fs/jbd2/commit.c

   1 /*
   2  * linux/fs/jbd2/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/jbd2.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 = jbd2_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(JBD2_MAGIC_NUMBER);
 124                 tmp->h_blocktype = cpu_to_be32(JBD2_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 & JBD2_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 &= ~JBD2_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         jbd2_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                         __jbd2_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                 }
 252                 else {
 253                         BUFFER_TRACE(bh, "writeout complete: unfile");
 254                         __jbd2_journal_unfile_buffer(jh);
 255                         jbd_unlock_bh_state(bh);
 256                         if (locked)
 257                                 unlock_buffer(bh);
 258                         jbd2_journal_remove_journal_head(bh);
 259                         /* Once for our safety reference, once for
 260                          * jbd2_journal_remove_journal_head() */
 261                         put_bh(bh);
 262                         put_bh(bh);
 263                 }
 264
 265                 if (lock_need_resched(&journal->j_list_lock)) {
 266                         spin_unlock(&journal->j_list_lock);
 267                         goto write_out_data;
 268                 }
 269         }
 270         spin_unlock(&journal->j_list_lock);
 271         journal_do_submit_data(wbuf, bufs);
 272 }
 273
 274 static inline void write_tag_block(int tag_bytes, journal_block_tag_t *tag,
 275                                    unsigned long long block)
 276 {
 277         tag->t_blocknr = cpu_to_be32(block & (u32)~0);
 278         if (tag_bytes > JBD_TAG_SIZE32)
 279                 tag->t_blocknr_high = cpu_to_be32((block >> 31) >> 1);
 280 }
 281
 282 /*
 283  * jbd2_journal_commit_transaction
 284  *
 285  * The primary function for committing a transaction to the log.  This
 286  * function is called by the journal thread to begin a complete commit.
 287  */
 288 void jbd2_journal_commit_transaction(journal_t *journal)
 289 {
 290         transaction_t *commit_transaction;
 291         struct journal_head *jh, *new_jh, *descriptor;
 292         struct buffer_head **wbuf = journal->j_wbuf;
 293         int bufs;
 294         int flags;
 295         int err;
 296         unsigned long long blocknr;
 297         char *tagp = NULL;
 298         journal_header_t *header;
 299         journal_block_tag_t *tag = NULL;
 300         int space_left = 0;
 301         int first_tag = 0;
 302         int tag_flag;
 303         int i;
 304         int tag_bytes = journal_tag_bytes(journal);
 305
 306         /*
 307          * First job: lock down the current transaction and wait for
 308          * all outstanding updates to complete.
 309          */
 310
 311 #ifdef COMMIT_STATS
 312         spin_lock(&journal->j_list_lock);
 313         summarise_journal_usage(journal);
 314         spin_unlock(&journal->j_list_lock);
 315 #endif
 316
 317         /* Do we need to erase the effects of a prior jbd2_journal_flush? */
 318         if (journal->j_flags & JBD2_FLUSHED) {
 319                 jbd_debug(3, "super block updated\n");
 320                 jbd2_journal_update_superblock(journal, 1);
 321         } else {
 322                 jbd_debug(3, "superblock not updated\n");
 323         }
 324
 325         J_ASSERT(journal->j_running_transaction != NULL);
 326         J_ASSERT(journal->j_committing_transaction == NULL);
 327
 328         commit_transaction = journal->j_running_transaction;
 329         J_ASSERT(commit_transaction->t_state == T_RUNNING);
 330
 331         jbd_debug(1, "JBD: starting commit of transaction %d\n",
 332                         commit_transaction->t_tid);
 333
 334         spin_lock(&journal->j_state_lock);
 335         commit_transaction->t_state = T_LOCKED;
 336
 337         spin_lock(&commit_transaction->t_handle_lock);
 338         while (commit_transaction->t_updates) {
 339                 DEFINE_WAIT(wait);
 340
 341                 prepare_to_wait(&journal->j_wait_updates, &wait,
 342                                         TASK_UNINTERRUPTIBLE);
 343                 if (commit_transaction->t_updates) {
 344                         spin_unlock(&commit_transaction->t_handle_lock);
 345                         spin_unlock(&journal->j_state_lock);
 346                         schedule();
 347                         spin_lock(&journal->j_state_lock);
 348                         spin_lock(&commit_transaction->t_handle_lock);
 349                 }
 350                 finish_wait(&journal->j_wait_updates, &wait);
 351         }
 352         spin_unlock(&commit_transaction->t_handle_lock);
 353
 354         J_ASSERT (commit_transaction->t_outstanding_credits <=
 355                         journal->j_max_transaction_buffers);
 356
 357         /*
 358          * First thing we are allowed to do is to discard any remaining
 359          * BJ_Reserved buffers.  Note, it is _not_ permissible to assume
 360          * that there are no such buffers: if a large filesystem
 361          * operation like a truncate needs to split itself over multiple
 362          * transactions, then it may try to do a jbd2_journal_restart() while
 363          * there are still BJ_Reserved buffers outstanding.  These must
 364          * be released cleanly from the current transaction.
