fs/jbd2/checkpoint.c

   1 /*
   2  * linux/fs/jbd2/checkpoint.c
   3  *
   4  * Written by Stephen C. Tweedie <sct@redhat.com>, 1999
   5  *
   6  * Copyright 1999 Red Hat Software --- All Rights Reserved
   7  *
   8  * This file is part of the Linux kernel and is made available under
   9  * the terms of the GNU General Public License, version 2, or at your
  10  * option, any later version, incorporated herein by reference.
  11  *
  12  * Checkpoint routines for the generic filesystem journaling code.
  13  * Part of the ext2fs journaling system.
  14  *
  15  * Checkpointing is the process of ensuring that a section of the log is
  16  * committed fully to disk, so that that portion of the log can be
  17  * reused.
  18  */
  19
  20 #include <linux/time.h>
  21 #include <linux/fs.h>
  22 #include <linux/jbd2.h>
  23 #include <linux/errno.h>
  24 #include <linux/slab.h>
  25 #include <linux/blkdev.h>
  26 #include <trace/events/jbd2.h>
  27
  28 /*
  29  * Unlink a buffer from a transaction checkpoint list.
  30  *
  31  * Called with j_list_lock held.
  32  */
  33 static inline void __buffer_unlink_first(struct journal_head *jh)
  34 {
  35         transaction_t *transaction = jh->b_cp_transaction;
  36
  37         jh->b_cpnext->b_cpprev = jh->b_cpprev;
  38         jh->b_cpprev->b_cpnext = jh->b_cpnext;
  39         if (transaction->t_checkpoint_list == jh) {
  40                 transaction->t_checkpoint_list = jh->b_cpnext;
  41                 if (transaction->t_checkpoint_list == jh)
  42                         transaction->t_checkpoint_list = NULL;
  43         }
  44 }
  45
  46 /*
  47  * Unlink a buffer from a transaction checkpoint(io) list.
  48  *
  49  * Called with j_list_lock held.
  50  */
  51 static inline void __buffer_unlink(struct journal_head *jh)
  52 {
  53         transaction_t *transaction = jh->b_cp_transaction;
  54
  55         __buffer_unlink_first(jh);
  56         if (transaction->t_checkpoint_io_list == jh) {
  57                 transaction->t_checkpoint_io_list = jh->b_cpnext;
  58                 if (transaction->t_checkpoint_io_list == jh)
  59                         transaction->t_checkpoint_io_list = NULL;
  60         }
  61 }
  62
  63 /*
  64  * Move a buffer from the checkpoint list to the checkpoint io list
  65  *
  66  * Called with j_list_lock held
  67  */
  68 static inline void __buffer_relink_io(struct journal_head *jh)
  69 {
  70         transaction_t *transaction = jh->b_cp_transaction;
  71
  72         __buffer_unlink_first(jh);
  73
  74         if (!transaction->t_checkpoint_io_list) {
  75                 jh->b_cpnext = jh->b_cpprev = jh;
  76         } else {
  77                 jh->b_cpnext = transaction->t_checkpoint_io_list;
  78                 jh->b_cpprev = transaction->t_checkpoint_io_list->b_cpprev;
  79                 jh->b_cpprev->b_cpnext = jh;
  80                 jh->b_cpnext->b_cpprev = jh;
  81         }
  82         transaction->t_checkpoint_io_list = jh;
  83 }
  84
  85 /*
  86  * Try to release a checkpointed buffer from its transaction.
  87  * Returns 1 if we released it and 2 if we also released the
  88  * whole transaction.
