fs/jbd2/checkpoint.c

   1 /*
   2  * linux/fs/jbd2/checkpoint.c
   3  *
   4  * Written by Stephen C. Tweedie <sct@redhat.com>, 1999
   5  *
   6  * Copyright 1999 Red Hat Software --- All Rights Reserved
   7  *
   8  * This file is part of the Linux kernel and is made available under
   9  * the terms of the GNU General Public License, version 2, or at your
  10  * option, any later version, incorporated herein by reference.
  11  *
  12  * Checkpoint routines for the generic filesystem journaling code.
  13  * Part of the ext2fs journaling system.
  14  *
  15  * Checkpointing is the process of ensuring that a section of the log is
  16  * committed fully to disk, so that that portion of the log can be
  17  * reused.
  18  */
  19
  20 #include <linux/time.h>
  21 #include <linux/fs.h>
  22 #include <linux/jbd2.h>
  23 #include <linux/marker.h>
  24 #include <linux/errno.h>
  25 #include <linux/slab.h>
  26
  27 /*
  28  * Unlink a buffer from a transaction checkpoint list.
  29  *
  30  * Called with j_list_lock held.
  31  */
  32 static inline void __buffer_unlink_first(struct journal_head *jh)
  33 {
  34         transaction_t *transaction = jh->b_cp_transaction;
  35
  36         jh->b_cpnext->b_cpprev = jh->b_cpprev;
  37         jh->b_cpprev->b_cpnext = jh->b_cpnext;
  38         if (transaction->t_checkpoint_list == jh) {
  39                 transaction->t_checkpoint_list = jh->b_cpnext;
  40                 if (transaction->t_checkpoint_list == jh)
  41                         transaction->t_checkpoint_list = NULL;
  42         }
  43 }
  44
  45 /*
  46  * Unlink a buffer from a transaction checkpoint(io) list.
  47  *
  48  * Called with j_list_lock held.
  49  */
  50 static inline void __buffer_unlink(struct journal_head *jh)
  51 {
  52         transaction_t *transaction = jh->b_cp_transaction;
  53
  54         __buffer_unlink_first(jh);
  55         if (transaction->t_checkpoint_io_list == jh) {
  56                 transaction->t_checkpoint_io_list = jh->b_cpnext;
  57                 if (transaction->t_checkpoint_io_list == jh)
  58                         transaction->t_checkpoint_io_list = NULL;
  59         }
  60 }
  61
  62 /*
  63  * Move a buffer from the checkpoint list to the checkpoint io list
  64  *
  65  * Called with j_list_lock held
  66  */
  67 static inline void __buffer_relink_io(struct journal_head *jh)
  68 {
  69         transaction_t *transaction = jh->b_cp_transaction;
  70
  71         __buffer_unlink_first(jh);
  72
  73         if (!transaction->t_checkpoint_io_list) {
  74                 jh->b_cpnext = jh->b_cpprev = jh;
  75         } else {
  76                 jh->b_cpnext = transaction->t_checkpoint_io_list;
  77                 jh->b_cpprev = transaction->t_checkpoint_io_list->b_cpprev;
  78                 jh->b_cpprev->b_cpnext = jh;
  79                 jh->b_cpnext->b_cpprev = jh;
  80         }
  81         transaction->t_checkpoint_io_list = jh;
  82 }
  83
  84 /*
  85  * Try to release a checkpointed buffer from its transaction.
  86  * Returns 1 if we released it and 2 if we also released the
  87  * whole transaction.
  88  *
  89  * Requires j_list_lock
  90  * Called under jbd_lock_bh_state(jh2bh(jh)), and drops it
  91  */
  92 static int __try_to_free_cp_buf(struct journal_head *jh)
  93 {
  94         int ret = 0;
  95         struct buffer_head *bh = jh2bh(jh);
  96
  97         if (jh->b_jlist == BJ_None && !buffer_locked(bh) &&
  98             !buffer_dirty(bh) && !buffer_write_io_error(bh)) {
  99                 JBUFFER_TRACE(jh, "remove from checkpoint list");
 100                 ret = __jbd2_journal_remove_checkpoint(jh) + 1;
 101                 jbd_unlock_bh_state(bh);
 102                 jbd2_journal_remove_journal_head(bh);
 103                 BUFFER_TRACE(bh, "release");
 104                 __brelse(bh);
 105         } else {
 106                 jbd_unlock_bh_state(bh);
 107         }
 108         return ret;
 109 }
 110
 111 /*
 112  * __jbd2_log_wait_for_space: wait until there is space in the journal.
