fs/jbd/commit.c

   1 /*
   2  * linux/fs/commit.c
   3  *
   4  * Written by Stephen C. Tweedie <sct@redhat.com>, 1998
   5  *
   6  * Copyright 1998 Red Hat corp --- 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  * Journal commit routines for the generic filesystem journaling code;
  13  * part of the ext2fs journaling system.
  14  */
  15
  16 #include <linux/time.h>
  17 #include <linux/fs.h>
  18 #include <linux/jbd.h>
  19 #include <linux/errno.h>
  20 #include <linux/slab.h>
  21 #include <linux/mm.h>
  22 #include <linux/pagemap.h>
  23 #include <linux/smp_lock.h>
  24
  25 /*
  26  * Default IO end handler for temporary BJ_IO buffer_heads.
  27  */
  28 static void journal_end_buffer_io_sync(struct buffer_head *bh, int uptodate)
  29 {
  30         BUFFER_TRACE(bh, "");
  31         if (uptodate)
  32                 set_buffer_uptodate(bh);
  33         else
  34                 clear_buffer_uptodate(bh);
  35         unlock_buffer(bh);
  36 }
  37
  38 /*
  39  * When an ext3-ordered file is truncated, it is possible that many pages are
  40  * not sucessfully freed, because they are attached to a committing transaction.
  41  * After the transaction commits, these pages are left on the LRU, with no
  42  * ->mapping, and with attached buffers.  These pages are trivially reclaimable
  43  * by the VM, but their apparent absence upsets the VM accounting, and it makes
  44  * the numbers in /proc/meminfo look odd.
  45  *
  46  * So here, we have a buffer which has just come off the forget list.  Look to
  47  * see if we can strip all buffers from the backing page.
  48  *
  49  * Called under lock_journal(), and possibly under journal_datalist_lock.  The
  50  * caller provided us with a ref against the buffer, and we drop that here.
  51  */
  52 static void release_buffer_page(struct buffer_head *bh)
  53 {
  54         struct page *page;
  55
  56         if (buffer_dirty(bh))
  57                 goto nope;
  58         if (atomic_read(&bh->b_count) != 1)
  59                 goto nope;
  60         page = bh->b_page;
  61         if (!page)
  62                 goto nope;
  63         if (page->mapping)
  64                 goto nope;
  65
  66         /* OK, it's a truncated page */
  67         if (TestSetPageLocked(page))
  68                 goto nope;
  69
  70         page_cache_get(page);
  71         __brelse(bh);
  72         try_to_free_buffers(page);
  73         unlock_page(page);
  74         page_cache_release(page);
  75         return;
  76
  77 nope:
  78         __brelse(bh);
  79 }
  80
  81 /*
  82  * Try to acquire jbd_lock_bh_state() against the buffer, when j_list_lock is
  83  * held.  For ranking reasons we must trylock.  If we lose, schedule away and
  84  * return 0.  j_list_lock is dropped in this case.
  85  */
  86 static int inverted_lock(journal_t *journal, struct buffer_head *bh)
  87 {
  88         if (!jbd_trylock_bh_state(bh)) {
  89                 spin_unlock(&journal->j_list_lock);
  90                 schedule();
  91                 return 0;
  92         }
  93         return 1;
  94 }
  95
  96 /* Done it all: now write the commit record.  We should have
  97  * cleaned up our previous buffers by now, so if we are in abort
  98  * mode we can now just skip the rest of the journal write
  99  * entirely.
