git.oblomov.eu Git - linux-2.6/blob - 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 #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 static void journal_do_submit_data(struct buffer_head **wbuf, int bufs)
 164 {
 165         int i;
 166
 167         for (i = 0; i < bufs; i++) {
 168                 wbuf[i]->b_end_io = end_buffer_write_sync;
 169                 /* We use-up our safety reference in submit_bh() */
 170                 submit_bh(WRITE, wbuf[i]);
 171         }
 172 }
 173
 174 /*
 175  *  Submit all the data buffers to disk
 176  */
 177 static void journal_submit_data_buffers(journal_t *journal,
 178                                 transaction_t *commit_transaction)
 179 {
 180         struct journal_head *jh;
 181         struct buffer_head *bh;
 182         int locked;
 183         int bufs = 0;
 184         struct buffer_head **wbuf = journal->j_wbuf;
 185
 186         /*
 187          * Whenever we unlock the journal and sleep, things can get added
 188          * onto ->t_sync_datalist, so we have to keep looping back to
 189          * write_out_data until we *know* that the list is empty.
 190          *
 191          * Cleanup any flushed data buffers from the data list.  Even in
 192          * abort mode, we want to flush this out as soon as possible.
 193          */
 194 write_out_data:
 195         cond_resched();
 196         spin_lock(&journal->j_list_lock);
 197
 198         while (commit_transaction->t_sync_datalist) {
 199                 jh = commit_transaction->t_sync_datalist;
 200                 bh = jh2bh(jh);
 201                 locked = 0;
 202
 203                 /* Get reference just to make sure buffer does not disappear
 204                  * when we are forced to drop various locks */
 205                 get_bh(bh);
 206                 /* If the buffer is dirty, we need to submit IO and hence
 207                  * we need the buffer lock. We try to lock the buffer without
 208                  * blocking. If we fail, we need to drop j_list_lock and do
 209                  * blocking lock_buffer().
 210                  */
 211                 if (buffer_dirty(bh)) {
 212                         if (test_set_buffer_locked(bh)) {
 213                                 BUFFER_TRACE(bh, "needs blocking lock");
 214                                 spin_unlock(&journal->j_list_lock);
 215                                 /* Write out all data to prevent deadlocks */
 216                                 journal_do_submit_data(wbuf, bufs);
 217                                 bufs = 0;
 218                                 lock_buffer(bh);
 219                                 spin_lock(&journal->j_list_lock);
 220                         }
 221                         locked = 1;
 222                 }
 223                 /* We have to get bh_state lock. Again out of order, sigh. */
 224                 if (!inverted_lock(journal, bh)) {
 225                         jbd_lock_bh_state(bh);
 226                         spin_lock(&journal->j_list_lock);
 227                 }
 228                 /* Someone already cleaned up the buffer? */
 229                 if (!buffer_jbd(bh)
 230                         || jh->b_transaction != commit_transaction
 231                         || jh->b_jlist != BJ_SyncData) {
 232                         jbd_unlock_bh_state(bh);
 233                         if (locked)
 234                                 unlock_buffer(bh);
 235                         BUFFER_TRACE(bh, "already cleaned up");
 236                         put_bh(bh);
 237                         continue;
 238                 }
 239                 if (locked && test_clear_buffer_dirty(bh)) {
 240                         BUFFER_TRACE(bh, "needs writeout, adding to array");
 241                         wbuf[bufs++] = bh;
 242                         __journal_file_buffer(jh, commit_transaction,
 243                                                 BJ_Locked);
 244                         jbd_unlock_bh_state(bh);
 245                         if (bufs == journal->j_wbufsize) {
 246                                 spin_unlock(&journal->j_list_lock);
 247                                 journal_do_submit_data(wbuf, bufs);
 248                                 bufs = 0;
 249                                 goto write_out_data;
 250                         }
 251                 } else if (!locked && buffer_locked(bh)) {
 252                         __journal_file_buffer(jh, commit_transaction,
 253                                                 BJ_Locked);
 254                         jbd_unlock_bh_state(bh);
 255                         put_bh(bh);
 256                 } else {
 257                         BUFFER_TRACE(bh, "writeout complete: unfile");
 258                         __journal_unfile_buffer(jh);
 259                         jbd_unlock_bh_state(bh);
 260                         if (locked)
 261                                 unlock_buffer(bh);
 262                         journal_remove_journal_head(bh);
 263                         /* Once for our safety reference, once for
 264                          * journal_remove_journal_head() */
 265                         put_bh(bh);
 266                         put_bh(bh);
 267                 }
 268
 269                 if (lock_need_resched(&journal->j_list_lock)) {
 270                         spin_unlock(&journal->j_list_lock);
 271                         goto write_out_data;
 272                 }
 273         }
 274         spin_unlock(&journal->j_list_lock);
 275         journal_do_submit_data(wbuf, bufs);
 276 }
 277
 278 /*
 279  * journal_commit_transaction
 280  *
 281  * The primary function for committing a transaction to the log.  This
 282  * function is called by the journal thread to begin a complete commit.
