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 successfully 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 journal->j_list_lock.  The caller provided us with a ref
  49  * 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  * Decrement reference counter for data buffer. If it has been marked
  82  * 'BH_Freed', release it and the page to which it belongs if possible.
  83  */
  84 static void release_data_buffer(struct buffer_head *bh)
  85 {
  86         if (buffer_freed(bh)) {
  87                 clear_buffer_freed(bh);
  88                 release_buffer_page(bh);
  89         } else
  90                 put_bh(bh);
  91 }
  92
  93 /*
  94  * Try to acquire jbd_lock_bh_state() against the buffer, when j_list_lock is
  95  * held.  For ranking reasons we must trylock.  If we lose, schedule away and
  96  * return 0.  j_list_lock is dropped in this case.
  97  */
  98 static int inverted_lock(journal_t *journal, struct buffer_head *bh)
  99 {
 100         if (!jbd_trylock_bh_state(bh)) {
 101                 spin_unlock(&journal->j_list_lock);
 102                 schedule();
 103                 return 0;
 104         }
 105         return 1;
 106 }
 107
 108 /* Done it all: now write the commit record.  We should have
 109  * cleaned up our previous buffers by now, so if we are in abort
 110  * mode we can now just skip the rest of the journal write
 111  * entirely.
 112  *
 113  * Returns 1 if the journal needs to be aborted or 0 on success
 114  */
 115 static int journal_write_commit_record(journal_t *journal,
 116                                         transaction_t *commit_transaction)
 117 {
 118         struct journal_head *descriptor;
 119         struct buffer_head *bh;
 120         journal_header_t *header;
 121         int ret;
 122         int barrier_done = 0;
 123
 124         if (is_journal_aborted(journal))
 125                 return 0;
 126
 127         descriptor = journal_get_descriptor_buffer(journal);
 128         if (!descriptor)
 129                 return 1;
 130
 131         bh = jh2bh(descriptor);
 132
 133         header = (journal_header_t *)(bh->b_data);
 134         header->h_magic = cpu_to_be32(JFS_MAGIC_NUMBER);
 135         header->h_blocktype = cpu_to_be32(JFS_COMMIT_BLOCK);
 136         header->h_sequence = cpu_to_be32(commit_transaction->t_tid);
 137
 138         JBUFFER_TRACE(descriptor, "write commit block");
 139         set_buffer_dirty(bh);
 140         if (journal->j_flags & JFS_BARRIER) {
 141                 set_buffer_ordered(bh);
 142                 barrier_done = 1;
 143         }
 144         ret = sync_dirty_buffer(bh);
 145         if (barrier_done)
 146                 clear_buffer_ordered(bh);
 147         /* is it possible for another commit to fail at roughly
 148          * the same time as this one?  If so, we don't want to
 149          * trust the barrier flag in the super, but instead want
 150          * to remember if we sent a barrier request
 151          */
 152         if (ret == -EOPNOTSUPP && barrier_done) {
 153                 char b[BDEVNAME_SIZE];
 154
 155                 printk(KERN_WARNING
 156                         "JBD: barrier-based sync failed on %s - "
 157                         "disabling barriers\n",
 158                         bdevname(journal->j_dev, b));
 159                 spin_lock(&journal->j_state_lock);
 160                 journal->j_flags &= ~JFS_BARRIER;
 161                 spin_unlock(&journal->j_state_lock);
 162
 163                 /* And try again, without the barrier */
 164                 set_buffer_uptodate(bh);
 165                 set_buffer_dirty(bh);
 166                 ret = sync_dirty_buffer(bh);
 167         }
 168         put_bh(bh);             /* One for getblk() */
 169         journal_put_journal_head(descriptor);
 170
 171         return (ret == -EIO);
 172 }
 173
 174 static void journal_do_submit_data(struct buffer_head **wbuf, int bufs)
 175 {
 176         int i;
 177
 178         for (i = 0; i < bufs; i++) {
 179                 wbuf[i]->b_end_io = end_buffer_write_sync;
 180                 /* We use-up our safety reference in submit_bh() */
 181                 submit_bh(WRITE, wbuf[i]);
 182         }
 183 }
 184
 185 /*
 186  *  Submit all the data buffers to disk
 187  */
 188 static void journal_submit_data_buffers(journal_t *journal,
 189                                 transaction_t *commit_transaction)
 190 {
 191         struct journal_head *jh;
 192         struct buffer_head *bh;
 193         int locked;
 194         int bufs = 0;
 195         struct buffer_head **wbuf = journal->j_wbuf;
 196
 197         /*
 198          * Whenever we unlock the journal and sleep, things can get added
 199          * onto ->t_sync_datalist, so we have to keep looping back to
 200          * write_out_data until we *know* that the list is empty.