 113  *
 114  * Called under j-state_lock *only*.  It will be unlocked if we have to wait
 115  * for a checkpoint to free up some space in the log.
 116  */
 117 void __jbd2_log_wait_for_space(journal_t *journal)
 118 {
 119         int nblocks, space_left;
 120         assert_spin_locked(&journal->j_state_lock);
 121
 122         nblocks = jbd_space_needed(journal);
 123         while (__jbd2_log_space_left(journal) < nblocks) {
 124                 if (journal->j_flags & JBD2_ABORT)
 125                         return;
 126                 spin_unlock(&journal->j_state_lock);
 127                 mutex_lock(&journal->j_checkpoint_mutex);
 128
 129                 /*
 130                  * Test again, another process may have checkpointed while we
 131                  * were waiting for the checkpoint lock. If there are no
 132                  * transactions ready to be checkpointed, try to recover
 133                  * journal space by calling cleanup_journal_tail(), and if
 134                  * that doesn't work, by waiting for the currently committing
 135                  * transaction to complete.  If there is absolutely no way
 136                  * to make progress, this is either a BUG or corrupted
 137                  * filesystem, so abort the journal and leave a stack
 138                  * trace for forensic evidence.
 139                  */
 140                 spin_lock(&journal->j_state_lock);
 141                 spin_lock(&journal->j_list_lock);
 142                 nblocks = jbd_space_needed(journal);
 143                 space_left = __jbd2_log_space_left(journal);
 144                 if (space_left < nblocks) {
 145                         int chkpt = journal->j_checkpoint_transactions != NULL;
 146                         tid_t tid = 0;
 147
 148                         if (journal->j_committing_transaction)
 149                                 tid = journal->j_committing_transaction->t_tid;
 150                         spin_unlock(&journal->j_list_lock);
 151                         spin_unlock(&journal->j_state_lock);
 152                         if (chkpt) {
 153                                 jbd2_log_do_checkpoint(journal);
 154                         } else if (jbd2_cleanup_journal_tail(journal) == 0) {
 155                                 /* We were able to recover space; yay! */
 156                                 ;
 157                         } else if (tid) {
 158                                 jbd2_log_wait_commit(journal, tid);
 159                         } else {
 160                                 printk(KERN_ERR "%s: needed %d blocks and "
 161                                        "only had %d space available\n",
 162                                        __func__, nblocks, space_left);
 163                                 printk(KERN_ERR "%s: no way to get more "
 164                                        "journal space in %s\n", __func__,
 165                                        journal->j_devname);
 166                                 WARN_ON(1);
 167                                 jbd2_journal_abort(journal, 0);
 168                         }
 169                         spin_lock(&journal->j_state_lock);
 170                 } else {
 171                         spin_unlock(&journal->j_list_lock);
 172                 }
 173                 mutex_unlock(&journal->j_checkpoint_mutex);
 174         }
 175 }
 176
 177 /*
 178  * We were unable to perform jbd_trylock_bh_state() inside j_list_lock.
 179  * The caller must restart a list walk.  Wait for someone else to run
 180  * jbd_unlock_bh_state().
 181  */
 182 static void jbd_sync_bh(journal_t *journal, struct buffer_head *bh)
 183         __releases(journal->j_list_lock)
 184 {
 185         get_bh(bh);
 186         spin_unlock(&journal->j_list_lock);
 187         jbd_lock_bh_state(bh);
 188         jbd_unlock_bh_state(bh);
 189         put_bh(bh);
 190 }
 191
 192 /*
 193  * Clean up transaction's list of buffers submitted for io.
 194  * We wait for any pending IO to complete and remove any clean
 195  * buffers. Note that we take the buffers in the opposite ordering
 196  * from the one in which they were submitted for IO.
 197  *
 198  * Return 0 on success, and return <0 if some buffers have failed
 199  * to be written out.
 200  *
 201  * Called with j_list_lock held.