 100  *
 101  * Returns 1 if the journal needs to be aborted or 0 on success
 102  */
 103 static int journal_write_commit_record(journal_t *journal,
 104                                         transaction_t *commit_transaction)
 105 {
 106         struct journal_head *descriptor;
 107         struct buffer_head *bh;
 108         int i, ret;
 109         int barrier_done = 0;
 110
 111         if (is_journal_aborted(journal))
 112                 return 0;
 113
 114         descriptor = journal_get_descriptor_buffer(journal);
 115         if (!descriptor)
 116                 return 1;
 117
 118         bh = jh2bh(descriptor);
 119
 120         /* AKPM: buglet - add `i' to tmp! */
 121         for (i = 0; i < bh->b_size; i += 512) {
 122                 journal_header_t *tmp = (journal_header_t*)bh->b_data;
 123                 tmp->h_magic = cpu_to_be32(JFS_MAGIC_NUMBER);
 124                 tmp->h_blocktype = cpu_to_be32(JFS_COMMIT_BLOCK);
 125                 tmp->h_sequence = cpu_to_be32(commit_transaction->t_tid);
 126         }
 127
 128         JBUFFER_TRACE(descriptor, "write commit block");
 129         set_buffer_dirty(bh);
 130         if (journal->j_flags & JFS_BARRIER) {
 131                 set_buffer_ordered(bh);
 132                 barrier_done = 1;
 133         }
 134         ret = sync_dirty_buffer(bh);
 135         /* is it possible for another commit to fail at roughly
 136          * the same time as this one?  If so, we don't want to
 137          * trust the barrier flag in the super, but instead want
 138          * to remember if we sent a barrier request
 139          */
 140         if (ret == -EOPNOTSUPP && barrier_done) {
 141                 char b[BDEVNAME_SIZE];
 142
 143                 printk(KERN_WARNING
 144                         "JBD: barrier-based sync failed on %s - "
 145                         "disabling barriers\n",
 146                         bdevname(journal->j_dev, b));
 147                 spin_lock(&journal->j_state_lock);
 148                 journal->j_flags &= ~JFS_BARRIER;
 149                 spin_unlock(&journal->j_state_lock);
 150
 151                 /* And try again, without the barrier */
 152                 clear_buffer_ordered(bh);
 153                 set_buffer_uptodate(bh);
 154                 set_buffer_dirty(bh);
 155                 ret = sync_dirty_buffer(bh);
 156         }
 157         put_bh(bh);             /* One for getblk() */
 158         journal_put_journal_head(descriptor);
 159
 160         return (ret == -EIO);
 161 }
 162
 163 /*
 164  * journal_commit_transaction
 165  *
 166  * The primary function for committing a transaction to the log.  This
 167  * function is called by the journal thread to begin a complete commit.
 168  */
 169 void journal_commit_transaction(journal_t *journal)
 170 {
 171         transaction_t *commit_transaction;
 172         struct journal_head *jh, *new_jh, *descriptor;
 173         struct buffer_head **wbuf = journal->j_wbuf;
 174         int bufs;
 175         int flags;
 176         int err;
 177         unsigned long blocknr;
 178         char *tagp = NULL;
 179         journal_header_t *header;
 180         journal_block_tag_t *tag = NULL;
 181         int space_left = 0;
 182         int first_tag = 0;
 183         int tag_flag;
 184         int i;
 185
 186         /*
 187          * First job: lock down the current transaction and wait for
 188          * all outstanding updates to complete.
 189          */
 190
 191 #ifdef COMMIT_STATS
 192         spin_lock(&journal->j_list_lock);
 193         summarise_journal_usage(journal);
 194         spin_unlock(&journal->j_list_lock);
 195 #endif
 196
 197         /* Do we need to erase the effects of a prior journal_flush? */
 198         if (journal->j_flags & JFS_FLUSHED) {
 199                 jbd_debug(3, "super block updated\n");
 200                 journal_update_superblock(journal, 1);
 201         } else {
 202                 jbd_debug(3, "superblock not updated\n");
 203         }
 204
 205         J_ASSERT(journal->j_running_transaction != NULL);
 206         J_ASSERT(journal->j_committing_transaction == NULL);
 207
 208         commit_transaction = journal->j_running_transaction;
 209         J_ASSERT(commit_transaction->t_state == T_RUNNING);
 210
 211         jbd_debug(1, "JBD: starting commit of transaction %d\n",
 212                         commit_transaction->t_tid);
 213
 214         spin_lock(&journal->j_state_lock);
 215         commit_transaction->t_state = T_LOCKED;
 216
 217         spin_lock(&commit_transaction->t_handle_lock);
 218         while (commit_transaction->t_updates) {
 219                 DEFINE_WAIT(wait);
 220
 221                 prepare_to_wait(&journal->j_wait_updates, &wait,
 222                                         TASK_UNINTERRUPTIBLE);
 223                 if (commit_transaction->t_updates) {
 224                         spin_unlock(&commit_transaction->t_handle_lock);
 225                         spin_unlock(&journal->j_state_lock);
 226                         schedule();
 227                         spin_lock(&journal->j_state_lock);
 228                         spin_lock(&commit_transaction->t_handle_lock);
 229                 }
 230                 finish_wait(&journal->j_wait_updates, &wait);
 231         }
 232         spin_unlock(&commit_transaction->t_handle_lock);
 233
 234         J_ASSERT (commit_transaction->t_outstanding_credits <=
 235                         journal->j_max_transaction_buffers);
 236
 237         /*
 238          * First thing we are allowed to do is to discard any remaining
 239          * BJ_Reserved buffers.  Note, it is _not_ permissible to assume
 240          * that there are no such buffers: if a large filesystem
 241          * operation like a truncate needs to split itself over multiple
 242          * transactions, then it may try to do a journal_restart() while
 243          * there are still BJ_Reserved buffers outstanding.  These must
 244          * be released cleanly from the current transaction.