 365          *
 366          * In this case, the filesystem must still reserve write access
 367          * again before modifying the buffer in the new transaction, but
 368          * we do not require it to remember exactly which old buffers it
 369          * has reserved.  This is consistent with the existing behaviour
 370          * that multiple jbd2_journal_get_write_access() calls to the same
 371          * buffer are perfectly permissable.
 372          */
 373         while (commit_transaction->t_reserved_list) {
 374                 jh = commit_transaction->t_reserved_list;
 375                 JBUFFER_TRACE(jh, "reserved, unused: refile");
 376                 /*
 377                  * A jbd2_journal_get_undo_access()+jbd2_journal_release_buffer() may
 378                  * leave undo-committed data.
 379                  */
 380                 if (jh->b_committed_data) {
 381                         struct buffer_head *bh = jh2bh(jh);
 382
 383                         jbd_lock_bh_state(bh);
 384                         jbd2_slab_free(jh->b_committed_data, bh->b_size);
 385                         jh->b_committed_data = NULL;
 386                         jbd_unlock_bh_state(bh);
 387                 }
 388                 jbd2_journal_refile_buffer(journal, jh);
 389         }
 390
 391         /*
 392          * Now try to drop any written-back buffers from the journal's
 393          * checkpoint lists.  We do this *before* commit because it potentially
 394          * frees some memory
 395          */
 396         spin_lock(&journal->j_list_lock);
 397         __jbd2_journal_clean_checkpoint_list(journal);
 398         spin_unlock(&journal->j_list_lock);
 399
 400         jbd_debug (3, "JBD: commit phase 1\n");
 401
 402         /*
 403          * Switch to a new revoke table.
 404          */
 405         jbd2_journal_switch_revoke_table(journal);
 406
 407         commit_transaction->t_state = T_FLUSH;
 408         journal->j_committing_transaction = commit_transaction;
 409         journal->j_running_transaction = NULL;
 410         commit_transaction->t_log_start = journal->j_head;
 411         wake_up(&journal->j_wait_transaction_locked);
 412         spin_unlock(&journal->j_state_lock);
 413
 414         jbd_debug (3, "JBD: commit phase 2\n");
 415
 416         /*
 417          * First, drop modified flag: all accesses to the buffers
 418          * will be tracked for a new trasaction only -bzzz
 419          */
 420         spin_lock(&journal->j_list_lock);
 421         if (commit_transaction->t_buffers) {
 422                 new_jh = jh = commit_transaction->t_buffers->b_tnext;
 423                 do {
 424                         J_ASSERT_JH(new_jh, new_jh->b_modified == 1 ||
 425                                         new_jh->b_modified == 0);
 426                         new_jh->b_modified = 0;
 427                         new_jh = new_jh->b_tnext;
 428                 } while (new_jh != jh);
 429         }
 430         spin_unlock(&journal->j_list_lock);
 431
 432         /*
 433          * Now start flushing things to disk, in the order they appear
 434          * on the transaction lists.  Data blocks go first.
 435          */
 436         err = 0;
 437         journal_submit_data_buffers(journal, commit_transaction);
 438
 439         /*
 440          * Wait for all previously submitted IO to complete.
