mm/z3fold.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * z3fold.c
   4  *
   5  * Author: Vitaly Wool <vitaly.wool@konsulko.com>
   6  * Copyright (C) 2016, Sony Mobile Communications Inc.
   7  *
   8  * This implementation is based on zbud written by Seth Jennings.
   9  *
  10  * z3fold is an special purpose allocator for storing compressed pages. It
  11  * can store up to three compressed pages per page which improves the
  12  * compression ratio of zbud while retaining its main concepts (e. g. always
  13  * storing an integral number of objects per page) and simplicity.
  14  * It still has simple and deterministic reclaim properties that make it
  15  * preferable to a higher density approach (with no requirement on integral
  16  * number of object per page) when reclaim is used.
  17  *
  18  * As in zbud, pages are divided into "chunks".  The size of the chunks is
  19  * fixed at compile time and is determined by NCHUNKS_ORDER below.
  20  *
  21  * z3fold doesn't export any API and is meant to be used via zpool API.
  22  */
  23
  24 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  25
  26 #include <linux/atomic.h>
  27 #include <linux/sched.h>
  28 #include <linux/cpumask.h>
  29 #include <linux/list.h>
  30 #include <linux/mm.h>
  31 #include <linux/module.h>
  32 #include <linux/page-flags.h>
  33 #include <linux/migrate.h>
  34 #include <linux/node.h>
  35 #include <linux/compaction.h>
  36 #include <linux/percpu.h>
  37 #include <linux/mount.h>
  38 #include <linux/pseudo_fs.h>
  39 #include <linux/fs.h>
  40 #include <linux/preempt.h>
  41 #include <linux/workqueue.h>
  42 #include <linux/slab.h>
  43 #include <linux/spinlock.h>
  44 #include <linux/zpool.h>
  45 #include <linux/magic.h>
  46 #include <linux/kmemleak.h>
  47
  48 /*
  49  * NCHUNKS_ORDER determines the internal allocation granularity, effectively
  50  * adjusting internal fragmentation.  It also determines the number of
  51  * freelists maintained in each pool. NCHUNKS_ORDER of 6 means that the
  52  * allocation granularity will be in chunks of size PAGE_SIZE/64. Some chunks
  53  * in the beginning of an allocated page are occupied by z3fold header, so
  54  * NCHUNKS will be calculated to 63 (or 62 in case CONFIG_DEBUG_SPINLOCK=y),
  55  * which shows the max number of free chunks in z3fold page, also there will
  56  * be 63, or 62, respectively, freelists per pool.
  57  */
  58 #define NCHUNKS_ORDER   6
  59
  60 #define CHUNK_SHIFT     (PAGE_SHIFT - NCHUNKS_ORDER)
  61 #define CHUNK_SIZE      (1 << CHUNK_SHIFT)
  62 #define ZHDR_SIZE_ALIGNED round_up(sizeof(struct z3fold_header), CHUNK_SIZE)
  63 #define ZHDR_CHUNKS     (ZHDR_SIZE_ALIGNED >> CHUNK_SHIFT)
  64 #define TOTAL_CHUNKS    (PAGE_SIZE >> CHUNK_SHIFT)
  65 #define NCHUNKS         (TOTAL_CHUNKS - ZHDR_CHUNKS)
  66
  67 #define BUDDY_MASK      (0x3)
  68 #define BUDDY_SHIFT     2
  69 #define SLOTS_ALIGN     (0x40)
  70
  71 /*****************
  72  * Structures
  73 *****************/
  74 struct z3fold_pool;
  75 struct z3fold_ops {
  76         int (*evict)(struct z3fold_pool *pool, unsigned long handle);
  77 };
  78
  79 enum buddy {
  80         HEADLESS = 0,
  81         FIRST,
  82         MIDDLE,
  83         LAST,
  84         BUDDIES_MAX = LAST
  85 };
  86
  87 struct z3fold_buddy_slots {
  88         /*
  89          * we are using BUDDY_MASK in handle_to_buddy etc. so there should
  90          * be enough slots to hold all possible variants
  91          */
  92         unsigned long slot[BUDDY_MASK + 1];
  93         unsigned long pool; /* back link */
  94         rwlock_t lock;
  95 };
  96 #define HANDLE_FLAG_MASK        (0x03)
  97
  98 /*
  99  * struct z3fold_header - z3fold page metadata occupying first chunks of each
 100  *                      z3fold page, except for HEADLESS pages
 101  * @buddy:              links the z3fold page into the relevant list in the
 102  *                      pool
 103  * @page_lock:          per-page lock
 104  * @refcount:           reference count for the z3fold page
 105  * @work:               work_struct for page layout optimization
 106  * @slots:              pointer to the structure holding buddy slots
 107  * @pool:               pointer to the containing pool
 108  * @cpu:                CPU which this page "belongs" to
 109  * @first_chunks:       the size of the first buddy in chunks, 0 if free
 110  * @middle_chunks:      the size of the middle buddy in chunks, 0 if free
 111  * @last_chunks:        the size of the last buddy in chunks, 0 if free
 112  * @first_num:          the starting number (for the first handle)
 113  * @mapped_count:       the number of objects currently mapped
 114  */
 115 struct z3fold_header {
 116         struct list_head buddy;
 117         spinlock_t page_lock;
 118         struct kref refcount;
 119         struct work_struct work;
 120         struct z3fold_buddy_slots *slots;
 121         struct z3fold_pool *pool;
 122         short cpu;
 123         unsigned short first_chunks;
 124         unsigned short middle_chunks;
 125         unsigned short last_chunks;
 126         unsigned short start_middle;
 127         unsigned short first_num:2;
 128         unsigned short mapped_count:2;
 129         unsigned short foreign_handles:2;
 130 };
 131
 132 /**
 133  * struct z3fold_pool - stores metadata for each z3fold pool
 134  * @name:       pool name
 135  * @lock:       protects pool unbuddied/lru lists
 136  * @stale_lock: protects pool stale page list
 137  * @unbuddied:  per-cpu array of lists tracking z3fold pages that contain 2-
 138  *              buddies; the list each z3fold page is added to depends on
 139  *              the size of its free region.
 140  * @lru:        list tracking the z3fold pages in LRU order by most recently
 141  *              added buddy.
 142  * @stale:      list of pages marked for freeing
 143  * @pages_nr:   number of z3fold pages in the pool.
 144  * @c_handle:   cache for z3fold_buddy_slots allocation
 145  * @ops:        pointer to a structure of user defined operations specified at
 146  *              pool creation time.
 147  * @zpool:      zpool driver
 148  * @zpool_ops:  zpool operations structure with an evict callback
 149  * @compact_wq: workqueue for page layout background optimization
 150  * @release_wq: workqueue for safe page release
 151  * @work:       work_struct for safe page release
 152  * @inode:      inode for z3fold pseudo filesystem
 153  *
 154  * This structure is allocated at pool creation time and maintains metadata
 155  * pertaining to a particular z3fold pool.
