fs/jbd2/checkpoint.c

   1 /*
   2  * linux/fs/jbd2/checkpoint.c
   3  *
   4  * Written by Stephen C. Tweedie <sct@redhat.com>, 1999
   5  *
   6  * Copyright 1999 Red Hat Software --- All Rights Reserved
   7  *
   8  * This file is part of the Linux kernel and is made available under
   9  * the terms of the GNU General Public License, version 2, or at your
  10  * option, any later version, incorporated herein by reference.
  11  *
  12  * Checkpoint routines for the generic filesystem journaling code.
  13  * Part of the ext2fs journaling system.
  14  *
  15  * Checkpointing is the process of ensuring that a section of the log is
  16  * committed fully to disk, so that that portion of the log can be
  17  * reused.
  18  */
  19
  20 #include <linux/time.h>
  21 #include <linux/fs.h>
  22 #include <linux/jbd2.h>
  23 #include <linux/errno.h>
  24 #include <linux/slab.h>
  25 #include <linux/blkdev.h>
  26 #include <trace/events/jbd2.h>
  27
  28 /*
  29  * Unlink a buffer from a transaction checkpoint list.
  30  *
  31  * Called with j_list_lock held.
  32  */
  33 static inline void __buffer_unlink_first(struct journal_head *jh)
  34 {
  35         transaction_t *transaction = jh->b_cp_transaction;
  36
  37         jh->b_cpnext->b_cpprev = jh->b_cpprev;
  38         jh->b_cpprev->b_cpnext = jh->b_cpnext;
  39         if (transaction->t_checkpoint_list == jh) {
  40                 transaction->t_checkpoint_list = jh->b_cpnext;
  41                 if (transaction->t_checkpoint_list == jh)
  42                         transaction->t_checkpoint_list = NULL;
  43         }
  44 }
  45
  46 /*
  47  * Unlink a buffer from a transaction checkpoint(io) list.
  48  *
  49  * Called with j_list_lock held.
  50  */
  51 static inline void __buffer_unlink(struct journal_head *jh)
  52 {
  53         transaction_t *transaction = jh->b_cp_transaction;
  54
  55         __buffer_unlink_first(jh);
  56         if (transaction->t_checkpoint_io_list == jh) {
  57                 transaction->t_checkpoint_io_list = jh->b_cpnext;
  58                 if (transaction->t_checkpoint_io_list == jh)
  59                         transaction->t_checkpoint_io_list = NULL;
  60         }
  61 }
  62
  63 /*
  64  * Move a buffer from the checkpoint list to the checkpoint io list
  65  *
  66  * Called with j_list_lock held
  67  */
  68 static inline void __buffer_relink_io(struct journal_head *jh)
  69 {
  70         transaction_t *transaction = jh->b_cp_transaction;
  71
  72         __buffer_unlink_first(jh);
  73
  74         if (!transaction->t_checkpoint_io_list) {
  75                 jh->b_cpnext = jh->b_cpprev = jh;
  76         } else {
  77                 jh->b_cpnext = transaction->t_checkpoint_io_list;
  78                 jh->b_cpprev = transaction->t_checkpoint_io_list->b_cpprev;
  79                 jh->b_cpprev->b_cpnext = jh;
  80                 jh->b_cpnext->b_cpprev = jh;
  81         }
  82         transaction->t_checkpoint_io_list = jh;
  83 }
  84
  85 /*
  86  * Try to release a checkpointed buffer from its transaction.
  87  * Returns 1 if we released it and 2 if we also released the
  88  * whole transaction.
  89  *
  90  * Requires j_list_lock
  91  * Called under jbd_lock_bh_state(jh2bh(jh)), and drops it
  92  */
  93 static int __try_to_free_cp_buf(struct journal_head *jh)
  94 {
  95         int ret = 0;
  96         struct buffer_head *bh = jh2bh(jh);
  97
  98         if (jh->b_jlist == BJ_None && !buffer_locked(bh) &&
  99             !buffer_dirty(bh) && !buffer_write_io_error(bh)) {
 100                 JBUFFER_TRACE(jh, "remove from checkpoint list");
 101                 ret = __jbd2_journal_remove_checkpoint(jh) + 1;
 102                 jbd_unlock_bh_state(bh);
 103                 jbd2_journal_remove_journal_head(bh);
 104                 BUFFER_TRACE(bh, "release");
 105                 __brelse(bh);
 106         } else {
 107                 jbd_unlock_bh_state(bh);
 108         }
 109         return ret;
 110 }
 111
 112 /*
 113  * __jbd2_log_wait_for_space: wait until there is space in the journal.
