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/errno.h>
  24 #include <linux/slab.h>
  25 #include <linux/blkdev.h>
  26 #include <trace/events/jbd2.h>
  27
  28 /*
  29  * Unlink a buffer from a transaction checkpoint list.
  30  *
  31  * Called with j_list_lock held.
  32  */
  33 static inline void __buffer_unlink_first(struct journal_head *jh)
  34 {
  35         transaction_t *transaction = jh->b_cp_transaction;
  36
  37         jh->b_cpnext->b_cpprev = jh->b_cpprev;
  38         jh->b_cpprev->b_cpnext = jh->b_cpnext;
  39         if (transaction->t_checkpoint_list == jh) {
  40                 transaction->t_checkpoint_list = jh->b_cpnext;
  41                 if (transaction->t_checkpoint_list == jh)
  42                         transaction->t_checkpoint_list = NULL;
  43         }
  44 }
  45
  46 /*
  47  * Unlink a buffer from a transaction checkpoint(io) list.
  48  *
  49  * Called with j_list_lock held.
  50  */
  51 static inline void __buffer_unlink(struct journal_head *jh)
  52 {
  53         transaction_t *transaction = jh->b_cp_transaction;
  54
  55         __buffer_unlink_first(jh);
  56         if (transaction->t_checkpoint_io_list == jh) {
  57                 transaction->t_checkpoint_io_list = jh->b_cpnext;
  58                 if (transaction->t_checkpoint_io_list == jh)
  59                         transaction->t_checkpoint_io_list = NULL;
  60         }
  61 }
  62
  63 /*
  64  * Move a buffer from the checkpoint list to the checkpoint io list
  65  *
  66  * Called with j_list_lock held
  67  */
  68 static inline void __buffer_relink_io(struct journal_head *jh)
  69 {
  70         transaction_t *transaction = jh->b_cp_transaction;
  71
  72         __buffer_unlink_first(jh);
  73
  74         if (!transaction->t_checkpoint_io_list) {
  75                 jh->b_cpnext = jh->b_cpprev = jh;
  76         } else {
  77                 jh->b_cpnext = transaction->t_checkpoint_io_list;
  78                 jh->b_cpprev = transaction->t_checkpoint_io_list->b_cpprev;
  79                 jh->b_cpprev->b_cpnext = jh;
  80                 jh->b_cpnext->b_cpprev = jh;
  81         }
  82         transaction->t_checkpoint_io_list = jh;
  83 }
  84
  85 /*
  86  * Try to release a checkpointed buffer from its transaction.
  87  * Returns 1 if we released it and 2 if we also released the
  88  * whole transaction.
  89  *
  90  * Requires j_list_lock
  91  * Called under jbd_lock_bh_state(jh2bh(jh)), and drops it
  92  */
  93 static int __try_to_free_cp_buf(struct journal_head *jh)
  94 {
  95         int ret = 0;
  96         struct buffer_head *bh = jh2bh(jh);
  97
  98         if (jh->b_jlist == BJ_None && !buffer_locked(bh) &&
  99             !buffer_dirty(bh) && !buffer_write_io_error(bh)) {
 100                 /*
 101                  * Get our reference so that bh cannot be freed before
 102                  * we unlock it
 103                  */
 104                 get_bh(bh);
 105                 JBUFFER_TRACE(jh, "remove from checkpoint list");
 106                 ret = __jbd2_journal_remove_checkpoint(jh) + 1;
 107                 jbd_unlock_bh_state(bh);
 108                 BUFFER_TRACE(bh, "release");
 109                 __brelse(bh);
 110         } else {
 111                 jbd_unlock_bh_state(bh);
 112         }
 113         return ret;
 114 }
 115
 116 /*
 117  * __jbd2_log_wait_for_space: wait until there is space in the journal.
 118  *
 119  * Called under j-state_lock *only*.  It will be unlocked if we have to wait
 120  * for a checkpoint to free up some space in the log.