 245          *
 246          * In this case, the filesystem must still reserve write access
 247          * again before modifying the buffer in the new transaction, but
 248          * we do not require it to remember exactly which old buffers it
 249          * has reserved.  This is consistent with the existing behaviour
 250          * that multiple journal_get_write_access() calls to the same
 251          * buffer are perfectly permissable.
 252          */
 253         while (commit_transaction->t_reserved_list) {
 254                 jh = commit_transaction->t_reserved_list;
 255                 JBUFFER_TRACE(jh, "reserved, unused: refile");
 256                 /*
 257                  * A journal_get_undo_access()+journal_release_buffer() may
 258                  * leave undo-committed data.
 259                  */
 260                 if (jh->b_committed_data) {
 261                         struct buffer_head *bh = jh2bh(jh);
 262
 263                         jbd_lock_bh_state(bh);
 264                         if (jh->b_committed_data) {
 265                                 kfree(jh->b_committed_data);
 266                                 jh->b_committed_data = NULL;
 267                         }
 268                         jbd_unlock_bh_state(bh);
 269                 }
 270                 journal_refile_buffer(journal, jh);
 271         }
 272
 273         /*
 274          * Now try to drop any written-back buffers from the journal's
 275          * checkpoint lists.  We do this *before* commit because it potentially
 276          * frees some memory
 277          */
 278         spin_lock(&journal->j_list_lock);
 279         __journal_clean_checkpoint_list(journal);
 280         spin_unlock(&journal->j_list_lock);
 281
 282         jbd_debug (3, "JBD: commit phase 1\n");
 283
 284         /*
 285          * Switch to a new revoke table.
 286          */
 287         journal_switch_revoke_table(journal);
 288
 289         commit_transaction->t_state = T_FLUSH;
 290         journal->j_committing_transaction = commit_transaction;
 291         journal->j_running_transaction = NULL;
 292         commit_transaction->t_log_start = journal->j_head;
 293         wake_up(&journal->j_wait_transaction_locked);
 294         spin_unlock(&journal->j_state_lock);
 295
 296         jbd_debug (3, "JBD: commit phase 2\n");
 297
 298         /*
 299          * First, drop modified flag: all accesses to the buffers
 300          * will be tracked for a new trasaction only -bzzz
 301          */
 302         spin_lock(&journal->j_list_lock);
 303         if (commit_transaction->t_buffers) {
 304                 new_jh = jh = commit_transaction->t_buffers->b_tnext;
 305                 do {
 306                         J_ASSERT_JH(new_jh, new_jh->b_modified == 1 ||
 307                                         new_jh->b_modified == 0);
 308                         new_jh->b_modified = 0;
 309                         new_jh = new_jh->b_tnext;
 310                 } while (new_jh != jh);
 311         }
 312         spin_unlock(&journal->j_list_lock);
 313
 314         /*
 315          * Now start flushing things to disk, in the order they appear
 316          * on the transaction lists.  Data blocks go first.
 317          */
 318
 319         err = 0;
 320         /*
 321          * Whenever we unlock the journal and sleep, things can get added
 322          * onto ->t_sync_datalist, so we have to keep looping back to
 323          * write_out_data until we *know* that the list is empty.
 324          */
 325         bufs = 0;
 326         /*
 327          * Cleanup any flushed data buffers from the data list.  Even in
 328          * abort mode, we want to flush this out as soon as possible.
