sys/vfs/hammer/hammer_flusher.c

   1 /*
   2  * Copyright (c) 2008 The DragonFly Project.  All rights reserved.
   3  *
   4  * This code is derived from software contributed to The DragonFly Project
   5  * by Matthew Dillon <dillon@backplane.com>
   6  *
   7  * Redistribution and use in source and binary forms, with or without
   8  * modification, are permitted provided that the following conditions
   9  * are met:
  10  *
  11  * 1. Redistributions of source code must retain the above copyright
  12  *    notice, this list of conditions and the following disclaimer.
  13  * 2. Redistributions in binary form must reproduce the above copyright
  14  *    notice, this list of conditions and the following disclaimer in
  15  *    the documentation and/or other materials provided with the
  16  *    distribution.
  17  * 3. Neither the name of The DragonFly Project nor the names of its
  18  *    contributors may be used to endorse or promote products derived
  19  *    from this software without specific, prior written permission.
  20  *
  21  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  24  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
  25  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  26  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
  27  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  28  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  29  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
  30  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
  31  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  32  * SUCH DAMAGE.
  33  *
  34  * $DragonFly: src/sys/vfs/hammer/hammer_flusher.c,v 1.27 2008/06/14 01:42:13 dillon Exp $
  35  */
  36 /*
  37  * HAMMER dependancy flusher thread
  38  *
  39  * Meta data updates create buffer dependancies which are arranged as a
  40  * hierarchy of lists.
  41  */
  42
  43 #include "hammer.h"
  44
  45 static void hammer_flusher_master_thread(void *arg);
  46 static void hammer_flusher_slave_thread(void *arg);
  47 static void hammer_flusher_clean_loose_ios(hammer_mount_t hmp);
  48 static void hammer_flusher_flush(hammer_mount_t hmp);
  49 static void hammer_flusher_flush_inode(hammer_inode_t ip,
  50                                         hammer_transaction_t trans);
  51 static int hammer_must_finalize_undo(hammer_mount_t hmp);
  52 static void hammer_flusher_finalize(hammer_transaction_t trans, int final);
  53
  54 /*
  55  * Support structures for the flusher threads.
  56  */
  57 struct hammer_flusher_info {
  58         struct hammer_mount *hmp;
  59         thread_t        td;
  60         int             startit;
  61         TAILQ_HEAD(,hammer_inode) work_list;
  62 };
  63
  64 typedef struct hammer_flusher_info *hammer_flusher_info_t;
  65
  66 /*
  67  * Sync all inodes pending on the flusher.  This routine may have to be
  68  * called twice to get them all as some may be queued to a later flush group.
  69  */
  70 void
  71 hammer_flusher_sync(hammer_mount_t hmp)
  72 {
  73         int seq;
  74
  75         if (hmp->flusher.td) {
  76                 seq = hmp->flusher.next;
  77                 if (hmp->flusher.signal++ == 0)
  78                         wakeup(&hmp->flusher.signal);
  79                 while ((int)(seq - hmp->flusher.done) > 0)
  80                         tsleep(&hmp->flusher.done, 0, "hmrfls", 0);
  81         }
  82 }
  83
  84 /*
  85  * Sync all inodes pending on the flusher - return immediately.
