fs/jbd/commit.c

   1 /*
   2  * linux/fs/commit.c
   3  *
   4  * Written by Stephen C. Tweedie <sct@redhat.com>, 1998
   5  *
   6  * Copyright 1998 Red Hat corp --- All Rights Reserved
   7  *
   8  * This file is part of the Linux kernel and is made available under
   9  * the terms of the GNU General Public License, version 2, or at your
  10  * option, any later version, incorporated herein by reference.
  11  *
  12  * Journal commit routines for the generic filesystem journaling code;
  13  * part of the ext2fs journaling system.
  14  */
  15
  16 #include <linux/time.h>
  17 #include <linux/fs.h>
  18 #include <linux/jbd.h>
  19 #include <linux/errno.h>
  20 #include <linux/slab.h>
  21 #include <linux/mm.h>
  22 #include <linux/pagemap.h>
  23 #include <linux/smp_lock.h>
  24
  25 /*
  26  * Default IO end handler for temporary BJ_IO buffer_heads.
  27  */
  28 static void journal_end_buffer_io_sync(struct buffer_head *bh, int uptodate)
  29 {
  30         BUFFER_TRACE(bh, "");
  31         if (uptodate)
  32                 set_buffer_uptodate(bh);
  33         else
  34                 clear_buffer_uptodate(bh);
  35         unlock_buffer(bh);
  36 }
  37
  38 /*
  39  * When an ext3-ordered file is truncated, it is possible that many pages are
  40  * not sucessfully freed, because they are attached to a committing transaction.
  41  * After the transaction commits, these pages are left on the LRU, with no
  42  * ->mapping, and with attached buffers.  These pages are trivially reclaimable
  43  * by the VM, but their apparent absence upsets the VM accounting, and it makes
  44  * the numbers in /proc/meminfo look odd.
  45  *
  46  * So here, we have a buffer which has just come off the forget list.  Look to
  47  * see if we can strip all buffers from the backing page.
  48  *
  49  * Called under lock_journal(), and possibly under journal_datalist_lock.  The
  50  * caller provided us with a ref against the buffer, and we drop that here.
  51  */
  52 static void release_buffer_page(struct buffer_head *bh)
  53 {
  54         struct page *page;
  55
  56         if (buffer_dirty(bh))
  57                 goto nope;
  58         if (atomic_read(&bh->b_count) != 1)
  59                 goto nope;
  60         page = bh->b_page;
  61         if (!page)
  62                 goto nope;
  63         if (page->mapping)
  64                 goto nope;
  65
  66         /* OK, it's a truncated page */
  67         if (TestSetPageLocked(page))
  68                 goto nope;
  69
  70         page_cache_get(page);
  71         __brelse(bh);
  72         try_to_free_buffers(page);
  73         unlock_page(page);
  74         page_cache_release(page);
  75         return;
  76
  77 nope:
  78         __brelse(bh);
  79 }
  80
  81 /*
  82  * Try to acquire jbd_lock_bh_state() against the buffer, when j_list_lock is
  83  * held.  For ranking reasons we must trylock.  If we lose, schedule away and
  84  * return 0.  j_list_lock is dropped in this case.
  85  */
  86 static int inverted_lock(journal_t *journal, struct buffer_head *bh)
  87 {
  88         if (!jbd_trylock_bh_state(bh)) {
  89                 spin_unlock(&journal->j_list_lock);
  90                 schedule();
  91                 return 0;
  92         }
  93         return 1;
  94 }
  95
  96 /*
  97  * journal_commit_transaction
  98  *
  99  * The primary function for committing a transaction to the log.  This
 100  * function is called by the journal thread to begin a complete commit.
 101  */
 102 void journal_commit_transaction(journal_t *journal)
 103 {
 104         transaction_t *commit_transaction;
 105         struct journal_head *jh, *new_jh, *descriptor;
 106         struct buffer_head *wbuf[64];
 107         int bufs;
 108         int flags;
 109         int err;
 110         unsigned long blocknr;
 111         char *tagp = NULL;
 112         journal_header_t *header;
 113         journal_block_tag_t *tag = NULL;
 114         int space_left = 0;
 115         int first_tag = 0;
 116         int tag_flag;
 117         int i;
 118
 119         /*
 120          * First job: lock down the current transaction and wait for
 121          * all outstanding updates to complete.