 329          */
 330 write_out_data:
 331         cond_resched();
 332         spin_lock(&journal->j_list_lock);
 333
 334         while (commit_transaction->t_sync_datalist) {
 335                 struct buffer_head *bh;
 336
 337                 jh = commit_transaction->t_sync_datalist;
 338                 commit_transaction->t_sync_datalist = jh->b_tnext;
 339                 bh = jh2bh(jh);
 340                 if (buffer_locked(bh)) {
 341                         BUFFER_TRACE(bh, "locked");
 342                         if (!inverted_lock(journal, bh))
 343                                 goto write_out_data;
 344                         __journal_temp_unlink_buffer(jh);
 345                         __journal_file_buffer(jh, commit_transaction,
 346                                                 BJ_Locked);
 347                         jbd_unlock_bh_state(bh);
 348                         if (lock_need_resched(&journal->j_list_lock)) {
 349                                 spin_unlock(&journal->j_list_lock);
 350                                 goto write_out_data;
 351                         }
 352                 } else {
 353                         if (buffer_dirty(bh)) {
 354                                 BUFFER_TRACE(bh, "start journal writeout");
 355                                 get_bh(bh);
 356                                 wbuf[bufs++] = bh;
 357                                 if (bufs == journal->j_wbufsize) {
 358                                         jbd_debug(2, "submit %d writes\n",
 359                                                         bufs);
 360                                         spin_unlock(&journal->j_list_lock);
 361                                         ll_rw_block(WRITE, bufs, wbuf);
 362                                         journal_brelse_array(wbuf, bufs);
 363                                         bufs = 0;
 364                                         goto write_out_data;
 365                                 }
 366                         } else {
 367                                 BUFFER_TRACE(bh, "writeout complete: unfile");
 368                                 if (!inverted_lock(journal, bh))
 369                                         goto write_out_data;
 370                                 __journal_unfile_buffer(jh);
 371                                 jbd_unlock_bh_state(bh);
 372                                 journal_remove_journal_head(bh);
 373                                 put_bh(bh);
 374                                 if (lock_need_resched(&journal->j_list_lock)) {
 375                                         spin_unlock(&journal->j_list_lock);
 376                                         goto write_out_data;
 377                                 }
 378                         }
 379                 }
 380         }
 381
 382         if (bufs) {
 383                 spin_unlock(&journal->j_list_lock);
 384                 ll_rw_block(WRITE, bufs, wbuf);
 385                 journal_brelse_array(wbuf, bufs);
 386                 spin_lock(&journal->j_list_lock);
 387         }
 388
 389         /*
 390          * Wait for all previously submitted IO to complete.
 391          */
 392         while (commit_transaction->t_locked_list) {
 393                 struct buffer_head *bh;
 394
 395                 jh = commit_transaction->t_locked_list->b_tprev;
 396                 bh = jh2bh(jh);
 397                 get_bh(bh);
 398                 if (buffer_locked(bh)) {
 399                         spin_unlock(&journal->j_list_lock);
 400                         wait_on_buffer(bh);
 401                         if (unlikely(!buffer_uptodate(bh)))
 402                                 err = -EIO;
 403                         spin_lock(&journal->j_list_lock);
 404                 }
 405                 if (!inverted_lock(journal, bh)) {
 406                         put_bh(bh);
 407                         spin_lock(&journal->j_list_lock);
 408                         continue;
 409                 }
 410                 if (buffer_jbd(bh) && jh->b_jlist == BJ_Locked) {
 411                         __journal_unfile_buffer(jh);
 412                         jbd_unlock_bh_state(bh);
 413                         journal_remove_journal_head(bh);
 414                         put_bh(bh);
 415                 } else {
 416                         jbd_unlock_bh_state(bh);
 417                 }
 418                 put_bh(bh);
 419                 cond_resched_lock(&journal->j_list_lock);
 420         }
 421         spin_unlock(&journal->j_list_lock);
 422
 423         if (err)
 424                 __journal_abort_hard(journal);
 425
 426         journal_write_revoke_records(journal, commit_transaction);
 427
 428         jbd_debug(3, "JBD: commit phase 2\n");
 429
 430         /*
 431          * If we found any dirty or locked buffers, then we should have
 432          * looped back up to the write_out_data label.  If there weren't
 433          * any then journal_clean_data_list should have wiped the list
 434          * clean by now, so check that it is in fact empty.