  89  *
  90  * Requires j_list_lock
  91  * Called under jbd_lock_bh_state(jh2bh(jh)), and drops it
  92  */
  93 static int __try_to_free_cp_buf(struct journal_head *jh)
  94 {
  95         int ret = 0;
  96         struct buffer_head *bh = jh2bh(jh);
  97
  98         if (jh->b_jlist == BJ_None && !buffer_locked(bh) &&
  99             !buffer_dirty(bh) && !buffer_write_io_error(bh)) {
 100                 /*
 101                  * Get our reference so that bh cannot be freed before
 102                  * we unlock it
 103                  */
 104                 get_bh(bh);
 105                 JBUFFER_TRACE(jh, "remove from checkpoint list");
 106                 ret = __jbd2_journal_remove_checkpoint(jh) + 1;
 107                 jbd_unlock_bh_state(bh);
 108                 BUFFER_TRACE(bh, "release");
 109                 __brelse(bh);
 110         } else {
 111                 jbd_unlock_bh_state(bh);
 112         }
 113         return ret;
 114 }
 115
 116 /*
 117  * __jbd2_log_wait_for_space: wait until there is space in the journal.
 118  *
 119  * Called under j-state_lock *only*.  It will be unlocked if we have to wait
 120  * for a checkpoint to free up some space in the log.
 121  */
 122 void __jbd2_log_wait_for_space(journal_t *journal)
 123 {
 124         int nblocks, space_left;
 125         /* assert_spin_locked(&journal->j_state_lock); */
 126
 127         nblocks = jbd_space_needed(journal);
 128         while (__jbd2_log_space_left(journal) < nblocks) {
 129                 if (journal->j_flags & JBD2_ABORT)
 130                         return;
 131                 write_unlock(&journal->j_state_lock);
 132                 mutex_lock(&journal->j_checkpoint_mutex);
 133
 134                 /*
 135                  * Test again, another process may have checkpointed while we
 136                  * were waiting for the checkpoint lock. If there are no
 137                  * transactions ready to be checkpointed, try to recover
 138                  * journal space by calling cleanup_journal_tail(), and if
 139                  * that doesn't work, by waiting for the currently committing
 140                  * transaction to complete.  If there is absolutely no way
 141                  * to make progress, this is either a BUG or corrupted
 142                  * filesystem, so abort the journal and leave a stack
 143                  * trace for forensic evidence.
 144                  */
 145                 write_lock(&journal->j_state_lock);
 146                 spin_lock(&journal->j_list_lock);
 147                 nblocks = jbd_space_needed(journal);
 148                 space_left = __jbd2_log_space_left(journal);
 149                 if (space_left < nblocks) {
 150                         int chkpt = journal->j_checkpoint_transactions != NULL;
 151                         tid_t tid = 0;
 152
 153                         if (journal->j_committing_transaction)
 154                                 tid = journal->j_committing_transaction->t_tid;
 155                         spin_unlock(&journal->j_list_lock);
 156                         write_unlock(&journal->j_state_lock);
 157                         if (chkpt) {
 158                                 jbd2_log_do_checkpoint(journal);
 159                         } else if (jbd2_cleanup_journal_tail(journal) == 0) {
 160                                 /* We were able to recover space; yay! */
 161                                 ;
 162                         } else if (tid) {
 163                                 jbd2_log_wait_commit(journal, tid);
 164                         } else {
 165                                 printk(KERN_ERR "%s: needed %d blocks and "
 166                                        "only had %d space available\n",
 167                                        __func__, nblocks, space_left);
 168                                 printk(KERN_ERR "%s: no way to get more "
 169                                        "journal space in %s\n", __func__,
 170                                        journal->j_devname);
 171                                 WARN_ON(1);
 172                                 jbd2_journal_abort(journal, 0);
 173                         }
 174                         write_lock(&journal->j_state_lock);
 175                 } else {
 176                         spin_unlock(&journal->j_list_lock);
 177                 }
 178                 mutex_unlock(&journal->j_checkpoint_mutex);
 179         }
 180 }
 181
 182 /*
 183  * We were unable to perform jbd_trylock_bh_state() inside j_list_lock.
 184  * The caller must restart a list walk.  Wait for someone else to run
 185  * jbd_unlock_bh_state().
 186  */
 187 static void jbd_sync_bh(journal_t *journal, struct buffer_head *bh)
 188         __releases(journal->j_list_lock)
 189 {
 190         get_bh(bh);
 191         spin_unlock(&journal->j_list_lock);
 192         jbd_lock_bh_state(bh);
 193         jbd_unlock_bh_state(bh);
 194         put_bh(bh);
 195 }
 196
 197 /*
 198  * Clean up transaction's list of buffers submitted for io.