 283  */
 284 void journal_commit_transaction(journal_t *journal)
 285 {
 286         transaction_t *commit_transaction;
 287         struct journal_head *jh, *new_jh, *descriptor;
 288         struct buffer_head **wbuf = journal->j_wbuf;
 289         int bufs;
 290         int flags;
 291         int err;
 292         unsigned long blocknr;
 293         char *tagp = NULL;
 294         journal_header_t *header;
 295         journal_block_tag_t *tag = NULL;
 296         int space_left = 0;
 297         int first_tag = 0;
 298         int tag_flag;
 299         int i;
 300
 301         /*
 302          * First job: lock down the current transaction and wait for
 303          * all outstanding updates to complete.
 304          */
 305
 306 #ifdef COMMIT_STATS
 307         spin_lock(&journal->j_list_lock);
 308         summarise_journal_usage(journal);
 309         spin_unlock(&journal->j_list_lock);
 310 #endif
 311
 312         /* Do we need to erase the effects of a prior journal_flush? */
 313         if (journal->j_flags & JFS_FLUSHED) {
 314                 jbd_debug(3, "super block updated\n");
 315                 journal_update_superblock(journal, 1);
 316         } else {
 317                 jbd_debug(3, "superblock not updated\n");
 318         }
 319
 320         J_ASSERT(journal->j_running_transaction != NULL);
 321         J_ASSERT(journal->j_committing_transaction == NULL);
 322
 323         commit_transaction = journal->j_running_transaction;
 324         J_ASSERT(commit_transaction->t_state == T_RUNNING);
 325
 326         jbd_debug(1, "JBD: starting commit of transaction %d\n",
 327                         commit_transaction->t_tid);
 328
 329         spin_lock(&journal->j_state_lock);
 330         commit_transaction->t_state = T_LOCKED;
 331
 332         spin_lock(&commit_transaction->t_handle_lock);
 333         while (commit_transaction->t_updates) {
 334                 DEFINE_WAIT(wait);
 335
 336                 prepare_to_wait(&journal->j_wait_updates, &wait,
 337                                         TASK_UNINTERRUPTIBLE);
 338                 if (commit_transaction->t_updates) {
 339                         spin_unlock(&commit_transaction->t_handle_lock);
 340                         spin_unlock(&journal->j_state_lock);
 341                         schedule();
 342                         spin_lock(&journal->j_state_lock);
 343                         spin_lock(&commit_transaction->t_handle_lock);
 344                 }
 345                 finish_wait(&journal->j_wait_updates, &wait);
 346         }
 347         spin_unlock(&commit_transaction->t_handle_lock);
 348
 349         J_ASSERT (commit_transaction->t_outstanding_credits <=
 350                         journal->j_max_transaction_buffers);
 351
 352         /*
 353          * First thing we are allowed to do is to discard any remaining
 354          * BJ_Reserved buffers.  Note, it is _not_ permissible to assume
 355          * that there are no such buffers: if a large filesystem
 356          * operation like a truncate needs to split itself over multiple
 357          * transactions, then it may try to do a journal_restart() while
 358          * there are still BJ_Reserved buffers outstanding.  These must
 359          * be released cleanly from the current transaction.
