 2; /* D D P Q D */
3124                         ddf_layout = 1;
3125                         break;
3126
3127                 case ALGORITHM_ROTATING_N_RESTART:
3128                         /* Same a left_asymmetric, by first stripe is
3129                          * D D D P Q  rather than
3130                          * Q D D D P
3131                          */
3132                         stripe2 += 1;
3133                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
3134                         qd_idx = pd_idx + 1;
3135                         if (pd_idx == raid_disks-1) {
3136                                 (*dd_idx)++;    /* Q D D D P */
3137                                 qd_idx = 0;
3138                         } else if (*dd_idx >= pd_idx)
3139                                 (*dd_idx) += 2; /* D D P Q D */
3140                         ddf_layout = 1;
3141                         break;
3142
3143                 case ALGORITHM_ROTATING_N_CONTINUE:
3144                         /* Same as left_symmetric but Q is before P */
3145                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
3146                         qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
3147                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
3148                         ddf_layout = 1;
3149                         break;
3150
3151                 case ALGORITHM_LEFT_ASYMMETRIC_6:
3152                         /* RAID5 left_asymmetric, with Q on last device */
3153                         pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
3154                         if (*dd_idx >= pd_idx)
3155                                 (*dd_idx)++;
3156                         qd_idx = raid_disks - 1;
3157                         break;
3158
3159                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
3160                         pd_idx = sector_div(stripe2, raid_disks-1);
3161                         if (*dd_idx >= pd_idx)
3162                                 (*dd_idx)++;
3163                         qd_idx = raid_disks - 1;
3164                         break;
3165
3166                 case ALGORITHM_LEFT_SYMMETRIC_6:
3167                         pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
3168                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
3169                         qd_idx = raid_disks - 1;
3170                         break;
3171
3172                 case ALGORITHM_RIGHT_SYMMETRIC_6:
3173                         pd_idx = sector_div(stripe2, raid_disks-1);
3174                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
3175                         qd_idx = raid_disks - 1;
3176                         break;
3177
3178                 case ALGORITHM_PARITY_0_6:
3179                         pd_idx = 0;
3180                         (*dd_idx)++;
3181                         qd_idx = raid_disks - 1;
3182                         break;
3183
3184                 default:
3185                         BUG();
3186                 }
3187                 break;
3188         }
3189
3190         if (sh) {
3191                 sh->pd_idx = pd_idx;
3192                 sh->qd_idx = qd_idx;
3193                 sh->ddf_layout = ddf_layout;
3194         }
3195         /*
3196          * Finally, compute the new sector number
3197          */
3198         new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
3199         return new_sector;
3200 }
3201
3202 sector_t raid5_compute_blocknr(struct stripe_head *sh, int i, int previous)
3203 {
3204         struct r5conf *conf = sh->raid_conf;
3205         int raid_disks = sh->disks;
3206         int data_disks = raid_disks - conf->max_degraded;
3207         sector_t new_sector = sh->sector, check;
3208         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
3209                                          : conf->chunk_sectors;
3210         int algorithm = previous ? conf->prev_algo
3211                                  : conf->algorithm;
3212         sector_t stripe;
3213         int chunk_offset;
3214         sector_t chunk_number;
3215         int dummy1, dd_idx = i;
3216         sector_t r_sector;
3217         struct stripe_head sh2;
3218
3219         chunk_offset = sector_div(new_sector, sectors_per_chunk);
3220         stripe = new_sector;
3221
3222         if (i == sh->pd_idx)
3223                 return 0;
3224         switch(conf->level) {
3225         case 4: break;
3226         case 5:
3227                 switch (algorithm) {
3228                 case ALGORITHM_LEFT_ASYMMETRIC:
3229                 case ALGORITHM_RIGHT_ASYMMETRIC:
3230                         if (i > sh->pd_idx)
3231                                 i--;
3232                         break;
3233                 case ALGORITHM_LEFT_SYMMETRIC:
3234                 case ALGORITHM_RIGHT_SYMMETRIC:
3235                         if (i < sh->pd_idx)
3236                                 i += raid_disks;
3237                         i -= (sh->pd_idx + 1);
3238                         break;
3239                 case ALGORITHM_PARITY_0:
3240                         i -= 1;
3241                         break;
3242                 case ALGORITHM_PARITY_N:
3243                         break;
3244                 default:
3245                         BUG();
3246                 }
3247                 break;
3248         case 6:
3249                 if (i == sh->qd_idx)
3250                         return 0; /* It is the Q disk */
3251                 switch (algorithm) {
3252                 case ALGORITHM_LEFT_ASYMMETRIC:
3253                 case ALGORITHM_RIGHT_ASYMMETRIC:
3254                 case ALGORITHM_ROTATING_ZERO_RESTART:
3255                 case ALGORITHM_ROTATING_N_RESTART:
3256                         if (sh->pd_idx == raid_disks-1)
3257                                 i--;    /* Q D D D P */
3258                         else if (i > sh->pd_idx)
3259                                 i -= 2; /* D D P Q D */
3260                         break;
3261                 case ALGORITHM_LEFT_SYMMETRIC:
3262                 case ALGORITHM_RIGHT_SYMMETRIC:
3263                         if (sh->pd_idx == raid_disks-1)
3264                                 i--; /* Q D D D P */
3265                         else {
3266                                 /* D D P Q D */
3267                                 if (i < sh->pd_idx)
3268                                         i += raid_disks;
3269                                 i -= (sh->pd_idx + 2);
3270                         }
3271                         break;
3272                 case ALGORITHM_PARITY_0:
3273                         i -= 2;
3274                         break;
3275                 case ALGORITHM_PARITY_N:
3276                         break;
3277                 case ALGORITHM_ROTATING_N_CONTINUE:
3278                         /* Like left_symmetric, but P is before Q */
3279                         if (sh->pd_idx == 0)
3280                                 i--;    /* P D D D Q */
3281                         else {
3282                                 /* D D Q P D */
3283                                 if (i < sh->pd_idx)
3284                                         i += raid_disks;
3285                                 i -= (sh->pd_idx + 1);
3286                         }
3287                         break;
3288                 case ALGORITHM_LEFT_ASYMMETRIC_6:
3289                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
3290                         if (i > sh->pd_idx)
3291                                 i--;
3292                         break;
3293                 case ALGORITHM_LEFT_SYMMETRIC_6:
3294                 case ALGORITHM_RIGHT_SYMMETRIC_6:
3295                         if (i < sh->pd_idx)
3296                                 i += data_disks + 1;
3297                         i -= (sh->pd_idx + 1);
3298                         break;
3299                 case ALGORITHM_PARITY_0_6:
3300                         i -= 1;
3301                         break;
3302                 default:
3303                         BUG();
3304                 }
3305                 break;
3306         }
3307
3308         chunk_number = stripe * data_disks + i;
3309         r_sector = chunk_number * sectors_per_chunk + chunk_offset;
3310
3311         check = raid5_compute_sector(conf, r_sector,
3312                                      previous, &dummy1, &sh2);
3313         if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
3314                 || sh2.qd_idx != sh->qd_idx) {
3315                 pr_warn("md/raid:%s: compute_blocknr: map not correct\n",
3316                         mdname(conf->mddev));
3317                 return 0;
3318         }
3319         return r_sector;
3320 }
3321
3322 /*
3323  * There are cases where we want handle_stripe_dirtying() and
3324  * schedule_reconstruction() to delay towrite to some dev of a stripe.
3325  *
3326  * This function checks whether we want to delay the towrite. Specifically,
3327  * we delay the towrite when:
3328  *
3329  *   1. degraded stripe has a non-overwrite to the missing dev, AND this
3330  *      stripe has data in journal (for other devices).
3331  *
3332  *      In this case, when reading data for the non-overwrite dev, it is
3333  *      necessary to handle complex rmw of write back cache (prexor with
3334  *      orig_page, and xor with page). To keep read path simple, we would
3335  *      like to flush data in journal to RAID disks first, so complex rmw
3336  *      is handled in the write patch (handle_stripe_dirtying).
3337  *
3338  *   2. when journal space is critical (R5C_LOG_CRITICAL=1)
3339  *
3340  *      It is important to be able to flush all stripes in raid5-cache.
3341  *      Therefore, we need reserve some space on the journal device for
3342  *      these flushes. If flush operation includes pending writes to the
3343  *      stripe, we need to reserve (conf->raid_disk + 1) pages per stripe
3344  *      for the flush out. If we exclude these pending writes from flush
3345  *      operation, we only need (conf->max_degraded + 1) pages per stripe.
3346  *      Therefore, excluding pending writes in these cases enables more
3347  *      efficient use of the journal device.
3348  *
3349  *      Note: To make sure the stripe makes progress, we only delay
3350  *      towrite for stripes with data already in journal (injournal > 0).
3351  *      When LOG_CRITICAL, stripes with injournal == 0 will be sent to
3352  *      no_space_stripes list.
3353  *
3354  *   3. during journal failure
3355  *      In journal failure, we try to flush all cached data to raid disks
3356  *      based on data in stripe cache. The array is read-only to upper
3357  *      layers, so we would skip all pending writes.
3358  *
3359  */
3360 static inline bool delay_towrite(struct r5conf *conf,
3361                                  struct r5dev *dev,
3362                                  struct stripe_head_state *s)
3363 {
3364         /* case 1 above */
3365         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3366             !test_bit(R5_Insync, &dev->flags) && s->injournal)
3367                 return true;
3368         /* case 2 above */
3369         if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
3370             s->injournal > 0)
3371                 return true;
3372         /* case 3 above */
3373         if (s->log_failed && s->injournal)
3374                 return true;
3375         return false;
3376 }
3377
3378 static void
3379 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
3380                          int rcw, int expand)
3381 {
3382         int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx, disks = sh->disks;
3383         struct r5conf *conf = sh->raid_conf;
3384         int level = conf->level;
3385
3386         if (rcw) {
3387                 /*
3388                  * In some cases, handle_stripe_dirtying initially decided to
3389                  * run rmw and allocates extra page for prexor. However, rcw is
3390                  * cheaper later on. We need to free the extra page now,
3391                  * because we won't be able to do that in ops_complete_prexor().
3392                  */
3393                 r5c_release_extra_page(sh);
3394
3395                 for (i = disks; i--; ) {
3396                         struct r5dev *dev = &sh->dev[i];
3397
3398                         if (dev->towrite && !delay_towrite(conf, dev, s)) {
3399                                 set_bit(R5_LOCKED, &dev->flags);
3400                                 set_bit(R5_Wantdrain, &dev->flags);
3401                                 if (!expand)
3402                                         clear_bit(R5_UPTODATE, &dev->flags);
3403                                 s->locked++;
3404                         } else if (test_bit(R5_InJournal, &dev->flags)) {
3405                                 set_bit(R5_LOCKED, &dev->flags);
3406                                 s->locked++;
3407                         }
3408                 }
3409                 /* if we are not expanding this is a proper write request, and
3410                  * there will be bios with new data to be drained into the
3411                  * stripe cache
3412                  */
3413                 if (!expand) {
3414                         if (!s->locked)
3415                                 /* False alarm, nothing to do */
3416                                 return;
3417                         sh->reconstruct_state = reconstruct_state_drain_run;
3418                         set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
3419                 } else
3420                         sh->reconstruct_state = reconstruct_state_run;
3421
3422                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
3423
3424                 if (s->locked + conf->max_degraded == disks)
3425                         if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
3426                                 atomic_inc(&conf->pending_full_writes);
3427         } else {
3428                 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
3429                         test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
3430                 BUG_ON(level == 6 &&
3431                         (!(test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags) ||
3432                            test_bit(R5_Wantcompute, &sh->dev[qd_idx].flags))));
3433
3434                 for (i = disks; i--; ) {
3435                         struct r5dev *dev = &sh->dev[i];
3436                         if (i == pd_idx || i == qd_idx)
3437                                 continue;
3438
3439                         if (dev->towrite &&
3440                             (test_bit(R5_UPTODATE, &dev->flags) ||
3441                              test_bit(R5_Wantcompute, &dev->flags))) {
3442                                 set_bit(R5_Wantdrain, &dev->flags);
3443                                 set_bit(R5_LOCKED, &dev->flags);
3444                                 clear_bit(R5_UPTODATE, &dev->flags);
3445                                 s->locked++;
3446                         } else if (test_bit(R5_InJournal, &dev->flags)) {
3447                                 set_bit(R5_LOCKED, &dev->flags);
3448                                 s->locked++;
3449                         }
3450                 }
3451                 if (!s->locked)
3452                         /* False alarm - nothing to do */
3453                         return;
3454                 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
3455                 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
3456                 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
3457                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
3458         }
3459
3460         /* keep the parity disk(s) locked while asynchronous operations
3461          * are in flight
3462          */
3463         set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
3464         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3465         s->locked++;
3466
3467         if (level == 6) {
3468                 int qd_idx = sh->qd_idx;
3469                 struct r5dev *dev = &sh->dev[qd_idx];
3470
3471                 set_bit(R5_LOCKED, &dev->flags);
3472                 clear_bit(R5_UPTODATE, &dev->flags);
3473                 s->locked++;
3474         }
3475
3476         if (raid5_has_ppl(sh->raid_conf) && sh->ppl_page &&
3477             test_bit(STRIPE_OP_BIODRAIN, &s->ops_request) &&
3478             !test_bit(STRIPE_FULL_WRITE, &sh->state) &&
3479             test_bit(R5_Insync, &sh->dev[pd_idx].flags))
3480                 set_bit(STRIPE_OP_PARTIAL_PARITY, &s->ops_request);
3481
3482         pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
3483                 __func__, (unsigned long long)sh->sector,
3484                 s->locked, s->ops_request);
3485 }
3486
3487 static bool stripe_bio_overlaps(struct stripe_head *sh, struct bio *bi,
3488                                 int dd_idx, int forwrite)
3489 {
3490         struct r5conf *conf = sh->raid_conf;
3491         struct bio **bip;
3492
3493         pr_debug("checking bi b#%llu to stripe s#%llu\n",
3494                  bi->bi_iter.bi_sector, sh->sector);
3495
3496         /* Don't allow new IO added to stripes in batch list */
3497         if (sh->batch_head)
3498                 return true;
3499
3500         if (forwrite)
3501                 bip = &sh->dev[dd_idx].towrite;
3502         else
3503                 bip = &sh->dev[dd_idx].toread;
3504
3505         while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector) {
3506                 if (bio_end_sector(*bip) > bi->bi_iter.bi_sector)
3507                         return true;
3508                 bip = &(*bip)->bi_next;
3509         }
3510
3511         if (*bip && (*bip)->bi_iter.bi_sector < bio_end_sector(bi))
3512                 return true;
3513
3514         if (forwrite && raid5_has_ppl(conf)) {
3515                 /*
3516                  * With PPL only writes to consecutive data chunks within a
3517                  * stripe are allowed because for a single stripe_head we can
3518                  * only have one PPL entry at a time, which describes one data
3519                  * range. Not really an overlap, but wait_for_overlap can be
3520                  * used to handle this.
3521                  */
3522                 sector_t sector;
3523                 sector_t first = 0;
3524                 sector_t last = 0;
3525                 int count = 0;
3526                 int i;
3527
3528                 for (i = 0; i < sh->disks; i++) {
3529                         if (i != sh->pd_idx &&
3530                             (i == dd_idx || sh->dev[i].towrite)) {
3531                                 sector = sh->dev[i].sector;
3532                                 if (count == 0 || sector < first)
3533                                         first = sector;
3534                                 if (sector > last)
3535                                         last = sector;
3536                                 count++;
3537                         }
3538                 }
3539
3540                 if (first + conf->chunk_sectors * (count - 1) != last)
3541                         return true;
3542         }
3543
3544         return false;
3545 }
3546
3547 static void __add_stripe_bio(struct stripe_head *sh, struct bio *bi,
3548                              int dd_idx, int forwrite, int previous)
3549 {
3550         struct r5conf *conf = sh->raid_conf;
3551         struct bio **bip;
3552         int firstwrite = 0;
3553
3554         if (forwrite) {
3555                 bip = &sh->dev[dd_idx].towrite;
3556                 if (!*bip)
3557                         firstwrite = 1;
3558         } else {
3559                 bip = &sh->dev[dd_idx].toread;
3560         }
3561
3562         while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector)
3563                 bip = &(*bip)->bi_next;
3564
3565         if (!forwrite || previous)
3566                 clear_bit(STRIPE_BATCH_READY, &sh->state);
3567
3568         BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
3569         if (*bip)
3570                 bi->bi_next = *bip;
3571         *bip = bi;
3572         bio_inc_remaining(bi);
3573         md_write_inc(conf->mddev, bi);
3574
3575         if (forwrite) {
3576                 /* check if page is covered */
3577                 sector_t sector = sh->dev[dd_idx].sector;
3578                 for (bi=sh->dev[dd_idx].towrite;
3579                      sector < sh->dev[dd_idx].sector + RAID5_STRIPE_SECTORS(conf) &&
3580                              bi && bi->bi_iter.bi_sector <= sector;
3581                      bi = r5_next_bio(conf, bi, sh->dev[dd_idx].sector)) {
3582                         if (bio_end_sector(bi) >= sector)
3583                                 sector = bio_end_sector(bi);
3584                 }
3585                 if (sector >= sh->dev[dd_idx].sector + RAID5_STRIPE_SECTORS(conf))
3586                         if (!test_and_set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags))
3587                                 sh->overwrite_disks++;
3588         }
3589
3590         pr_debug("added bi b#%llu to stripe s#%llu, disk %d, logical %llu\n",
3591                  (*bip)->bi_iter.bi_sector, sh->sector, dd_idx,
3592                  sh->dev[dd_idx].sector);
3593
3594         if (conf->mddev->bitmap && firstwrite) {
3595                 /* Cannot hold spinlock over bitmap_startwrite,
3596                  * but must ensure this isn't added to a batch until
3597                  * we have added to the bitmap and set bm_seq.
3598                  * So set STRIPE_BITMAP_PENDING to prevent
3599                  * batching.
3600                  * If multiple __add_stripe_bio() calls race here they
3601                  * much all set STRIPE_BITMAP_PENDING.  So only the first one
3602                  * to complete "bitmap_startwrite" gets to set
3603                  * STRIPE_BIT_DELAY.  This is important as once a stripe
3604                  * is added to a batch, STRIPE_BIT_DELAY cannot be changed
3605                  * any more.
3606                  */
3607                 set_bit(STRIPE_BITMAP_PENDING, &sh->state);
3608                 spin_unlock_irq(&sh->stripe_lock);
3609                 md_bitmap_startwrite(conf->mddev->bitmap, sh->sector,
3610                                      RAID5_STRIPE_SECTORS(conf), 0);
3611                 spin_lock_irq(&sh->stripe_lock);
3612                 clear_bit(STRIPE_BITMAP_PENDING, &sh->state);
3613                 if (!sh->batch_head) {
3614                         sh->bm_seq = conf->seq_flush+1;
3615                         set_bit(STRIPE_BIT_DELAY, &sh->state);
3616                 }
3617         }
3618 }
3619
3620 /*
3621  * Each stripe/dev can have one or more bios attached.
3622  * toread/towrite point to the first in a chain.
3623  * The bi_next chain must be in order.
3624  */
3625 static bool add_stripe_bio(struct stripe_head *sh, struct bio *bi,
3626                            int dd_idx, int forwrite, int previous)
3627 {
3628         spin_lock_irq(&sh->stripe_lock);
3629
3630         if (stripe_bio_overlaps(sh, bi, dd_idx, forwrite)) {
3631                 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
3632                 spin_unlock_irq(&sh->stripe_lock);
3633                 return false;
3634         }
3635
3636         __add_stripe_bio(sh, bi, dd_idx, forwrite, previous);
3637         spin_unlock_irq(&sh->stripe_lock);
3638         return true;
3639 }
3640
3641 static void end_reshape(struct r5conf *conf);
3642
3643 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
3644                             struct stripe_head *sh)
3645 {
3646         int sectors_per_chunk =
3647                 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
3648         int dd_idx;
3649         int chunk_offset = sector_div(stripe, sectors_per_chunk);
3650         int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
3651
3652         raid5_compute_sector(conf,
3653                              stripe * (disks - conf->max_degraded)
3654                              *sectors_per_chunk + chunk_offset,
3655                              previous,
3656                              &dd_idx, sh);
3657 }
3658
3659 static void
3660 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
3661                      struct stripe_head_state *s, int disks)
3662 {
3663         int i;
3664         BUG_ON(sh->batch_head);
3665         for (i = disks; i--; ) {
3666                 struct bio *bi;
3667                 int bitmap_end = 0;
3668
3669                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
3670                         struct md_rdev *rdev;
3671                         rcu_read_lock();
3672                         rdev = rcu_dereference(conf->disks[i].rdev);
3673                         if (rdev && test_bit(In_sync, &rdev->flags) &&
3674                             !test_bit(Faulty, &rdev->flags))
3675                                 atomic_inc(&rdev->nr_pending);
3676                         else
3677                                 rdev = NULL;
3678                         rcu_read_unlock();
3679                         if (rdev) {
3680                                 if (!rdev_set_badblocks(
3681                                             rdev,
3682                                             sh->sector,
3683                                             RAID5_STRIPE_SECTORS(conf), 0))
3684                                         md_error(conf->mddev, rdev);
3685                                 rdev_dec_pending(rdev, conf->mddev);
3686                         }
3687                 }
3688                 spin_lock_irq(&sh->stripe_lock);
3689                 /* fail all writes first */
3690                 bi = sh->dev[i].towrite;
3691                 sh->dev[i].towrite = NULL;
3692                 sh->overwrite_disks = 0;
3693                 spin_unlock_irq(&sh->stripe_lock);
3694                 if (bi)
3695                         bitmap_end = 1;
3696
3697                 log_stripe_write_finished(sh);
3698
3699                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3700                         wake_up(&conf->wait_for_overlap);
3701
3702                 while (bi && bi->bi_iter.bi_sector <
3703                         sh->dev[i].sector + RAID5_STRIPE_SECTORS(conf)) {
3704                         struct bio *nextbi = r5_next_bio(conf, bi, sh->dev[i].sector);
3705
3706                         md_write_end(conf->mddev);
3707                         bio_io_error(bi);
3708                         bi = nextbi;
3709                 }
3710                 if (bitmap_end)
3711                         md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3712                                            RAID5_STRIPE_SECTORS(conf), 0, 0);
3713                 bitmap_end = 0;
3714                 /* and fail all 'written' */
3715                 bi = sh->dev[i].written;
3716                 sh->dev[i].written = NULL;
3717                 if (test_and_clear_bit(R5_SkipCopy, &sh->dev[i].flags)) {
3718                         WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
3719                         sh->dev[i].page = sh->dev[i].orig_page;
3720                 }
3721
3722                 if (bi) bitmap_end = 1;
3723                 while (bi && bi->bi_iter.bi_sector <
3724                        sh->dev[i].sector + RAID5_STRIPE_SECTORS(conf)) {
3725                         struct bio *bi2 = r5_next_bio(conf, bi, sh->dev[i].sector);
3726
3727                         md_write_end(conf->mddev);
3728                         bio_io_error(bi);
3729                         bi = bi2;
3730                 }
3731
3732                 /* fail any reads if this device is non-operational and
3733                  * the data has not reached the cache yet.
3734                  */
3735                 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
3736                     s->failed > conf->max_degraded &&
3737                     (!test_bit(R5_Insync, &sh->dev[i].flags) ||
3738                       test_bit(R5_ReadError, &sh->dev[i].flags))) {
3739                         spin_lock_irq(&sh->stripe_lock);
3740                         bi = sh->dev[i].toread;
3741                         sh->dev[i].toread = NULL;
3742                         spin_unlock_irq(&sh->stripe_lock);
3743                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3744                                 wake_up(&conf->wait_for_overlap);
3745                         if (bi)
3746                                 s->to_read--;
3747                         while (bi && bi->bi_iter.bi_sector <
3748                                sh->dev[i].sector + RAID5_STRIPE_SECTORS(conf)) {
3749                                 struct bio *nextbi =
3750                                         r5_next_bio(conf, bi, sh->dev[i].sector);
3751
3752                                 bio_io_error(bi);
3753                                 bi = nextbi;
3754                         }
3755                 }
3756                 if (bitmap_end)
3757                         md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3758                                            RAID5_STRIPE_SECTORS(conf), 0, 0);
3759                 /* If we were in the middle of a write the parity block might
3760                  * still be locked - so just clear all R5_LOCKED flags
3761                  */
3762                 clear_bit(R5_LOCKED, &sh->dev[i].flags);
3763         }
3764         s->to_write = 0;
3765         s->written = 0;
3766
3767         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3768                 if (atomic_dec_and_test(&conf->pending_full_writes))
3769                         md_wakeup_thread(conf->mddev->thread);
3770 }
3771
3772 static void
3773 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
3774                    struct stripe_head_state *s)
3775 {
3776         int abort = 0;
3777         int i;
3778
3779         BUG_ON(sh->batch_head);
3780         clear_bit(STRIPE_SYNCING, &sh->state);
3781         if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3782                 wake_up(&conf->wait_for_overlap);
3783         s->syncing = 0;
3784         s->replacing = 0;
3785         /* There is nothing more to do for sync/check/repair.
3786          * Don't even need to abort as that is handled elsewhere
3787          * if needed, and not always wanted e.g. if there is a known
3788          * bad block here.
3789          * For recover/replace we need to record a bad block on all
3790          * non-sync devices, or abort the recovery
3791          */
3792         if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
3793                 /* During recovery devices cannot be removed, so
3794                  * locking and refcounting of rdevs is not needed
3795                  */
3796                 rcu_read_lock();
3797                 for (i = 0; i < conf->raid_disks; i++) {
3798                         struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
3799                         if (rdev
3800                             && !test_bit(Faulty, &rdev->flags)
3801                             && !test_bit(In_sync, &rdev->flags)
3802                             && !rdev_set_badblocks(rdev, sh->sector,
3803                                                    RAID5_STRIPE_SECTORS(conf), 0))
3804                                 abort = 1;
3805                         rdev = rcu_dereference(conf->disks[i].replacement);
3806                         if (rdev
3807                             && !test_bit(Faulty, &rdev->flags)
3808                             && !test_bit(In_sync, &rdev->flags)
3809                             && !rdev_set_badblocks(rdev, sh->sector,
3810                                                    RAID5_STRIPE_SECTORS(conf), 0))
3811                                 abort = 1;
3812                 }
3813                 rcu_read_unlock();
3814                 if (abort)
3815                         conf->recovery_disabled =
3816                                 conf->mddev->recovery_disabled;
3817         }
3818         md_done_sync(conf->mddev, RAID5_STRIPE_SECTORS(conf), !abort);
3819 }
3820
3821 static int want_replace(struct stripe_head *sh, int disk_idx)
3822 {
3823         struct md_rdev *rdev;
3824         int rv = 0;
3825
3826         rcu_read_lock();
3827         rdev = rcu_dereference(sh->raid_conf->disks[disk_idx].replacement);
3828         if (rdev
3829             && !test_bit(Faulty, &rdev->flags)
3830             && !test_bit(In_sync, &rdev->flags)
3831             && (rdev->recovery_offset <= sh->sector
3832                 || rdev->mddev->recovery_cp <= sh->sector))
3833                 rv = 1;
3834         rcu_read_unlock();
3835         return rv;
3836 }
3837
3838 static int need_this_block(struct stripe_head *sh, struct stripe_head_state *s,
3839                            int disk_idx, int disks)
3840 {
3841         struct r5dev *dev = &sh->dev[disk_idx];
3842         struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
3843                                   &sh->dev[s->failed_num[1]] };
3844         int i;
3845         bool force_rcw = (sh->raid_conf->rmw_level == PARITY_DISABLE_RMW);
3846
3847
3848         if (test_bit(R5_LOCKED, &dev->flags) ||
3849             test_bit(R5_UPTODATE, &dev->flags))
3850                 /* No point reading this as we already have it or have
3851                  * decided to get it.
3852                  */
3853                 return 0;
3854
3855         if (dev->toread ||
3856             (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)))
3857                 /* We need this block to directly satisfy a request */
3858                 return 1;
3859
3860         if (s->syncing || s->expanding ||
3861             (s->replacing && want_replace(sh, disk_idx)))
3862                 /* When syncing, or expanding we read everything.
3863                  * When replacing, we need the replaced block.
3864                  */
3865                 return 1;
3866
3867         if ((s->failed >= 1 && fdev[0]->toread) ||
3868             (s->failed >= 2 && fdev[1]->toread))
3869                 /* If we want to read from a failed device, then
3870                  * we need to actually read every other device.
3871                  */
3872                 return 1;
3873
3874         /* Sometimes neither read-modify-write nor reconstruct-write
3875          * cycles can work.  In those cases we read every block we
3876          * can.  Then the parity-update is certain to have enough to
3877          * work with.
3878          * This can only be a problem when we need to write something,
3879          * and some device has failed.  If either of those tests
3880          * fail we need look no further.
3881          */
3882         if (!s->failed || !s->to_write)
3883                 return 0;
3884
3885         if (test_bit(R5_Insync, &dev->flags) &&
3886             !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3887                 /* Pre-reads at not permitted until after short delay
3888                  * to gather multiple requests.  However if this
3889                  * device is no Insync, the block could only be computed
3890                  * and there is no need to delay that.