 199  * We wait for any pending IO to complete and remove any clean
 200  * buffers. Note that we take the buffers in the opposite ordering
 201  * from the one in which they were submitted for IO.
 202  *
 203  * Return 0 on success, and return <0 if some buffers have failed
 204  * to be written out.
 205  *
 206  * Called with j_list_lock held.
 207  */
 208 static int __wait_cp_io(journal_t *journal, transaction_t *transaction)
 209 {
 210         struct journal_head *jh;
 211         struct buffer_head *bh;
 212         tid_t this_tid;
 213         int released = 0;
 214         int ret = 0;
 215
 216         this_tid = transaction->t_tid;
 217 restart:
 218         /* Did somebody clean up the transaction in the meanwhile? */
 219         if (journal->j_checkpoint_transactions != transaction ||
 220                         transaction->t_tid != this_tid)
 221                 return ret;
 222         while (!released && transaction->t_checkpoint_io_list) {
 223                 jh = transaction->t_checkpoint_io_list;
 224                 bh = jh2bh(jh);
 225                 if (!jbd_trylock_bh_state(bh)) {
 226                         jbd_sync_bh(journal, bh);
 227                         spin_lock(&journal->j_list_lock);
 228                         goto restart;
 229                 }
 230                 get_bh(bh);
 231                 if (buffer_locked(bh)) {
 232                         spin_unlock(&journal->j_list_lock);
 233                         jbd_unlock_bh_state(bh);
 234                         wait_on_buffer(bh);
 235                         /* the journal_head may have gone by now */
 236                         BUFFER_TRACE(bh, "brelse");
 237                         __brelse(bh);
 238                         spin_lock(&journal->j_list_lock);
 239                         goto restart;
 240                 }
 241                 if (unlikely(buffer_write_io_error(bh)))
 242                         ret = -EIO;
 243
 244                 /*
 245                  * Now in whatever state the buffer currently is, we know that
 246                  * it has been written out and so we can drop it from the list
 247                  */
 248                 released = __jbd2_journal_remove_checkpoint(jh);
 249                 jbd_unlock_bh_state(bh);
 250                 __brelse(bh);
 251         }
 252
 253         return ret;
 254 }
 255
 256 static void
 257 __flush_batch(journal_t *journal, int *batch_count)
 258 {
 259         int i;
 260
 261         for (i = 0; i < *batch_count; i++)
 262                 write_dirty_buffer(journal->j_chkpt_bhs[i], WRITE);
 263
 264         for (i = 0; i < *batch_count; i++) {
 265                 struct buffer_head *bh = journal->j_chkpt_bhs[i];
 266                 clear_buffer_jwrite(bh);
 267                 BUFFER_TRACE(bh, "brelse");
 268                 __brelse(bh);
 269         }
 270         *batch_count = 0;
 271 }
 272
 273 /*
 274  * Try to flush one buffer from the checkpoint list to disk.
 275  *
 276  * Return 1 if something happened which requires us to abort the current
 277  * scan of the checkpoint list.  Return <0 if the buffer has failed to
 278  * be written out.
 279  *
 280  * Called with j_list_lock held and drops it if 1 is returned
 281  * Called under jbd_lock_bh_state(jh2bh(jh)), and drops it
 282  */
 283 static int __process_buffer(journal_t *journal, struct journal_head *jh,
 284                             int *batch_count, transaction_t *transaction)
 285 {
 286         struct buffer_head *bh = jh2bh(jh);
 287         int ret = 0;
 288
 289         if (buffer_locked(bh)) {
 290                 get_bh(bh);
 291                 spin_unlock(&journal->j_list_lock);
 292                 jbd_unlock_bh_state(bh);
 293                 wait_on_buffer(bh);
 294                 /* the journal_head may have gone by now */
 295                 BUFFER_TRACE(bh, "brelse");
 296                 __brelse(bh);
 297                 ret = 1;
 298         } else if (jh->b_transaction != NULL) {
 299                 transaction_t *t = jh->b_transaction;
 300                 tid_t tid = t->t_tid;
 301
 302                 transaction->t_chp_stats.cs_forced_to_close++;
 303                 spin_unlock(&journal->j_list_lock);
 304                 jbd_unlock_bh_state(bh);
 305                 if (unlikely(journal->j_flags & JBD2_UNMOUNT))
 306                         /*
 307                          * The journal thread is dead; so starting and
 308                          * waiting for a commit to finish will cause
 309                          * us to wait for a _very_ long time.