 360          *
 361          * In this case, the filesystem must still reserve write access
 362          * again before modifying the buffer in the new transaction, but
 363          * we do not require it to remember exactly which old buffers it
 364          * has reserved.  This is consistent with the existing behaviour
 365          * that multiple journal_get_write_access() calls to the same
 366          * buffer are perfectly permissable.
 367          */
 368         while (commit_transaction->t_reserved_list) {
 369                 jh = commit_transaction->t_reserved_list;
 370                 JBUFFER_TRACE(jh, "reserved, unused: refile");
 371                 /*
 372                  * A journal_get_undo_access()+journal_release_buffer() may
 373                  * leave undo-committed data.
 374                  */
 375                 if (jh->b_committed_data) {
 376                         struct buffer_head *bh = jh2bh(jh);
 377
 378                         jbd_lock_bh_state(bh);
 379                         jbd_slab_free(jh->b_committed_data, bh->b_size);
 380                         jh->b_committed_data = NULL;
 381                         jbd_unlock_bh_state(bh);
 382                 }
 383                 journal_refile_buffer(journal, jh);
 384         }
 385
 386         /*
 387          * Now try to drop any written-back buffers from the journal's
 388          * checkpoint lists.  We do this *before* commit because it potentially
 389          * frees some memory
 390          */
 391         spin_lock(&journal->j_list_lock);
 392         __journal_clean_checkpoint_list(journal);
 393         spin_unlock(&journal->j_list_lock);
 394
 395         jbd_debug (3, "JBD: commit phase 1\n");
 396
 397         /*
 398          * Switch to a new revoke table.
 399          */
 400         journal_switch_revoke_table(journal);
 401
 402         commit_transaction->t_state = T_FLUSH;
 403         journal->j_committing_transaction = commit_transaction;
 404         journal->j_running_transaction = NULL;
 405         commit_transaction->t_log_start = journal->j_head;
 406         wake_up(&journal->j_wait_transaction_locked);
 407         spin_unlock(&journal->j_state_lock);
 408
 409         jbd_debug (3, "JBD: commit phase 2\n");
 410
 411         /*
 412          * First, drop modified flag: all accesses to the buffers
 413          * will be tracked for a new trasaction only -bzzz
 414          */
 415         spin_lock(&journal->j_list_lock);
 416         if (commit_transaction->t_buffers) {
 417                 new_jh = jh = commit_transaction->t_buffers->b_tnext;
 418                 do {
 419                         J_ASSERT_JH(new_jh, new_jh->b_modified == 1 ||
 420                                         new_jh->b_modified == 0);
 421                         new_jh->b_modified = 0;
 422                         new_jh = new_jh->b_tnext;
 423                 } while (new_jh != jh);
 424         }
 425         spin_unlock(&journal->j_list_lock);
 426
 427         /*
 428          * Now start flushing things to disk, in the order they appear
 429          * on the transaction lists.  Data blocks go first.
 430          */
 431         err = 0;
 432         journal_submit_data_buffers(journal, commit_transaction);
 433
 434         /*
 435          * Wait for all previously submitted IO to complete.