  86  */
  87 void
  88 hammer_flusher_async(hammer_mount_t hmp)
  89 {
  90         if (hmp->flusher.td) {
  91                 if (hmp->flusher.signal++ == 0)
  92                         wakeup(&hmp->flusher.signal);
  93         }
  94 }
  95
  96 void
  97 hammer_flusher_create(hammer_mount_t hmp)
  98 {
  99         hammer_flusher_info_t info;
 100         int i;
 101
 102         hmp->flusher.signal = 0;
 103         hmp->flusher.act = 0;
 104         hmp->flusher.done = 0;
 105         hmp->flusher.next = 1;
 106         hmp->flusher.count = 0;
 107         hammer_ref(&hmp->flusher.finalize_lock);
 108
 109         lwkt_create(hammer_flusher_master_thread, hmp,
 110                     &hmp->flusher.td, NULL, 0, -1, "hammer-M");
 111         for (i = 0; i < HAMMER_MAX_FLUSHERS; ++i) {
 112                 info = kmalloc(sizeof(*info), M_HAMMER, M_WAITOK|M_ZERO);
 113                 info->hmp = hmp;
 114                 TAILQ_INIT(&info->work_list);
 115                 ++hmp->flusher.count;
 116                 hmp->flusher.info[i] = info;
 117                 lwkt_create(hammer_flusher_slave_thread, info,
 118                             &info->td, NULL, 0, -1, "hammer-S%d", i);
 119         }
 120 }
 121
 122 void
 123 hammer_flusher_destroy(hammer_mount_t hmp)
 124 {
 125         hammer_flusher_info_t info;
 126         int i;
 127
 128         /*
 129          * Kill the master
 130          */
 131         hmp->flusher.exiting = 1;
 132         while (hmp->flusher.td) {
 133                 ++hmp->flusher.signal;
 134                 wakeup(&hmp->flusher.signal);
 135                 tsleep(&hmp->flusher.exiting, 0, "hmrwex", hz);
 136         }
 137
 138         /*
 139          * Kill the slaves
 140          */
 141         for (i = 0; i < HAMMER_MAX_FLUSHERS; ++i) {
 142                 if ((info = hmp->flusher.info[i]) != NULL) {
 143                         KKASSERT(info->startit == 0);
 144                         info->startit = -1;
 145                         wakeup(&info->startit);
 146                         while (info->td) {
 147                                 tsleep(&info->td, 0, "hmrwwc", 0);
 148                         }
 149                         hmp->flusher.info[i] = NULL;
 150                         kfree(info, M_HAMMER);
 151                         --hmp->flusher.count;
 152                 }
 153         }
 154         KKASSERT(hmp->flusher.count == 0);
 155 }
 156
 157 /*
 158  * The master flusher thread manages the flusher sequence id and
 159  * synchronization with the slave work threads.
 160  */
 161 static void
 162 hammer_flusher_master_thread(void *arg)
 163 {
 164         hammer_mount_t hmp = arg;
 165
 166         for (;;) {
 167                 while (hmp->flusher.group_lock)
 168                         tsleep(&hmp->flusher.group_lock, 0, "hmrhld", 0);
 169                 kprintf("S");
 170                 hmp->flusher.act = hmp->flusher.next;
 171                 ++hmp->flusher.next;
 172                 hammer_flusher_clean_loose_ios(hmp);
 173                 hammer_flusher_flush(hmp);
 174                 hmp->flusher.done = hmp->flusher.act;
 175                 wakeup(&hmp->flusher.done);
 176
 177                 /*
 178                  * Wait for activity.
 179                  */
 180                 if (hmp->flusher.exiting && TAILQ_EMPTY(&hmp->flush_list))
 181                         break;
 182
 183                 /*
 184                  * This is a hack until we can dispose of frontend buffer
 185                  * cache buffers on the frontend.
 186                  */
 187                 while (hmp->flusher.signal == 0)
 188                         tsleep(&hmp->flusher.signal, 0, "hmrwwa", 0);
 189                 hmp->flusher.signal = 0;
 190         }
 191
 192         /*
 193          * And we are done.
 194          */
 195         hmp->flusher.td = NULL;
 196         wakeup(&hmp->flusher.exiting);
 197         lwkt_exit();
 198 }
 199
 200 /*
 201  * The slave flusher thread pulls work off the master flush_list until no
 202  * work is left.