 122          */
 123
 124 #ifdef COMMIT_STATS
 125         spin_lock(&journal->j_list_lock);
 126         summarise_journal_usage(journal);
 127         spin_unlock(&journal->j_list_lock);
 128 #endif
 129
 130         /* Do we need to erase the effects of a prior journal_flush? */
 131         if (journal->j_flags & JFS_FLUSHED) {
 132                 jbd_debug(3, "super block updated\n");
 133                 journal_update_superblock(journal, 1);
 134         } else {
 135                 jbd_debug(3, "superblock not updated\n");
 136         }
 137
 138         J_ASSERT(journal->j_running_transaction != NULL);
 139         J_ASSERT(journal->j_committing_transaction == NULL);
 140
 141         commit_transaction = journal->j_running_transaction;
 142         J_ASSERT(commit_transaction->t_state == T_RUNNING);
 143
 144         jbd_debug(1, "JBD: starting commit of transaction %d\n",
 145                         commit_transaction->t_tid);
 146
 147         spin_lock(&journal->j_state_lock);
 148         commit_transaction->t_state = T_LOCKED;
 149
 150         spin_lock(&commit_transaction->t_handle_lock);
 151         while (commit_transaction->t_updates) {
 152                 DEFINE_WAIT(wait);
 153
 154                 prepare_to_wait(&journal->j_wait_updates, &wait,
 155                                         TASK_UNINTERRUPTIBLE);
 156                 if (commit_transaction->t_updates) {
 157                         spin_unlock(&commit_transaction->t_handle_lock);
 158                         spin_unlock(&journal->j_state_lock);
 159                         schedule();
 160                         spin_lock(&journal->j_state_lock);
 161                         spin_lock(&commit_transaction->t_handle_lock);
 162                 }
 163                 finish_wait(&journal->j_wait_updates, &wait);
 164         }
 165         spin_unlock(&commit_transaction->t_handle_lock);
 166
 167         J_ASSERT (commit_transaction->t_outstanding_credits <=
 168                         journal->j_max_transaction_buffers);
 169
 170         /*
 171          * First thing we are allowed to do is to discard any remaining
 172          * BJ_Reserved buffers.  Note, it is _not_ permissible to assume
 173          * that there are no such buffers: if a large filesystem
 174          * operation like a truncate needs to split itself over multiple
 175          * transactions, then it may try to do a journal_restart() while
 176          * there are still BJ_Reserved buffers outstanding.  These must
 177          * be released cleanly from the current transaction.
 178          *
 179          * In this case, the filesystem must still reserve write access
 180          * again before modifying the buffer in the new transaction, but
 181          * we do not require it to remember exactly which old buffers it
 182          * has reserved.  This is consistent with the existing behaviour
 183          * that multiple journal_get_write_access() calls to the same
 184          * buffer are perfectly permissable.
 185          */
 186         while (commit_transaction->t_reserved_list) {
 187                 jh = commit_transaction->t_reserved_list;
 188                 JBUFFER_TRACE(jh, "reserved, unused: refile");
 189                 /*
 190                  * A journal_get_undo_access()+journal_release_buffer() may
 191                  * leave undo-committed data.
 192                  */
 193                 if (jh->b_committed_data) {
 194                         struct buffer_head *bh = jh2bh(jh);
 195
 196                         jbd_lock_bh_state(bh);
 197                         if (jh->b_committed_data) {
 198                                 kfree(jh->b_committed_data);
 199                                 jh->b_committed_data = NULL;
 200                         }
 201                         jbd_unlock_bh_state(bh);
 202                 }
 203                 journal_refile_buffer(journal, jh);
 204         }
 205
 206         /*
 207          * Now try to drop any written-back buffers from the journal's
 208          * checkpoint lists.  We do this *before* commit because it potentially
 209          * frees some memory
 210          */
 211         spin_lock(&journal->j_list_lock);
 212         __journal_clean_checkpoint_list(journal);
 213         spin_unlock(&journal->j_list_lock);
 214
 215         jbd_debug (3, "JBD: commit phase 1\n");
 216
 217         /*
 218          * Switch to a new revoke table.
