fs/jbd/commit.c

   1 /*
   2  * linux/fs/jbd/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
  24 /*
  25  * Default IO end handler for temporary BJ_IO buffer_heads.
  26  */
  27 static void journal_end_buffer_io_sync(struct buffer_head *bh, int uptodate)
  28 {
  29         BUFFER_TRACE(bh, "");
  30         if (uptodate)
  31                 set_buffer_uptodate(bh);
  32         else
  33                 clear_buffer_uptodate(bh);
  34         unlock_buffer(bh);
  35 }
  36
  37 /*
  38  * When an ext3-ordered file is truncated, it is possible that many pages are
  39  * not sucessfully freed, because they are attached to a committing transaction.
  40  * After the transaction commits, these pages are left on the LRU, with no
  41  * ->mapping, and with attached buffers.  These pages are trivially reclaimable
  42  * by the VM, but their apparent absence upsets the VM accounting, and it makes
  43  * the numbers in /proc/meminfo look odd.
  44  *
  45  * So here, we have a buffer which has just come off the forget list.  Look to
  46  * see if we can strip all buffers from the backing page.
  47  *
  48  * Called under lock_journal(), and possibly under journal_datalist_lock.  The
  49  * caller provided us with a ref against the buffer, and we drop that here.
  50  */
  51 static void release_buffer_page(struct buffer_head *bh)
  52 {
  53         struct page *page;
  54
  55         if (buffer_dirty(bh))
  56                 goto nope;
  57         if (atomic_read(&bh->b_count) != 1)
  58                 goto nope;
  59         page = bh->b_page;
  60         if (!page)
  61                 goto nope;
  62         if (page->mapping)
  63                 goto nope;
  64
  65         /* OK, it's a truncated page */
  66         if (TestSetPageLocked(page))
  67                 goto nope;
  68
  69         page_cache_get(page);
  70         __brelse(bh);
  71         try_to_free_buffers(page);
  72         unlock_page(page);
  73         page_cache_release(page);
  74         return;
  75
  76 nope:
  77         __brelse(bh);
  78 }
  79
  80 /*
  81  * Try to acquire jbd_lock_bh_state() against the buffer, when j_list_lock is
  82  * held.  For ranking reasons we must trylock.  If we lose, schedule away and
  83  * return 0.  j_list_lock is dropped in this case.
  84  */
  85 static int inverted_lock(journal_t *journal, struct buffer_head *bh)
  86 {
  87         if (!jbd_trylock_bh_state(bh)) {
  88                 spin_unlock(&journal->j_list_lock);
  89                 schedule();
  90                 return 0;
  91         }
  92         return 1;
  93 }
  94
  95 /* Done it all: now write the commit record.  We should have
  96  * cleaned up our previous buffers by now, so if we are in abort
  97  * mode we can now just skip the rest of the journal write
  98  * entirely.
  99  *
 100  * Returns 1 if the journal needs to be aborted or 0 on success
 101  */
 102 static int journal_write_commit_record(journal_t *journal,
 103                                         transaction_t *commit_transaction)
 104 {
 105         struct journal_head *descriptor;
 106         struct buffer_head *bh;
 107         journal_header_t *header;
 108         int 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         header = (journal_header_t *)(bh->b_data);
 121         header->h_magic = cpu_to_be32(JFS_MAGIC_NUMBER);
 122         header->h_blocktype = cpu_to_be32(JFS_COMMIT_BLOCK);
 123         header->h_sequence = cpu_to_be32(commit_transaction->t_tid);
 124
 125         JBUFFER_TRACE(descriptor, "write commit block");
 126         set_buffer_dirty(bh);
 127         if (journal->j_flags & JFS_BARRIER) {
 128                 set_buffer_ordered(bh);
 129                 barrier_done = 1;
 130         }
 131         ret = sync_dirty_buffer(bh);
 132         if (barrier_done)
 133                 clear_buffer_ordered(bh);
 134         /* is it possible for another commit to fail at roughly
 135          * the same time as this one?  If so, we don't want to
 136          * trust the barrier flag in the super, but instead want
 137          * to remember if we sent a barrier request
 138          */
 139         if (ret == -EOPNOTSUPP && barrier_done) {
 140                 char b[BDEVNAME_SIZE];
 141
 142                 printk(KERN_WARNING
 143                         "JBD: barrier-based sync failed on %s - "
 144                         "disabling barriers\n",
 145                         bdevname(journal->j_dev, b));
 146                 spin_lock(&journal->j_state_lock);
 147                 journal->j_flags &= ~JFS_BARRIER;
 148                 spin_unlock(&journal->j_state_lock);
 149
 150                 /* And try again, without the barrier */
 151                 set_buffer_uptodate(bh);
 152                 set_buffer_dirty(bh);
 153                 ret = sync_dirty_buffer(bh);
 154         }
 155         put_bh(bh);             /* One for getblk() */
 156         journal_put_journal_head(descriptor);
 157
 158         return (ret == -EIO);
 159 }
 160
 161 static void journal_do_submit_data(struct buffer_head **wbuf, int bufs)
 162 {
 163         int i;
 164
 165         for (i = 0; i < bufs; i++) {
 166                 wbuf[i]->b_end_io = end_buffer_write_sync;
 167                 /* We use-up our safety reference in submit_bh() */
 168                 submit_bh(WRITE, wbuf[i]);