 156  */
 157 struct z3fold_pool {
 158         const char *name;
 159         spinlock_t lock;
 160         spinlock_t stale_lock;
 161         struct list_head *unbuddied;
 162         struct list_head lru;
 163         struct list_head stale;
 164         atomic64_t pages_nr;
 165         struct kmem_cache *c_handle;
 166         const struct z3fold_ops *ops;
 167         struct zpool *zpool;
 168         const struct zpool_ops *zpool_ops;
 169         struct workqueue_struct *compact_wq;
 170         struct workqueue_struct *release_wq;
 171         struct work_struct work;
 172         struct inode *inode;
 173 };
 174
 175 /*
 176  * Internal z3fold page flags
 177  */
 178 enum z3fold_page_flags {
 179         PAGE_HEADLESS = 0,
 180         MIDDLE_CHUNK_MAPPED,
 181         NEEDS_COMPACTING,
 182         PAGE_STALE,
 183         PAGE_CLAIMED, /* by either reclaim or free */
 184         PAGE_MIGRATED, /* page is migrated and soon to be released */
 185 };
 186
 187 /*
 188  * handle flags, go under HANDLE_FLAG_MASK
 189  */
 190 enum z3fold_handle_flags {
 191         HANDLES_NOFREE = 0,
 192 };
 193
 194 /*
 195  * Forward declarations
 196  */
 197 static struct z3fold_header *__z3fold_alloc(struct z3fold_pool *, size_t, bool);
 198 static void compact_page_work(struct work_struct *w);
 199
 200 /*****************
 201  * Helpers
 202 *****************/
 203
 204 /* Converts an allocation size in bytes to size in z3fold chunks */
 205 static int size_to_chunks(size_t size)
 206 {
 207         return (size + CHUNK_SIZE - 1) >> CHUNK_SHIFT;
 208 }
 209
 210 #define for_each_unbuddied_list(_iter, _begin) \
 211         for ((_iter) = (_begin); (_iter) < NCHUNKS; (_iter)++)
 212
 213 static inline struct z3fold_buddy_slots *alloc_slots(struct z3fold_pool *pool,
 214                                                         gfp_t gfp)
 215 {
 216         struct z3fold_buddy_slots *slots = kmem_cache_zalloc(pool->c_handle,
 217                                                              gfp);
 218
 219         if (slots) {
 220                 /* It will be freed separately in free_handle(). */
 221                 kmemleak_not_leak(slots);
 222                 slots->pool = (unsigned long)pool;
 223                 rwlock_init(&slots->lock);
 224         }
 225
 226         return slots;
 227 }
 228
 229 static inline struct z3fold_pool *slots_to_pool(struct z3fold_buddy_slots *s)
 230 {
 231         return (struct z3fold_pool *)(s->pool & ~HANDLE_FLAG_MASK);
 232 }
 233
 234 static inline struct z3fold_buddy_slots *handle_to_slots(unsigned long handle)
 235 {
 236         return (struct z3fold_buddy_slots *)(handle & ~(SLOTS_ALIGN - 1));
 237 }
 238
 239 /* Lock a z3fold page */
 240 static inline void z3fold_page_lock(struct z3fold_header *zhdr)
 241 {
 242         spin_lock(&zhdr->page_lock);
 243 }
 244
 245 /* Try to lock a z3fold page */
 246 static inline int z3fold_page_trylock(struct z3fold_header *zhdr)
 247 {
 248         return spin_trylock(&zhdr->page_lock);
 249 }
 250
 251 /* Unlock a z3fold page */
 252 static inline void z3fold_page_unlock(struct z3fold_header *zhdr)
 253 {
 254         spin_unlock(&zhdr->page_lock);
 255 }
 256
 257 /* return locked z3fold page if it's not headless */
 258 static inline struct z3fold_header *get_z3fold_header(unsigned long handle)
 259 {
 260         struct z3fold_buddy_slots *slots;
 261         struct z3fold_header *zhdr;
 262         int locked = 0;
 263
 264         if (!(handle & (1 << PAGE_HEADLESS))) {
 265                 slots = handle_to_slots(handle);
 266                 do {
 267                         unsigned long addr;
 268
 269                         read_lock(&slots->lock);
 270                         addr = *(unsigned long *)handle;
 271                         zhdr = (struct z3fold_header *)(addr & PAGE_MASK);
 272                         locked = z3fold_page_trylock(zhdr);
 273                         read_unlock(&slots->lock);
 274                         if (locked) {
 275                                 struct page *page = virt_to_page(zhdr);
 276
 277                                 if (!test_bit(PAGE_MIGRATED, &page->private))
 278                                         break;
 279                                 z3fold_page_unlock(zhdr);
 280                         }
 281                         cpu_relax();
 282                 } while (true);
 283         } else {
 284                 zhdr = (struct z3fold_header *)(handle & PAGE_MASK);
 285         }
 286
 287         return zhdr;
 288 }
 289
 290 static inline void put_z3fold_header(struct z3fold_header *zhdr)
 291 {
 292         struct page *page = virt_to_page(zhdr);
 293
 294         if (!test_bit(PAGE_HEADLESS, &page->private))
 295                 z3fold_page_unlock(zhdr);
 296 }
 297
 298 static inline void free_handle(unsigned long handle, struct z3fold_header *zhdr)
 299 {
 300         struct z3fold_buddy_slots *slots;
 301         int i;
 302         bool is_free;
 303
 304         if (WARN_ON(*(unsigned long *)handle == 0))
 305                 return;
 306
 307         slots = handle_to_slots(handle);
 308         write_lock(&slots->lock);
 309         *(unsigned long *)handle = 0;
 310
 311         if (test_bit(HANDLES_NOFREE, &slots->pool)) {
 312                 write_unlock(&slots->lock);
 313                 return; /* simple case, nothing else to do */
 314         }
 315
 316         if (zhdr->slots != slots)
 317                 zhdr->foreign_handles--;
 318
 319         is_free = true;
 320         for (i = 0; i <= BUDDY_MASK; i++) {
 321                 if (slots->slot[i]) {
 322                         is_free = false;
 323                         break;
 324                 }
 325         }
 326         write_unlock(&slots->lock);
 327
 328         if (is_free) {
 329                 struct z3fold_pool *pool = slots_to_pool(slots);
 330
 331                 if (zhdr->slots == slots)
 332                         zhdr->slots = NULL;
 333                 kmem_cache_free(pool->c_handle, slots);
 334         }
 335 }
 336
 337 static int z3fold_init_fs_context(struct fs_context *fc)
 338 {
 339         return init_pseudo(fc, Z3FOLD_MAGIC) ? 0 : -ENOMEM;
 340 }
 341
 342 static struct file_system_type z3fold_fs = {
 343         .name           = "z3fold",
 344         .init_fs_context = z3fold_init_fs_context,
 345         .kill_sb        = kill_anon_super,
 346 };
 347
 348 static struct vfsmount *z3fold_mnt;
 349 static int __init z3fold_mount(void)
 350 {
 351         int ret = 0;
 352
 353         z3fold_mnt = kern_mount(&z3fold_fs);
 354         if (IS_ERR(z3fold_mnt))
 355                 ret = PTR_ERR(z3fold_mnt);
 356
 357         return ret;
 358 }
 359
 360 static void z3fold_unmount(void)
 361 {
 362         kern_unmount(z3fold_mnt);
 363 }
 364
 365 static const struct address_space_operations z3fold_aops;
 366 static int z3fold_register_migration(struct z3fold_pool *pool)
 367 {
 368         pool->inode = alloc_anon_inode(z3fold_mnt->mnt_sb);
 369         if (IS_ERR(pool->inode)) {
 370                 pool->inode = NULL;
 371                 return 1;
 372         }
 373
 374         pool->inode->i_mapping->private_data = pool;
 375         pool->inode->i_mapping->a_ops = &z3fold_aops;
 376         return 0;
 377 }
 378
 379 static void z3fold_unregister_migration(struct z3fold_pool *pool)
 380 {
 381         if (pool->inode)
 382                 iput(pool->inode);
 383 }
 384
 385 /* Initializes the z3fold header of a newly allocated z3fold page */
 386 static struct z3fold_header *init_z3fold_page(struct page *page, bool headless,
 387                                         struct z3fold_pool *pool, gfp_t gfp)
 388 {
 389         struct z3fold_header *zhdr = page_address(page);
 390         struct z3fold_buddy_slots *slots;
 391
 392         INIT_LIST_HEAD(&page->lru);
 393         clear_bit(PAGE_HEADLESS, &page->private);
 394         clear_bit(MIDDLE_CHUNK_MAPPED, &page->private);
 395         clear_bit(NEEDS_COMPACTING, &page->private);
 396         clear_bit(PAGE_STALE, &page->private);
 397         clear_bit(PAGE_CLAIMED, &page->private);
 398         clear_bit(PAGE_MIGRATED, &page->private);
 399         if (headless)
 400                 return zhdr;
 401
 402         slots = alloc_slots(pool, gfp);
 403         if (!slots)
 404                 return NULL;
 405
 406         memset(zhdr, 0, sizeof(*zhdr));
 407         spin_lock_init(&zhdr->page_lock);
 408         kref_init(&zhdr->refcount);
 409         zhdr->cpu = -1;
 410         zhdr->slots = slots;
 411         zhdr->pool = pool;
 412         INIT_LIST_HEAD(&zhdr->buddy);
 413         INIT_WORK(&zhdr->work, compact_page_work);
 414         return zhdr;
 415 }
 416
 417 /* Resets the struct page fields and frees the page */
 418 static void free_z3fold_page(struct page *page, bool headless)
 419 {
 420         if (!headless) {
 421                 lock_page(page);
 422                 __ClearPageMovable(page);
 423                 unlock_page(page);
 424         }
 425         __free_page(page);
 426 }
 427
 428 /* Helper function to build the index */
 429 static inline int __idx(struct z3fold_header *zhdr, enum buddy bud)
 430 {
 431         return (bud + zhdr->first_num) & BUDDY_MASK;
 432 }
 433
 434 /*
 435  * Encodes the handle of a particular buddy within a z3fold page
 436  * Pool lock should be held as this function accesses first_num
 437  */
 438 static unsigned long __encode_handle(struct z3fold_header *zhdr,
 439                                 struct z3fold_buddy_slots *slots,
 440                                 enum buddy bud)
 441 {
 442         unsigned long h = (unsigned long)zhdr;
 443         int idx = 0;
 444
 445         /*
 446          * For a headless page, its handle is its pointer with the extra
 447          * PAGE_HEADLESS bit set
 448          */
 449         if (bud == HEADLESS)
 450                 return h | (1 << PAGE_HEADLESS);
 451
 452         /* otherwise, return pointer to encoded handle */
 453         idx = __idx(zhdr, bud);
 454         h += idx;
 455         if (bud == LAST)
 456                 h |= (zhdr->last_chunks << BUDDY_SHIFT);
 457
 458         write_lock(&slots->lock);
 459         slots->slot[idx] = h;
 460         write_unlock(&slots->lock);
 461         return (unsigned long)&slots->slot[idx];
 462 }
 463
 464 static unsigned long encode_handle(struct z3fold_header *zhdr, enum buddy bud)
 465 {
 466         return __encode_handle(zhdr, zhdr->slots, bud);
 467 }
 468
 469 /* only for LAST bud, returns zero otherwise */
 470 static unsigned short handle_to_chunks(unsigned long handle)
 471 {
 472         struct z3fold_buddy_slots *slots = handle_to_slots(handle);
 473         unsigned long addr;
 474
 475         read_lock(&slots->lock);
 476         addr = *(unsigned long *)handle;
 477         read_unlock(&slots->lock);
 478         return (addr & ~PAGE_MASK) >> BUDDY_SHIFT;
 479 }
 480
 481 /*
 482  * (handle & BUDDY_MASK) < zhdr->first_num is possible in encode_handle
 483  *  but that doesn't matter. because the masking will result in the
 484  *  correct buddy number.