 121  */
 122 void __jbd2_log_wait_for_space(journal_t *journal)
 123 {
 124         int nblocks, space_left;
 125         /* assert_spin_locked(&journal->j_state_lock); */
 126
 127         nblocks = jbd_space_needed(journal);
 128         while (__jbd2_log_space_left(journal) < nblocks) {
 129                 if (journal->j_flags & JBD2_ABORT)
 130                         return;
 131                 write_unlock(&journal->j_state_lock);
 132                 mutex_lock(&journal->j_checkpoint_mutex);
 133
 134                 /*
 135                  * Test again, another process may have checkpointed while we
 136                  * were waiting for the checkpoint lock. If there are no
 137                  * transactions ready to be checkpointed, try to recover
 138                  * journal space by calling cleanup_journal_tail(), and if
 139                  * that doesn't work, by waiting for the currently committing
 140                  * transaction to complete.  If there is absolutely no way
 141                  * to make progress, this is either a BUG or corrupted
 142                  * filesystem, so abort the journal and leave a stack
 143                  * trace for forensic evidence.
 144                  */
 145                 write_lock(&journal->j_state_lock);
 146                 spin_lock(&journal->j_list_lock);
 147                 nblocks = jbd_space_needed(journal);
 148                 space_left = __jbd2_log_space_left(journal);
 149                 if (space_left < nblocks) {
 150                         int chkpt = journal->j_checkpoint_transactions != NULL;
 151                         tid_t tid = 0;
 152
 153                         if (journal->j_committing_transaction)
 154                                 tid = journal->j_committing_transaction->t_tid;
 155                         spin_unlock(&journal->j_list_lock);
 156                         write_unlock(&journal->j_state_lock);
 157                         if (chkpt) {
 158                                 jbd2_log_do_checkpoint(journal);
 159                         } else if (jbd2_cleanup_journal_tail(journal) == 0) {
 160                                 /* We were able to recover space; yay! */
 161                                 ;
 162                         } else if (tid) {
 163                                 jbd2_log_wait_commit(journal, tid);
 164                         } else {
 165                                 printk(KERN_ERR "%s: needed %d blocks and "
 166                                        "only had %d space available\n",
 167                                        __func__, nblocks, space_left);
 168                                 printk(KERN_ERR "%s: no way to get more "
 169                                        "journal space in %s\n", __func__,
 170                                        journal->j_devname);
 171                                 WARN_ON(1);
 172                                 jbd2_journal_abort(journal, 0);
 173                         }
 174                         write_lock(&journal->j_state_lock);
 175                 } else {
 176                         spin_unlock(&journal->j_list_lock);
 177                 }
 178                 mutex_unlock(&journal->j_checkpoint_mutex);
 179         }
 180 }
 181
 182 /*
 183  * We were unable to perform jbd_trylock_bh_state() inside j_list_lock.
 184  * The caller must restart a list walk.  Wait for someone else to run
 185  * jbd_unlock_bh_state().
 186  */
 187 static void jbd_sync_bh(journal_t *journal, struct buffer_head *bh)
 188         __releases(journal->j_list_lock)
 189 {
 190         get_bh(bh);
 191         spin_unlock(&journal->j_list_lock);
 192         jbd_lock_bh_state(bh);
 193         jbd_unlock_bh_state(bh);
 194         put_bh(bh);
 195 }
 196
 197 /*
 198  * Clean up transaction's list of buffers submitted for io.
 199  * We wait for any pending IO to complete and remove any clean
 200  * buffers. Note that we take the buffers in the opposite ordering
 201  * from the one in which they were submitted for IO.
 202  *
 203  * Return 0 on success, and return <0 if some buffers have failed
 204  * to be written out.
 205  *
 206  * Called with j_list_lock held.