 202  */
 203 static int __wait_cp_io(journal_t *journal, transaction_t *transaction)
 204 {
 205         struct journal_head *jh;
 206         struct buffer_head *bh;
 207         tid_t this_tid;
 208         int released = 0;
 209         int ret = 0;
 210
 211         this_tid = transaction->t_tid;
 212 restart:
 213         /* Did somebody clean up the transaction in the meanwhile? */
 214         if (journal->j_checkpoint_transactions != transaction ||
 215                         transaction->t_tid != this_tid)
 216                 return ret;
 217         while (!released && transaction->t_checkpoint_io_list) {
 218                 jh = transaction->t_checkpoint_io_list;
 219                 bh = jh2bh(jh);
 220                 if (!jbd_trylock_bh_state(bh)) {
 221                         jbd_sync_bh(journal, bh);
 222                         spin_lock(&journal->j_list_lock);
 223                         goto restart;
 224                 }
 225                 if (buffer_locked(bh)) {
 226                         atomic_inc(&bh->b_count);
 227                         spin_unlock(&journal->j_list_lock);
 228                         jbd_unlock_bh_state(bh);
 229                         wait_on_buffer(bh);
 230                         /* the journal_head may have gone by now */
 231                         BUFFER_TRACE(bh, "brelse");
 232                         __brelse(bh);
 233                         spin_lock(&journal->j_list_lock);
 234                         goto restart;
 235                 }
 236                 if (unlikely(buffer_write_io_error(bh)))
 237                         ret = -EIO;
 238
 239                 /*
 240                  * Now in whatever state the buffer currently is, we know that
 241                  * it has been written out and so we can drop it from the list
 242                  */
 243                 released = __jbd2_journal_remove_checkpoint(jh);
 244                 jbd_unlock_bh_state(bh);
 245                 jbd2_journal_remove_journal_head(bh);
 246                 __brelse(bh);
 247         }
 248
 249         return ret;
 250 }
 251
 252 #define NR_BATCH        64
 253
 254 static void
 255 __flush_batch(journal_t *journal, struct buffer_head **bhs, int *batch_count)
 256 {
 257         int i;
 258
 259         ll_rw_block(SWRITE, *batch_count, bhs);
 260         for (i = 0; i < *batch_count; i++) {
 261                 struct buffer_head *bh = bhs[i];
 262                 clear_buffer_jwrite(bh);
 263                 BUFFER_TRACE(bh, "brelse");
 264                 __brelse(bh);
 265         }
 266         *batch_count = 0;
 267 }
 268
 269 /*
 270  * Try to flush one buffer from the checkpoint list to disk.
 271  *
 272  * Return 1 if something happened which requires us to abort the current
 273  * scan of the checkpoint list.  Return <0 if the buffer has failed to
 274  * be written out.
 275  *
 276  * Called with j_list_lock held and drops it if 1 is returned
 277  * Called under jbd_lock_bh_state(jh2bh(jh)), and drops it
 278  */
 279 static int __process_buffer(journal_t *journal, struct journal_head *jh,
 280                         struct buffer_head **bhs, int *batch_count,
 281                         transaction_t *transaction)
 282 {
 283         struct buffer_head *bh = jh2bh(jh);
 284         int ret = 0;
 285
 286         if (buffer_locked(bh)) {
 287                 atomic_inc(&bh->b_count);
 288                 spin_unlock(&journal->j_list_lock);
 289                 jbd_unlock_bh_state(bh);
 290                 wait_on_buffer(bh);
 291                 /* the journal_head may have gone by now */
 292                 BUFFER_TRACE(bh, "brelse");
 293                 __brelse(bh);
 294                 ret = 1;
 295         } else if (jh->b_transaction != NULL) {
 296                 transaction_t *t = jh->b_transaction;
 297                 tid_t tid = t->t_tid;
 298
 299                 transaction->t_chp_stats.cs_forced_to_close++;
 300                 spin_unlock(&journal->j_list_lock);
 301                 jbd_unlock_bh_state(bh);
 302                 jbd2_log_start_commit(journal, tid);
 303                 jbd2_log_wait_commit(journal, tid);
 304                 ret = 1;
 305         } else if (!buffer_dirty(bh)) {
 306                 ret = 1;
 307                 if (unlikely(buffer_write_io_error(bh)))
 308                         ret = -EIO;
 309                 J_ASSERT_JH(jh, !buffer_jbddirty(bh));
 310                 BUFFER_TRACE(bh, "remove from checkpoint");
 311                 __jbd2_journal_remove_checkpoint(jh);
 312                 spin_unlock(&journal->j_list_lock);
 313                 jbd_unlock_bh_state(bh);
 314                 jbd2_journal_remove_journal_head(bh);
 315                 __brelse(bh);
 316         } else {
 317                 /*
 318                  * Important: we are about to write the buffer, and
 319                  * possibly block, while still holding the journal lock.
