fs/jbd2/checkpoint.c

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