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/mm.h>
  21 #include <linux/pagemap.h>
  22 #include <linux/bio.h>
  23 #include <linux/blkdev.h>
  24 #include <trace/events/jbd.h>
  25
  26 /*
  27  * Default IO end handler for temporary BJ_IO buffer_heads.
  28  */
  29 static void journal_end_buffer_io_sync(struct buffer_head *bh, int uptodate)
  30 {
  31         BUFFER_TRACE(bh, "");
  32         if (uptodate)
  33                 set_buffer_uptodate(bh);
  34         else
  35                 clear_buffer_uptodate(bh);
  36         unlock_buffer(bh);
  37 }
  38
  39 /*
  40  * When an ext3-ordered file is truncated, it is possible that many pages are
  41  * not successfully freed, because they are attached to a committing transaction.
  42  * After the transaction commits, these pages are left on the LRU, with no
  43  * ->mapping, and with attached buffers.  These pages are trivially reclaimable
  44  * by the VM, but their apparent absence upsets the VM accounting, and it makes
  45  * the numbers in /proc/meminfo look odd.
  46  *
  47  * So here, we have a buffer which has just come off the forget list.  Look to
  48  * see if we can strip all buffers from the backing page.
  49  *
  50  * Called under journal->j_list_lock.  The caller provided us with a ref
  51  * against the buffer, and we drop that here.
  52  */
  53 static void release_buffer_page(struct buffer_head *bh)
  54 {
  55         struct page *page;
  56
  57         if (buffer_dirty(bh))
  58                 goto nope;
  59         if (atomic_read(&bh->b_count) != 1)
  60                 goto nope;
  61         page = bh->b_page;
  62         if (!page)
  63                 goto nope;
  64         if (page->mapping)
  65                 goto nope;
  66
  67         /* OK, it's a truncated page */
  68         if (!trylock_page(page))
  69                 goto nope;
  70
  71         page_cache_get(page);
  72         __brelse(bh);
  73         try_to_free_buffers(page);
  74         unlock_page(page);
  75         page_cache_release(page);
  76         return;
  77
  78 nope:
  79         __brelse(bh);
  80 }
  81
  82 /*
  83  * Decrement reference counter for data buffer. If it has been marked
  84  * 'BH_Freed', release it and the page to which it belongs if possible.
  85  */
  86 static void release_data_buffer(struct buffer_head *bh)
  87 {
  88         if (buffer_freed(bh)) {
  89                 clear_buffer_freed(bh);
  90                 release_buffer_page(bh);
  91         } else
  92                 put_bh(bh);
  93 }
  94
  95 /*
  96  * Try to acquire jbd_lock_bh_state() against the buffer, when j_list_lock is
  97  * held.  For ranking reasons we must trylock.  If we lose, schedule away and
  98  * return 0.  j_list_lock is dropped in this case.
  99  */
 100 static int inverted_lock(journal_t *journal, struct buffer_head *bh)
 101 {
 102         if (!jbd_trylock_bh_state(bh)) {
 103                 spin_unlock(&journal->j_list_lock);
 104                 schedule();
 105                 return 0;
 106         }
 107         return 1;
 108 }
 109
 110 /* Done it all: now write the commit record.  We should have
 111  * cleaned up our previous buffers by now, so if we are in abort
 112  * mode we can now just skip the rest of the journal write
 113  * entirely.