 207  */
 208 static int __wait_cp_io(journal_t *journal, transaction_t *transaction)
 209 {
 210         struct journal_head *jh;
 211         struct buffer_head *bh;
 212         tid_t this_tid;
 213         int released = 0;
 214         int ret = 0;
 215
 216         this_tid = transaction->t_tid;
 217 restart:
 218         /* Did somebody clean up the transaction in the meanwhile? */
 219         if (journal->j_checkpoint_transactions != transaction ||
 220                         transaction->t_tid != this_tid)
 221                 return ret;
 222         while (!released && transaction->t_checkpoint_io_list) {
 223                 jh = transaction->t_checkpoint_io_list;
 224                 bh = jh2bh(jh);
 225                 if (!jbd_trylock_bh_state(bh)) {
 226                         jbd_sync_bh(journal, bh);
 227                         spin_lock(&journal->j_list_lock);
 228                         goto restart;
 229                 }
 230                 get_bh(bh);
 231                 if (buffer_locked(bh)) {
 232                         spin_unlock(&journal->j_list_lock);
 233                         jbd_unlock_bh_state(bh);
 234                         wait_on_buffer(bh);
 235                         /* the journal_head may have gone by now */
 236                         BUFFER_TRACE(bh, "brelse");
 237                         __brelse(bh);
 238                         spin_lock(&journal->j_list_lock);
 239                         goto restart;
 240                 }
 241                 if (unlikely(buffer_write_io_error(bh)))
 242                         ret = -EIO;
 243
 244                 /*
 245                  * Now in whatever state the buffer currently is, we know that
 246                  * it has been written out and so we can drop it from the list
 247                  */
 248                 released = __jbd2_journal_remove_checkpoint(jh);
 249                 jbd_unlock_bh_state(bh);
 250                 __brelse(bh);
 251         }
 252
 253         return ret;
 254 }
 255
 256 static void
 257 __flush_batch(journal_t *journal, int *batch_count)
 258 {
 259         int i;
 260         struct blk_plug plug;
 261
 262         blk_start_plug(&plug);
 263         for (i = 0; i < *batch_count; i++)
 264                 write_dirty_buffer(journal->j_chkpt_bhs[i], WRITE_SYNC);
 265         blk_finish_plug(&plug);
 266
 267         for (i = 0; i < *batch_count; i++) {
 268                 struct buffer_head *bh = journal->j_chkpt_bhs[i];
 269                 BUFFER_TRACE(bh, "brelse");
 270                 __brelse(bh);
 271         }
 272         *batch_count = 0;
 273 }
 274
 275 /*
 276  * Try to flush one buffer from the checkpoint list to disk.
 277  *
 278  * Return 1 if something happened which requires us to abort the current
 279  * scan of the checkpoint list.  Return <0 if the buffer has failed to
 280  * be written out.
 281  *
 282  * Called with j_list_lock held and drops it if 1 is returned
 283  * Called under jbd_lock_bh_state(jh2bh(jh)), and drops it
 284  */
 285 static int __process_buffer(journal_t *journal, struct journal_head *jh,
 286                             int *batch_count, transaction_t *transaction)
 287 {
 288         struct buffer_head *bh = jh2bh(jh);
 289         int ret = 0;
 290
 291         if (buffer_locked(bh)) {
 292                 get_bh(bh);
 293                 spin_unlock(&journal->j_list_lock);
 294                 jbd_unlock_bh_state(bh);
 295                 wait_on_buffer(bh);
 296                 /* the journal_head may have gone by now */
 297                 BUFFER_TRACE(bh, "brelse");
 298                 __brelse(bh);
 299                 ret = 1;
 300         } else if (jh->b_transaction != NULL) {
 301                 transaction_t *t = jh->b_transaction;
 302                 tid_t tid = t->t_tid;
 303
 304                 transaction->t_chp_stats.cs_forced_to_close++;
 305                 spin_unlock(&journal->j_list_lock);
 306                 jbd_unlock_bh_state(bh);
 307                 if (unlikely(journal->j_flags & JBD2_UNMOUNT))
 308                         /*
 309                          * The journal thread is dead; so starting and
 310                          * waiting for a commit to finish will cause
 311                          * us to wait for a _very_ long time.
 312                          */
 313                         printk(KERN_ERR "JBD2: %s: "
 314                                "Waiting for Godot: block %llu\n",
 315                                journal->j_devname,
 316                                (unsigned long long) bh->b_blocknr);
 317                 jbd2_log_start_commit(journal, tid);
 318                 jbd2_log_wait_commit(journal, tid);
 319                 ret = 1;
 320         } else if (!buffer_dirty(bh)) {
 321                 ret = 1;
 322                 if (unlikely(buffer_write_io_error(bh)))
 323                         ret = -EIO;
 324                 get_bh(bh);
 325                 J_ASSERT_JH(jh, !buffer_jbddirty(bh));
 326                 BUFFER_TRACE(bh, "remove from checkpoint");
 327                 __jbd2_journal_remove_checkpoint(jh);
 328                 spin_unlock(&journal->j_list_lock);
 329                 jbd_unlock_bh_state(bh);
 330                 __brelse(bh);
 331         } else {
 332                 /*
 333                  * Important: we are about to write the buffer, and
 334                  * possibly block, while still holding the journal lock.