 201          *
 202          * Cleanup any flushed data buffers from the data list.  Even in
 203          * abort mode, we want to flush this out as soon as possible.
 204          */
 205 write_out_data:
 206         cond_resched();
 207         spin_lock(&journal->j_list_lock);
 208
 209         while (commit_transaction->t_sync_datalist) {
 210                 jh = commit_transaction->t_sync_datalist;
 211                 bh = jh2bh(jh);
 212                 locked = 0;
 213
 214                 /* Get reference just to make sure buffer does not disappear
 215                  * when we are forced to drop various locks */
 216                 get_bh(bh);
 217                 /* If the buffer is dirty, we need to submit IO and hence
 218                  * we need the buffer lock. We try to lock the buffer without
 219                  * blocking. If we fail, we need to drop j_list_lock and do
 220                  * blocking lock_buffer().
 221                  */
 222                 if (buffer_dirty(bh)) {
 223                         if (test_set_buffer_locked(bh)) {
 224                                 BUFFER_TRACE(bh, "needs blocking lock");
 225                                 spin_unlock(&journal->j_list_lock);
 226                                 /* Write out all data to prevent deadlocks */
 227                                 journal_do_submit_data(wbuf, bufs);
 228                                 bufs = 0;
 229                                 lock_buffer(bh);
 230                                 spin_lock(&journal->j_list_lock);
 231                         }
 232                         locked = 1;
 233                 }
 234                 /* We have to get bh_state lock. Again out of order, sigh. */
 235                 if (!inverted_lock(journal, bh)) {
 236                         jbd_lock_bh_state(bh);
 237                         spin_lock(&journal->j_list_lock);
 238                 }
 239                 /* Someone already cleaned up the buffer? */
 240                 if (!buffer_jbd(bh)
 241                         || jh->b_transaction != commit_transaction
 242                         || jh->b_jlist != BJ_SyncData) {
 243                         jbd_unlock_bh_state(bh);
 244                         if (locked)
 245                                 unlock_buffer(bh);
 246                         BUFFER_TRACE(bh, "already cleaned up");
 247                         release_data_buffer(bh);
 248                         continue;
 249                 }
 250                 if (locked && test_clear_buffer_dirty(bh)) {
 251                         BUFFER_TRACE(bh, "needs writeout, adding to array");
 252                         wbuf[bufs++] = bh;
 253                         __journal_file_buffer(jh, commit_transaction,
 254                                                 BJ_Locked);
 255                         jbd_unlock_bh_state(bh);
 256                         if (bufs == journal->j_wbufsize) {
 257                                 spin_unlock(&journal->j_list_lock);
 258                                 journal_do_submit_data(wbuf, bufs);
 259                                 bufs = 0;
 260                                 goto write_out_data;
 261                         }
 262                 } else if (!locked && buffer_locked(bh)) {
 263                         __journal_file_buffer(jh, commit_transaction,
 264                                                 BJ_Locked);
 265                         jbd_unlock_bh_state(bh);
 266                         put_bh(bh);
 267                 } else {
 268                         BUFFER_TRACE(bh, "writeout complete: unfile");
 269                         __journal_unfile_buffer(jh);
 270                         jbd_unlock_bh_state(bh);
 271                         if (locked)
 272                                 unlock_buffer(bh);
 273                         journal_remove_journal_head(bh);
 274                         /* One for our safety reference, other for
 275                          * journal_remove_journal_head() */
 276                         put_bh(bh);
 277                         release_data_buffer(bh);
 278                 }
 279
 280                 if (need_resched() || spin_needbreak(&journal->j_list_lock)) {
 281                         spin_unlock(&journal->j_list_lock);
 282                         goto write_out_data;
 283                 }
 284         }
 285         spin_unlock(&journal->j_list_lock);
 286         journal_do_submit_data(wbuf, bufs);
 287 }
 288
 289 /*
 290  * journal_commit_transaction
 291  *
 292  * The primary function for committing a transaction to the log.  This
 293  * function is called by the journal thread to begin a complete commit.