 320                  * We cannot afford to let the transaction logic start
 321                  * messing around with this buffer before we write it to
 322                  * disk, as that would break recoverability.
 323                  */
 324                 BUFFER_TRACE(bh, "queue");
 325                 get_bh(bh);
 326                 J_ASSERT_BH(bh, !buffer_jwrite(bh));
 327                 set_buffer_jwrite(bh);
 328                 bhs[*batch_count] = bh;
 329                 __buffer_relink_io(jh);
 330                 jbd_unlock_bh_state(bh);
 331                 transaction->t_chp_stats.cs_written++;
 332                 (*batch_count)++;
 333                 if (*batch_count == NR_BATCH) {
 334                         spin_unlock(&journal->j_list_lock);
 335                         __flush_batch(journal, bhs, batch_count);
 336                         ret = 1;
 337                 }
 338         }
 339         return ret;
 340 }
 341
 342 /*
 343  * Perform an actual checkpoint. We take the first transaction on the
 344  * list of transactions to be checkpointed and send all its buffers
 345  * to disk. We submit larger chunks of data at once.
 346  *
 347  * The journal should be locked before calling this function.
 348  * Called with j_checkpoint_mutex held.
 349  */
 350 int jbd2_log_do_checkpoint(journal_t *journal)
 351 {
 352         transaction_t *transaction;
 353         tid_t this_tid;
 354         int result;
 355
 356         jbd_debug(1, "Start checkpoint\n");
 357
 358         /*
 359          * First thing: if there are any transactions in the log which
 360          * don't need checkpointing, just eliminate them from the
 361          * journal straight away.
 362          */
 363         result = jbd2_cleanup_journal_tail(journal);
 364         trace_mark(jbd2_checkpoint, "dev %s need_checkpoint %d",
 365                    journal->j_devname, result);
 366         jbd_debug(1, "cleanup_journal_tail returned %d\n", result);
 367         if (result <= 0)
 368                 return result;
 369
 370         /*
 371          * OK, we need to start writing disk blocks.  Take one transaction
 372          * and write it.
 373          */
 374         result = 0;
 375         spin_lock(&journal->j_list_lock);
 376         if (!journal->j_checkpoint_transactions)
 377                 goto out;
 378         transaction = journal->j_checkpoint_transactions;
 379         if (transaction->t_chp_stats.cs_chp_time == 0)
 380                 transaction->t_chp_stats.cs_chp_time = jiffies;
 381         this_tid = transaction->t_tid;
 382 restart:
 383         /*
 384          * If someone cleaned up this transaction while we slept, we're
 385          * done (maybe it's a new transaction, but it fell at the same
 386          * address).
 387          */
 388         if (journal->j_checkpoint_transactions == transaction &&
 389                         transaction->t_tid == this_tid) {
 390                 int batch_count = 0;
 391                 struct buffer_head *bhs[NR_BATCH];
 392                 struct journal_head *jh;
 393                 int retry = 0, err;
 394
 395                 while (!retry && transaction->t_checkpoint_list) {
 396                         struct buffer_head *bh;
 397
 398                         jh = transaction->t_checkpoint_list;
 399                         bh = jh2bh(jh);
 400                         if (!jbd_trylock_bh_state(bh)) {
 401                                 jbd_sync_bh(journal, bh);
 402                                 retry = 1;
 403                                 break;
 404                         }
 405                         retry = __process_buffer(journal, jh, bhs, &batch_count,
 406                                                  transaction);
 407                         if (retry < 0 && !result)
 408                                 result = retry;
 409                         if (!retry && (need_resched() ||
 410                                 spin_needbreak(&journal->j_list_lock))) {
 411                                 spin_unlock(&journal->j_list_lock);
 412                                 retry = 1;
 413                                 break;
 414                         }
 415                 }
 416
 417                 if (batch_count) {
 418                         if (!retry) {
 419                                 spin_unlock(&journal->j_list_lock);
 420                                 retry = 1;
 421                         }
 422                         __flush_batch(journal, bhs, &batch_count);
 423                 }
 424
 425                 if (retry) {
 426                         spin_lock(&journal->j_list_lock);
 427                         goto restart;
 428                 }
 429                 /*
 430                  * Now we have cleaned up the first transaction's checkpoint
 431                  * list. Let's clean up the second one
 432                  */
 433                 err = __wait_cp_io(journal, transaction);
 434                 if (!result)
 435                         result = err;
 436         }
 437 out:
 438         spin_unlock(&journal->j_list_lock);
 439         if (result < 0)
 440                 jbd2_journal_abort(journal, result);
 441         else
 442                 result = jbd2_cleanup_journal_tail(journal);
 443
 444         return (result < 0) ? result : 0;
 445 }
 446
 447 /*
 448  * Check the list of checkpoint transactions for the journal to see if
 449  * we have already got rid of any since the last update of the log tail
 450  * in the journal superblock.  If so, we can instantly roll the
 451  * superblock forward to remove those transactions from the log.