 436          */
 437         spin_lock(&journal->j_list_lock);
 438         while (commit_transaction->t_locked_list) {
 439                 struct buffer_head *bh;
 440
 441                 jh = commit_transaction->t_locked_list->b_tprev;
 442                 bh = jh2bh(jh);
 443                 get_bh(bh);
 444                 if (buffer_locked(bh)) {
 445                         spin_unlock(&journal->j_list_lock);
 446                         wait_on_buffer(bh);
 447                         if (unlikely(!buffer_uptodate(bh)))
 448                                 err = -EIO;
 449                         spin_lock(&journal->j_list_lock);
 450                 }
 451                 if (!inverted_lock(journal, bh)) {
 452                         put_bh(bh);
 453                         spin_lock(&journal->j_list_lock);
 454                         continue;
 455                 }
 456                 if (buffer_jbd(bh) && jh->b_jlist == BJ_Locked) {
 457                         __journal_unfile_buffer(jh);
 458                         jbd_unlock_bh_state(bh);
 459                         journal_remove_journal_head(bh);
 460                         put_bh(bh);
 461                 } else {
 462                         jbd_unlock_bh_state(bh);
 463                 }
 464                 put_bh(bh);
 465                 cond_resched_lock(&journal->j_list_lock);
 466         }
 467         spin_unlock(&journal->j_list_lock);
 468
 469         if (err)
 470                 __journal_abort_hard(journal);
 471
 472         journal_write_revoke_records(journal, commit_transaction);
 473
 474         jbd_debug(3, "JBD: commit phase 2\n");
 475
 476         /*
 477          * If we found any dirty or locked buffers, then we should have
 478          * looped back up to the write_out_data label.  If there weren't
 479          * any then journal_clean_data_list should have wiped the list
 480          * clean by now, so check that it is in fact empty.
 481          */
 482         J_ASSERT (commit_transaction->t_sync_datalist == NULL);
 483
 484         jbd_debug (3, "JBD: commit phase 3\n");
 485
 486         /*
 487          * Way to go: we have now written out all of the data for a
 488          * transaction!  Now comes the tricky part: we need to write out
 489          * metadata.  Loop over the transaction's entire buffer list:
 490          */
 491         commit_transaction->t_state = T_COMMIT;
 492
 493         descriptor = NULL;
 494         bufs = 0;
 495         while (commit_transaction->t_buffers) {
 496
 497                 /* Find the next buffer to be journaled... */
 498
 499                 jh = commit_transaction->t_buffers;
 500
 501                 /* If we're in abort mode, we just un-journal the buffer and
 502                    release it for background writing. */
 503
 504                 if (is_journal_aborted(journal)) {
 505                         JBUFFER_TRACE(jh, "journal is aborting: refile");
 506                         journal_refile_buffer(journal, jh);
 507                         /* If that was the last one, we need to clean up
 508                          * any descriptor buffers which may have been
 509                          * already allocated, even if we are now
 510                          * aborting. */
 511                         if (!commit_transaction->t_buffers)
 512                                 goto start_journal_io;
 513                         continue;
 514                 }
 515
 516                 /* Make sure we have a descriptor block in which to
 517                    record the metadata buffer. */
 518
 519                 if (!descriptor) {
 520                         struct buffer_head *bh;
 521
 522                         J_ASSERT (bufs == 0);
 523
 524                         jbd_debug(4, "JBD: get descriptor\n");
 525
 526                         descriptor = journal_get_descriptor_buffer(journal);
 527                         if (!descriptor) {
 528                                 __journal_abort_hard(journal);
 529                                 continue;
 530                         }
 531
 532                         bh = jh2bh(descriptor);
 533                         jbd_debug(4, "JBD: got buffer %llu (%p)\n",
 534                                 (unsigned long long)bh->b_blocknr, bh->b_data);
 535                         header = (journal_header_t *)&bh->b_data[0];
 536                         header->h_magic     = cpu_to_be32(JFS_MAGIC_NUMBER);
 537                         header->h_blocktype = cpu_to_be32(JFS_DESCRIPTOR_BLOCK);
 538                         header->h_sequence  = cpu_to_be32(commit_transaction->t_tid);
 539
 540                         tagp = &bh->b_data[sizeof(journal_header_t)];
 541                         space_left = bh->b_size - sizeof(journal_header_t);
 542                         first_tag = 1;
 543                         set_buffer_jwrite(bh);
 544                         set_buffer_dirty(bh);
 545                         wbuf[bufs++] = bh;
 546
 547                         /* Record it so that we can wait for IO
 548                            completion later */
 549                         BUFFER_TRACE(bh, "ph3: file as descriptor");
 550                         journal_file_buffer(descriptor, commit_transaction,
 551                                         BJ_LogCtl);
 552                 }
 553
 554                 /* Where is the buffer to be written? */
 555
 556                 err = journal_next_log_block(journal, &blocknr);
 557                 /* If the block mapping failed, just abandon the buffer
 558                    and repeat this loop: we'll fall into the
 559                    refile-on-abort condition above. */
 560                 if (err) {
 561                         __journal_abort_hard(journal);
 562                         continue;
 563                 }
 564
 565                 /*
 566                  * start_this_handle() uses t_outstanding_credits to determine
 567                  * the free space in the log, but this counter is changed
 568                  * by journal_next_log_block() also.