 294  */
 295 void journal_commit_transaction(journal_t *journal)
 296 {
 297         transaction_t *commit_transaction;
 298         struct journal_head *jh, *new_jh, *descriptor;
 299         struct buffer_head **wbuf = journal->j_wbuf;
 300         int bufs;
 301         int flags;
 302         int err;
 303         unsigned long blocknr;
 304         char *tagp = NULL;
 305         journal_header_t *header;
 306         journal_block_tag_t *tag = NULL;
 307         int space_left = 0;
 308         int first_tag = 0;
 309         int tag_flag;
 310         int i;
 311
 312         /*
 313          * First job: lock down the current transaction and wait for
 314          * all outstanding updates to complete.
 315          */
 316
 317 #ifdef COMMIT_STATS
 318         spin_lock(&journal->j_list_lock);
 319         summarise_journal_usage(journal);
 320         spin_unlock(&journal->j_list_lock);
 321 #endif
 322
 323         /* Do we need to erase the effects of a prior journal_flush? */
 324         if (journal->j_flags & JFS_FLUSHED) {
 325                 jbd_debug(3, "super block updated\n");
 326                 journal_update_superblock(journal, 1);
 327         } else {
 328                 jbd_debug(3, "superblock not updated\n");
 329         }
 330
 331         J_ASSERT(journal->j_running_transaction != NULL);
 332         J_ASSERT(journal->j_committing_transaction == NULL);
 333
 334         commit_transaction = journal->j_running_transaction;
 335         J_ASSERT(commit_transaction->t_state == T_RUNNING);
 336
 337         jbd_debug(1, "JBD: starting commit of transaction %d\n",
 338                         commit_transaction->t_tid);
 339
 340         spin_lock(&journal->j_state_lock);
 341         commit_transaction->t_state = T_LOCKED;
 342
 343         spin_lock(&commit_transaction->t_handle_lock);
 344         while (commit_transaction->t_updates) {
 345                 DEFINE_WAIT(wait);
 346
 347                 prepare_to_wait(&journal->j_wait_updates, &wait,
 348                                         TASK_UNINTERRUPTIBLE);
 349                 if (commit_transaction->t_updates) {
 350                         spin_unlock(&commit_transaction->t_handle_lock);
 351                         spin_unlock(&journal->j_state_lock);
 352                         schedule();
 353                         spin_lock(&journal->j_state_lock);
 354                         spin_lock(&commit_transaction->t_handle_lock);
 355                 }
 356                 finish_wait(&journal->j_wait_updates, &wait);
 357         }
 358         spin_unlock(&commit_transaction->t_handle_lock);
 359
 360         J_ASSERT (commit_transaction->t_outstanding_credits <=
 361                         journal->j_max_transaction_buffers);
 362
 363         /*
 364          * First thing we are allowed to do is to discard any remaining
 365          * BJ_Reserved buffers.  Note, it is _not_ permissible to assume
 366          * that there are no such buffers: if a large filesystem
 367          * operation like a truncate needs to split itself over multiple
 368          * transactions, then it may try to do a journal_restart() while
 369          * there are still BJ_Reserved buffers outstanding.  These must
 370          * be released cleanly from the current transaction.
 371          *
 372          * In this case, the filesystem must still reserve write access
 373          * again before modifying the buffer in the new transaction, but
 374          * we do not require it to remember exactly which old buffers it
 375          * has reserved.  This is consistent with the existing behaviour
 376          * that multiple journal_get_write_access() calls to the same
 377          * buffer are perfectly permissable.
 378          */
 379         while (commit_transaction->t_reserved_list) {
 380                 jh = commit_transaction->t_reserved_list;
 381                 JBUFFER_TRACE(jh, "reserved, unused: refile");
 382                 /*
 383                  * A journal_get_undo_access()+journal_release_buffer() may
 384                  * leave undo-committed data.
 385                  */
 386                 if (jh->b_committed_data) {
 387                         struct buffer_head *bh = jh2bh(jh);
 388
 389                         jbd_lock_bh_state(bh);
 390                         jbd_free(jh->b_committed_data, bh->b_size);
 391                         jh->b_committed_data = NULL;
 392                         jbd_unlock_bh_state(bh);
 393                 }
 394                 journal_refile_buffer(journal, jh);
 395         }
 396
 397         /*
 398          * Now try to drop any written-back buffers from the journal's
 399          * checkpoint lists.  We do this *before* commit because it potentially
 400          * frees some memory
 401          */
 402         spin_lock(&journal->j_list_lock);
 403         __journal_clean_checkpoint_list(journal);
 404         spin_unlock(&journal->j_list_lock);
 405
 406         jbd_debug (3, "JBD: commit phase 1\n");
 407
 408         /*
 409          * Switch to a new revoke table.