 441          */
 442         spin_lock(&journal->j_list_lock);
 443         while (commit_transaction->t_locked_list) {
 444                 struct buffer_head *bh;
 445
 446                 jh = commit_transaction->t_locked_list->b_tprev;
 447                 bh = jh2bh(jh);
 448                 get_bh(bh);
 449                 if (buffer_locked(bh)) {
 450                         spin_unlock(&journal->j_list_lock);
 451                         wait_on_buffer(bh);
 452                         if (unlikely(!buffer_uptodate(bh)))
 453                                 err = -EIO;
 454                         spin_lock(&journal->j_list_lock);
 455                 }
 456                 if (!inverted_lock(journal, bh)) {
 457                         put_bh(bh);
 458                         spin_lock(&journal->j_list_lock);
 459                         continue;
 460                 }
 461                 if (buffer_jbd(bh) && jh->b_jlist == BJ_Locked) {
 462                         __jbd2_journal_unfile_buffer(jh);
 463                         jbd_unlock_bh_state(bh);
 464                         jbd2_journal_remove_journal_head(bh);
 465                         put_bh(bh);
 466                 } else {
 467                         jbd_unlock_bh_state(bh);
 468                 }
 469                 put_bh(bh);
 470                 cond_resched_lock(&journal->j_list_lock);
 471         }
 472         spin_unlock(&journal->j_list_lock);
 473
 474         if (err)
 475                 __jbd2_journal_abort_hard(journal);
 476
 477         jbd2_journal_write_revoke_records(journal, commit_transaction);
 478
 479         jbd_debug(3, "JBD: commit phase 2\n");
 480
 481         /*
 482          * If we found any dirty or locked buffers, then we should have
 483          * looped back up to the write_out_data label.  If there weren't
 484          * any then journal_clean_data_list should have wiped the list
 485          * clean by now, so check that it is in fact empty.
 486          */
 487         J_ASSERT (commit_transaction->t_sync_datalist == NULL);
 488
 489         jbd_debug (3, "JBD: commit phase 3\n");
 490
 491         /*
 492          * Way to go: we have now written out all of the data for a
 493          * transaction!  Now comes the tricky part: we need to write out
 494          * metadata.  Loop over the transaction's entire buffer list:
 495          */
 496         commit_transaction->t_state = T_COMMIT;
 497
 498         descriptor = NULL;
 499         bufs = 0;
 500         while (commit_transaction->t_buffers) {
 501
 502                 /* Find the next buffer to be journaled... */
 503
 504                 jh = commit_transaction->t_buffers;
 505
 506                 /* If we're in abort mode, we just un-journal the buffer and
 507                    release it for background writing. */
 508
 509                 if (is_journal_aborted(journal)) {
 510                         JBUFFER_TRACE(jh, "journal is aborting: refile");
 511                         jbd2_journal_refile_buffer(journal, jh);
 512                         /* If that was the last one, we need to clean up
 513                          * any descriptor buffers which may have been
 514                          * already allocated, even if we are now
 515                          * aborting. */
 516                         if (!commit_transaction->t_buffers)
 517                                 goto start_journal_io;
 518                         continue;
 519                 }
 520
 521                 /* Make sure we have a descriptor block in which to
 522                    record the metadata buffer. */
 523
 524                 if (!descriptor) {
 525                         struct buffer_head *bh;
 526
 527                         J_ASSERT (bufs == 0);
 528
 529                         jbd_debug(4, "JBD: get descriptor\n");
 530
 531                         descriptor = jbd2_journal_get_descriptor_buffer(journal);
 532                         if (!descriptor) {
 533                                 __jbd2_journal_abort_hard(journal);
 534                                 continue;
 535                         }
 536
 537                         bh = jh2bh(descriptor);
 538                         jbd_debug(4, "JBD: got buffer %llu (%p)\n",
 539                                 (unsigned long long)bh->b_blocknr, bh->b_data);
 540                         header = (journal_header_t *)&bh->b_data[0];
 541                         header->h_magic     = cpu_to_be32(JBD2_MAGIC_NUMBER);
 542                         header->h_blocktype = cpu_to_be32(JBD2_DESCRIPTOR_BLOCK);
 543                         header->h_sequence  = cpu_to_be32(commit_transaction->t_tid);
 544
 545                         tagp = &bh->b_data[sizeof(journal_header_t)];
 546                         space_left = bh->b_size - sizeof(journal_header_t);
 547                         first_tag = 1;
 548                         set_buffer_jwrite(bh);
 549                         set_buffer_dirty(bh);
 550                         wbuf[bufs++] = bh;
 551
 552                         /* Record it so that we can wait for IO
 553                            completion later */
 554                         BUFFER_TRACE(bh, "ph3: file as descriptor");
 555                         jbd2_journal_file_buffer(descriptor, commit_transaction,
 556                                         BJ_LogCtl);
 557                 }
 558
 559                 /* Where is the buffer to be written? */
 560
 561                 err = jbd2_journal_next_log_block(journal, &blocknr);
 562                 /* If the block mapping failed, just abandon the buffer
 563                    and repeat this loop: we'll fall into the
 564                    refile-on-abort condition above. */
 565                 if (err) {
 566                         __jbd2_journal_abort_hard(journal);
 567                         continue;
 568                 }
 569
 570                 /*
 571                  * start_this_handle() uses t_outstanding_credits to determine
 572                  * the free space in the log, but this counter is changed
 573                  * by jbd2_journal_next_log_block() also.