3891                  */
3892                 return 0;
3893
3894         for (i = 0; i < s->failed && i < 2; i++) {
3895                 if (fdev[i]->towrite &&
3896                     !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3897                     !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3898                         /* If we have a partial write to a failed
3899                          * device, then we will need to reconstruct
3900                          * the content of that device, so all other
3901                          * devices must be read.
3902                          */
3903                         return 1;
3904
3905                 if (s->failed >= 2 &&
3906                     (fdev[i]->towrite ||
3907                      s->failed_num[i] == sh->pd_idx ||
3908                      s->failed_num[i] == sh->qd_idx) &&
3909                     !test_bit(R5_UPTODATE, &fdev[i]->flags))
3910                         /* In max degraded raid6, If the failed disk is P, Q,
3911                          * or we want to read the failed disk, we need to do
3912                          * reconstruct-write.
3913                          */
3914                         force_rcw = true;
3915         }
3916
3917         /* If we are forced to do a reconstruct-write, because parity
3918          * cannot be trusted and we are currently recovering it, there
3919          * is extra need to be careful.
3920          * If one of the devices that we would need to read, because
3921          * it is not being overwritten (and maybe not written at all)
3922          * is missing/faulty, then we need to read everything we can.
3923          */
3924         if (!force_rcw &&
3925             sh->sector < sh->raid_conf->mddev->recovery_cp)
3926                 /* reconstruct-write isn't being forced */
3927                 return 0;
3928         for (i = 0; i < s->failed && i < 2; i++) {
3929                 if (s->failed_num[i] != sh->pd_idx &&
3930                     s->failed_num[i] != sh->qd_idx &&
3931                     !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3932                     !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3933                         return 1;
3934         }
3935
3936         return 0;
3937 }
3938
3939 /* fetch_block - checks the given member device to see if its data needs
3940  * to be read or computed to satisfy a request.
3941  *
3942  * Returns 1 when no more member devices need to be checked, otherwise returns
3943  * 0 to tell the loop in handle_stripe_fill to continue
3944  */
3945 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
3946                        int disk_idx, int disks)
3947 {
3948         struct r5dev *dev = &sh->dev[disk_idx];
3949
3950         /* is the data in this block needed, and can we get it? */
3951         if (need_this_block(sh, s, disk_idx, disks)) {
3952                 /* we would like to get this block, possibly by computing it,
3953                  * otherwise read it if the backing disk is insync
3954                  */
3955                 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
3956                 BUG_ON(test_bit(R5_Wantread, &dev->flags));
3957                 BUG_ON(sh->batch_head);
3958
3959                 /*
3960                  * In the raid6 case if the only non-uptodate disk is P
3961                  * then we already trusted P to compute the other failed
3962                  * drives. It is safe to compute rather than re-read P.
3963                  * In other cases we only compute blocks from failed
3964                  * devices, otherwise check/repair might fail to detect
3965                  * a real inconsistency.
3966                  */
3967
3968                 if ((s->uptodate == disks - 1) &&
3969                     ((sh->qd_idx >= 0 && sh->pd_idx == disk_idx) ||
3970                     (s->failed && (disk_idx == s->failed_num[0] ||
3971                                    disk_idx == s->failed_num[1])))) {
3972                         /* have disk failed, and we're requested to fetch it;
3973                          * do compute it
3974                          */
3975                         pr_debug("Computing stripe %llu block %d\n",
3976                                (unsigned long long)sh->sector, disk_idx);
3977                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3978                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3979                         set_bit(R5_Wantcompute, &dev->flags);
3980                         sh->ops.target = disk_idx;
3981                         sh->ops.target2 = -1; /* no 2nd target */
3982                         s->req_compute = 1;
3983                         /* Careful: from this point on 'uptodate' is in the eye
3984                          * of raid_run_ops which services 'compute' operations
3985                          * before writes. R5_Wantcompute flags a block that will
3986                          * be R5_UPTODATE by the time it is needed for a
3987                          * subsequent operation.
3988                          */
3989                         s->uptodate++;
3990                         return 1;
3991                 } else if (s->uptodate == disks-2 && s->failed >= 2) {
3992                         /* Computing 2-failure is *very* expensive; only
3993                          * do it if failed >= 2
3994                          */
3995                         int other;
3996                         for (other = disks; other--; ) {
3997                                 if (other == disk_idx)
3998                                         continue;
3999                                 if (!test_bit(R5_UPTODATE,
4000                                       &sh->dev[other].flags))
4001                                         break;
4002                         }
4003                         BUG_ON(other < 0);
4004                         pr_debug("Computing stripe %llu blocks %d,%d\n",
4005                                (unsigned long long)sh->sector,
4006                                disk_idx, other);
4007                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
4008                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
4009                         set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
4010                         set_bit(R5_Wantcompute, &sh->dev[other].flags);
4011                         sh->ops.target = disk_idx;
4012                         sh->ops.target2 = other;
4013                         s->uptodate += 2;
4014                         s->req_compute = 1;
4015                         return 1;
4016                 } else if (test_bit(R5_Insync, &dev->flags)) {
4017                         set_bit(R5_LOCKED, &dev->flags);
4018                         set_bit(R5_Wantread, &dev->flags);
4019                         s->locked++;
4020                         pr_debug("Reading block %d (sync=%d)\n",
4021                                 disk_idx, s->syncing);
4022                 }
4023         }
4024
4025         return 0;
4026 }
4027
4028 /*
4029  * handle_stripe_fill - read or compute data to satisfy pending requests.
4030  */
4031 static void handle_stripe_fill(struct stripe_head *sh,
4032                                struct stripe_head_state *s,
4033                                int disks)
4034 {
4035         int i;
4036
4037         /* look for blocks to read/compute, skip this if a compute
4038          * is already in flight, or if the stripe contents are in the
4039          * midst of changing due to a write
4040          */
4041         if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
4042             !sh->reconstruct_state) {
4043
4044                 /*
4045                  * For degraded stripe with data in journal, do not handle
4046                  * read requests yet, instead, flush the stripe to raid
4047                  * disks first, this avoids handling complex rmw of write
4048                  * back cache (prexor with orig_page, and then xor with
4049                  * page) in the read path
4050                  */
4051                 if (s->to_read && s->injournal && s->failed) {
4052                         if (test_bit(STRIPE_R5C_CACHING, &sh->state))
4053                                 r5c_make_stripe_write_out(sh);
4054                         goto out;
4055                 }
4056
4057                 for (i = disks; i--; )
4058                         if (fetch_block(sh, s, i, disks))
4059                                 break;
4060         }
4061 out:
4062         set_bit(STRIPE_HANDLE, &sh->state);
4063 }
4064
4065 static void break_stripe_batch_list(struct stripe_head *head_sh,
4066                                     unsigned long handle_flags);
4067 /* handle_stripe_clean_event
4068  * any written block on an uptodate or failed drive can be returned.
4069  * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
4070  * never LOCKED, so we don't need to test 'failed' directly.
4071  */
4072 static void handle_stripe_clean_event(struct r5conf *conf,
4073         struct stripe_head *sh, int disks)
4074 {
4075         int i;
4076         struct r5dev *dev;
4077         int discard_pending = 0;
4078         struct stripe_head *head_sh = sh;
4079         bool do_endio = false;
4080
4081         for (i = disks; i--; )
4082                 if (sh->dev[i].written) {
4083                         dev = &sh->dev[i];
4084                         if (!test_bit(R5_LOCKED, &dev->flags) &&
4085                             (test_bit(R5_UPTODATE, &dev->flags) ||
4086                              test_bit(R5_Discard, &dev->flags) ||
4087                              test_bit(R5_SkipCopy, &dev->flags))) {
4088                                 /* We can return any write requests */
4089                                 struct bio *wbi, *wbi2;
4090                                 pr_debug("Return write for disc %d\n", i);
4091                                 if (test_and_clear_bit(R5_Discard, &dev->flags))
4092                                         clear_bit(R5_UPTODATE, &dev->flags);
4093                                 if (test_and_clear_bit(R5_SkipCopy, &dev->flags)) {
4094                                         WARN_ON(test_bit(R5_UPTODATE, &dev->flags));
4095                                 }
4096                                 do_endio = true;
4097
4098 returnbi:
4099                                 dev->page = dev->orig_page;
4100                                 wbi = dev->written;
4101                                 dev->written = NULL;
4102                                 while (wbi && wbi->bi_iter.bi_sector <
4103                                         dev->sector + RAID5_STRIPE_SECTORS(conf)) {
4104                                         wbi2 = r5_next_bio(conf, wbi, dev->sector);
4105                                         md_write_end(conf->mddev);
4106                                         bio_endio(wbi);
4107                                         wbi = wbi2;
4108                                 }
4109                                 md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
4110                                                    RAID5_STRIPE_SECTORS(conf),
4111                                                    !test_bit(STRIPE_DEGRADED, &sh->state),
4112                                                    0);
4113                                 if (head_sh->batch_head) {
4114                                         sh = list_first_entry(&sh->batch_list,
4115                                                               struct stripe_head,
4116                                                               batch_list);
4117                                         if (sh != head_sh) {
4118                                                 dev = &sh->dev[i];
4119                                                 goto returnbi;
4120                                         }
4121                                 }
4122                                 sh = head_sh;
4123                                 dev = &sh->dev[i];
4124                         } else if (test_bit(R5_Discard, &dev->flags))
4125                                 discard_pending = 1;
4126                 }
4127
4128         log_stripe_write_finished(sh);
4129
4130         if (!discard_pending &&
4131             test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
4132                 int hash;
4133                 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
4134                 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
4135                 if (sh->qd_idx >= 0) {
4136                         clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
4137                         clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
4138                 }
4139                 /* now that discard is done we can proceed with any sync */
4140                 clear_bit(STRIPE_DISCARD, &sh->state);
4141                 /*
4142                  * SCSI discard will change some bio fields and the stripe has
4143                  * no updated data, so remove it from hash list and the stripe
4144                  * will be reinitialized
4145                  */
4146 unhash:
4147                 hash = sh->hash_lock_index;
4148                 spin_lock_irq(conf->hash_locks + hash);
4149                 remove_hash(sh);
4150                 spin_unlock_irq(conf->hash_locks + hash);
4151                 if (head_sh->batch_head) {
4152                         sh = list_first_entry(&sh->batch_list,
4153                                               struct stripe_head, batch_list);
4154                         if (sh != head_sh)
4155                                         goto unhash;
4156                 }
4157                 sh = head_sh;
4158
4159                 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
4160                         set_bit(STRIPE_HANDLE, &sh->state);
4161
4162         }
4163
4164         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
4165                 if (atomic_dec_and_test(&conf->pending_full_writes))
4166                         md_wakeup_thread(conf->mddev->thread);
4167
4168         if (head_sh->batch_head && do_endio)
4169                 break_stripe_batch_list(head_sh, STRIPE_EXPAND_SYNC_FLAGS);
4170 }
4171
4172 /*
4173  * For RMW in write back cache, we need extra page in prexor to store the
4174  * old data. This page is stored in dev->orig_page.
4175  *
4176  * This function checks whether we have data for prexor. The exact logic
4177  * is:
4178  *       R5_UPTODATE && (!R5_InJournal || R5_OrigPageUPTDODATE)
4179  */
4180 static inline bool uptodate_for_rmw(struct r5dev *dev)
4181 {
4182         return (test_bit(R5_UPTODATE, &dev->flags)) &&
4183                 (!test_bit(R5_InJournal, &dev->flags) ||
4184                  test_bit(R5_OrigPageUPTDODATE, &dev->flags));
4185 }
4186
4187 static int handle_stripe_dirtying(struct r5conf *conf,
4188                                   struct stripe_head *sh,
4189                                   struct stripe_head_state *s,
4190                                   int disks)
4191 {
4192         int rmw = 0, rcw = 0, i;
4193         sector_t recovery_cp = conf->mddev->recovery_cp;
4194
4195         /* Check whether resync is now happening or should start.
4196          * If yes, then the array is dirty (after unclean shutdown or
4197          * initial creation), so parity in some stripes might be inconsistent.
4198          * In this case, we need to always do reconstruct-write, to ensure
4199          * that in case of drive failure or read-error correction, we
4200          * generate correct data from the parity.
4201          */
4202         if (conf->rmw_level == PARITY_DISABLE_RMW ||
4203             (recovery_cp < MaxSector && sh->sector >= recovery_cp &&
4204              s->failed == 0)) {
4205                 /* Calculate the real rcw later - for now make it
4206                  * look like rcw is cheaper
4207                  */
4208                 rcw = 1; rmw = 2;
4209                 pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
4210                          conf->rmw_level, (unsigned long long)recovery_cp,
4211                          (unsigned long long)sh->sector);
4212         } else for (i = disks; i--; ) {
4213                 /* would I have to read this buffer for read_modify_write */
4214                 struct r5dev *dev = &sh->dev[i];
4215                 if (((dev->towrite && !delay_towrite(conf, dev, s)) ||
4216                      i == sh->pd_idx || i == sh->qd_idx ||
4217                      test_bit(R5_InJournal, &dev->flags)) &&
4218                     !test_bit(R5_LOCKED, &dev->flags) &&
4219                     !(uptodate_for_rmw(dev) ||
4220                       test_bit(R5_Wantcompute, &dev->flags))) {
4221                         if (test_bit(R5_Insync, &dev->flags))
4222                                 rmw++;
4223                         else
4224                                 rmw += 2*disks;  /* cannot read it */
4225                 }
4226                 /* Would I have to read this buffer for reconstruct_write */
4227                 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
4228                     i != sh->pd_idx && i != sh->qd_idx &&
4229                     !test_bit(R5_LOCKED, &dev->flags) &&
4230                     !(test_bit(R5_UPTODATE, &dev->flags) ||
4231                       test_bit(R5_Wantcompute, &dev->flags))) {
4232                         if (test_bit(R5_Insync, &dev->flags))
4233                                 rcw++;
4234                         else
4235                                 rcw += 2*disks;
4236                 }
4237         }
4238
4239         pr_debug("for sector %llu state 0x%lx, rmw=%d rcw=%d\n",
4240                  (unsigned long long)sh->sector, sh->state, rmw, rcw);
4241         set_bit(STRIPE_HANDLE, &sh->state);
4242         if ((rmw < rcw || (rmw == rcw && conf->rmw_level == PARITY_PREFER_RMW)) && rmw > 0) {
4243                 /* prefer read-modify-write, but need to get some data */
4244                 if (conf->mddev->queue)
4245                         blk_add_trace_msg(conf->mddev->queue,
4246                                           "raid5 rmw %llu %d",
4247                                           (unsigned long long)sh->sector, rmw);
4248                 for (i = disks; i--; ) {
4249                         struct r5dev *dev = &sh->dev[i];
4250                         if (test_bit(R5_InJournal, &dev->flags) &&
4251                             dev->page == dev->orig_page &&
4252                             !test_bit(R5_LOCKED, &sh->dev[sh->pd_idx].flags)) {
4253                                 /* alloc page for prexor */
4254                                 struct page *p = alloc_page(GFP_NOIO);
4255
4256                                 if (p) {
4257                                         dev->orig_page = p;
4258                                         continue;
4259                                 }
4260
4261                                 /*
4262                                  * alloc_page() failed, try use
4263                                  * disk_info->extra_page
4264                                  */
4265                                 if (!test_and_set_bit(R5C_EXTRA_PAGE_IN_USE,
4266                                                       &conf->cache_state)) {
4267                                         r5c_use_extra_page(sh);
4268                                         break;
4269                                 }
4270
4271                                 /* extra_page in use, add to delayed_list */
4272                                 set_bit(STRIPE_DELAYED, &sh->state);
4273                                 s->waiting_extra_page = 1;
4274                                 return -EAGAIN;
4275                         }
4276                 }
4277
4278                 for (i = disks; i--; ) {
4279                         struct r5dev *dev = &sh->dev[i];
4280                         if (((dev->towrite && !delay_towrite(conf, dev, s)) ||
4281                              i == sh->pd_idx || i == sh->qd_idx ||
4282                              test_bit(R5_InJournal, &dev->flags)) &&
4283                             !test_bit(R5_LOCKED, &dev->flags) &&
4284                             !(uptodate_for_rmw(dev) ||
4285                               test_bit(R5_Wantcompute, &dev->flags)) &&
4286                             test_bit(R5_Insync, &dev->flags)) {
4287                                 if (test_bit(STRIPE_PREREAD_ACTIVE,
4288                                              &sh->state)) {
4289                                         pr_debug("Read_old block %d for r-m-w\n",
4290                                                  i);
4291                                         set_bit(R5_LOCKED, &dev->flags);
4292                                         set_bit(R5_Wantread, &dev->flags);
4293                                         s->locked++;
4294                                 } else
4295                                         set_bit(STRIPE_DELAYED, &sh->state);
4296                         }
4297                 }
4298         }
4299         if ((rcw < rmw || (rcw == rmw && conf->rmw_level != PARITY_PREFER_RMW)) && rcw > 0) {
4300                 /* want reconstruct write, but need to get some data */
4301                 int qread =0;
4302                 rcw = 0;
4303                 for (i = disks; i--; ) {
4304                         struct r5dev *dev = &sh->dev[i];
4305                         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
4306                             i != sh->pd_idx && i != sh->qd_idx &&
4307                             !test_bit(R5_LOCKED, &dev->flags) &&
4308                             !(test_bit(R5_UPTODATE, &dev->flags) ||
4309                               test_bit(R5_Wantcompute, &dev->flags))) {
4310                                 rcw++;
4311                                 if (test_bit(R5_Insync, &dev->flags) &&
4312                                     test_bit(STRIPE_PREREAD_ACTIVE,
4313                                              &sh->state)) {
4314                                         pr_debug("Read_old block "
4315                                                 "%d for Reconstruct\n", i);
4316                                         set_bit(R5_LOCKED, &dev->flags);
4317                                         set_bit(R5_Wantread, &dev->flags);
4318                                         s->locked++;
4319                                         qread++;
4320                                 } else
4321                                         set_bit(STRIPE_DELAYED, &sh->state);
4322                         }
4323                 }
4324                 if (rcw && conf->mddev->queue)
4325                         blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
4326                                           (unsigned long long)sh->sector,
4327                                           rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
4328         }
4329
4330         if (rcw > disks && rmw > disks &&
4331             !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4332                 set_bit(STRIPE_DELAYED, &sh->state);
4333
4334         /* now if nothing is locked, and if we have enough data,
4335          * we can start a write request
4336          */
4337         /* since handle_stripe can be called at any time we need to handle the
4338          * case where a compute block operation has been submitted and then a
4339          * subsequent call wants to start a write request.  raid_run_ops only
4340          * handles the case where compute block and reconstruct are requested
4341          * simultaneously.  If this is not the case then new writes need to be
4342          * held off until the compute completes.
4343          */
4344         if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
4345             (s->locked == 0 && (rcw == 0 || rmw == 0) &&
4346              !test_bit(STRIPE_BIT_DELAY, &sh->state)))
4347                 schedule_reconstruction(sh, s, rcw == 0, 0);
4348         return 0;
4349 }
4350
4351 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
4352                                 struct stripe_head_state *s, int disks)
4353 {
4354         struct r5dev *dev = NULL;
4355
4356         BUG_ON(sh->batch_head);
4357         set_bit(STRIPE_HANDLE, &sh->state);
4358
4359         switch (sh->check_state) {
4360         case check_state_idle:
4361                 /* start a new check operation if there are no failures */
4362                 if (s->failed == 0) {
4363                         BUG_ON(s->uptodate != disks);
4364                         sh->check_state = check_state_run;
4365                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
4366                         clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
4367                         s->uptodate--;
4368                         break;
4369                 }
4370                 dev = &sh->dev[s->failed_num[0]];
4371                 fallthrough;
4372         case check_state_compute_result:
4373                 sh->check_state = check_state_idle;
4374                 if (!dev)
4375                         dev = &sh->dev[sh->pd_idx];
4376
4377                 /* check that a write has not made the stripe insync */
4378                 if (test_bit(STRIPE_INSYNC, &sh->state))
4379                         break;
4380
4381                 /* either failed parity check, or recovery is happening */
4382                 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
4383                 BUG_ON(s->uptodate != disks);
4384
4385                 set_bit(R5_LOCKED, &dev->flags);
4386                 s->locked++;
4387                 set_bit(R5_Wantwrite, &dev->flags);
4388
4389                 clear_bit(STRIPE_DEGRADED, &sh->state);
4390                 set_bit(STRIPE_INSYNC, &sh->state);
4391                 break;
4392         case check_state_run:
4393                 break; /* we will be called again upon completion */
4394         case check_state_check_result:
4395                 sh->check_state = check_state_idle;
4396
4397                 /* if a failure occurred during the check operation, leave
4398                  * STRIPE_INSYNC not set and let the stripe be handled again
4399                  */
4400                 if (s->failed)
4401                         break;
4402
4403                 /* handle a successful check operation, if parity is correct
4404                  * we are done.  Otherwise update the mismatch count and repair
4405                  * parity if !MD_RECOVERY_CHECK
4406                  */
4407                 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
4408                         /* parity is correct (on disc,
4409                          * not in buffer any more)
4410                          */
4411                         set_bit(STRIPE_INSYNC, &sh->state);
4412                 else {
4413                         atomic64_add(RAID5_STRIPE_SECTORS(conf), &conf->mddev->resync_mismatches);
4414                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery)) {
4415                                 /* don't try to repair!! */
4416                                 set_bit(STRIPE_INSYNC, &sh->state);
4417                                 pr_warn_ratelimited("%s: mismatch sector in range "
4418                                                     "%llu-%llu\n", mdname(conf->mddev),
4419                                                     (unsigned long long) sh->sector,
4420                                                     (unsigned long long) sh->sector +
4421                                                     RAID5_STRIPE_SECTORS(conf));
4422                         } else {
4423                                 sh->check_state = check_state_compute_run;
4424                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
4425                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
4426                                 set_bit(R5_Wantcompute,
4427                                         &sh->dev[sh->pd_idx].flags);
4428                                 sh->ops.target = sh->pd_idx;
4429                                 sh->ops.target2 = -1;
4430                                 s->uptodate++;
4431                         }
4432                 }
4433                 break;
4434         case check_state_compute_run:
4435                 break;
4436         default:
4437                 pr_err("%s: unknown check_state: %d sector: %llu\n",
4438                        __func__, sh->check_state,
4439                        (unsigned long long) sh->sector);
4440                 BUG();
4441         }
4442 }
4443
4444 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
4445                                   struct stripe_head_state *s,
4446                                   int disks)
4447 {
4448         int pd_idx = sh->pd_idx;
4449         int qd_idx = sh->qd_idx;
4450         struct r5dev *dev;
4451
4452         BUG_ON(sh->batch_head);
4453         set_bit(STRIPE_HANDLE, &sh->state);
4454
4455         BUG_ON(s->failed > 2);
4456
4457         /* Want to check and possibly repair P and Q.
4458          * However there could be one 'failed' device, in which
4459          * case we can only check one of them, possibly using the
4460          * other to generate missing data
4461          */
4462
4463         switch (sh->check_state) {
4464         case check_state_idle:
4465                 /* start a new check operation if there are < 2 failures */
4466                 if (s->failed == s->q_failed) {
4467                         /* The only possible failed device holds Q, so it
4468                          * makes sense to check P (If anything else were failed,
4469                          * we would have used P to recreate it).
4470                          */
4471                         sh->check_state = check_state_run;
4472                 }
4473                 if (!s->q_failed && s->failed < 2) {
4474                         /* Q is not failed, and we didn't use it to generate
4475                          * anything, so it makes sense to check it
4476                          */
4477                         if (sh->check_state == check_state_run)
4478                                 sh->check_state = check_state_run_pq;
4479                         else
4480                                 sh->check_state = check_state_run_q;
4481                 }
4482
4483                 /* discard potentially stale zero_sum_result */
4484                 sh->ops.zero_sum_result = 0;
4485
4486                 if (sh->check_state == check_state_run) {
4487                         /* async_xor_zero_sum destroys the contents of P */
4488                         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
4489                         s->uptodate--;
4490                 }
4491                 if (sh->check_state >= check_state_run &&
4492                     sh->check_state <= check_state_run_pq) {
4493                         /* async_syndrome_zero_sum preserves P and Q, so
4494                          * no need to mark them !uptodate here
4495                          */
4496                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
4497                         break;
4498                 }
4499
4500                 /* we have 2-disk failure */
4501                 BUG_ON(s->failed != 2);
4502                 fallthrough;
4503         case check_state_compute_result:
4504                 sh->check_state = check_state_idle;
4505
4506                 /* check that a write has not made the stripe insync */
4507                 if (test_bit(STRIPE_INSYNC, &sh->state))
4508                         break;
4509
4510                 /* now write out any block on a failed drive,
4511                  * or P or Q if they were recomputed
4512                  */
4513                 dev = NULL;
4514                 if (s->failed == 2) {
4515                         dev = &sh->dev[s->failed_num[1]];
4516                         s->locked++;
4517                         set_bit(R5_LOCKED, &dev->flags);
4518                         set_bit(R5_Wantwrite, &dev->flags);
4519                 }
4520                 if (s->failed >= 1) {
4521                         dev = &sh->dev[s->failed_num[0]];
4522                         s->locked++;
4523                         set_bit(R5_LOCKED, &dev->flags);
4524                         set_bit(R5_Wantwrite, &dev->flags);
4525                 }
4526                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
4527                         dev = &sh->dev[pd_idx];
4528                         s->locked++;
4529                         set_bit(R5_LOCKED, &dev->flags);
4530                         set_bit(R5_Wantwrite, &dev->flags);
4531                 }
4532                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
4533                         dev = &sh->dev[qd_idx];
4534                         s->locked++;
4535                         set_bit(R5_LOCKED, &dev->flags);
4536                         set_bit(R5_Wantwrite, &dev->flags);
4537                 }
4538                 if (WARN_ONCE(dev && !test_bit(R5_UPTODATE, &dev->flags),
4539                               "%s: disk%td not up to date\n",
4540                               mdname(conf->mddev),
4541                               dev - (struct r5dev *) &sh->dev)) {
4542                         clear_bit(R5_LOCKED, &dev->flags);
4543                         clear_bit(R5_Wantwrite, &dev->flags);
4544                         s->locked--;
4545                 }
4546                 clear_bit(STRIPE_DEGRADED, &sh->state);
4547
4548                 set_bit(STRIPE_INSYNC, &sh->state);
4549                 break;
4550         case check_state_run:
4551         case check_state_run_q:
4552         case check_state_run_pq:
4553                 break; /* we will be called again upon completion */
4554         case check_state_check_result:
4555                 sh->check_state = check_state_idle;
4556
4557                 /* handle a successful check operation, if parity is correct
4558                  * we are done.  Otherwise update the mismatch count and repair
4559                  * parity if !MD_RECOVERY_CHECK
4560                  */
4561                 if (sh->ops.zero_sum_result == 0) {
4562                         /* both parities are correct */
4563                         if (!s->failed)
4564                                 set_bit(STRIPE_INSYNC, &sh->state);
4565                         else {
4566                                 /* in contrast to the raid5 case we can validate
4567                                  * parity, but still have a failure to write
4568                                  * back
4569                                  */
4570                                 sh->check_state = check_state_compute_result;
4571                                 /* Returning at this point means that we may go
4572                                  * off and bring p and/or q uptodate again so
4573                                  * we make sure to check zero_sum_result again
4574                                  * to verify if p or q need writeback
4575                                  */
4576                         }
4577                 } else {
4578                         atomic64_add(RAID5_STRIPE_SECTORS(conf), &conf->mddev->resync_mismatches);
4579                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery)) {
4580                                 /* don't try to repair!! */
4581                                 set_bit(STRIPE_INSYNC, &sh->state);
4582                                 pr_warn_ratelimited("%s: mismatch sector in range "
4583                                                     "%llu-%llu\n", mdname(conf->mddev),
4584                                                     (unsigned long long) sh->sector,
4585                                                     (unsigned long long) sh->sector +
4586                                                     RAID5_STRIPE_SECTORS(conf));
4587                         } else {
4588                                 int *target = &sh->ops.target;
4589
4590                                 sh->ops.target = -1;
4591                                 sh->ops.target2 = -1;
4592                                 sh->check_state = check_state_compute_run;
4593                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
4594                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
4595                                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
4596                                         set_bit(R5_Wantcompute,
4597                                                 &sh->dev[pd_idx].flags);
4598                                         *target = pd_idx;
4599                                         target = &sh->ops.target2;
4600                                         s->uptodate++;
4601                                 }
4602                                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
4603                                         set_bit(R5_Wantcompute,
4604                                                 &sh->dev[qd_idx].flags);
4605                                         *target = qd_idx;
4606                                         s->uptodate++;
4607                                 }
4608                         }
4609                 }
4610                 break;
4611         case check_state_compute_run:
4612                 break;
4613         default:
4614                 pr_warn("%s: unknown check_state: %d sector: %llu\n",
4615                         __func__, sh->check_state,
4616                         (unsigned long long) sh->sector);
4617                 BUG();
4618         }
4619 }
4620
4621 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
4622 {
4623         int i;
4624
4625         /* We have read all the blocks in this stripe and now we need to
4626          * copy some of them into a target stripe for expand.