 410          */
 411         journal_switch_revoke_table(journal);
 412
 413         commit_transaction->t_state = T_FLUSH;
 414         journal->j_committing_transaction = commit_transaction;
 415         journal->j_running_transaction = NULL;
 416         commit_transaction->t_log_start = journal->j_head;
 417         wake_up(&journal->j_wait_transaction_locked);
 418         spin_unlock(&journal->j_state_lock);
 419
 420         jbd_debug (3, "JBD: commit phase 2\n");
 421
 422         /*
 423          * Now start flushing things to disk, in the order they appear
 424          * on the transaction lists.  Data blocks go first.
 425          */
 426         err = 0;
 427         journal_submit_data_buffers(journal, commit_transaction);
 428
 429         /*
 430          * Wait for all previously submitted IO to complete.
 431          */
 432         spin_lock(&journal->j_list_lock);
 433         while (commit_transaction->t_locked_list) {
 434                 struct buffer_head *bh;
 435
 436                 jh = commit_transaction->t_locked_list->b_tprev;
 437                 bh = jh2bh(jh);
 438                 get_bh(bh);
 439                 if (buffer_locked(bh)) {
 440                         spin_unlock(&journal->j_list_lock);
 441                         wait_on_buffer(bh);
 442                         if (unlikely(!buffer_uptodate(bh)))
 443                                 err = -EIO;
 444                         spin_lock(&journal->j_list_lock);
 445                 }
 446                 if (!inverted_lock(journal, bh)) {
 447                         put_bh(bh);
 448                         spin_lock(&journal->j_list_lock);
 449                         continue;
 450                 }
 451                 if (buffer_jbd(bh) && jh->b_jlist == BJ_Locked) {
 452                         __journal_unfile_buffer(jh);
 453                         jbd_unlock_bh_state(bh);
 454                         journal_remove_journal_head(bh);
 455                         put_bh(bh);
 456                 } else {
 457                         jbd_unlock_bh_state(bh);
 458                 }
 459                 release_data_buffer(bh);
 460                 cond_resched_lock(&journal->j_list_lock);
 461         }
 462         spin_unlock(&journal->j_list_lock);
 463
 464         if (err)
 465                 journal_abort(journal, err);
 466
 467         journal_write_revoke_records(journal, commit_transaction);
 468
 469         /*
 470          * If we found any dirty or locked buffers, then we should have
 471          * looped back up to the write_out_data label.  If there weren't
 472          * any then journal_clean_data_list should have wiped the list
 473          * clean by now, so check that it is in fact empty.
 474          */
 475         J_ASSERT (commit_transaction->t_sync_datalist == NULL);
 476
 477         jbd_debug (3, "JBD: commit phase 3\n");
 478
 479         /*
 480          * Way to go: we have now written out all of the data for a
 481          * transaction!  Now comes the tricky part: we need to write out
 482          * metadata.  Loop over the transaction's entire buffer list:
 483          */
 484         spin_lock(&journal->j_state_lock);
 485         commit_transaction->t_state = T_COMMIT;
 486         spin_unlock(&journal->j_state_lock);
 487
 488         J_ASSERT(commit_transaction->t_nr_buffers <=
 489                  commit_transaction->t_outstanding_credits);
 490
 491         descriptor = NULL;
 492         bufs = 0;
 493         while (commit_transaction->t_buffers) {
 494
 495                 /* Find the next buffer to be journaled... */
 496
 497                 jh = commit_transaction->t_buffers;
 498
 499                 /* If we're in abort mode, we just un-journal the buffer and
 500                    release it for background writing. */
 501
 502                 if (is_journal_aborted(journal)) {
 503                         JBUFFER_TRACE(jh, "journal is aborting: refile");
 504                         journal_refile_buffer(journal, jh);
 505                         /* If that was the last one, we need to clean up
 506                          * any descriptor buffers which may have been
 507                          * already allocated, even if we are now
 508                          * aborting. */
 509                         if (!commit_transaction->t_buffers)
