Pileus Git - ~andy/linux/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 #include <trace/events/jbd.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 = __journal_remove_checkpoint(jh) + 1;
 101                 jbd_unlock_bh_state(bh);
 102                 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  * __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 __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 (__log_space_left(journal) < nblocks) {
 124                 if (journal->j_flags & JFS_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 = __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                                 log_do_checkpoint(journal);
 154                         } else if (cleanup_journal_tail(journal) == 0) {
 155                                 /* We were able to recover space; yay! */
 156                                 ;
 157                         } else if (tid) {
 158                                 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\n", __func__);
 165                                 WARN_ON(1);
 166                                 journal_abort(journal, 0);
 167                         }
 168                         spin_lock(&journal->j_state_lock);
 169                 } else {
 170                         spin_unlock(&journal->j_list_lock);
 171                 }
 172                 mutex_unlock(&journal->j_checkpoint_mutex);
 173         }
 174 }
 175
 176 /*
 177  * We were unable to perform jbd_trylock_bh_state() inside j_list_lock.
 178  * The caller must restart a list walk.  Wait for someone else to run
 179  * jbd_unlock_bh_state().
 180  */
 181 static void jbd_sync_bh(journal_t *journal, struct buffer_head *bh)
 182         __releases(journal->j_list_lock)
 183 {
 184         get_bh(bh);
 185         spin_unlock(&journal->j_list_lock);
 186         jbd_lock_bh_state(bh);
 187         jbd_unlock_bh_state(bh);
 188         put_bh(bh);
 189 }
 190
 191 /*
 192  * Clean up transaction's list of buffers submitted for io.
 193  * We wait for any pending IO to complete and remove any clean
 194  * buffers. Note that we take the buffers in the opposite ordering
 195  * from the one in which they were submitted for IO.
 196  *
 197  * Return 0 on success, and return <0 if some buffers have failed
 198  * to be written out.
 199  *
 200  * Called with j_list_lock held.
 201  */
 202 static int __wait_cp_io(journal_t *journal, transaction_t *transaction)
 203 {
 204         struct journal_head *jh;
 205         struct buffer_head *bh;
 206         tid_t this_tid;
 207         int released = 0;
 208         int ret = 0;
 209
 210         this_tid = transaction->t_tid;
 211 restart:
 212         /* Did somebody clean up the transaction in the meanwhile? */
 213         if (journal->j_checkpoint_transactions != transaction ||
 214                         transaction->t_tid != this_tid)
 215                 return ret;
 216         while (!released && transaction->t_checkpoint_io_list) {
 217                 jh = transaction->t_checkpoint_io_list;
 218                 bh = jh2bh(jh);
 219                 if (!jbd_trylock_bh_state(bh)) {
 220                         jbd_sync_bh(journal, bh);
 221                         spin_lock(&journal->j_list_lock);
 222                         goto restart;
 223                 }
 224                 if (buffer_locked(bh)) {
 225                         get_bh(bh);
 226                         spin_unlock(&journal->j_list_lock);
 227                         jbd_unlock_bh_state(bh);
 228                         wait_on_buffer(bh);
 229                         /* the journal_head may have gone by now */
 230                         BUFFER_TRACE(bh, "brelse");
 231                         __brelse(bh);
 232                         spin_lock(&journal->j_list_lock);
 233                         goto restart;
 234                 }
 235                 if (unlikely(buffer_write_io_error(bh)))
 236                         ret = -EIO;
 237
 238                 /*
 239                  * Now in whatever state the buffer currently is, we know that
 240                  * it has been written out and so we can drop it from the list
 241                  */
 242                 released = __journal_remove_checkpoint(jh);
 243                 jbd_unlock_bh_state(bh);
 244                 journal_remove_journal_head(bh);
 245                 __brelse(bh);
 246         }
 247
 248         return ret;
 249 }
 250
 251 #define NR_BATCH        64
 252
 253 static void
 254 __flush_batch(journal_t *journal, struct buffer_head **bhs, int *batch_count)
 255 {
 256         int i;
 257
 258         for (i = 0; i < *batch_count; i++)
 259                 write_dirty_buffer(bhs[i], WRITE);
 260
 261         for (i = 0; i < *batch_count; i++) {
 262                 struct buffer_head *bh = bhs[i];
 263                 clear_buffer_jwrite(bh);
 264                 BUFFER_TRACE(bh, "brelse");
 265                 __brelse(bh);
 266         }
 267         *batch_count = 0;
 268 }
 269
 270 /*
 271  * Try to flush one buffer from the checkpoint list to disk.