 335                  * We cannot afford to let the transaction logic start
 336                  * messing around with this buffer before we write it to
 337                  * disk, as that would break recoverability.
 338                  */
 339                 BUFFER_TRACE(bh, "queue");
 340                 get_bh(bh);
 341                 J_ASSERT_BH(bh, !buffer_jwrite(bh));
 342                 journal->j_chkpt_bhs[*batch_count] = bh;
 343                 __buffer_relink_io(jh);
 344                 jbd_unlock_bh_state(bh);
 345                 transaction->t_chp_stats.cs_written++;
 346                 (*batch_count)++;
 347                 if (*batch_count == JBD2_NR_BATCH) {
 348                         spin_unlock(&journal->j_list_lock);
 349                         __flush_batch(journal, batch_count);
 350                         ret = 1;
 351                 }
 352         }
 353         return ret;
 354 }
 355
 356 /*
 357  * Perform an actual checkpoint. We take the first transaction on the
 358  * list of transactions to be checkpointed and send all its buffers
 359  * to disk. We submit larger chunks of data at once.
 360  *
 361  * The journal should be locked before calling this function.
 362  * Called with j_checkpoint_mutex held.
 363  */
 364 int jbd2_log_do_checkpoint(journal_t *journal)
 365 {
 366         transaction_t *transaction;
 367         tid_t this_tid;
 368         int result;
 369
 370         jbd_debug(1, "Start checkpoint\n");
 371
 372         /*
 373          * First thing: if there are any transactions in the log which
 374          * don't need checkpointing, just eliminate them from the
 375          * journal straight away.
 376          */
 377         result = jbd2_cleanup_journal_tail(journal);
 378         trace_jbd2_checkpoint(journal, result);
 379         jbd_debug(1, "cleanup_journal_tail returned %d\n", result);
 380         if (result <= 0)
 381                 return result;
 382
 383         /*
 384          * OK, we need to start writing disk blocks.  Take one transaction
 385          * and write it.
 386          */
 387         result = 0;
 388         spin_lock(&journal->j_list_lock);
 389         if (!journal->j_checkpoint_transactions)
 390                 goto out;
 391         transaction = journal->j_checkpoint_transactions;
 392         if (transaction->t_chp_stats.cs_chp_time == 0)
 393                 transaction->t_chp_stats.cs_chp_time = jiffies;
 394         this_tid = transaction->t_tid;
 395 restart:
 396         /*
 397          * If someone cleaned up this transaction while we slept, we're
 398          * done (maybe it's a new transaction, but it fell at the same
 399          * address).