 510                                 goto start_journal_io;
 511                         continue;
 512                 }
 513
 514                 /* Make sure we have a descriptor block in which to
 515                    record the metadata buffer. */
 516
 517                 if (!descriptor) {
 518                         struct buffer_head *bh;
 519
 520                         J_ASSERT (bufs == 0);
 521
 522                         jbd_debug(4, "JBD: get descriptor\n");
 523
 524                         descriptor = journal_get_descriptor_buffer(journal);
 525                         if (!descriptor) {
 526                                 journal_abort(journal, -EIO);
 527                                 continue;
 528                         }
 529
 530                         bh = jh2bh(descriptor);
 531                         jbd_debug(4, "JBD: got buffer %llu (%p)\n",
 532                                 (unsigned long long)bh->b_blocknr, bh->b_data);
 533                         header = (journal_header_t *)&bh->b_data[0];
 534                         header->h_magic     = cpu_to_be32(JFS_MAGIC_NUMBER);
 535                         header->h_blocktype = cpu_to_be32(JFS_DESCRIPTOR_BLOCK);
 536                         header->h_sequence  = cpu_to_be32(commit_transaction->t_tid);
 537
 538                         tagp = &bh->b_data[sizeof(journal_header_t)];
 539                         space_left = bh->b_size - sizeof(journal_header_t);
 540                         first_tag = 1;
 541                         set_buffer_jwrite(bh);
 542                         set_buffer_dirty(bh);
 543                         wbuf[bufs++] = bh;
 544
 545                         /* Record it so that we can wait for IO
 546                            completion later */
 547                         BUFFER_TRACE(bh, "ph3: file as descriptor");
 548                         journal_file_buffer(descriptor, commit_transaction,
 549                                         BJ_LogCtl);
 550                 }
 551
 552                 /* Where is the buffer to be written? */
 553
 554                 err = journal_next_log_block(journal, &blocknr);
 555                 /* If the block mapping failed, just abandon the buffer
 556                    and repeat this loop: we'll fall into the
 557                    refile-on-abort condition above. */
 558                 if (err) {
 559                         journal_abort(journal, err);
 560                         continue;
 561                 }
 562
 563                 /*
 564                  * start_this_handle() uses t_outstanding_credits to determine
 565                  * the free space in the log, but this counter is changed
 566                  * by journal_next_log_block() also.
 567                  */
 568                 commit_transaction->t_outstanding_credits--;
 569
 570                 /* Bump b_count to prevent truncate from stumbling over
 571                    the shadowed buffer!  @@@ This can go if we ever get
 572                    rid of the BJ_IO/BJ_Shadow pairing of buffers. */
 573                 atomic_inc(&jh2bh(jh)->b_count);
 574
 575                 /* Make a temporary IO buffer with which to write it out
 576                    (this will requeue both the metadata buffer and the
 577                    temporary IO buffer). new_bh goes on BJ_IO*/
 578
 579                 set_bit(BH_JWrite, &jh2bh(jh)->b_state);
 580                 /*
 581                  * akpm: journal_write_metadata_buffer() sets
 582                  * new_bh->b_transaction to commit_transaction.
 583                  * We need to clean this up before we release new_bh
 584                  * (which is of type BJ_IO)
 585                  */
 586                 JBUFFER_TRACE(jh, "ph3: write metadata");
 587                 flags = journal_write_metadata_buffer(commit_transaction,
 588                                                       jh, &new_jh, blocknr);
 589                 set_bit(BH_JWrite, &jh2bh(new_jh)->b_state);
 590                 wbuf[bufs++] = jh2bh(new_jh);
 591
 592                 /* Record the new block's tag in the current descriptor
 593                    buffer */
 594
 595                 tag_flag = 0;
 596                 if (flags & 1)
 597                         tag_flag |= JFS_FLAG_ESCAPE;