4627          */
4628         struct dma_async_tx_descriptor *tx = NULL;
4629         BUG_ON(sh->batch_head);
4630         clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4631         for (i = 0; i < sh->disks; i++)
4632                 if (i != sh->pd_idx && i != sh->qd_idx) {
4633                         int dd_idx, j;
4634                         struct stripe_head *sh2;
4635                         struct async_submit_ctl submit;
4636
4637                         sector_t bn = raid5_compute_blocknr(sh, i, 1);
4638                         sector_t s = raid5_compute_sector(conf, bn, 0,
4639                                                           &dd_idx, NULL);
4640                         sh2 = raid5_get_active_stripe(conf, NULL, s,
4641                                 R5_GAS_NOBLOCK | R5_GAS_NOQUIESCE);
4642                         if (sh2 == NULL)
4643                                 /* so far only the early blocks of this stripe
4644                                  * have been requested.  When later blocks
4645                                  * get requested, we will try again
4646                                  */
4647                                 continue;
4648                         if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
4649                            test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
4650                                 /* must have already done this block */
4651                                 raid5_release_stripe(sh2);
4652                                 continue;
4653                         }
4654
4655                         /* place all the copies on one channel */
4656                         init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
4657                         tx = async_memcpy(sh2->dev[dd_idx].page,
4658                                           sh->dev[i].page, sh2->dev[dd_idx].offset,
4659                                           sh->dev[i].offset, RAID5_STRIPE_SIZE(conf),
4660                                           &submit);
4661
4662                         set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
4663                         set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
4664                         for (j = 0; j < conf->raid_disks; j++)
4665                                 if (j != sh2->pd_idx &&
4666                                     j != sh2->qd_idx &&
4667                                     !test_bit(R5_Expanded, &sh2->dev[j].flags))
4668                                         break;
4669                         if (j == conf->raid_disks) {
4670                                 set_bit(STRIPE_EXPAND_READY, &sh2->state);
4671                                 set_bit(STRIPE_HANDLE, &sh2->state);
4672                         }
4673                         raid5_release_stripe(sh2);
4674
4675                 }
4676         /* done submitting copies, wait for them to complete */
4677         async_tx_quiesce(&tx);
4678 }
4679
4680 /*
4681  * handle_stripe - do things to a stripe.
4682  *
4683  * We lock the stripe by setting STRIPE_ACTIVE and then examine the
4684  * state of various bits to see what needs to be done.
4685  * Possible results:
4686  *    return some read requests which now have data
4687  *    return some write requests which are safely on storage
4688  *    schedule a read on some buffers
4689  *    schedule a write of some buffers
4690  *    return confirmation of parity correctness
4691  *
4692  */
4693
4694 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
4695 {
4696         struct r5conf *conf = sh->raid_conf;
4697         int disks = sh->disks;
4698         struct r5dev *dev;
4699         int i;
4700         int do_recovery = 0;
4701
4702         memset(s, 0, sizeof(*s));
4703
4704         s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state) && !sh->batch_head;
4705         s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state) && !sh->batch_head;
4706         s->failed_num[0] = -1;
4707         s->failed_num[1] = -1;
4708         s->log_failed = r5l_log_disk_error(conf);
4709
4710         /* Now to look around and see what can be done */
4711         rcu_read_lock();
4712         for (i=disks; i--; ) {
4713                 struct md_rdev *rdev;
4714                 sector_t first_bad;
4715                 int bad_sectors;
4716                 int is_bad = 0;
4717
4718                 dev = &sh->dev[i];
4719
4720                 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4721                          i, dev->flags,
4722                          dev->toread, dev->towrite, dev->written);
4723                 /* maybe we can reply to a read
4724                  *
4725                  * new wantfill requests are only permitted while
4726                  * ops_complete_biofill is guaranteed to be inactive
4727                  */
4728                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
4729                     !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
4730                         set_bit(R5_Wantfill, &dev->flags);
4731
4732                 /* now count some things */
4733                 if (test_bit(R5_LOCKED, &dev->flags))
4734                         s->locked++;
4735                 if (test_bit(R5_UPTODATE, &dev->flags))
4736                         s->uptodate++;
4737                 if (test_bit(R5_Wantcompute, &dev->flags)) {
4738                         s->compute++;
4739                         BUG_ON(s->compute > 2);
4740                 }
4741
4742                 if (test_bit(R5_Wantfill, &dev->flags))
4743                         s->to_fill++;
4744                 else if (dev->toread)
4745                         s->to_read++;
4746                 if (dev->towrite) {
4747                         s->to_write++;
4748                         if (!test_bit(R5_OVERWRITE, &dev->flags))
4749                                 s->non_overwrite++;
4750                 }
4751                 if (dev->written)
4752                         s->written++;
4753                 /* Prefer to use the replacement for reads, but only
4754                  * if it is recovered enough and has no bad blocks.
4755                  */
4756                 rdev = rcu_dereference(conf->disks[i].replacement);
4757                 if (rdev && !test_bit(Faulty, &rdev->flags) &&
4758                     rdev->recovery_offset >= sh->sector + RAID5_STRIPE_SECTORS(conf) &&
4759                     !is_badblock(rdev, sh->sector, RAID5_STRIPE_SECTORS(conf),
4760                                  &first_bad, &bad_sectors))
4761                         set_bit(R5_ReadRepl, &dev->flags);
4762                 else {
4763                         if (rdev && !test_bit(Faulty, &rdev->flags))
4764                                 set_bit(R5_NeedReplace, &dev->flags);
4765                         else
4766                                 clear_bit(R5_NeedReplace, &dev->flags);
4767                         rdev = rcu_dereference(conf->disks[i].rdev);
4768                         clear_bit(R5_ReadRepl, &dev->flags);
4769                 }
4770                 if (rdev && test_bit(Faulty, &rdev->flags))
4771                         rdev = NULL;
4772                 if (rdev) {
4773                         is_bad = is_badblock(rdev, sh->sector, RAID5_STRIPE_SECTORS(conf),
4774                                              &first_bad, &bad_sectors);
4775                         if (s->blocked_rdev == NULL
4776                             && (test_bit(Blocked, &rdev->flags)
4777                                 || is_bad < 0)) {
4778                                 if (is_bad < 0)
4779                                         set_bit(BlockedBadBlocks,
4780                                                 &rdev->flags);
4781                                 s->blocked_rdev = rdev;
4782                                 atomic_inc(&rdev->nr_pending);
4783                         }
4784                 }
4785                 clear_bit(R5_Insync, &dev->flags);
4786                 if (!rdev)
4787                         /* Not in-sync */;
4788                 else if (is_bad) {
4789                         /* also not in-sync */
4790                         if (!test_bit(WriteErrorSeen, &rdev->flags) &&
4791                             test_bit(R5_UPTODATE, &dev->flags)) {
4792                                 /* treat as in-sync, but with a read error
4793                                  * which we can now try to correct
4794                                  */
4795                                 set_bit(R5_Insync, &dev->flags);
4796                                 set_bit(R5_ReadError, &dev->flags);
4797                         }
4798                 } else if (test_bit(In_sync, &rdev->flags))
4799                         set_bit(R5_Insync, &dev->flags);
4800                 else if (sh->sector + RAID5_STRIPE_SECTORS(conf) <= rdev->recovery_offset)
4801                         /* in sync if before recovery_offset */
4802                         set_bit(R5_Insync, &dev->flags);
4803                 else if (test_bit(R5_UPTODATE, &dev->flags) &&
4804                          test_bit(R5_Expanded, &dev->flags))
4805                         /* If we've reshaped into here, we assume it is Insync.
4806                          * We will shortly update recovery_offset to make
4807                          * it official.
4808                          */
4809                         set_bit(R5_Insync, &dev->flags);
4810
4811                 if (test_bit(R5_WriteError, &dev->flags)) {
4812                         /* This flag does not apply to '.replacement'
4813                          * only to .rdev, so make sure to check that*/
4814                         struct md_rdev *rdev2 = rcu_dereference(
4815                                 conf->disks[i].rdev);
4816                         if (rdev2 == rdev)
4817                                 clear_bit(R5_Insync, &dev->flags);
4818                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4819                                 s->handle_bad_blocks = 1;
4820                                 atomic_inc(&rdev2->nr_pending);
4821                         } else
4822                                 clear_bit(R5_WriteError, &dev->flags);
4823                 }
4824                 if (test_bit(R5_MadeGood, &dev->flags)) {
4825                         /* This flag does not apply to '.replacement'
4826                          * only to .rdev, so make sure to check that*/
4827                         struct md_rdev *rdev2 = rcu_dereference(
4828                                 conf->disks[i].rdev);
4829                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4830                                 s->handle_bad_blocks = 1;
4831                                 atomic_inc(&rdev2->nr_pending);
4832                         } else
4833                                 clear_bit(R5_MadeGood, &dev->flags);
4834                 }
4835                 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
4836                         struct md_rdev *rdev2 = rcu_dereference(
4837                                 conf->disks[i].replacement);
4838                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4839                                 s->handle_bad_blocks = 1;
4840                                 atomic_inc(&rdev2->nr_pending);
4841                         } else
4842                                 clear_bit(R5_MadeGoodRepl, &dev->flags);
4843                 }
4844                 if (!test_bit(R5_Insync, &dev->flags)) {
4845                         /* The ReadError flag will just be confusing now */
4846                         clear_bit(R5_ReadError, &dev->flags);
4847                         clear_bit(R5_ReWrite, &dev->flags);
4848                 }
4849                 if (test_bit(R5_ReadError, &dev->flags))
4850                         clear_bit(R5_Insync, &dev->flags);
4851                 if (!test_bit(R5_Insync, &dev->flags)) {
4852                         if (s->failed < 2)
4853                                 s->failed_num[s->failed] = i;
4854                         s->failed++;
4855                         if (rdev && !test_bit(Faulty, &rdev->flags))
4856                                 do_recovery = 1;
4857                         else if (!rdev) {
4858                                 rdev = rcu_dereference(
4859                                     conf->disks[i].replacement);
4860                                 if (rdev && !test_bit(Faulty, &rdev->flags))
4861                                         do_recovery = 1;
4862                         }
4863                 }
4864
4865                 if (test_bit(R5_InJournal, &dev->flags))
4866                         s->injournal++;
4867                 if (test_bit(R5_InJournal, &dev->flags) && dev->written)
4868                         s->just_cached++;
4869         }
4870         if (test_bit(STRIPE_SYNCING, &sh->state)) {
4871                 /* If there is a failed device being replaced,
4872                  *     we must be recovering.
4873                  * else if we are after recovery_cp, we must be syncing
4874                  * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4875                  * else we can only be replacing
4876                  * sync and recovery both need to read all devices, and so
4877                  * use the same flag.
4878                  */
4879                 if (do_recovery ||
4880                     sh->sector >= conf->mddev->recovery_cp ||
4881                     test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
4882                         s->syncing = 1;
4883                 else
4884                         s->replacing = 1;
4885         }
4886         rcu_read_unlock();
4887 }
4888
4889 /*
4890  * Return '1' if this is a member of batch, or '0' if it is a lone stripe or
4891  * a head which can now be handled.
4892  */
4893 static int clear_batch_ready(struct stripe_head *sh)
4894 {
4895         struct stripe_head *tmp;
4896         if (!test_and_clear_bit(STRIPE_BATCH_READY, &sh->state))
4897                 return (sh->batch_head && sh->batch_head != sh);
4898         spin_lock(&sh->stripe_lock);
4899         if (!sh->batch_head) {
4900                 spin_unlock(&sh->stripe_lock);
4901                 return 0;
4902         }
4903
4904         /*
4905          * this stripe could be added to a batch list before we check
4906          * BATCH_READY, skips it
4907          */
4908         if (sh->batch_head != sh) {
4909                 spin_unlock(&sh->stripe_lock);
4910                 return 1;
4911         }
4912         spin_lock(&sh->batch_lock);
4913         list_for_each_entry(tmp, &sh->batch_list, batch_list)
4914                 clear_bit(STRIPE_BATCH_READY, &tmp->state);
4915         spin_unlock(&sh->batch_lock);
4916         spin_unlock(&sh->stripe_lock);
4917
4918         /*
4919          * BATCH_READY is cleared, no new stripes can be added.
4920          * batch_list can be accessed without lock
4921          */
4922         return 0;
4923 }
4924
4925 static void break_stripe_batch_list(struct stripe_head *head_sh,
4926                                     unsigned long handle_flags)
4927 {
4928         struct stripe_head *sh, *next;
4929         int i;
4930         int do_wakeup = 0;
4931
4932         list_for_each_entry_safe(sh, next, &head_sh->batch_list, batch_list) {
4933
4934                 list_del_init(&sh->batch_list);
4935
4936                 WARN_ONCE(sh->state & ((1 << STRIPE_ACTIVE) |
4937                                           (1 << STRIPE_SYNCING) |
4938                                           (1 << STRIPE_REPLACED) |
4939                                           (1 << STRIPE_DELAYED) |
4940                                           (1 << STRIPE_BIT_DELAY) |
4941                                           (1 << STRIPE_FULL_WRITE) |
4942                                           (1 << STRIPE_BIOFILL_RUN) |
4943                                           (1 << STRIPE_COMPUTE_RUN)  |
4944                                           (1 << STRIPE_DISCARD) |
4945                                           (1 << STRIPE_BATCH_READY) |
4946                                           (1 << STRIPE_BATCH_ERR) |
4947                                           (1 << STRIPE_BITMAP_PENDING)),
4948                         "stripe state: %lx\n", sh->state);
4949                 WARN_ONCE(head_sh->state & ((1 << STRIPE_DISCARD) |
4950                                               (1 << STRIPE_REPLACED)),
4951                         "head stripe state: %lx\n", head_sh->state);
4952
4953                 set_mask_bits(&sh->state, ~(STRIPE_EXPAND_SYNC_FLAGS |
4954                                             (1 << STRIPE_PREREAD_ACTIVE) |
4955                                             (1 << STRIPE_DEGRADED) |
4956                                             (1 << STRIPE_ON_UNPLUG_LIST)),
4957                               head_sh->state & (1 << STRIPE_INSYNC));
4958
4959                 sh->check_state = head_sh->check_state;
4960                 sh->reconstruct_state = head_sh->reconstruct_state;
4961                 spin_lock_irq(&sh->stripe_lock);
4962                 sh->batch_head = NULL;
4963                 spin_unlock_irq(&sh->stripe_lock);
4964                 for (i = 0; i < sh->disks; i++) {
4965                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
4966                                 do_wakeup = 1;
4967                         sh->dev[i].flags = head_sh->dev[i].flags &
4968                                 (~((1 << R5_WriteError) | (1 << R5_Overlap)));
4969                 }
4970                 if (handle_flags == 0 ||
4971                     sh->state & handle_flags)
4972                         set_bit(STRIPE_HANDLE, &sh->state);
4973                 raid5_release_stripe(sh);
4974         }
4975         spin_lock_irq(&head_sh->stripe_lock);
4976         head_sh->batch_head = NULL;
4977         spin_unlock_irq(&head_sh->stripe_lock);
4978         for (i = 0; i < head_sh->disks; i++)
4979                 if (test_and_clear_bit(R5_Overlap, &head_sh->dev[i].flags))
4980                         do_wakeup = 1;
4981         if (head_sh->state & handle_flags)
4982                 set_bit(STRIPE_HANDLE, &head_sh->state);
4983
4984         if (do_wakeup)
4985                 wake_up(&head_sh->raid_conf->wait_for_overlap);
4986 }
4987
4988 static void handle_stripe(struct stripe_head *sh)
4989 {
4990         struct stripe_head_state s;
4991         struct r5conf *conf = sh->raid_conf;
4992         int i;
4993         int prexor;
4994         int disks = sh->disks;
4995         struct r5dev *pdev, *qdev;
4996
4997         clear_bit(STRIPE_HANDLE, &sh->state);
4998
4999         /*
5000          * handle_stripe should not continue handle the batched stripe, only
5001          * the head of batch list or lone stripe can continue. Otherwise we
5002          * could see break_stripe_batch_list warns about the STRIPE_ACTIVE
5003          * is set for the batched stripe.
5004          */
5005         if (clear_batch_ready(sh))
5006                 return;
5007
5008         if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
5009                 /* already being handled, ensure it gets handled
5010                  * again when current action finishes */
5011                 set_bit(STRIPE_HANDLE, &sh->state);
5012                 return;
5013         }
5014
5015         if (test_and_clear_bit(STRIPE_BATCH_ERR, &sh->state))
5016                 break_stripe_batch_list(sh, 0);
5017
5018         if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) && !sh->batch_head) {
5019                 spin_lock(&sh->stripe_lock);
5020                 /*
5021                  * Cannot process 'sync' concurrently with 'discard'.
5022                  * Flush data in r5cache before 'sync'.
5023                  */
5024                 if (!test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) &&
5025                     !test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state) &&
5026                     !test_bit(STRIPE_DISCARD, &sh->state) &&
5027                     test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
5028                         set_bit(STRIPE_SYNCING, &sh->state);
5029                         clear_bit(STRIPE_INSYNC, &sh->state);
5030                         clear_bit(STRIPE_REPLACED, &sh->state);
5031                 }
5032                 spin_unlock(&sh->stripe_lock);
5033         }
5034         clear_bit(STRIPE_DELAYED, &sh->state);
5035
5036         pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
5037                 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
5038                (unsigned long long)sh->sector, sh->state,
5039                atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
5040                sh->check_state, sh->reconstruct_state);
5041
5042         analyse_stripe(sh, &s);
5043
5044         if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
5045                 goto finish;
5046
5047         if (s.handle_bad_blocks ||
5048             test_bit(MD_SB_CHANGE_PENDING, &conf->mddev->sb_flags)) {
5049                 set_bit(STRIPE_HANDLE, &sh->state);
5050                 goto finish;
5051         }
5052
5053         if (unlikely(s.blocked_rdev)) {
5054                 if (s.syncing || s.expanding || s.expanded ||
5055                     s.replacing || s.to_write || s.written) {
5056                         set_bit(STRIPE_HANDLE, &sh->state);
5057                         goto finish;
5058                 }
5059                 /* There is nothing for the blocked_rdev to block */
5060                 rdev_dec_pending(s.blocked_rdev, conf->mddev);
5061                 s.blocked_rdev = NULL;
5062         }
5063
5064         if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
5065                 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
5066                 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
5067         }
5068
5069         pr_debug("locked=%d uptodate=%d to_read=%d"
5070                " to_write=%d failed=%d failed_num=%d,%d\n",
5071                s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
5072                s.failed_num[0], s.failed_num[1]);
5073         /*
5074          * check if the array has lost more than max_degraded devices and,
5075          * if so, some requests might need to be failed.
5076          *
5077          * When journal device failed (log_failed), we will only process
5078          * the stripe if there is data need write to raid disks
5079          */
5080         if (s.failed > conf->max_degraded ||
5081             (s.log_failed && s.injournal == 0)) {
5082                 sh->check_state = 0;
5083                 sh->reconstruct_state = 0;
5084                 break_stripe_batch_list(sh, 0);
5085                 if (s.to_read+s.to_write+s.written)
5086                         handle_failed_stripe(conf, sh, &s, disks);
5087                 if (s.syncing + s.replacing)
5088                         handle_failed_sync(conf, sh, &s);
5089         }
5090
5091         /* Now we check to see if any write operations have recently
5092          * completed
5093          */
5094         prexor = 0;
5095         if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
5096                 prexor = 1;
5097         if (sh->reconstruct_state == reconstruct_state_drain_result ||
5098             sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
5099                 sh->reconstruct_state = reconstruct_state_idle;
5100
5101                 /* All the 'written' buffers and the parity block are ready to
5102                  * be written back to disk
5103                  */
5104                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
5105                        !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
5106                 BUG_ON(sh->qd_idx >= 0 &&
5107                        !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
5108                        !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
5109                 for (i = disks; i--; ) {
5110                         struct r5dev *dev = &sh->dev[i];
5111                         if (test_bit(R5_LOCKED, &dev->flags) &&
5112                                 (i == sh->pd_idx || i == sh->qd_idx ||
5113                                  dev->written || test_bit(R5_InJournal,
5114                                                           &dev->flags))) {
5115                                 pr_debug("Writing block %d\n", i);
5116                                 set_bit(R5_Wantwrite, &dev->flags);
5117                                 if (prexor)
5118                                         continue;
5119                                 if (s.failed > 1)
5120                                         continue;
5121                                 if (!test_bit(R5_Insync, &dev->flags) ||
5122                                     ((i == sh->pd_idx || i == sh->qd_idx)  &&
5123                                      s.failed == 0))
5124                                         set_bit(STRIPE_INSYNC, &sh->state);
5125                         }
5126                 }
5127                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5128                         s.dec_preread_active = 1;
5129         }
5130
5131         /*
5132          * might be able to return some write requests if the parity blocks
5133          * are safe, or on a failed drive
5134          */
5135         pdev = &sh->dev[sh->pd_idx];
5136         s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
5137                 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
5138         qdev = &sh->dev[sh->qd_idx];
5139         s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
5140                 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
5141                 || conf->level < 6;
5142
5143         if (s.written &&
5144             (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
5145                              && !test_bit(R5_LOCKED, &pdev->flags)
5146                              && (test_bit(R5_UPTODATE, &pdev->flags) ||
5147                                  test_bit(R5_Discard, &pdev->flags))))) &&
5148             (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
5149                              && !test_bit(R5_LOCKED, &qdev->flags)
5150                              && (test_bit(R5_UPTODATE, &qdev->flags) ||
5151                                  test_bit(R5_Discard, &qdev->flags))))))
5152                 handle_stripe_clean_event(conf, sh, disks);
5153
5154         if (s.just_cached)
5155                 r5c_handle_cached_data_endio(conf, sh, disks);
5156         log_stripe_write_finished(sh);
5157
5158         /* Now we might consider reading some blocks, either to check/generate
5159          * parity, or to satisfy requests
5160          * or to load a block that is being partially written.
5161          */
5162         if (s.to_read || s.non_overwrite
5163             || (s.to_write && s.failed)
5164             || (s.syncing && (s.uptodate + s.compute < disks))
5165             || s.replacing
5166             || s.expanding)
5167                 handle_stripe_fill(sh, &s, disks);
5168
5169         /*
5170          * When the stripe finishes full journal write cycle (write to journal
5171          * and raid disk), this is the clean up procedure so it is ready for
5172          * next operation.
5173          */
5174         r5c_finish_stripe_write_out(conf, sh, &s);
5175
5176         /*
5177          * Now to consider new write requests, cache write back and what else,
5178          * if anything should be read.  We do not handle new writes when:
5179          * 1/ A 'write' operation (copy+xor) is already in flight.
5180          * 2/ A 'check' operation is in flight, as it may clobber the parity
5181          *    block.
5182          * 3/ A r5c cache log write is in flight.
5183          */
5184
5185         if (!sh->reconstruct_state && !sh->check_state && !sh->log_io) {
5186                 if (!r5c_is_writeback(conf->log)) {
5187                         if (s.to_write)
5188                                 handle_stripe_dirtying(conf, sh, &s, disks);
5189                 } else { /* write back cache */
5190                         int ret = 0;
5191
5192                         /* First, try handle writes in caching phase */
5193                         if (s.to_write)
5194                                 ret = r5c_try_caching_write(conf, sh, &s,
5195                                                             disks);
5196                         /*
5197                          * If caching phase failed: ret == -EAGAIN
5198                          *    OR
5199                          * stripe under reclaim: !caching && injournal
5200                          *
5201                          * fall back to handle_stripe_dirtying()
5202                          */
5203                         if (ret == -EAGAIN ||
5204                             /* stripe under reclaim: !caching && injournal */
5205                             (!test_bit(STRIPE_R5C_CACHING, &sh->state) &&
5206                              s.injournal > 0)) {
5207                                 ret = handle_stripe_dirtying(conf, sh, &s,
5208                                                              disks);
5209                                 if (ret == -EAGAIN)
5210                                         goto finish;
5211                         }
5212                 }
5213         }
5214
5215         /* maybe we need to check and possibly fix the parity for this stripe
5216          * Any reads will already have been scheduled, so we just see if enough
5217          * data is available.  The parity check is held off while parity
5218          * dependent operations are in flight.
5219          */
5220         if (sh->check_state ||
5221             (s.syncing && s.locked == 0 &&
5222              !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
5223              !test_bit(STRIPE_INSYNC, &sh->state))) {
5224                 if (conf->level == 6)
5225                         handle_parity_checks6(conf, sh, &s, disks);
5226                 else
5227                         handle_parity_checks5(conf, sh, &s, disks);
5228         }
5229
5230         if ((s.replacing || s.syncing) && s.locked == 0
5231             && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
5232             && !test_bit(STRIPE_REPLACED, &sh->state)) {
5233                 /* Write out to replacement devices where possible */
5234                 for (i = 0; i < conf->raid_disks; i++)
5235                         if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
5236                                 WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
5237                                 set_bit(R5_WantReplace, &sh->dev[i].flags);
5238                                 set_bit(R5_LOCKED, &sh->dev[i].flags);
5239                                 s.locked++;
5240                         }
5241                 if (s.replacing)
5242                         set_bit(STRIPE_INSYNC, &sh->state);
5243                 set_bit(STRIPE_REPLACED, &sh->state);
5244         }
5245         if ((s.syncing || s.replacing) && s.locked == 0 &&
5246             !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
5247             test_bit(STRIPE_INSYNC, &sh->state)) {
5248                 md_done_sync(conf->mddev, RAID5_STRIPE_SECTORS(conf), 1);
5249                 clear_bit(STRIPE_SYNCING, &sh->state);
5250                 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
5251                         wake_up(&conf->wait_for_overlap);
5252         }
5253
5254         /* If the failed drives are just a ReadError, then we might need
5255          * to progress the repair/check process
5256          */
5257         if (s.failed <= conf->max_degraded && !conf->mddev->ro)
5258                 for (i = 0; i < s.failed; i++) {
5259                         struct r5dev *dev = &sh->dev[s.failed_num[i]];
5260                         if (test_bit(R5_ReadError, &dev->flags)
5261                             && !test_bit(R5_LOCKED, &dev->flags)
5262                             && test_bit(R5_UPTODATE, &dev->flags)
5263                                 ) {
5264                                 if (!test_bit(R5_ReWrite, &dev->flags)) {
5265                                         set_bit(R5_Wantwrite, &dev->flags);
5266                                         set_bit(R5_ReWrite, &dev->flags);
5267                                 } else
5268                                         /* let's read it back */
5269                                         set_bit(R5_Wantread, &dev->flags);
5270                                 set_bit(R5_LOCKED, &dev->flags);
5271                                 s.locked++;
5272                         }
5273                 }
5274
5275         /* Finish reconstruct operations initiated by the expansion process */
5276         if (sh->reconstruct_state == reconstruct_state_result) {
5277                 struct stripe_head *sh_src
5278                         = raid5_get_active_stripe(conf, NULL, sh->sector,
5279                                         R5_GAS_PREVIOUS | R5_GAS_NOBLOCK |
5280                                         R5_GAS_NOQUIESCE);
5281                 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
5282                         /* sh cannot be written until sh_src has been read.
5283                          * so arrange for sh to be delayed a little
5284                          */
5285                         set_bit(STRIPE_DELAYED, &sh->state);
5286                         set_bit(STRIPE_HANDLE, &sh->state);
5287                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
5288                                               &sh_src->state))
5289                                 atomic_inc(&conf->preread_active_stripes);
5290                         raid5_release_stripe(sh_src);
5291                         goto finish;
5292                 }
5293                 if (sh_src)
5294                         raid5_release_stripe(sh_src);
5295
5296                 sh->reconstruct_state = reconstruct_state_idle;
5297                 clear_bit(STRIPE_EXPANDING, &sh->state);
5298                 for (i = conf->raid_disks; i--; ) {
5299                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
5300                         set_bit(R5_LOCKED, &sh->dev[i].flags);
5301                         s.locked++;
5302                 }
5303         }
5304
5305         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
5306             !sh->reconstruct_state) {
5307                 /* Need to write out all blocks after computing parity */
5308                 sh->disks = conf->raid_disks;
5309                 stripe_set_idx(sh->sector, conf, 0, sh);
5310                 schedule_reconstruction(sh, &s, 1, 1);
5311         } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
5312                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
5313                 atomic_dec(&conf->reshape_stripes);
5314                 wake_up(&conf->wait_for_overlap);
5315                 md_done_sync(conf->mddev, RAID5_STRIPE_SECTORS(conf), 1);
5316         }
5317
5318         if (s.expanding && s.locked == 0 &&
5319             !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
5320                 handle_stripe_expansion(conf, sh);
5321
5322 finish:
5323         /* wait for this device to become unblocked */
5324         if (unlikely(s.blocked_rdev)) {
5325                 if (conf->mddev->external)
5326                         md_wait_for_blocked_rdev(s.blocked_rdev,
5327                                                  conf->mddev);
5328                 else
5329                         /* Internal metadata will immediately
5330                          * be written by raid5d, so we don't
5331                          * need to wait here.