 272  *
 273  * Return 1 if something happened which requires us to abort the current
 274  * scan of the checkpoint list.  Return <0 if the buffer has failed to
 275  * be written out.
 276  *
 277  * Called with j_list_lock held and drops it if 1 is returned
 278  * Called under jbd_lock_bh_state(jh2bh(jh)), and drops it
 279  */
 280 static int __process_buffer(journal_t *journal, struct journal_head *jh,
 281                         struct buffer_head **bhs, int *batch_count)
 282 {
 283         struct buffer_head *bh = jh2bh(jh);
 284         int ret = 0;
 285
 286         if (buffer_locked(bh)) {
 287                 get_bh(bh);
 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                 spin_unlock(&journal->j_list_lock);
 300                 jbd_unlock_bh_state(bh);
 301                 log_start_commit(journal, tid);
 302                 log_wait_commit(journal, tid);
 303                 ret = 1;
 304         } else if (!buffer_dirty(bh)) {
 305                 ret = 1;
 306                 if (unlikely(buffer_write_io_error(bh)))
 307                         ret = -EIO;
 308                 J_ASSERT_JH(jh, !buffer_jbddirty(bh));
 309                 BUFFER_TRACE(bh, "remove from checkpoint");
 310                 __journal_remove_checkpoint(jh);
 311                 spin_unlock(&journal->j_list_lock);
 312                 jbd_unlock_bh_state(bh);
 313                 journal_remove_journal_head(bh);
 314                 __brelse(bh);
 315         } else {
 316                 /*
 317                  * Important: we are about to write the buffer, and
 318                  * possibly block, while still holding the journal lock.
 319                  * We cannot afford to let the transaction logic start
 320                  * messing around with this buffer before we write it to
 321                  * disk, as that would break recoverability.
 322                  */
 323                 BUFFER_TRACE(bh, "queue");
 324                 get_bh(bh);
 325                 J_ASSERT_BH(bh, !buffer_jwrite(bh));
 326                 set_buffer_jwrite(bh);
 327                 bhs[*batch_count] = bh;
 328                 __buffer_relink_io(jh);
 329                 jbd_unlock_bh_state(bh);
 330                 (*batch_count)++;
 331                 if (*batch_count == NR_BATCH) {
 332                         spin_unlock(&journal->j_list_lock);
 333                         __flush_batch(journal, bhs, batch_count);
 334                         ret = 1;
 335                 }
 336         }
 337         return ret;
 338 }
 339
 340 /*
 341  * Perform an actual checkpoint. We take the first transaction on the
 342  * list of transactions to be checkpointed and send all its buffers
 343  * to disk. We submit larger chunks of data at once.
 344  *
 345  * The journal should be locked before calling this function.
 346  * Called with j_checkpoint_mutex held.
 347  */
 348 int log_do_checkpoint(journal_t *journal)
 349 {
 350         transaction_t *transaction;
 351         tid_t this_tid;
 352         int result;
 353
 354         jbd_debug(1, "Start checkpoint\n");
 355
 356         /*
 357          * First thing: if there are any transactions in the log which
 358          * don't need checkpointing, just eliminate them from the
 359          * journal straight away.
 360          */
 361         result = cleanup_journal_tail(journal);
 362         trace_jbd_checkpoint(journal, result);
 363         jbd_debug(1, "cleanup_journal_tail returned %d\n", result);
 364         if (result <= 0)
 365                 return result;
 366
 367         /*
 368          * OK, we need to start writing disk blocks.  Take one transaction
 369          * and write it.
 370          */
 371         result = 0;
 372         spin_lock(&journal->j_list_lock);
 373         if (!journal->j_checkpoint_transactions)
 374                 goto out;
 375         transaction = journal->j_checkpoint_transactions;
 376         this_tid = transaction->t_tid;
 377 restart:
 378         /*
 379          * If someone cleaned up this transaction while we slept, we're
 380          * done (maybe it's a new transaction, but it fell at the same
 381          * address).