 114  *
 115  * Called under j-state_lock *only*.  It will be unlocked if we have to wait
 116  * for a checkpoint to free up some space in the log.
 117  */
 118 void __jbd2_log_wait_for_space(journal_t *journal)
 119 {
 120         int nblocks, space_left;
 121         /* assert_spin_locked(&journal->j_state_lock); */
 122
 123         nblocks = jbd_space_needed(journal);
 124         while (__jbd2_log_space_left(journal) < nblocks) {
 125                 if (journal->j_flags & JBD2_ABORT)
 126                         return;
 127                 write_unlock(&journal->j_state_lock);
 128                 mutex_lock(&journal->j_checkpoint_mutex);
 129
 130                 /*
 131                  * Test again, another process may have checkpointed while we
 132                  * were waiting for the checkpoint lock. If there are no
 133                  * transactions ready to be checkpointed, try to recover
 134                  * journal space by calling cleanup_journal_tail(), and if
 135                  * that doesn't work, by waiting for the currently committing
 136                  * transaction to complete.  If there is absolutely no way
 137                  * to make progress, this is either a BUG or corrupted
 138                  * filesystem, so abort the journal and leave a stack
 139                  * trace for forensic evidence.
 140                  */
 141                 write_lock(&journal->j_state_lock);
 142                 spin_lock(&journal->j_list_lock);
 143                 nblocks = jbd_space_needed(journal);
 144                 space_left = __jbd2_log_space_left(journal);
 145                 if (space_left < nblocks) {
 146                         int chkpt = journal->j_checkpoint_transactions != NULL;
 147                         tid_t tid = 0;
 148
 149                         if (journal->j_committing_transaction)
 150                                 tid = journal->j_committing_transaction->t_tid;
 151                         spin_unlock(&journal->j_list_lock);
 152                         write_unlock(&journal->j_state_lock);
 153                         if (chkpt) {
 154                                 jbd2_log_do_checkpoint(journal);
 155                         } else if (jbd2_cleanup_journal_tail(journal) == 0) {
 156                                 /* We were able to recover space; yay! */
 157                                 ;
 158                         } else if (tid) {
 159                                 jbd2_log_wait_commit(journal, tid);
 160                         } else {
 161                                 printk(KERN_ERR "%s: needed %d blocks and "
 162                                        "only had %d space available\n",
 163                                        __func__, nblocks, space_left);
 164                                 printk(KERN_ERR "%s: no way to get more "
 165                                        "journal space in %s\n", __func__,
 166                                        journal->j_devname);
 167                                 WARN_ON(1);
 168                                 jbd2_journal_abort(journal, 0);
 169                         }
 170                         write_lock(&journal->j_state_lock);
 171                 } else {
 172                         spin_unlock(&journal->j_list_lock);
 173                 }
 174                 mutex_unlock(&journal->j_checkpoint_mutex);
 175         }
 176 }
 177
 178 /*
 179  * We were unable to perform jbd_trylock_bh_state() inside j_list_lock.
 180  * The caller must restart a list walk.  Wait for someone else to run
 181  * jbd_unlock_bh_state().
 182  */
 183 static void jbd_sync_bh(journal_t *journal, struct buffer_head *bh)
 184         __releases(journal->j_list_lock)
 185 {
 186         get_bh(bh);
 187         spin_unlock(&journal->j_list_lock);
 188         jbd_lock_bh_state(bh);
 189         jbd_unlock_bh_state(bh);
 190         put_bh(bh);
 191 }
 192
 193 /*
 194  * Clean up transaction's list of buffers submitted for io.
 195  * We wait for any pending IO to complete and remove any clean
 196  * buffers. Note that we take the buffers in the opposite ordering
 197  * from the one in which they were submitted for IO.
 198  *
 199  * Return 0 on success, and return <0 if some buffers have failed
 200  * to be written out.
 201  *
 202  * Called with j_list_lock held.