 114  *
 115  * Returns 1 if the journal needs to be aborted or 0 on success
 116  */
 117 static int journal_write_commit_record(journal_t *journal,
 118                                         transaction_t *commit_transaction)
 119 {
 120         struct journal_head *descriptor;
 121         struct buffer_head *bh;
 122         journal_header_t *header;
 123         int ret;
 124
 125         if (is_journal_aborted(journal))
 126                 return 0;
 127
 128         descriptor = journal_get_descriptor_buffer(journal);
 129         if (!descriptor)
 130                 return 1;
 131
 132         bh = jh2bh(descriptor);
 133
 134         header = (journal_header_t *)(bh->b_data);
 135         header->h_magic = cpu_to_be32(JFS_MAGIC_NUMBER);
 136         header->h_blocktype = cpu_to_be32(JFS_COMMIT_BLOCK);
 137         header->h_sequence = cpu_to_be32(commit_transaction->t_tid);
 138
 139         JBUFFER_TRACE(descriptor, "write commit block");
 140         set_buffer_dirty(bh);
 141
 142         if (journal->j_flags & JFS_BARRIER)
 143                 ret = __sync_dirty_buffer(bh, WRITE_SYNC | WRITE_FLUSH_FUA);
 144         else
 145                 ret = sync_dirty_buffer(bh);
 146
 147         put_bh(bh);             /* One for getblk() */
 148         journal_put_journal_head(descriptor);
 149
 150         return (ret == -EIO);
 151 }
 152
 153 static void journal_do_submit_data(struct buffer_head **wbuf, int bufs,
 154                                    int write_op)
 155 {
 156         int i;
 157
 158         for (i = 0; i < bufs; i++) {
 159                 wbuf[i]->b_end_io = end_buffer_write_sync;
 160                 /* We use-up our safety reference in submit_bh() */
 161                 submit_bh(write_op, wbuf[i]);
 162         }
 163 }
 164
 165 /*
 166  *  Submit all the data buffers to disk
 167  */
 168 static int journal_submit_data_buffers(journal_t *journal,
 169                                        transaction_t *commit_transaction,
 170                                        int write_op)
 171 {
 172         struct journal_head *jh;
 173         struct buffer_head *bh;
 174         int locked;
 175         int bufs = 0;
 176         struct buffer_head **wbuf = journal->j_wbuf;
 177         int err = 0;
 178
 179         /*
 180          * Whenever we unlock the journal and sleep, things can get added
 181          * onto ->t_sync_datalist, so we have to keep looping back to
 182          * write_out_data until we *know* that the list is empty.
 183          *
 184          * Cleanup any flushed data buffers from the data list.  Even in
 185          * abort mode, we want to flush this out as soon as possible.
 186          */
 187 write_out_data:
 188         cond_resched();
 189         spin_lock(&journal->j_list_lock);
 190
 191         while (commit_transaction->t_sync_datalist) {
 192                 jh = commit_transaction->t_sync_datalist;
 193                 bh = jh2bh(jh);
 194                 locked = 0;
 195
 196                 /* Get reference just to make sure buffer does not disappear
 197                  * when we are forced to drop various locks */
 198                 get_bh(bh);
 199                 /* If the buffer is dirty, we need to submit IO and hence
 200                  * we need the buffer lock. We try to lock the buffer without
 201                  * blocking. If we fail, we need to drop j_list_lock and do
 202                  * blocking lock_buffer().
 203                  */
 204                 if (buffer_dirty(bh)) {
 205                         if (!trylock_buffer(bh)) {
 206                                 BUFFER_TRACE(bh, "needs blocking lock");
 207                                 spin_unlock(&journal->j_list_lock);
 208                                 trace_jbd_do_submit_data(journal,
 209                                                      commit_transaction);
 210                                 /* Write out all data to prevent deadlocks */
 211                                 journal_do_submit_data(wbuf, bufs, write_op);
 212                                 bufs = 0;
 213                                 lock_buffer(bh);