 452  *
 453  * Return <0 on error, 0 on success, 1 if there was nothing to clean up.
 454  *
 455  * Called with the journal lock held.
 456  *
 457  * This is the only part of the journaling code which really needs to be
 458  * aware of transaction aborts.  Checkpointing involves writing to the
 459  * main filesystem area rather than to the journal, so it can proceed
 460  * even in abort state, but we must not update the super block if
 461  * checkpointing may have failed.  Otherwise, we would lose some metadata
 462  * buffers which should be written-back to the filesystem.
 463  */
 464
 465 int jbd2_cleanup_journal_tail(journal_t *journal)
 466 {
 467         transaction_t * transaction;
 468         tid_t           first_tid;
 469         unsigned long   blocknr, freed;
 470
 471         if (is_journal_aborted(journal))
 472                 return 1;
 473
 474         /* OK, work out the oldest transaction remaining in the log, and
 475          * the log block it starts at.
 476          *
 477          * If the log is now empty, we need to work out which is the
 478          * next transaction ID we will write, and where it will
 479          * start. */
 480
 481         spin_lock(&journal->j_state_lock);
 482         spin_lock(&journal->j_list_lock);
 483         transaction = journal->j_checkpoint_transactions;
 484         if (transaction) {
 485                 first_tid = transaction->t_tid;
 486                 blocknr = transaction->t_log_start;
 487         } else if ((transaction = journal->j_committing_transaction) != NULL) {
 488                 first_tid = transaction->t_tid;
 489                 blocknr = transaction->t_log_start;
 490         } else if ((transaction = journal->j_running_transaction) != NULL) {
 491                 first_tid = transaction->t_tid;
 492                 blocknr = journal->j_head;
 493         } else {
 494                 first_tid = journal->j_transaction_sequence;
 495                 blocknr = journal->j_head;
 496         }
 497         spin_unlock(&journal->j_list_lock);
 498         J_ASSERT(blocknr != 0);
 499
 500         /* If the oldest pinned transaction is at the tail of the log
 501            already then there's not much we can do right now. */
 502         if (journal->j_tail_sequence == first_tid) {
 503                 spin_unlock(&journal->j_state_lock);
 504                 return 1;
 505         }
 506
 507         /* OK, update the superblock to recover the freed space.
 508          * Physical blocks come first: have we wrapped beyond the end of
 509          * the log?  */
 510         freed = blocknr - journal->j_tail;
 511         if (blocknr < journal->j_tail)
 512                 freed = freed + journal->j_last - journal->j_first;
 513
 514         jbd_debug(1,
 515                   "Cleaning journal tail from %d to %d (offset %lu), "
 516                   "freeing %lu\n",
 517                   journal->j_tail_sequence, first_tid, blocknr, freed);
 518
 519         journal->j_free += freed;
 520         journal->j_tail_sequence = first_tid;
 521         journal->j_tail = blocknr;
 522         spin_unlock(&journal->j_state_lock);
 523         if (!(journal->j_flags & JBD2_ABORT))
 524                 jbd2_journal_update_superblock(journal, 1);
 525         return 0;
 526 }
 527
 528
 529 /* Checkpoint list management */
 530
 531 /*
 532  * journal_clean_one_cp_list
 533  *
 534  * Find all the written-back checkpoint buffers in the given list and release them.
 535  *
 536  * Called with the journal locked.
 537  * Called with j_list_lock held.