 485  */
 486 static enum buddy handle_to_buddy(unsigned long handle)
 487 {
 488         struct z3fold_header *zhdr;
 489         struct z3fold_buddy_slots *slots = handle_to_slots(handle);
 490         unsigned long addr;
 491
 492         read_lock(&slots->lock);
 493         WARN_ON(handle & (1 << PAGE_HEADLESS));
 494         addr = *(unsigned long *)handle;
 495         read_unlock(&slots->lock);
 496         zhdr = (struct z3fold_header *)(addr & PAGE_MASK);
 497         return (addr - zhdr->first_num) & BUDDY_MASK;
 498 }
 499
 500 static inline struct z3fold_pool *zhdr_to_pool(struct z3fold_header *zhdr)
 501 {
 502         return zhdr->pool;
 503 }
 504
 505 static void __release_z3fold_page(struct z3fold_header *zhdr, bool locked)
 506 {
 507         struct page *page = virt_to_page(zhdr);
 508         struct z3fold_pool *pool = zhdr_to_pool(zhdr);
 509
 510         WARN_ON(!list_empty(&zhdr->buddy));
 511         set_bit(PAGE_STALE, &page->private);
 512         clear_bit(NEEDS_COMPACTING, &page->private);
 513         spin_lock(&pool->lock);
 514         if (!list_empty(&page->lru))
 515                 list_del_init(&page->lru);
 516         spin_unlock(&pool->lock);
 517
 518         if (locked)
 519                 z3fold_page_unlock(zhdr);
 520
 521         spin_lock(&pool->stale_lock);
 522         list_add(&zhdr->buddy, &pool->stale);
 523         queue_work(pool->release_wq, &pool->work);
 524         spin_unlock(&pool->stale_lock);
 525
 526         atomic64_dec(&pool->pages_nr);
 527 }
 528
 529 static void release_z3fold_page_locked(struct kref *ref)
 530 {
 531         struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
 532                                                 refcount);
 533         WARN_ON(z3fold_page_trylock(zhdr));
 534         __release_z3fold_page(zhdr, true);
 535 }
 536
 537 static void release_z3fold_page_locked_list(struct kref *ref)
 538 {
 539         struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
 540                                                refcount);
 541         struct z3fold_pool *pool = zhdr_to_pool(zhdr);
 542
 543         spin_lock(&pool->lock);
 544         list_del_init(&zhdr->buddy);
 545         spin_unlock(&pool->lock);
 546
 547         WARN_ON(z3fold_page_trylock(zhdr));
 548         __release_z3fold_page(zhdr, true);
 549 }
 550
 551 static void free_pages_work(struct work_struct *w)
 552 {
 553         struct z3fold_pool *pool = container_of(w, struct z3fold_pool, work);
 554
 555         spin_lock(&pool->stale_lock);
 556         while (!list_empty(&pool->stale)) {
 557                 struct z3fold_header *zhdr = list_first_entry(&pool->stale,
 558                                                 struct z3fold_header, buddy);
 559                 struct page *page = virt_to_page(zhdr);
 560
 561                 list_del(&zhdr->buddy);
 562                 if (WARN_ON(!test_bit(PAGE_STALE, &page->private)))
 563                         continue;
 564                 spin_unlock(&pool->stale_lock);
 565                 cancel_work_sync(&zhdr->work);
 566                 free_z3fold_page(page, false);
 567                 cond_resched();
 568                 spin_lock(&pool->stale_lock);
 569         }
 570         spin_unlock(&pool->stale_lock);
 571 }
 572
 573 /*
 574  * Returns the number of free chunks in a z3fold page.
 575  * NB: can't be used with HEADLESS pages.
 576  */
 577 static int num_free_chunks(struct z3fold_header *zhdr)
 578 {
 579         int nfree;
 580         /*
 581          * If there is a middle object, pick up the bigger free space
 582          * either before or after it. Otherwise just subtract the number
 583          * of chunks occupied by the first and the last objects.
 584          */
 585         if (zhdr->middle_chunks != 0) {
 586                 int nfree_before = zhdr->first_chunks ?
 587                         0 : zhdr->start_middle - ZHDR_CHUNKS;
 588                 int nfree_after = zhdr->last_chunks ?
 589                         0 : TOTAL_CHUNKS -
 590                                 (zhdr->start_middle + zhdr->middle_chunks);
 591                 nfree = max(nfree_before, nfree_after);
 592         } else
 593                 nfree = NCHUNKS - zhdr->first_chunks - zhdr->last_chunks;
 594         return nfree;
 595 }
 596
 597 /* Add to the appropriate unbuddied list */
 598 static inline void add_to_unbuddied(struct z3fold_pool *pool,
 599                                 struct z3fold_header *zhdr)
 600 {
 601         if (zhdr->first_chunks == 0 || zhdr->last_chunks == 0 ||
 602                         zhdr->middle_chunks == 0) {
 603                 struct list_head *unbuddied;
 604                 int freechunks = num_free_chunks(zhdr);
 605
 606                 migrate_disable();
 607                 unbuddied = this_cpu_ptr(pool->unbuddied);
 608                 spin_lock(&pool->lock);
 609                 list_add(&zhdr->buddy, &unbuddied[freechunks]);
 610                 spin_unlock(&pool->lock);
 611                 zhdr->cpu = smp_processor_id();
 612                 migrate_enable();
 613         }
 614 }
 615
 616 static inline enum buddy get_free_buddy(struct z3fold_header *zhdr, int chunks)
 617 {
 618         enum buddy bud = HEADLESS;
 619
 620         if (zhdr->middle_chunks) {
 621                 if (!zhdr->first_chunks &&
 622                     chunks <= zhdr->start_middle - ZHDR_CHUNKS)
 623                         bud = FIRST;
 624                 else if (!zhdr->last_chunks)
 625                         bud = LAST;
 626         } else {
 627                 if (!zhdr->first_chunks)
 628                         bud = FIRST;
 629                 else if (!zhdr->last_chunks)
 630                         bud = LAST;
 631                 else
 632                         bud = MIDDLE;
 633         }
 634
 635         return bud;
 636 }
 637
 638 static inline void *mchunk_memmove(struct z3fold_header *zhdr,
 639                                 unsigned short dst_chunk)
 640 {
 641         void *beg = zhdr;
 642         return memmove(beg + (dst_chunk << CHUNK_SHIFT),
 643                        beg + (zhdr->start_middle << CHUNK_SHIFT),
 644                        zhdr->middle_chunks << CHUNK_SHIFT);
 645 }
 646
 647 static inline bool buddy_single(struct z3fold_header *zhdr)
 648 {
 649         return !((zhdr->first_chunks && zhdr->middle_chunks) ||
 650                         (zhdr->first_chunks && zhdr->last_chunks) ||
 651                         (zhdr->middle_chunks && zhdr->last_chunks));
 652 }
 653
 654 static struct z3fold_header *compact_single_buddy(struct z3fold_header *zhdr)
 655 {
 656         struct z3fold_pool *pool = zhdr_to_pool(zhdr);
 657         void *p = zhdr;
 658         unsigned long old_handle = 0;
 659         size_t sz = 0;
 660         struct z3fold_header *new_zhdr = NULL;
 661         int first_idx = __idx(zhdr, FIRST);
 662         int middle_idx = __idx(zhdr, MIDDLE);
 663         int last_idx = __idx(zhdr, LAST);
 664         unsigned short *moved_chunks = NULL;
 665
 666         /*
 667          * No need to protect slots here -- all the slots are "local" and
 668          * the page lock is already taken
 669          */
 670         if (zhdr->first_chunks && zhdr->slots->slot[first_idx]) {
 671                 p += ZHDR_SIZE_ALIGNED;
 672                 sz = zhdr->first_chunks << CHUNK_SHIFT;
 673                 old_handle = (unsigned long)&zhdr->slots->slot[first_idx];
 674                 moved_chunks = &zhdr->first_chunks;
 675         } else if (zhdr->middle_chunks && zhdr->slots->slot[middle_idx]) {
 676                 p += zhdr->start_middle << CHUNK_SHIFT;
 677                 sz = zhdr->middle_chunks << CHUNK_SHIFT;
 678                 old_handle = (unsigned long)&zhdr->slots->slot[middle_idx];
 679                 moved_chunks = &zhdr->middle_chunks;