 569                  */
 570                 commit_transaction->t_outstanding_credits--;
 571
 572                 /* Bump b_count to prevent truncate from stumbling over
 573                    the shadowed buffer!  @@@ This can go if we ever get
 574                    rid of the BJ_IO/BJ_Shadow pairing of buffers. */
 575                 atomic_inc(&jh2bh(jh)->b_count);
 576
 577                 /* Make a temporary IO buffer with which to write it out
 578                    (this will requeue both the metadata buffer and the
 579                    temporary IO buffer). new_bh goes on BJ_IO*/
 580
 581                 set_bit(BH_JWrite, &jh2bh(jh)->b_state);
 582                 /*
 583                  * akpm: journal_write_metadata_buffer() sets
 584                  * new_bh->b_transaction to commit_transaction.
 585                  * We need to clean this up before we release new_bh
 586                  * (which is of type BJ_IO)
 587                  */
 588                 JBUFFER_TRACE(jh, "ph3: write metadata");
 589                 flags = journal_write_metadata_buffer(commit_transaction,
 590                                                       jh, &new_jh, blocknr);
 591                 set_bit(BH_JWrite, &jh2bh(new_jh)->b_state);
 592                 wbuf[bufs++] = jh2bh(new_jh);
 593
 594                 /* Record the new block's tag in the current descriptor
 595                    buffer */
 596
 597                 tag_flag = 0;
 598                 if (flags & 1)
 599                         tag_flag |= JFS_FLAG_ESCAPE;
 600                 if (!first_tag)
 601                         tag_flag |= JFS_FLAG_SAME_UUID;
 602
 603                 tag = (journal_block_tag_t *) tagp;
 604                 tag->t_blocknr = cpu_to_be32(jh2bh(jh)->b_blocknr);
 605                 tag->t_flags = cpu_to_be32(tag_flag);
 606                 tagp += sizeof(journal_block_tag_t);
 607                 space_left -= sizeof(journal_block_tag_t);
 608
 609                 if (first_tag) {
 610                         memcpy (tagp, journal->j_uuid, 16);
 611                         tagp += 16;
 612                         space_left -= 16;
 613                         first_tag = 0;
 614                 }
 615
 616                 /* If there's no more to do, or if the descriptor is full,
 617                    let the IO rip! */
 618
 619                 if (bufs == journal->j_wbufsize ||
 620                     commit_transaction->t_buffers == NULL ||
 621                     space_left < sizeof(journal_block_tag_t) + 16) {
 622
 623                         jbd_debug(4, "JBD: Submit %d IOs\n", bufs);
 624
 625                         /* Write an end-of-descriptor marker before
 626                            submitting the IOs.  "tag" still points to
 627                            the last tag we set up. */
 628
 629                         tag->t_flags |= cpu_to_be32(JFS_FLAG_LAST_TAG);
 630
 631 start_journal_io:
 632                         for (i = 0; i < bufs; i++) {
 633                                 struct buffer_head *bh = wbuf[i];
 634                                 lock_buffer(bh);
 635                                 clear_buffer_dirty(bh);
 636                                 set_buffer_uptodate(bh);
 637                                 bh->b_end_io = journal_end_buffer_io_sync;
 638                                 submit_bh(WRITE, bh);
 639                         }
 640                         cond_resched();
 641
 642                         /* Force a new descriptor to be generated next
 643                            time round the loop. */
 644                         descriptor = NULL;
 645                         bufs = 0;
 646                 }
 647         }
 648
 649         /* Lo and behold: we have just managed to send a transaction to
 650            the log.  Before we can commit it, wait for the IO so far to
 651            complete.  Control buffers being written are on the
 652            transaction's t_log_list queue, and metadata buffers are on
 653            the t_iobuf_list queue.