 435          */
 436         J_ASSERT (commit_transaction->t_sync_datalist == NULL);
 437
 438         jbd_debug (3, "JBD: commit phase 3\n");
 439
 440         /*
 441          * Way to go: we have now written out all of the data for a
 442          * transaction!  Now comes the tricky part: we need to write out
 443          * metadata.  Loop over the transaction's entire buffer list:
 444          */
 445         commit_transaction->t_state = T_COMMIT;
 446
 447         descriptor = NULL;
 448         bufs = 0;
 449         while (commit_transaction->t_buffers) {
 450
 451                 /* Find the next buffer to be journaled... */
 452
 453                 jh = commit_transaction->t_buffers;
 454
 455                 /* If we're in abort mode, we just un-journal the buffer and
 456                    release it for background writing. */
 457
 458                 if (is_journal_aborted(journal)) {
 459                         JBUFFER_TRACE(jh, "journal is aborting: refile");
 460                         journal_refile_buffer(journal, jh);
 461                         /* If that was the last one, we need to clean up
 462                          * any descriptor buffers which may have been
 463                          * already allocated, even if we are now
 464                          * aborting. */
 465                         if (!commit_transaction->t_buffers)
 466                                 goto start_journal_io;
 467                         continue;
 468                 }
 469
 470                 /* Make sure we have a descriptor block in which to
 471                    record the metadata buffer. */
 472
 473                 if (!descriptor) {
 474                         struct buffer_head *bh;
 475
 476                         J_ASSERT (bufs == 0);
 477
 478                         jbd_debug(4, "JBD: get descriptor\n");
 479
 480                         descriptor = journal_get_descriptor_buffer(journal);
 481                         if (!descriptor) {
 482                                 __journal_abort_hard(journal);
 483                                 continue;
 484                         }
 485
 486                         bh = jh2bh(descriptor);
 487                         jbd_debug(4, "JBD: got buffer %llu (%p)\n",
 488                                 (unsigned long long)bh->b_blocknr, bh->b_data);
 489                         header = (journal_header_t *)&bh->b_data[0];
 490                         header->h_magic     = cpu_to_be32(JFS_MAGIC_NUMBER);
 491                         header->h_blocktype = cpu_to_be32(JFS_DESCRIPTOR_BLOCK);
 492                         header->h_sequence  = cpu_to_be32(commit_transaction->t_tid);
 493
 494                         tagp = &bh->b_data[sizeof(journal_header_t)];
 495                         space_left = bh->b_size - sizeof(journal_header_t);
 496                         first_tag = 1;
 497                         set_buffer_jwrite(bh);
 498                         set_buffer_dirty(bh);
 499                         wbuf[bufs++] = bh;
 500
 501                         /* Record it so that we can wait for IO
 502                            completion later */
 503                         BUFFER_TRACE(bh, "ph3: file as descriptor");
 504                         journal_file_buffer(descriptor, commit_transaction,
 505                                         BJ_LogCtl);
 506                 }
 507
 508                 /* Where is the buffer to be written? */
 509
 510                 err = journal_next_log_block(journal, &blocknr);
 511                 /* If the block mapping failed, just abandon the buffer
 512                    and repeat this loop: we'll fall into the
 513                    refile-on-abort condition above. */
 514                 if (err) {
 515                         __journal_abort_hard(journal);
 516                         continue;
 517                 }
 518
 519                 /*
 520                  * start_this_handle() uses t_outstanding_credits to determine
 521                  * the free space in the log, but this counter is changed
 522                  * by journal_next_log_block() also.
 523                  */
 524                 commit_transaction->t_outstanding_credits--;
 525
 526                 /* Bump b_count to prevent truncate from stumbling over
 527                    the shadowed buffer!  @@@ This can go if we ever get
 528                    rid of the BJ_IO/BJ_Shadow pairing of buffers. */
 529                 atomic_inc(&jh2bh(jh)->b_count);
 530
 531                 /* Make a temporary IO buffer with which to write it out
 532                    (this will requeue both the metadata buffer and the
 533                    temporary IO buffer). new_bh goes on BJ_IO*/
 534
 535                 set_bit(BH_JWrite, &jh2bh(jh)->b_state);
 536                 /*
 537                  * akpm: journal_write_metadata_buffer() sets
 538                  * new_bh->b_transaction to commit_transaction.