 169         }
 170 }
 171
 172 /*
 173  *  Submit all the data buffers to disk
 174  */
 175 static void journal_submit_data_buffers(journal_t *journal,
 176                                 transaction_t *commit_transaction)
 177 {
 178         struct journal_head *jh;
 179         struct buffer_head *bh;
 180         int locked;
 181         int bufs = 0;
 182         struct buffer_head **wbuf = journal->j_wbuf;
 183
 184         /*
 185          * Whenever we unlock the journal and sleep, things can get added
 186          * onto ->t_sync_datalist, so we have to keep looping back to
 187          * write_out_data until we *know* that the list is empty.
 188          *
 189          * Cleanup any flushed data buffers from the data list.  Even in
 190          * abort mode, we want to flush this out as soon as possible.
 191          */
 192 write_out_data:
 193         cond_resched();
 194         spin_lock(&journal->j_list_lock);
 195
 196         while (commit_transaction->t_sync_datalist) {
 197                 jh = commit_transaction->t_sync_datalist;
 198                 bh = jh2bh(jh);
 199                 locked = 0;
 200
 201                 /* Get reference just to make sure buffer does not disappear
 202                  * when we are forced to drop various locks */
 203                 get_bh(bh);
 204                 /* If the buffer is dirty, we need to submit IO and hence
 205                  * we need the buffer lock. We try to lock the buffer without
 206                  * blocking. If we fail, we need to drop j_list_lock and do
 207                  * blocking lock_buffer().
 208                  */
 209                 if (buffer_dirty(bh)) {
 210                         if (test_set_buffer_locked(bh)) {
 211                                 BUFFER_TRACE(bh, "needs blocking lock");
 212                                 spin_unlock(&journal->j_list_lock);
 213                                 /* Write out all data to prevent deadlocks */
 214                                 journal_do_submit_data(wbuf, bufs);
 215                                 bufs = 0;
 216                                 lock_buffer(bh);
 217                                 spin_lock(&journal->j_list_lock);
 218                         }
 219                         locked = 1;
 220                 }
 221                 /* We have to get bh_state lock. Again out of order, sigh. */
 222                 if (!inverted_lock(journal, bh)) {
 223                         jbd_lock_bh_state(bh);
 224                         spin_lock(&journal->j_list_lock);
 225                 }
 226                 /* Someone already cleaned up the buffer? */
 227                 if (!buffer_jbd(bh)
 228                         || jh->b_transaction != commit_transaction
 229                         || jh->b_jlist != BJ_SyncData) {
 230                         jbd_unlock_bh_state(bh);
 231                         if (locked)
 232                                 unlock_buffer(bh);
 233                         BUFFER_TRACE(bh, "already cleaned up");
 234                         put_bh(bh);
 235                         continue;
 236                 }
 237                 if (locked && test_clear_buffer_dirty(bh)) {
 238                         BUFFER_TRACE(bh, "needs writeout, adding to array");
 239                         wbuf[bufs++] = bh;
 240                         __journal_file_buffer(jh, commit_transaction,
 241                                                 BJ_Locked);
 242                         jbd_unlock_bh_state(bh);
 243                         if (bufs == journal->j_wbufsize) {
 244                                 spin_unlock(&journal->j_list_lock);
 245                                 journal_do_submit_data(wbuf, bufs);
 246                                 bufs = 0;
 247                                 goto write_out_data;
 248                         }
 249                 } else if (!locked && buffer_locked(bh)) {
 250                         __journal_file_buffer(jh, commit_transaction,
 251                                                 BJ_Locked);
 252                         jbd_unlock_bh_state(bh);
 253                         put_bh(bh);
 254                 } else {
 255                         BUFFER_TRACE(bh, "writeout complete: unfile");
 256                         __journal_unfile_buffer(jh);
 257                         jbd_unlock_bh_state(bh);
 258                         if (locked)
 259                                 unlock_buffer(bh);
 260                         journal_remove_journal_head(bh);
 261                         /* Once for our safety reference, once for
 262                          * journal_remove_journal_head() */
 263                         put_bh(bh);
 264                         put_bh(bh);
 265                 }
 266
 267                 if (need_resched() || spin_needbreak(&journal->j_list_lock)) {
 268                         spin_unlock(&journal->j_list_lock);
 269                         goto write_out_data;
 270                 }
 271         }
 272         spin_unlock(&journal->j_list_lock);
 273         journal_do_submit_data(wbuf, bufs);
 274 }
 275
 276 /*
 277  * journal_commit_transaction
 278  *
 279  * The primary function for committing a transaction to the log.  This
 280  * function is called by the journal thread to begin a complete commit.