 203  */
 204 static void
 205 hammer_flusher_slave_thread(void *arg)
 206 {
 207         hammer_flusher_info_t info;
 208         hammer_mount_t hmp;
 209         hammer_inode_t ip;
 210
 211         info = arg;
 212         hmp = info->hmp;
 213
 214         for (;;) {
 215                 while (info->startit == 0)
 216                         tsleep(&info->startit, 0, "hmrssw", 0);
 217                 if (info->startit < 0)
 218                         break;
 219                 info->startit = 0;
 220                 while ((ip = TAILQ_FIRST(&hmp->flush_list)) != NULL) {
 221                         if (ip->flush_group != hmp->flusher.act)
 222                                 break;
 223                         TAILQ_REMOVE(&hmp->flush_list, ip, flush_entry);
 224                         hammer_flusher_flush_inode(ip, &hmp->flusher.trans);
 225                 }
 226                 if (--hmp->flusher.running == 0)
 227                         wakeup(&hmp->flusher.running);
 228         }
 229         info->td = NULL;
 230         wakeup(&info->td);
 231         lwkt_exit();
 232 }
 233
 234 static void
 235 hammer_flusher_clean_loose_ios(hammer_mount_t hmp)
 236 {
 237         hammer_buffer_t buffer;
 238         hammer_io_t io;
 239         int panic_count = 1000000;
 240
 241         /*
 242          * loose ends - buffers without bp's aren't tracked by the kernel
 243          * and can build up, so clean them out.  This can occur when an
 244          * IO completes on a buffer with no references left.
 245          */
 246         crit_enter();   /* biodone() race */
 247         while ((io = TAILQ_FIRST(&hmp->lose_list)) != NULL) {
 248                 KKASSERT(--panic_count > 0);
 249                 KKASSERT(io->mod_list == &hmp->lose_list);
 250                 TAILQ_REMOVE(&hmp->lose_list, io, mod_entry);
 251                 io->mod_list = NULL;
 252                 if (io->lock.refs == 0)
 253                         ++hammer_count_refedbufs;
 254                 hammer_ref(&io->lock);
 255                 buffer = (void *)io;
 256                 hammer_rel_buffer(buffer, 0);
 257         }
 258         crit_exit();
 259 }
 260
 261 /*
 262  * Flush all inodes in the current flush group.
 263  */
 264 static void
 265 hammer_flusher_flush(hammer_mount_t hmp)
 266 {
 267         hammer_flusher_info_t info;
 268         hammer_reserve_t resv;
 269         int i;
 270         int n;
 271
 272         hammer_start_transaction_fls(&hmp->flusher.trans, hmp);
 273
 274         /*
 275          * Start work threads.
 276          */
 277         i = 0;
 278         n = hmp->count_iqueued / 64;
 279         if (TAILQ_FIRST(&hmp->flush_list)) {
 280                 for (i = 0; i <= hmp->count_iqueued / 64; ++i) {
 281                         if (i == HAMMER_MAX_FLUSHERS ||
 282                             hmp->flusher.info[i] == NULL) {
 283                                 break;
 284                         }
 285                         info = hmp->flusher.info[i];
 286                         if (info->startit == 0) {
 287                                 ++hmp->flusher.running;
 288                                 info->startit = 1;
 289                                 wakeup(&info->startit);
 290                         }
 291                 }
 292         }
 293         while (hmp->flusher.running)
 294                 tsleep(&hmp->flusher.running, 0, "hmrfcc", 0);
 295
 296         hammer_flusher_finalize(&hmp->flusher.trans, 1);
 297         hmp->flusher.tid = hmp->flusher.trans.tid;
 298
 299         /*
 300          * Clean up any freed big-blocks (typically zone-2).
 301          * resv->flush_group is typically set several flush groups ahead
 302          * of the free to ensure that the freed block is not reused until
 303          * it can no longer be reused.
 304          */
 305         while ((resv = TAILQ_FIRST(&hmp->delay_list)) != NULL) {
 306                 if (resv->flush_group != hmp->flusher.act)
 307                         break;
 308                 TAILQ_REMOVE(&hmp->delay_list, resv, delay_entry);
 309                 hammer_blockmap_reserve_complete(hmp, resv);
 310         }
 311         hammer_done_transaction(&hmp->flusher.trans);
 312 }
 313
 314 /*
 315  * Flush a single inode that is part of a flush group.