 382          */
 383         if (journal->j_checkpoint_transactions == transaction &&
 384                         transaction->t_tid == this_tid) {
 385                 int batch_count = 0;
 386                 struct buffer_head *bhs[NR_BATCH];
 387                 struct journal_head *jh;
 388                 int retry = 0, err;
 389
 390                 while (!retry && transaction->t_checkpoint_list) {
 391                         struct buffer_head *bh;
 392
 393                         jh = transaction->t_checkpoint_list;
 394                         bh = jh2bh(jh);
 395                         if (!jbd_trylock_bh_state(bh)) {
 396                                 jbd_sync_bh(journal, bh);
 397                                 retry = 1;
 398                                 break;
 399                         }
 400                         retry = __process_buffer(journal, jh, bhs,&batch_count);
 401                         if (retry < 0 && !result)
 402                                 result = retry;
 403                         if (!retry && (need_resched() ||
 404                                 spin_needbreak(&journal->j_list_lock))) {
 405                                 spin_unlock(&journal->j_list_lock);
 406                                 retry = 1;
 407                                 break;
 408                         }
 409                 }
 410
 411                 if (batch_count) {
 412                         if (!retry) {
 413                                 spin_unlock(&journal->j_list_lock);
 414                                 retry = 1;
 415                         }
 416                         __flush_batch(journal, bhs, &batch_count);
 417                 }
 418
 419                 if (retry) {
 420                         spin_lock(&journal->j_list_lock);
 421                         goto restart;
 422                 }
 423                 /*
 424                  * Now we have cleaned up the first transaction's checkpoint
 425                  * list. Let's clean up the second one
 426                  */
 427                 err = __wait_cp_io(journal, transaction);
 428                 if (!result)
 429                         result = err;
 430         }
 431 out:
 432         spin_unlock(&journal->j_list_lock);
 433         if (result < 0)
 434                 journal_abort(journal, result);
 435         else
 436                 result = cleanup_journal_tail(journal);
 437
 438         return (result < 0) ? result : 0;
 439 }
 440
 441 /*
 442  * Check the list of checkpoint transactions for the journal to see if
 443  * we have already got rid of any since the last update of the log tail
 444  * in the journal superblock.  If so, we can instantly roll the
 445  * superblock forward to remove those transactions from the log.
 446  *
 447  * Return <0 on error, 0 on success, 1 if there was nothing to clean up.
 448  *
 449  * Called with the journal lock held.
 450  *
 451  * This is the only part of the journaling code which really needs to be
 452  * aware of transaction aborts.  Checkpointing involves writing to the
 453  * main filesystem area rather than to the journal, so it can proceed
 454  * even in abort state, but we must not update the super block if
 455  * checkpointing may have failed.  Otherwise, we would lose some metadata
 456  * buffers which should be written-back to the filesystem.
 457  */
 458
 459 int cleanup_journal_tail(journal_t *journal)
 460 {
 461         transaction_t * transaction;
 462         tid_t           first_tid;
 463         unsigned int    blocknr, freed;
 464
 465         if (is_journal_aborted(journal))
 466                 return 1;
 467
 468         /* OK, work out the oldest transaction remaining in the log, and
 469          * the log block it starts at.
 470          *
 471          * If the log is now empty, we need to work out which is the
 472          * next transaction ID we will write, and where it will
 473          * start. */
 474
 475         spin_lock(&journal->j_state_lock);
 476         spin_lock(&journal->j_list_lock);
 477         transaction = journal->j_checkpoint_transactions;
 478         if (transaction) {
 479                 first_tid = transaction->t_tid;
 480                 blocknr = transaction->t_log_start;
 481         } else if ((transaction = journal->j_committing_transaction) != NULL) {
 482                 first_tid = transaction->t_tid;
 483                 blocknr = transaction->t_log_start;
 484         } else if ((transaction = journal->j_running_transaction) != NULL) {
 485                 first_tid = transaction->t_tid;
 486                 blocknr = journal->j_head;
 487         } else {
 488                 first_tid = journal->j_transaction_sequence;
 489                 blocknr = journal->j_head;
 490         }
 491         spin_unlock(&journal->j_list_lock);
 492         J_ASSERT(blocknr != 0);
 493
 494         /* If the oldest pinned transaction is at the tail of the log
 495            already then there's not much we can do right now. */
 496         if (journal->j_tail_sequence == first_tid) {
 497                 spin_unlock(&journal->j_state_lock);
 498                 return 1;
 499         }
 500
 501         /* OK, update the superblock to recover the freed space.
 502          * Physical blocks come first: have we wrapped beyond the end of
 503          * the log?  */
 504         freed = blocknr - journal->j_tail;
 505         if (blocknr < journal->j_tail)
 506                 freed = freed + journal->j_last - journal->j_first;
 507
 508         trace_jbd_cleanup_journal_tail(journal, first_tid, blocknr, freed);
 509         jbd_debug(1,
 510                   "Cleaning journal tail from %d to %d (offset %u), "
 511                   "freeing %u\n",
 512                   journal->j_tail_sequence, first_tid, blocknr, freed);
 513
 514         journal->j_free += freed;
 515         journal->j_tail_sequence = first_tid;
 516         journal->j_tail = blocknr;
 517         spin_unlock(&journal->j_state_lock);
 518         if (!(journal->j_flags & JFS_ABORT))
 519                 journal_update_superblock(journal, 1);
 520         return 0;
 521 }
 522
 523
 524 /* Checkpoint list management */
 525
 526 /*
 527  * journal_clean_one_cp_list
 528  *
 529  * Find all the written-back checkpoint buffers in the given list and release them.