 203  */
 204 static int __wait_cp_io(journal_t *journal, transaction_t *transaction)
 205 {
 206         struct journal_head *jh;
 207         struct buffer_head *bh;
 208         tid_t this_tid;
 209         int released = 0;
 210         int ret = 0;
 211
 212         this_tid = transaction->t_tid;
 213 restart:
 214         /* Did somebody clean up the transaction in the meanwhile? */
 215         if (journal->j_checkpoint_transactions != transaction ||
 216                         transaction->t_tid != this_tid)
 217                 return ret;
 218         while (!released && transaction->t_checkpoint_io_list) {
 219                 jh = transaction->t_checkpoint_io_list;
 220                 bh = jh2bh(jh);
 221                 if (!jbd_trylock_bh_state(bh)) {
 222                         jbd_sync_bh(journal, bh);
 223                         spin_lock(&journal->j_list_lock);
 224                         goto restart;
 225                 }
 226                 if (buffer_locked(bh)) {
 227                         atomic_inc(&bh->b_count);
 228                         spin_unlock(&journal->j_list_lock);
 229                         jbd_unlock_bh_state(bh);
 230                         wait_on_buffer(bh);
 231                         /* the journal_head may have gone by now */
 232                         BUFFER_TRACE(bh, "brelse");
 233                         __brelse(bh);
 234                         spin_lock(&journal->j_list_lock);
 235                         goto restart;
 236                 }
 237                 if (unlikely(buffer_write_io_error(bh)))
 238                         ret = -EIO;
 239
 240                 /*
 241                  * Now in whatever state the buffer currently is, we know that
 242                  * it has been written out and so we can drop it from the list
 243                  */
 244                 released = __jbd2_journal_remove_checkpoint(jh);
 245                 jbd_unlock_bh_state(bh);
 246                 jbd2_journal_remove_journal_head(bh);
 247                 __brelse(bh);
 248         }
 249
 250         return ret;
 251 }
 252
 253 static void
 254 __flush_batch(journal_t *journal, int *batch_count)
 255 {
 256         int i;
 257
 258         for (i = 0; i < *batch_count; i++)
 259                 write_dirty_buffer(journal->j_chkpt_bhs[i], WRITE);
 260
 261         for (i = 0; i < *batch_count; i++) {
 262                 struct buffer_head *bh = journal->j_chkpt_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                             int *batch_count, transaction_t *transaction)
 282 {
 283         struct buffer_head *bh = jh2bh(jh);
 284         int ret = 0;
 285
 286         if (buffer_locked(bh)) {
 287                 atomic_inc(&bh->b_count);
 288                 spin_unlock(&journal->j_list_lock);
 289                 jbd_unlock_bh_state(bh);
 290                 wait_on_buffer(bh);
 291                 /* the journal_head may have gone by now */
 292                 BUFFER_TRACE(bh, "brelse");
 293                 __brelse(bh);
 294                 ret = 1;
 295         } else if (jh->b_transaction != NULL) {
 296                 transaction_t *t = jh->b_transaction;
 297                 tid_t tid = t->t_tid;
 298
 299                 transaction->t_chp_stats.cs_forced_to_close++;
 300                 spin_unlock(&journal->j_list_lock);
 301                 jbd_unlock_bh_state(bh);
 302                 jbd2_log_start_commit(journal, tid);
 303                 jbd2_log_wait_commit(journal, tid);
 304                 ret = 1;
 305         } else if (!buffer_dirty(bh)) {
 306                 ret = 1;
 307                 if (unlikely(buffer_write_io_error(bh)))
 308                         ret = -EIO;
 309                 J_ASSERT_JH(jh, !buffer_jbddirty(bh));
 310                 BUFFER_TRACE(bh, "remove from checkpoint");
 311                 __jbd2_journal_remove_checkpoint(jh);
 312                 spin_unlock(&journal->j_list_lock);
 313                 jbd_unlock_bh_state(bh);
 314                 jbd2_journal_remove_journal_head(bh);
 315                 __brelse(bh);
 316         } else {
 317                 /*
 318                  * Important: we are about to write the buffer, and
 319                  * possibly block, while still holding the journal lock.
 320                  * We cannot afford to let the transaction logic start
 321                  * messing around with this buffer before we write it to
 322                  * disk, as that would break recoverability.