 214                                 spin_lock(&journal->j_list_lock);
 215                         }
 216                         locked = 1;
 217                 }
 218                 /* We have to get bh_state lock. Again out of order, sigh. */
 219                 if (!inverted_lock(journal, bh)) {
 220                         jbd_lock_bh_state(bh);
 221                         spin_lock(&journal->j_list_lock);
 222                 }
 223                 /* Someone already cleaned up the buffer? */
 224                 if (!buffer_jbd(bh) || bh2jh(bh) != jh
 225                         || jh->b_transaction != commit_transaction
 226                         || jh->b_jlist != BJ_SyncData) {
 227                         jbd_unlock_bh_state(bh);
 228                         if (locked)
 229                                 unlock_buffer(bh);
 230                         BUFFER_TRACE(bh, "already cleaned up");
 231                         release_data_buffer(bh);
 232                         continue;
 233                 }
 234                 if (locked && test_clear_buffer_dirty(bh)) {
 235                         BUFFER_TRACE(bh, "needs writeout, adding to array");
 236                         wbuf[bufs++] = bh;
 237                         __journal_file_buffer(jh, commit_transaction,
 238                                                 BJ_Locked);
 239                         jbd_unlock_bh_state(bh);
 240                         if (bufs == journal->j_wbufsize) {
 241                                 spin_unlock(&journal->j_list_lock);
 242                                 trace_jbd_do_submit_data(journal,
 243                                                      commit_transaction);
 244                                 journal_do_submit_data(wbuf, bufs, write_op);
 245                                 bufs = 0;
 246                                 goto write_out_data;
 247                         }
 248                 } else if (!locked && buffer_locked(bh)) {
 249                         __journal_file_buffer(jh, commit_transaction,
 250                                                 BJ_Locked);
 251                         jbd_unlock_bh_state(bh);
 252                         put_bh(bh);
 253                 } else {
 254                         BUFFER_TRACE(bh, "writeout complete: unfile");
 255                         if (unlikely(!buffer_uptodate(bh)))
 256                                 err = -EIO;
 257                         __journal_unfile_buffer(jh);
 258                         jbd_unlock_bh_state(bh);
 259                         if (locked)
 260                                 unlock_buffer(bh);
 261                         release_data_buffer(bh);
 262                 }
 263
 264                 if (need_resched() || spin_needbreak(&journal->j_list_lock)) {
 265                         spin_unlock(&journal->j_list_lock);
 266                         goto write_out_data;
 267                 }
 268         }
 269         spin_unlock(&journal->j_list_lock);
 270         trace_jbd_do_submit_data(journal, commit_transaction);
 271         journal_do_submit_data(wbuf, bufs, write_op);
 272
 273         return err;
 274 }
 275
 276 /*
 277  * journal_commit_transaction
 278  *
 279  * The primary function for committing a transaction to the log.  This
 280  * function is called by the journal thread to begin a complete commit.
 281  */
 282 void journal_commit_transaction(journal_t *journal)
 283 {
 284         transaction_t *commit_transaction;
 285         struct journal_head *jh, *new_jh, *descriptor;
 286         struct buffer_head **wbuf = journal->j_wbuf;
 287         int bufs;
 288         int flags;
 289         int err;
 290         unsigned int blocknr;
 291         ktime_t start_time;
 292         u64 commit_time;
 293         char *tagp = NULL;
 294         journal_header_t *header;
 295         journal_block_tag_t *tag = NULL;
 296         int space_left = 0;
 297         int first_tag = 0;
 298         int tag_flag;
 299         int i;
 300         struct blk_plug plug;
 301
 302         /*
 303          * First job: lock down the current transaction and wait for
 304          * all outstanding updates to complete.