 316  */
 317 static
 318 void
 319 hammer_flusher_flush_inode(hammer_inode_t ip, hammer_transaction_t trans)
 320 {
 321         hammer_mount_t hmp = ip->hmp;
 322
 323         hammer_lock_sh(&hmp->flusher.finalize_lock);
 324         ip->error = hammer_sync_inode(ip);
 325         hammer_flush_inode_done(ip);
 326         hammer_unlock(&hmp->flusher.finalize_lock);
 327         while (hmp->flusher.finalize_want)
 328                 tsleep(&hmp->flusher.finalize_want, 0, "hmrsxx", 0);
 329         if (hammer_must_finalize_undo(hmp)) {
 330                 hmp->flusher.finalize_want = 1;
 331                 hammer_lock_ex(&hmp->flusher.finalize_lock);
 332                 kprintf("HAMMER: Warning: UNDO area too small!");
 333                 hammer_flusher_finalize(trans, 1);
 334                 hammer_unlock(&hmp->flusher.finalize_lock);
 335                 hmp->flusher.finalize_want = 0;
 336                 wakeup(&hmp->flusher.finalize_want);
 337         } else if (trans->hmp->locked_dirty_count +
 338                    trans->hmp->io_running_count > hammer_limit_dirtybufs) {
 339                 hmp->flusher.finalize_want = 1;
 340                 hammer_lock_ex(&hmp->flusher.finalize_lock);
 341                 kprintf("t");
 342                 hammer_flusher_finalize(trans, 0);
 343                 hammer_unlock(&hmp->flusher.finalize_lock);
 344                 hmp->flusher.finalize_want = 0;
 345                 wakeup(&hmp->flusher.finalize_want);
 346         }
 347 }
 348
 349 /*
 350  * If the UNDO area gets over half full we have to flush it.  We can't
 351  * afford the UNDO area becoming completely full as that would break
 352  * the crash recovery atomicy.
 353  */
 354 static
 355 int
 356 hammer_must_finalize_undo(hammer_mount_t hmp)
 357 {
 358         if (hammer_undo_space(hmp) < hammer_undo_max(hmp) / 2) {
 359                 hkprintf("*");
 360                 return(1);
 361         } else {
 362                 return(0);
 363         }
 364 }
 365
 366 /*
 367  * Flush all pending UNDOs, wait for write completion, update the volume
 368  * header with the new UNDO end position, and flush it.  Then
 369  * asynchronously flush the meta-data.
 370  *
 371  * If this is the last finalization in a flush group we also synchronize
 372  * our cached blockmap and set hmp->flusher_undo_start and our cached undo
 373  * fifo first_offset so the next flush resets the FIFO pointers.
 374  */
 375 static
 376 void
 377 hammer_flusher_finalize(hammer_transaction_t trans, int final)
 378 {
 379         hammer_volume_t root_volume;
 380         hammer_blockmap_t cundomap, dundomap;
 381         hammer_mount_t hmp;
 382         hammer_io_t io;
 383         int count;
 384         int i;
 385
 386         hmp = trans->hmp;
 387         root_volume = trans->rootvol;
 388
 389         /*
 390          * Flush data buffers.  This can occur asynchronously and at any
 391          * time.  We must interlock against the frontend direct-data write
 392          * but do not have to acquire the sync-lock yet.
 393          */
 394         count = 0;
 395         while ((io = TAILQ_FIRST(&hmp->data_list)) != NULL) {
 396                 if (io->lock.refs == 0)
 397                         ++hammer_count_refedbufs;
 398                 hammer_ref(&io->lock);
 399                 hammer_io_write_interlock(io);
 400                 KKASSERT(io->type != HAMMER_STRUCTURE_VOLUME);
 401                 hammer_io_flush(io);
 402                 hammer_io_done_interlock(io);
 403                 hammer_rel_buffer((hammer_buffer_t)io, 0);
 404                 ++count;
 405         }
 406
 407         /*
 408          * The sync-lock is required for the remaining sequence.  This lock
 409          * prevents meta-data from being modified.
 410          */
 411         hammer_sync_lock_ex(trans);
 412
 413         /*
 414          * If we have been asked to finalize the volume header sync the
 415          * cached blockmap to the on-disk blockmap.  Generate an UNDO
 416          * record for the update.