 323                  */
 324                 BUFFER_TRACE(bh, "queue");
 325                 get_bh(bh);
 326                 J_ASSERT_BH(bh, !buffer_jwrite(bh));
 327                 set_buffer_jwrite(bh);
 328                 journal->j_chkpt_bhs[*batch_count] = bh;
 329                 __buffer_relink_io(jh);
 330                 jbd_unlock_bh_state(bh);
 331                 transaction->t_chp_stats.cs_written++;
 332                 (*batch_count)++;
 333                 if (*batch_count == JBD2_NR_BATCH) {
 334                         spin_unlock(&journal->j_list_lock);
 335                         __flush_batch(journal, batch_count);
 336                         ret = 1;
 337                 }
 338         }
 339         return ret;
 340 }
 341
 342 /*
 343  * Perform an actual checkpoint. We take the first transaction on the
 344  * list of transactions to be checkpointed and send all its buffers
 345  * to disk. We submit larger chunks of data at once.
 346  *
 347  * The journal should be locked before calling this function.
 348  * Called with j_checkpoint_mutex held.
 349  */
 350 int jbd2_log_do_checkpoint(journal_t *journal)
 351 {
 352         transaction_t *transaction;
 353         tid_t this_tid;
 354         int result;
 355
 356         jbd_debug(1, "Start checkpoint\n");
 357
 358         /*
 359          * First thing: if there are any transactions in the log which
 360          * don't need checkpointing, just eliminate them from the
 361          * journal straight away.
 362          */
 363         result = jbd2_cleanup_journal_tail(journal);
 364         trace_jbd2_checkpoint(journal, result);
 365         jbd_debug(1, "cleanup_journal_tail returned %d\n", result);
 366         if (result <= 0)
 367                 return result;
 368
 369         /*
 370          * OK, we need to start writing disk blocks.  Take one transaction
 371          * and write it.
 372          */
 373         result = 0;
 374         spin_lock(&journal->j_list_lock);
 375         if (!journal->j_checkpoint_transactions)
 376                 goto out;
 377         transaction = journal->j_checkpoint_transactions;
 378         if (transaction->t_chp_stats.cs_chp_time == 0)
 379                 transaction->t_chp_stats.cs_chp_time = jiffies;
 380         this_tid = transaction->t_tid;
 381 restart:
 382         /*
 383          * If someone cleaned up this transaction while we slept, we're
 384          * done (maybe it's a new transaction, but it fell at the same
 385          * address).
 386          */
 387         if (journal->j_checkpoint_transactions == transaction &&
 388                         transaction->t_tid == this_tid) {
 389                 int batch_count = 0;
 390                 struct journal_head *jh;
 391                 int retry = 0, err;
 392
 393                 while (!retry && transaction->t_checkpoint_list) {
 394                         struct buffer_head *bh;
 395
 396                         jh = transaction->t_checkpoint_list;
 397                         bh = jh2bh(jh);
 398                         if (!jbd_trylock_bh_state(bh)) {
 399                                 jbd_sync_bh(journal, bh);
 400                                 retry = 1;
 401                                 break;
 402                         }
 403                         retry = __process_buffer(journal, jh, &batch_count,
 404                                                  transaction);
 405                         if (retry < 0 && !result)
 406                                 result = retry;
 407                         if (!retry && (need_resched() ||
 408                                 spin_needbreak(&journal->j_list_lock))) {
 409                                 spin_unlock(&journal->j_list_lock);
 410                                 retry = 1;
 411                                 break;
 412                         }
 413                 }
 414
 415                 if (batch_count) {
 416                         if (!retry) {
 417                                 spin_unlock(&journal->j_list_lock);
 418                                 retry = 1;
 419                         }
 420                         __flush_batch(journal, &batch_count);
 421                 }
 422
 423                 if (retry) {
 424                         spin_lock(&journal->j_list_lock);
 425                         goto restart;
 426                 }
 427                 /*
 428                  * Now we have cleaned up the first transaction's checkpoint
 429                  * list. Let's clean up the second one
 430                  */
 431                 err = __wait_cp_io(journal, transaction);
 432                 if (!result)
 433                         result = err;
 434         }
 435 out:
 436         spin_unlock(&journal->j_list_lock);
 437         if (result < 0)
 438                 jbd2_journal_abort(journal, result);
 439         else
 440                 result = jbd2_cleanup_journal_tail(journal);
 441
 442         return (result < 0) ? result : 0;
 443 }
 444
 445 /*
 446  * Check the list of checkpoint transactions for the journal to see if
 447  * we have already got rid of any since the last update of the log tail
 448  * in the journal superblock.  If so, we can instantly roll the
 449  * superblock forward to remove those transactions from the log.
 450  *
 451  * Return <0 on error, 0 on success, 1 if there was nothing to clean up.
 452  *
 453  * Called with the journal lock held.