 538  * Returns number of bufers reaped (for debug)
 539  */
 540
 541 static int journal_clean_one_cp_list(struct journal_head *jh, int *released)
 542 {
 543         struct journal_head *last_jh;
 544         struct journal_head *next_jh = jh;
 545         int ret, freed = 0;
 546
 547         *released = 0;
 548         if (!jh)
 549                 return 0;
 550
 551         last_jh = jh->b_cpprev;
 552         do {
 553                 jh = next_jh;
 554                 next_jh = jh->b_cpnext;
 555                 /* Use trylock because of the ranking */
 556                 if (jbd_trylock_bh_state(jh2bh(jh))) {
 557                         ret = __try_to_free_cp_buf(jh);
 558                         if (ret) {
 559                                 freed++;
 560                                 if (ret == 2) {
 561                                         *released = 1;
 562                                         return freed;
 563                                 }
 564                         }
 565                 }
 566                 /*
 567                  * This function only frees up some memory
 568                  * if possible so we dont have an obligation
 569                  * to finish processing. Bail out if preemption
 570                  * requested:
 571                  */
 572                 if (need_resched())
 573                         return freed;
 574         } while (jh != last_jh);
 575
 576         return freed;
 577 }
 578
 579 /*
 580  * journal_clean_checkpoint_list
 581  *
 582  * Find all the written-back checkpoint buffers in the journal and release them.
 583  *
 584  * Called with the journal locked.
 585  * Called with j_list_lock held.
 586  * Returns number of buffers reaped (for debug)
 587  */
 588
 589 int __jbd2_journal_clean_checkpoint_list(journal_t *journal)
 590 {
 591         transaction_t *transaction, *last_transaction, *next_transaction;
 592         int ret = 0;
 593         int released;
 594
 595         transaction = journal->j_checkpoint_transactions;
 596         if (!transaction)
 597                 goto out;
 598
 599         last_transaction = transaction->t_cpprev;
 600         next_transaction = transaction;
 601         do {
 602                 transaction = next_transaction;
 603                 next_transaction = transaction->t_cpnext;
 604                 ret += journal_clean_one_cp_list(transaction->
 605                                 t_checkpoint_list, &released);
 606                 /*
 607                  * This function only frees up some memory if possible so we
 608                  * dont have an obligation to finish processing. Bail out if
 609                  * preemption requested:
 610                  */
 611                 if (need_resched())
 612                         goto out;
 613                 if (released)
 614                         continue;
 615                 /*
 616                  * It is essential that we are as careful as in the case of
 617                  * t_checkpoint_list with removing the buffer from the list as
 618                  * we can possibly see not yet submitted buffers on io_list
 619                  */
 620                 ret += journal_clean_one_cp_list(transaction->
 621                                 t_checkpoint_io_list, &released);
 622                 if (need_resched())
 623                         goto out;
 624         } while (transaction != last_transaction);
 625 out:
 626         return ret;
 627 }
 628
 629 /*
 630  * journal_remove_checkpoint: called after a buffer has been committed
 631  * to disk (either by being write-back flushed to disk, or being
 632  * committed to the log).
 633  *
 634  * We cannot safely clean a transaction out of the log until all of the
 635  * buffer updates committed in that transaction have safely been stored
 636  * elsewhere on disk.  To achieve this, all of the buffers in a
 637  * transaction need to be maintained on the transaction's checkpoint
 638  * lists until they have been rewritten, at which point this function is
 639  * called to remove the buffer from the existing transaction's
 640  * checkpoint lists.
 641  *
 642  * The function returns 1 if it frees the transaction, 0 otherwise.
 643  *
 644  * This function is called with the journal locked.
 645  * This function is called with j_list_lock held.
 646  * This function is called with jbd_lock_bh_state(jh2bh(jh))
 647  */
 648
 649 int __jbd2_journal_remove_checkpoint(struct journal_head *jh)
 650 {
 651         transaction_t *transaction;
 652         journal_t *journal;
 653         int ret = 0;
 654
 655         JBUFFER_TRACE(jh, "entry");
 656
 657         if ((transaction = jh->b_cp_transaction) == NULL) {
 658                 JBUFFER_TRACE(jh, "not on transaction");
 659                 goto out;
 660         }
 661         journal = transaction->t_journal;
 662
 663         __buffer_unlink(jh);
 664         jh->b_cp_transaction = NULL;
 665
 666         if (transaction->t_checkpoint_list != NULL ||
 667             transaction->t_checkpoint_io_list != NULL)
 668                 goto out;
 669         JBUFFER_TRACE(jh, "transaction has no more buffers");
 670
 671         /*
 672          * There is one special case to worry about: if we have just pulled the
 673          * buffer off a running or committing transaction's checkpoing list,
 674          * then even if the checkpoint list is empty, the transaction obviously
 675          * cannot be dropped!