 417          */
 418         if (final) {
 419                 cundomap = &hmp->blockmap[0];
 420                 dundomap = &root_volume->ondisk->vol0_blockmap[0];
 421                 if (root_volume->io.modified) {
 422                         hammer_modify_volume(trans, root_volume,
 423                                              dundomap, sizeof(hmp->blockmap));
 424                         for (i = 0; i < HAMMER_MAX_ZONES; ++i)
 425                                 hammer_crc_set_blockmap(&cundomap[i]);
 426                         bcopy(cundomap, dundomap, sizeof(hmp->blockmap));
 427                         hammer_modify_volume_done(root_volume);
 428                 }
 429         }
 430
 431         /*
 432          * Flush UNDOs
 433          */
 434         count = 0;
 435         while ((io = TAILQ_FIRST(&hmp->undo_list)) != NULL) {
 436                 KKASSERT(io->modify_refs == 0);
 437                 if (io->lock.refs == 0)
 438                         ++hammer_count_refedbufs;
 439                 hammer_ref(&io->lock);
 440                 KKASSERT(io->type != HAMMER_STRUCTURE_VOLUME);
 441                 hammer_io_flush(io);
 442                 hammer_rel_buffer((hammer_buffer_t)io, 0);
 443                 ++count;
 444         }
 445
 446         /*
 447          * Wait for I/Os to complete
 448          */
 449         hammer_flusher_clean_loose_ios(hmp);
 450         hammer_io_wait_all(hmp, "hmrfl1");
 451
 452         /*
 453          * Update the on-disk volume header with new UNDO FIFO end position
 454          * (do not generate new UNDO records for this change).  We have to
 455          * do this for the UNDO FIFO whether (final) is set or not.
 456          *
 457          * Also update the on-disk next_tid field.  This does not require
 458          * an UNDO.  However, because our TID is generated before we get
 459          * the sync lock another sync may have beat us to the punch.
 460          *
 461          * The volume header will be flushed out synchronously.
 462          */
 463         dundomap = &root_volume->ondisk->vol0_blockmap[HAMMER_ZONE_UNDO_INDEX];
 464         cundomap = &hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
 465
 466         if (dundomap->first_offset != cundomap->first_offset ||
 467             dundomap->next_offset != cundomap->next_offset) {
 468                 hammer_modify_volume(NULL, root_volume, NULL, 0);
 469                 dundomap->first_offset = cundomap->first_offset;
 470                 dundomap->next_offset = cundomap->next_offset;
 471                 hammer_crc_set_blockmap(dundomap);
 472                 hammer_crc_set_volume(root_volume->ondisk);
 473                 if (root_volume->ondisk->vol0_next_tid < trans->tid)
 474                         root_volume->ondisk->vol0_next_tid = trans->tid;
 475                 hammer_modify_volume_done(root_volume);
 476         }
 477
 478         if (root_volume->io.modified) {
 479                 hammer_io_flush(&root_volume->io);
 480         }
 481
 482         /*
 483          * Wait for I/Os to complete
 484          */
 485         hammer_flusher_clean_loose_ios(hmp);
 486         hammer_io_wait_all(hmp, "hmrfl2");
 487
 488         /*
 489          * Flush meta-data.  The meta-data will be undone if we crash
 490          * so we can safely flush it asynchronously.
 491          *
 492          * Repeated catchups will wind up flushing this update's meta-data
 493          * and the UNDO buffers for the next update simultaniously.  This
 494          * is ok.
 495          */
 496         count = 0;
 497         while ((io = TAILQ_FIRST(&hmp->meta_list)) != NULL) {
 498                 KKASSERT(io->modify_refs == 0);
 499                 if (io->lock.refs == 0)
 500                         ++hammer_count_refedbufs;
 501                 hammer_ref(&io->lock);
 502                 KKASSERT(io->type != HAMMER_STRUCTURE_VOLUME);
 503                 hammer_io_flush(io);
 504                 hammer_rel_buffer((hammer_buffer_t)io, 0);
 505                 ++count;
 506         }
 507
 508         /*
 509          * If this is the final finalization for the flush group set
 510          * up for the next sequence by setting a new first_offset in
 511          * our cached blockmap and
 512          * clearing the undo history.
 513          */
 514         if (final) {
 515                 cundomap = &hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
 516                 cundomap->first_offset = cundomap->next_offset;
 517                 hammer_clear_undo_history(hmp);
 518         }
 519
 520         hammer_sync_unlock(trans);
 521 }
 522