 281  */
 282 void journal_commit_transaction(journal_t *journal)
 283 {
 284         transaction_t *commit_transaction;
 285         struct journal_head *jh, *new_jh, *descriptor;
 286         struct buffer_head **wbuf = journal->j_wbuf;
 287         int bufs;
 288         int flags;
 289         int err;
 290         unsigned long blocknr;
 291         char *tagp = NULL;
 292         journal_header_t *header;
 293         journal_block_tag_t *tag = NULL;
 294         int space_left = 0;
 295         int first_tag = 0;
 296         int tag_flag;
 297         int i;
 298
 299         /*
 300          * First job: lock down the current transaction and wait for
 301          * all outstanding updates to complete.
 302          */
 303
 304 #ifdef COMMIT_STATS
 305         spin_lock(&journal->j_list_lock);
 306         summarise_journal_usage(journal);
 307         spin_unlock(&journal->j_list_lock);
 308 #endif
 309
 310         /* Do we need to erase the effects of a prior journal_flush? */
 311         if (journal->j_flags & JFS_FLUSHED) {
 312                 jbd_debug(3, "super block updated\n");
 313                 journal_update_superblock(journal, 1);
 314         } else {
 315                 jbd_debug(3, "superblock not updated\n");
 316         }
 317
 318         J_ASSERT(journal->j_running_transaction != NULL);
 319         J_ASSERT(journal->j_committing_transaction == NULL);
 320
 321         commit_transaction = journal->j_running_transaction;
 322         J_ASSERT(commit_transaction->t_state == T_RUNNING);
 323
 324         jbd_debug(1, "JBD: starting commit of transaction %d\n",
 325                         commit_transaction->t_tid);
 326
 327         spin_lock(&journal->j_state_lock);
 328         commit_transaction->t_state = T_LOCKED;
 329
 330         spin_lock(&commit_transaction->t_handle_lock);
 331         while (commit_transaction->t_updates) {
 332                 DEFINE_WAIT(wait);
 333
 334                 prepare_to_wait(&journal->j_wait_updates, &wait,
 335                                         TASK_UNINTERRUPTIBLE);
 336                 if (commit_transaction->t_updates) {
 337                         spin_unlock(&commit_transaction->t_handle_lock);
 338                         spin_unlock(&journal->j_state_lock);
 339                         schedule();
 340                         spin_lock(&journal->j_state_lock);
 341                         spin_lock(&commit_transaction->t_handle_lock);
 342                 }
 343                 finish_wait(&journal->j_wait_updates, &wait);
 344         }
 345         spin_unlock(&commit_transaction->t_handle_lock);
 346
 347         J_ASSERT (commit_transaction->t_outstanding_credits <=
 348                         journal->j_max_transaction_buffers);
 349
 350         /*
 351          * First thing we are allowed to do is to discard any remaining
 352          * BJ_Reserved buffers.  Note, it is _not_ permissible to assume
 353          * that there are no such buffers: if a large filesystem
 354          * operation like a truncate needs to split itself over multiple
 355          * transactions, then it may try to do a journal_restart() while
 356          * there are still BJ_Reserved buffers outstanding.  These must
 357          * be released cleanly from the current transaction.
 358          *
 359          * In this case, the filesystem must still reserve write access
 360          * again before modifying the buffer in the new transaction, but
 361          * we do not require it to remember exactly which old buffers it
 362          * has reserved.  This is consistent with the existing behaviour
 363          * that multiple journal_get_write_access() calls to the same
 364          * buffer are perfectly permissable.