 574                  */
 575                 commit_transaction->t_outstanding_credits--;
 576
 577                 /* Bump b_count to prevent truncate from stumbling over
 578                    the shadowed buffer!  @@@ This can go if we ever get
 579                    rid of the BJ_IO/BJ_Shadow pairing of buffers. */
 580                 atomic_inc(&jh2bh(jh)->b_count);
 581
 582                 /* Make a temporary IO buffer with which to write it out
 583                    (this will requeue both the metadata buffer and the
 584                    temporary IO buffer). new_bh goes on BJ_IO*/
 585
 586                 set_bit(BH_JWrite, &jh2bh(jh)->b_state);
 587                 /*
 588                  * akpm: jbd2_journal_write_metadata_buffer() sets
 589                  * new_bh->b_transaction to commit_transaction.
 590                  * We need to clean this up before we release new_bh
 591                  * (which is of type BJ_IO)
 592                  */
 593                 JBUFFER_TRACE(jh, "ph3: write metadata");
 594                 flags = jbd2_journal_write_metadata_buffer(commit_transaction,
 595                                                       jh, &new_jh, blocknr);
 596                 set_bit(BH_JWrite, &jh2bh(new_jh)->b_state);
 597                 wbuf[bufs++] = jh2bh(new_jh);
 598
 599                 /* Record the new block's tag in the current descriptor
 600                    buffer */
 601
 602                 tag_flag = 0;
 603                 if (flags & 1)
 604                         tag_flag |= JBD2_FLAG_ESCAPE;
 605                 if (!first_tag)
 606                         tag_flag |= JBD2_FLAG_SAME_UUID;
 607
 608                 tag = (journal_block_tag_t *) tagp;
 609                 write_tag_block(tag_bytes, tag, jh2bh(jh)->b_blocknr);
 610                 tag->t_flags = cpu_to_be32(tag_flag);
 611                 tagp += tag_bytes;
 612                 space_left -= tag_bytes;
 613
 614                 if (first_tag) {
 615                         memcpy (tagp, journal->j_uuid, 16);
 616                         tagp += 16;
 617                         space_left -= 16;
 618                         first_tag = 0;
 619                 }
 620
 621                 /* If there's no more to do, or if the descriptor is full,
 622                    let the IO rip! */
 623
 624                 if (bufs == journal->j_wbufsize ||
 625                     commit_transaction->t_buffers == NULL ||
 626                     space_left < tag_bytes + 16) {
 627
 628                         jbd_debug(4, "JBD: Submit %d IOs\n", bufs);
 629
 630                         /* Write an end-of-descriptor marker before
 631                            submitting the IOs.  "tag" still points to
 632                            the last tag we set up. */
 633
 634                         tag->t_flags |= cpu_to_be32(JBD2_FLAG_LAST_TAG);
 635
 636 start_journal_io:
 637                         for (i = 0; i < bufs; i++) {
 638                                 struct buffer_head *bh = wbuf[i];
 639                                 lock_buffer(bh);
 640                                 clear_buffer_dirty(bh);
 641                                 set_buffer_uptodate(bh);
 642                                 bh->b_end_io = journal_end_buffer_io_sync;
 643                                 submit_bh(WRITE, bh);
 644                         }
 645                         cond_resched();
 646
 647                         /* Force a new descriptor to be generated next
 648                            time round the loop. */
 649                         descriptor = NULL;
 650                         bufs = 0;
 651                 }
 652         }
 653
 654         /* Lo and behold: we have just managed to send a transaction to
 655            the log.  Before we can commit it, wait for the IO so far to
 656            complete.  Control buffers being written are on the
 657            transaction's t_log_list queue, and metadata buffers are on
 658            the t_iobuf_list queue.