 310                          */
 311                         printk(KERN_ERR "JBD2: %s: "
 312                                "Waiting for Godot: block %llu\n",
 313                                journal->j_devname,
 314                                (unsigned long long) bh->b_blocknr);
 315                 jbd2_log_start_commit(journal, tid);
 316                 jbd2_log_wait_commit(journal, tid);
 317                 ret = 1;
 318         } else if (!buffer_dirty(bh)) {
 319                 ret = 1;
 320                 if (unlikely(buffer_write_io_error(bh)))
 321                         ret = -EIO;
 322                 get_bh(bh);
 323                 J_ASSERT_JH(jh, !buffer_jbddirty(bh));
 324                 BUFFER_TRACE(bh, "remove from checkpoint");
 325                 __jbd2_journal_remove_checkpoint(jh);
 326                 spin_unlock(&journal->j_list_lock);
 327                 jbd_unlock_bh_state(bh);
 328                 __brelse(bh);
 329         } else {
 330                 /*
 331                  * Important: we are about to write the buffer, and
 332                  * possibly block, while still holding the journal lock.
 333                  * We cannot afford to let the transaction logic start
 334                  * messing around with this buffer before we write it to
 335                  * disk, as that would break recoverability.
 336                  */
 337                 BUFFER_TRACE(bh, "queue");
 338                 get_bh(bh);
 339                 J_ASSERT_BH(bh, !buffer_jwrite(bh));
 340                 set_buffer_jwrite(bh);
 341                 journal->j_chkpt_bhs[*batch_count] = bh;
 342                 __buffer_relink_io(jh);
 343                 jbd_unlock_bh_state(bh);
 344                 transaction->t_chp_stats.cs_written++;
 345                 (*batch_count)++;
 346                 if (*batch_count == JBD2_NR_BATCH) {
 347                         spin_unlock(&journal->j_list_lock);
 348                         __flush_batch(journal, batch_count);
 349                         ret = 1;
 350                 }
 351         }
 352         return ret;
 353 }
 354
 355 /*
 356  * Perform an actual checkpoint. We take the first transaction on the
 357  * list of transactions to be checkpointed and send all its buffers
 358  * to disk. We submit larger chunks of data at once.
 359  *
 360  * The journal should be locked before calling this function.
 361  * Called with j_checkpoint_mutex held.
 362  */
 363 int jbd2_log_do_checkpoint(journal_t *journal)
 364 {
 365         transaction_t *transaction;
 366         tid_t this_tid;
 367         int result;
 368
 369         jbd_debug(1, "Start checkpoint\n");
 370
 371         /*
 372          * First thing: if there are any transactions in the log which
 373          * don't need checkpointing, just eliminate them from the
 374          * journal straight away.
 375          */
 376         result = jbd2_cleanup_journal_tail(journal);
 377         trace_jbd2_checkpoint(journal, result);
 378         jbd_debug(1, "cleanup_journal_tail returned %d\n", result);
 379         if (result <= 0)
 380                 return result;
 381
 382         /*
 383          * OK, we need to start writing disk blocks.  Take one transaction
 384          * and write it.