5332                          */
5333                         rdev_dec_pending(s.blocked_rdev,
5334                                          conf->mddev);
5335         }
5336
5337         if (s.handle_bad_blocks)
5338                 for (i = disks; i--; ) {
5339                         struct md_rdev *rdev;
5340                         struct r5dev *dev = &sh->dev[i];
5341                         if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
5342                                 /* We own a safe reference to the rdev */
5343                                 rdev = rdev_pend_deref(conf->disks[i].rdev);
5344                                 if (!rdev_set_badblocks(rdev, sh->sector,
5345                                                         RAID5_STRIPE_SECTORS(conf), 0))
5346                                         md_error(conf->mddev, rdev);
5347                                 rdev_dec_pending(rdev, conf->mddev);
5348                         }
5349                         if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
5350                                 rdev = rdev_pend_deref(conf->disks[i].rdev);
5351                                 rdev_clear_badblocks(rdev, sh->sector,
5352                                                      RAID5_STRIPE_SECTORS(conf), 0);
5353                                 rdev_dec_pending(rdev, conf->mddev);
5354                         }
5355                         if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
5356                                 rdev = rdev_pend_deref(conf->disks[i].replacement);
5357                                 if (!rdev)
5358                                         /* rdev have been moved down */
5359                                         rdev = rdev_pend_deref(conf->disks[i].rdev);
5360                                 rdev_clear_badblocks(rdev, sh->sector,
5361                                                      RAID5_STRIPE_SECTORS(conf), 0);
5362                                 rdev_dec_pending(rdev, conf->mddev);
5363                         }
5364                 }
5365
5366         if (s.ops_request)
5367                 raid_run_ops(sh, s.ops_request);
5368
5369         ops_run_io(sh, &s);
5370
5371         if (s.dec_preread_active) {
5372                 /* We delay this until after ops_run_io so that if make_request
5373                  * is waiting on a flush, it won't continue until the writes
5374                  * have actually been submitted.
5375                  */
5376                 atomic_dec(&conf->preread_active_stripes);
5377                 if (atomic_read(&conf->preread_active_stripes) <
5378                     IO_THRESHOLD)
5379                         md_wakeup_thread(conf->mddev->thread);
5380         }
5381
5382         clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
5383 }
5384
5385 static void raid5_activate_delayed(struct r5conf *conf)
5386         __must_hold(&conf->device_lock)
5387 {
5388         if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
5389                 while (!list_empty(&conf->delayed_list)) {
5390                         struct list_head *l = conf->delayed_list.next;
5391                         struct stripe_head *sh;
5392                         sh = list_entry(l, struct stripe_head, lru);
5393                         list_del_init(l);
5394                         clear_bit(STRIPE_DELAYED, &sh->state);
5395                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5396                                 atomic_inc(&conf->preread_active_stripes);
5397                         list_add_tail(&sh->lru, &conf->hold_list);
5398                         raid5_wakeup_stripe_thread(sh);
5399                 }
5400         }
5401 }
5402
5403 static void activate_bit_delay(struct r5conf *conf,
5404                 struct list_head *temp_inactive_list)
5405         __must_hold(&conf->device_lock)
5406 {
5407         struct list_head head;
5408         list_add(&head, &conf->bitmap_list);
5409         list_del_init(&conf->bitmap_list);
5410         while (!list_empty(&head)) {
5411                 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
5412                 int hash;
5413                 list_del_init(&sh->lru);
5414                 atomic_inc(&sh->count);
5415                 hash = sh->hash_lock_index;
5416                 __release_stripe(conf, sh, &temp_inactive_list[hash]);
5417         }
5418 }
5419
5420 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
5421 {
5422         struct r5conf *conf = mddev->private;
5423         sector_t sector = bio->bi_iter.bi_sector;
5424         unsigned int chunk_sectors;
5425         unsigned int bio_sectors = bio_sectors(bio);
5426
5427         chunk_sectors = min(conf->chunk_sectors, conf->prev_chunk_sectors);
5428         return  chunk_sectors >=
5429                 ((sector & (chunk_sectors - 1)) + bio_sectors);
5430 }
5431
5432 /*
5433  *  add bio to the retry LIFO  ( in O(1) ... we are in interrupt )
5434  *  later sampled by raid5d.
5435  */
5436 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
5437 {
5438         unsigned long flags;
5439
5440         spin_lock_irqsave(&conf->device_lock, flags);
5441
5442         bi->bi_next = conf->retry_read_aligned_list;
5443         conf->retry_read_aligned_list = bi;
5444
5445         spin_unlock_irqrestore(&conf->device_lock, flags);
5446         md_wakeup_thread(conf->mddev->thread);
5447 }
5448
5449 static struct bio *remove_bio_from_retry(struct r5conf *conf,
5450                                          unsigned int *offset)
5451 {
5452         struct bio *bi;
5453
5454         bi = conf->retry_read_aligned;
5455         if (bi) {
5456                 *offset = conf->retry_read_offset;
5457                 conf->retry_read_aligned = NULL;
5458                 return bi;
5459         }
5460         bi = conf->retry_read_aligned_list;
5461         if(bi) {
5462                 conf->retry_read_aligned_list = bi->bi_next;
5463                 bi->bi_next = NULL;
5464                 *offset = 0;
5465         }
5466
5467         return bi;
5468 }
5469
5470 /*
5471  *  The "raid5_align_endio" should check if the read succeeded and if it
5472  *  did, call bio_endio on the original bio (having bio_put the new bio
5473  *  first).
5474  *  If the read failed..
5475  */
5476 static void raid5_align_endio(struct bio *bi)
5477 {
5478         struct bio *raid_bi = bi->bi_private;
5479         struct md_rdev *rdev = (void *)raid_bi->bi_next;
5480         struct mddev *mddev = rdev->mddev;
5481         struct r5conf *conf = mddev->private;
5482         blk_status_t error = bi->bi_status;
5483
5484         bio_put(bi);
5485         raid_bi->bi_next = NULL;
5486         rdev_dec_pending(rdev, conf->mddev);
5487
5488         if (!error) {
5489                 bio_endio(raid_bi);
5490                 if (atomic_dec_and_test(&conf->active_aligned_reads))
5491                         wake_up(&conf->wait_for_quiescent);
5492                 return;
5493         }
5494
5495         pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
5496
5497         add_bio_to_retry(raid_bi, conf);
5498 }
5499
5500 static int raid5_read_one_chunk(struct mddev *mddev, struct bio *raid_bio)
5501 {
5502         struct r5conf *conf = mddev->private;
5503         struct bio *align_bio;
5504         struct md_rdev *rdev;
5505         sector_t sector, end_sector, first_bad;
5506         int bad_sectors, dd_idx;
5507         bool did_inc;
5508
5509         if (!in_chunk_boundary(mddev, raid_bio)) {
5510                 pr_debug("%s: non aligned\n", __func__);
5511                 return 0;
5512         }
5513
5514         sector = raid5_compute_sector(conf, raid_bio->bi_iter.bi_sector, 0,
5515                                       &dd_idx, NULL);
5516         end_sector = sector + bio_sectors(raid_bio);
5517
5518         rcu_read_lock();
5519         if (r5c_big_stripe_cached(conf, sector))
5520                 goto out_rcu_unlock;
5521
5522         rdev = rcu_dereference(conf->disks[dd_idx].replacement);
5523         if (!rdev || test_bit(Faulty, &rdev->flags) ||
5524             rdev->recovery_offset < end_sector) {
5525                 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
5526                 if (!rdev)
5527                         goto out_rcu_unlock;
5528                 if (test_bit(Faulty, &rdev->flags) ||
5529                     !(test_bit(In_sync, &rdev->flags) ||
5530                       rdev->recovery_offset >= end_sector))
5531                         goto out_rcu_unlock;
5532         }
5533
5534         atomic_inc(&rdev->nr_pending);
5535         rcu_read_unlock();
5536
5537         if (is_badblock(rdev, sector, bio_sectors(raid_bio), &first_bad,
5538                         &bad_sectors)) {
5539                 rdev_dec_pending(rdev, mddev);
5540                 return 0;
5541         }
5542
5543         md_account_bio(mddev, &raid_bio);
5544         raid_bio->bi_next = (void *)rdev;
5545
5546         align_bio = bio_alloc_clone(rdev->bdev, raid_bio, GFP_NOIO,
5547                                     &mddev->bio_set);
5548         align_bio->bi_end_io = raid5_align_endio;
5549         align_bio->bi_private = raid_bio;
5550         align_bio->bi_iter.bi_sector = sector;
5551
5552         /* No reshape active, so we can trust rdev->data_offset */
5553         align_bio->bi_iter.bi_sector += rdev->data_offset;
5554
5555         did_inc = false;
5556         if (conf->quiesce == 0) {
5557                 atomic_inc(&conf->active_aligned_reads);
5558                 did_inc = true;
5559         }
5560         /* need a memory barrier to detect the race with raid5_quiesce() */
5561         if (!did_inc || smp_load_acquire(&conf->quiesce) != 0) {
5562                 /* quiesce is in progress, so we need to undo io activation and wait
5563                  * for it to finish
5564                  */
5565                 if (did_inc && atomic_dec_and_test(&conf->active_aligned_reads))
5566                         wake_up(&conf->wait_for_quiescent);
5567                 spin_lock_irq(&conf->device_lock);
5568                 wait_event_lock_irq(conf->wait_for_quiescent, conf->quiesce == 0,
5569                                     conf->device_lock);
5570                 atomic_inc(&conf->active_aligned_reads);
5571                 spin_unlock_irq(&conf->device_lock);
5572         }
5573
5574         if (mddev->gendisk)
5575                 trace_block_bio_remap(align_bio, disk_devt(mddev->gendisk),
5576                                       raid_bio->bi_iter.bi_sector);
5577         submit_bio_noacct(align_bio);
5578         return 1;
5579
5580 out_rcu_unlock:
5581         rcu_read_unlock();
5582         return 0;
5583 }
5584
5585 static struct bio *chunk_aligned_read(struct mddev *mddev, struct bio *raid_bio)
5586 {
5587         struct bio *split;
5588         sector_t sector = raid_bio->bi_iter.bi_sector;
5589         unsigned chunk_sects = mddev->chunk_sectors;
5590         unsigned sectors = chunk_sects - (sector & (chunk_sects-1));
5591
5592         if (sectors < bio_sectors(raid_bio)) {
5593                 struct r5conf *conf = mddev->private;
5594                 split = bio_split(raid_bio, sectors, GFP_NOIO, &conf->bio_split);
5595                 bio_chain(split, raid_bio);
5596                 submit_bio_noacct(raid_bio);
5597                 raid_bio = split;
5598         }
5599
5600         if (!raid5_read_one_chunk(mddev, raid_bio))
5601                 return raid_bio;
5602
5603         return NULL;
5604 }
5605
5606 /* __get_priority_stripe - get the next stripe to process
5607  *
5608  * Full stripe writes are allowed to pass preread active stripes up until
5609  * the bypass_threshold is exceeded.  In general the bypass_count
5610  * increments when the handle_list is handled before the hold_list; however, it
5611  * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
5612  * stripe with in flight i/o.  The bypass_count will be reset when the
5613  * head of the hold_list has changed, i.e. the head was promoted to the
5614  * handle_list.
5615  */
5616 static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
5617         __must_hold(&conf->device_lock)
5618 {
5619         struct stripe_head *sh, *tmp;
5620         struct list_head *handle_list = NULL;
5621         struct r5worker_group *wg;
5622         bool second_try = !r5c_is_writeback(conf->log) &&
5623                 !r5l_log_disk_error(conf);
5624         bool try_loprio = test_bit(R5C_LOG_TIGHT, &conf->cache_state) ||
5625                 r5l_log_disk_error(conf);
5626
5627 again:
5628         wg = NULL;
5629         sh = NULL;
5630         if (conf->worker_cnt_per_group == 0) {
5631                 handle_list = try_loprio ? &conf->loprio_list :
5632                                         &conf->handle_list;
5633         } else if (group != ANY_GROUP) {
5634                 handle_list = try_loprio ? &conf->worker_groups[group].loprio_list :
5635                                 &conf->worker_groups[group].handle_list;
5636                 wg = &conf->worker_groups[group];
5637         } else {
5638                 int i;
5639                 for (i = 0; i < conf->group_cnt; i++) {
5640                         handle_list = try_loprio ? &conf->worker_groups[i].loprio_list :
5641                                 &conf->worker_groups[i].handle_list;
5642                         wg = &conf->worker_groups[i];
5643                         if (!list_empty(handle_list))
5644                                 break;
5645                 }
5646         }
5647
5648         pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
5649                   __func__,
5650                   list_empty(handle_list) ? "empty" : "busy",
5651                   list_empty(&conf->hold_list) ? "empty" : "busy",
5652                   atomic_read(&conf->pending_full_writes), conf->bypass_count);
5653
5654         if (!list_empty(handle_list)) {
5655                 sh = list_entry(handle_list->next, typeof(*sh), lru);
5656
5657                 if (list_empty(&conf->hold_list))
5658                         conf->bypass_count = 0;
5659                 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
5660                         if (conf->hold_list.next == conf->last_hold)
5661                                 conf->bypass_count++;
5662                         else {
5663                                 conf->last_hold = conf->hold_list.next;
5664                                 conf->bypass_count -= conf->bypass_threshold;
5665                                 if (conf->bypass_count < 0)
5666                                         conf->bypass_count = 0;
5667                         }
5668                 }
5669         } else if (!list_empty(&conf->hold_list) &&
5670                    ((conf->bypass_threshold &&
5671                      conf->bypass_count > conf->bypass_threshold) ||
5672                     atomic_read(&conf->pending_full_writes) == 0)) {
5673
5674                 list_for_each_entry(tmp, &conf->hold_list,  lru) {
5675                         if (conf->worker_cnt_per_group == 0 ||
5676                             group == ANY_GROUP ||
5677                             !cpu_online(tmp->cpu) ||
5678                             cpu_to_group(tmp->cpu) == group) {
5679                                 sh = tmp;
5680                                 break;
5681                         }
5682                 }
5683
5684                 if (sh) {
5685                         conf->bypass_count -= conf->bypass_threshold;
5686                         if (conf->bypass_count < 0)
5687                                 conf->bypass_count = 0;
5688                 }
5689                 wg = NULL;
5690         }
5691
5692         if (!sh) {
5693                 if (second_try)
5694                         return NULL;
5695                 second_try = true;
5696                 try_loprio = !try_loprio;
5697                 goto again;
5698         }
5699
5700         if (wg) {
5701                 wg->stripes_cnt--;
5702                 sh->group = NULL;
5703         }
5704         list_del_init(&sh->lru);
5705         BUG_ON(atomic_inc_return(&sh->count) != 1);
5706         return sh;
5707 }
5708
5709 struct raid5_plug_cb {
5710         struct blk_plug_cb      cb;
5711         struct list_head        list;
5712         struct list_head        temp_inactive_list[NR_STRIPE_HASH_LOCKS];
5713 };
5714
5715 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
5716 {
5717         struct raid5_plug_cb *cb = container_of(
5718                 blk_cb, struct raid5_plug_cb, cb);
5719         struct stripe_head *sh;
5720         struct mddev *mddev = cb->cb.data;
5721         struct r5conf *conf = mddev->private;
5722         int cnt = 0;
5723         int hash;
5724
5725         if (cb->list.next && !list_empty(&cb->list)) {
5726                 spin_lock_irq(&conf->device_lock);
5727                 while (!list_empty(&cb->list)) {
5728                         sh = list_first_entry(&cb->list, struct stripe_head, lru);
5729                         list_del_init(&sh->lru);
5730                         /*
5731                          * avoid race release_stripe_plug() sees
5732                          * STRIPE_ON_UNPLUG_LIST clear but the stripe
5733                          * is still in our list
5734                          */
5735                         smp_mb__before_atomic();
5736                         clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
5737                         /*
5738                          * STRIPE_ON_RELEASE_LIST could be set here. In that
5739                          * case, the count is always > 1 here
5740                          */
5741                         hash = sh->hash_lock_index;
5742                         __release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
5743                         cnt++;
5744                 }
5745                 spin_unlock_irq(&conf->device_lock);
5746         }
5747         release_inactive_stripe_list(conf, cb->temp_inactive_list,
5748                                      NR_STRIPE_HASH_LOCKS);
5749         if (mddev->queue)
5750                 trace_block_unplug(mddev->queue, cnt, !from_schedule);
5751         kfree(cb);
5752 }
5753
5754 static void release_stripe_plug(struct mddev *mddev,
5755                                 struct stripe_head *sh)
5756 {
5757         struct blk_plug_cb *blk_cb = blk_check_plugged(
5758                 raid5_unplug, mddev,
5759                 sizeof(struct raid5_plug_cb));
5760         struct raid5_plug_cb *cb;
5761
5762         if (!blk_cb) {
5763                 raid5_release_stripe(sh);
5764                 return;
5765         }
5766
5767         cb = container_of(blk_cb, struct raid5_plug_cb, cb);
5768
5769         if (cb->list.next == NULL) {
5770                 int i;
5771                 INIT_LIST_HEAD(&cb->list);
5772                 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5773                         INIT_LIST_HEAD(cb->temp_inactive_list + i);
5774         }
5775
5776         if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
5777                 list_add_tail(&sh->lru, &cb->list);
5778         else
5779                 raid5_release_stripe(sh);
5780 }
5781
5782 static void make_discard_request(struct mddev *mddev, struct bio *bi)
5783 {
5784         struct r5conf *conf = mddev->private;
5785         sector_t logical_sector, last_sector;
5786         struct stripe_head *sh;
5787         int stripe_sectors;
5788
5789         /* We need to handle this when io_uring supports discard/trim */
5790         if (WARN_ON_ONCE(bi->bi_opf & REQ_NOWAIT))
5791                 return;
5792
5793         if (mddev->reshape_position != MaxSector)
5794                 /* Skip discard while reshape is happening */
5795                 return;
5796
5797         logical_sector = bi->bi_iter.bi_sector & ~((sector_t)RAID5_STRIPE_SECTORS(conf)-1);
5798         last_sector = bio_end_sector(bi);
5799
5800         bi->bi_next = NULL;
5801
5802         stripe_sectors = conf->chunk_sectors *
5803                 (conf->raid_disks - conf->max_degraded);
5804         logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
5805                                                stripe_sectors);
5806         sector_div(last_sector, stripe_sectors);
5807
5808         logical_sector *= conf->chunk_sectors;
5809         last_sector *= conf->chunk_sectors;
5810
5811         for (; logical_sector < last_sector;
5812              logical_sector += RAID5_STRIPE_SECTORS(conf)) {
5813                 DEFINE_WAIT(w);
5814                 int d;
5815         again:
5816                 sh = raid5_get_active_stripe(conf, NULL, logical_sector, 0);
5817                 prepare_to_wait(&conf->wait_for_overlap, &w,
5818                                 TASK_UNINTERRUPTIBLE);
5819                 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5820                 if (test_bit(STRIPE_SYNCING, &sh->state)) {
5821                         raid5_release_stripe(sh);
5822                         schedule();
5823                         goto again;
5824                 }
5825                 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5826                 spin_lock_irq(&sh->stripe_lock);
5827                 for (d = 0; d < conf->raid_disks; d++) {
5828                         if (d == sh->pd_idx || d == sh->qd_idx)
5829                                 continue;
5830                         if (sh->dev[d].towrite || sh->dev[d].toread) {
5831                                 set_bit(R5_Overlap, &sh->dev[d].flags);
5832                                 spin_unlock_irq(&sh->stripe_lock);
5833                                 raid5_release_stripe(sh);
5834                                 schedule();
5835                                 goto again;
5836                         }
5837                 }
5838                 set_bit(STRIPE_DISCARD, &sh->state);
5839                 finish_wait(&conf->wait_for_overlap, &w);
5840                 sh->overwrite_disks = 0;
5841                 for (d = 0; d < conf->raid_disks; d++) {
5842                         if (d == sh->pd_idx || d == sh->qd_idx)
5843                                 continue;
5844                         sh->dev[d].towrite = bi;
5845                         set_bit(R5_OVERWRITE, &sh->dev[d].flags);
5846                         bio_inc_remaining(bi);
5847                         md_write_inc(mddev, bi);
5848                         sh->overwrite_disks++;
5849                 }
5850                 spin_unlock_irq(&sh->stripe_lock);
5851                 if (conf->mddev->bitmap) {
5852                         for (d = 0;
5853                              d < conf->raid_disks - conf->max_degraded;
5854                              d++)
5855                                 md_bitmap_startwrite(mddev->bitmap,
5856                                                      sh->sector,
5857                                                      RAID5_STRIPE_SECTORS(conf),
5858                                                      0);
5859                         sh->bm_seq = conf->seq_flush + 1;
5860                         set_bit(STRIPE_BIT_DELAY, &sh->state);
5861                 }
5862
5863                 set_bit(STRIPE_HANDLE, &sh->state);
5864                 clear_bit(STRIPE_DELAYED, &sh->state);
5865                 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5866                         atomic_inc(&conf->preread_active_stripes);
5867                 release_stripe_plug(mddev, sh);
5868         }
5869
5870         bio_endio(bi);
5871 }
5872
5873 static bool ahead_of_reshape(struct mddev *mddev, sector_t sector,
5874                              sector_t reshape_sector)
5875 {
5876         return mddev->reshape_backwards ? sector < reshape_sector :
5877                                           sector >= reshape_sector;
5878 }
5879
5880 static bool range_ahead_of_reshape(struct mddev *mddev, sector_t min,
5881                                    sector_t max, sector_t reshape_sector)
5882 {
5883         return mddev->reshape_backwards ? max < reshape_sector :
5884                                           min >= reshape_sector;
5885 }
5886
5887 static bool stripe_ahead_of_reshape(struct mddev *mddev, struct r5conf *conf,
5888                                     struct stripe_head *sh)
5889 {
5890         sector_t max_sector = 0, min_sector = MaxSector;
5891         bool ret = false;
5892         int dd_idx;
5893
5894         for (dd_idx = 0; dd_idx < sh->disks; dd_idx++) {
5895                 if (dd_idx == sh->pd_idx)
5896                         continue;
5897
5898                 min_sector = min(min_sector, sh->dev[dd_idx].sector);
5899                 max_sector = min(max_sector, sh->dev[dd_idx].sector);
5900         }
5901
5902         spin_lock_irq(&conf->device_lock);
5903
5904         if (!range_ahead_of_reshape(mddev, min_sector, max_sector,
5905                                      conf->reshape_progress))
5906                 /* mismatch, need to try again */
5907                 ret = true;
5908
5909         spin_unlock_irq(&conf->device_lock);
5910
5911         return ret;
5912 }
5913
5914 static int add_all_stripe_bios(struct r5conf *conf,
5915                 struct stripe_request_ctx *ctx, struct stripe_head *sh,
5916                 struct bio *bi, int forwrite, int previous)
5917 {
5918         int dd_idx;
5919         int ret = 1;
5920
5921         spin_lock_irq(&sh->stripe_lock);
5922
5923         for (dd_idx = 0; dd_idx < sh->disks; dd_idx++) {
5924                 struct r5dev *dev = &sh->dev[dd_idx];
5925
5926                 if (dd_idx == sh->pd_idx || dd_idx == sh->qd_idx)
5927                         continue;
5928
5929                 if (dev->sector < ctx->first_sector ||
5930                     dev->sector >= ctx->last_sector)
5931                         continue;
5932
5933                 if (stripe_bio_overlaps(sh, bi, dd_idx, forwrite)) {
5934                         set_bit(R5_Overlap, &dev->flags);
5935                         ret = 0;
5936                         continue;
5937                 }
5938         }
5939
5940         if (!ret)
5941                 goto out;
5942
5943         for (dd_idx = 0; dd_idx < sh->disks; dd_idx++) {
5944                 struct r5dev *dev = &sh->dev[dd_idx];
5945
5946                 if (dd_idx == sh->pd_idx || dd_idx == sh->qd_idx)
5947                         continue;
5948
5949                 if (dev->sector < ctx->first_sector ||
5950                     dev->sector >= ctx->last_sector)
5951                         continue;
5952
5953                 __add_stripe_bio(sh, bi, dd_idx, forwrite, previous);
5954                 clear_bit((dev->sector - ctx->first_sector) >>
5955                           RAID5_STRIPE_SHIFT(conf), ctx->sectors_to_do);
5956         }
5957
5958 out:
5959         spin_unlock_irq(&sh->stripe_lock);
5960         return ret;
5961 }
5962
5963 static enum stripe_result make_stripe_request(struct mddev *mddev,
5964                 struct r5conf *conf, struct stripe_request_ctx *ctx,
5965                 sector_t logical_sector, struct bio *bi)
5966 {
5967         const int rw = bio_data_dir(bi);
5968         enum stripe_result ret;
5969         struct stripe_head *sh;
5970         sector_t new_sector;
5971         int previous = 0, flags = 0;
5972         int seq, dd_idx;
5973
5974         seq = read_seqcount_begin(&conf->gen_lock);
5975
5976         if (unlikely(conf->reshape_progress != MaxSector)) {
5977                 /*
5978                  * Spinlock is needed as reshape_progress may be
5979                  * 64bit on a 32bit platform, and so it might be
5980                  * possible to see a half-updated value
5981                  * Of course reshape_progress could change after
5982                  * the lock is dropped, so once we get a reference
5983                  * to the stripe that we think it is, we will have
5984                  * to check again.
5985                  */
5986                 spin_lock_irq(&conf->device_lock);
5987                 if (ahead_of_reshape(mddev, logical_sector,
5988                                      conf->reshape_progress)) {
5989                         previous = 1;
5990                 } else {
5991                         if (ahead_of_reshape(mddev, logical_sector,
5992                                              conf->reshape_safe)) {
5993                                 spin_unlock_irq(&conf->device_lock);
5994                                 return STRIPE_SCHEDULE_AND_RETRY;
5995                         }
5996                 }
5997                 spin_unlock_irq(&conf->device_lock);
5998         }
5999
6000         new_sector = raid5_compute_sector(conf, logical_sector, previous,
6001                                           &dd_idx, NULL);
6002         pr_debug("raid456: %s, sector %llu logical %llu\n", __func__,
6003                  new_sector, logical_sector);
6004
6005         if (previous)
6006                 flags |= R5_GAS_PREVIOUS;
6007         if (bi->bi_opf & REQ_RAHEAD)
6008                 flags |= R5_GAS_NOBLOCK;
6009         sh = raid5_get_active_stripe(conf, ctx, new_sector, flags);
6010         if (unlikely(!sh)) {
6011                 /* cannot get stripe, just give-up */
6012                 bi->bi_status = BLK_STS_IOERR;
6013                 return STRIPE_FAIL;
6014         }
6015
6016         if (unlikely(previous) &&
6017             stripe_ahead_of_reshape(mddev, conf, sh)) {
6018                 /*
6019                  * Expansion moved on while waiting for a stripe.
6020                  * Expansion could still move past after this
6021                  * test, but as we are holding a reference to
6022                  * 'sh', we know that if that happens,
6023                  *  STRIPE_EXPANDING will get set and the expansion
6024                  * won't proceed until we finish with the stripe.
6025                  */
6026                 ret = STRIPE_SCHEDULE_AND_RETRY;
6027                 goto out_release;
6028         }
6029
6030         if (read_seqcount_retry(&conf->gen_lock, seq)) {
6031                 /* Might have got the wrong stripe_head by accident */
6032                 ret = STRIPE_RETRY;
6033                 goto out_release;
6034         }
6035
6036         if (test_bit(STRIPE_EXPANDING, &sh->state) ||
6037             !add_all_stripe_bios(conf, ctx, sh, bi, rw, previous)) {
6038                 /*
6039                  * Stripe is busy expanding or add failed due to
6040                  * overlap. Flush everything and wait a while.
6041                  */
6042                 md_wakeup_thread(mddev->thread);
6043                 ret = STRIPE_SCHEDULE_AND_RETRY;
6044                 goto out_release;
6045         }
6046
6047         if (stripe_can_batch(sh)) {
6048                 stripe_add_to_batch_list(conf, sh, ctx->batch_last);
6049                 if (ctx->batch_last)
6050                         raid5_release_stripe(ctx->batch_last);
6051                 atomic_inc(&sh->count);
6052                 ctx->batch_last = sh;
6053         }
6054
6055         if (ctx->do_flush) {
6056                 set_bit(STRIPE_R5C_PREFLUSH, &sh->state);
6057                 /* we only need flush for one stripe */
6058                 ctx->do_flush = false;
6059         }
6060
6061         set_bit(STRIPE_HANDLE, &sh->state);
6062         clear_bit(STRIPE_DELAYED, &sh->state);
6063         if ((!sh->batch_head || sh == sh->batch_head) &&
6064             (bi->bi_opf & REQ_SYNC) &&
6065             !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
6066                 atomic_inc(&conf->preread_active_stripes);
6067
6068         release_stripe_plug(mddev, sh);
6069         return STRIPE_SUCCESS;
6070
6071 out_release:
6072         raid5_release_stripe(sh);
6073         return ret;
6074 }
6075
6076 /*
6077  * If the bio covers multiple data disks, find sector within the bio that has
6078  * the lowest chunk offset in the first chunk.
6079  */
6080 static sector_t raid5_bio_lowest_chunk_sector(struct r5conf *conf,
6081                                               struct bio *bi)
6082 {
6083         int sectors_per_chunk = conf->chunk_sectors;
6084         int raid_disks = conf->raid_disks;
6085         int dd_idx;
6086         struct stripe_head sh;
6087         unsigned int chunk_offset;
6088         sector_t r_sector = bi->bi_iter.bi_sector & ~((sector_t)RAID5_STRIPE_SECTORS(conf)-1);
6089         sector_t sector;
6090
6091         /* We pass in fake stripe_head to get back parity disk numbers */
6092         sector = raid5_compute_sector(conf, r_sector, 0, &dd_idx, &sh);
6093         chunk_offset = sector_div(sector, sectors_per_chunk);
6094         if (sectors_per_chunk - chunk_offset >= bio_sectors(bi))
6095                 return r_sector;
6096         /*
6097          * Bio crosses to the next data disk. Check whether it's in the same
6098          * chunk.