 305          */
 306
 307         /* Do we need to erase the effects of a prior journal_flush? */
 308         if (journal->j_flags & JFS_FLUSHED) {
 309                 jbd_debug(3, "super block updated\n");
 310                 journal_update_superblock(journal, 1);
 311         } else {
 312                 jbd_debug(3, "superblock not updated\n");
 313         }
 314
 315         J_ASSERT(journal->j_running_transaction != NULL);
 316         J_ASSERT(journal->j_committing_transaction == NULL);
 317
 318         commit_transaction = journal->j_running_transaction;
 319         J_ASSERT(commit_transaction->t_state == T_RUNNING);
 320
 321         trace_jbd_start_commit(journal, commit_transaction);
 322         jbd_debug(1, "JBD: starting commit of transaction %d\n",
 323                         commit_transaction->t_tid);
 324
 325         spin_lock(&journal->j_state_lock);
 326         commit_transaction->t_state = T_LOCKED;
 327
 328         trace_jbd_commit_locking(journal, commit_transaction);
 329         spin_lock(&commit_transaction->t_handle_lock);
 330         while (commit_transaction->t_updates) {
 331                 DEFINE_WAIT(wait);
 332
 333                 prepare_to_wait(&journal->j_wait_updates, &wait,
 334                                         TASK_UNINTERRUPTIBLE);
 335                 if (commit_transaction->t_updates) {
 336                         spin_unlock(&commit_transaction->t_handle_lock);
 337                         spin_unlock(&journal->j_state_lock);
 338                         schedule();
 339                         spin_lock(&journal->j_state_lock);
 340                         spin_lock(&commit_transaction->t_handle_lock);
 341                 }
 342                 finish_wait(&journal->j_wait_updates, &wait);
 343         }
 344         spin_unlock(&commit_transaction->t_handle_lock);
 345
 346         J_ASSERT (commit_transaction->t_outstanding_credits <=
 347                         journal->j_max_transaction_buffers);
 348
 349         /*
 350          * First thing we are allowed to do is to discard any remaining
 351          * BJ_Reserved buffers.  Note, it is _not_ permissible to assume
 352          * that there are no such buffers: if a large filesystem
 353          * operation like a truncate needs to split itself over multiple
 354          * transactions, then it may try to do a journal_restart() while
 355          * there are still BJ_Reserved buffers outstanding.  These must
 356          * be released cleanly from the current transaction.
 357          *
 358          * In this case, the filesystem must still reserve write access
 359          * again before modifying the buffer in the new transaction, but
 360          * we do not require it to remember exactly which old buffers it
 361          * has reserved.  This is consistent with the existing behaviour
 362          * that multiple journal_get_write_access() calls to the same
 363          * buffer are perfectly permissible.
 364          */
 365         while (commit_transaction->t_reserved_list) {
 366                 jh = commit_transaction->t_reserved_list;
 367                 JBUFFER_TRACE(jh, "reserved, unused: refile");
 368                 /*
 369                  * A journal_get_undo_access()+journal_release_buffer() may
 370                  * leave undo-committed data.
 371                  */
 372                 if (jh->b_committed_data) {
 373                         struct buffer_head *bh = jh2bh(jh);
 374
 375                         jbd_lock_bh_state(bh);
 376                         jbd_free(jh->b_committed_data, bh->b_size);
 377                         jh->b_committed_data = NULL;
 378                         jbd_unlock_bh_state(bh);
 379                 }
 380                 journal_refile_buffer(journal, jh);
 381         }
 382
 383         /*
 384          * Now try to drop any written-back buffers from the journal's
 385          * checkpoint lists.  We do this *before* commit because it potentially
 386          * frees some memory
 387          */
 388         spin_lock(&journal->j_list_lock);
 389         __journal_clean_checkpoint_list(journal);
 390         spin_unlock(&journal->j_list_lock);
 391
 392         jbd_debug (3, "JBD: commit phase 1\n");
 393
 394         /*
 395          * Clear revoked flag to reflect there is no revoked buffers
 396          * in the next transaction which is going to be started.
 397          */
 398         journal_clear_buffer_revoked_flags(journal);
 399
 400         /*
 401          * Switch to a new revoke table.
 402          */
 403         journal_switch_revoke_table(journal);
 404
 405         trace_jbd_commit_flushing(journal, commit_transaction);
 406         commit_transaction->t_state = T_FLUSH;
 407         journal->j_committing_transaction = commit_transaction;
 408         journal->j_running_transaction = NULL;
 409         start_time = ktime_get();
 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          * Now start flushing things to disk, in the order they appear
 418          * on the transaction lists.  Data blocks go first.
 419          */
 420         blk_start_plug(&plug);
 421         err = journal_submit_data_buffers(journal, commit_transaction,
 422                                           WRITE_SYNC);
 423         blk_finish_plug(&plug);
 424
 425         /*
 426          * Wait for all previously submitted IO to complete.