 219          */
 220         journal_switch_revoke_table(journal);
 221
 222         commit_transaction->t_state = T_FLUSH;
 223         journal->j_committing_transaction = commit_transaction;
 224         journal->j_running_transaction = NULL;
 225         commit_transaction->t_log_start = journal->j_head;
 226         wake_up(&journal->j_wait_transaction_locked);
 227         spin_unlock(&journal->j_state_lock);
 228
 229         jbd_debug (3, "JBD: commit phase 2\n");
 230
 231         /*
 232          * Now start flushing things to disk, in the order they appear
 233          * on the transaction lists.  Data blocks go first.
 234          */
 235
 236         err = 0;
 237         /*
 238          * Whenever we unlock the journal and sleep, things can get added
 239          * onto ->t_sync_datalist, so we have to keep looping back to
 240          * write_out_data until we *know* that the list is empty.
 241          */
 242         bufs = 0;
 243         /*
 244          * Cleanup any flushed data buffers from the data list.  Even in
 245          * abort mode, we want to flush this out as soon as possible.
 246          */
 247 write_out_data:
 248         cond_resched();
 249         spin_lock(&journal->j_list_lock);
 250
 251         while (commit_transaction->t_sync_datalist) {
 252                 struct buffer_head *bh;
 253
 254                 jh = commit_transaction->t_sync_datalist;
 255                 commit_transaction->t_sync_datalist = jh->b_tnext;
 256                 bh = jh2bh(jh);
 257                 if (buffer_locked(bh)) {
 258                         BUFFER_TRACE(bh, "locked");
 259                         if (!inverted_lock(journal, bh))
 260                                 goto write_out_data;
 261                         __journal_unfile_buffer(jh);
 262                         __journal_file_buffer(jh, commit_transaction,
 263                                                 BJ_Locked);
 264                         jbd_unlock_bh_state(bh);
 265                         if (need_resched()) {
 266                                 spin_unlock(&journal->j_list_lock);
 267                                 goto write_out_data;
 268                         }
 269                 } else {
 270                         if (buffer_dirty(bh)) {
 271                                 BUFFER_TRACE(bh, "start journal writeout");
 272                                 get_bh(bh);
 273                                 wbuf[bufs++] = bh;
 274                                 if (bufs == ARRAY_SIZE(wbuf)) {
 275                                         jbd_debug(2, "submit %d writes\n",
 276                                                         bufs);
 277                                         spin_unlock(&journal->j_list_lock);
 278                                         ll_rw_block(WRITE, bufs, wbuf);
 279                                         journal_brelse_array(wbuf, bufs);
 280                                         bufs = 0;
 281                                         goto write_out_data;
 282                                 }
 283                         } else {
 284                                 BUFFER_TRACE(bh, "writeout complete: unfile");
 285                                 if (!inverted_lock(journal, bh))
 286                                         goto write_out_data;
 287                                 __journal_unfile_buffer(jh);
 288                                 jbd_unlock_bh_state(bh);
 289                                 journal_remove_journal_head(bh);
 290                                 put_bh(bh);
 291                                 if (need_resched()) {
 292                                         spin_unlock(&journal->j_list_lock);
 293                                         goto write_out_data;
 294                                 }
 295                         }
 296                 }
 297         }
 298
 299         if (bufs) {
 300                 spin_unlock(&journal->j_list_lock);
 301                 ll_rw_block(WRITE, bufs, wbuf);
 302                 journal_brelse_array(wbuf, bufs);
 303                 spin_lock(&journal->j_list_lock);
 304         }
 305
 306         /*
 307          * Wait for all previously submitted IO to complete.
 308          */
 309         while (commit_transaction->t_locked_list) {
 310                 struct buffer_head *bh;
 311
 312                 jh = commit_transaction->t_locked_list->b_tprev;
 313                 bh = jh2bh(jh);
 314                 get_bh(bh);
 315                 if (buffer_locked(bh)) {
 316                         spin_unlock(&journal->j_list_lock);
 317                         wait_on_buffer(bh);
 318                         if (unlikely(!buffer_uptodate(bh)))
 319                                 err = -EIO;
 320                         spin_lock(&journal->j_list_lock);
 321                 }
 322                 if (!inverted_lock(journal, bh)) {
 323                         put_bh(bh);
 324                         spin_lock(&journal->j_list_lock);
 325                         continue;
 326                 }
 327                 if (buffer_jbd(bh) && jh->b_jlist == BJ_Locked) {
 328                         __journal_unfile_buffer(jh);
 329                         jbd_unlock_bh_state(bh);
 330                         journal_remove_journal_head(bh);
 331                         put_bh(bh);
 332                 } else {
 333                         jbd_unlock_bh_state(bh);
 334                 }
 335                 put_bh(bh);
 336                 if (need_resched()) {
 337                         spin_unlock(&journal->j_list_lock);
 338                         cond_resched();
 339                         spin_lock(&journal->j_list_lock);
 340                 }
 341         }
 342         spin_unlock(&journal->j_list_lock);
 343
 344         journal_write_revoke_records(journal, commit_transaction);
 345
 346         jbd_debug(3, "JBD: commit phase 2\n");
 347
 348         /*
 349          * If we found any dirty or locked buffers, then we should have
 350          * looped back up to the write_out_data label.  If there weren't
 351          * any then journal_clean_data_list should have wiped the list
 352          * clean by now, so check that it is in fact empty.