6099          */
6100         dd_idx++;
6101         while (dd_idx == sh.pd_idx || dd_idx == sh.qd_idx)
6102                 dd_idx++;
6103         if (dd_idx >= raid_disks)
6104                 return r_sector;
6105         return r_sector + sectors_per_chunk - chunk_offset;
6106 }
6107
6108 static bool raid5_make_request(struct mddev *mddev, struct bio * bi)
6109 {
6110         DEFINE_WAIT_FUNC(wait, woken_wake_function);
6111         struct r5conf *conf = mddev->private;
6112         sector_t logical_sector;
6113         struct stripe_request_ctx ctx = {};
6114         const int rw = bio_data_dir(bi);
6115         enum stripe_result res;
6116         int s, stripe_cnt;
6117
6118         if (unlikely(bi->bi_opf & REQ_PREFLUSH)) {
6119                 int ret = log_handle_flush_request(conf, bi);
6120
6121                 if (ret == 0)
6122                         return true;
6123                 if (ret == -ENODEV) {
6124                         if (md_flush_request(mddev, bi))
6125                                 return true;
6126                 }
6127                 /* ret == -EAGAIN, fallback */
6128                 /*
6129                  * if r5l_handle_flush_request() didn't clear REQ_PREFLUSH,
6130                  * we need to flush journal device
6131                  */
6132                 ctx.do_flush = bi->bi_opf & REQ_PREFLUSH;
6133         }
6134
6135         if (!md_write_start(mddev, bi))
6136                 return false;
6137         /*
6138          * If array is degraded, better not do chunk aligned read because
6139          * later we might have to read it again in order to reconstruct
6140          * data on failed drives.
6141          */
6142         if (rw == READ && mddev->degraded == 0 &&
6143             mddev->reshape_position == MaxSector) {
6144                 bi = chunk_aligned_read(mddev, bi);
6145                 if (!bi)
6146                         return true;
6147         }
6148
6149         if (unlikely(bio_op(bi) == REQ_OP_DISCARD)) {
6150                 make_discard_request(mddev, bi);
6151                 md_write_end(mddev);
6152                 return true;
6153         }
6154
6155         logical_sector = bi->bi_iter.bi_sector & ~((sector_t)RAID5_STRIPE_SECTORS(conf)-1);
6156         ctx.first_sector = logical_sector;
6157         ctx.last_sector = bio_end_sector(bi);
6158         bi->bi_next = NULL;
6159
6160         stripe_cnt = DIV_ROUND_UP_SECTOR_T(ctx.last_sector - logical_sector,
6161                                            RAID5_STRIPE_SECTORS(conf));
6162         bitmap_set(ctx.sectors_to_do, 0, stripe_cnt);
6163
6164         pr_debug("raid456: %s, logical %llu to %llu\n", __func__,
6165                  bi->bi_iter.bi_sector, ctx.last_sector);
6166
6167         /* Bail out if conflicts with reshape and REQ_NOWAIT is set */
6168         if ((bi->bi_opf & REQ_NOWAIT) &&
6169             (conf->reshape_progress != MaxSector) &&
6170             !ahead_of_reshape(mddev, logical_sector, conf->reshape_progress) &&
6171             ahead_of_reshape(mddev, logical_sector, conf->reshape_safe)) {
6172                 bio_wouldblock_error(bi);
6173                 if (rw == WRITE)
6174                         md_write_end(mddev);
6175                 return true;
6176         }
6177         md_account_bio(mddev, &bi);
6178
6179         /*
6180          * Lets start with the stripe with the lowest chunk offset in the first
6181          * chunk. That has the best chances of creating IOs adjacent to
6182          * previous IOs in case of sequential IO and thus creates the most
6183          * sequential IO pattern. We don't bother with the optimization when
6184          * reshaping as the performance benefit is not worth the complexity.
6185          */
6186         if (likely(conf->reshape_progress == MaxSector))
6187                 logical_sector = raid5_bio_lowest_chunk_sector(conf, bi);
6188         s = (logical_sector - ctx.first_sector) >> RAID5_STRIPE_SHIFT(conf);
6189
6190         add_wait_queue(&conf->wait_for_overlap, &wait);
6191         while (1) {
6192                 res = make_stripe_request(mddev, conf, &ctx, logical_sector,
6193                                           bi);
6194                 if (res == STRIPE_FAIL)
6195                         break;
6196
6197                 if (res == STRIPE_RETRY)
6198                         continue;
6199
6200                 if (res == STRIPE_SCHEDULE_AND_RETRY) {
6201                         /*
6202                          * Must release the reference to batch_last before
6203                          * scheduling and waiting for work to be done,
6204                          * otherwise the batch_last stripe head could prevent
6205                          * raid5_activate_delayed() from making progress
6206                          * and thus deadlocking.
6207                          */
6208                         if (ctx.batch_last) {
6209                                 raid5_release_stripe(ctx.batch_last);
6210                                 ctx.batch_last = NULL;
6211                         }
6212
6213                         wait_woken(&wait, TASK_UNINTERRUPTIBLE,
6214                                    MAX_SCHEDULE_TIMEOUT);
6215                         continue;
6216                 }
6217
6218                 s = find_next_bit_wrap(ctx.sectors_to_do, stripe_cnt, s);
6219                 if (s == stripe_cnt)
6220                         break;
6221
6222                 logical_sector = ctx.first_sector +
6223                         (s << RAID5_STRIPE_SHIFT(conf));
6224         }
6225         remove_wait_queue(&conf->wait_for_overlap, &wait);
6226
6227         if (ctx.batch_last)
6228                 raid5_release_stripe(ctx.batch_last);
6229
6230         if (rw == WRITE)
6231                 md_write_end(mddev);
6232         bio_endio(bi);
6233         return true;
6234 }
6235
6236 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
6237
6238 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
6239 {
6240         /* reshaping is quite different to recovery/resync so it is
6241          * handled quite separately ... here.
6242          *
6243          * On each call to sync_request, we gather one chunk worth of
6244          * destination stripes and flag them as expanding.
6245          * Then we find all the source stripes and request reads.
6246          * As the reads complete, handle_stripe will copy the data
6247          * into the destination stripe and release that stripe.
6248          */
6249         struct r5conf *conf = mddev->private;
6250         struct stripe_head *sh;
6251         struct md_rdev *rdev;
6252         sector_t first_sector, last_sector;
6253         int raid_disks = conf->previous_raid_disks;
6254         int data_disks = raid_disks - conf->max_degraded;
6255         int new_data_disks = conf->raid_disks - conf->max_degraded;
6256         int i;
6257         int dd_idx;
6258         sector_t writepos, readpos, safepos;
6259         sector_t stripe_addr;
6260         int reshape_sectors;
6261         struct list_head stripes;
6262         sector_t retn;
6263
6264         if (sector_nr == 0) {
6265                 /* If restarting in the middle, skip the initial sectors */
6266                 if (mddev->reshape_backwards &&
6267                     conf->reshape_progress < raid5_size(mddev, 0, 0)) {
6268                         sector_nr = raid5_size(mddev, 0, 0)
6269                                 - conf->reshape_progress;
6270                 } else if (mddev->reshape_backwards &&
6271                            conf->reshape_progress == MaxSector) {
6272                         /* shouldn't happen, but just in case, finish up.*/
6273                         sector_nr = MaxSector;
6274                 } else if (!mddev->reshape_backwards &&
6275                            conf->reshape_progress > 0)
6276                         sector_nr = conf->reshape_progress;
6277                 sector_div(sector_nr, new_data_disks);
6278                 if (sector_nr) {
6279                         mddev->curr_resync_completed = sector_nr;
6280                         sysfs_notify_dirent_safe(mddev->sysfs_completed);
6281                         *skipped = 1;
6282                         retn = sector_nr;
6283                         goto finish;
6284                 }
6285         }
6286
6287         /* We need to process a full chunk at a time.
6288          * If old and new chunk sizes differ, we need to process the
6289          * largest of these
6290          */
6291
6292         reshape_sectors = max(conf->chunk_sectors, conf->prev_chunk_sectors);
6293
6294         /* We update the metadata at least every 10 seconds, or when
6295          * the data about to be copied would over-write the source of
6296          * the data at the front of the range.  i.e. one new_stripe
6297          * along from reshape_progress new_maps to after where
6298          * reshape_safe old_maps to
6299          */
6300         writepos = conf->reshape_progress;
6301         sector_div(writepos, new_data_disks);
6302         readpos = conf->reshape_progress;
6303         sector_div(readpos, data_disks);
6304         safepos = conf->reshape_safe;
6305         sector_div(safepos, data_disks);
6306         if (mddev->reshape_backwards) {
6307                 BUG_ON(writepos < reshape_sectors);
6308                 writepos -= reshape_sectors;
6309                 readpos += reshape_sectors;
6310                 safepos += reshape_sectors;
6311         } else {
6312                 writepos += reshape_sectors;
6313                 /* readpos and safepos are worst-case calculations.
6314                  * A negative number is overly pessimistic, and causes
6315                  * obvious problems for unsigned storage.  So clip to 0.
6316                  */
6317                 readpos -= min_t(sector_t, reshape_sectors, readpos);
6318                 safepos -= min_t(sector_t, reshape_sectors, safepos);
6319         }
6320
6321         /* Having calculated the 'writepos' possibly use it
6322          * to set 'stripe_addr' which is where we will write to.
6323          */
6324         if (mddev->reshape_backwards) {
6325                 BUG_ON(conf->reshape_progress == 0);
6326                 stripe_addr = writepos;
6327                 BUG_ON((mddev->dev_sectors &
6328                         ~((sector_t)reshape_sectors - 1))
6329                        - reshape_sectors - stripe_addr
6330                        != sector_nr);
6331         } else {
6332                 BUG_ON(writepos != sector_nr + reshape_sectors);
6333                 stripe_addr = sector_nr;
6334         }
6335
6336         /* 'writepos' is the most advanced device address we might write.
6337          * 'readpos' is the least advanced device address we might read.
6338          * 'safepos' is the least address recorded in the metadata as having
6339          *     been reshaped.
6340          * If there is a min_offset_diff, these are adjusted either by
6341          * increasing the safepos/readpos if diff is negative, or
6342          * increasing writepos if diff is positive.
6343          * If 'readpos' is then behind 'writepos', there is no way that we can
6344          * ensure safety in the face of a crash - that must be done by userspace
6345          * making a backup of the data.  So in that case there is no particular
6346          * rush to update metadata.
6347          * Otherwise if 'safepos' is behind 'writepos', then we really need to
6348          * update the metadata to advance 'safepos' to match 'readpos' so that
6349          * we can be safe in the event of a crash.
6350          * So we insist on updating metadata if safepos is behind writepos and
6351          * readpos is beyond writepos.
6352          * In any case, update the metadata every 10 seconds.
6353          * Maybe that number should be configurable, but I'm not sure it is
6354          * worth it.... maybe it could be a multiple of safemode_delay???
6355          */
6356         if (conf->min_offset_diff < 0) {
6357                 safepos += -conf->min_offset_diff;
6358                 readpos += -conf->min_offset_diff;
6359         } else
6360                 writepos += conf->min_offset_diff;
6361
6362         if ((mddev->reshape_backwards
6363              ? (safepos > writepos && readpos < writepos)
6364              : (safepos < writepos && readpos > writepos)) ||
6365             time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
6366                 /* Cannot proceed until we've updated the superblock... */
6367                 wait_event(conf->wait_for_overlap,
6368                            atomic_read(&conf->reshape_stripes)==0
6369                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
6370                 if (atomic_read(&conf->reshape_stripes) != 0)
6371                         return 0;
6372                 mddev->reshape_position = conf->reshape_progress;
6373                 mddev->curr_resync_completed = sector_nr;
6374                 if (!mddev->reshape_backwards)
6375                         /* Can update recovery_offset */
6376                         rdev_for_each(rdev, mddev)
6377                                 if (rdev->raid_disk >= 0 &&
6378                                     !test_bit(Journal, &rdev->flags) &&
6379                                     !test_bit(In_sync, &rdev->flags) &&
6380                                     rdev->recovery_offset < sector_nr)
6381                                         rdev->recovery_offset = sector_nr;
6382
6383                 conf->reshape_checkpoint = jiffies;
6384                 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
6385                 md_wakeup_thread(mddev->thread);
6386                 wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
6387                            test_bit(MD_RECOVERY_INTR, &mddev->recovery));
6388                 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
6389                         return 0;
6390                 spin_lock_irq(&conf->device_lock);
6391                 conf->reshape_safe = mddev->reshape_position;
6392                 spin_unlock_irq(&conf->device_lock);
6393                 wake_up(&conf->wait_for_overlap);
6394                 sysfs_notify_dirent_safe(mddev->sysfs_completed);
6395         }
6396
6397         INIT_LIST_HEAD(&stripes);
6398         for (i = 0; i < reshape_sectors; i += RAID5_STRIPE_SECTORS(conf)) {
6399                 int j;
6400                 int skipped_disk = 0;
6401                 sh = raid5_get_active_stripe(conf, NULL, stripe_addr+i,
6402                                              R5_GAS_NOQUIESCE);
6403                 set_bit(STRIPE_EXPANDING, &sh->state);
6404                 atomic_inc(&conf->reshape_stripes);
6405                 /* If any of this stripe is beyond the end of the old
6406                  * array, then we need to zero those blocks
6407                  */
6408                 for (j=sh->disks; j--;) {
6409                         sector_t s;
6410                         if (j == sh->pd_idx)
6411                                 continue;
6412                         if (conf->level == 6 &&
6413                             j == sh->qd_idx)
6414                                 continue;
6415                         s = raid5_compute_blocknr(sh, j, 0);
6416                         if (s < raid5_size(mddev, 0, 0)) {
6417                                 skipped_disk = 1;
6418                                 continue;
6419                         }
6420                         memset(page_address(sh->dev[j].page), 0, RAID5_STRIPE_SIZE(conf));
6421                         set_bit(R5_Expanded, &sh->dev[j].flags);
6422                         set_bit(R5_UPTODATE, &sh->dev[j].flags);
6423                 }
6424                 if (!skipped_disk) {
6425                         set_bit(STRIPE_EXPAND_READY, &sh->state);
6426                         set_bit(STRIPE_HANDLE, &sh->state);
6427                 }
6428                 list_add(&sh->lru, &stripes);
6429         }
6430         spin_lock_irq(&conf->device_lock);
6431         if (mddev->reshape_backwards)
6432                 conf->reshape_progress -= reshape_sectors * new_data_disks;
6433         else
6434                 conf->reshape_progress += reshape_sectors * new_data_disks;
6435         spin_unlock_irq(&conf->device_lock);
6436         /* Ok, those stripe are ready. We can start scheduling
6437          * reads on the source stripes.
6438          * The source stripes are determined by mapping the first and last
6439          * block on the destination stripes.
6440          */
6441         first_sector =
6442                 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
6443                                      1, &dd_idx, NULL);
6444         last_sector =
6445                 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
6446                                             * new_data_disks - 1),
6447                                      1, &dd_idx, NULL);
6448         if (last_sector >= mddev->dev_sectors)
6449                 last_sector = mddev->dev_sectors - 1;
6450         while (first_sector <= last_sector) {
6451                 sh = raid5_get_active_stripe(conf, NULL, first_sector,
6452                                 R5_GAS_PREVIOUS | R5_GAS_NOQUIESCE);
6453                 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
6454                 set_bit(STRIPE_HANDLE, &sh->state);
6455                 raid5_release_stripe(sh);
6456                 first_sector += RAID5_STRIPE_SECTORS(conf);
6457         }
6458         /* Now that the sources are clearly marked, we can release
6459          * the destination stripes
6460          */
6461         while (!list_empty(&stripes)) {
6462                 sh = list_entry(stripes.next, struct stripe_head, lru);
6463                 list_del_init(&sh->lru);
6464                 raid5_release_stripe(sh);
6465         }
6466         /* If this takes us to the resync_max point where we have to pause,
6467          * then we need to write out the superblock.
6468          */
6469         sector_nr += reshape_sectors;
6470         retn = reshape_sectors;
6471 finish:
6472         if (mddev->curr_resync_completed > mddev->resync_max ||
6473             (sector_nr - mddev->curr_resync_completed) * 2
6474             >= mddev->resync_max - mddev->curr_resync_completed) {
6475                 /* Cannot proceed until we've updated the superblock... */
6476                 wait_event(conf->wait_for_overlap,
6477                            atomic_read(&conf->reshape_stripes) == 0
6478                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
6479                 if (atomic_read(&conf->reshape_stripes) != 0)
6480                         goto ret;
6481                 mddev->reshape_position = conf->reshape_progress;
6482                 mddev->curr_resync_completed = sector_nr;
6483                 if (!mddev->reshape_backwards)
6484                         /* Can update recovery_offset */
6485                         rdev_for_each(rdev, mddev)
6486                                 if (rdev->raid_disk >= 0 &&
6487                                     !test_bit(Journal, &rdev->flags) &&
6488                                     !test_bit(In_sync, &rdev->flags) &&
6489                                     rdev->recovery_offset < sector_nr)
6490                                         rdev->recovery_offset = sector_nr;
6491                 conf->reshape_checkpoint = jiffies;
6492                 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
6493                 md_wakeup_thread(mddev->thread);
6494                 wait_event(mddev->sb_wait,
6495                            !test_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags)
6496                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
6497                 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
6498                         goto ret;
6499                 spin_lock_irq(&conf->device_lock);
6500                 conf->reshape_safe = mddev->reshape_position;
6501                 spin_unlock_irq(&conf->device_lock);
6502                 wake_up(&conf->wait_for_overlap);
6503                 sysfs_notify_dirent_safe(mddev->sysfs_completed);
6504         }
6505 ret:
6506         return retn;
6507 }
6508
6509 static inline sector_t raid5_sync_request(struct mddev *mddev, sector_t sector_nr,
6510                                           int *skipped)
6511 {
6512         struct r5conf *conf = mddev->private;
6513         struct stripe_head *sh;
6514         sector_t max_sector = mddev->dev_sectors;
6515         sector_t sync_blocks;
6516         int still_degraded = 0;
6517         int i;
6518
6519         if (sector_nr >= max_sector) {
6520                 /* just being told to finish up .. nothing much to do */
6521
6522                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
6523                         end_reshape(conf);
6524                         return 0;
6525                 }
6526
6527                 if (mddev->curr_resync < max_sector) /* aborted */
6528                         md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
6529                                            &sync_blocks, 1);
6530                 else /* completed sync */
6531                         conf->fullsync = 0;
6532                 md_bitmap_close_sync(mddev->bitmap);
6533
6534                 return 0;
6535         }
6536
6537         /* Allow raid5_quiesce to complete */
6538         wait_event(conf->wait_for_overlap, conf->quiesce != 2);
6539
6540         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
6541                 return reshape_request(mddev, sector_nr, skipped);
6542
6543         /* No need to check resync_max as we never do more than one
6544          * stripe, and as resync_max will always be on a chunk boundary,
6545          * if the check in md_do_sync didn't fire, there is no chance
6546          * of overstepping resync_max here
6547          */
6548
6549         /* if there is too many failed drives and we are trying
6550          * to resync, then assert that we are finished, because there is
6551          * nothing we can do.
6552          */
6553         if (mddev->degraded >= conf->max_degraded &&
6554             test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
6555                 sector_t rv = mddev->dev_sectors - sector_nr;
6556                 *skipped = 1;
6557                 return rv;
6558         }
6559         if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
6560             !conf->fullsync &&
6561             !md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
6562             sync_blocks >= RAID5_STRIPE_SECTORS(conf)) {
6563                 /* we can skip this block, and probably more */
6564                 do_div(sync_blocks, RAID5_STRIPE_SECTORS(conf));
6565                 *skipped = 1;
6566                 /* keep things rounded to whole stripes */
6567                 return sync_blocks * RAID5_STRIPE_SECTORS(conf);
6568         }
6569
6570         md_bitmap_cond_end_sync(mddev->bitmap, sector_nr, false);
6571
6572         sh = raid5_get_active_stripe(conf, NULL, sector_nr,
6573                                      R5_GAS_NOBLOCK);
6574         if (sh == NULL) {
6575                 sh = raid5_get_active_stripe(conf, NULL, sector_nr, 0);
6576                 /* make sure we don't swamp the stripe cache if someone else
6577                  * is trying to get access
6578                  */
6579                 schedule_timeout_uninterruptible(1);
6580         }
6581         /* Need to check if array will still be degraded after recovery/resync
6582          * Note in case of > 1 drive failures it's possible we're rebuilding
6583          * one drive while leaving another faulty drive in array.
6584          */
6585         rcu_read_lock();
6586         for (i = 0; i < conf->raid_disks; i++) {
6587                 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
6588
6589                 if (rdev == NULL || test_bit(Faulty, &rdev->flags))
6590                         still_degraded = 1;
6591         }
6592         rcu_read_unlock();
6593
6594         md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
6595
6596         set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
6597         set_bit(STRIPE_HANDLE, &sh->state);
6598
6599         raid5_release_stripe(sh);
6600
6601         return RAID5_STRIPE_SECTORS(conf);
6602 }
6603
6604 static int  retry_aligned_read(struct r5conf *conf, struct bio *raid_bio,
6605                                unsigned int offset)
6606 {
6607         /* We may not be able to submit a whole bio at once as there
6608          * may not be enough stripe_heads available.
6609          * We cannot pre-allocate enough stripe_heads as we may need
6610          * more than exist in the cache (if we allow ever large chunks).
6611          * So we do one stripe head at a time and record in
6612          * ->bi_hw_segments how many have been done.
6613          *
6614          * We *know* that this entire raid_bio is in one chunk, so
6615          * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
6616          */
6617         struct stripe_head *sh;
6618         int dd_idx;
6619         sector_t sector, logical_sector, last_sector;
6620         int scnt = 0;
6621         int handled = 0;
6622
6623         logical_sector = raid_bio->bi_iter.bi_sector &
6624                 ~((sector_t)RAID5_STRIPE_SECTORS(conf)-1);
6625         sector = raid5_compute_sector(conf, logical_sector,
6626                                       0, &dd_idx, NULL);
6627         last_sector = bio_end_sector(raid_bio);
6628
6629         for (; logical_sector < last_sector;
6630              logical_sector += RAID5_STRIPE_SECTORS(conf),
6631                      sector += RAID5_STRIPE_SECTORS(conf),
6632                      scnt++) {
6633
6634                 if (scnt < offset)
6635                         /* already done this stripe */
6636                         continue;
6637
6638                 sh = raid5_get_active_stripe(conf, NULL, sector,
6639                                 R5_GAS_NOBLOCK | R5_GAS_NOQUIESCE);
6640                 if (!sh) {
6641                         /* failed to get a stripe - must wait */
6642                         conf->retry_read_aligned = raid_bio;
6643                         conf->retry_read_offset = scnt;
6644                         return handled;
6645                 }
6646
6647                 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0, 0)) {
6648                         raid5_release_stripe(sh);
6649                         conf->retry_read_aligned = raid_bio;
6650                         conf->retry_read_offset = scnt;
6651                         return handled;
6652                 }
6653
6654                 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
6655                 handle_stripe(sh);
6656                 raid5_release_stripe(sh);
6657                 handled++;
6658         }
6659
6660         bio_endio(raid_bio);
6661
6662         if (atomic_dec_and_test(&conf->active_aligned_reads))
6663                 wake_up(&conf->wait_for_quiescent);
6664         return handled;
6665 }
6666
6667 static int handle_active_stripes(struct r5conf *conf, int group,
6668                                  struct r5worker *worker,
6669                                  struct list_head *temp_inactive_list)
6670                 __must_hold(&conf->device_lock)
6671 {
6672         struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
6673         int i, batch_size = 0, hash;
6674         bool release_inactive = false;
6675
6676         while (batch_size < MAX_STRIPE_BATCH &&
6677                         (sh = __get_priority_stripe(conf, group)) != NULL)
6678                 batch[batch_size++] = sh;
6679
6680         if (batch_size == 0) {
6681                 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6682                         if (!list_empty(temp_inactive_list + i))
6683                                 break;
6684                 if (i == NR_STRIPE_HASH_LOCKS) {
6685                         spin_unlock_irq(&conf->device_lock);
6686                         log_flush_stripe_to_raid(conf);
6687                         spin_lock_irq(&conf->device_lock);
6688                         return batch_size;
6689                 }
6690                 release_inactive = true;
6691         }
6692         spin_unlock_irq(&conf->device_lock);
6693
6694         release_inactive_stripe_list(conf, temp_inactive_list,
6695                                      NR_STRIPE_HASH_LOCKS);
6696
6697         r5l_flush_stripe_to_raid(conf->log);
6698         if (release_inactive) {
6699                 spin_lock_irq(&conf->device_lock);
6700                 return 0;
6701         }
6702
6703         for (i = 0; i < batch_size; i++)
6704                 handle_stripe(batch[i]);
6705         log_write_stripe_run(conf);
6706
6707         cond_resched();
6708
6709         spin_lock_irq(&conf->device_lock);
6710         for (i = 0; i < batch_size; i++) {
6711                 hash = batch[i]->hash_lock_index;
6712                 __release_stripe(conf, batch[i], &temp_inactive_list[hash]);
6713         }
6714         return batch_size;
6715 }
6716
6717 static void raid5_do_work(struct work_struct *work)
6718 {
6719         struct r5worker *worker = container_of(work, struct r5worker, work);
6720         struct r5worker_group *group = worker->group;
6721         struct r5conf *conf = group->conf;
6722         struct mddev *mddev = conf->mddev;
6723         int group_id = group - conf->worker_groups;
6724         int handled;
6725         struct blk_plug plug;
6726
6727         pr_debug("+++ raid5worker active\n");
6728
6729         blk_start_plug(&plug);
6730         handled = 0;
6731         spin_lock_irq(&conf->device_lock);
6732         while (1) {
6733                 int batch_size, released;
6734
6735                 released = release_stripe_list(conf, worker->temp_inactive_list);
6736
6737                 batch_size = handle_active_stripes(conf, group_id, worker,
6738                                                    worker->temp_inactive_list);
6739                 worker->working = false;
6740                 if (!batch_size && !released)
6741                         break;
6742                 handled += batch_size;
6743                 wait_event_lock_irq(mddev->sb_wait,
6744                         !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags),
6745                         conf->device_lock);
6746         }
6747         pr_debug("%d stripes handled\n", handled);
6748
6749         spin_unlock_irq(&conf->device_lock);
6750
6751         flush_deferred_bios(conf);
6752
6753         r5l_flush_stripe_to_raid(conf->log);
6754
6755         async_tx_issue_pending_all();
6756         blk_finish_plug(&plug);
6757
6758         pr_debug("--- raid5worker inactive\n");
6759 }
6760
6761 /*
6762  * This is our raid5 kernel thread.
6763  *
6764  * We scan the hash table for stripes which can be handled now.
6765  * During the scan, completed stripes are saved for us by the interrupt
6766  * handler, so that they will not have to wait for our next wakeup.
6767  */
6768 static void raid5d(struct md_thread *thread)
6769 {
6770         struct mddev *mddev = thread->mddev;
6771         struct r5conf *conf = mddev->private;
6772         int handled;
6773         struct blk_plug plug;
6774
6775         pr_debug("+++ raid5d active\n");
6776
6777         md_check_recovery(mddev);
6778
6779         blk_start_plug(&plug);
6780         handled = 0;
6781         spin_lock_irq(&conf->device_lock);
6782         while (1) {
6783                 struct bio *bio;
6784                 int batch_size, released;
6785                 unsigned int offset;
6786
6787                 released = release_stripe_list(conf, conf->temp_inactive_list);
6788                 if (released)
6789                         clear_bit(R5_DID_ALLOC, &conf->cache_state);
6790
6791                 if (
6792                     !list_empty(&conf->bitmap_list)) {
6793                         /* Now is a good time to flush some bitmap updates */
6794                         conf->seq_flush++;
6795                         spin_unlock_irq(&conf->device_lock);
6796                         md_bitmap_unplug(mddev->bitmap);
6797                         spin_lock_irq(&conf->device_lock);
6798                         conf->seq_write = conf->seq_flush;
6799                         activate_bit_delay(conf, conf->temp_inactive_list);
6800                 }
6801                 raid5_activate_delayed(conf);
6802
6803                 while ((bio = remove_bio_from_retry(conf, &offset))) {
6804                         int ok;
6805                         spin_unlock_irq(&conf->device_lock);
6806                         ok = retry_aligned_read(conf, bio, offset);
6807                         spin_lock_irq(&conf->device_lock);
6808                         if (!ok)
6809                                 break;
6810                         handled++;
6811                 }
6812
6813                 batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
6814                                                    conf->temp_inactive_list);
6815                 if (!batch_size && !released)
6816                         break;
6817                 handled += batch_size;
6818
6819                 if (mddev->sb_flags & ~(1 << MD_SB_CHANGE_PENDING)) {
6820                         spin_unlock_irq(&conf->device_lock);
6821                         md_check_recovery(mddev);
6822                         spin_lock_irq(&conf->device_lock);
6823
6824                         /*
6825                          * Waiting on MD_SB_CHANGE_PENDING below may deadlock
6826                          * seeing md_check_recovery() is needed to clear
6827                          * the flag when using mdmon.