 385          */
 386         result = 0;
 387         spin_lock(&journal->j_list_lock);
 388         if (!journal->j_checkpoint_transactions)
 389                 goto out;
 390         transaction = journal->j_checkpoint_transactions;
 391         if (transaction->t_chp_stats.cs_chp_time == 0)
 392                 transaction->t_chp_stats.cs_chp_time = jiffies;
 393         this_tid = transaction->t_tid;
 394 restart:
 395         /*
 396          * If someone cleaned up this transaction while we slept, we're
 397          * done (maybe it's a new transaction, but it fell at the same
 398          * address).
 399          */
 400         if (journal->j_checkpoint_transactions == transaction &&
 401                         transaction->t_tid == this_tid) {
 402                 int batch_count = 0;
 403                 struct journal_head *jh;
 404                 int retry = 0, err;
 405
 406                 while (!retry && transaction->t_checkpoint_list) {
 407                         struct buffer_head *bh;
 408
 409                         jh = transaction->t_checkpoint_list;
 410                         bh = jh2bh(jh);
 411                         if (!jbd_trylock_bh_state(bh)) {
 412                                 jbd_sync_bh(journal, bh);
 413                                 retry = 1;
 414                                 break;
 415                         }
 416                         retry = __process_buffer(journal, jh, &batch_count,
 417                                                  transaction);
 418                         if (retry < 0 && !result)
 419                                 result = retry;
 420                         if (!retry && (need_resched() ||
 421                                 spin_needbreak(&journal->j_list_lock))) {
 422                                 spin_unlock(&journal->j_list_lock);
 423                                 retry = 1;
 424                                 break;
 425                         }
 426                 }
 427
 428                 if (batch_count) {
 429                         if (!retry) {
 430                                 spin_unlock(&journal->j_list_lock);
 431                                 retry = 1;
 432                         }
 433                         __flush_batch(journal, &batch_count);
 434                 }
 435
 436                 if (retry) {
 437                         spin_lock(&journal->j_list_lock);
 438                         goto restart;
 439                 }
 440                 /*
 441                  * Now we have cleaned up the first transaction's checkpoint
 442                  * list. Let's clean up the second one
 443                  */
 444                 err = __wait_cp_io(journal, transaction);
 445                 if (!result)
 446                         result = err;
 447         }
 448 out:
 449         spin_unlock(&journal->j_list_lock);
 450         if (result < 0)
 451                 jbd2_journal_abort(journal, result);
 452         else
 453                 result = jbd2_cleanup_journal_tail(journal);
 454
 455         return (result < 0) ? result : 0;
 456 }
 457
 458 /*
 459  * Check the list of checkpoint transactions for the journal to see if
 460  * we have already got rid of any since the last update of the log tail
 461  * in the journal superblock.  If so, we can instantly roll the
 462  * superblock forward to remove those transactions from the log.
 463  *
 464  * Return <0 on error, 0 on success, 1 if there was nothing to clean up.
 465  *
 466  * Called with the journal lock held.
 467  *
 468  * This is the only part of the journaling code which really needs to be
 469  * aware of transaction aborts.  Checkpointing involves writing to the
 470  * main filesystem area rather than to the journal, so it can proceed
 471  * even in abort state, but we must not update the super block if
 472  * checkpointing may have failed.  Otherwise, we would lose some metadata
 473  * buffers which should be written-back to the filesystem.
 474  */
 475
 476 int jbd2_cleanup_journal_tail(journal_t *journal)
 477 {
 478         transaction_t * transaction;
 479         tid_t           first_tid;
 480         unsigned long   blocknr, freed;
 481
 482         if (is_journal_aborted(journal))
 483                 return 1;
 484
 485         /* OK, work out the oldest transaction remaining in the log, and
 486          * the log block it starts at.