 353          */
 354         J_ASSERT (commit_transaction->t_sync_datalist == NULL);
 355
 356         jbd_debug (3, "JBD: commit phase 3\n");
 357
 358         /*
 359          * Way to go: we have now written out all of the data for a
 360          * transaction!  Now comes the tricky part: we need to write out
 361          * metadata.  Loop over the transaction's entire buffer list:
 362          */
 363         commit_transaction->t_state = T_COMMIT;
 364
 365         descriptor = NULL;
 366         bufs = 0;
 367         while (commit_transaction->t_buffers) {
 368
 369                 /* Find the next buffer to be journaled... */
 370
 371                 jh = commit_transaction->t_buffers;
 372
 373                 /* If we're in abort mode, we just un-journal the buffer and
 374                    release it for background writing. */
 375
 376                 if (is_journal_aborted(journal)) {
 377                         JBUFFER_TRACE(jh, "journal is aborting: refile");
 378                         journal_refile_buffer(journal, jh);
 379                         /* If that was the last one, we need to clean up
 380                          * any descriptor buffers which may have been
 381                          * already allocated, even if we are now
 382                          * aborting. */
 383                         if (!commit_transaction->t_buffers)
 384                                 goto start_journal_io;
 385                         continue;
 386                 }
 387
 388                 /* Make sure we have a descriptor block in which to
 389                    record the metadata buffer. */
 390
 391                 if (!descriptor) {
 392                         struct buffer_head *bh;
 393
 394                         J_ASSERT (bufs == 0);
 395
 396                         jbd_debug(4, "JBD: get descriptor\n");
 397
 398                         descriptor = journal_get_descriptor_buffer(journal);
 399                         if (!descriptor) {
 400                                 __journal_abort_hard(journal);
 401                                 continue;
 402                         }
 403
 404                         bh = jh2bh(descriptor);
 405                         jbd_debug(4, "JBD: got buffer %llu (%p)\n",
 406                                 (unsigned long long)bh->b_blocknr, bh->b_data);
 407                         header = (journal_header_t *)&bh->b_data[0];
 408                         header->h_magic     = cpu_to_be32(JFS_MAGIC_NUMBER);
 409                         header->h_blocktype = cpu_to_be32(JFS_DESCRIPTOR_BLOCK);
 410                         header->h_sequence  = cpu_to_be32(commit_transaction->t_tid);
 411
 412                         tagp = &bh->b_data[sizeof(journal_header_t)];
 413                         space_left = bh->b_size - sizeof(journal_header_t);
 414                         first_tag = 1;
 415                         set_buffer_jwrite(bh);
 416                         set_buffer_dirty(bh);
 417                         wbuf[bufs++] = bh;
 418
 419                         /* Record it so that we can wait for IO
 420                            completion later */
 421                         BUFFER_TRACE(bh, "ph3: file as descriptor");
 422                         journal_file_buffer(descriptor, commit_transaction,
 423                                         BJ_LogCtl);
 424                 }
 425
 426                 /* Where is the buffer to be written? */
 427
 428                 err = journal_next_log_block(journal, &blocknr);
 429                 /* If the block mapping failed, just abandon the buffer
 430                    and repeat this loop: we'll fall into the
 431                    refile-on-abort condition above. */
 432                 if (err) {
 433                         __journal_abort_hard(journal);
 434                         continue;
 435                 }
 436
 437                 /*
 438                  * start_this_handle() uses t_outstanding_credits to determine
 439                  * the free space in the log, but this counter is changed
 440                  * by journal_next_log_block() also.