6828                          */
6829                         continue;
6830                 }
6831
6832                 wait_event_lock_irq(mddev->sb_wait,
6833                         !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags),
6834                         conf->device_lock);
6835         }
6836         pr_debug("%d stripes handled\n", handled);
6837
6838         spin_unlock_irq(&conf->device_lock);
6839         if (test_and_clear_bit(R5_ALLOC_MORE, &conf->cache_state) &&
6840             mutex_trylock(&conf->cache_size_mutex)) {
6841                 grow_one_stripe(conf, __GFP_NOWARN);
6842                 /* Set flag even if allocation failed.  This helps
6843                  * slow down allocation requests when mem is short
6844                  */
6845                 set_bit(R5_DID_ALLOC, &conf->cache_state);
6846                 mutex_unlock(&conf->cache_size_mutex);
6847         }
6848
6849         flush_deferred_bios(conf);
6850
6851         r5l_flush_stripe_to_raid(conf->log);
6852
6853         async_tx_issue_pending_all();
6854         blk_finish_plug(&plug);
6855
6856         pr_debug("--- raid5d inactive\n");
6857 }
6858
6859 static ssize_t
6860 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
6861 {
6862         struct r5conf *conf;
6863         int ret = 0;
6864         spin_lock(&mddev->lock);
6865         conf = mddev->private;
6866         if (conf)
6867                 ret = sprintf(page, "%d\n", conf->min_nr_stripes);
6868         spin_unlock(&mddev->lock);
6869         return ret;
6870 }
6871
6872 int
6873 raid5_set_cache_size(struct mddev *mddev, int size)
6874 {
6875         int result = 0;
6876         struct r5conf *conf = mddev->private;
6877
6878         if (size <= 16 || size > 32768)
6879                 return -EINVAL;
6880
6881         conf->min_nr_stripes = size;
6882         mutex_lock(&conf->cache_size_mutex);
6883         while (size < conf->max_nr_stripes &&
6884                drop_one_stripe(conf))
6885                 ;
6886         mutex_unlock(&conf->cache_size_mutex);
6887
6888         md_allow_write(mddev);
6889
6890         mutex_lock(&conf->cache_size_mutex);
6891         while (size > conf->max_nr_stripes)
6892                 if (!grow_one_stripe(conf, GFP_KERNEL)) {
6893                         conf->min_nr_stripes = conf->max_nr_stripes;
6894                         result = -ENOMEM;
6895                         break;
6896                 }
6897         mutex_unlock(&conf->cache_size_mutex);
6898
6899         return result;
6900 }
6901 EXPORT_SYMBOL(raid5_set_cache_size);
6902
6903 static ssize_t
6904 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
6905 {
6906         struct r5conf *conf;
6907         unsigned long new;
6908         int err;
6909
6910         if (len >= PAGE_SIZE)
6911                 return -EINVAL;
6912         if (kstrtoul(page, 10, &new))
6913                 return -EINVAL;
6914         err = mddev_lock(mddev);
6915         if (err)
6916                 return err;
6917         conf = mddev->private;
6918         if (!conf)
6919                 err = -ENODEV;
6920         else
6921                 err = raid5_set_cache_size(mddev, new);
6922         mddev_unlock(mddev);
6923
6924         return err ?: len;
6925 }
6926
6927 static struct md_sysfs_entry
6928 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
6929                                 raid5_show_stripe_cache_size,
6930                                 raid5_store_stripe_cache_size);
6931
6932 static ssize_t
6933 raid5_show_rmw_level(struct mddev  *mddev, char *page)
6934 {
6935         struct r5conf *conf = mddev->private;
6936         if (conf)
6937                 return sprintf(page, "%d\n", conf->rmw_level);
6938         else
6939                 return 0;
6940 }
6941
6942 static ssize_t
6943 raid5_store_rmw_level(struct mddev  *mddev, const char *page, size_t len)
6944 {
6945         struct r5conf *conf = mddev->private;
6946         unsigned long new;
6947
6948         if (!conf)
6949                 return -ENODEV;
6950
6951         if (len >= PAGE_SIZE)
6952                 return -EINVAL;
6953
6954         if (kstrtoul(page, 10, &new))
6955                 return -EINVAL;
6956
6957         if (new != PARITY_DISABLE_RMW && !raid6_call.xor_syndrome)
6958                 return -EINVAL;
6959
6960         if (new != PARITY_DISABLE_RMW &&
6961             new != PARITY_ENABLE_RMW &&
6962             new != PARITY_PREFER_RMW)
6963                 return -EINVAL;
6964
6965         conf->rmw_level = new;
6966         return len;
6967 }
6968
6969 static struct md_sysfs_entry
6970 raid5_rmw_level = __ATTR(rmw_level, S_IRUGO | S_IWUSR,
6971                          raid5_show_rmw_level,
6972                          raid5_store_rmw_level);
6973
6974 static ssize_t
6975 raid5_show_stripe_size(struct mddev  *mddev, char *page)
6976 {
6977         struct r5conf *conf;
6978         int ret = 0;
6979
6980         spin_lock(&mddev->lock);
6981         conf = mddev->private;
6982         if (conf)
6983                 ret = sprintf(page, "%lu\n", RAID5_STRIPE_SIZE(conf));
6984         spin_unlock(&mddev->lock);
6985         return ret;
6986 }
6987
6988 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
6989 static ssize_t
6990 raid5_store_stripe_size(struct mddev  *mddev, const char *page, size_t len)
6991 {
6992         struct r5conf *conf;
6993         unsigned long new;
6994         int err;
6995         int size;
6996
6997         if (len >= PAGE_SIZE)
6998                 return -EINVAL;
6999         if (kstrtoul(page, 10, &new))
7000                 return -EINVAL;
7001
7002         /*
7003          * The value should not be bigger than PAGE_SIZE. It requires to
7004          * be multiple of DEFAULT_STRIPE_SIZE and the value should be power
7005          * of two.
7006          */
7007         if (new % DEFAULT_STRIPE_SIZE != 0 ||
7008                         new > PAGE_SIZE || new == 0 ||
7009                         new != roundup_pow_of_two(new))
7010                 return -EINVAL;
7011
7012         err = mddev_suspend_and_lock(mddev);
7013         if (err)
7014                 return err;
7015
7016         conf = mddev->private;
7017         if (!conf) {
7018                 err = -ENODEV;
7019                 goto out_unlock;
7020         }
7021
7022         if (new == conf->stripe_size)
7023                 goto out_unlock;
7024
7025         pr_debug("md/raid: change stripe_size from %lu to %lu\n",
7026                         conf->stripe_size, new);
7027
7028         if (mddev->sync_thread ||
7029                 test_bit(MD_RECOVERY_RUNNING, &mddev->recovery) ||
7030                 mddev->reshape_position != MaxSector ||
7031                 mddev->sysfs_active) {
7032                 err = -EBUSY;
7033                 goto out_unlock;
7034         }
7035
7036         mutex_lock(&conf->cache_size_mutex);
7037         size = conf->max_nr_stripes;
7038
7039         shrink_stripes(conf);
7040
7041         conf->stripe_size = new;
7042         conf->stripe_shift = ilog2(new) - 9;
7043         conf->stripe_sectors = new >> 9;
7044         if (grow_stripes(conf, size)) {
7045                 pr_warn("md/raid:%s: couldn't allocate buffers\n",
7046                                 mdname(mddev));
7047                 err = -ENOMEM;
7048         }
7049         mutex_unlock(&conf->cache_size_mutex);
7050
7051 out_unlock:
7052         mddev_unlock_and_resume(mddev);
7053         return err ?: len;
7054 }
7055
7056 static struct md_sysfs_entry
7057 raid5_stripe_size = __ATTR(stripe_size, 0644,
7058                          raid5_show_stripe_size,
7059                          raid5_store_stripe_size);
7060 #else
7061 static struct md_sysfs_entry
7062 raid5_stripe_size = __ATTR(stripe_size, 0444,
7063                          raid5_show_stripe_size,
7064                          NULL);
7065 #endif
7066
7067 static ssize_t
7068 raid5_show_preread_threshold(struct mddev *mddev, char *page)
7069 {
7070         struct r5conf *conf;
7071         int ret = 0;
7072         spin_lock(&mddev->lock);
7073         conf = mddev->private;
7074         if (conf)
7075                 ret = sprintf(page, "%d\n", conf->bypass_threshold);
7076         spin_unlock(&mddev->lock);
7077         return ret;
7078 }
7079
7080 static ssize_t
7081 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
7082 {
7083         struct r5conf *conf;
7084         unsigned long new;
7085         int err;
7086
7087         if (len >= PAGE_SIZE)
7088                 return -EINVAL;
7089         if (kstrtoul(page, 10, &new))
7090                 return -EINVAL;
7091
7092         err = mddev_lock(mddev);
7093         if (err)
7094                 return err;
7095         conf = mddev->private;
7096         if (!conf)
7097                 err = -ENODEV;
7098         else if (new > conf->min_nr_stripes)
7099                 err = -EINVAL;
7100         else
7101                 conf->bypass_threshold = new;
7102         mddev_unlock(mddev);
7103         return err ?: len;
7104 }
7105
7106 static struct md_sysfs_entry
7107 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
7108                                         S_IRUGO | S_IWUSR,
7109                                         raid5_show_preread_threshold,
7110                                         raid5_store_preread_threshold);
7111
7112 static ssize_t
7113 raid5_show_skip_copy(struct mddev *mddev, char *page)
7114 {
7115         struct r5conf *conf;
7116         int ret = 0;
7117         spin_lock(&mddev->lock);
7118         conf = mddev->private;
7119         if (conf)
7120                 ret = sprintf(page, "%d\n", conf->skip_copy);
7121         spin_unlock(&mddev->lock);
7122         return ret;
7123 }
7124
7125 static ssize_t
7126 raid5_store_skip_copy(struct mddev *mddev, const char *page, size_t len)
7127 {
7128         struct r5conf *conf;
7129         unsigned long new;
7130         int err;
7131
7132         if (len >= PAGE_SIZE)
7133                 return -EINVAL;
7134         if (kstrtoul(page, 10, &new))
7135                 return -EINVAL;
7136         new = !!new;
7137
7138         err = mddev_suspend_and_lock(mddev);
7139         if (err)
7140                 return err;
7141         conf = mddev->private;
7142         if (!conf)
7143                 err = -ENODEV;
7144         else if (new != conf->skip_copy) {
7145                 struct request_queue *q = mddev->queue;
7146
7147                 conf->skip_copy = new;
7148                 if (new)
7149                         blk_queue_flag_set(QUEUE_FLAG_STABLE_WRITES, q);
7150                 else
7151                         blk_queue_flag_clear(QUEUE_FLAG_STABLE_WRITES, q);
7152         }
7153         mddev_unlock_and_resume(mddev);
7154         return err ?: len;
7155 }
7156
7157 static struct md_sysfs_entry
7158 raid5_skip_copy = __ATTR(skip_copy, S_IRUGO | S_IWUSR,
7159                                         raid5_show_skip_copy,
7160                                         raid5_store_skip_copy);
7161
7162 static ssize_t
7163 stripe_cache_active_show(struct mddev *mddev, char *page)
7164 {
7165         struct r5conf *conf = mddev->private;
7166         if (conf)
7167                 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
7168         else
7169                 return 0;
7170 }
7171
7172 static struct md_sysfs_entry
7173 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
7174
7175 static ssize_t
7176 raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
7177 {
7178         struct r5conf *conf;
7179         int ret = 0;
7180         spin_lock(&mddev->lock);
7181         conf = mddev->private;
7182         if (conf)
7183                 ret = sprintf(page, "%d\n", conf->worker_cnt_per_group);
7184         spin_unlock(&mddev->lock);
7185         return ret;
7186 }
7187
7188 static int alloc_thread_groups(struct r5conf *conf, int cnt,
7189                                int *group_cnt,
7190                                struct r5worker_group **worker_groups);
7191 static ssize_t
7192 raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
7193 {
7194         struct r5conf *conf;
7195         unsigned int new;
7196         int err;
7197         struct r5worker_group *new_groups, *old_groups;
7198         int group_cnt;
7199
7200         if (len >= PAGE_SIZE)
7201                 return -EINVAL;
7202         if (kstrtouint(page, 10, &new))
7203                 return -EINVAL;
7204         /* 8192 should be big enough */
7205         if (new > 8192)
7206                 return -EINVAL;
7207
7208         err = mddev_suspend_and_lock(mddev);
7209         if (err)
7210                 return err;
7211         conf = mddev->private;
7212         if (!conf)
7213                 err = -ENODEV;
7214         else if (new != conf->worker_cnt_per_group) {
7215                 old_groups = conf->worker_groups;
7216                 if (old_groups)
7217                         flush_workqueue(raid5_wq);
7218
7219                 err = alloc_thread_groups(conf, new, &group_cnt, &new_groups);
7220                 if (!err) {
7221                         spin_lock_irq(&conf->device_lock);
7222                         conf->group_cnt = group_cnt;
7223                         conf->worker_cnt_per_group = new;
7224                         conf->worker_groups = new_groups;
7225                         spin_unlock_irq(&conf->device_lock);
7226
7227                         if (old_groups)
7228                                 kfree(old_groups[0].workers);
7229                         kfree(old_groups);
7230                 }
7231         }
7232         mddev_unlock_and_resume(mddev);
7233
7234         return err ?: len;
7235 }
7236
7237 static struct md_sysfs_entry
7238 raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
7239                                 raid5_show_group_thread_cnt,
7240                                 raid5_store_group_thread_cnt);
7241
7242 static struct attribute *raid5_attrs[] =  {
7243         &raid5_stripecache_size.attr,
7244         &raid5_stripecache_active.attr,
7245         &raid5_preread_bypass_threshold.attr,
7246         &raid5_group_thread_cnt.attr,
7247         &raid5_skip_copy.attr,
7248         &raid5_rmw_level.attr,
7249         &raid5_stripe_size.attr,
7250         &r5c_journal_mode.attr,
7251         &ppl_write_hint.attr,
7252         NULL,
7253 };
7254 static const struct attribute_group raid5_attrs_group = {
7255         .name = NULL,
7256         .attrs = raid5_attrs,
7257 };
7258
7259 static int alloc_thread_groups(struct r5conf *conf, int cnt, int *group_cnt,
7260                                struct r5worker_group **worker_groups)
7261 {
7262         int i, j, k;
7263         ssize_t size;
7264         struct r5worker *workers;
7265
7266         if (cnt == 0) {
7267                 *group_cnt = 0;
7268                 *worker_groups = NULL;
7269                 return 0;
7270         }
7271         *group_cnt = num_possible_nodes();
7272         size = sizeof(struct r5worker) * cnt;
7273         workers = kcalloc(size, *group_cnt, GFP_NOIO);
7274         *worker_groups = kcalloc(*group_cnt, sizeof(struct r5worker_group),
7275                                  GFP_NOIO);
7276         if (!*worker_groups || !workers) {
7277                 kfree(workers);
7278                 kfree(*worker_groups);
7279                 return -ENOMEM;
7280         }
7281
7282         for (i = 0; i < *group_cnt; i++) {
7283                 struct r5worker_group *group;
7284
7285                 group = &(*worker_groups)[i];
7286                 INIT_LIST_HEAD(&group->handle_list);
7287                 INIT_LIST_HEAD(&group->loprio_list);
7288                 group->conf = conf;
7289                 group->workers = workers + i * cnt;
7290
7291                 for (j = 0; j < cnt; j++) {
7292                         struct r5worker *worker = group->workers + j;
7293                         worker->group = group;
7294                         INIT_WORK(&worker->work, raid5_do_work);
7295
7296                         for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
7297                                 INIT_LIST_HEAD(worker->temp_inactive_list + k);
7298                 }
7299         }
7300
7301         return 0;
7302 }
7303
7304 static void free_thread_groups(struct r5conf *conf)
7305 {
7306         if (conf->worker_groups)
7307                 kfree(conf->worker_groups[0].workers);
7308         kfree(conf->worker_groups);
7309         conf->worker_groups = NULL;
7310 }
7311
7312 static sector_t
7313 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
7314 {
7315         struct r5conf *conf = mddev->private;
7316
7317         if (!sectors)
7318                 sectors = mddev->dev_sectors;
7319         if (!raid_disks)
7320                 /* size is defined by the smallest of previous and new size */
7321                 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
7322
7323         sectors &= ~((sector_t)conf->chunk_sectors - 1);
7324         sectors &= ~((sector_t)conf->prev_chunk_sectors - 1);
7325         return sectors * (raid_disks - conf->max_degraded);
7326 }
7327
7328 static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
7329 {
7330         safe_put_page(percpu->spare_page);
7331         percpu->spare_page = NULL;
7332         kvfree(percpu->scribble);
7333         percpu->scribble = NULL;
7334 }
7335
7336 static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
7337 {
7338         if (conf->level == 6 && !percpu->spare_page) {
7339                 percpu->spare_page = alloc_page(GFP_KERNEL);
7340                 if (!percpu->spare_page)
7341                         return -ENOMEM;
7342         }
7343
7344         if (scribble_alloc(percpu,
7345                            max(conf->raid_disks,
7346                                conf->previous_raid_disks),
7347                            max(conf->chunk_sectors,
7348                                conf->prev_chunk_sectors)
7349                            / RAID5_STRIPE_SECTORS(conf))) {
7350                 free_scratch_buffer(conf, percpu);
7351                 return -ENOMEM;
7352         }
7353
7354         local_lock_init(&percpu->lock);
7355         return 0;
7356 }
7357
7358 static int raid456_cpu_dead(unsigned int cpu, struct hlist_node *node)
7359 {
7360         struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
7361
7362         free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
7363         return 0;
7364 }
7365
7366 static void raid5_free_percpu(struct r5conf *conf)
7367 {
7368         if (!conf->percpu)
7369                 return;
7370
7371         cpuhp_state_remove_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
7372         free_percpu(conf->percpu);
7373 }
7374
7375 static void free_conf(struct r5conf *conf)
7376 {
7377         int i;
7378
7379         log_exit(conf);
7380
7381         shrinker_free(conf->shrinker);
7382         free_thread_groups(conf);
7383         shrink_stripes(conf);
7384         raid5_free_percpu(conf);
7385         for (i = 0; i < conf->pool_size; i++)
7386                 if (conf->disks[i].extra_page)
7387                         put_page(conf->disks[i].extra_page);
7388         kfree(conf->disks);
7389         bioset_exit(&conf->bio_split);
7390         kfree(conf->stripe_hashtbl);
7391         kfree(conf->pending_data);
7392         kfree(conf);
7393 }
7394
7395 static int raid456_cpu_up_prepare(unsigned int cpu, struct hlist_node *node)
7396 {
7397         struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
7398         struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
7399
7400         if (alloc_scratch_buffer(conf, percpu)) {
7401                 pr_warn("%s: failed memory allocation for cpu%u\n",
7402                         __func__, cpu);
7403                 return -ENOMEM;
7404         }
7405         return 0;
7406 }
7407
7408 static int raid5_alloc_percpu(struct r5conf *conf)
7409 {
7410         int err = 0;
7411
7412         conf->percpu = alloc_percpu(struct raid5_percpu);
7413         if (!conf->percpu)
7414                 return -ENOMEM;
7415
7416         err = cpuhp_state_add_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
7417         if (!err) {
7418                 conf->scribble_disks = max(conf->raid_disks,
7419                         conf->previous_raid_disks);
7420                 conf->scribble_sectors = max(conf->chunk_sectors,
7421                         conf->prev_chunk_sectors);
7422         }
7423         return err;
7424 }
7425
7426 static unsigned long raid5_cache_scan(struct shrinker *shrink,
7427                                       struct shrink_control *sc)
7428 {
7429         struct r5conf *conf = shrink->private_data;
7430         unsigned long ret = SHRINK_STOP;
7431
7432         if (mutex_trylock(&conf->cache_size_mutex)) {
7433                 ret= 0;
7434                 while (ret < sc->nr_to_scan &&
7435                        conf->max_nr_stripes > conf->min_nr_stripes) {
7436                         if (drop_one_stripe(conf) == 0) {
7437                                 ret = SHRINK_STOP;
7438                                 break;
7439                         }
7440                         ret++;
7441                 }
7442                 mutex_unlock(&conf->cache_size_mutex);
7443         }
7444         return ret;
7445 }
7446
7447 static unsigned long raid5_cache_count(struct shrinker *shrink,
7448                                        struct shrink_control *sc)
7449 {
7450         struct r5conf *conf = shrink->private_data;
7451
7452         if (conf->max_nr_stripes < conf->min_nr_stripes)
7453                 /* unlikely, but not impossible */
7454                 return 0;
7455         return conf->max_nr_stripes - conf->min_nr_stripes;
7456 }
7457
7458 static struct r5conf *setup_conf(struct mddev *mddev)
7459 {
7460         struct r5conf *conf;
7461         int raid_disk, memory, max_disks;
7462         struct md_rdev *rdev;
7463         struct disk_info *disk;
7464         char pers_name[6];
7465         int i;
7466         int group_cnt;
7467         struct r5worker_group *new_group;
7468         int ret = -ENOMEM;
7469
7470         if (mddev->new_level != 5
7471             && mddev->new_level != 4
7472             && mddev->new_level != 6) {
7473                 pr_warn("md/raid:%s: raid level not set to 4/5/6 (%d)\n",
7474                         mdname(mddev), mddev->new_level);
7475                 return ERR_PTR(-EIO);
7476         }
7477         if ((mddev->new_level == 5
7478              && !algorithm_valid_raid5(mddev->new_layout)) ||
7479             (mddev->new_level == 6
7480              && !algorithm_valid_raid6(mddev->new_layout))) {
7481                 pr_warn("md/raid:%s: layout %d not supported\n",
7482                         mdname(mddev), mddev->new_layout);
7483                 return ERR_PTR(-EIO);
7484         }
7485         if (mddev->new_level == 6 && mddev->raid_disks < 4) {
7486                 pr_warn("md/raid:%s: not enough configured devices (%d, minimum 4)\n",
7487                         mdname(mddev), mddev->raid_disks);
7488                 return ERR_PTR(-EINVAL);
7489         }
7490
7491         if (!mddev->new_chunk_sectors ||
7492             (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
7493             !is_power_of_2(mddev->new_chunk_sectors)) {
7494                 pr_warn("md/raid:%s: invalid chunk size %d\n",
7495                         mdname(mddev), mddev->new_chunk_sectors << 9);
7496                 return ERR_PTR(-EINVAL);
7497         }
7498
7499         conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
7500         if (conf == NULL)
7501                 goto abort;
7502
7503 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
7504         conf->stripe_size = DEFAULT_STRIPE_SIZE;
7505         conf->stripe_shift = ilog2(DEFAULT_STRIPE_SIZE) - 9;
7506         conf->stripe_sectors = DEFAULT_STRIPE_SIZE >> 9;
7507 #endif
7508         INIT_LIST_HEAD(&conf->free_list);
7509         INIT_LIST_HEAD(&conf->pending_list);
7510         conf->pending_data = kcalloc(PENDING_IO_MAX,
7511                                      sizeof(struct r5pending_data),
7512                                      GFP_KERNEL);
7513         if (!conf->pending_data)
7514                 goto abort;
7515         for (i = 0; i < PENDING_IO_MAX; i++)
7516                 list_add(&conf->pending_data[i].sibling, &conf->free_list);
7517         /* Don't enable multi-threading by default*/
7518         if (!alloc_thread_groups(conf, 0, &group_cnt, &new_group)) {
7519                 conf->group_cnt = group_cnt;
7520                 conf->worker_cnt_per_group = 0;
7521                 conf->worker_groups = new_group;
7522         } else
7523                 goto abort;
7524         spin_lock_init(&conf->device_lock);
7525         seqcount_spinlock_init(&conf->gen_lock, &conf->device_lock);
7526         mutex_init(&conf->cache_size_mutex);
7527
7528         init_waitqueue_head(&conf->wait_for_quiescent);
7529         init_waitqueue_head(&conf->wait_for_stripe);
7530         init_waitqueue_head(&conf->wait_for_overlap);
7531         INIT_LIST_HEAD(&conf->handle_list);
7532         INIT_LIST_HEAD(&conf->loprio_list);
7533         INIT_LIST_HEAD(&conf->hold_list);
7534         INIT_LIST_HEAD(&conf->delayed_list);
7535         INIT_LIST_HEAD(&conf->bitmap_list);
7536         init_llist_head(&conf->released_stripes);
7537         atomic_set(&conf->active_stripes, 0);
7538         atomic_set(&conf->preread_active_stripes, 0);
7539         atomic_set(&conf->active_aligned_reads, 0);
7540         spin_lock_init(&conf->pending_bios_lock);
7541         conf->batch_bio_dispatch = true;
7542         rdev_for_each(rdev, mddev) {
7543                 if (test_bit(Journal, &rdev->flags))
7544                         continue;
7545                 if (bdev_nonrot(rdev->bdev)) {
7546                         conf->batch_bio_dispatch = false;
7547                         break;
7548                 }
7549         }
7550
7551         conf->bypass_threshold = BYPASS_THRESHOLD;
7552         conf->recovery_disabled = mddev->recovery_disabled - 1;
7553
7554         conf->raid_disks = mddev->raid_disks;
7555         if (mddev->reshape_position == MaxSector)
7556                 conf->previous_raid_disks = mddev->raid_disks;
7557         else
7558                 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
7559         max_disks = max(conf->raid_disks, conf->previous_raid_disks);
7560
7561         conf->disks = kcalloc(max_disks, sizeof(struct disk_info),
7562                               GFP_KERNEL);
7563
7564         if (!conf->disks)
7565                 goto abort;
7566
7567         for (i = 0; i < max_disks; i++) {
7568                 conf->disks[i].extra_page = alloc_page(GFP_KERNEL);
7569                 if (!conf->disks[i].extra_page)
7570                         goto abort;
7571         }
7572
7573         ret = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
7574         if (ret)
7575                 goto abort;
7576         conf->mddev = mddev;
7577
7578         ret = -ENOMEM;
7579         conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL);
7580         if (!conf->stripe_hashtbl)
7581                 goto abort;
7582
7583         /* We init hash_locks[0] separately to that it can be used
7584          * as the reference lock in the spin_lock_nest_lock() call
7585          * in lock_all_device_hash_locks_irq in order to convince
7586          * lockdep that we know what we are doing.
7587          */
7588         spin_lock_init(conf->hash_locks);
7589         for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
7590                 spin_lock_init(conf->hash_locks + i);
7591
7592         for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
7593                 INIT_LIST_HEAD(conf->inactive_list + i);
7594
7595         for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
7596                 INIT_LIST_HEAD(conf->temp_inactive_list + i);
7597
7598         atomic_set(&conf->r5c_cached_full_stripes, 0);
7599         INIT_LIST_HEAD(&conf->r5c_full_stripe_list);
7600         atomic_set(&conf->r5c_cached_partial_stripes, 0);
7601         INIT_LIST_HEAD(&conf->r5c_partial_stripe_list);
7602         atomic_set(&conf->r5c_flushing_full_stripes, 0);
7603         atomic_set(&conf->r5c_flushing_partial_stripes, 0);
7604
7605         conf->level = mddev->new_level;
7606         conf->chunk_sectors = mddev->new_chunk_sectors;
7607         ret = raid5_alloc_percpu(conf);
7608         if (ret)
7609                 goto abort;
7610
7611         pr_debug("raid456: run(%s) called.\n", mdname(mddev));
7612
7613         ret = -EIO;
7614         rdev_for_each(rdev, mddev) {
7615                 raid_disk = rdev->raid_disk;
7616                 if (raid_disk >= max_disks
7617                     || raid_disk < 0 || test_bit(Journal, &rdev->flags))
7618                         continue;
7619                 disk = conf->disks + raid_disk;
7620
7621                 if (test_bit(Replacement, &rdev->flags)) {
7622                         if (disk->replacement)
7623                                 goto abort;
7624                         RCU_INIT_POINTER(disk->replacement, rdev);
7625                 } else {
7626                         if (disk->rdev)
7627                                 goto abort;
7628                         RCU_INIT_POINTER(disk->rdev, rdev);
7629                 }
7630
7631                 if (test_bit(In_sync, &rdev->flags)) {
7632                         pr_info("md/raid:%s: device %pg operational as raid disk %d\n",
7633                                 mdname(mddev), rdev->bdev, raid_disk);
7634                 } else if (rdev->saved_raid_disk != raid_disk)
7635                         /* Cannot rely on bitmap to complete recovery */
7636                         conf->fullsync = 1;
7637         }
7638
7639         conf->level = mddev->new_level;
7640         if (conf->level == 6) {
7641                 conf->max_degraded = 2;
7642                 if (raid6_call.xor_syndrome)
7643                         conf->rmw_level = PARITY_ENABLE_RMW;
7644                 else
7645                         conf->rmw_level = PARITY_DISABLE_RMW;
7646         } else {
7647                 conf->max_degraded = 1;
7648                 conf->rmw_level = PARITY_ENABLE_RMW;
7649         }
7650         conf->algorithm = mddev->new_layout;
7651         conf->reshape_progress = mddev->reshape_position;
7652         if (conf->reshape_progress != MaxSector) {
7653                 conf->prev_chunk_sectors = mddev->chunk_sectors;
7654                 conf->prev_algo = mddev->layout;
7655         } else {
7656                 conf->prev_chunk_sectors = conf->chunk_sectors;
7657                 conf->prev_algo = conf->algorithm;
7658         }
7659
7660         conf->min_nr_stripes = NR_STRIPES;
7661         if (mddev->reshape_position != MaxSector) {
7662                 int stripes = max_t(int,
7663                         ((mddev->chunk_sectors << 9) / RAID5_STRIPE_SIZE(conf)) * 4,
7664                         ((mddev->new_chunk_sectors << 9) / RAID5_STRIPE_SIZE(conf)) * 4);
7665                 conf->min_nr_stripes = max(NR_STRIPES, stripes);
7666                 if (conf->min_nr_stripes != NR_STRIPES)
7667                         pr_info("md/raid:%s: force stripe size %d for reshape\n",
7668                                 mdname(mddev), conf->min_nr_stripes);
7669         }
7670         memory = conf->min_nr_stripes * (sizeof(struct stripe_head) +
7671                  max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
7672         atomic_set(&conf->empty_inactive_list_nr, NR_STRIPE_HASH_LOCKS);
7673         if (grow_stripes(conf, conf->min_nr_stripes)) {
7674                 pr_warn("md/raid:%s: couldn't allocate %dkB for buffers\n",
7675                         mdname(mddev), memory);
7676                 ret = -ENOMEM;
7677                 goto abort;
7678         } else
7679                 pr_debug("md/raid:%s: allocated %dkB\n", mdname(mddev), memory);
7680         /*
7681          * Losing a stripe head costs more than the time to refill it,
7682          * it reduces the queue depth and so can hurt throughput.