 487          *
 488          * If the log is now empty, we need to work out which is the
 489          * next transaction ID we will write, and where it will
 490          * start. */
 491
 492         write_lock(&journal->j_state_lock);
 493         spin_lock(&journal->j_list_lock);
 494         transaction = journal->j_checkpoint_transactions;
 495         if (transaction) {
 496                 first_tid = transaction->t_tid;
 497                 blocknr = transaction->t_log_start;
 498         } else if ((transaction = journal->j_committing_transaction) != NULL) {
 499                 first_tid = transaction->t_tid;
 500                 blocknr = transaction->t_log_start;
 501         } else if ((transaction = journal->j_running_transaction) != NULL) {
 502                 first_tid = transaction->t_tid;
 503                 blocknr = journal->j_head;
 504         } else {
 505                 first_tid = journal->j_transaction_sequence;
 506                 blocknr = journal->j_head;
 507         }
 508         spin_unlock(&journal->j_list_lock);
 509         J_ASSERT(blocknr != 0);
 510
 511         /* If the oldest pinned transaction is at the tail of the log
 512            already then there's not much we can do right now. */
 513         if (journal->j_tail_sequence == first_tid) {
 514                 write_unlock(&journal->j_state_lock);
 515                 return 1;
 516         }
 517
 518         /* OK, update the superblock to recover the freed space.
 519          * Physical blocks come first: have we wrapped beyond the end of
 520          * the log?  */
 521         freed = blocknr - journal->j_tail;
 522         if (blocknr < journal->j_tail)
 523                 freed = freed + journal->j_last - journal->j_first;
 524
 525         trace_jbd2_cleanup_journal_tail(journal, first_tid, blocknr, freed);
 526         jbd_debug(1,
 527                   "Cleaning journal tail from %d to %d (offset %lu), "
 528                   "freeing %lu\n",
 529                   journal->j_tail_sequence, first_tid, blocknr, freed);
 530
 531         journal->j_free += freed;
 532         journal->j_tail_sequence = first_tid;
 533         journal->j_tail = blocknr;
 534         write_unlock(&journal->j_state_lock);
 535
 536         /*
 537          * If there is an external journal, we need to make sure that
 538          * any data blocks that were recently written out --- perhaps
 539          * by jbd2_log_do_checkpoint() --- are flushed out before we
 540          * drop the transactions from the external journal.  It's
 541          * unlikely this will be necessary, especially with a
 542          * appropriately sized journal, but we need this to guarantee
 543          * correctness.  Fortunately jbd2_cleanup_journal_tail()
 544          * doesn't get called all that often.
 545          */
 546         if ((journal->j_fs_dev != journal->j_dev) &&
 547             (journal->j_flags & JBD2_BARRIER))
 548                 blkdev_issue_flush(journal->j_fs_dev, GFP_KERNEL, NULL);
 549         if (!(journal->j_flags & JBD2_ABORT))
 550                 jbd2_journal_update_superblock(journal, 1);
 551         return 0;
 552 }
 553
 554
 555 /* Checkpoint list management */
 556
 557 /*
 558  * journal_clean_one_cp_list
 559  *
 560  * Find all the written-back checkpoint buffers in the given list and
 561  * release them.
 562  *
 563  * Called with the journal locked.
 564  * Called with j_list_lock held.
 565  * Returns number of bufers reaped (for debug)
 566  */
 567
 568 static int journal_clean_one_cp_list(struct journal_head *jh, int *released)
 569 {
 570         struct journal_head *last_jh;
 571         struct journal_head *next_jh = jh;
 572         int ret, freed = 0;
 573
 574         *released = 0;
 575         if (!jh)
 576                 return 0;
 577
 578         last_jh = jh->b_cpprev;
 579         do {
 580                 jh = next_jh;
 581                 next_jh = jh->b_cpnext;
 582                 /* Use trylock because of the ranking */
 583                 if (jbd_trylock_bh_state(jh2bh(jh))) {
 584                         ret = __try_to_free_cp_buf(jh);
 585                         if (ret) {
 586                                 freed++;
 587                                 if (ret == 2) {
 588                                         *released = 1;
 589                                         return freed;
 590                                 }
 591                         }
 592                 }
 593                 /*
 594                  * This function only frees up some memory
 595                  * if possible so we dont have an obligation
 596                  * to finish processing. Bail out if preemption
 597                  * requested:
 598                  */
 599                 if (need_resched())
 600                         return freed;
 601         } while (jh != last_jh);
 602
 603         return freed;
 604 }
 605
 606 /*
 607  * journal_clean_checkpoint_list
 608  *
 609  * Find all the written-back checkpoint buffers in the journal and release them.