 400          */
 401         if (journal->j_checkpoint_transactions == transaction &&
 402                         transaction->t_tid == this_tid) {
 403                 int batch_count = 0;
 404                 struct journal_head *jh;
 405                 int retry = 0, err;
 406
 407                 while (!retry && transaction->t_checkpoint_list) {
 408                         struct buffer_head *bh;
 409
 410                         jh = transaction->t_checkpoint_list;
 411                         bh = jh2bh(jh);
 412                         if (!jbd_trylock_bh_state(bh)) {
 413                                 jbd_sync_bh(journal, bh);
 414                                 retry = 1;
 415                                 break;
 416                         }
 417                         retry = __process_buffer(journal, jh, &batch_count,
 418                                                  transaction);
 419                         if (retry < 0 && !result)
 420                                 result = retry;
 421                         if (!retry && (need_resched() ||
 422                                 spin_needbreak(&journal->j_list_lock))) {
 423                                 spin_unlock(&journal->j_list_lock);
 424                                 retry = 1;
 425                                 break;
 426                         }
 427                 }
 428
 429                 if (batch_count) {
 430                         if (!retry) {
 431                                 spin_unlock(&journal->j_list_lock);
 432                                 retry = 1;
 433                         }
 434                         __flush_batch(journal, &batch_count);
 435                 }
 436
 437                 if (retry) {
 438                         spin_lock(&journal->j_list_lock);
 439                         goto restart;
 440                 }
 441                 /*
 442                  * Now we have cleaned up the first transaction's checkpoint
 443                  * list. Let's clean up the second one
 444                  */
 445                 err = __wait_cp_io(journal, transaction);
 446                 if (!result)
 447                         result = err;
 448         }
 449 out:
 450         spin_unlock(&journal->j_list_lock);
 451         if (result < 0)
 452                 jbd2_journal_abort(journal, result);
 453         else
 454                 result = jbd2_cleanup_journal_tail(journal);
 455
 456         return (result < 0) ? result : 0;
 457 }
 458
 459 /*
 460  * Check the list of checkpoint transactions for the journal to see if
 461  * we have already got rid of any since the last update of the log tail
 462  * in the journal superblock.  If so, we can instantly roll the
 463  * superblock forward to remove those transactions from the log.
 464  *
 465  * Return <0 on error, 0 on success, 1 if there was nothing to clean up.
 466  *
 467  * Called with the journal lock held.
 468  *
 469  * This is the only part of the journaling code which really needs to be
 470  * aware of transaction aborts.  Checkpointing involves writing to the
 471  * main filesystem area rather than to the journal, so it can proceed
 472  * even in abort state, but we must not update the super block if
 473  * checkpointing may have failed.  Otherwise, we would lose some metadata
 474  * buffers which should be written-back to the filesystem.
 475  */
 476
 477 int jbd2_cleanup_journal_tail(journal_t *journal)
 478 {
 479         tid_t           first_tid;
 480         unsigned long   blocknr;
 481
 482         if (is_journal_aborted(journal))
 483                 return 1;
 484
 485         if (!jbd2_journal_get_log_tail(journal, &first_tid, &blocknr))
 486                 return 1;
 487         J_ASSERT(blocknr != 0);
 488
 489         /*
 490          * We need to make sure that any blocks that were recently written out
 491          * --- perhaps by jbd2_log_do_checkpoint() --- are flushed out before
 492          * we drop the transactions from the journal. It's unlikely this will
 493          * be necessary, especially with an appropriately sized journal, but we
 494          * need this to guarantee correctness.  Fortunately
 495          * jbd2_cleanup_journal_tail() doesn't get called all that often.
 496          */
 497         if (journal->j_flags & JBD2_BARRIER)
 498                 blkdev_issue_flush(journal->j_fs_dev, GFP_KERNEL, NULL);
 499
 500         __jbd2_update_log_tail(journal, first_tid, blocknr);
 501         return 0;
 502 }
 503
 504
 505 /* Checkpoint list management */
 506
 507 /*
 508  * journal_clean_one_cp_list
 509  *
 510  * Find all the written-back checkpoint buffers in the given list and
 511  * release them.
 512  *
 513  * Called with the journal locked.
 514  * Called with j_list_lock held.
 515  * Returns number of buffers reaped (for debug)
 516  */
 517
 518 static int journal_clean_one_cp_list(struct journal_head *jh, int *released)
 519 {
 520         struct journal_head *last_jh;
 521         struct journal_head *next_jh = jh;
 522         int ret, freed = 0;
 523
 524         *released = 0;
 525         if (!jh)
 526                 return 0;
 527
 528         last_jh = jh->b_cpprev;
 529         do {
 530                 jh = next_jh;
 531                 next_jh = jh->b_cpnext;
 532                 /* Use trylock because of the ranking */
 533                 if (jbd_trylock_bh_state(jh2bh(jh))) {
 534                         ret = __try_to_free_cp_buf(jh);
 535                         if (ret) {
 536                                 freed++;
 537                                 if (ret == 2) {
 538                                         *released = 1;
 539                                         return freed;
 540                                 }
 541                         }
 542                 }
 543                 /*
 544                  * This function only frees up some memory
 545                  * if possible so we dont have an obligation
 546                  * to finish processing. Bail out if preemption
 547                  * requested:
 548                  */
 549                 if (need_resched())
 550                         return freed;
 551         } while (jh != last_jh);
 552
 553         return freed;
 554 }
 555
 556 /*
 557  * journal_clean_checkpoint_list
 558  *
 559  * Find all the written-back checkpoint buffers in the journal and release them.