 530  *
 531  * Called with the journal locked.
 532  * Called with j_list_lock held.
 533  * Returns number of bufers reaped (for debug)
 534  */
 535
 536 static int journal_clean_one_cp_list(struct journal_head *jh, int *released)
 537 {
 538         struct journal_head *last_jh;
 539         struct journal_head *next_jh = jh;
 540         int ret, freed = 0;
 541
 542         *released = 0;
 543         if (!jh)
 544                 return 0;
 545
 546         last_jh = jh->b_cpprev;
 547         do {
 548                 jh = next_jh;
 549                 next_jh = jh->b_cpnext;
 550                 /* Use trylock because of the ranking */
 551                 if (jbd_trylock_bh_state(jh2bh(jh))) {
 552                         ret = __try_to_free_cp_buf(jh);
 553                         if (ret) {
 554                                 freed++;
 555                                 if (ret == 2) {
 556                                         *released = 1;
 557                                         return freed;
 558                                 }
 559                         }
 560                 }
 561                 /*
 562                  * This function only frees up some memory
 563                  * if possible so we dont have an obligation
 564                  * to finish processing. Bail out if preemption
 565                  * requested:
 566                  */
 567                 if (need_resched())
 568                         return freed;
 569         } while (jh != last_jh);
 570
 571         return freed;
 572 }
 573
 574 /*
 575  * journal_clean_checkpoint_list
 576  *
 577  * Find all the written-back checkpoint buffers in the journal and release them.
 578  *
 579  * Called with the journal locked.
 580  * Called with j_list_lock held.
 581  * Returns number of buffers reaped (for debug)
 582  */
 583
 584 int __journal_clean_checkpoint_list(journal_t *journal)
 585 {
 586         transaction_t *transaction, *last_transaction, *next_transaction;
 587         int ret = 0;
 588         int released;
 589
 590         transaction = journal->j_checkpoint_transactions;
 591         if (!transaction)
 592                 goto out;
 593
 594         last_transaction = transaction->t_cpprev;
 595         next_transaction = transaction;
 596         do {
 597                 transaction = next_transaction;
 598                 next_transaction = transaction->t_cpnext;
 599                 ret += journal_clean_one_cp_list(transaction->
 600                                 t_checkpoint_list, &released);
 601                 /*
 602                  * This function only frees up some memory if possible so we
 603                  * dont have an obligation to finish processing. Bail out if
 604                  * preemption requested:
 605                  */
 606                 if (need_resched())
 607                         goto out;
 608                 if (released)
 609                         continue;
 610                 /*
 611                  * It is essential that we are as careful as in the case of
 612                  * t_checkpoint_list with removing the buffer from the list as
 613                  * we can possibly see not yet submitted buffers on io_list
 614                  */
 615                 ret += journal_clean_one_cp_list(transaction->
 616                                 t_checkpoint_io_list, &released);
 617                 if (need_resched())
 618                         goto out;
 619         } while (transaction != last_transaction);
 620 out:
 621         return ret;
 622 }
 623
 624 /*
 625  * journal_remove_checkpoint: called after a buffer has been committed
 626  * to disk (either by being write-back flushed to disk, or being
 627  * committed to the log).
 628  *
 629  * We cannot safely clean a transaction out of the log until all of the
 630  * buffer updates committed in that transaction have safely been stored
 631  * elsewhere on disk.  To achieve this, all of the buffers in a
 632  * transaction need to be maintained on the transaction's checkpoint
 633  * lists until they have been rewritten, at which point this function is
 634  * called to remove the buffer from the existing transaction's
 635  * checkpoint lists.
 636  *
 637  * The function returns 1 if it frees the transaction, 0 otherwise.
 638  *
 639  * This function is called with the journal locked.
 640  * This function is called with j_list_lock held.