 654
 655            Wait for the buffers in reverse order.  That way we are
 656            less likely to be woken up until all IOs have completed, and
 657            so we incur less scheduling load.
 658         */
 659
 660         jbd_debug(3, "JBD: commit phase 4\n");
 661
 662         /*
 663          * akpm: these are BJ_IO, and j_list_lock is not needed.
 664          * See __journal_try_to_free_buffer.
 665          */
 666 wait_for_iobuf:
 667         while (commit_transaction->t_iobuf_list != NULL) {
 668                 struct buffer_head *bh;
 669
 670                 jh = commit_transaction->t_iobuf_list->b_tprev;
 671                 bh = jh2bh(jh);
 672                 if (buffer_locked(bh)) {
 673                         wait_on_buffer(bh);
 674                         goto wait_for_iobuf;
 675                 }
 676                 if (cond_resched())
 677                         goto wait_for_iobuf;
 678
 679                 if (unlikely(!buffer_uptodate(bh)))
 680                         err = -EIO;
 681
 682                 clear_buffer_jwrite(bh);
 683
 684                 JBUFFER_TRACE(jh, "ph4: unfile after journal write");
 685                 journal_unfile_buffer(journal, jh);
 686
 687                 /*
 688                  * ->t_iobuf_list should contain only dummy buffer_heads
 689                  * which were created by journal_write_metadata_buffer().
 690                  */
 691                 BUFFER_TRACE(bh, "dumping temporary bh");
 692                 journal_put_journal_head(jh);
 693                 __brelse(bh);
 694                 J_ASSERT_BH(bh, atomic_read(&bh->b_count) == 0);
 695                 free_buffer_head(bh);
 696
 697                 /* We also have to unlock and free the corresponding
 698                    shadowed buffer */
 699                 jh = commit_transaction->t_shadow_list->b_tprev;
 700                 bh = jh2bh(jh);
 701                 clear_bit(BH_JWrite, &bh->b_state);
 702                 J_ASSERT_BH(bh, buffer_jbddirty(bh));
 703
 704                 /* The metadata is now released for reuse, but we need
 705                    to remember it against this transaction so that when
 706                    we finally commit, we can do any checkpointing
 707                    required. */
 708                 JBUFFER_TRACE(jh, "file as BJ_Forget");
 709                 journal_file_buffer(jh, commit_transaction, BJ_Forget);
 710                 /* Wake up any transactions which were waiting for this
 711                    IO to complete */
 712                 wake_up_bit(&bh->b_state, BH_Unshadow);
 713                 JBUFFER_TRACE(jh, "brelse shadowed buffer");
 714                 __brelse(bh);
 715         }
 716
 717         J_ASSERT (commit_transaction->t_shadow_list == NULL);
 718
 719         jbd_debug(3, "JBD: commit phase 5\n");
 720
 721         /* Here we wait for the revoke record and descriptor record buffers */
 722  wait_for_ctlbuf:
 723         while (commit_transaction->t_log_list != NULL) {
 724                 struct buffer_head *bh;
 725
 726                 jh = commit_transaction->t_log_list->b_tprev;
 727                 bh = jh2bh(jh);
 728                 if (buffer_locked(bh)) {
 729                         wait_on_buffer(bh);
 730                         goto wait_for_ctlbuf;
 731                 }
 732                 if (cond_resched())
 733                         goto wait_for_ctlbuf;
 734
 735                 if (unlikely(!buffer_uptodate(bh)))
 736                         err = -EIO;
 737
 738                 BUFFER_TRACE(bh, "ph5: control buffer writeout done: unfile");
 739                 clear_buffer_jwrite(bh);
 740                 journal_unfile_buffer(journal, jh);
 741                 journal_put_journal_head(jh);
 742                 __brelse(bh);           /* One for getblk */
 743                 /* AKPM: bforget here */
 744         }
 745
 746         jbd_debug(3, "JBD: commit phase 6\n");
 747
 748         if (journal_write_commit_record(journal, commit_transaction))
 749                 err = -EIO;
 750
 751         if (err)
 752                 __journal_abort_hard(journal);
 753