 598                 if (!first_tag)
 599                         tag_flag |= JFS_FLAG_SAME_UUID;
 600
 601                 tag = (journal_block_tag_t *) tagp;
 602                 tag->t_blocknr = cpu_to_be32(jh2bh(jh)->b_blocknr);
 603                 tag->t_flags = cpu_to_be32(tag_flag);
 604                 tagp += sizeof(journal_block_tag_t);
 605                 space_left -= sizeof(journal_block_tag_t);
 606
 607                 if (first_tag) {
 608                         memcpy (tagp, journal->j_uuid, 16);
 609                         tagp += 16;
 610                         space_left -= 16;
 611                         first_tag = 0;
 612                 }
 613
 614                 /* If there's no more to do, or if the descriptor is full,
 615                    let the IO rip! */
 616
 617                 if (bufs == journal->j_wbufsize ||
 618                     commit_transaction->t_buffers == NULL ||
 619                     space_left < sizeof(journal_block_tag_t) + 16) {
 620
 621                         jbd_debug(4, "JBD: Submit %d IOs\n", bufs);
 622
 623                         /* Write an end-of-descriptor marker before
 624                            submitting the IOs.  "tag" still points to
 625                            the last tag we set up. */
 626
 627                         tag->t_flags |= cpu_to_be32(JFS_FLAG_LAST_TAG);
 628
 629 start_journal_io:
 630                         for (i = 0; i < bufs; i++) {
 631                                 struct buffer_head *bh = wbuf[i];
 632                                 lock_buffer(bh);
 633                                 clear_buffer_dirty(bh);
 634                                 set_buffer_uptodate(bh);
 635                                 bh->b_end_io = journal_end_buffer_io_sync;
 636                                 submit_bh(WRITE, bh);
 637                         }
 638                         cond_resched();
 639
 640                         /* Force a new descriptor to be generated next
 641                            time round the loop. */
 642                         descriptor = NULL;
 643                         bufs = 0;
 644                 }
 645         }
 646
 647         /* Lo and behold: we have just managed to send a transaction to
 648            the log.  Before we can commit it, wait for the IO so far to
 649            complete.  Control buffers being written are on the
 650            transaction's t_log_list queue, and metadata buffers are on
 651            the t_iobuf_list queue.
 652
 653            Wait for the buffers in reverse order.  That way we are
 654            less likely to be woken up until all IOs have completed, and
 655            so we incur less scheduling load.
 656         */
 657
 658         jbd_debug(3, "JBD: commit phase 4\n");
 659
 660         /*
 661          * akpm: these are BJ_IO, and j_list_lock is not needed.
 662          * See __journal_try_to_free_buffer.
 663          */
 664 wait_for_iobuf:
 665         while (commit_transaction->t_iobuf_list != NULL) {
 666                 struct buffer_head *bh;
 667
 668                 jh = commit_transaction->t_iobuf_list->b_tprev;
 669                 bh = jh2bh(jh);
 670                 if (buffer_locked(bh)) {
 671                         wait_on_buffer(bh);
 672                         goto wait_for_iobuf;
 673                 }
 674                 if (cond_resched())
 675                         goto wait_for_iobuf;
 676
 677                 if (unlikely(!buffer_uptodate(bh)))
 678                         err = -EIO;
 679
 680                 clear_buffer_jwrite(bh);
 681
 682                 JBUFFER_TRACE(jh, "ph4: unfile after journal write");
 683                 journal_unfile_buffer(journal, jh);
 684
 685                 /*
 686                  * ->t_iobuf_list should contain only dummy buffer_heads
 687                  * which were created by journal_write_metadata_buffer().
 688                  */
 689                 BUFFER_TRACE(bh, "dumping temporary bh");
 690                 journal_put_journal_head(jh);
 691                 __brelse(bh);
 692                 J_ASSERT_BH(bh, atomic_read(&bh->b_count) == 0);
 693                 free_buffer_head(bh);
 694
 695                 /* We also have to unlock and free the corresponding
 696                    shadowed buffer */