 680         } else if (zhdr->last_chunks && zhdr->slots->slot[last_idx]) {
 681                 p += PAGE_SIZE - (zhdr->last_chunks << CHUNK_SHIFT);
 682                 sz = zhdr->last_chunks << CHUNK_SHIFT;
 683                 old_handle = (unsigned long)&zhdr->slots->slot[last_idx];
 684                 moved_chunks = &zhdr->last_chunks;
 685         }
 686
 687         if (sz > 0) {
 688                 enum buddy new_bud = HEADLESS;
 689                 short chunks = size_to_chunks(sz);
 690                 void *q;
 691
 692                 new_zhdr = __z3fold_alloc(pool, sz, false);
 693                 if (!new_zhdr)
 694                         return NULL;
 695
 696                 if (WARN_ON(new_zhdr == zhdr))
 697                         goto out_fail;
 698
 699                 new_bud = get_free_buddy(new_zhdr, chunks);
 700                 q = new_zhdr;
 701                 switch (new_bud) {
 702                 case FIRST:
 703                         new_zhdr->first_chunks = chunks;
 704                         q += ZHDR_SIZE_ALIGNED;
 705                         break;
 706                 case MIDDLE:
 707                         new_zhdr->middle_chunks = chunks;
 708                         new_zhdr->start_middle =
 709                                 new_zhdr->first_chunks + ZHDR_CHUNKS;
 710                         q += new_zhdr->start_middle << CHUNK_SHIFT;
 711                         break;
 712                 case LAST:
 713                         new_zhdr->last_chunks = chunks;
 714                         q += PAGE_SIZE - (new_zhdr->last_chunks << CHUNK_SHIFT);
 715                         break;
 716                 default:
 717                         goto out_fail;
 718                 }
 719                 new_zhdr->foreign_handles++;
 720                 memcpy(q, p, sz);
 721                 write_lock(&zhdr->slots->lock);
 722                 *(unsigned long *)old_handle = (unsigned long)new_zhdr +
 723                         __idx(new_zhdr, new_bud);
 724                 if (new_bud == LAST)
 725                         *(unsigned long *)old_handle |=
 726                                         (new_zhdr->last_chunks << BUDDY_SHIFT);
 727                 write_unlock(&zhdr->slots->lock);
 728                 add_to_unbuddied(pool, new_zhdr);
 729                 z3fold_page_unlock(new_zhdr);
 730
 731                 *moved_chunks = 0;
 732         }
 733
 734         return new_zhdr;
 735
 736 out_fail:
 737         if (new_zhdr && !kref_put(&new_zhdr->refcount, release_z3fold_page_locked)) {
 738                 add_to_unbuddied(pool, new_zhdr);
 739                 z3fold_page_unlock(new_zhdr);
 740         }
 741         return NULL;
 742
 743 }
 744
 745 #define BIG_CHUNK_GAP   3
 746 /* Has to be called with lock held */
 747 static int z3fold_compact_page(struct z3fold_header *zhdr)
 748 {
 749         struct page *page = virt_to_page(zhdr);
 750
 751         if (test_bit(MIDDLE_CHUNK_MAPPED, &page->private))
 752                 return 0; /* can't move middle chunk, it's used */
 753
 754         if (unlikely(PageIsolated(page)))
 755                 return 0;
 756
 757         if (zhdr->middle_chunks == 0)
 758                 return 0; /* nothing to compact */
 759
 760         if (zhdr->first_chunks == 0 && zhdr->last_chunks == 0) {
 761                 /* move to the beginning */
 762                 mchunk_memmove(zhdr, ZHDR_CHUNKS);
 763                 zhdr->first_chunks = zhdr->middle_chunks;
 764                 zhdr->middle_chunks = 0;
 765                 zhdr->start_middle = 0;
 766                 zhdr->first_num++;
 767                 return 1;
 768         }
 769
 770         /*
 771          * moving data is expensive, so let's only do that if
 772          * there's substantial gain (at least BIG_CHUNK_GAP chunks)
 773          */
 774         if (zhdr->first_chunks != 0 && zhdr->last_chunks == 0 &&
 775             zhdr->start_middle - (zhdr->first_chunks + ZHDR_CHUNKS) >=
 776                         BIG_CHUNK_GAP) {
 777                 mchunk_memmove(zhdr, zhdr->first_chunks + ZHDR_CHUNKS);
 778                 zhdr->start_middle = zhdr->first_chunks + ZHDR_CHUNKS;
 779                 return 1;
 780         } else if (zhdr->last_chunks != 0 && zhdr->first_chunks == 0 &&
 781                    TOTAL_CHUNKS - (zhdr->last_chunks + zhdr->start_middle
 782                                         + zhdr->middle_chunks) >=
 783                         BIG_CHUNK_GAP) {
 784                 unsigned short new_start = TOTAL_CHUNKS - zhdr->last_chunks -
 785                         zhdr->middle_chunks;
 786                 mchunk_memmove(zhdr, new_start);
 787                 zhdr->start_middle = new_start;
 788                 return 1;
 789         }
 790
 791         return 0;
 792 }
 793
 794 static void do_compact_page(struct z3fold_header *zhdr, bool locked)
 795 {
 796         struct z3fold_pool *pool = zhdr_to_pool(zhdr);
 797         struct page *page;
 798
 799         page = virt_to_page(zhdr);
 800         if (locked)
 801                 WARN_ON(z3fold_page_trylock(zhdr));
 802         else
 803                 z3fold_page_lock(zhdr);
 804         if (WARN_ON(!test_and_clear_bit(NEEDS_COMPACTING, &page->private))) {
 805                 z3fold_page_unlock(zhdr);
 806                 return;
 807         }
 808         spin_lock(&pool->lock);
 809         list_del_init(&zhdr->buddy);
 810         spin_unlock(&pool->lock);
 811
 812         if (kref_put(&zhdr->refcount, release_z3fold_page_locked))
 813                 return;
 814
 815         if (test_bit(PAGE_STALE, &page->private) ||
 816             test_and_set_bit(PAGE_CLAIMED, &page->private)) {
 817                 z3fold_page_unlock(zhdr);
 818                 return;
 819         }
 820
 821         if (!zhdr->foreign_handles && buddy_single(zhdr) &&
 822             zhdr->mapped_count == 0 && compact_single_buddy(zhdr)) {
 823                 if (!kref_put(&zhdr->refcount, release_z3fold_page_locked)) {
 824                         clear_bit(PAGE_CLAIMED, &page->private);
 825                         z3fold_page_unlock(zhdr);
 826                 }
 827                 return;
 828         }
 829
 830         z3fold_compact_page(zhdr);
 831         add_to_unbuddied(pool, zhdr);
 832         clear_bit(PAGE_CLAIMED, &page->private);
 833         z3fold_page_unlock(zhdr);
 834 }
 835
 836 static void compact_page_work(struct work_struct *w)
 837 {
 838         struct z3fold_header *zhdr = container_of(w, struct z3fold_header,
 839                                                 work);
 840
 841         do_compact_page(zhdr, false);
 842 }
 843
 844 /* returns _locked_ z3fold page header or NULL */
 845 static inline struct z3fold_header *__z3fold_alloc(struct z3fold_pool *pool,
 846                                                 size_t size, bool can_sleep)
 847 {
 848         struct z3fold_header *zhdr = NULL;
 849         struct page *page;
 850         struct list_head *unbuddied;
 851         int chunks = size_to_chunks(size), i;
 852
 853 lookup:
 854         migrate_disable();
 855         /* First, try to find an unbuddied z3fold page. */
 856         unbuddied = this_cpu_ptr(pool->unbuddied);
 857         for_each_unbuddied_list(i, chunks) {
 858                 struct list_head *l = &unbuddied[i];
 859
 860                 zhdr = list_first_entry_or_null(READ_ONCE(l),
 861                                         struct z3fold_header, buddy);
 862
 863                 if (!zhdr)
 864                         continue;
 865
 866                 /* Re-check under lock. */
 867                 spin_lock(&pool->lock);
 868                 if (unlikely(zhdr != list_first_entry(READ_ONCE(l),
 869                                                 struct z3fold_header, buddy)) ||
 870                     !z3fold_page_trylock(zhdr)) {