 754         /* End of a transaction!  Finally, we can do checkpoint
 755            processing: any buffers committed as a result of this
 756            transaction can be removed from any checkpoint list it was on
 757            before. */
 758
 759         jbd_debug(3, "JBD: commit phase 7\n");
 760
 761         J_ASSERT(commit_transaction->t_sync_datalist == NULL);
 762         J_ASSERT(commit_transaction->t_buffers == NULL);
 763         J_ASSERT(commit_transaction->t_checkpoint_list == NULL);
 764         J_ASSERT(commit_transaction->t_iobuf_list == NULL);
 765         J_ASSERT(commit_transaction->t_shadow_list == NULL);
 766         J_ASSERT(commit_transaction->t_log_list == NULL);
 767
 768 restart_loop:
 769         /*
 770          * As there are other places (journal_unmap_buffer()) adding buffers
 771          * to this list we have to be careful and hold the j_list_lock.
 772          */
 773         spin_lock(&journal->j_list_lock);
 774         while (commit_transaction->t_forget) {
 775                 transaction_t *cp_transaction;
 776                 struct buffer_head *bh;
 777
 778                 jh = commit_transaction->t_forget;
 779                 spin_unlock(&journal->j_list_lock);
 780                 bh = jh2bh(jh);
 781                 jbd_lock_bh_state(bh);
 782                 J_ASSERT_JH(jh, jh->b_transaction == commit_transaction ||
 783                         jh->b_transaction == journal->j_running_transaction);
 784
 785                 /*
 786                  * If there is undo-protected committed data against
 787                  * this buffer, then we can remove it now.  If it is a
 788                  * buffer needing such protection, the old frozen_data
 789                  * field now points to a committed version of the
 790                  * buffer, so rotate that field to the new committed
 791                  * data.
 792                  *
 793                  * Otherwise, we can just throw away the frozen data now.
 794                  */
 795                 if (jh->b_committed_data) {
 796                         jbd_slab_free(jh->b_committed_data, bh->b_size);
 797                         jh->b_committed_data = NULL;
 798                         if (jh->b_frozen_data) {
 799                                 jh->b_committed_data = jh->b_frozen_data;
 800                                 jh->b_frozen_data = NULL;
 801                         }
 802                 } else if (jh->b_frozen_data) {
 803                         jbd_slab_free(jh->b_frozen_data, bh->b_size);
 804                         jh->b_frozen_data = NULL;
 805                 }
 806
 807                 spin_lock(&journal->j_list_lock);
 808                 cp_transaction = jh->b_cp_transaction;
 809                 if (cp_transaction) {
 810                         JBUFFER_TRACE(jh, "remove from old cp transaction");
 811                         __journal_remove_checkpoint(jh);
 812                 }
 813
 814                 /* Only re-checkpoint the buffer_head if it is marked
 815                  * dirty.  If the buffer was added to the BJ_Forget list
 816                  * by journal_forget, it may no longer be dirty and
 817                  * there's no point in keeping a checkpoint record for
 818                  * it. */
 819
 820                 /* A buffer which has been freed while still being
 821                  * journaled by a previous transaction may end up still
 822                  * being dirty here, but we want to avoid writing back
 823                  * that buffer in the future now that the last use has
 824                  * been committed.  That's not only a performance gain,
 825                  * it also stops aliasing problems if the buffer is left
 826                  * behind for writeback and gets reallocated for another
 827                  * use in a different page. */
 828                 if (buffer_freed(bh)) {
 829                         clear_buffer_freed(bh);
 830                         clear_buffer_jbddirty(bh);
 831                 }
 832
 833                 if (buffer_jbddirty(bh)) {
 834                         JBUFFER_TRACE(jh, "add to new checkpointing trans");