 659
 660            Wait for the buffers in reverse order.  That way we are
 661            less likely to be woken up until all IOs have completed, and
 662            so we incur less scheduling load.
 663         */
 664
 665         jbd_debug(3, "JBD: commit phase 4\n");
 666
 667         /*
 668          * akpm: these are BJ_IO, and j_list_lock is not needed.
 669          * See __journal_try_to_free_buffer.
 670          */
 671 wait_for_iobuf:
 672         while (commit_transaction->t_iobuf_list != NULL) {
 673                 struct buffer_head *bh;
 674
 675                 jh = commit_transaction->t_iobuf_list->b_tprev;
 676                 bh = jh2bh(jh);
 677                 if (buffer_locked(bh)) {
 678                         wait_on_buffer(bh);
 679                         goto wait_for_iobuf;
 680                 }
 681                 if (cond_resched())
 682                         goto wait_for_iobuf;
 683
 684                 if (unlikely(!buffer_uptodate(bh)))
 685                         err = -EIO;
 686
 687                 clear_buffer_jwrite(bh);
 688
 689                 JBUFFER_TRACE(jh, "ph4: unfile after journal write");
 690                 jbd2_journal_unfile_buffer(journal, jh);
 691
 692                 /*
 693                  * ->t_iobuf_list should contain only dummy buffer_heads
 694                  * which were created by jbd2_journal_write_metadata_buffer().
 695                  */
 696                 BUFFER_TRACE(bh, "dumping temporary bh");
 697                 jbd2_journal_put_journal_head(jh);
 698                 __brelse(bh);
 699                 J_ASSERT_BH(bh, atomic_read(&bh->b_count) == 0);
 700                 free_buffer_head(bh);
 701
 702                 /* We also have to unlock and free the corresponding
 703                    shadowed buffer */
 704                 jh = commit_transaction->t_shadow_list->b_tprev;
 705                 bh = jh2bh(jh);
 706                 clear_bit(BH_JWrite, &bh->b_state);
 707                 J_ASSERT_BH(bh, buffer_jbddirty(bh));
 708
 709                 /* The metadata is now released for reuse, but we need
 710                    to remember it against this transaction so that when
 711                    we finally commit, we can do any checkpointing
 712                    required. */
 713                 JBUFFER_TRACE(jh, "file as BJ_Forget");
 714                 jbd2_journal_file_buffer(jh, commit_transaction, BJ_Forget);
 715                 /* Wake up any transactions which were waiting for this
 716                    IO to complete */
 717                 wake_up_bit(&bh->b_state, BH_Unshadow);
 718                 JBUFFER_TRACE(jh, "brelse shadowed buffer");
 719                 __brelse(bh);
 720         }
 721
 722         J_ASSERT (commit_transaction->t_shadow_list == NULL);
 723
 724         jbd_debug(3, "JBD: commit phase 5\n");
 725
 726         /* Here we wait for the revoke record and descriptor record buffers */
 727  wait_for_ctlbuf:
 728         while (commit_transaction->t_log_list != NULL) {
 729                 struct buffer_head *bh;
 730
 731                 jh = commit_transaction->t_log_list->b_tprev;
 732                 bh = jh2bh(jh);
 733                 if (buffer_locked(bh)) {
 734                         wait_on_buffer(bh);
 735                         goto wait_for_ctlbuf;
 736                 }
 737                 if (cond_resched())
 738                         goto wait_for_ctlbuf;
 739
 740                 if (unlikely(!buffer_uptodate(bh)))
 741                         err = -EIO;
 742
 743                 BUFFER_TRACE(bh, "ph5: control buffer writeout done: unfile");
 744                 clear_buffer_jwrite(bh);
 745                 jbd2_journal_unfile_buffer(journal, jh);
 746                 jbd2_journal_put_journal_head(jh);
 747                 __brelse(bh);           /* One for getblk */
 748                 /* AKPM: bforget here */
 749         }
 750
 751         jbd_debug(3, "JBD: commit phase 6\n");
 752
 753         if (journal_write_commit_record(journal, commit_transaction))
 754                 err = -EIO;
 755
 756         if (err)
 757                 __jbd2_journal_abort_hard(journal);
 758