 871                         spin_unlock(&pool->lock);
 872                         zhdr = NULL;
 873                         migrate_enable();
 874                         if (can_sleep)
 875                                 cond_resched();
 876                         goto lookup;
 877                 }
 878                 list_del_init(&zhdr->buddy);
 879                 zhdr->cpu = -1;
 880                 spin_unlock(&pool->lock);
 881
 882                 page = virt_to_page(zhdr);
 883                 if (test_bit(NEEDS_COMPACTING, &page->private) ||
 884                     test_bit(PAGE_CLAIMED, &page->private)) {
 885                         z3fold_page_unlock(zhdr);
 886                         zhdr = NULL;
 887                         migrate_enable();
 888                         if (can_sleep)
 889                                 cond_resched();
 890                         goto lookup;
 891                 }
 892
 893                 /*
 894                  * this page could not be removed from its unbuddied
 895                  * list while pool lock was held, and then we've taken
 896                  * page lock so kref_put could not be called before
 897                  * we got here, so it's safe to just call kref_get()
 898                  */
 899                 kref_get(&zhdr->refcount);
 900                 break;
 901         }
 902         migrate_enable();
 903
 904         if (!zhdr) {
 905                 int cpu;
 906
 907                 /* look for _exact_ match on other cpus' lists */
 908                 for_each_online_cpu(cpu) {
 909                         struct list_head *l;
 910
 911                         unbuddied = per_cpu_ptr(pool->unbuddied, cpu);
 912                         spin_lock(&pool->lock);
 913                         l = &unbuddied[chunks];
 914
 915                         zhdr = list_first_entry_or_null(READ_ONCE(l),
 916                                                 struct z3fold_header, buddy);
 917
 918                         if (!zhdr || !z3fold_page_trylock(zhdr)) {
 919                                 spin_unlock(&pool->lock);
 920                                 zhdr = NULL;
 921                                 continue;
 922                         }
 923                         list_del_init(&zhdr->buddy);
 924                         zhdr->cpu = -1;
 925                         spin_unlock(&pool->lock);
 926
 927                         page = virt_to_page(zhdr);
 928                         if (test_bit(NEEDS_COMPACTING, &page->private) ||
 929                             test_bit(PAGE_CLAIMED, &page->private)) {
 930                                 z3fold_page_unlock(zhdr);
 931                                 zhdr = NULL;
 932                                 if (can_sleep)
 933                                         cond_resched();
 934                                 continue;
 935                         }
 936                         kref_get(&zhdr->refcount);
 937                         break;
 938                 }
 939         }
 940
 941         if (zhdr && !zhdr->slots) {
 942                 zhdr->slots = alloc_slots(pool, GFP_ATOMIC);
 943                 if (!zhdr->slots)
 944                         goto out_fail;
 945         }
 946         return zhdr;
 947
 948 out_fail:
 949         if (!kref_put(&zhdr->refcount, release_z3fold_page_locked)) {
 950                 add_to_unbuddied(pool, zhdr);
 951                 z3fold_page_unlock(zhdr);
 952         }
 953         return NULL;
 954 }
 955
 956 /*
 957  * API Functions
 958  */
 959
 960 /**
 961  * z3fold_create_pool() - create a new z3fold pool
 962  * @name:       pool name
 963  * @gfp:        gfp flags when allocating the z3fold pool structure
 964  * @ops:        user-defined operations for the z3fold pool
 965  *
 966  * Return: pointer to the new z3fold pool or NULL if the metadata allocation
 967  * failed.
 968  */
 969 static struct z3fold_pool *z3fold_create_pool(const char *name, gfp_t gfp,
 970                 const struct z3fold_ops *ops)
 971 {
 972         struct z3fold_pool *pool = NULL;
 973         int i, cpu;
 974
 975         pool = kzalloc(sizeof(struct z3fold_pool), gfp);
 976         if (!pool)
 977                 goto out;
 978         pool->c_handle = kmem_cache_create("z3fold_handle",
 979                                 sizeof(struct z3fold_buddy_slots),
 980                                 SLOTS_ALIGN, 0, NULL);
 981         if (!pool->c_handle)
 982                 goto out_c;
 983         spin_lock_init(&pool->lock);
 984         spin_lock_init(&pool->stale_lock);
 985         pool->unbuddied = __alloc_percpu(sizeof(struct list_head) * NCHUNKS,
 986                                          __alignof__(struct list_head));
 987         if (!pool->unbuddied)
 988                 goto out_pool;
 989         for_each_possible_cpu(cpu) {
 990                 struct list_head *unbuddied =
 991                                 per_cpu_ptr(pool->unbuddied, cpu);
 992                 for_each_unbuddied_list(i, 0)
 993                         INIT_LIST_HEAD(&unbuddied[i]);
 994         }
 995         INIT_LIST_HEAD(&pool->lru);
 996         INIT_LIST_HEAD(&pool->stale);
 997         atomic64_set(&pool->pages_nr, 0);
 998         pool->name = name;
 999         pool->compact_wq = create_singlethread_workqueue(pool->name);
1000         if (!pool->compact_wq)
1001                 goto out_unbuddied;
1002         pool->release_wq = create_singlethread_workqueue(pool->name);
1003         if (!pool->release_wq)
1004                 goto out_wq;
1005         if (z3fold_register_migration(pool))
1006                 goto out_rwq;
1007         INIT_WORK(&pool->work, free_pages_work);
1008         pool->ops = ops;
1009         return pool;
1010
1011 out_rwq:
1012         destroy_workqueue(pool->release_wq);
1013 out_wq:
1014         destroy_workqueue(pool->compact_wq);
1015 out_unbuddied:
1016         free_percpu(pool->unbuddied);
1017 out_pool:
1018         kmem_cache_destroy(pool->c_handle);
1019 out_c:
1020         kfree(pool);
1021 out:
1022         return NULL;
1023 }
1024
1025 /**
1026  * z3fold_destroy_pool() - destroys an existing z3fold pool
1027  * @pool:       the z3fold pool to be destroyed
1028  *
1029  * The pool should be emptied before this function is called.
1030  */
1031 static void z3fold_destroy_pool(struct z3fold_pool *pool)
1032 {
1033         kmem_cache_destroy(pool->c_handle);
1034
1035         /*
1036          * We need to destroy pool->compact_wq before pool->release_wq,
1037          * as any pending work on pool->compact_wq will call
1038          * queue_work(pool->release_wq, &pool->work).
1039          *
1040          * There are still outstanding pages until both workqueues are drained,
1041          * so we cannot unregister migration until then.
1042          */
1043
1044         destroy_workqueue(pool->compact_wq);
1045         destroy_workqueue(pool->release_wq);
1046         z3fold_unregister_migration(pool);
1047         free_percpu(pool->unbuddied);
1048         kfree(pool);
1049 }
1050
1051 /**
1052  * z3fold_alloc() - allocates a region of a given size
1053  * @pool:       z3fold pool from which to allocate
1054  * @size:       size in bytes of the desired allocation
1055  * @gfp:        gfp flags used if the pool needs to grow
1056  * @handle:     handle of the new allocation
1057  *
1058  * This function will attempt to find a free region in the pool large enough to
1059  * satisfy the allocation request.  A search of the unbuddied lists is
1060  * performed first. If no suitable free region is found, then a new page is
1061  * allocated and added to the pool to satisfy the request.
1062  *
1063  * Return: 0 if success and handle is set, otherwise -EINVAL if the size or
1064  * gfp arguments are invalid or -ENOMEM if the pool was unable to allocate
1065  * a new page.