 539                  * We need to clean this up before we release new_bh
 540                  * (which is of type BJ_IO)
 541                  */
 542                 JBUFFER_TRACE(jh, "ph3: write metadata");
 543                 flags = journal_write_metadata_buffer(commit_transaction,
 544                                                       jh, &new_jh, blocknr);
 545                 set_bit(BH_JWrite, &jh2bh(new_jh)->b_state);
 546                 wbuf[bufs++] = jh2bh(new_jh);
 547
 548                 /* Record the new block's tag in the current descriptor
 549                    buffer */
 550
 551                 tag_flag = 0;
 552                 if (flags & 1)
 553                         tag_flag |= JFS_FLAG_ESCAPE;
 554                 if (!first_tag)
 555                         tag_flag |= JFS_FLAG_SAME_UUID;
 556
 557                 tag = (journal_block_tag_t *) tagp;
 558                 tag->t_blocknr = cpu_to_be32(jh2bh(jh)->b_blocknr);
 559                 tag->t_flags = cpu_to_be32(tag_flag);
 560                 tagp += sizeof(journal_block_tag_t);
 561                 space_left -= sizeof(journal_block_tag_t);
 562
 563                 if (first_tag) {
 564                         memcpy (tagp, journal->j_uuid, 16);
 565                         tagp += 16;
 566                         space_left -= 16;
 567                         first_tag = 0;
 568                 }
 569
 570                 /* If there's no more to do, or if the descriptor is full,
 571                    let the IO rip! */
 572
 573                 if (bufs == journal->j_wbufsize ||
 574                     commit_transaction->t_buffers == NULL ||
 575                     space_left < sizeof(journal_block_tag_t) + 16) {
 576
 577                         jbd_debug(4, "JBD: Submit %d IOs\n", bufs);
 578
 579                         /* Write an end-of-descriptor marker before
 580                            submitting the IOs.  "tag" still points to
 581                            the last tag we set up. */
 582
 583                         tag->t_flags |= cpu_to_be32(JFS_FLAG_LAST_TAG);
 584
 585 start_journal_io:
 586                         for (i = 0; i < bufs; i++) {
 587                                 struct buffer_head *bh = wbuf[i];
 588                                 lock_buffer(bh);
 589                                 clear_buffer_dirty(bh);
 590                                 set_buffer_uptodate(bh);
 591                                 bh->b_end_io = journal_end_buffer_io_sync;
 592                                 submit_bh(WRITE, bh);
 593                         }
 594                         cond_resched();
 595
 596                         /* Force a new descriptor to be generated next
 597                            time round the loop. */
 598                         descriptor = NULL;
 599                         bufs = 0;
 600                 }
 601         }
 602
 603         /* Lo and behold: we have just managed to send a transaction to
 604            the log.  Before we can commit it, wait for the IO so far to
 605            complete.  Control buffers being written are on the
 606            transaction's t_log_list queue, and metadata buffers are on
 607            the t_iobuf_list queue.
 608
 609            Wait for the buffers in reverse order.  That way we are
 610            less likely to be woken up until all IOs have completed, and
 611            so we incur less scheduling load.
 612         */
 613
 614         jbd_debug(3, "JBD: commit phase 4\n");
 615
 616         /*
 617          * akpm: these are BJ_IO, and j_list_lock is not needed.
 618          * See __journal_try_to_free_buffer.
 619          */
 620 wait_for_iobuf:
 621         while (commit_transaction->t_iobuf_list != NULL) {
 622                 struct buffer_head *bh;
 623
 624                 jh = commit_transaction->t_iobuf_list->b_tprev;
 625                 bh = jh2bh(jh);
 626                 if (buffer_locked(bh)) {
 627                         wait_on_buffer(bh);
 628                         goto wait_for_iobuf;
 629                 }
 630                 if (cond_resched())
 631                         goto wait_for_iobuf;
 632
 633                 if (unlikely(!buffer_uptodate(bh)))
 634                         err = -EIO;
 635
 636                 clear_buffer_jwrite(bh);
 637
 638                 JBUFFER_TRACE(jh, "ph4: unfile after journal write");
 639                 journal_unfile_buffer(journal, jh);
 640
 641                 /*
 642                  * ->t_iobuf_list should contain only dummy buffer_heads
 643                  * which were created by journal_write_metadata_buffer().