 610  *
 611  * Called with the journal locked.
 612  * Called with j_list_lock held.
 613  * Returns number of buffers reaped (for debug)
 614  */
 615
 616 int __jbd2_journal_clean_checkpoint_list(journal_t *journal)
 617 {
 618         transaction_t *transaction, *last_transaction, *next_transaction;
 619         int ret = 0;
 620         int released;
 621
 622         transaction = journal->j_checkpoint_transactions;
 623         if (!transaction)
 624                 goto out;
 625
 626         last_transaction = transaction->t_cpprev;
 627         next_transaction = transaction;
 628         do {
 629                 transaction = next_transaction;
 630                 next_transaction = transaction->t_cpnext;
 631                 ret += journal_clean_one_cp_list(transaction->
 632                                 t_checkpoint_list, &released);
 633                 /*
 634                  * This function only frees up some memory if possible so we
 635                  * dont have an obligation to finish processing. Bail out if
 636                  * preemption requested:
 637                  */
 638                 if (need_resched())
 639                         goto out;
 640                 if (released)
 641                         continue;
 642                 /*
 643                  * It is essential that we are as careful as in the case of
 644                  * t_checkpoint_list with removing the buffer from the list as
 645                  * we can possibly see not yet submitted buffers on io_list
 646                  */
 647                 ret += journal_clean_one_cp_list(transaction->
 648                                 t_checkpoint_io_list, &released);
 649                 if (need_resched())
 650                         goto out;
 651         } while (transaction != last_transaction);
 652 out:
 653         return ret;
 654 }
 655
 656 /*
 657  * journal_remove_checkpoint: called after a buffer has been committed
 658  * to disk (either by being write-back flushed to disk, or being
 659  * committed to the log).
 660  *
 661  * We cannot safely clean a transaction out of the log until all of the
 662  * buffer updates committed in that transaction have safely been stored
 663  * elsewhere on disk.  To achieve this, all of the buffers in a
 664  * transaction need to be maintained on the transaction's checkpoint
 665  * lists until they have been rewritten, at which point this function is
 666  * called to remove the buffer from the existing transaction's
 667  * checkpoint lists.
 668  *
 669  * The function returns 1 if it frees the transaction, 0 otherwise.
 670  * The function can free jh and bh.
 671  *
 672  * This function is called with j_list_lock held.
 673  * This function is called with jbd_lock_bh_state(jh2bh(jh))
 674  */
 675
 676 int __jbd2_journal_remove_checkpoint(struct journal_head *jh)
 677 {
 678         struct transaction_chp_stats_s *stats;
 679         transaction_t *transaction;
 680         journal_t *journal;
 681         int ret = 0;
 682
 683         JBUFFER_TRACE(jh, "entry");
 684
 685         if ((transaction = jh->b_cp_transaction) == NULL) {
 686                 JBUFFER_TRACE(jh, "not on transaction");
 687                 goto out;
 688         }
 689         journal = transaction->t_journal;
 690
 691         JBUFFER_TRACE(jh, "removing from transaction");
 692         __buffer_unlink(jh);
 693         jh->b_cp_transaction = NULL;
 694         jbd2_journal_put_journal_head(jh);
 695
 696         if (transaction->t_checkpoint_list != NULL ||
 697             transaction->t_checkpoint_io_list != NULL)
 698                 goto out;
 699
 700         /*
 701          * There is one special case to worry about: if we have just pulled the
 702          * buffer off a running or committing transaction's checkpoing list,
 703          * then even if the checkpoint list is empty, the transaction obviously
 704          * cannot be dropped!
 705          *
 706          * The locking here around t_state is a bit sleazy.
 707          * See the comment at the end of jbd2_journal_commit_transaction().