 697                 jh = commit_transaction->t_shadow_list->b_tprev;
 698                 bh = jh2bh(jh);
 699                 clear_bit(BH_JWrite, &bh->b_state);
 700                 J_ASSERT_BH(bh, buffer_jbddirty(bh));
 701
 702                 /* The metadata is now released for reuse, but we need
 703                    to remember it against this transaction so that when
 704                    we finally commit, we can do any checkpointing
 705                    required. */
 706                 JBUFFER_TRACE(jh, "file as BJ_Forget");
 707                 journal_file_buffer(jh, commit_transaction, BJ_Forget);
 708                 /* Wake up any transactions which were waiting for this
 709                    IO to complete */
 710                 wake_up_bit(&bh->b_state, BH_Unshadow);
 711                 JBUFFER_TRACE(jh, "brelse shadowed buffer");
 712                 __brelse(bh);
 713         }
 714
 715         J_ASSERT (commit_transaction->t_shadow_list == NULL);
 716
 717         jbd_debug(3, "JBD: commit phase 5\n");
 718
 719         /* Here we wait for the revoke record and descriptor record buffers */
 720  wait_for_ctlbuf:
 721         while (commit_transaction->t_log_list != NULL) {
 722                 struct buffer_head *bh;
 723
 724                 jh = commit_transaction->t_log_list->b_tprev;
 725                 bh = jh2bh(jh);
 726                 if (buffer_locked(bh)) {
 727                         wait_on_buffer(bh);
 728                         goto wait_for_ctlbuf;
 729                 }
 730                 if (cond_resched())
 731                         goto wait_for_ctlbuf;
 732
 733                 if (unlikely(!buffer_uptodate(bh)))
 734                         err = -EIO;
 735
 736                 BUFFER_TRACE(bh, "ph5: control buffer writeout done: unfile");
 737                 clear_buffer_jwrite(bh);
 738                 journal_unfile_buffer(journal, jh);
 739                 journal_put_journal_head(jh);
 740                 __brelse(bh);           /* One for getblk */
 741                 /* AKPM: bforget here */
 742         }
 743
 744         jbd_debug(3, "JBD: commit phase 6\n");
 745
 746         if (journal_write_commit_record(journal, commit_transaction))
 747                 err = -EIO;
 748
 749         if (err)
 750                 journal_abort(journal, err);
 751
 752         /* End of a transaction!  Finally, we can do checkpoint
 753            processing: any buffers committed as a result of this
 754            transaction can be removed from any checkpoint list it was on
 755            before. */
 756
 757         jbd_debug(3, "JBD: commit phase 7\n");
 758
 759         J_ASSERT(commit_transaction->t_sync_datalist == NULL);
 760         J_ASSERT(commit_transaction->t_buffers == NULL);
 761         J_ASSERT(commit_transaction->t_checkpoint_list == NULL);
 762         J_ASSERT(commit_transaction->t_iobuf_list == NULL);
 763         J_ASSERT(commit_transaction->t_shadow_list == NULL);
 764         J_ASSERT(commit_transaction->t_log_list == NULL);
 765
 766 restart_loop:
 767         /*
 768          * As there are other places (journal_unmap_buffer()) adding buffers
 769          * to this list we have to be careful and hold the j_list_lock.
 770          */
 771         spin_lock(&journal->j_list_lock);
 772         while (commit_transaction->t_forget) {
 773                 transaction_t *cp_transaction;
 774                 struct buffer_head *bh;
 775
 776                 jh = commit_transaction->t_forget;
 777                 spin_unlock(&journal->j_list_lock);
 778                 bh = jh2bh(jh);
 779                 jbd_lock_bh_state(bh);
 780                 J_ASSERT_JH(jh, jh->b_transaction == commit_transaction ||
 781                         jh->b_transaction == journal->j_running_transaction);
 782
 783                 /*
 784                  * If there is undo-protected committed data against
 785                  * this buffer, then we can remove it now.  If it is a
 786                  * buffer needing such protection, the old frozen_data
 787                  * field now points to a committed version of the
 788                  * buffer, so rotate that field to the new committed
 789                  * data.
 790                  *
 791                  * Otherwise, we can just throw away the frozen data now.