 454  *
 455  * This is the only part of the journaling code which really needs to be
 456  * aware of transaction aborts.  Checkpointing involves writing to the
 457  * main filesystem area rather than to the journal, so it can proceed
 458  * even in abort state, but we must not update the super block if
 459  * checkpointing may have failed.  Otherwise, we would lose some metadata
 460  * buffers which should be written-back to the filesystem.
 461  */
 462
 463 int jbd2_cleanup_journal_tail(journal_t *journal)
 464 {
 465         transaction_t * transaction;
 466         tid_t           first_tid;
 467         unsigned long   blocknr, freed;
 468
 469         if (is_journal_aborted(journal))
 470                 return 1;
 471
 472         /* OK, work out the oldest transaction remaining in the log, and
 473          * the log block it starts at.
 474          *
 475          * If the log is now empty, we need to work out which is the
 476          * next transaction ID we will write, and where it will
 477          * start. */
 478
 479         write_lock(&journal->j_state_lock);
 480         spin_lock(&journal->j_list_lock);
 481         transaction = journal->j_checkpoint_transactions;
 482         if (transaction) {
 483                 first_tid = transaction->t_tid;
 484                 blocknr = transaction->t_log_start;
 485         } else if ((transaction = journal->j_committing_transaction) != NULL) {
 486                 first_tid = transaction->t_tid;
 487                 blocknr = transaction->t_log_start;
 488         } else if ((transaction = journal->j_running_transaction) != NULL) {
 489                 first_tid = transaction->t_tid;
 490                 blocknr = journal->j_head;
 491         } else {
 492                 first_tid = journal->j_transaction_sequence;
 493                 blocknr = journal->j_head;
 494         }
 495         spin_unlock(&journal->j_list_lock);
 496         J_ASSERT(blocknr != 0);
 497
 498         /* If the oldest pinned transaction is at the tail of the log
 499            already then there's not much we can do right now. */
 500         if (journal->j_tail_sequence == first_tid) {
 501                 write_unlock(&journal->j_state_lock);
 502                 return 1;
 503         }
 504
 505         /* OK, update the superblock to recover the freed space.
 506          * Physical blocks come first: have we wrapped beyond the end of
 507          * the log?  */
 508         freed = blocknr - journal->j_tail;
 509         if (blocknr < journal->j_tail)
 510                 freed = freed + journal->j_last - journal->j_first;
 511
 512         trace_jbd2_cleanup_journal_tail(journal, first_tid, blocknr, freed);
 513         jbd_debug(1,
 514                   "Cleaning journal tail from %d to %d (offset %lu), "
 515                   "freeing %lu\n",
 516                   journal->j_tail_sequence, first_tid, blocknr, freed);
 517
 518         journal->j_free += freed;
 519         journal->j_tail_sequence = first_tid;
 520         journal->j_tail = blocknr;
 521         write_unlock(&journal->j_state_lock);
 522
 523         /*
 524          * If there is an external journal, we need to make sure that
 525          * any data blocks that were recently written out --- perhaps
 526          * by jbd2_log_do_checkpoint() --- are flushed out before we
 527          * drop the transactions from the external journal.  It's
 528          * unlikely this will be necessary, especially with a
 529          * appropriately sized journal, but we need this to guarantee
 530          * correctness.  Fortunately jbd2_cleanup_journal_tail()
 531          * doesn't get called all that often.
 532          */
 533         if ((journal->j_fs_dev != journal->j_dev) &&
 534             (journal->j_flags & JBD2_BARRIER))
 535                 blkdev_issue_flush(journal->j_fs_dev, GFP_KERNEL, NULL);
 536         if (!(journal->j_flags & JBD2_ABORT))
 537                 jbd2_journal_update_superblock(journal, 1);
 538         return 0;
 539 }
 540
 541
 542 /* Checkpoint list management */
 543
 544 /*
 545  * journal_clean_one_cp_list
 546  *
 547  * Find all the written-back checkpoint buffers in the given list and release them.
 548  *
 549  * Called with the journal locked.
 550  * Called with j_list_lock held.