 708          */
 709         if (transaction->t_state != T_FINISHED)
 710                 goto out;
 711
 712         /* OK, that was the last buffer for the transaction: we can now
 713            safely remove this transaction from the log */
 714         stats = &transaction->t_chp_stats;
 715         if (stats->cs_chp_time)
 716                 stats->cs_chp_time = jbd2_time_diff(stats->cs_chp_time,
 717                                                     jiffies);
 718         trace_jbd2_checkpoint_stats(journal->j_fs_dev->bd_dev,
 719                                     transaction->t_tid, stats);
 720
 721         __jbd2_journal_drop_transaction(journal, transaction);
 722         kfree(transaction);
 723
 724         /* Just in case anybody was waiting for more transactions to be
 725            checkpointed... */
 726         wake_up(&journal->j_wait_logspace);
 727         ret = 1;
 728 out:
 729         return ret;
 730 }
 731
 732 /*
 733  * journal_insert_checkpoint: put a committed buffer onto a checkpoint
 734  * list so that we know when it is safe to clean the transaction out of
 735  * the log.
 736  *
 737  * Called with the journal locked.
 738  * Called with j_list_lock held.
 739  */
 740 void __jbd2_journal_insert_checkpoint(struct journal_head *jh,
 741                                transaction_t *transaction)
 742 {
 743         JBUFFER_TRACE(jh, "entry");
 744         J_ASSERT_JH(jh, buffer_dirty(jh2bh(jh)) || buffer_jbddirty(jh2bh(jh)));
 745         J_ASSERT_JH(jh, jh->b_cp_transaction == NULL);
 746
 747         /* Get reference for checkpointing transaction */
 748         jbd2_journal_grab_journal_head(jh2bh(jh));
 749         jh->b_cp_transaction = transaction;
 750
 751         if (!transaction->t_checkpoint_list) {
 752                 jh->b_cpnext = jh->b_cpprev = jh;
 753         } else {
 754                 jh->b_cpnext = transaction->t_checkpoint_list;
 755                 jh->b_cpprev = transaction->t_checkpoint_list->b_cpprev;
 756                 jh->b_cpprev->b_cpnext = jh;
 757                 jh->b_cpnext->b_cpprev = jh;
 758         }
 759         transaction->t_checkpoint_list = jh;
 760 }
 761
 762 /*
 763  * We've finished with this transaction structure: adios...
 764  *
 765  * The transaction must have no links except for the checkpoint by this
 766  * point.
 767  *
 768  * Called with the journal locked.
 769  * Called with j_list_lock held.
 770  */
 771
 772 void __jbd2_journal_drop_transaction(journal_t *journal, transaction_t *transaction)
 773 {
 774         assert_spin_locked(&journal->j_list_lock);
 775         if (transaction->t_cpnext) {
 776                 transaction->t_cpnext->t_cpprev = transaction->t_cpprev;
 777                 transaction->t_cpprev->t_cpnext = transaction->t_cpnext;
 778                 if (journal->j_checkpoint_transactions == transaction)
 779                         journal->j_checkpoint_transactions =
 780                                 transaction->t_cpnext;
 781                 if (journal->j_checkpoint_transactions == transaction)
 782                         journal->j_checkpoint_transactions = NULL;
 783         }
 784
 785         J_ASSERT(transaction->t_state == T_FINISHED);
 786         J_ASSERT(transaction->t_buffers == NULL);
 787         J_ASSERT(transaction->t_forget == NULL);
 788         J_ASSERT(transaction->t_iobuf_list == NULL);
 789         J_ASSERT(transaction->t_shadow_list == NULL);
 790         J_ASSERT(transaction->t_log_list == NULL);
 791         J_ASSERT(transaction->t_checkpoint_list == NULL);
 792         J_ASSERT(transaction->t_checkpoint_io_list == NULL);
 793         J_ASSERT(atomic_read(&transaction->t_updates) == 0);
 794         J_ASSERT(journal->j_committing_transaction != transaction);
 795         J_ASSERT(journal->j_running_transaction != transaction);
 796
 797         jbd_debug(1, "Dropping transaction %d, all done\n", transaction->t_tid);
 798 }