 427          */
 428         spin_lock(&journal->j_list_lock);
 429         while (commit_transaction->t_locked_list) {
 430                 struct buffer_head *bh;
 431
 432                 jh = commit_transaction->t_locked_list->b_tprev;
 433                 bh = jh2bh(jh);
 434                 get_bh(bh);
 435                 if (buffer_locked(bh)) {
 436                         spin_unlock(&journal->j_list_lock);
 437                         wait_on_buffer(bh);
 438                         spin_lock(&journal->j_list_lock);
 439                 }
 440                 if (unlikely(!buffer_uptodate(bh))) {
 441                         if (!trylock_page(bh->b_page)) {
 442                                 spin_unlock(&journal->j_list_lock);
 443                                 lock_page(bh->b_page);
 444                                 spin_lock(&journal->j_list_lock);
 445                         }
 446                         if (bh->b_page->mapping)
 447                                 set_bit(AS_EIO, &bh->b_page->mapping->flags);
 448
 449                         unlock_page(bh->b_page);
 450                         SetPageError(bh->b_page);
 451                         err = -EIO;
 452                 }
 453                 if (!inverted_lock(journal, bh)) {
 454                         put_bh(bh);
 455                         spin_lock(&journal->j_list_lock);
 456                         continue;
 457                 }
 458                 if (buffer_jbd(bh) && bh2jh(bh) == jh &&
 459                     jh->b_transaction == commit_transaction &&
 460                     jh->b_jlist == BJ_Locked)
 461                         __journal_unfile_buffer(jh);
 462                 jbd_unlock_bh_state(bh);
 463                 release_data_buffer(bh);
 464                 cond_resched_lock(&journal->j_list_lock);
 465         }
 466         spin_unlock(&journal->j_list_lock);
 467
 468         if (err) {
 469                 char b[BDEVNAME_SIZE];
 470
 471                 printk(KERN_WARNING
 472                         "JBD: Detected IO errors while flushing file data "
 473                         "on %s\n", bdevname(journal->j_fs_dev, b));
 474                 if (journal->j_flags & JFS_ABORT_ON_SYNCDATA_ERR)
 475                         journal_abort(journal, err);
 476                 err = 0;
 477         }
 478
 479         blk_start_plug(&plug);
 480
 481         journal_write_revoke_records(journal, commit_transaction, WRITE_SYNC);
 482
 483         /*
 484          * If we found any dirty or locked buffers, then we should have
 485          * looped back up to the write_out_data label.  If there weren't
 486          * any then journal_clean_data_list should have wiped the list
 487          * clean by now, so check that it is in fact empty.
 488          */
 489         J_ASSERT (commit_transaction->t_sync_datalist == NULL);
 490
 491         jbd_debug (3, "JBD: commit phase 3\n");
 492
 493         /*
 494          * Way to go: we have now written out all of the data for a
 495          * transaction!  Now comes the tricky part: we need to write out
 496          * metadata.  Loop over the transaction's entire buffer list:
 497          */
 498         spin_lock(&journal->j_state_lock);
 499         commit_transaction->t_state = T_COMMIT;
 500         spin_unlock(&journal->j_state_lock);
 501
 502         trace_jbd_commit_logging(journal, commit_transaction);
 503         J_ASSERT(commit_transaction->t_nr_buffers <=
 504                  commit_transaction->t_outstanding_credits);
 505
 506         descriptor = NULL;
 507         bufs = 0;
 508         while (commit_transaction->t_buffers) {
 509
 510                 /* Find the next buffer to be journaled... */
 511
 512                 jh = commit_transaction->t_buffers;
 513
 514                 /* If we're in abort mode, we just un-journal the buffer and
 515                    release it. */
 516
 517                 if (is_journal_aborted(journal)) {
 518                         clear_buffer_jbddirty(jh2bh(jh));
 519                         JBUFFER_TRACE(jh, "journal is aborting: refile");
 520                         journal_refile_buffer(journal, jh);
 521                         /* If that was the last one, we need to clean up
 522                          * any descriptor buffers which may have been
 523                          * already allocated, even if we are now
 524                          * aborting. */
 525                         if (!commit_transaction->t_buffers)