 365          */
 366         while (commit_transaction->t_reserved_list) {
 367                 jh = commit_transaction->t_reserved_list;
 368                 JBUFFER_TRACE(jh, "reserved, unused: refile");
 369                 /*
 370                  * A journal_get_undo_access()+journal_release_buffer() may
 371                  * leave undo-committed data.
 372                  */
 373                 if (jh->b_committed_data) {
 374                         struct buffer_head *bh = jh2bh(jh);
 375
 376                         jbd_lock_bh_state(bh);
 377                         jbd_free(jh->b_committed_data, bh->b_size);
 378                         jh->b_committed_data = NULL;
 379                         jbd_unlock_bh_state(bh);
 380                 }
 381                 journal_refile_buffer(journal, jh);
 382         }
 383
 384         /*
 385          * Now try to drop any written-back buffers from the journal's
 386          * checkpoint lists.  We do this *before* commit because it potentially
 387          * frees some memory
 388          */
 389         spin_lock(&journal->j_list_lock);
 390         __journal_clean_checkpoint_list(journal);
 391         spin_unlock(&journal->j_list_lock);
 392
 393         jbd_debug (3, "JBD: commit phase 1\n");
 394
 395         /*
 396          * Switch to a new revoke table.
 397          */
 398         journal_switch_revoke_table(journal);
 399
 400         commit_transaction->t_state = T_FLUSH;
 401         journal->j_committing_transaction = commit_transaction;
 402         journal->j_running_transaction = NULL;
 403         commit_transaction->t_log_start = journal->j_head;
 404         wake_up(&journal->j_wait_transaction_locked);
 405         spin_unlock(&journal->j_state_lock);
 406
 407         jbd_debug (3, "JBD: commit phase 2\n");
 408
 409         /*
 410          * Now start flushing things to disk, in the order they appear
 411          * on the transaction lists.  Data blocks go first.
 412          */
 413         err = 0;
 414         journal_submit_data_buffers(journal, commit_transaction);
 415
 416         /*
 417          * Wait for all previously submitted IO to complete.
 418          */
 419         spin_lock(&journal->j_list_lock);
 420         while (commit_transaction->t_locked_list) {
 421                 struct buffer_head *bh;
 422
 423                 jh = commit_transaction->t_locked_list->b_tprev;
 424                 bh = jh2bh(jh);
 425                 get_bh(bh);
 426                 if (buffer_locked(bh)) {
 427                         spin_unlock(&journal->j_list_lock);
 428                         wait_on_buffer(bh);
 429                         if (unlikely(!buffer_uptodate(bh)))
 430                                 err = -EIO;
 431                         spin_lock(&journal->j_list_lock);
 432                 }
 433                 if (!inverted_lock(journal, bh)) {
 434                         put_bh(bh);
 435                         spin_lock(&journal->j_list_lock);
 436                         continue;
 437                 }
 438                 if (buffer_jbd(bh) && jh->b_jlist == BJ_Locked) {
 439                         __journal_unfile_buffer(jh);
 440                         jbd_unlock_bh_state(bh);
 441                         journal_remove_journal_head(bh);
 442                         put_bh(bh);
 443                 } else {
 444                         jbd_unlock_bh_state(bh);
 445                 }
 446                 put_bh(bh);
 447                 cond_resched_lock(&journal->j_list_lock);
 448         }
 449         spin_unlock(&journal->j_list_lock);
 450
 451         if (err)
 452                 journal_abort(journal, err);
 453
 454         journal_write_revoke_records(journal, commit_transaction);
 455
 456         jbd_debug(3, "JBD: commit phase 2\n");
 457
 458         /*
 459          * If we found any dirty or locked buffers, then we should have
 460          * looped back up to the write_out_data label.  If there weren't
 461          * any then journal_clean_data_list should have wiped the list
 462          * clean by now, so check that it is in fact empty.