 560  *
 561  * Called with the journal locked.
 562  * Called with j_list_lock held.
 563  * Returns number of buffers reaped (for debug)
 564  */
 565
 566 int __jbd2_journal_clean_checkpoint_list(journal_t *journal)
 567 {
 568         transaction_t *transaction, *last_transaction, *next_transaction;
 569         int ret = 0;
 570         int released;
 571
 572         transaction = journal->j_checkpoint_transactions;
 573         if (!transaction)
 574                 goto out;
 575
 576         last_transaction = transaction->t_cpprev;
 577         next_transaction = transaction;
 578         do {
 579                 transaction = next_transaction;
 580                 next_transaction = transaction->t_cpnext;
 581                 ret += journal_clean_one_cp_list(transaction->
 582                                 t_checkpoint_list, &released);
 583                 /*
 584                  * This function only frees up some memory if possible so we
 585                  * dont have an obligation to finish processing. Bail out if
 586                  * preemption requested:
 587                  */
 588                 if (need_resched())
 589                         goto out;
 590                 if (released)
 591                         continue;
 592                 /*
 593                  * It is essential that we are as careful as in the case of
 594                  * t_checkpoint_list with removing the buffer from the list as
 595                  * we can possibly see not yet submitted buffers on io_list
 596                  */
 597                 ret += journal_clean_one_cp_list(transaction->
 598                                 t_checkpoint_io_list, &released);
 599                 if (need_resched())
 600                         goto out;
 601         } while (transaction != last_transaction);
 602 out:
 603         return ret;
 604 }
 605
 606 /*
 607  * journal_remove_checkpoint: called after a buffer has been committed
 608  * to disk (either by being write-back flushed to disk, or being
 609  * committed to the log).
 610  *
 611  * We cannot safely clean a transaction out of the log until all of the
 612  * buffer updates committed in that transaction have safely been stored
 613  * elsewhere on disk.  To achieve this, all of the buffers in a
 614  * transaction need to be maintained on the transaction's checkpoint
 615  * lists until they have been rewritten, at which point this function is
 616  * called to remove the buffer from the existing transaction's
 617  * checkpoint lists.
 618  *
 619  * The function returns 1 if it frees the transaction, 0 otherwise.
 620  * The function can free jh and bh.
 621  *
 622  * This function is called with j_list_lock held.
 623  * This function is called with jbd_lock_bh_state(jh2bh(jh))
 624  */
 625
 626 int __jbd2_journal_remove_checkpoint(struct journal_head *jh)
 627 {
 628         struct transaction_chp_stats_s *stats;
 629         transaction_t *transaction;
 630         journal_t *journal;
 631         int ret = 0;
 632
 633         JBUFFER_TRACE(jh, "entry");
 634
 635         if ((transaction = jh->b_cp_transaction) == NULL) {
 636                 JBUFFER_TRACE(jh, "not on transaction");
 637                 goto out;
 638         }
 639         journal = transaction->t_journal;
 640
 641         JBUFFER_TRACE(jh, "removing from transaction");
 642         __buffer_unlink(jh);
 643         jh->b_cp_transaction = NULL;
 644         jbd2_journal_put_journal_head(jh);
 645
 646         if (transaction->t_checkpoint_list != NULL ||
 647             transaction->t_checkpoint_io_list != NULL)
 648                 goto out;
 649
 650         /*
 651          * There is one special case to worry about: if we have just pulled the
 652          * buffer off a running or committing transaction's checkpoing list,
 653          * then even if the checkpoint list is empty, the transaction obviously
 654          * cannot be dropped!
 655          *
 656          * The locking here around t_state is a bit sleazy.
 657          * See the comment at the end of jbd2_journal_commit_transaction().