 644                  */
 645                 BUFFER_TRACE(bh, "dumping temporary bh");
 646                 journal_put_journal_head(jh);
 647                 __brelse(bh);
 648                 J_ASSERT_BH(bh, atomic_read(&bh->b_count) == 0);
 649                 free_buffer_head(bh);
 650
 651                 /* We also have to unlock and free the corresponding
 652                    shadowed buffer */
 653                 jh = commit_transaction->t_shadow_list->b_tprev;
 654                 bh = jh2bh(jh);
 655                 clear_bit(BH_JWrite, &bh->b_state);
 656                 J_ASSERT_BH(bh, buffer_jbddirty(bh));
 657
 658                 /* The metadata is now released for reuse, but we need
 659                    to remember it against this transaction so that when
 660                    we finally commit, we can do any checkpointing
 661                    required. */
 662                 JBUFFER_TRACE(jh, "file as BJ_Forget");
 663                 journal_file_buffer(jh, commit_transaction, BJ_Forget);
 664                 /* Wake up any transactions which were waiting for this
 665                    IO to complete */
 666                 wake_up_bit(&bh->b_state, BH_Unshadow);
 667                 JBUFFER_TRACE(jh, "brelse shadowed buffer");
 668                 __brelse(bh);
 669         }
 670
 671         J_ASSERT (commit_transaction->t_shadow_list == NULL);
 672
 673         jbd_debug(3, "JBD: commit phase 5\n");
 674
 675         /* Here we wait for the revoke record and descriptor record buffers */
 676  wait_for_ctlbuf:
 677         while (commit_transaction->t_log_list != NULL) {
 678                 struct buffer_head *bh;
 679
 680                 jh = commit_transaction->t_log_list->b_tprev;
 681                 bh = jh2bh(jh);
 682                 if (buffer_locked(bh)) {
 683                         wait_on_buffer(bh);
 684                         goto wait_for_ctlbuf;
 685                 }
 686                 if (cond_resched())
 687                         goto wait_for_ctlbuf;
 688
 689                 if (unlikely(!buffer_uptodate(bh)))
 690                         err = -EIO;
 691
 692                 BUFFER_TRACE(bh, "ph5: control buffer writeout done: unfile");
 693                 clear_buffer_jwrite(bh);
 694                 journal_unfile_buffer(journal, jh);
 695                 journal_put_journal_head(jh);
 696                 __brelse(bh);           /* One for getblk */
 697                 /* AKPM: bforget here */
 698         }
 699
 700         jbd_debug(3, "JBD: commit phase 6\n");
 701
 702         if (journal_write_commit_record(journal, commit_transaction))
 703                 err = -EIO;
 704
 705         if (err)
 706                 __journal_abort_hard(journal);
 707
 708         /* End of a transaction!  Finally, we can do checkpoint
 709            processing: any buffers committed as a result of this
 710            transaction can be removed from any checkpoint list it was on
 711            before. */
 712
 713         jbd_debug(3, "JBD: commit phase 7\n");
 714
 715         J_ASSERT(commit_transaction->t_sync_datalist == NULL);
 716         J_ASSERT(commit_transaction->t_buffers == NULL);
 717         J_ASSERT(commit_transaction->t_checkpoint_list == NULL);
 718         J_ASSERT(commit_transaction->t_iobuf_list == NULL);
 719         J_ASSERT(commit_transaction->t_shadow_list == NULL);
 720         J_ASSERT(commit_transaction->t_log_list == NULL);
 721
 722 restart_loop:
 723         while (commit_transaction->t_forget) {
 724                 transaction_t *cp_transaction;
 725                 struct buffer_head *bh;
 726
 727                 jh = commit_transaction->t_forget;
 728                 bh = jh2bh(jh);
 729                 jbd_lock_bh_state(bh);
 730                 J_ASSERT_JH(jh, jh->b_transaction == commit_transaction ||
 731                         jh->b_transaction == journal->j_running_transaction);
 732
 733                 /*
 734                  * If there is undo-protected committed data against
 735                  * this buffer, then we can remove it now.  If it is a
 736                  * buffer needing such protection, the old frozen_data
 737                  * field now points to a committed version of the
 738                  * buffer, so rotate that field to the new committed
 739                  * data.
 740                  *
 741                  * Otherwise, we can just throw away the frozen data now.