 759         /* End of a transaction!  Finally, we can do checkpoint
 760            processing: any buffers committed as a result of this
 761            transaction can be removed from any checkpoint list it was on
 762            before. */
 763
 764         jbd_debug(3, "JBD: commit phase 7\n");
 765
 766         J_ASSERT(commit_transaction->t_sync_datalist == NULL);
 767         J_ASSERT(commit_transaction->t_buffers == NULL);
 768         J_ASSERT(commit_transaction->t_checkpoint_list == NULL);
 769         J_ASSERT(commit_transaction->t_iobuf_list == NULL);
 770         J_ASSERT(commit_transaction->t_shadow_list == NULL);
 771         J_ASSERT(commit_transaction->t_log_list == NULL);
 772
 773 restart_loop:
 774         /*
 775          * As there are other places (journal_unmap_buffer()) adding buffers
 776          * to this list we have to be careful and hold the j_list_lock.
 777          */
 778         spin_lock(&journal->j_list_lock);
 779         while (commit_transaction->t_forget) {
 780                 transaction_t *cp_transaction;
 781                 struct buffer_head *bh;
 782
 783                 jh = commit_transaction->t_forget;
 784                 spin_unlock(&journal->j_list_lock);
 785                 bh = jh2bh(jh);
 786                 jbd_lock_bh_state(bh);
 787                 J_ASSERT_JH(jh, jh->b_transaction == commit_transaction ||
 788                         jh->b_transaction == journal->j_running_transaction);
 789
 790                 /*
 791                  * If there is undo-protected committed data against
 792                  * this buffer, then we can remove it now.  If it is a
 793                  * buffer needing such protection, the old frozen_data
 794                  * field now points to a committed version of the
 795                  * buffer, so rotate that field to the new committed
 796                  * data.
 797                  *
 798                  * Otherwise, we can just throw away the frozen data now.
 799                  */
 800                 if (jh->b_committed_data) {
 801                         jbd2_slab_free(jh->b_committed_data, bh->b_size);
 802                         jh->b_committed_data = NULL;
 803                         if (jh->b_frozen_data) {
 804                                 jh->b_committed_data = jh->b_frozen_data;
 805                                 jh->b_frozen_data = NULL;
 806                         }
 807                 } else if (jh->b_frozen_data) {
 808                         jbd2_slab_free(jh->b_frozen_data, bh->b_size);
 809                         jh->b_frozen_data = NULL;
 810                 }
 811
 812                 spin_lock(&journal->j_list_lock);
 813                 cp_transaction = jh->b_cp_transaction;
 814                 if (cp_transaction) {
 815                         JBUFFER_TRACE(jh, "remove from old cp transaction");
 816                         __jbd2_journal_remove_checkpoint(jh);
 817                 }
 818
 819                 /* Only re-checkpoint the buffer_head if it is marked
 820                  * dirty.  If the buffer was added to the BJ_Forget list
 821                  * by jbd2_journal_forget, it may no longer be dirty and
 822                  * there's no point in keeping a checkpoint record for
 823                  * it. */
 824
 825                 /* A buffer which has been freed while still being
 826                  * journaled by a previous transaction may end up still
 827                  * being dirty here, but we want to avoid writing back
 828                  * that buffer in the future now that the last use has
 829                  * been committed.  That's not only a performance gain,
 830                  * it also stops aliasing problems if the buffer is left
 831                  * behind for writeback and gets reallocated for another
 832                  * use in a different page. */
 833                 if (buffer_freed(bh)) {
 834                         clear_buffer_freed(bh);
 835                         clear_buffer_jbddirty(bh);
 836                 }
 837
 838                 if (buffer_jbddirty(bh)) {
 839                         JBUFFER_TRACE(jh, "add to new checkpointing trans");
