git.oblomov.eu Git - linux-2.6/blob - fs/jbd/checkpoint.c

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