fs/jbd/checkpoint.c

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