 551  * Returns number of bufers reaped (for debug)
 552  */
 553
 554 static int journal_clean_one_cp_list(struct journal_head *jh, int *released)
 555 {
 556         struct journal_head *last_jh;
 557         struct journal_head *next_jh = jh;
 558         int ret, freed = 0;
 559
 560         *released = 0;
 561         if (!jh)
 562                 return 0;
 563
 564         last_jh = jh->b_cpprev;
 565         do {
 566                 jh = next_jh;
 567                 next_jh = jh->b_cpnext;
 568                 /* Use trylock because of the ranking */
 569                 if (jbd_trylock_bh_state(jh2bh(jh))) {
 570                         ret = __try_to_free_cp_buf(jh);
 571                         if (ret) {
 572                                 freed++;
 573                                 if (ret == 2) {
 574                                         *released = 1;
 575                                         return freed;
 576                                 }
 577                         }
 578                 }
 579                 /*
 580                  * This function only frees up some memory
 581                  * if possible so we dont have an obligation
 582                  * to finish processing. Bail out if preemption
 583                  * requested:
 584                  */
 585                 if (need_resched())
 586                         return freed;
 587         } while (jh != last_jh);
 588
 589         return freed;
 590 }
 591
 592 /*
 593  * journal_clean_checkpoint_list
 594  *
 595  * Find all the written-back checkpoint buffers in the journal and release them.
 596  *
 597  * Called with the journal locked.
 598  * Called with j_list_lock held.
 599  * Returns number of buffers reaped (for debug)
 600  */
 601
 602 int __jbd2_journal_clean_checkpoint_list(journal_t *journal)
 603 {
 604         transaction_t *transaction, *last_transaction, *next_transaction;
 605         int ret = 0;
 606         int released;
 607
 608         transaction = journal->j_checkpoint_transactions;
 609         if (!transaction)
 610                 goto out;
 611
 612         last_transaction = transaction->t_cpprev;
 613         next_transaction = transaction;
 614         do {
 615                 transaction = next_transaction;
 616                 next_transaction = transaction->t_cpnext;
 617                 ret += journal_clean_one_cp_list(transaction->
 618                                 t_checkpoint_list, &released);
 619                 /*
 620                  * This function only frees up some memory if possible so we
 621                  * dont have an obligation to finish processing. Bail out if
 622                  * preemption requested:
 623                  */
 624                 if (need_resched())
 625                         goto out;
 626                 if (released)
 627                         continue;
 628                 /*
 629                  * It is essential that we are as careful as in the case of
 630                  * t_checkpoint_list with removing the buffer from the list as
 631                  * we can possibly see not yet submitted buffers on io_list
 632                  */
 633                 ret += journal_clean_one_cp_list(transaction->
 634                                 t_checkpoint_io_list, &released);
 635                 if (need_resched())
 636                         goto out;
 637         } while (transaction != last_transaction);
 638 out:
 639         return ret;
 640 }
 641
 642 /*
 643  * journal_remove_checkpoint: called after a buffer has been committed
 644  * to disk (either by being write-back flushed to disk, or being
 645  * committed to the log).
 646  *
 647  * We cannot safely clean a transaction out of the log until all of the
 648  * buffer updates committed in that transaction have safely been stored
 649  * elsewhere on disk.  To achieve this, all of the buffers in a
 650  * transaction need to be maintained on the transaction's checkpoint
 651  * lists until they have been rewritten, at which point this function is
 652  * called to remove the buffer from the existing transaction's
 653  * checkpoint lists.
 654  *
 655  * The function returns 1 if it frees the transaction, 0 otherwise.
 656  *
 657  * This function is called with the journal locked.
 658  * This function is called with j_list_lock held.
 659  * This function is called with jbd_lock_bh_state(jh2bh(jh))
 660  */
 661
 662 int __jbd2_journal_remove_checkpoint(struct journal_head *jh)
 663 {
 664         struct transaction_chp_stats_s *stats;
 665         transaction_t *transaction;
 666         journal_t *journal;
 667         int ret = 0;
 668
 669         JBUFFER_TRACE(jh, "entry");
 670
 671         if ((transaction = jh->b_cp_transaction) == NULL) {
 672                 JBUFFER_TRACE(jh, "not on transaction");
 673                 goto out;
 674         }
 675         journal = transaction->t_journal;
 676
 677         __buffer_unlink(jh);
 678         jh->b_cp_transaction = NULL;
 679
 680         if (transaction->t_checkpoint_list != NULL ||
 681             transaction->t_checkpoint_io_list != NULL)
 682                 goto out;
 683         JBUFFER_TRACE(jh, "transaction has no more buffers");
 684
 685         /*
 686          * There is one special case to worry about: if we have just pulled the
 687          * buffer off a running or committing transaction's checkpoing list,
 688          * then even if the checkpoint list is empty, the transaction obviously
 689          * cannot be dropped!