1066  */
1067 static int z3fold_alloc(struct z3fold_pool *pool, size_t size, gfp_t gfp,
1068                         unsigned long *handle)
1069 {
1070         int chunks = size_to_chunks(size);
1071         struct z3fold_header *zhdr = NULL;
1072         struct page *page = NULL;
1073         enum buddy bud;
1074         bool can_sleep = gfpflags_allow_blocking(gfp);
1075
1076         if (!size || (gfp & __GFP_HIGHMEM))
1077                 return -EINVAL;
1078
1079         if (size > PAGE_SIZE)
1080                 return -ENOSPC;
1081
1082         if (size > PAGE_SIZE - ZHDR_SIZE_ALIGNED - CHUNK_SIZE)
1083                 bud = HEADLESS;
1084         else {
1085 retry:
1086                 zhdr = __z3fold_alloc(pool, size, can_sleep);
1087                 if (zhdr) {
1088                         bud = get_free_buddy(zhdr, chunks);
1089                         if (bud == HEADLESS) {
1090                                 if (!kref_put(&zhdr->refcount,
1091                                              release_z3fold_page_locked))
1092                                         z3fold_page_unlock(zhdr);
1093                                 pr_err("No free chunks in unbuddied\n");
1094                                 WARN_ON(1);
1095                                 goto retry;
1096                         }
1097                         page = virt_to_page(zhdr);
1098                         goto found;
1099                 }
1100                 bud = FIRST;
1101         }
1102
1103         page = alloc_page(gfp);
1104         if (!page)
1105                 return -ENOMEM;
1106
1107         zhdr = init_z3fold_page(page, bud == HEADLESS, pool, gfp);
1108         if (!zhdr) {
1109                 __free_page(page);
1110                 return -ENOMEM;
1111         }
1112         atomic64_inc(&pool->pages_nr);
1113
1114         if (bud == HEADLESS) {
1115                 set_bit(PAGE_HEADLESS, &page->private);
1116                 goto headless;
1117         }
1118         if (can_sleep) {
1119                 lock_page(page);
1120                 __SetPageMovable(page, pool->inode->i_mapping);
1121                 unlock_page(page);
1122         } else {
1123                 WARN_ON(!trylock_page(page));
1124                 __SetPageMovable(page, pool->inode->i_mapping);
1125                 unlock_page(page);
1126         }
1127         z3fold_page_lock(zhdr);
1128
1129 found:
1130         if (bud == FIRST)
1131                 zhdr->first_chunks = chunks;
1132         else if (bud == LAST)
1133                 zhdr->last_chunks = chunks;
1134         else {
1135                 zhdr->middle_chunks = chunks;
1136                 zhdr->start_middle = zhdr->first_chunks + ZHDR_CHUNKS;
1137         }
1138         add_to_unbuddied(pool, zhdr);
1139
1140 headless:
1141         spin_lock(&pool->lock);
1142         /* Add/move z3fold page to beginning of LRU */
1143         if (!list_empty(&page->lru))
1144                 list_del(&page->lru);
1145
1146         list_add(&page->lru, &pool->lru);
1147
1148         *handle = encode_handle(zhdr, bud);
1149         spin_unlock(&pool->lock);
1150         if (bud != HEADLESS)
1151                 z3fold_page_unlock(zhdr);
1152
1153         return 0;
1154 }
1155
1156 /**
1157  * z3fold_free() - frees the allocation associated with the given handle
1158  * @pool:       pool in which the allocation resided
1159  * @handle:     handle associated with the allocation returned by z3fold_alloc()
1160  *
1161  * In the case that the z3fold page in which the allocation resides is under
1162  * reclaim, as indicated by the PAGE_CLAIMED flag being set, this function
1163  * only sets the first|middle|last_chunks to 0.  The page is actually freed
1164  * once all buddies are evicted (see z3fold_reclaim_page() below).
1165  */
1166 static void z3fold_free(struct z3fold_pool *pool, unsigned long handle)
1167 {
1168         struct z3fold_header *zhdr;
1169         struct page *page;
1170         enum buddy bud;
1171         bool page_claimed;
1172
1173         zhdr = get_z3fold_header(handle);
1174         page = virt_to_page(zhdr);
1175         page_claimed = test_and_set_bit(PAGE_CLAIMED, &page->private);
1176
1177         if (test_bit(PAGE_HEADLESS, &page->private)) {
1178                 /* if a headless page is under reclaim, just leave.
1179                  * NB: we use test_and_set_bit for a reason: if the bit
1180                  * has not been set before, we release this page
1181                  * immediately so we don't care about its value any more.
1182                  */
1183                 if (!page_claimed) {
1184                         spin_lock(&pool->lock);
1185                         list_del(&page->lru);
1186                         spin_unlock(&pool->lock);
1187                         put_z3fold_header(zhdr);
1188                         free_z3fold_page(page, true);
1189                         atomic64_dec(&pool->pages_nr);
1190                 }
1191                 return;
1192         }
1193
1194         /* Non-headless case */
1195         bud = handle_to_buddy(handle);
1196
1197         switch (bud) {
1198         case FIRST:
1199                 zhdr->first_chunks = 0;
1200                 break;
1201         case MIDDLE:
1202                 zhdr->middle_chunks = 0;
1203                 break;
1204         case LAST:
1205                 zhdr->last_chunks = 0;
1206                 break;
1207         default:
1208                 pr_err("%s: unknown bud %d\n", __func__, bud);
1209                 WARN_ON(1);
1210                 put_z3fold_header(zhdr);
1211                 return;
1212         }
1213
1214         if (!page_claimed)
1215                 free_handle(handle, zhdr);
1216         if (kref_put(&zhdr->refcount, release_z3fold_page_locked_list))
1217                 return;
1218         if (page_claimed) {
1219                 /* the page has not been claimed by us */
1220                 put_z3fold_header(zhdr);
1221                 return;
1222         }
1223         if (test_and_set_bit(NEEDS_COMPACTING, &page->private)) {
1224                 clear_bit(PAGE_CLAIMED, &page->private);
1225                 put_z3fold_header(zhdr);
1226                 return;
1227         }
1228         if (zhdr->cpu < 0 || !cpu_online(zhdr->cpu)) {
1229                 zhdr->cpu = -1;
1230                 kref_get(&zhdr->refcount);
1231                 clear_bit(PAGE_CLAIMED, &page->private);
1232                 do_compact_page(zhdr, true);
1233                 return;
1234         }
1235         kref_get(&zhdr->refcount);
1236         clear_bit(PAGE_CLAIMED, &page->private);
1237         queue_work_on(zhdr->cpu, pool->compact_wq, &zhdr->work);
1238         put_z3fold_header(zhdr);
1239 }
1240
1241 /**
1242  * z3fold_reclaim_page() - evicts allocations from a pool page and frees it
1243  * @pool:       pool from which a page will attempt to be evicted
1244  * @retries:    number of pages on the LRU list for which eviction will
1245  *              be attempted before failing
1246  *
1247  * z3fold reclaim is different from normal system reclaim in that it is done
1248  * from the bottom, up. This is because only the bottom layer, z3fold, has
1249  * information on how the allocations are organized within each z3fold page.
1250  * This has the potential to create interesting locking situations between
1251  * z3fold and the user, however.
1252  *
1253  * To avoid these, this is how z3fold_reclaim_page() should be called:
1254  *
1255  * The user detects a page should be reclaimed and calls z3fold_reclaim_page().
1256  * z3fold_reclaim_page() will remove a z3fold page from the pool LRU list and
1257  * call the user-defined eviction handler with the pool and handle as
1258  * arguments.
1259  *
1260  * If the handle can not be evicted, the eviction handler should return
1261  * non-zero. z3fold_reclaim_page() will add the z3fold page back to the
1262  * appropriate list and try the next z3fold page on the LRU up to
1263  * a user defined number of retries.
1264  *
1265  * If the handle is successfully evicted, the eviction handler should
1266  * return 0 _and_ should have called z3fold_free() on the handle. z3fold_free()
1267  * contains logic to delay freeing the page if the page is under reclaim,
1268  * as indicated by the setting of the PG_reclaim flag on the underlying page.
1269  *
1270  * If all buddies in the z3fold page are successfully evicted, then the
1271  * z3fold page can be freed.
1272  *
1273  * Returns: 0 if page is successfully freed, otherwise -EINVAL if there are
1274  * no pages to evict or an eviction handler is not registered, -EAGAIN if
1275  * the retry limit was hit.
1276  */
1277 static int z3fold_reclaim_page(struct z3fold_pool *pool, unsigned int retries)
1278 {
1279         int i, ret = -1;
1280         struct z3fold_header *zhdr = NULL;
1281         struct page *page = NULL;
1282         struct list_head *pos;
1283         unsigned long first_handle = 0, middle_handle = 0, last_handle = 0;
1284         struct z3fold_buddy_slots slots __attribute__((aligned(SLOTS_ALIGN)));
1285
1286         rwlock_init(&slots.lock);
1287         slots.pool = (unsigned long)pool | (1 << HANDLES_NOFREE);
1288
1289         spin_lock(&pool->lock);
1290         if (!pool->ops || !pool->ops->evict || retries == 0) {
1291                 spin_unlock(&pool->lock);
1292                 return -EINVAL;
1293         }
1294         for (i = 0; i < retries; i++) {
1295                 if (list_empty(&pool->lru)) {
1296                         spin_unlock(&pool->lock);
1297                         return -EINVAL;
1298                 }
1299                 list_for_each_prev(pos, &pool->lru) {
1300                         page = list_entry(pos, struct page, lru);
1301
1302                         zhdr = page_address(page);
1303                         if (test_bit(PAGE_HEADLESS, &page->private)) {
1304                                 /*
1305                                  * For non-headless pages, we wait to do this
1306                                  * until we have the page lock to avoid racing
1307                                  * with __z3fold_alloc(). Headless pages don't
1308                                  * have a lock (and __z3fold_alloc() will never
1309                                  * see them), but we still need to test and set
1310                                  * PAGE_CLAIMED to avoid racing with
1311                                  * z3fold_free(), so just do it now before
1312                                  * leaving the loop.