7683          * So set it rather large, scaled by number of devices.
7684          */
7685         conf->shrinker = shrinker_alloc(0, "md-raid5:%s", mdname(mddev));
7686         if (!conf->shrinker) {
7687                 ret = -ENOMEM;
7688                 pr_warn("md/raid:%s: couldn't allocate shrinker.\n",
7689                         mdname(mddev));
7690                 goto abort;
7691         }
7692
7693         conf->shrinker->seeks = DEFAULT_SEEKS * conf->raid_disks * 4;
7694         conf->shrinker->scan_objects = raid5_cache_scan;
7695         conf->shrinker->count_objects = raid5_cache_count;
7696         conf->shrinker->batch = 128;
7697         conf->shrinker->private_data = conf;
7698
7699         shrinker_register(conf->shrinker);
7700
7701         sprintf(pers_name, "raid%d", mddev->new_level);
7702         rcu_assign_pointer(conf->thread,
7703                            md_register_thread(raid5d, mddev, pers_name));
7704         if (!conf->thread) {
7705                 pr_warn("md/raid:%s: couldn't allocate thread.\n",
7706                         mdname(mddev));
7707                 ret = -ENOMEM;
7708                 goto abort;
7709         }
7710
7711         return conf;
7712
7713  abort:
7714         if (conf)
7715                 free_conf(conf);
7716         return ERR_PTR(ret);
7717 }
7718
7719 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
7720 {
7721         switch (algo) {
7722         case ALGORITHM_PARITY_0:
7723                 if (raid_disk < max_degraded)
7724                         return 1;
7725                 break;
7726         case ALGORITHM_PARITY_N:
7727                 if (raid_disk >= raid_disks - max_degraded)
7728                         return 1;
7729                 break;
7730         case ALGORITHM_PARITY_0_6:
7731                 if (raid_disk == 0 ||
7732                     raid_disk == raid_disks - 1)
7733                         return 1;
7734                 break;
7735         case ALGORITHM_LEFT_ASYMMETRIC_6:
7736         case ALGORITHM_RIGHT_ASYMMETRIC_6:
7737         case ALGORITHM_LEFT_SYMMETRIC_6:
7738         case ALGORITHM_RIGHT_SYMMETRIC_6:
7739                 if (raid_disk == raid_disks - 1)
7740                         return 1;
7741         }
7742         return 0;
7743 }
7744
7745 static void raid5_set_io_opt(struct r5conf *conf)
7746 {
7747         blk_queue_io_opt(conf->mddev->queue, (conf->chunk_sectors << 9) *
7748                          (conf->raid_disks - conf->max_degraded));
7749 }
7750
7751 static int raid5_run(struct mddev *mddev)
7752 {
7753         struct r5conf *conf;
7754         int dirty_parity_disks = 0;
7755         struct md_rdev *rdev;
7756         struct md_rdev *journal_dev = NULL;
7757         sector_t reshape_offset = 0;
7758         int i;
7759         long long min_offset_diff = 0;
7760         int first = 1;
7761
7762         if (mddev->recovery_cp != MaxSector)
7763                 pr_notice("md/raid:%s: not clean -- starting background reconstruction\n",
7764                           mdname(mddev));
7765
7766         rdev_for_each(rdev, mddev) {
7767                 long long diff;
7768
7769                 if (test_bit(Journal, &rdev->flags)) {
7770                         journal_dev = rdev;
7771                         continue;
7772                 }
7773                 if (rdev->raid_disk < 0)
7774                         continue;
7775                 diff = (rdev->new_data_offset - rdev->data_offset);
7776                 if (first) {
7777                         min_offset_diff = diff;
7778                         first = 0;
7779                 } else if (mddev->reshape_backwards &&
7780                          diff < min_offset_diff)
7781                         min_offset_diff = diff;
7782                 else if (!mddev->reshape_backwards &&
7783                          diff > min_offset_diff)
7784                         min_offset_diff = diff;
7785         }
7786
7787         if ((test_bit(MD_HAS_JOURNAL, &mddev->flags) || journal_dev) &&
7788             (mddev->bitmap_info.offset || mddev->bitmap_info.file)) {
7789                 pr_notice("md/raid:%s: array cannot have both journal and bitmap\n",
7790                           mdname(mddev));
7791                 return -EINVAL;
7792         }
7793
7794         if (mddev->reshape_position != MaxSector) {
7795                 /* Check that we can continue the reshape.
7796                  * Difficulties arise if the stripe we would write to
7797                  * next is at or after the stripe we would read from next.
7798                  * For a reshape that changes the number of devices, this
7799                  * is only possible for a very short time, and mdadm makes
7800                  * sure that time appears to have past before assembling
7801                  * the array.  So we fail if that time hasn't passed.
7802                  * For a reshape that keeps the number of devices the same
7803                  * mdadm must be monitoring the reshape can keeping the
7804                  * critical areas read-only and backed up.  It will start
7805                  * the array in read-only mode, so we check for that.
7806                  */
7807                 sector_t here_new, here_old;
7808                 int old_disks;
7809                 int max_degraded = (mddev->level == 6 ? 2 : 1);
7810                 int chunk_sectors;
7811                 int new_data_disks;
7812
7813                 if (journal_dev) {
7814                         pr_warn("md/raid:%s: don't support reshape with journal - aborting.\n",
7815                                 mdname(mddev));
7816                         return -EINVAL;
7817                 }
7818
7819                 if (mddev->new_level != mddev->level) {
7820                         pr_warn("md/raid:%s: unsupported reshape required - aborting.\n",
7821                                 mdname(mddev));
7822                         return -EINVAL;
7823                 }
7824                 old_disks = mddev->raid_disks - mddev->delta_disks;
7825                 /* reshape_position must be on a new-stripe boundary, and one
7826                  * further up in new geometry must map after here in old
7827                  * geometry.
7828                  * If the chunk sizes are different, then as we perform reshape
7829                  * in units of the largest of the two, reshape_position needs
7830                  * be a multiple of the largest chunk size times new data disks.
7831                  */
7832                 here_new = mddev->reshape_position;
7833                 chunk_sectors = max(mddev->chunk_sectors, mddev->new_chunk_sectors);
7834                 new_data_disks = mddev->raid_disks - max_degraded;
7835                 if (sector_div(here_new, chunk_sectors * new_data_disks)) {
7836                         pr_warn("md/raid:%s: reshape_position not on a stripe boundary\n",
7837                                 mdname(mddev));
7838                         return -EINVAL;
7839                 }
7840                 reshape_offset = here_new * chunk_sectors;
7841                 /* here_new is the stripe we will write to */
7842                 here_old = mddev->reshape_position;
7843                 sector_div(here_old, chunk_sectors * (old_disks-max_degraded));
7844                 /* here_old is the first stripe that we might need to read
7845                  * from */
7846                 if (mddev->delta_disks == 0) {
7847                         /* We cannot be sure it is safe to start an in-place
7848                          * reshape.  It is only safe if user-space is monitoring
7849                          * and taking constant backups.
7850                          * mdadm always starts a situation like this in
7851                          * readonly mode so it can take control before
7852                          * allowing any writes.  So just check for that.
7853                          */
7854                         if (abs(min_offset_diff) >= mddev->chunk_sectors &&
7855                             abs(min_offset_diff) >= mddev->new_chunk_sectors)
7856                                 /* not really in-place - so OK */;
7857                         else if (mddev->ro == 0) {
7858                                 pr_warn("md/raid:%s: in-place reshape must be started in read-only mode - aborting\n",
7859                                         mdname(mddev));
7860                                 return -EINVAL;
7861                         }
7862                 } else if (mddev->reshape_backwards
7863                     ? (here_new * chunk_sectors + min_offset_diff <=
7864                        here_old * chunk_sectors)
7865                     : (here_new * chunk_sectors >=
7866                        here_old * chunk_sectors + (-min_offset_diff))) {
7867                         /* Reading from the same stripe as writing to - bad */
7868                         pr_warn("md/raid:%s: reshape_position too early for auto-recovery - aborting.\n",
7869                                 mdname(mddev));
7870                         return -EINVAL;
7871                 }
7872                 pr_debug("md/raid:%s: reshape will continue\n", mdname(mddev));
7873                 /* OK, we should be able to continue; */
7874         } else {
7875                 BUG_ON(mddev->level != mddev->new_level);
7876                 BUG_ON(mddev->layout != mddev->new_layout);
7877                 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
7878                 BUG_ON(mddev->delta_disks != 0);
7879         }
7880
7881         if (test_bit(MD_HAS_JOURNAL, &mddev->flags) &&
7882             test_bit(MD_HAS_PPL, &mddev->flags)) {
7883                 pr_warn("md/raid:%s: using journal device and PPL not allowed - disabling PPL\n",
7884                         mdname(mddev));
7885                 clear_bit(MD_HAS_PPL, &mddev->flags);
7886                 clear_bit(MD_HAS_MULTIPLE_PPLS, &mddev->flags);
7887         }
7888
7889         if (mddev->private == NULL)
7890                 conf = setup_conf(mddev);
7891         else
7892                 conf = mddev->private;
7893
7894         if (IS_ERR(conf))
7895                 return PTR_ERR(conf);
7896
7897         if (test_bit(MD_HAS_JOURNAL, &mddev->flags)) {
7898                 if (!journal_dev) {
7899                         pr_warn("md/raid:%s: journal disk is missing, force array readonly\n",
7900                                 mdname(mddev));
7901                         mddev->ro = 1;
7902                         set_disk_ro(mddev->gendisk, 1);
7903                 } else if (mddev->recovery_cp == MaxSector)
7904                         set_bit(MD_JOURNAL_CLEAN, &mddev->flags);
7905         }
7906
7907         conf->min_offset_diff = min_offset_diff;
7908         rcu_assign_pointer(mddev->thread, conf->thread);
7909         rcu_assign_pointer(conf->thread, NULL);
7910         mddev->private = conf;
7911
7912         for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
7913              i++) {
7914                 rdev = rdev_mdlock_deref(mddev, conf->disks[i].rdev);
7915                 if (!rdev && conf->disks[i].replacement) {
7916                         /* The replacement is all we have yet */
7917                         rdev = rdev_mdlock_deref(mddev,
7918                                                  conf->disks[i].replacement);
7919                         conf->disks[i].replacement = NULL;
7920                         clear_bit(Replacement, &rdev->flags);
7921                         rcu_assign_pointer(conf->disks[i].rdev, rdev);
7922                 }
7923                 if (!rdev)
7924                         continue;
7925                 if (rcu_access_pointer(conf->disks[i].replacement) &&
7926                     conf->reshape_progress != MaxSector) {
7927                         /* replacements and reshape simply do not mix. */
7928                         pr_warn("md: cannot handle concurrent replacement and reshape.\n");
7929                         goto abort;
7930                 }
7931                 if (test_bit(In_sync, &rdev->flags))
7932                         continue;
7933                 /* This disc is not fully in-sync.  However if it
7934                  * just stored parity (beyond the recovery_offset),
7935                  * when we don't need to be concerned about the
7936                  * array being dirty.
7937                  * When reshape goes 'backwards', we never have
7938                  * partially completed devices, so we only need
7939                  * to worry about reshape going forwards.
7940                  */
7941                 /* Hack because v0.91 doesn't store recovery_offset properly. */
7942                 if (mddev->major_version == 0 &&
7943                     mddev->minor_version > 90)
7944                         rdev->recovery_offset = reshape_offset;
7945
7946                 if (rdev->recovery_offset < reshape_offset) {
7947                         /* We need to check old and new layout */
7948                         if (!only_parity(rdev->raid_disk,
7949                                          conf->algorithm,
7950                                          conf->raid_disks,
7951                                          conf->max_degraded))
7952                                 continue;
7953                 }
7954                 if (!only_parity(rdev->raid_disk,
7955                                  conf->prev_algo,
7956                                  conf->previous_raid_disks,
7957                                  conf->max_degraded))
7958                         continue;
7959                 dirty_parity_disks++;
7960         }
7961
7962         /*
7963          * 0 for a fully functional array, 1 or 2 for a degraded array.
7964          */
7965         mddev->degraded = raid5_calc_degraded(conf);
7966
7967         if (has_failed(conf)) {
7968                 pr_crit("md/raid:%s: not enough operational devices (%d/%d failed)\n",
7969                         mdname(mddev), mddev->degraded, conf->raid_disks);
7970                 goto abort;
7971         }
7972
7973         /* device size must be a multiple of chunk size */
7974         mddev->dev_sectors &= ~((sector_t)mddev->chunk_sectors - 1);
7975         mddev->resync_max_sectors = mddev->dev_sectors;
7976
7977         if (mddev->degraded > dirty_parity_disks &&
7978             mddev->recovery_cp != MaxSector) {
7979                 if (test_bit(MD_HAS_PPL, &mddev->flags))
7980                         pr_crit("md/raid:%s: starting dirty degraded array with PPL.\n",
7981                                 mdname(mddev));
7982                 else if (mddev->ok_start_degraded)
7983                         pr_crit("md/raid:%s: starting dirty degraded array - data corruption possible.\n",
7984                                 mdname(mddev));
7985                 else {
7986                         pr_crit("md/raid:%s: cannot start dirty degraded array.\n",
7987                                 mdname(mddev));
7988                         goto abort;
7989                 }
7990         }
7991
7992         pr_info("md/raid:%s: raid level %d active with %d out of %d devices, algorithm %d\n",
7993                 mdname(mddev), conf->level,
7994                 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
7995                 mddev->new_layout);
7996
7997         print_raid5_conf(conf);
7998
7999         if (conf->reshape_progress != MaxSector) {
8000                 conf->reshape_safe = conf->reshape_progress;
8001                 atomic_set(&conf->reshape_stripes, 0);
8002                 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
8003                 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
8004                 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
8005                 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
8006                 rcu_assign_pointer(mddev->sync_thread,
8007                         md_register_thread(md_do_sync, mddev, "reshape"));
8008                 if (!mddev->sync_thread)
8009                         goto abort;
8010         }
8011
8012         /* Ok, everything is just fine now */
8013         if (mddev->to_remove == &raid5_attrs_group)
8014                 mddev->to_remove = NULL;
8015         else if (mddev->kobj.sd &&
8016             sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
8017                 pr_warn("raid5: failed to create sysfs attributes for %s\n",
8018                         mdname(mddev));
8019         md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
8020
8021         if (mddev->queue) {
8022                 int chunk_size;
8023                 /* read-ahead size must cover two whole stripes, which
8024                  * is 2 * (datadisks) * chunksize where 'n' is the
8025                  * number of raid devices
8026                  */
8027                 int data_disks = conf->previous_raid_disks - conf->max_degraded;
8028                 int stripe = data_disks *
8029                         ((mddev->chunk_sectors << 9) / PAGE_SIZE);
8030
8031                 chunk_size = mddev->chunk_sectors << 9;
8032                 blk_queue_io_min(mddev->queue, chunk_size);
8033                 raid5_set_io_opt(conf);
8034                 mddev->queue->limits.raid_partial_stripes_expensive = 1;
8035                 /*
8036                  * We can only discard a whole stripe. It doesn't make sense to
8037                  * discard data disk but write parity disk
8038                  */
8039                 stripe = stripe * PAGE_SIZE;
8040                 stripe = roundup_pow_of_two(stripe);
8041                 mddev->queue->limits.discard_granularity = stripe;
8042
8043                 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
8044
8045                 rdev_for_each(rdev, mddev) {
8046                         disk_stack_limits(mddev->gendisk, rdev->bdev,
8047                                           rdev->data_offset << 9);
8048                         disk_stack_limits(mddev->gendisk, rdev->bdev,
8049                                           rdev->new_data_offset << 9);
8050                 }
8051
8052                 /*
8053                  * zeroing is required, otherwise data
8054                  * could be lost. Consider a scenario: discard a stripe
8055                  * (the stripe could be inconsistent if
8056                  * discard_zeroes_data is 0); write one disk of the
8057                  * stripe (the stripe could be inconsistent again
8058                  * depending on which disks are used to calculate
8059                  * parity); the disk is broken; The stripe data of this
8060                  * disk is lost.
8061                  *
8062                  * We only allow DISCARD if the sysadmin has confirmed that
8063                  * only safe devices are in use by setting a module parameter.
8064                  * A better idea might be to turn DISCARD into WRITE_ZEROES
8065                  * requests, as that is required to be safe.
8066                  */
8067                 if (!devices_handle_discard_safely ||
8068                     mddev->queue->limits.max_discard_sectors < (stripe >> 9) ||
8069                     mddev->queue->limits.discard_granularity < stripe)
8070                         blk_queue_max_discard_sectors(mddev->queue, 0);
8071
8072                 /*
8073                  * Requests require having a bitmap for each stripe.
8074                  * Limit the max sectors based on this.
8075                  */
8076                 blk_queue_max_hw_sectors(mddev->queue,
8077                         RAID5_MAX_REQ_STRIPES << RAID5_STRIPE_SHIFT(conf));
8078
8079                 /* No restrictions on the number of segments in the request */
8080                 blk_queue_max_segments(mddev->queue, USHRT_MAX);
8081         }
8082
8083         if (log_init(conf, journal_dev, raid5_has_ppl(conf)))
8084                 goto abort;
8085
8086         return 0;
8087 abort:
8088         md_unregister_thread(mddev, &mddev->thread);
8089         print_raid5_conf(conf);
8090         free_conf(conf);
8091         mddev->private = NULL;
8092         pr_warn("md/raid:%s: failed to run raid set.\n", mdname(mddev));
8093         return -EIO;
8094 }
8095
8096 static void raid5_free(struct mddev *mddev, void *priv)
8097 {
8098         struct r5conf *conf = priv;
8099
8100         free_conf(conf);
8101         mddev->to_remove = &raid5_attrs_group;
8102 }
8103
8104 static void raid5_status(struct seq_file *seq, struct mddev *mddev)
8105 {
8106         struct r5conf *conf = mddev->private;
8107         int i;
8108
8109         seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
8110                 conf->chunk_sectors / 2, mddev->layout);
8111         seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
8112         rcu_read_lock();
8113         for (i = 0; i < conf->raid_disks; i++) {
8114                 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
8115                 seq_printf (seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
8116         }
8117         rcu_read_unlock();
8118         seq_printf (seq, "]");
8119 }
8120
8121 static void print_raid5_conf (struct r5conf *conf)
8122 {
8123         struct md_rdev *rdev;
8124         int i;
8125
8126         pr_debug("RAID conf printout:\n");
8127         if (!conf) {
8128                 pr_debug("(conf==NULL)\n");
8129                 return;
8130         }
8131         pr_debug(" --- level:%d rd:%d wd:%d\n", conf->level,
8132                conf->raid_disks,
8133                conf->raid_disks - conf->mddev->degraded);
8134
8135         rcu_read_lock();
8136         for (i = 0; i < conf->raid_disks; i++) {
8137                 rdev = rcu_dereference(conf->disks[i].rdev);
8138                 if (rdev)
8139                         pr_debug(" disk %d, o:%d, dev:%pg\n",
8140                                i, !test_bit(Faulty, &rdev->flags),
8141                                rdev->bdev);
8142         }
8143         rcu_read_unlock();
8144 }
8145
8146 static int raid5_spare_active(struct mddev *mddev)
8147 {
8148         int i;
8149         struct r5conf *conf = mddev->private;
8150         struct md_rdev *rdev, *replacement;
8151         int count = 0;
8152         unsigned long flags;
8153
8154         for (i = 0; i < conf->raid_disks; i++) {
8155                 rdev = rdev_mdlock_deref(mddev, conf->disks[i].rdev);
8156                 replacement = rdev_mdlock_deref(mddev,
8157                                                 conf->disks[i].replacement);
8158                 if (replacement
8159                     && replacement->recovery_offset == MaxSector
8160                     && !test_bit(Faulty, &replacement->flags)
8161                     && !test_and_set_bit(In_sync, &replacement->flags)) {
8162                         /* Replacement has just become active. */
8163                         if (!rdev
8164                             || !test_and_clear_bit(In_sync, &rdev->flags))
8165                                 count++;
8166                         if (rdev) {
8167                                 /* Replaced device not technically faulty,
8168                                  * but we need to be sure it gets removed
8169                                  * and never re-added.
8170                                  */
8171                                 set_bit(Faulty, &rdev->flags);
8172                                 sysfs_notify_dirent_safe(
8173                                         rdev->sysfs_state);
8174                         }
8175                         sysfs_notify_dirent_safe(replacement->sysfs_state);
8176                 } else if (rdev
8177                     && rdev->recovery_offset == MaxSector
8178                     && !test_bit(Faulty, &rdev->flags)
8179                     && !test_and_set_bit(In_sync, &rdev->flags)) {
8180                         count++;
8181                         sysfs_notify_dirent_safe(rdev->sysfs_state);
8182                 }
8183         }
8184         spin_lock_irqsave(&conf->device_lock, flags);
8185         mddev->degraded = raid5_calc_degraded(conf);
8186         spin_unlock_irqrestore(&conf->device_lock, flags);
8187         print_raid5_conf(conf);
8188         return count;
8189 }
8190
8191 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
8192 {
8193         struct r5conf *conf = mddev->private;
8194         int err = 0;
8195         int number = rdev->raid_disk;
8196         struct md_rdev __rcu **rdevp;
8197         struct disk_info *p;
8198         struct md_rdev *tmp;
8199
8200         print_raid5_conf(conf);
8201         if (test_bit(Journal, &rdev->flags) && conf->log) {
8202                 /*
8203                  * we can't wait pending write here, as this is called in
8204                  * raid5d, wait will deadlock.
8205                  * neilb: there is no locking about new writes here,
8206                  * so this cannot be safe.
8207                  */
8208                 if (atomic_read(&conf->active_stripes) ||
8209                     atomic_read(&conf->r5c_cached_full_stripes) ||
8210                     atomic_read(&conf->r5c_cached_partial_stripes)) {
8211                         return -EBUSY;
8212                 }
8213                 log_exit(conf);
8214                 return 0;
8215         }
8216         if (unlikely(number >= conf->pool_size))
8217                 return 0;
8218         p = conf->disks + number;
8219         if (rdev == rcu_access_pointer(p->rdev))
8220                 rdevp = &p->rdev;
8221         else if (rdev == rcu_access_pointer(p->replacement))
8222                 rdevp = &p->replacement;
8223         else
8224                 return 0;
8225
8226         if (number >= conf->raid_disks &&
8227             conf->reshape_progress == MaxSector)
8228                 clear_bit(In_sync, &rdev->flags);
8229
8230         if (test_bit(In_sync, &rdev->flags) ||
8231             atomic_read(&rdev->nr_pending)) {
8232                 err = -EBUSY;
8233                 goto abort;
8234         }
8235         /* Only remove non-faulty devices if recovery
8236          * isn't possible.
8237          */
8238         if (!test_bit(Faulty, &rdev->flags) &&
8239             mddev->recovery_disabled != conf->recovery_disabled &&
8240             !has_failed(conf) &&
8241             (!rcu_access_pointer(p->replacement) ||
8242              rcu_access_pointer(p->replacement) == rdev) &&
8243             number < conf->raid_disks) {
8244                 err = -EBUSY;
8245                 goto abort;
8246         }
8247         *rdevp = NULL;
8248         if (!test_bit(RemoveSynchronized, &rdev->flags)) {
8249                 lockdep_assert_held(&mddev->reconfig_mutex);
8250                 synchronize_rcu();
8251                 if (atomic_read(&rdev->nr_pending)) {
8252                         /* lost the race, try later */
8253                         err = -EBUSY;
8254                         rcu_assign_pointer(*rdevp, rdev);
8255                 }
8256         }
8257         if (!err) {
8258                 err = log_modify(conf, rdev, false);
8259                 if (err)
8260                         goto abort;
8261         }
8262
8263         tmp = rcu_access_pointer(p->replacement);
8264         if (tmp) {
8265                 /* We must have just cleared 'rdev' */
8266                 rcu_assign_pointer(p->rdev, tmp);
8267                 clear_bit(Replacement, &tmp->flags);
8268                 smp_mb(); /* Make sure other CPUs may see both as identical
8269                            * but will never see neither - if they are careful
8270                            */
8271                 rcu_assign_pointer(p->replacement, NULL);
8272
8273                 if (!err)
8274                         err = log_modify(conf, tmp, true);
8275         }
8276
8277         clear_bit(WantReplacement, &rdev->flags);
8278 abort:
8279
8280         print_raid5_conf(conf);
8281         return err;
8282 }
8283
8284 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
8285 {
8286         struct r5conf *conf = mddev->private;
8287         int ret, err = -EEXIST;
8288         int disk;
8289         struct disk_info *p;
8290         struct md_rdev *tmp;
8291         int first = 0;
8292         int last = conf->raid_disks - 1;
8293
8294         if (test_bit(Journal, &rdev->flags)) {
8295                 if (conf->log)
8296                         return -EBUSY;
8297
8298                 rdev->raid_disk = 0;
8299                 /*
8300                  * The array is in readonly mode if journal is missing, so no
8301                  * write requests running. We should be safe
8302                  */
8303                 ret = log_init(conf, rdev, false);
8304                 if (ret)
8305                         return ret;
8306
8307                 ret = r5l_start(conf->log);
8308                 if (ret)
8309                         return ret;
8310
8311                 return 0;
8312         }
8313         if (mddev->recovery_disabled == conf->recovery_disabled)
8314                 return -EBUSY;
8315
8316         if (rdev->saved_raid_disk < 0 && has_failed(conf))
8317                 /* no point adding a device */
8318                 return -EINVAL;
8319
8320         if (rdev->raid_disk >= 0)
8321                 first = last = rdev->raid_disk;
8322
8323         /*
8324          * find the disk ... but prefer rdev->saved_raid_disk
8325          * if possible.
8326          */
8327         if (rdev->saved_raid_disk >= first &&
8328             rdev->saved_raid_disk <= last &&
8329             conf->disks[rdev->saved_raid_disk].rdev == NULL)
8330                 first = rdev->saved_raid_disk;
8331
8332         for (disk = first; disk <= last; disk++) {
8333                 p = conf->disks + disk;
8334                 if (p->rdev == NULL) {
8335                         clear_bit(In_sync, &rdev->flags);
8336                         rdev->raid_disk = disk;
8337                         if (rdev->saved_raid_disk != disk)
8338                                 conf->fullsync = 1;
8339                         rcu_assign_pointer(p->rdev, rdev);
8340
8341                         err = log_modify(conf, rdev, true);
8342
8343                         goto out;
8344                 }
8345         }
8346         for (disk = first; disk <= last; disk++) {
8347                 p = conf->disks + disk;
8348                 tmp = rdev_mdlock_deref(mddev, p->rdev);
8349                 if (test_bit(WantReplacement, &tmp->flags) &&
8350                     mddev->reshape_position == MaxSector &&
8351                     p->replacement == NULL) {
8352                         clear_bit(In_sync, &rdev->flags);
8353                         set_bit(Replacement, &rdev->flags);
8354                         rdev->raid_disk = disk;
8355                         err = 0;
8356                         conf->fullsync = 1;
8357                         rcu_assign_pointer(p->replacement, rdev);
8358                         break;
8359                 }
8360         }
8361 out:
8362         print_raid5_conf(conf);
8363         return err;
8364 }
8365
8366 static int raid5_resize(struct mddev *mddev, sector_t sectors)
8367 {
8368         /* no resync is happening, and there is enough space
8369          * on all devices, so we can resize.
8370          * We need to make sure resync covers any new space.
8371          * If the array is shrinking we should possibly wait until
8372          * any io in the removed space completes, but it hardly seems
8373          * worth it.