 840                         __jbd2_journal_insert_checkpoint(jh, commit_transaction);
 841                         JBUFFER_TRACE(jh, "refile for checkpoint writeback");
 842                         __jbd2_journal_refile_buffer(jh);
 843                         jbd_unlock_bh_state(bh);
 844                 } else {
 845                         J_ASSERT_BH(bh, !buffer_dirty(bh));
 846                         /* The buffer on BJ_Forget list and not jbddirty means
 847                          * it has been freed by this transaction and hence it
 848                          * could not have been reallocated until this
 849                          * transaction has committed. *BUT* it could be
 850                          * reallocated once we have written all the data to
 851                          * disk and before we process the buffer on BJ_Forget
 852                          * list. */
 853                         JBUFFER_TRACE(jh, "refile or unfile freed buffer");
 854                         __jbd2_journal_refile_buffer(jh);
 855                         if (!jh->b_transaction) {
 856                                 jbd_unlock_bh_state(bh);
 857                                  /* needs a brelse */
 858                                 jbd2_journal_remove_journal_head(bh);
 859                                 release_buffer_page(bh);
 860                         } else
 861                                 jbd_unlock_bh_state(bh);
 862                 }
 863                 cond_resched_lock(&journal->j_list_lock);
 864         }
 865         spin_unlock(&journal->j_list_lock);
 866         /*
 867          * This is a bit sleazy.  We borrow j_list_lock to protect
 868          * journal->j_committing_transaction in __jbd2_journal_remove_checkpoint.
 869          * Really, __jbd2_journal_remove_checkpoint should be using j_state_lock but
 870          * it's a bit hassle to hold that across __jbd2_journal_remove_checkpoint
 871          */
 872         spin_lock(&journal->j_state_lock);
 873         spin_lock(&journal->j_list_lock);
 874         /*
 875          * Now recheck if some buffers did not get attached to the transaction
 876          * while the lock was dropped...
 877          */
 878         if (commit_transaction->t_forget) {
 879                 spin_unlock(&journal->j_list_lock);
 880                 spin_unlock(&journal->j_state_lock);
 881                 goto restart_loop;
 882         }
 883
 884         /* Done with this transaction! */
 885
 886         jbd_debug(3, "JBD: commit phase 8\n");
 887
 888         J_ASSERT(commit_transaction->t_state == T_COMMIT);
 889
 890         commit_transaction->t_state = T_FINISHED;
 891         J_ASSERT(commit_transaction == journal->j_committing_transaction);
 892         journal->j_commit_sequence = commit_transaction->t_tid;
 893         journal->j_committing_transaction = NULL;
 894         spin_unlock(&journal->j_state_lock);
 895
 896         if (commit_transaction->t_checkpoint_list == NULL) {
 897                 __jbd2_journal_drop_transaction(journal, commit_transaction);
 898         } else {
 899                 if (journal->j_checkpoint_transactions == NULL) {
 900                         journal->j_checkpoint_transactions = commit_transaction;
 901                         commit_transaction->t_cpnext = commit_transaction;
 902                         commit_transaction->t_cpprev = commit_transaction;
 903                 } else {
 904                         commit_transaction->t_cpnext =
 905                                 journal->j_checkpoint_transactions;
 906                         commit_transaction->t_cpprev =
 907                                 commit_transaction->t_cpnext->t_cpprev;
 908                         commit_transaction->t_cpnext->t_cpprev =
 909                                 commit_transaction;
 910                         commit_transaction->t_cpprev->t_cpnext =
 911                                 commit_transaction;
 912                 }
 913         }
 914         spin_unlock(&journal->j_list_lock);
 915
 916         jbd_debug(1, "JBD: commit %d complete, head %d\n",
 917                   journal->j_commit_sequence, journal->j_tail_sequence);
 918
 919         wake_up(&journal->j_wait_done_commit);
 920 }