 441                  */
 442                 commit_transaction->t_outstanding_credits--;
 443
 444                 /* Bump b_count to prevent truncate from stumbling over
 445                    the shadowed buffer!  @@@ This can go if we ever get
 446                    rid of the BJ_IO/BJ_Shadow pairing of buffers. */
 447                 atomic_inc(&jh2bh(jh)->b_count);
 448
 449                 /* Make a temporary IO buffer with which to write it out
 450                    (this will requeue both the metadata buffer and the
 451                    temporary IO buffer). new_bh goes on BJ_IO*/
 452
 453                 set_bit(BH_JWrite, &jh2bh(jh)->b_state);
 454                 /*
 455                  * akpm: journal_write_metadata_buffer() sets
 456                  * new_bh->b_transaction to commit_transaction.
 457                  * We need to clean this up before we release new_bh
 458                  * (which is of type BJ_IO)
 459                  */
 460                 JBUFFER_TRACE(jh, "ph3: write metadata");
 461                 flags = journal_write_metadata_buffer(commit_transaction,
 462                                                       jh, &new_jh, blocknr);
 463                 set_bit(BH_JWrite, &jh2bh(new_jh)->b_state);
 464                 wbuf[bufs++] = jh2bh(new_jh);
 465
 466                 /* Record the new block's tag in the current descriptor
 467                    buffer */
 468
 469                 tag_flag = 0;
 470                 if (flags & 1)
 471                         tag_flag |= JFS_FLAG_ESCAPE;
 472                 if (!first_tag)
 473                         tag_flag |= JFS_FLAG_SAME_UUID;
 474
 475                 tag = (journal_block_tag_t *) tagp;
 476                 tag->t_blocknr = cpu_to_be32(jh2bh(jh)->b_blocknr);
 477                 tag->t_flags = cpu_to_be32(tag_flag);
 478                 tagp += sizeof(journal_block_tag_t);
 479                 space_left -= sizeof(journal_block_tag_t);
 480
 481                 if (first_tag) {
 482                         memcpy (tagp, journal->j_uuid, 16);
 483                         tagp += 16;
 484                         space_left -= 16;
 485                         first_tag = 0;
 486                 }
 487
 488                 /* If there's no more to do, or if the descriptor is full,
 489                    let the IO rip! */
 490
 491                 if (bufs == ARRAY_SIZE(wbuf) ||
 492                     commit_transaction->t_buffers == NULL ||
 493                     space_left < sizeof(journal_block_tag_t) + 16) {
 494
 495                         jbd_debug(4, "JBD: Submit %d IOs\n", bufs);
 496
 497                         /* Write an end-of-descriptor marker before
 498                            submitting the IOs.  "tag" still points to
 499                            the last tag we set up. */
 500
 501                         tag->t_flags |= cpu_to_be32(JFS_FLAG_LAST_TAG);
 502
 503 start_journal_io:
 504                         for (i = 0; i < bufs; i++) {
 505                                 struct buffer_head *bh = wbuf[i];
 506                                 lock_buffer(bh);
 507                                 clear_buffer_dirty(bh);
 508                                 set_buffer_uptodate(bh);
 509                                 bh->b_end_io = journal_end_buffer_io_sync;
 510                                 submit_bh(WRITE, bh);
 511                         }
 512                         cond_resched();
 513
 514                         /* Force a new descriptor to be generated next
 515                            time round the loop. */
 516                         descriptor = NULL;
 517                         bufs = 0;
 518                 }
 519         }
 520
 521         /* Lo and behold: we have just managed to send a transaction to
 522            the log.  Before we can commit it, wait for the IO so far to
 523            complete.  Control buffers being written are on the
 524            transaction's t_log_list queue, and metadata buffers are on
 525            the t_iobuf_list queue.
 526
 527            Wait for the buffers in reverse order.  That way we are
 528            less likely to be woken up until all IOs have completed, and
 529            so we incur less scheduling load.
 530         */
 531
 532         jbd_debug(3, "JBD: commit phase 4\n");
 533
 534         /*
 535          * akpm: these are BJ_IO, and j_list_lock is not needed.
 536          * See __journal_try_to_free_buffer.