8374          */
8375         sector_t newsize;
8376         struct r5conf *conf = mddev->private;
8377
8378         if (raid5_has_log(conf) || raid5_has_ppl(conf))
8379                 return -EINVAL;
8380         sectors &= ~((sector_t)conf->chunk_sectors - 1);
8381         newsize = raid5_size(mddev, sectors, mddev->raid_disks);
8382         if (mddev->external_size &&
8383             mddev->array_sectors > newsize)
8384                 return -EINVAL;
8385         if (mddev->bitmap) {
8386                 int ret = md_bitmap_resize(mddev->bitmap, sectors, 0, 0);
8387                 if (ret)
8388                         return ret;
8389         }
8390         md_set_array_sectors(mddev, newsize);
8391         if (sectors > mddev->dev_sectors &&
8392             mddev->recovery_cp > mddev->dev_sectors) {
8393                 mddev->recovery_cp = mddev->dev_sectors;
8394                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
8395         }
8396         mddev->dev_sectors = sectors;
8397         mddev->resync_max_sectors = sectors;
8398         return 0;
8399 }
8400
8401 static int check_stripe_cache(struct mddev *mddev)
8402 {
8403         /* Can only proceed if there are plenty of stripe_heads.
8404          * We need a minimum of one full stripe,, and for sensible progress
8405          * it is best to have about 4 times that.
8406          * If we require 4 times, then the default 256 4K stripe_heads will
8407          * allow for chunk sizes up to 256K, which is probably OK.
8408          * If the chunk size is greater, user-space should request more
8409          * stripe_heads first.
8410          */
8411         struct r5conf *conf = mddev->private;
8412         if (((mddev->chunk_sectors << 9) / RAID5_STRIPE_SIZE(conf)) * 4
8413             > conf->min_nr_stripes ||
8414             ((mddev->new_chunk_sectors << 9) / RAID5_STRIPE_SIZE(conf)) * 4
8415             > conf->min_nr_stripes) {
8416                 pr_warn("md/raid:%s: reshape: not enough stripes.  Needed %lu\n",
8417                         mdname(mddev),
8418                         ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
8419                          / RAID5_STRIPE_SIZE(conf))*4);
8420                 return 0;
8421         }
8422         return 1;
8423 }
8424
8425 static int check_reshape(struct mddev *mddev)
8426 {
8427         struct r5conf *conf = mddev->private;
8428
8429         if (raid5_has_log(conf) || raid5_has_ppl(conf))
8430                 return -EINVAL;
8431         if (mddev->delta_disks == 0 &&
8432             mddev->new_layout == mddev->layout &&
8433             mddev->new_chunk_sectors == mddev->chunk_sectors)
8434                 return 0; /* nothing to do */
8435         if (has_failed(conf))
8436                 return -EINVAL;
8437         if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
8438                 /* We might be able to shrink, but the devices must
8439                  * be made bigger first.
8440                  * For raid6, 4 is the minimum size.
8441                  * Otherwise 2 is the minimum
8442                  */
8443                 int min = 2;
8444                 if (mddev->level == 6)
8445                         min = 4;
8446                 if (mddev->raid_disks + mddev->delta_disks < min)
8447                         return -EINVAL;
8448         }
8449
8450         if (!check_stripe_cache(mddev))
8451                 return -ENOSPC;
8452
8453         if (mddev->new_chunk_sectors > mddev->chunk_sectors ||
8454             mddev->delta_disks > 0)
8455                 if (resize_chunks(conf,
8456                                   conf->previous_raid_disks
8457                                   + max(0, mddev->delta_disks),
8458                                   max(mddev->new_chunk_sectors,
8459                                       mddev->chunk_sectors)
8460                             ) < 0)
8461                         return -ENOMEM;
8462
8463         if (conf->previous_raid_disks + mddev->delta_disks <= conf->pool_size)
8464                 return 0; /* never bother to shrink */
8465         return resize_stripes(conf, (conf->previous_raid_disks
8466                                      + mddev->delta_disks));
8467 }
8468
8469 static int raid5_start_reshape(struct mddev *mddev)
8470 {
8471         struct r5conf *conf = mddev->private;
8472         struct md_rdev *rdev;
8473         int spares = 0;
8474         int i;
8475         unsigned long flags;
8476
8477         if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
8478                 return -EBUSY;
8479
8480         if (!check_stripe_cache(mddev))
8481                 return -ENOSPC;
8482
8483         if (has_failed(conf))
8484                 return -EINVAL;
8485
8486         /* raid5 can't handle concurrent reshape and recovery */
8487         if (mddev->recovery_cp < MaxSector)
8488                 return -EBUSY;
8489         for (i = 0; i < conf->raid_disks; i++)
8490                 if (rdev_mdlock_deref(mddev, conf->disks[i].replacement))
8491                         return -EBUSY;
8492
8493         rdev_for_each(rdev, mddev) {
8494                 if (!test_bit(In_sync, &rdev->flags)
8495                     && !test_bit(Faulty, &rdev->flags))
8496                         spares++;
8497         }
8498
8499         if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
8500                 /* Not enough devices even to make a degraded array
8501                  * of that size
8502                  */
8503                 return -EINVAL;
8504
8505         /* Refuse to reduce size of the array.  Any reductions in
8506          * array size must be through explicit setting of array_size
8507          * attribute.
8508          */
8509         if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
8510             < mddev->array_sectors) {
8511                 pr_warn("md/raid:%s: array size must be reduced before number of disks\n",
8512                         mdname(mddev));
8513                 return -EINVAL;
8514         }
8515
8516         atomic_set(&conf->reshape_stripes, 0);
8517         spin_lock_irq(&conf->device_lock);
8518         write_seqcount_begin(&conf->gen_lock);
8519         conf->previous_raid_disks = conf->raid_disks;
8520         conf->raid_disks += mddev->delta_disks;
8521         conf->prev_chunk_sectors = conf->chunk_sectors;
8522         conf->chunk_sectors = mddev->new_chunk_sectors;
8523         conf->prev_algo = conf->algorithm;
8524         conf->algorithm = mddev->new_layout;
8525         conf->generation++;
8526         /* Code that selects data_offset needs to see the generation update
8527          * if reshape_progress has been set - so a memory barrier needed.
8528          */
8529         smp_mb();
8530         if (mddev->reshape_backwards)
8531                 conf->reshape_progress = raid5_size(mddev, 0, 0);
8532         else
8533                 conf->reshape_progress = 0;
8534         conf->reshape_safe = conf->reshape_progress;
8535         write_seqcount_end(&conf->gen_lock);
8536         spin_unlock_irq(&conf->device_lock);
8537
8538         /* Now make sure any requests that proceeded on the assumption
8539          * the reshape wasn't running - like Discard or Read - have
8540          * completed.
8541          */
8542         raid5_quiesce(mddev, true);
8543         raid5_quiesce(mddev, false);
8544
8545         /* Add some new drives, as many as will fit.
8546          * We know there are enough to make the newly sized array work.
8547          * Don't add devices if we are reducing the number of
8548          * devices in the array.  This is because it is not possible
8549          * to correctly record the "partially reconstructed" state of
8550          * such devices during the reshape and confusion could result.
8551          */
8552         if (mddev->delta_disks >= 0) {
8553                 rdev_for_each(rdev, mddev)
8554                         if (rdev->raid_disk < 0 &&
8555                             !test_bit(Faulty, &rdev->flags)) {
8556                                 if (raid5_add_disk(mddev, rdev) == 0) {
8557                                         if (rdev->raid_disk
8558                                             >= conf->previous_raid_disks)
8559                                                 set_bit(In_sync, &rdev->flags);
8560                                         else
8561                                                 rdev->recovery_offset = 0;
8562
8563                                         /* Failure here is OK */
8564                                         sysfs_link_rdev(mddev, rdev);
8565                                 }
8566                         } else if (rdev->raid_disk >= conf->previous_raid_disks
8567                                    && !test_bit(Faulty, &rdev->flags)) {
8568                                 /* This is a spare that was manually added */
8569                                 set_bit(In_sync, &rdev->flags);
8570                         }
8571
8572                 /* When a reshape changes the number of devices,
8573                  * ->degraded is measured against the larger of the
8574                  * pre and post number of devices.
8575                  */
8576                 spin_lock_irqsave(&conf->device_lock, flags);
8577                 mddev->degraded = raid5_calc_degraded(conf);
8578                 spin_unlock_irqrestore(&conf->device_lock, flags);
8579         }
8580         mddev->raid_disks = conf->raid_disks;
8581         mddev->reshape_position = conf->reshape_progress;
8582         set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
8583
8584         clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
8585         clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
8586         clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
8587         set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
8588         set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
8589         rcu_assign_pointer(mddev->sync_thread,
8590                            md_register_thread(md_do_sync, mddev, "reshape"));
8591         if (!mddev->sync_thread) {
8592                 mddev->recovery = 0;
8593                 spin_lock_irq(&conf->device_lock);
8594                 write_seqcount_begin(&conf->gen_lock);
8595                 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
8596                 mddev->new_chunk_sectors =
8597                         conf->chunk_sectors = conf->prev_chunk_sectors;
8598                 mddev->new_layout = conf->algorithm = conf->prev_algo;
8599                 rdev_for_each(rdev, mddev)
8600                         rdev->new_data_offset = rdev->data_offset;
8601                 smp_wmb();
8602                 conf->generation --;
8603                 conf->reshape_progress = MaxSector;
8604                 mddev->reshape_position = MaxSector;
8605                 write_seqcount_end(&conf->gen_lock);
8606                 spin_unlock_irq(&conf->device_lock);
8607                 return -EAGAIN;
8608         }
8609         conf->reshape_checkpoint = jiffies;
8610         md_wakeup_thread(mddev->sync_thread);
8611         md_new_event();
8612         return 0;
8613 }
8614
8615 /* This is called from the reshape thread and should make any
8616  * changes needed in 'conf'
8617  */
8618 static void end_reshape(struct r5conf *conf)
8619 {
8620
8621         if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
8622                 struct md_rdev *rdev;
8623
8624                 spin_lock_irq(&conf->device_lock);
8625                 conf->previous_raid_disks = conf->raid_disks;
8626                 md_finish_reshape(conf->mddev);
8627                 smp_wmb();
8628                 conf->reshape_progress = MaxSector;
8629                 conf->mddev->reshape_position = MaxSector;
8630                 rdev_for_each(rdev, conf->mddev)
8631                         if (rdev->raid_disk >= 0 &&
8632                             !test_bit(Journal, &rdev->flags) &&
8633                             !test_bit(In_sync, &rdev->flags))
8634                                 rdev->recovery_offset = MaxSector;
8635                 spin_unlock_irq(&conf->device_lock);
8636                 wake_up(&conf->wait_for_overlap);
8637
8638                 if (conf->mddev->queue)
8639                         raid5_set_io_opt(conf);
8640         }
8641 }
8642
8643 /* This is called from the raid5d thread with mddev_lock held.
8644  * It makes config changes to the device.
8645  */
8646 static void raid5_finish_reshape(struct mddev *mddev)
8647 {
8648         struct r5conf *conf = mddev->private;
8649         struct md_rdev *rdev;
8650
8651         if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
8652
8653                 if (mddev->delta_disks <= 0) {
8654                         int d;
8655                         spin_lock_irq(&conf->device_lock);
8656                         mddev->degraded = raid5_calc_degraded(conf);
8657                         spin_unlock_irq(&conf->device_lock);
8658                         for (d = conf->raid_disks ;
8659                              d < conf->raid_disks - mddev->delta_disks;
8660                              d++) {
8661                                 rdev = rdev_mdlock_deref(mddev,
8662                                                          conf->disks[d].rdev);
8663                                 if (rdev)
8664                                         clear_bit(In_sync, &rdev->flags);
8665                                 rdev = rdev_mdlock_deref(mddev,
8666                                                 conf->disks[d].replacement);
8667                                 if (rdev)
8668                                         clear_bit(In_sync, &rdev->flags);
8669                         }
8670                 }
8671                 mddev->layout = conf->algorithm;
8672                 mddev->chunk_sectors = conf->chunk_sectors;
8673                 mddev->reshape_position = MaxSector;
8674                 mddev->delta_disks = 0;
8675                 mddev->reshape_backwards = 0;
8676         }
8677 }
8678
8679 static void raid5_quiesce(struct mddev *mddev, int quiesce)
8680 {
8681         struct r5conf *conf = mddev->private;
8682
8683         if (quiesce) {
8684                 /* stop all writes */
8685                 lock_all_device_hash_locks_irq(conf);
8686                 /* '2' tells resync/reshape to pause so that all
8687                  * active stripes can drain
8688                  */
8689                 r5c_flush_cache(conf, INT_MAX);
8690                 /* need a memory barrier to make sure read_one_chunk() sees
8691                  * quiesce started and reverts to slow (locked) path.
8692                  */
8693                 smp_store_release(&conf->quiesce, 2);
8694                 wait_event_cmd(conf->wait_for_quiescent,
8695                                     atomic_read(&conf->active_stripes) == 0 &&
8696                                     atomic_read(&conf->active_aligned_reads) == 0,
8697                                     unlock_all_device_hash_locks_irq(conf),
8698                                     lock_all_device_hash_locks_irq(conf));
8699                 conf->quiesce = 1;
8700                 unlock_all_device_hash_locks_irq(conf);
8701                 /* allow reshape to continue */
8702                 wake_up(&conf->wait_for_overlap);
8703         } else {
8704                 /* re-enable writes */
8705                 lock_all_device_hash_locks_irq(conf);
8706                 conf->quiesce = 0;
8707                 wake_up(&conf->wait_for_quiescent);
8708                 wake_up(&conf->wait_for_overlap);
8709                 unlock_all_device_hash_locks_irq(conf);
8710         }
8711         log_quiesce(conf, quiesce);
8712 }
8713
8714 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
8715 {
8716         struct r0conf *raid0_conf = mddev->private;
8717         sector_t sectors;
8718
8719         /* for raid0 takeover only one zone is supported */
8720         if (raid0_conf->nr_strip_zones > 1) {
8721                 pr_warn("md/raid:%s: cannot takeover raid0 with more than one zone.\n",
8722                         mdname(mddev));
8723                 return ERR_PTR(-EINVAL);
8724         }
8725
8726         sectors = raid0_conf->strip_zone[0].zone_end;
8727         sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
8728         mddev->dev_sectors = sectors;
8729         mddev->new_level = level;
8730         mddev->new_layout = ALGORITHM_PARITY_N;
8731         mddev->new_chunk_sectors = mddev->chunk_sectors;
8732         mddev->raid_disks += 1;
8733         mddev->delta_disks = 1;
8734         /* make sure it will be not marked as dirty */
8735         mddev->recovery_cp = MaxSector;
8736
8737         return setup_conf(mddev);
8738 }
8739
8740 static void *raid5_takeover_raid1(struct mddev *mddev)
8741 {
8742         int chunksect;
8743         void *ret;
8744
8745         if (mddev->raid_disks != 2 ||
8746             mddev->degraded > 1)
8747                 return ERR_PTR(-EINVAL);
8748
8749         /* Should check if there are write-behind devices? */
8750
8751         chunksect = 64*2; /* 64K by default */
8752
8753         /* The array must be an exact multiple of chunksize */
8754         while (chunksect && (mddev->array_sectors & (chunksect-1)))
8755                 chunksect >>= 1;
8756
8757         if ((chunksect<<9) < RAID5_STRIPE_SIZE((struct r5conf *)mddev->private))
8758                 /* array size does not allow a suitable chunk size */
8759                 return ERR_PTR(-EINVAL);
8760
8761         mddev->new_level = 5;
8762         mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
8763         mddev->new_chunk_sectors = chunksect;
8764
8765         ret = setup_conf(mddev);
8766         if (!IS_ERR(ret))
8767                 mddev_clear_unsupported_flags(mddev,
8768                         UNSUPPORTED_MDDEV_FLAGS);
8769         return ret;
8770 }
8771
8772 static void *raid5_takeover_raid6(struct mddev *mddev)
8773 {
8774         int new_layout;
8775
8776         switch (mddev->layout) {
8777         case ALGORITHM_LEFT_ASYMMETRIC_6:
8778                 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
8779                 break;
8780         case ALGORITHM_RIGHT_ASYMMETRIC_6:
8781                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
8782                 break;
8783         case ALGORITHM_LEFT_SYMMETRIC_6:
8784                 new_layout = ALGORITHM_LEFT_SYMMETRIC;
8785                 break;
8786         case ALGORITHM_RIGHT_SYMMETRIC_6:
8787                 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
8788                 break;
8789         case ALGORITHM_PARITY_0_6:
8790                 new_layout = ALGORITHM_PARITY_0;
8791                 break;
8792         case ALGORITHM_PARITY_N:
8793                 new_layout = ALGORITHM_PARITY_N;
8794                 break;
8795         default:
8796                 return ERR_PTR(-EINVAL);
8797         }
8798         mddev->new_level = 5;
8799         mddev->new_layout = new_layout;
8800         mddev->delta_disks = -1;
8801         mddev->raid_disks -= 1;
8802         return setup_conf(mddev);
8803 }
8804
8805 static int raid5_check_reshape(struct mddev *mddev)
8806 {
8807         /* For a 2-drive array, the layout and chunk size can be changed
8808          * immediately as not restriping is needed.
8809          * For larger arrays we record the new value - after validation
8810          * to be used by a reshape pass.
8811          */
8812         struct r5conf *conf = mddev->private;
8813         int new_chunk = mddev->new_chunk_sectors;
8814
8815         if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
8816                 return -EINVAL;
8817         if (new_chunk > 0) {
8818                 if (!is_power_of_2(new_chunk))
8819                         return -EINVAL;
8820                 if (new_chunk < (PAGE_SIZE>>9))
8821                         return -EINVAL;
8822                 if (mddev->array_sectors & (new_chunk-1))
8823                         /* not factor of array size */
8824                         return -EINVAL;
8825         }
8826
8827         /* They look valid */
8828
8829         if (mddev->raid_disks == 2) {
8830                 /* can make the change immediately */
8831                 if (mddev->new_layout >= 0) {
8832                         conf->algorithm = mddev->new_layout;
8833                         mddev->layout = mddev->new_layout;
8834                 }
8835                 if (new_chunk > 0) {
8836                         conf->chunk_sectors = new_chunk ;
8837                         mddev->chunk_sectors = new_chunk;
8838                 }
8839                 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
8840                 md_wakeup_thread(mddev->thread);
8841         }
8842         return check_reshape(mddev);
8843 }
8844
8845 static int raid6_check_reshape(struct mddev *mddev)
8846 {
8847         int new_chunk = mddev->new_chunk_sectors;
8848
8849         if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
8850                 return -EINVAL;
8851         if (new_chunk > 0) {
8852                 if (!is_power_of_2(new_chunk))
8853                         return -EINVAL;
8854                 if (new_chunk < (PAGE_SIZE >> 9))
8855                         return -EINVAL;
8856                 if (mddev->array_sectors & (new_chunk-1))
8857                         /* not factor of array size */
8858                         return -EINVAL;
8859         }
8860
8861         /* They look valid */
8862         return check_reshape(mddev);
8863 }
8864
8865 static void *raid5_takeover(struct mddev *mddev)
8866 {
8867         /* raid5 can take over:
8868          *  raid0 - if there is only one strip zone - make it a raid4 layout
8869          *  raid1 - if there are two drives.  We need to know the chunk size
8870          *  raid4 - trivial - just use a raid4 layout.
8871          *  raid6 - Providing it is a *_6 layout
8872          */
8873         if (mddev->level == 0)
8874                 return raid45_takeover_raid0(mddev, 5);
8875         if (mddev->level == 1)
8876                 return raid5_takeover_raid1(mddev);
8877         if (mddev->level == 4) {
8878                 mddev->new_layout = ALGORITHM_PARITY_N;
8879                 mddev->new_level = 5;
8880                 return setup_conf(mddev);
8881         }
8882         if (mddev->level == 6)
8883                 return raid5_takeover_raid6(mddev);
8884
8885         return ERR_PTR(-EINVAL);
8886 }
8887
8888 static void *raid4_takeover(struct mddev *mddev)
8889 {
8890         /* raid4 can take over:
8891          *  raid0 - if there is only one strip zone
8892          *  raid5 - if layout is right
8893          */
8894         if (mddev->level == 0)
8895                 return raid45_takeover_raid0(mddev, 4);
8896         if (mddev->level == 5 &&
8897             mddev->layout == ALGORITHM_PARITY_N) {
8898                 mddev->new_layout = 0;
8899                 mddev->new_level = 4;
8900                 return setup_conf(mddev);
8901         }
8902         return ERR_PTR(-EINVAL);
8903 }
8904
8905 static struct md_personality raid5_personality;
8906
8907 static void *raid6_takeover(struct mddev *mddev)
8908 {
8909         /* Currently can only take over a raid5.  We map the
8910          * personality to an equivalent raid6 personality
8911          * with the Q block at the end.
8912          */
8913         int new_layout;
8914
8915         if (mddev->pers != &raid5_personality)
8916                 return ERR_PTR(-EINVAL);
8917         if (mddev->degraded > 1)
8918                 return ERR_PTR(-EINVAL);
8919         if (mddev->raid_disks > 253)
8920                 return ERR_PTR(-EINVAL);
8921         if (mddev->raid_disks < 3)
8922                 return ERR_PTR(-EINVAL);
8923
8924         switch (mddev->layout) {
8925         case ALGORITHM_LEFT_ASYMMETRIC:
8926                 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
8927                 break;
8928         case ALGORITHM_RIGHT_ASYMMETRIC:
8929                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
8930                 break;
8931         case ALGORITHM_LEFT_SYMMETRIC:
8932                 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
8933                 break;
8934         case ALGORITHM_RIGHT_SYMMETRIC:
8935                 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
8936                 break;
8937         case ALGORITHM_PARITY_0:
8938                 new_layout = ALGORITHM_PARITY_0_6;
8939                 break;
8940         case ALGORITHM_PARITY_N:
8941                 new_layout = ALGORITHM_PARITY_N;
8942                 break;
8943         default:
8944                 return ERR_PTR(-EINVAL);
8945         }
8946         mddev->new_level = 6;
8947         mddev->new_layout = new_layout;
8948         mddev->delta_disks = 1;
8949         mddev->raid_disks += 1;
8950         return setup_conf(mddev);
8951 }
8952
8953 static int raid5_change_consistency_policy(struct mddev *mddev, const char *buf)
8954 {
8955         struct r5conf *conf;
8956         int err;
8957
8958         err = mddev_suspend_and_lock(mddev);
8959         if (err)
8960                 return err;
8961         conf = mddev->private;
8962         if (!conf) {
8963                 mddev_unlock_and_resume(mddev);
8964                 return -ENODEV;
8965         }
8966
8967         if (strncmp(buf, "ppl", 3) == 0) {
8968                 /* ppl only works with RAID 5 */
8969                 if (!raid5_has_ppl(conf) && conf->level == 5) {
8970                         err = log_init(conf, NULL, true);
8971                         if (!err) {
8972                                 err = resize_stripes(conf, conf->pool_size);
8973                                 if (err)
8974                                         log_exit(conf);
8975                         }
8976                 } else
8977                         err = -EINVAL;
8978         } else if (strncmp(buf, "resync", 6) == 0) {
8979                 if (raid5_has_ppl(conf)) {
8980                         log_exit(conf);
8981                         err = resize_stripes(conf, conf->pool_size);
8982                 } else if (test_bit(MD_HAS_JOURNAL, &conf->mddev->flags) &&
8983                            r5l_log_disk_error(conf)) {
8984                         bool journal_dev_exists = false;
8985                         struct md_rdev *rdev;
8986
8987                         rdev_for_each(rdev, mddev)
8988                                 if (test_bit(Journal, &rdev->flags)) {
8989                                         journal_dev_exists = true;
8990                                         break;
8991                                 }
8992
8993                         if (!journal_dev_exists)
8994                                 clear_bit(MD_HAS_JOURNAL, &mddev->flags);
8995                         else  /* need remove journal device first */
8996                                 err = -EBUSY;
8997                 } else
8998                         err = -EINVAL;
8999         } else {
9000                 err = -EINVAL;
9001         }
9002
9003         if (!err)
9004                 md_update_sb(mddev, 1);
9005
9006         mddev_unlock_and_resume(mddev);
9007
9008         return err;
9009 }
9010
9011 static int raid5_start(struct mddev *mddev)
9012 {
9013         struct r5conf *conf = mddev->private;
9014
9015         return r5l_start(conf->log);
9016 }
9017
9018 static struct md_personality raid6_personality =
9019 {
9020         .name           = "raid6",
9021         .level          = 6,
9022         .owner          = THIS_MODULE,
9023         .make_request   = raid5_make_request,
9024         .run            = raid5_run,
9025         .start          = raid5_start,
9026         .free           = raid5_free,
9027         .status         = raid5_status,
9028         .error_handler  = raid5_error,
9029         .hot_add_disk   = raid5_add_disk,
9030         .hot_remove_disk= raid5_remove_disk,
9031         .spare_active   = raid5_spare_active,
9032         .sync_request   = raid5_sync_request,
9033         .resize         = raid5_resize,
9034         .size           = raid5_size,
9035         .check_reshape  = raid6_check_reshape,
9036         .start_reshape  = raid5_start_reshape,
9037         .finish_reshape = raid5_finish_reshape,
9038         .quiesce        = raid5_quiesce,
9039         .takeover       = raid6_takeover,
9040         .change_consistency_policy = raid5_change_consistency_policy,
9041 };
9042 static struct md_personality raid5_personality =
9043 {
9044         .name           = "raid5",
9045         .level          = 5,
9046         .owner          = THIS_MODULE,
9047         .make_request   = raid5_make_request,
9048         .run            = raid5_run,
9049         .start          = raid5_start,
9050         .free           = raid5_free,
9051         .status         = raid5_status,
9052         .error_handler  = raid5_error,
9053         .hot_add_disk   = raid5_add_disk,
9054         .hot_remove_disk= raid5_remove_disk,
9055         .spare_active   = raid5_spare_active,
9056         .sync_request   = raid5_sync_request,
9057         .resize         = raid5_resize,
9058         .size           = raid5_size,
9059         .check_reshape  = raid5_check_reshape,
9060         .start_reshape  = raid5_start_reshape,
9061         .finish_reshape = raid5_finish_reshape,
9062         .quiesce        = raid5_quiesce,
9063         .takeover       = raid5_takeover,
9064         .change_consistency_policy = raid5_change_consistency_policy,
9065 };
9066
9067 static struct md_personality raid4_personality =
9068 {
9069         .name           = "raid4",
9070         .level          = 4,
9071         .owner          = THIS_MODULE,
9072         .make_request   = raid5_make_request,
9073         .run            = raid5_run,
9074         .start          = raid5_start,
9075         .free           = raid5_free,
9076         .status         = raid5_status,
9077         .error_handler  = raid5_error,
9078         .hot_add_disk   = raid5_add_disk,
9079         .hot_remove_disk= raid5_remove_disk,
9080         .spare_active   = raid5_spare_active,
9081         .sync_request   = raid5_sync_request,
9082         .resize         = raid5_resize,
9083         .size           = raid5_size,
9084         .check_reshape  = raid5_check_reshape,
9085         .start_reshape  = raid5_start_reshape,
9086         .finish_reshape = raid5_finish_reshape,
9087         .quiesce        = raid5_quiesce,
9088         .takeover       = raid4_takeover,
9089         .change_consistency_policy = raid5_change_consistency_policy,
9090 };
9091
9092 static int __init raid5_init(void)
9093 {
9094         int ret;
9095
9096         raid5_wq = alloc_workqueue("raid5wq",
9097                 WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
9098         if (!raid5_wq)
9099                 return -ENOMEM;
9100
9101         ret = cpuhp_setup_state_multi(CPUHP_MD_RAID5_PREPARE,
9102                                       "md/raid5:prepare",
9103                                       raid456_cpu_up_prepare,
9104                                       raid456_cpu_dead);
9105         if (ret) {
9106                 destroy_workqueue(raid5_wq);
9107                 return ret;
9108         }
9109         register_md_personality(&raid6_personality);
9110         register_md_personality(&raid5_personality);
9111         register_md_personality(&raid4_personality);
9112         return 0;
9113 }
9114
9115 static void raid5_exit(void)
9116 {
9117         unregister_md_personality(&raid6_personality);
9118         unregister_md_personality(&raid5_personality);
9119         unregister_md_personality(&raid4_personality);
9120         cpuhp_remove_multi_state(CPUHP_MD_RAID5_PREPARE);
9121         destroy_workqueue(raid5_wq);
9122 }
9123
9124 module_init(raid5_init);
9125 module_exit(raid5_exit);
9126 MODULE_LICENSE("GPL");
9127 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
9128 MODULE_ALIAS("md-personality-4"); /* RAID5 */
9129 MODULE_ALIAS("md-raid5");
9130 MODULE_ALIAS("md-raid4");
9131 MODULE_ALIAS("md-level-5");
9132 MODULE_ALIAS("md-level-4");
9133 MODULE_ALIAS("md-personality-8"); /* RAID6 */
9134 MODULE_ALIAS("md-raid6");
9135 MODULE_ALIAS("md-level-6");
9136
9137 /* This used to be two separate modules, they were: */
9138 MODULE_ALIAS("raid5");
9139 MODULE_ALIAS("raid6");