1313                                  */
1314                                 if (test_and_set_bit(PAGE_CLAIMED, &page->private))
1315                                         continue;
1316
1317                                 break;
1318                         }
1319
1320                         if (!z3fold_page_trylock(zhdr)) {
1321                                 zhdr = NULL;
1322                                 continue; /* can't evict at this point */
1323                         }
1324
1325                         /* test_and_set_bit is of course atomic, but we still
1326                          * need to do it under page lock, otherwise checking
1327                          * that bit in __z3fold_alloc wouldn't make sense
1328                          */
1329                         if (zhdr->foreign_handles ||
1330                             test_and_set_bit(PAGE_CLAIMED, &page->private)) {
1331                                 z3fold_page_unlock(zhdr);
1332                                 zhdr = NULL;
1333                                 continue; /* can't evict such page */
1334                         }
1335                         list_del_init(&zhdr->buddy);
1336                         zhdr->cpu = -1;
1337                         /* See comment in __z3fold_alloc. */
1338                         kref_get(&zhdr->refcount);
1339                         break;
1340                 }
1341
1342                 if (!zhdr)
1343                         break;
1344
1345                 list_del_init(&page->lru);
1346                 spin_unlock(&pool->lock);
1347
1348                 if (!test_bit(PAGE_HEADLESS, &page->private)) {
1349                         /*
1350                          * We need encode the handles before unlocking, and
1351                          * use our local slots structure because z3fold_free
1352                          * can zero out zhdr->slots and we can't do much
1353                          * about that
1354                          */
1355                         first_handle = 0;
1356                         last_handle = 0;
1357                         middle_handle = 0;
1358                         memset(slots.slot, 0, sizeof(slots.slot));
1359                         if (zhdr->first_chunks)
1360                                 first_handle = __encode_handle(zhdr, &slots,
1361                                                                 FIRST);
1362                         if (zhdr->middle_chunks)
1363                                 middle_handle = __encode_handle(zhdr, &slots,
1364                                                                 MIDDLE);
1365                         if (zhdr->last_chunks)
1366                                 last_handle = __encode_handle(zhdr, &slots,
1367                                                                 LAST);
1368                         /*
1369                          * it's safe to unlock here because we hold a
1370                          * reference to this page
1371                          */
1372                         z3fold_page_unlock(zhdr);
1373                 } else {
1374                         first_handle = encode_handle(zhdr, HEADLESS);
1375                         last_handle = middle_handle = 0;
1376                 }
1377                 /* Issue the eviction callback(s) */
1378                 if (middle_handle) {
1379                         ret = pool->ops->evict(pool, middle_handle);
1380                         if (ret)
1381                                 goto next;
1382                 }
1383                 if (first_handle) {
1384                         ret = pool->ops->evict(pool, first_handle);
1385                         if (ret)
1386                                 goto next;
1387                 }
1388                 if (last_handle) {
1389                         ret = pool->ops->evict(pool, last_handle);
1390                         if (ret)
1391                                 goto next;
1392                 }
1393 next:
1394                 if (test_bit(PAGE_HEADLESS, &page->private)) {
1395                         if (ret == 0) {
1396                                 free_z3fold_page(page, true);
1397                                 atomic64_dec(&pool->pages_nr);
1398                                 return 0;
1399                         }
1400                         spin_lock(&pool->lock);
1401                         list_add(&page->lru, &pool->lru);
1402                         spin_unlock(&pool->lock);
1403                         clear_bit(PAGE_CLAIMED, &page->private);
1404                 } else {
1405                         struct z3fold_buddy_slots *slots = zhdr->slots;
1406                         z3fold_page_lock(zhdr);
1407                         if (kref_put(&zhdr->refcount,
1408                                         release_z3fold_page_locked)) {
1409                                 kmem_cache_free(pool->c_handle, slots);
1410                                 return 0;
1411                         }
1412                         /*
1413                          * if we are here, the page is still not completely
1414                          * free. Take the global pool lock then to be able
1415                          * to add it back to the lru list
1416                          */
1417                         spin_lock(&pool->lock);
1418                         list_add(&page->lru, &pool->lru);
1419                         spin_unlock(&pool->lock);
1420                         if (list_empty(&zhdr->buddy))
1421                                 add_to_unbuddied(pool, zhdr);
1422                         clear_bit(PAGE_CLAIMED, &page->private);
1423                         z3fold_page_unlock(zhdr);
1424                 }
1425
1426                 /* We started off locked to we need to lock the pool back */
1427                 spin_lock(&pool->lock);
1428         }
1429         spin_unlock(&pool->lock);
1430         return -EAGAIN;
1431 }
1432
1433 /**
1434  * z3fold_map() - maps the allocation associated with the given handle
1435  * @pool:       pool in which the allocation resides
1436  * @handle:     handle associated with the allocation to be mapped
1437  *
1438  * Extracts the buddy number from handle and constructs the pointer to the
1439  * correct starting chunk within the page.
1440  *
1441  * Returns: a pointer to the mapped allocation
1442  */
1443 static void *z3fold_map(struct z3fold_pool *pool, unsigned long handle)
1444 {
1445         struct z3fold_header *zhdr;
1446         struct page *page;
1447         void *addr;
1448         enum buddy buddy;
1449
1450         zhdr = get_z3fold_header(handle);
1451         addr = zhdr;
1452         page = virt_to_page(zhdr);
1453
1454         if (test_bit(PAGE_HEADLESS, &page->private))
1455                 goto out;
1456
1457         buddy = handle_to_buddy(handle);
1458         switch (buddy) {
1459         case FIRST:
1460                 addr += ZHDR_SIZE_ALIGNED;
1461                 break;
1462         case MIDDLE:
1463                 addr += zhdr->start_middle << CHUNK_SHIFT;
1464                 set_bit(MIDDLE_CHUNK_MAPPED, &page->private);
1465                 break;
1466         case LAST:
1467                 addr += PAGE_SIZE - (handle_to_chunks(handle) << CHUNK_SHIFT);
1468                 break;
1469         default:
1470                 pr_err("unknown buddy id %d\n", buddy);
1471                 WARN_ON(1);
1472                 addr = NULL;
1473                 break;
1474         }
1475
1476         if (addr)
1477                 zhdr->mapped_count++;
1478 out:
1479         put_z3fold_header(zhdr);
1480         return addr;
1481 }
1482
1483 /**
1484  * z3fold_unmap() - unmaps the allocation associated with the given handle
1485  * @pool:       pool in which the allocation resides
1486  * @handle:     handle associated with the allocation to be unmapped
1487  */
1488 static void z3fold_unmap(struct z3fold_pool *pool, unsigned long handle)
1489 {
1490         struct z3fold_header *zhdr;
1491         struct page *page;
1492         enum buddy buddy;
1493
1494         zhdr = get_z3fold_header(handle);
1495         page = virt_to_page(zhdr);
1496
1497         if (test_bit(PAGE_HEADLESS, &page->private))
1498                 return;
1499
1500         buddy = handle_to_buddy(handle);
1501         if (buddy == MIDDLE)
1502                 clear_bit(MIDDLE_CHUNK_MAPPED, &page->private);
1503         zhdr->mapped_count--;
1504         put_z3fold_header(zhdr);
1505 }
1506
1507 /**
1508  * z3fold_get_pool_size() - gets the z3fold pool size in pages
1509  * @pool:       pool whose size is being queried
1510  *
1511  * Returns: size in pages of the given pool.