 690          *
 691          * The locking here around t_state is a bit sleazy.
 692          * See the comment at the end of jbd2_journal_commit_transaction().
 693          */
 694         if (transaction->t_state != T_FINISHED) {
 695                 JBUFFER_TRACE(jh, "belongs to running/committing transaction");
 696                 goto out;
 697         }
 698
 699         /* OK, that was the last buffer for the transaction: we can now
 700            safely remove this transaction from the log */
 701         stats = &transaction->t_chp_stats;
 702         if (stats->cs_chp_time)
 703                 stats->cs_chp_time = jbd2_time_diff(stats->cs_chp_time,
 704                                                     jiffies);
 705         trace_jbd2_checkpoint_stats(journal->j_fs_dev->bd_dev,
 706                                     transaction->t_tid, stats);
 707
 708         __jbd2_journal_drop_transaction(journal, transaction);
 709         kfree(transaction);
 710
 711         /* Just in case anybody was waiting for more transactions to be
 712            checkpointed... */
 713         wake_up(&journal->j_wait_logspace);
 714         ret = 1;
 715 out:
 716         JBUFFER_TRACE(jh, "exit");
 717         return ret;
 718 }
 719
 720 /*
 721  * journal_insert_checkpoint: put a committed buffer onto a checkpoint
 722  * list so that we know when it is safe to clean the transaction out of
 723  * the log.
 724  *
 725  * Called with the journal locked.
 726  * Called with j_list_lock held.
 727  */
 728 void __jbd2_journal_insert_checkpoint(struct journal_head *jh,
 729                                transaction_t *transaction)
 730 {
 731         JBUFFER_TRACE(jh, "entry");
 732         J_ASSERT_JH(jh, buffer_dirty(jh2bh(jh)) || buffer_jbddirty(jh2bh(jh)));
 733         J_ASSERT_JH(jh, jh->b_cp_transaction == NULL);
 734
 735         jh->b_cp_transaction = transaction;
 736
 737         if (!transaction->t_checkpoint_list) {
 738                 jh->b_cpnext = jh->b_cpprev = jh;
 739         } else {
 740                 jh->b_cpnext = transaction->t_checkpoint_list;
 741                 jh->b_cpprev = transaction->t_checkpoint_list->b_cpprev;
 742                 jh->b_cpprev->b_cpnext = jh;
 743                 jh->b_cpnext->b_cpprev = jh;
 744         }
 745         transaction->t_checkpoint_list = jh;
 746 }
 747
 748 /*
 749  * We've finished with this transaction structure: adios...
 750  *
 751  * The transaction must have no links except for the checkpoint by this
 752  * point.
 753  *
 754  * Called with the journal locked.
 755  * Called with j_list_lock held.
 756  */
 757
 758 void __jbd2_journal_drop_transaction(journal_t *journal, transaction_t *transaction)
 759 {
 760         assert_spin_locked(&journal->j_list_lock);
 761         if (transaction->t_cpnext) {
 762                 transaction->t_cpnext->t_cpprev = transaction->t_cpprev;
 763                 transaction->t_cpprev->t_cpnext = transaction->t_cpnext;
 764                 if (journal->j_checkpoint_transactions == transaction)
 765                         journal->j_checkpoint_transactions =
 766                                 transaction->t_cpnext;
 767                 if (journal->j_checkpoint_transactions == transaction)
 768                         journal->j_checkpoint_transactions = NULL;
 769         }
 770
 771         J_ASSERT(transaction->t_state == T_FINISHED);
 772         J_ASSERT(transaction->t_buffers == NULL);
 773         J_ASSERT(transaction->t_forget == NULL);
 774         J_ASSERT(transaction->t_iobuf_list == NULL);
 775         J_ASSERT(transaction->t_shadow_list == NULL);
 776         J_ASSERT(transaction->t_log_list == NULL);
 777         J_ASSERT(transaction->t_checkpoint_list == NULL);
 778         J_ASSERT(transaction->t_checkpoint_io_list == NULL);
 779         J_ASSERT(atomic_read(&transaction->t_updates) == 0);
 780         J_ASSERT(journal->j_committing_transaction != transaction);
 781         J_ASSERT(journal->j_running_transaction != transaction);
 782
 783         jbd_debug(1, "Dropping transaction %d, all done\n", transaction->t_tid);
 784 }