 792                  */
 793                 if (jh->b_committed_data) {
 794                         jbd_free(jh->b_committed_data, bh->b_size);
 795                         jh->b_committed_data = NULL;
 796                         if (jh->b_frozen_data) {
 797                                 jh->b_committed_data = jh->b_frozen_data;
 798                                 jh->b_frozen_data = NULL;
 799                         }
 800                 } else if (jh->b_frozen_data) {
 801                         jbd_free(jh->b_frozen_data, bh->b_size);
 802                         jh->b_frozen_data = NULL;
 803                 }
 804
 805                 spin_lock(&journal->j_list_lock);
 806                 cp_transaction = jh->b_cp_transaction;
 807                 if (cp_transaction) {
 808                         JBUFFER_TRACE(jh, "remove from old cp transaction");
 809                         __journal_remove_checkpoint(jh);
 810                 }
 811
 812                 /* Only re-checkpoint the buffer_head if it is marked
 813                  * dirty.  If the buffer was added to the BJ_Forget list
 814                  * by journal_forget, it may no longer be dirty and
 815                  * there's no point in keeping a checkpoint record for
 816                  * it. */
 817
 818                 /* A buffer which has been freed while still being
 819                  * journaled by a previous transaction may end up still
 820                  * being dirty here, but we want to avoid writing back
 821                  * that buffer in the future now that the last use has
 822                  * been committed.  That's not only a performance gain,
 823                  * it also stops aliasing problems if the buffer is left
 824                  * behind for writeback and gets reallocated for another
 825                  * use in a different page. */
 826                 if (buffer_freed(bh)) {
 827                         clear_buffer_freed(bh);
 828                         clear_buffer_jbddirty(bh);
 829                 }
 830
 831                 if (buffer_jbddirty(bh)) {
 832                         JBUFFER_TRACE(jh, "add to new checkpointing trans");
 833                         __journal_insert_checkpoint(jh, commit_transaction);
 834                         JBUFFER_TRACE(jh, "refile for checkpoint writeback");
 835                         __journal_refile_buffer(jh);
 836                         jbd_unlock_bh_state(bh);
 837                 } else {
 838                         J_ASSERT_BH(bh, !buffer_dirty(bh));
 839                         /* The buffer on BJ_Forget list and not jbddirty means
 840                          * it has been freed by this transaction and hence it
 841                          * could not have been reallocated until this
 842                          * transaction has committed. *BUT* it could be
 843                          * reallocated once we have written all the data to
 844                          * disk and before we process the buffer on BJ_Forget
 845                          * list. */
 846                         JBUFFER_TRACE(jh, "refile or unfile freed buffer");
 847                         __journal_refile_buffer(jh);
 848                         if (!jh->b_transaction) {
 849                                 jbd_unlock_bh_state(bh);
 850                                  /* needs a brelse */
 851                                 journal_remove_journal_head(bh);
 852                                 release_buffer_page(bh);
 853                         } else
 854                                 jbd_unlock_bh_state(bh);
 855                 }
 856                 cond_resched_lock(&journal->j_list_lock);
 857         }
 858         spin_unlock(&journal->j_list_lock);
 859         /*
 860          * This is a bit sleazy.  We use j_list_lock to protect transition
 861          * of a transaction into T_FINISHED state and calling
 862          * __journal_drop_transaction(). Otherwise we could race with
 863          * other checkpointing code processing the transaction...
 864          */
 865         spin_lock(&journal->j_state_lock);
 866         spin_lock(&journal->j_list_lock);
 867         /*
 868          * Now recheck if some buffers did not get attached to the transaction
 869          * while the lock was dropped...
 870          */
 871         if (commit_transaction->t_forget) {
 872                 spin_unlock(&journal->j_list_lock);
 873                 spin_unlock(&journal->j_state_lock);
 874                 goto restart_loop;
 875         }
 876
 877         /* Done with this transaction! */
 878
 879         jbd_debug(3, "JBD: commit phase 8\n");
 880
 881         J_ASSERT(commit_transaction->t_state == T_COMMIT);
 882
 883         commit_transaction->t_state = T_FINISHED;
 884         J_ASSERT(commit_transaction == journal->j_committing_transaction);
 885         journal->j_commit_sequence = commit_transaction->t_tid;
 886         journal->j_committing_transaction = NULL;
 887         spin_unlock(&journal->j_state_lock);
 888
 889         if (commit_transaction->t_checkpoint_list == NULL &&
 890             commit_transaction->t_checkpoint_io_list == NULL) {
 891                 __journal_drop_transaction(journal, commit_transaction);
 892         } else {
 893                 if (journal->j_checkpoint_transactions == NULL) {
 894                         journal->j_checkpoint_transactions = commit_transaction;
 895                         commit_transaction->t_cpnext = commit_transaction;
 896                         commit_transaction->t_cpprev = commit_transaction;
 897                 } else {
 898                         commit_transaction->t_cpnext =
 899                                 journal->j_checkpoint_transactions;
 900                         commit_transaction->t_cpprev =
 901                                 commit_transaction->t_cpnext->t_cpprev;
 902                         commit_transaction->t_cpnext->t_cpprev =
 903                                 commit_transaction;
 904                         commit_transaction->t_cpprev->t_cpnext =
 905                                 commit_transaction;
 906                 }
 907         }
 908         spin_unlock(&journal->j_list_lock);
 909
 910         jbd_debug(1, "JBD: commit %d complete, head %d\n",
 911                   journal->j_commit_sequence, journal->j_tail_sequence);
 912
 913         wake_up(&journal->j_wait_done_commit);
 914 }