 537          */
 538 wait_for_iobuf:
 539         while (commit_transaction->t_iobuf_list != NULL) {
 540                 struct buffer_head *bh;
 541
 542                 jh = commit_transaction->t_iobuf_list->b_tprev;
 543                 bh = jh2bh(jh);
 544                 if (buffer_locked(bh)) {
 545                         wait_on_buffer(bh);
 546                         goto wait_for_iobuf;
 547                 }
 548
 549                 if (unlikely(!buffer_uptodate(bh)))
 550                         err = -EIO;
 551
 552                 clear_buffer_jwrite(bh);
 553
 554                 JBUFFER_TRACE(jh, "ph4: unfile after journal write");
 555                 journal_unfile_buffer(journal, jh);
 556
 557                 /*
 558                  * ->t_iobuf_list should contain only dummy buffer_heads
 559                  * which were created by journal_write_metadata_buffer().
 560                  */
 561                 BUFFER_TRACE(bh, "dumping temporary bh");
 562                 journal_put_journal_head(jh);
 563                 __brelse(bh);
 564                 J_ASSERT_BH(bh, atomic_read(&bh->b_count) == 0);
 565                 free_buffer_head(bh);
 566
 567                 /* We also have to unlock and free the corresponding
 568                    shadowed buffer */
 569                 jh = commit_transaction->t_shadow_list->b_tprev;
 570                 bh = jh2bh(jh);
 571                 clear_bit(BH_JWrite, &bh->b_state);
 572                 J_ASSERT_BH(bh, buffer_jbddirty(bh));
 573
 574                 /* The metadata is now released for reuse, but we need
 575                    to remember it against this transaction so that when
 576                    we finally commit, we can do any checkpointing
 577                    required. */
 578                 JBUFFER_TRACE(jh, "file as BJ_Forget");
 579                 journal_file_buffer(jh, commit_transaction, BJ_Forget);
 580                 /* Wake up any transactions which were waiting for this
 581                    IO to complete */
 582                 wake_up_buffer(bh);
 583                 JBUFFER_TRACE(jh, "brelse shadowed buffer");
 584                 __brelse(bh);
 585         }
 586
 587         J_ASSERT (commit_transaction->t_shadow_list == NULL);
 588
 589         jbd_debug(3, "JBD: commit phase 5\n");
 590
 591         /* Here we wait for the revoke record and descriptor record buffers */
 592  wait_for_ctlbuf:
 593         while (commit_transaction->t_log_list != NULL) {
 594                 struct buffer_head *bh;
 595
 596                 jh = commit_transaction->t_log_list->b_tprev;
 597                 bh = jh2bh(jh);
 598                 if (buffer_locked(bh)) {
 599                         wait_on_buffer(bh);
 600                         goto wait_for_ctlbuf;
 601                 }
 602
 603                 if (unlikely(!buffer_uptodate(bh)))
 604                         err = -EIO;
 605
 606                 BUFFER_TRACE(bh, "ph5: control buffer writeout done: unfile");
 607                 clear_buffer_jwrite(bh);
 608                 journal_unfile_buffer(journal, jh);
 609                 journal_put_journal_head(jh);
 610                 __brelse(bh);           /* One for getblk */
 611                 /* AKPM: bforget here */
 612         }
 613
 614         jbd_debug(3, "JBD: commit phase 6\n");
 615
 616         if (is_journal_aborted(journal))
 617                 goto skip_commit;
 618
 619         /* Done it all: now write the commit record.  We should have
 620          * cleaned up our previous buffers by now, so if we are in abort
 621          * mode we can now just skip the rest of the journal write
 622          * entirely. */
 623
 624         descriptor = journal_get_descriptor_buffer(journal);
 625         if (!descriptor) {
 626                 __journal_abort_hard(journal);
 627                 goto skip_commit;
 628         }
 629
 630         /* AKPM: buglet - add `i' to tmp! */
 631         for (i = 0; i < jh2bh(descriptor)->b_size; i += 512) {
 632                 journal_header_t *tmp =
 633                         (journal_header_t*)jh2bh(descriptor)->b_data;
 634                 tmp->h_magic = cpu_to_be32(JFS_MAGIC_NUMBER);
 635                 tmp->h_blocktype = cpu_to_be32(JFS_COMMIT_BLOCK);
 636                 tmp->h_sequence = cpu_to_be32(commit_transaction->t_tid);
 637         }
 638
 639         JBUFFER_TRACE(descriptor, "write commit block");
 640         {
 641                 struct buffer_head *bh = jh2bh(descriptor);
 642                 int ret;
 643                 int barrier_done = 0;
 644