 463          */
 464         J_ASSERT (commit_transaction->t_sync_datalist == NULL);
 465
 466         jbd_debug (3, "JBD: commit phase 3\n");
 467
 468         /*
 469          * Way to go: we have now written out all of the data for a
 470          * transaction!  Now comes the tricky part: we need to write out
 471          * metadata.  Loop over the transaction's entire buffer list:
 472          */
 473         commit_transaction->t_state = T_COMMIT;
 474
 475         descriptor = NULL;
 476         bufs = 0;
 477         while (commit_transaction->t_buffers) {
 478
 479                 /* Find the next buffer to be journaled... */
 480
 481                 jh = commit_transaction->t_buffers;
 482
 483                 /* If we're in abort mode, we just un-journal the buffer and
 484                    release it for background writing. */
 485
 486                 if (is_journal_aborted(journal)) {
 487                         JBUFFER_TRACE(jh, "journal is aborting: refile");
 488                         journal_refile_buffer(journal, jh);
 489                         /* If that was the last one, we need to clean up
 490                          * any descriptor buffers which may have been
 491                          * already allocated, even if we are now
 492                          * aborting. */
 493                         if (!commit_transaction->t_buffers)
 494                                 goto start_journal_io;
 495                         continue;
 496                 }
 497
 498                 /* Make sure we have a descriptor block in which to
 499                    record the metadata buffer. */
 500
 501                 if (!descriptor) {
 502                         struct buffer_head *bh;
 503
 504                         J_ASSERT (bufs == 0);
 505
 506                         jbd_debug(4, "JBD: get descriptor\n");
 507
 508                         descriptor = journal_get_descriptor_buffer(journal);
 509                         if (!descriptor) {
 510                                 journal_abort(journal, -EIO);
 511                                 continue;
 512                         }
 513
 514                         bh = jh2bh(descriptor);
 515                         jbd_debug(4, "JBD: got buffer %llu (%p)\n",
 516                                 (unsigned long long)bh->b_blocknr, bh->b_data);
 517                         header = (journal_header_t *)&bh->b_data[0];
 518                         header->h_magic     = cpu_to_be32(JFS_MAGIC_NUMBER);
 519                         header->h_blocktype = cpu_to_be32(JFS_DESCRIPTOR_BLOCK);
 520                         header->h_sequence  = cpu_to_be32(commit_transaction->t_tid);
 521
 522                         tagp = &bh->b_data[sizeof(journal_header_t)];
 523                         space_left = bh->b_size - sizeof(journal_header_t);
 524                         first_tag = 1;
 525                         set_buffer_jwrite(bh);
 526                         set_buffer_dirty(bh);
 527                         wbuf[bufs++] = bh;
 528
 529                         /* Record it so that we can wait for IO
 530                            completion later */
 531                         BUFFER_TRACE(bh, "ph3: file as descriptor");
 532                         journal_file_buffer(descriptor, commit_transaction,
 533                                         BJ_LogCtl);
 534                 }
 535
 536                 /* Where is the buffer to be written? */
 537
 538                 err = journal_next_log_block(journal, &blocknr);
 539                 /* If the block mapping failed, just abandon the buffer
 540                    and repeat this loop: we'll fall into the
 541                    refile-on-abort condition above. */
 542                 if (err) {
 543                         journal_abort(journal, err);
 544                         continue;
 545                 }
 546
 547                 /*
 548                  * start_this_handle() uses t_outstanding_credits to determine
 549                  * the free space in the log, but this counter is changed
 550                  * by journal_next_log_block() also.
 551                  */
 552                 commit_transaction->t_outstanding_credits--;
 553
 554                 /* Bump b_count to prevent truncate from stumbling over
 555                    the shadowed buffer!  @@@ This can go if we ever get
 556                    rid of the BJ_IO/BJ_Shadow pairing of buffers. */
 557                 atomic_inc(&jh2bh(jh)->b_count);
 558
 559                 /* Make a temporary IO buffer with which to write it out
 560                    (this will requeue both the metadata buffer and the
 561                    temporary IO buffer). new_bh goes on BJ_IO*/
 562
 563                 set_bit(BH_JWrite, &jh2bh(jh)->b_state);
 564                 /*
 565                  * akpm: journal_write_metadata_buffer() sets
 566                  * new_bh->b_transaction to commit_transaction.