 658          */
 659         if (transaction->t_state != T_FINISHED)
 660                 goto out;
 661
 662         /* OK, that was the last buffer for the transaction: we can now
 663            safely remove this transaction from the log */
 664         stats = &transaction->t_chp_stats;
 665         if (stats->cs_chp_time)
 666                 stats->cs_chp_time = jbd2_time_diff(stats->cs_chp_time,
 667                                                     jiffies);
 668         trace_jbd2_checkpoint_stats(journal->j_fs_dev->bd_dev,
 669                                     transaction->t_tid, stats);
 670
 671         __jbd2_journal_drop_transaction(journal, transaction);
 672         jbd2_journal_free_transaction(transaction);
 673
 674         /* Just in case anybody was waiting for more transactions to be
 675            checkpointed... */
 676         wake_up(&journal->j_wait_logspace);
 677         ret = 1;
 678 out:
 679         return ret;
 680 }
 681
 682 /*
 683  * journal_insert_checkpoint: put a committed buffer onto a checkpoint
 684  * list so that we know when it is safe to clean the transaction out of
 685  * the log.
 686  *
 687  * Called with the journal locked.
 688  * Called with j_list_lock held.
 689  */
 690 void __jbd2_journal_insert_checkpoint(struct journal_head *jh,
 691                                transaction_t *transaction)
 692 {
 693         JBUFFER_TRACE(jh, "entry");
 694         J_ASSERT_JH(jh, buffer_dirty(jh2bh(jh)) || buffer_jbddirty(jh2bh(jh)));
 695         J_ASSERT_JH(jh, jh->b_cp_transaction == NULL);
 696
 697         /* Get reference for checkpointing transaction */
 698         jbd2_journal_grab_journal_head(jh2bh(jh));
 699         jh->b_cp_transaction = transaction;
 700
 701         if (!transaction->t_checkpoint_list) {
 702                 jh->b_cpnext = jh->b_cpprev = jh;
 703         } else {
 704                 jh->b_cpnext = transaction->t_checkpoint_list;
 705                 jh->b_cpprev = transaction->t_checkpoint_list->b_cpprev;
 706                 jh->b_cpprev->b_cpnext = jh;
 707                 jh->b_cpnext->b_cpprev = jh;
 708         }
 709         transaction->t_checkpoint_list = jh;
 710 }
 711
 712 /*
 713  * We've finished with this transaction structure: adios...
 714  *
 715  * The transaction must have no links except for the checkpoint by this
 716  * point.
 717  *
 718  * Called with the journal locked.
 719  * Called with j_list_lock held.
 720  */
 721
 722 void __jbd2_journal_drop_transaction(journal_t *journal, transaction_t *transaction)
 723 {
 724         assert_spin_locked(&journal->j_list_lock);
 725         if (transaction->t_cpnext) {
 726                 transaction->t_cpnext->t_cpprev = transaction->t_cpprev;
 727                 transaction->t_cpprev->t_cpnext = transaction->t_cpnext;
 728                 if (journal->j_checkpoint_transactions == transaction)
 729                         journal->j_checkpoint_transactions =
 730                                 transaction->t_cpnext;
 731                 if (journal->j_checkpoint_transactions == transaction)
 732                         journal->j_checkpoint_transactions = NULL;
 733         }
 734
 735         J_ASSERT(transaction->t_state == T_FINISHED);
 736         J_ASSERT(transaction->t_buffers == NULL);
 737         J_ASSERT(transaction->t_forget == NULL);
 738         J_ASSERT(transaction->t_iobuf_list == NULL);
 739         J_ASSERT(transaction->t_shadow_list == NULL);
 740         J_ASSERT(transaction->t_log_list == NULL);
 741         J_ASSERT(transaction->t_checkpoint_list == NULL);
 742         J_ASSERT(transaction->t_checkpoint_io_list == NULL);
 743         J_ASSERT(atomic_read(&transaction->t_updates) == 0);
 744         J_ASSERT(journal->j_committing_transaction != transaction);
 745         J_ASSERT(journal->j_running_transaction != transaction);
 746
 747         trace_jbd2_drop_transaction(journal, transaction);
 748
 749         jbd_debug(1, "Dropping transaction %d, all done\n", transaction->t_tid);
 750 }