1512  */
1513 static u64 z3fold_get_pool_size(struct z3fold_pool *pool)
1514 {
1515         return atomic64_read(&pool->pages_nr);
1516 }
1517
1518 static bool z3fold_page_isolate(struct page *page, isolate_mode_t mode)
1519 {
1520         struct z3fold_header *zhdr;
1521         struct z3fold_pool *pool;
1522
1523         VM_BUG_ON_PAGE(!PageMovable(page), page);
1524         VM_BUG_ON_PAGE(PageIsolated(page), page);
1525
1526         if (test_bit(PAGE_HEADLESS, &page->private))
1527                 return false;
1528
1529         zhdr = page_address(page);
1530         z3fold_page_lock(zhdr);
1531         if (test_bit(NEEDS_COMPACTING, &page->private) ||
1532             test_bit(PAGE_STALE, &page->private))
1533                 goto out;
1534
1535         if (zhdr->mapped_count != 0 || zhdr->foreign_handles != 0)
1536                 goto out;
1537
1538         if (test_and_set_bit(PAGE_CLAIMED, &page->private))
1539                 goto out;
1540         pool = zhdr_to_pool(zhdr);
1541         spin_lock(&pool->lock);
1542         if (!list_empty(&zhdr->buddy))
1543                 list_del_init(&zhdr->buddy);
1544         if (!list_empty(&page->lru))
1545                 list_del_init(&page->lru);
1546         spin_unlock(&pool->lock);
1547
1548         kref_get(&zhdr->refcount);
1549         z3fold_page_unlock(zhdr);
1550         return true;
1551
1552 out:
1553         z3fold_page_unlock(zhdr);
1554         return false;
1555 }
1556
1557 static int z3fold_page_migrate(struct address_space *mapping, struct page *newpage,
1558                                struct page *page, enum migrate_mode mode)
1559 {
1560         struct z3fold_header *zhdr, *new_zhdr;
1561         struct z3fold_pool *pool;
1562         struct address_space *new_mapping;
1563
1564         VM_BUG_ON_PAGE(!PageMovable(page), page);
1565         VM_BUG_ON_PAGE(!PageIsolated(page), page);
1566         VM_BUG_ON_PAGE(!test_bit(PAGE_CLAIMED, &page->private), page);
1567         VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
1568
1569         zhdr = page_address(page);
1570         pool = zhdr_to_pool(zhdr);
1571
1572         if (!z3fold_page_trylock(zhdr))
1573                 return -EAGAIN;
1574         if (zhdr->mapped_count != 0 || zhdr->foreign_handles != 0) {
1575                 clear_bit(PAGE_CLAIMED, &page->private);
1576                 z3fold_page_unlock(zhdr);
1577                 return -EBUSY;
1578         }
1579         if (work_pending(&zhdr->work)) {
1580                 z3fold_page_unlock(zhdr);
1581                 return -EAGAIN;
1582         }
1583         new_zhdr = page_address(newpage);
1584         memcpy(new_zhdr, zhdr, PAGE_SIZE);
1585         newpage->private = page->private;
1586         set_bit(PAGE_MIGRATED, &page->private);
1587         z3fold_page_unlock(zhdr);
1588         spin_lock_init(&new_zhdr->page_lock);
1589         INIT_WORK(&new_zhdr->work, compact_page_work);
1590         /*
1591          * z3fold_page_isolate() ensures that new_zhdr->buddy is empty,
1592          * so we only have to reinitialize it.
1593          */
1594         INIT_LIST_HEAD(&new_zhdr->buddy);
1595         new_mapping = page_mapping(page);
1596         __ClearPageMovable(page);
1597
1598         get_page(newpage);
1599         z3fold_page_lock(new_zhdr);
1600         if (new_zhdr->first_chunks)
1601                 encode_handle(new_zhdr, FIRST);
1602         if (new_zhdr->last_chunks)
1603                 encode_handle(new_zhdr, LAST);
1604         if (new_zhdr->middle_chunks)
1605                 encode_handle(new_zhdr, MIDDLE);
1606         set_bit(NEEDS_COMPACTING, &newpage->private);
1607         new_zhdr->cpu = smp_processor_id();
1608         spin_lock(&pool->lock);
1609         list_add(&newpage->lru, &pool->lru);
1610         spin_unlock(&pool->lock);
1611         __SetPageMovable(newpage, new_mapping);
1612         z3fold_page_unlock(new_zhdr);
1613
1614         queue_work_on(new_zhdr->cpu, pool->compact_wq, &new_zhdr->work);
1615
1616         /* PAGE_CLAIMED and PAGE_MIGRATED are cleared now. */
1617         page->private = 0;
1618         put_page(page);
1619         return 0;
1620 }
1621
1622 static void z3fold_page_putback(struct page *page)
1623 {
1624         struct z3fold_header *zhdr;
1625         struct z3fold_pool *pool;
1626
1627         zhdr = page_address(page);
1628         pool = zhdr_to_pool(zhdr);
1629
1630         z3fold_page_lock(zhdr);
1631         if (!list_empty(&zhdr->buddy))
1632                 list_del_init(&zhdr->buddy);
1633         INIT_LIST_HEAD(&page->lru);
1634         if (kref_put(&zhdr->refcount, release_z3fold_page_locked))
1635                 return;
1636         spin_lock(&pool->lock);
1637         list_add(&page->lru, &pool->lru);
1638         spin_unlock(&pool->lock);
1639         if (list_empty(&zhdr->buddy))
1640                 add_to_unbuddied(pool, zhdr);
1641         clear_bit(PAGE_CLAIMED, &page->private);
1642         z3fold_page_unlock(zhdr);
1643 }
1644
1645 static const struct address_space_operations z3fold_aops = {
1646         .isolate_page = z3fold_page_isolate,
1647         .migratepage = z3fold_page_migrate,
1648         .putback_page = z3fold_page_putback,
1649 };
1650
1651 /*****************
1652  * zpool
1653  ****************/
1654
1655 static int z3fold_zpool_evict(struct z3fold_pool *pool, unsigned long handle)
1656 {
1657         if (pool->zpool && pool->zpool_ops && pool->zpool_ops->evict)
1658                 return pool->zpool_ops->evict(pool->zpool, handle);
1659         else
1660                 return -ENOENT;
1661 }
1662
1663 static const struct z3fold_ops z3fold_zpool_ops = {
1664         .evict =        z3fold_zpool_evict
1665 };
1666
1667 static void *z3fold_zpool_create(const char *name, gfp_t gfp,
1668                                const struct zpool_ops *zpool_ops,
1669                                struct zpool *zpool)
1670 {
1671         struct z3fold_pool *pool;
1672
1673         pool = z3fold_create_pool(name, gfp,
1674                                 zpool_ops ? &z3fold_zpool_ops : NULL);
1675         if (pool) {
1676                 pool->zpool = zpool;
1677                 pool->zpool_ops = zpool_ops;
1678         }
1679         return pool;
1680 }
1681
1682 static void z3fold_zpool_destroy(void *pool)
1683 {
1684         z3fold_destroy_pool(pool);
1685 }
1686
1687 static int z3fold_zpool_malloc(void *pool, size_t size, gfp_t gfp,
1688                         unsigned long *handle)
1689 {
1690         return z3fold_alloc(pool, size, gfp, handle);
1691 }
1692 static void z3fold_zpool_free(void *pool, unsigned long handle)
1693 {
1694         z3fold_free(pool, handle);
1695 }
1696
1697 static int z3fold_zpool_shrink(void *pool, unsigned int pages,
1698                         unsigned int *reclaimed)
1699 {
1700         unsigned int total = 0;
1701         int ret = -EINVAL;
1702
1703         while (total < pages) {
1704                 ret = z3fold_reclaim_page(pool, 8);
1705                 if (ret < 0)
1706                         break;
1707                 total++;
1708         }
1709
1710         if (reclaimed)
1711                 *reclaimed = total;
1712
1713         return ret;
1714 }
1715
1716 static void *z3fold_zpool_map(void *pool, unsigned long handle,
1717                         enum zpool_mapmode mm)
1718 {
1719         return z3fold_map(pool, handle);
1720 }
1721 static void z3fold_zpool_unmap(void *pool, unsigned long handle)
1722 {
1723         z3fold_unmap(pool, handle);
1724 }
1725
1726 static u64 z3fold_zpool_total_size(void *pool)
1727 {
1728         return z3fold_get_pool_size(pool) * PAGE_SIZE;
1729 }
1730
1731 static struct zpool_driver z3fold_zpool_driver = {
1732         .type =         "z3fold",
1733         .sleep_mapped = true,
1734         .owner =        THIS_MODULE,
1735         .create =       z3fold_zpool_create,
1736         .destroy =      z3fold_zpool_destroy,
1737         .malloc =       z3fold_zpool_malloc,
1738         .free =         z3fold_zpool_free,
1739         .shrink =       z3fold_zpool_shrink,
1740         .map =          z3fold_zpool_map,
1741         .unmap =        z3fold_zpool_unmap,
1742         .total_size =   z3fold_zpool_total_size,
1743 };
1744
1745 MODULE_ALIAS("zpool-z3fold");
1746
1747 static int __init init_z3fold(void)
1748 {
1749         int ret;
1750
1751         /*
1752          * Make sure the z3fold header is not larger than the page size and
1753          * there has remaining spaces for its buddy.
1754          */
1755         BUILD_BUG_ON(ZHDR_SIZE_ALIGNED > PAGE_SIZE - CHUNK_SIZE);
1756         ret = z3fold_mount();
1757         if (ret)
1758                 return ret;
1759
1760         zpool_register_driver(&z3fold_zpool_driver);
1761
1762         return 0;
1763 }
1764
1765 static void __exit exit_z3fold(void)
1766 {
1767         z3fold_unmount();
1768         zpool_unregister_driver(&z3fold_zpool_driver);
1769 }
1770
1771 module_init(init_z3fold);
1772 module_exit(exit_z3fold);
1773
1774 MODULE_LICENSE("GPL");
1775 MODULE_AUTHOR("Vitaly Wool <vitalywool@gmail.com>");
1776 MODULE_DESCRIPTION("3-Fold Allocator for Compressed Pages");