 567                  * We need to clean this up before we release new_bh
 568                  * (which is of type BJ_IO)
 569                  */
 570                 JBUFFER_TRACE(jh, "ph3: write metadata");
 571                 flags = journal_write_metadata_buffer(commit_transaction,
 572                                                       jh, &new_jh, blocknr);
 573                 set_bit(BH_JWrite, &jh2bh(new_jh)->b_state);
 574                 wbuf[bufs++] = jh2bh(new_jh);
 575
 576                 /* Record the new block's tag in the current descriptor
 577                    buffer */
 578
 579                 tag_flag = 0;
 580                 if (flags & 1)
 581                         tag_flag |= JFS_FLAG_ESCAPE;
 582                 if (!first_tag)
 583                         tag_flag |= JFS_FLAG_SAME_UUID;
 584
 585                 tag = (journal_block_tag_t *) tagp;
 586                 tag->t_blocknr = cpu_to_be32(jh2bh(jh)->b_blocknr);
 587                 tag->t_flags = cpu_to_be32(tag_flag);
 588                 tagp += sizeof(journal_block_tag_t);
 589                 space_left -= sizeof(journal_block_tag_t);
 590
 591                 if (first_tag) {
 592                         memcpy (tagp, journal->j_uuid, 16);
 593                         tagp += 16;
 594                         space_left -= 16;
 595                         first_tag = 0;
 596                 }
 597
 598                 /* If there's no more to do, or if the descriptor is full,
 599                    let the IO rip! */
 600
 601                 if (bufs == journal->j_wbufsize ||
 602                     commit_transaction->t_buffers == NULL ||
 603                     space_left < sizeof(journal_block_tag_t) + 16) {
 604
 605                         jbd_debug(4, "JBD: Submit %d IOs\n", bufs);
 606
 607                         /* Write an end-of-descriptor marker before
 608                            submitting the IOs.  "tag" still points to
 609                            the last tag we set up. */
 610
 611                         tag->t_flags |= cpu_to_be32(JFS_FLAG_LAST_TAG);
 612
 613 start_journal_io:
 614                         for (i = 0; i < bufs; i++) {
 615                                 struct buffer_head *bh = wbuf[i];
 616                                 lock_buffer(bh);
 617                                 clear_buffer_dirty(bh);
 618                                 set_buffer_uptodate(bh);
 619                                 bh->b_end_io = journal_end_buffer_io_sync;
 620                                 submit_bh(WRITE, bh);
 621                         }
 622                         cond_resched();
 623
 624                         /* Force a new descriptor to be generated next
 625                            time round the loop. */
 626                         descriptor = NULL;
 627                         bufs = 0;
 628                 }
 629         }
 630
 631         /* Lo and behold: we have just managed to send a transaction to
 632            the log.  Before we can commit it, wait for the IO so far to
 633            complete.  Control buffers being written are on the
 634            transaction's t_log_list queue, and metadata buffers are on
 635            the t_iobuf_list queue.
 636
 637            Wait for the buffers in reverse order.  That way we are
 638            less likely to be woken up until all IOs have completed, and
 639            so we incur less scheduling load.
 640         */
 641
 642         jbd_debug(3, "JBD: commit phase 4\n");
 643
 644         /*
 645          * akpm: these are BJ_IO, and j_list_lock is not needed.
 646          * See __journal_try_to_free_buffer.
 647          */
 648 wait_for_iobuf:
 649         while (commit_transaction->t_iobuf_list != NULL) {
 650                 struct buffer_head *bh;
 651
 652                 jh = commit_transaction->t_iobuf_list->b_tprev;
 653                 bh = jh2bh(jh);
 654                 if (buffer_locked(bh)) {
 655                         wait_on_buffer(bh);
 656                         goto wait_for_iobuf;
 657                 }
 658                 if (cond_resched())
 659                         goto wait_for_iobuf;
 660
 661                 if (unlikely(!buffer_uptodate(bh)))
 662                         err = -EIO;
 663
 664                 clear_buffer_jwrite(bh);
 665
 666                 JBUFFER_TRACE(jh, "ph4: unfile after journal write");
 667                 journal_unfile_buffer(journal, jh);
 668
 669                 /*
 670                  * ->t_iobuf_list should contain only dummy buffer_heads
 671                  * which were created by journal_write_metadata_buffer().