 742                  */
 743                 if (jh->b_committed_data) {
 744                         kfree(jh->b_committed_data);
 745                         jh->b_committed_data = NULL;
 746                         if (jh->b_frozen_data) {
 747                                 jh->b_committed_data = jh->b_frozen_data;
 748                                 jh->b_frozen_data = NULL;
 749                         }
 750                 } else if (jh->b_frozen_data) {
 751                         kfree(jh->b_frozen_data);
 752                         jh->b_frozen_data = NULL;
 753                 }
 754
 755                 spin_lock(&journal->j_list_lock);
 756                 cp_transaction = jh->b_cp_transaction;
 757                 if (cp_transaction) {
 758                         JBUFFER_TRACE(jh, "remove from old cp transaction");
 759                         __journal_remove_checkpoint(jh);
 760                 }
 761
 762                 /* Only re-checkpoint the buffer_head if it is marked
 763                  * dirty.  If the buffer was added to the BJ_Forget list
 764                  * by journal_forget, it may no longer be dirty and
 765                  * there's no point in keeping a checkpoint record for
 766                  * it. */
 767
 768                 /* A buffer which has been freed while still being
 769                  * journaled by a previous transaction may end up still
 770                  * being dirty here, but we want to avoid writing back
 771                  * that buffer in the future now that the last use has
 772                  * been committed.  That's not only a performance gain,
 773                  * it also stops aliasing problems if the buffer is left
 774                  * behind for writeback and gets reallocated for another
 775                  * use in a different page. */
 776                 if (buffer_freed(bh)) {
 777                         clear_buffer_freed(bh);
 778                         clear_buffer_jbddirty(bh);
 779                 }
 780
 781                 if (buffer_jbddirty(bh)) {
 782                         JBUFFER_TRACE(jh, "add to new checkpointing trans");
 783                         __journal_insert_checkpoint(jh, commit_transaction);
 784                         JBUFFER_TRACE(jh, "refile for checkpoint writeback");
 785                         __journal_refile_buffer(jh);
 786                         jbd_unlock_bh_state(bh);
 787                 } else {
 788                         J_ASSERT_BH(bh, !buffer_dirty(bh));
 789                         J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
 790                         __journal_unfile_buffer(jh);
 791                         jbd_unlock_bh_state(bh);
 792                         journal_remove_journal_head(bh);  /* needs a brelse */
 793                         release_buffer_page(bh);
 794                 }
 795                 spin_unlock(&journal->j_list_lock);
 796                 if (cond_resched())
 797                         goto restart_loop;
 798         }
 799
 800         /* Done with this transaction! */
 801
 802         jbd_debug(3, "JBD: commit phase 8\n");
 803
 804         J_ASSERT(commit_transaction->t_state == T_COMMIT);
 805
 806         /*
 807          * This is a bit sleazy.  We borrow j_list_lock to protect
 808          * journal->j_committing_transaction in __journal_remove_checkpoint.
 809          * Really, __jornal_remove_checkpoint should be using j_state_lock but
 810          * it's a bit hassle to hold that across __journal_remove_checkpoint
 811          */
 812         spin_lock(&journal->j_state_lock);
 813         spin_lock(&journal->j_list_lock);
 814         commit_transaction->t_state = T_FINISHED;
 815         J_ASSERT(commit_transaction == journal->j_committing_transaction);
 816         journal->j_commit_sequence = commit_transaction->t_tid;
 817         journal->j_committing_transaction = NULL;
 818         spin_unlock(&journal->j_state_lock);
 819
 820         if (commit_transaction->t_checkpoint_list == NULL) {
 821                 __journal_drop_transaction(journal, commit_transaction);
 822         } else {
 823                 if (journal->j_checkpoint_transactions == NULL) {
 824                         journal->j_checkpoint_transactions = commit_transaction;
 825                         commit_transaction->t_cpnext = commit_transaction;
 826                         commit_transaction->t_cpprev = commit_transaction;
 827                 } else {
 828                         commit_transaction->t_cpnext =
 829                                 journal->j_checkpoint_transactions;
 830                         commit_transaction->t_cpprev =
 831                                 commit_transaction->t_cpnext->t_cpprev;
 832                         commit_transaction->t_cpnext->t_cpprev =
 833                                 commit_transaction;
 834                         commit_transaction->t_cpprev->t_cpnext =
 835                                 commit_transaction;
 836                 }
 837         }
 838         spin_unlock(&journal->j_list_lock);
 839
 840         jbd_debug(1, "JBD: commit %d complete, head %d\n",
 841                   journal->j_commit_sequence, journal->j_tail_sequence);
 842
 843         wake_up(&journal->j_wait_done_commit);
 844 }