 835                         __journal_insert_checkpoint(jh, commit_transaction);
 836                         JBUFFER_TRACE(jh, "refile for checkpoint writeback");
 837                         __journal_refile_buffer(jh);
 838                         jbd_unlock_bh_state(bh);
 839                 } else {
 840                         J_ASSERT_BH(bh, !buffer_dirty(bh));
 841                         /* The buffer on BJ_Forget list and not jbddirty means
 842                          * it has been freed by this transaction and hence it
 843                          * could not have been reallocated until this
 844                          * transaction has committed. *BUT* it could be
 845                          * reallocated once we have written all the data to
 846                          * disk and before we process the buffer on BJ_Forget
 847                          * list. */
 848                         JBUFFER_TRACE(jh, "refile or unfile freed buffer");
 849                         __journal_refile_buffer(jh);
 850                         if (!jh->b_transaction) {
 851                                 jbd_unlock_bh_state(bh);
 852                                  /* needs a brelse */
 853                                 journal_remove_journal_head(bh);
 854                                 release_buffer_page(bh);
 855                         } else
 856                                 jbd_unlock_bh_state(bh);
 857                 }
 858                 cond_resched_lock(&journal->j_list_lock);
 859         }
 860         spin_unlock(&journal->j_list_lock);
 861         /*
 862          * This is a bit sleazy.  We borrow j_list_lock to protect
 863          * journal->j_committing_transaction in __journal_remove_checkpoint.
 864          * Really, __journal_remove_checkpoint should be using j_state_lock but
 865          * it's a bit hassle to hold that across __journal_remove_checkpoint
 866          */
 867         spin_lock(&journal->j_state_lock);
 868         spin_lock(&journal->j_list_lock);
 869         /*
 870          * Now recheck if some buffers did not get attached to the transaction
 871          * while the lock was dropped...
 872          */
 873         if (commit_transaction->t_forget) {
 874                 spin_unlock(&journal->j_list_lock);
 875                 spin_unlock(&journal->j_state_lock);
 876                 goto restart_loop;
 877         }
 878
 879         /* Done with this transaction! */
 880
 881         jbd_debug(3, "JBD: commit phase 8\n");
 882
 883         J_ASSERT(commit_transaction->t_state == T_COMMIT);
 884
 885         commit_transaction->t_state = T_FINISHED;
 886         J_ASSERT(commit_transaction == journal->j_committing_transaction);
 887         journal->j_commit_sequence = commit_transaction->t_tid;
 888         journal->j_committing_transaction = NULL;
 889         spin_unlock(&journal->j_state_lock);
 890
 891         if (commit_transaction->t_checkpoint_list == NULL) {
 892                 __journal_drop_transaction(journal, commit_transaction);
 893         } else {
 894                 if (journal->j_checkpoint_transactions == NULL) {
 895                         journal->j_checkpoint_transactions = commit_transaction;
 896                         commit_transaction->t_cpnext = commit_transaction;
 897                         commit_transaction->t_cpprev = commit_transaction;
 898                 } else {
 899                         commit_transaction->t_cpnext =
 900                                 journal->j_checkpoint_transactions;
 901                         commit_transaction->t_cpprev =
 902                                 commit_transaction->t_cpnext->t_cpprev;
 903                         commit_transaction->t_cpnext->t_cpprev =
 904                                 commit_transaction;
 905                         commit_transaction->t_cpprev->t_cpnext =
 906                                 commit_transaction;
 907                 }
 908         }
 909         spin_unlock(&journal->j_list_lock);
 910
 911         jbd_debug(1, "JBD: commit %d complete, head %d\n",
 912                   journal->j_commit_sequence, journal->j_tail_sequence);
 913
 914         wake_up(&journal->j_wait_done_commit);
 915 }