 641  * This function is called with jbd_lock_bh_state(jh2bh(jh))
 642  */
 643
 644 int __journal_remove_checkpoint(struct journal_head *jh)
 645 {
 646         transaction_t *transaction;
 647         journal_t *journal;
 648         int ret = 0;
 649
 650         JBUFFER_TRACE(jh, "entry");
 651
 652         if ((transaction = jh->b_cp_transaction) == NULL) {
 653                 JBUFFER_TRACE(jh, "not on transaction");
 654                 goto out;
 655         }
 656         journal = transaction->t_journal;
 657
 658         __buffer_unlink(jh);
 659         jh->b_cp_transaction = NULL;
 660
 661         if (transaction->t_checkpoint_list != NULL ||
 662             transaction->t_checkpoint_io_list != NULL)
 663                 goto out;
 664         JBUFFER_TRACE(jh, "transaction has no more buffers");
 665
 666         /*
 667          * There is one special case to worry about: if we have just pulled the
 668          * buffer off a running or committing transaction's checkpoing list,
 669          * then even if the checkpoint list is empty, the transaction obviously
 670          * cannot be dropped!
 671          *
 672          * The locking here around t_state is a bit sleazy.
 673          * See the comment at the end of journal_commit_transaction().
 674          */
 675         if (transaction->t_state != T_FINISHED) {
 676                 JBUFFER_TRACE(jh, "belongs to running/committing transaction");
 677                 goto out;
 678         }
 679
 680         /* OK, that was the last buffer for the transaction: we can now
 681            safely remove this transaction from the log */
 682
 683         __journal_drop_transaction(journal, transaction);
 684
 685         /* Just in case anybody was waiting for more transactions to be
 686            checkpointed... */
 687         wake_up(&journal->j_wait_logspace);
 688         ret = 1;
 689 out:
 690         JBUFFER_TRACE(jh, "exit");
 691         return ret;
 692 }
 693
 694 /*
 695  * journal_insert_checkpoint: put a committed buffer onto a checkpoint
 696  * list so that we know when it is safe to clean the transaction out of
 697  * the log.
 698  *
 699  * Called with the journal locked.
 700  * Called with j_list_lock held.
 701  */
 702 void __journal_insert_checkpoint(struct journal_head *jh,
 703                                transaction_t *transaction)
 704 {
 705         JBUFFER_TRACE(jh, "entry");
 706         J_ASSERT_JH(jh, buffer_dirty(jh2bh(jh)) || buffer_jbddirty(jh2bh(jh)));
 707         J_ASSERT_JH(jh, jh->b_cp_transaction == NULL);
 708
 709         jh->b_cp_transaction = transaction;
 710
 711         if (!transaction->t_checkpoint_list) {
 712                 jh->b_cpnext = jh->b_cpprev = jh;
 713         } else {
 714                 jh->b_cpnext = transaction->t_checkpoint_list;
 715                 jh->b_cpprev = transaction->t_checkpoint_list->b_cpprev;
 716                 jh->b_cpprev->b_cpnext = jh;
 717                 jh->b_cpnext->b_cpprev = jh;
 718         }
 719         transaction->t_checkpoint_list = jh;
 720 }
 721
 722 /*
 723  * We've finished with this transaction structure: adios...
 724  *
 725  * The transaction must have no links except for the checkpoint by this
 726  * point.
 727  *
 728  * Called with the journal locked.
 729  * Called with j_list_lock held.
 730  */
 731
 732 void __journal_drop_transaction(journal_t *journal, transaction_t *transaction)
 733 {
 734         assert_spin_locked(&journal->j_list_lock);
 735         if (transaction->t_cpnext) {
 736                 transaction->t_cpnext->t_cpprev = transaction->t_cpprev;
 737                 transaction->t_cpprev->t_cpnext = transaction->t_cpnext;
 738                 if (journal->j_checkpoint_transactions == transaction)
 739                         journal->j_checkpoint_transactions =
 740                                 transaction->t_cpnext;
 741                 if (journal->j_checkpoint_transactions == transaction)
 742                         journal->j_checkpoint_transactions = NULL;
 743         }
 744
 745         J_ASSERT(transaction->t_state == T_FINISHED);
 746         J_ASSERT(transaction->t_buffers == NULL);
 747         J_ASSERT(transaction->t_sync_datalist == NULL);
 748         J_ASSERT(transaction->t_forget == NULL);
 749         J_ASSERT(transaction->t_iobuf_list == NULL);
 750         J_ASSERT(transaction->t_shadow_list == NULL);
 751         J_ASSERT(transaction->t_log_list == NULL);
 752         J_ASSERT(transaction->t_checkpoint_list == NULL);
 753         J_ASSERT(transaction->t_checkpoint_io_list == NULL);
 754         J_ASSERT(transaction->t_updates == 0);
 755         J_ASSERT(journal->j_committing_transaction != transaction);
 756         J_ASSERT(journal->j_running_transaction != transaction);
 757
 758         trace_jbd_drop_transaction(journal, transaction);
 759         jbd_debug(1, "Dropping transaction %d, all done\n", transaction->t_tid);
 760         kfree(transaction);
 761 }