 645                 set_buffer_dirty(bh);
 646                 if (journal->j_flags & JFS_BARRIER) {
 647                         set_buffer_ordered(bh);
 648                         barrier_done = 1;
 649                 }
 650                 ret = sync_dirty_buffer(bh);
 651                 /* is it possible for another commit to fail at roughly
 652                  * the same time as this one?  If so, we don't want to
 653                  * trust the barrier flag in the super, but instead want
 654                  * to remember if we sent a barrier request
 655                  */
 656                 if (ret == -EOPNOTSUPP && barrier_done) {
 657                         char b[BDEVNAME_SIZE];
 658
 659                         printk(KERN_WARNING
 660                                 "JBD: barrier-based sync failed on %s - "
 661                                 "disabling barriers\n",
 662                                 bdevname(journal->j_dev, b));
 663                         spin_lock(&journal->j_state_lock);
 664                         journal->j_flags &= ~JFS_BARRIER;
 665                         spin_unlock(&journal->j_state_lock);
 666
 667                         /* And try again, without the barrier */
 668                         clear_buffer_ordered(bh);
 669                         set_buffer_uptodate(bh);
 670                         set_buffer_dirty(bh);
 671                         ret = sync_dirty_buffer(bh);
 672                 }
 673                 if (unlikely(ret == -EIO))
 674                         err = -EIO;
 675                 put_bh(bh);             /* One for getblk() */
 676                 journal_put_journal_head(descriptor);
 677         }
 678
 679         /* End of a transaction!  Finally, we can do checkpoint
 680            processing: any buffers committed as a result of this
 681            transaction can be removed from any checkpoint list it was on
 682            before. */
 683
 684 skip_commit: /* The journal should be unlocked by now. */
 685
 686         if (err)
 687                 __journal_abort_hard(journal);
 688
 689         /*
 690          * Call any callbacks that had been registered for handles in this
 691          * transaction.  It is up to the callback to free any allocated
 692          * memory.
 693          *
 694          * The spinlocking (t_jcb_lock) here is surely unnecessary...
 695          */
 696         spin_lock(&commit_transaction->t_jcb_lock);
 697         if (!list_empty(&commit_transaction->t_jcb)) {
 698                 struct list_head *p, *n;
 699                 int error = is_journal_aborted(journal);
 700
 701                 list_for_each_safe(p, n, &commit_transaction->t_jcb) {
 702                         struct journal_callback *jcb;
 703
 704                         jcb = list_entry(p, struct journal_callback, jcb_list);
 705                         list_del(p);
 706                         spin_unlock(&commit_transaction->t_jcb_lock);
 707                         jcb->jcb_func(jcb, error);
 708                         spin_lock(&commit_transaction->t_jcb_lock);
 709                 }
 710         }
 711         spin_unlock(&commit_transaction->t_jcb_lock);
 712
 713         jbd_debug(3, "JBD: commit phase 7\n");
 714
 715         J_ASSERT(commit_transaction->t_sync_datalist == NULL);
 716         J_ASSERT(commit_transaction->t_buffers == NULL);
 717         J_ASSERT(commit_transaction->t_checkpoint_list == NULL);
 718         J_ASSERT(commit_transaction->t_iobuf_list == NULL);
 719         J_ASSERT(commit_transaction->t_shadow_list == NULL);
 720         J_ASSERT(commit_transaction->t_log_list == NULL);
 721
 722         while (commit_transaction->t_forget) {
 723                 transaction_t *cp_transaction;
 724                 struct buffer_head *bh;
 725
 726                 jh = commit_transaction->t_forget;
 727                 bh = jh2bh(jh);
 728                 jbd_lock_bh_state(bh);
 729                 J_ASSERT_JH(jh, jh->b_transaction == commit_transaction ||
 730                         jh->b_transaction == journal->j_running_transaction);
 731
 732                 /*
 733                  * If there is undo-protected committed data against
 734                  * this buffer, then we can remove it now.  If it is a
 735                  * buffer needing such protection, the old frozen_data
 736                  * field now points to a committed version of the
 737                  * buffer, so rotate that field to the new committed
 738                  * data.
 739                  *
 740                  * Otherwise, we can just throw away the frozen data now.