 676          *
 677          * The locking here around t_state is a bit sleazy.
 678          * See the comment at the end of jbd2_journal_commit_transaction().
 679          */
 680         if (transaction->t_state != T_FINISHED) {
 681                 JBUFFER_TRACE(jh, "belongs to running/committing transaction");
 682                 goto out;
 683         }
 684
 685         /* OK, that was the last buffer for the transaction: we can now
 686            safely remove this transaction from the log */
 687
 688         __jbd2_journal_drop_transaction(journal, transaction);
 689
 690         /* Just in case anybody was waiting for more transactions to be
 691            checkpointed... */
 692         wake_up(&journal->j_wait_logspace);
 693         ret = 1;
 694 out:
 695         JBUFFER_TRACE(jh, "exit");
 696         return ret;
 697 }
 698
 699 /*
 700  * journal_insert_checkpoint: put a committed buffer onto a checkpoint
 701  * list so that we know when it is safe to clean the transaction out of
 702  * the log.
 703  *
 704  * Called with the journal locked.
 705  * Called with j_list_lock held.
 706  */
 707 void __jbd2_journal_insert_checkpoint(struct journal_head *jh,
 708                                transaction_t *transaction)
 709 {
 710         JBUFFER_TRACE(jh, "entry");
 711         J_ASSERT_JH(jh, buffer_dirty(jh2bh(jh)) || buffer_jbddirty(jh2bh(jh)));
 712         J_ASSERT_JH(jh, jh->b_cp_transaction == NULL);
 713
 714         jh->b_cp_transaction = transaction;
 715
 716         if (!transaction->t_checkpoint_list) {
 717                 jh->b_cpnext = jh->b_cpprev = jh;
 718         } else {
 719                 jh->b_cpnext = transaction->t_checkpoint_list;
 720                 jh->b_cpprev = transaction->t_checkpoint_list->b_cpprev;
 721                 jh->b_cpprev->b_cpnext = jh;
 722                 jh->b_cpnext->b_cpprev = jh;
 723         }
 724         transaction->t_checkpoint_list = jh;
 725 }
 726
 727 /*
 728  * We've finished with this transaction structure: adios...
 729  *
 730  * The transaction must have no links except for the checkpoint by this
 731  * point.
 732  *
 733  * Called with the journal locked.
 734  * Called with j_list_lock held.
 735  */
 736
 737 void __jbd2_journal_drop_transaction(journal_t *journal, transaction_t *transaction)
 738 {
 739         assert_spin_locked(&journal->j_list_lock);
 740         if (transaction->t_cpnext) {
 741                 transaction->t_cpnext->t_cpprev = transaction->t_cpprev;
 742                 transaction->t_cpprev->t_cpnext = transaction->t_cpnext;
 743                 if (journal->j_checkpoint_transactions == transaction)
 744                         journal->j_checkpoint_transactions =
 745                                 transaction->t_cpnext;
 746                 if (journal->j_checkpoint_transactions == transaction)
 747                         journal->j_checkpoint_transactions = NULL;
 748         }
 749
 750         J_ASSERT(transaction->t_state == T_FINISHED);
 751         J_ASSERT(transaction->t_buffers == NULL);
 752         J_ASSERT(transaction->t_forget == NULL);
 753         J_ASSERT(transaction->t_iobuf_list == NULL);
 754         J_ASSERT(transaction->t_shadow_list == NULL);
 755         J_ASSERT(transaction->t_log_list == NULL);
 756         J_ASSERT(transaction->t_checkpoint_list == NULL);
 757         J_ASSERT(transaction->t_checkpoint_io_list == NULL);
 758         J_ASSERT(transaction->t_updates == 0);
 759         J_ASSERT(journal->j_committing_transaction != transaction);
 760         J_ASSERT(journal->j_running_transaction != transaction);
 761
 762         jbd_debug(1, "Dropping transaction %d, all done\n", transaction->t_tid);
 763         kfree(transaction);
 764 }