 672                  */
 673                 BUFFER_TRACE(bh, "dumping temporary bh");
 674                 journal_put_journal_head(jh);
 675                 __brelse(bh);
 676                 J_ASSERT_BH(bh, atomic_read(&bh->b_count) == 0);
 677                 free_buffer_head(bh);
 678
 679                 /* We also have to unlock and free the corresponding
 680                    shadowed buffer */
 681                 jh = commit_transaction->t_shadow_list->b_tprev;
 682                 bh = jh2bh(jh);
 683                 clear_bit(BH_JWrite, &bh->b_state);
 684                 J_ASSERT_BH(bh, buffer_jbddirty(bh));
 685
 686                 /* The metadata is now released for reuse, but we need
 687                    to remember it against this transaction so that when
 688                    we finally commit, we can do any checkpointing
 689                    required. */
 690                 JBUFFER_TRACE(jh, "file as BJ_Forget");
 691                 journal_file_buffer(jh, commit_transaction, BJ_Forget);
 692                 /* Wake up any transactions which were waiting for this
 693                    IO to complete */
 694                 wake_up_bit(&bh->b_state, BH_Unshadow);
 695                 JBUFFER_TRACE(jh, "brelse shadowed buffer");
 696                 __brelse(bh);
 697         }
 698
 699         J_ASSERT (commit_transaction->t_shadow_list == NULL);
 700
 701         jbd_debug(3, "JBD: commit phase 5\n");
 702
 703         /* Here we wait for the revoke record and descriptor record buffers */
 704  wait_for_ctlbuf:
 705         while (commit_transaction->t_log_list != NULL) {
 706                 struct buffer_head *bh;
 707
 708                 jh = commit_transaction->t_log_list->b_tprev;
 709                 bh = jh2bh(jh);
 710                 if (buffer_locked(bh)) {
 711                         wait_on_buffer(bh);
 712                         goto wait_for_ctlbuf;
 713                 }
 714                 if (cond_resched())
 715                         goto wait_for_ctlbuf;
 716
 717                 if (unlikely(!buffer_uptodate(bh)))
 718                         err = -EIO;
 719
 720                 BUFFER_TRACE(bh, "ph5: control buffer writeout done: unfile");
 721                 clear_buffer_jwrite(bh);
 722                 journal_unfile_buffer(journal, jh);
 723                 journal_put_journal_head(jh);
 724                 __brelse(bh);           /* One for getblk */
 725                 /* AKPM: bforget here */
 726         }
 727
 728         jbd_debug(3, "JBD: commit phase 6\n");
 729
 730         if (journal_write_commit_record(journal, commit_transaction))
 731                 err = -EIO;
 732
 733         if (err)
 734                 journal_abort(journal, err);
 735
 736         /* End of a transaction!  Finally, we can do checkpoint
 737            processing: any buffers committed as a result of this
 738            transaction can be removed from any checkpoint list it was on
 739            before. */
 740
 741         jbd_debug(3, "JBD: commit phase 7\n");
 742
 743         J_ASSERT(commit_transaction->t_sync_datalist == NULL);
 744         J_ASSERT(commit_transaction->t_buffers == NULL);
 745         J_ASSERT(commit_transaction->t_checkpoint_list == NULL);
 746         J_ASSERT(commit_transaction->t_iobuf_list == NULL);
 747         J_ASSERT(commit_transaction->t_shadow_list == NULL);
 748         J_ASSERT(commit_transaction->t_log_list == NULL);
 749
 750 restart_loop:
 751         /*
 752          * As there are other places (journal_unmap_buffer()) adding buffers
 753          * to this list we have to be careful and hold the j_list_lock.
 754          */
 755         spin_lock(&journal->j_list_lock);
 756         while (commit_transaction->t_forget) {
 757                 transaction_t *cp_transaction;
 758                 struct buffer_head *bh;
 759
 760                 jh = commit_transaction->t_forget;
 761                 spin_unlock(&journal->j_list_lock);
 762                 bh = jh2bh(jh);
 763                 jbd_lock_bh_state(bh);
 764                 J_ASSERT_JH(jh, jh->b_transaction == commit_transaction ||
 765                         jh->b_transaction == journal->j_running_transaction);
 766
 767                 /*
 768                  * If there is undo-protected committed data against
 769                  * this buffer, then we can remove it now.  If it is a
 770                  * buffer needing such protection, the old frozen_data
 771                  * field now points to a committed version of the
 772                  * buffer, so rotate that field to the new committed
 773                  * data.
 774                  *
 775                  * Otherwise, we can just throw away the frozen data now.