 741                  */
 742                 if (jh->b_committed_data) {
 743                         kfree(jh->b_committed_data);
 744                         jh->b_committed_data = NULL;
 745                         if (jh->b_frozen_data) {
 746                                 jh->b_committed_data = jh->b_frozen_data;
 747                                 jh->b_frozen_data = NULL;
 748                         }
 749                 } else if (jh->b_frozen_data) {
 750                         kfree(jh->b_frozen_data);
 751                         jh->b_frozen_data = NULL;
 752                 }
 753
 754                 spin_lock(&journal->j_list_lock);
 755                 cp_transaction = jh->b_cp_transaction;
 756                 if (cp_transaction) {
 757                         JBUFFER_TRACE(jh, "remove from old cp transaction");
 758                         __journal_remove_checkpoint(jh);
 759                 }
 760
 761                 /* Only re-checkpoint the buffer_head if it is marked
 762                  * dirty.  If the buffer was added to the BJ_Forget list
 763                  * by journal_forget, it may no longer be dirty and
 764                  * there's no point in keeping a checkpoint record for
 765                  * it. */
 766
 767                 /* A buffer which has been freed while still being
 768                  * journaled by a previous transaction may end up still
 769                  * being dirty here, but we want to avoid writing back
 770                  * that buffer in the future now that the last use has
 771                  * been committed.  That's not only a performance gain,
 772                  * it also stops aliasing problems if the buffer is left
 773                  * behind for writeback and gets reallocated for another
 774                  * use in a different page. */
 775                 if (buffer_freed(bh)) {
 776                         clear_buffer_freed(bh);
 777                         clear_buffer_jbddirty(bh);
 778                 }
 779
 780                 if (buffer_jbddirty(bh)) {
 781                         JBUFFER_TRACE(jh, "add to new checkpointing trans");
 782                         __journal_insert_checkpoint(jh, commit_transaction);
 783                         JBUFFER_TRACE(jh, "refile for checkpoint writeback");
 784                         __journal_refile_buffer(jh);
 785                         jbd_unlock_bh_state(bh);
 786                 } else {
 787                         J_ASSERT_BH(bh, !buffer_dirty(bh));
 788                         J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
 789                         __journal_unfile_buffer(jh);
 790                         jbd_unlock_bh_state(bh);
 791                         journal_remove_journal_head(bh);  /* needs a brelse */
 792                         release_buffer_page(bh);
 793                 }
 794                 spin_unlock(&journal->j_list_lock);
 795         }
 796
 797         /* Done with this transaction! */
 798
 799         jbd_debug(3, "JBD: commit phase 8\n");
 800
 801         J_ASSERT(commit_transaction->t_state == T_COMMIT);
 802
 803         /*
 804          * This is a bit sleazy.  We borrow j_list_lock to protect
 805          * journal->j_committing_transaction in __journal_remove_checkpoint.
 806          * Really, __jornal_remove_checkpoint should be using j_state_lock but
 807          * it's a bit hassle to hold that across __journal_remove_checkpoint
 808          */
 809         spin_lock(&journal->j_state_lock);
 810         spin_lock(&journal->j_list_lock);
 811         commit_transaction->t_state = T_FINISHED;
 812         J_ASSERT(commit_transaction == journal->j_committing_transaction);
 813         journal->j_commit_sequence = commit_transaction->t_tid;
 814         journal->j_committing_transaction = NULL;
 815         spin_unlock(&journal->j_state_lock);
 816
 817         if (commit_transaction->t_checkpoint_list == NULL) {
 818                 __journal_drop_transaction(journal, commit_transaction);
 819         } else {
 820                 if (journal->j_checkpoint_transactions == NULL) {
 821                         journal->j_checkpoint_transactions = commit_transaction;
 822                         commit_transaction->t_cpnext = commit_transaction;
 823                         commit_transaction->t_cpprev = commit_transaction;
 824                 } else {
 825                         commit_transaction->t_cpnext =
 826                                 journal->j_checkpoint_transactions;
 827                         commit_transaction->t_cpprev =
 828                                 commit_transaction->t_cpnext->t_cpprev;
 829                         commit_transaction->t_cpnext->t_cpprev =
 830                                 commit_transaction;
 831                         commit_transaction->t_cpprev->t_cpnext =
 832                                 commit_transaction;
 833                 }
 834         }
 835         spin_unlock(&journal->j_list_lock);
 836
 837         jbd_debug(1, "JBD: commit %d complete, head %d\n",
 838                   journal->j_commit_sequence, journal->j_tail_sequence);
 839
 840         wake_up(&journal->j_wait_done_commit);
 841 }