 776                  */
 777                 if (jh->b_committed_data) {
 778                         jbd_free(jh->b_committed_data, bh->b_size);
 779                         jh->b_committed_data = NULL;
 780                         if (jh->b_frozen_data) {
 781                                 jh->b_committed_data = jh->b_frozen_data;
 782                                 jh->b_frozen_data = NULL;
 783                         }
 784                 } else if (jh->b_frozen_data) {
 785                         jbd_free(jh->b_frozen_data, bh->b_size);
 786                         jh->b_frozen_data = NULL;
 787                 }
 788
 789                 spin_lock(&journal->j_list_lock);
 790                 cp_transaction = jh->b_cp_transaction;
 791                 if (cp_transaction) {
 792                         JBUFFER_TRACE(jh, "remove from old cp transaction");
 793                         __journal_remove_checkpoint(jh);
 794                 }
 795
 796                 /* Only re-checkpoint the buffer_head if it is marked
 797                  * dirty.  If the buffer was added to the BJ_Forget list
 798                  * by journal_forget, it may no longer be dirty and
 799                  * there's no point in keeping a checkpoint record for
 800                  * it. */
 801
 802                 /* A buffer which has been freed while still being
 803                  * journaled by a previous transaction may end up still
 804                  * being dirty here, but we want to avoid writing back
 805                  * that buffer in the future now that the last use has
 806                  * been committed.  That's not only a performance gain,
 807                  * it also stops aliasing problems if the buffer is left
 808                  * behind for writeback and gets reallocated for another
 809                  * use in a different page. */
 810                 if (buffer_freed(bh)) {
 811                         clear_buffer_freed(bh);
 812                         clear_buffer_jbddirty(bh);
 813                 }
 814
 815                 if (buffer_jbddirty(bh)) {
 816                         JBUFFER_TRACE(jh, "add to new checkpointing trans");
 817                         __journal_insert_checkpoint(jh, commit_transaction);
 818                         JBUFFER_TRACE(jh, "refile for checkpoint writeback");
 819                         __journal_refile_buffer(jh);
 820                         jbd_unlock_bh_state(bh);
 821                 } else {
 822                         J_ASSERT_BH(bh, !buffer_dirty(bh));
 823                         /* The buffer on BJ_Forget list and not jbddirty means
 824                          * it has been freed by this transaction and hence it
 825                          * could not have been reallocated until this
 826                          * transaction has committed. *BUT* it could be
 827                          * reallocated once we have written all the data to
 828                          * disk and before we process the buffer on BJ_Forget
 829                          * list. */
 830                         JBUFFER_TRACE(jh, "refile or unfile freed buffer");
 831                         __journal_refile_buffer(jh);
 832                         if (!jh->b_transaction) {
 833                                 jbd_unlock_bh_state(bh);
 834                                  /* needs a brelse */
 835                                 journal_remove_journal_head(bh);
 836                                 release_buffer_page(bh);
 837                         } else
 838                                 jbd_unlock_bh_state(bh);
 839                 }
 840                 cond_resched_lock(&journal->j_list_lock);
 841         }
 842         spin_unlock(&journal->j_list_lock);
 843         /*
 844          * This is a bit sleazy.  We use j_list_lock to protect transition
 845          * of a transaction into T_FINISHED state and calling
 846          * __journal_drop_transaction(). Otherwise we could race with
 847          * other checkpointing code processing the transaction...
 848          */
 849         spin_lock(&journal->j_state_lock);
 850         spin_lock(&journal->j_list_lock);
 851         /*
 852          * Now recheck if some buffers did not get attached to the transaction
 853          * while the lock was dropped...
 854          */
 855         if (commit_transaction->t_forget) {
 856                 spin_unlock(&journal->j_list_lock);
 857                 spin_unlock(&journal->j_state_lock);
 858                 goto restart_loop;
 859         }
 860
 861         /* Done with this transaction! */
 862
 863         jbd_debug(3, "JBD: commit phase 8\n");
 864
 865         J_ASSERT(commit_transaction->t_state == T_COMMIT);
 866
 867         commit_transaction->t_state = T_FINISHED;
 868         J_ASSERT(commit_transaction == journal->j_committing_transaction);
 869         journal->j_commit_sequence = commit_transaction->t_tid;
 870         journal->j_committing_transaction = NULL;
 871         spin_unlock(&journal->j_state_lock);
 872
 873         if (commit_transaction->t_checkpoint_list == NULL &&
 874             commit_transaction->t_checkpoint_io_list == NULL) {
 875                 __journal_drop_transaction(journal, commit_transaction);
 876         } else {
 877                 if (journal->j_checkpoint_transactions == NULL) {
 878                         journal->j_checkpoint_transactions = commit_transaction;
 879                         commit_transaction->t_cpnext = commit_transaction;
 880                         commit_transaction->t_cpprev = commit_transaction;
 881                 } else {
 882                         commit_transaction->t_cpnext =
 883                                 journal->j_checkpoint_transactions;
 884                         commit_transaction->t_cpprev =
 885                                 commit_transaction->t_cpnext->t_cpprev;
 886                         commit_transaction->t_cpnext->t_cpprev =
 887                                 commit_transaction;
 888                         commit_transaction->t_cpprev->t_cpnext =
 889                                 commit_transaction;
 890                 }
 891         }
 892         spin_unlock(&journal->j_list_lock);
 893
 894         jbd_debug(1, "JBD: commit %d complete, head %d\n",
 895                   journal->j_commit_sequence, journal->j_tail_sequence);
 896
 897         wake_up(&journal->j_wait_done_commit);
 898 }