 526                                 goto start_journal_io;
 527                         continue;
 528                 }
 529
 530                 /* Make sure we have a descriptor block in which to
 531                    record the metadata buffer. */
 532
 533                 if (!descriptor) {
 534                         struct buffer_head *bh;
 535
 536                         J_ASSERT (bufs == 0);
 537
 538                         jbd_debug(4, "JBD: get descriptor\n");
 539
 540                         descriptor = journal_get_descriptor_buffer(journal);
 541                         if (!descriptor) {
 542                                 journal_abort(journal, -EIO);
 543                                 continue;
 544                         }
 545
 546                         bh = jh2bh(descriptor);
 547                         jbd_debug(4, "JBD: got buffer %llu (%p)\n",
 548                                 (unsigned long long)bh->b_blocknr, bh->b_data);
 549                         header = (journal_header_t *)&bh->b_data[0];
 550                         header->h_magic     = cpu_to_be32(JFS_MAGIC_NUMBER);
 551                         header->h_blocktype = cpu_to_be32(JFS_DESCRIPTOR_BLOCK);
 552                         header->h_sequence  = cpu_to_be32(commit_transaction->t_tid);
 553
 554                         tagp = &bh->b_data[sizeof(journal_header_t)];
 555                         space_left = bh->b_size - sizeof(journal_header_t);
 556                         first_tag = 1;
 557                         set_buffer_jwrite(bh);
 558                         set_buffer_dirty(bh);
 559                         wbuf[bufs++] = bh;
 560
 561                         /* Record it so that we can wait for IO
 562                            completion later */
 563                         BUFFER_TRACE(bh, "ph3: file as descriptor");
 564                         journal_file_buffer(descriptor, commit_transaction,
 565                                         BJ_LogCtl);
 566                 }
 567
 568                 /* Where is the buffer to be written? */
 569
 570                 err = journal_next_log_block(journal, &blocknr);
 571                 /* If the block mapping failed, just abandon the buffer
 572                    and repeat this loop: we'll fall into the
 573                    refile-on-abort condition above. */
 574                 if (err) {
 575                         journal_abort(journal, err);
 576                         continue;
 577                 }
 578
 579                 /*
 580                  * start_this_handle() uses t_outstanding_credits to determine
 581                  * the free space in the log, but this counter is changed
 582                  * by journal_next_log_block() also.
 583                  */
 584                 commit_transaction->t_outstanding_credits--;
 585
 586                 /* Bump b_count to prevent truncate from stumbling over
 587                    the shadowed buffer!  @@@ This can go if we ever get
 588                    rid of the BJ_IO/BJ_Shadow pairing of buffers. */
 589                 get_bh(jh2bh(jh));
 590
 591                 /* Make a temporary IO buffer with which to write it out
 592                    (this will requeue both the metadata buffer and the
 593                    temporary IO buffer). new_bh goes on BJ_IO*/
 594
 595                 set_buffer_jwrite(jh2bh(jh));
 596                 /*
 597                  * akpm: journal_write_metadata_buffer() sets
 598                  * new_bh->b_transaction to commit_transaction.
 599                  * We need to clean this up before we release new_bh
 600                  * (which is of type BJ_IO)
 601                  */
 602                 JBUFFER_TRACE(jh, "ph3: write metadata");
 603                 flags = journal_write_metadata_buffer(commit_transaction,
 604                                                       jh, &new_jh, blocknr);
 605                 set_buffer_jwrite(jh2bh(new_jh));
 606                 wbuf[bufs++] = jh2bh(new_jh);
 607
 608                 /* Record the new block's tag in the current descriptor
 609                    buffer */
 610
 611                 tag_flag = 0;
 612                 if (flags & 1)
 613                         tag_flag |= JFS_FLAG_ESCAPE;
 614                 if (!first_tag)
 615                         tag_flag |= JFS_FLAG_SAME_UUID;
 616
 617                 tag = (journal_block_tag_t *) tagp;
 618                 tag->t_blocknr = cpu_to_be32(jh2bh(jh)->b_blocknr);
 619                 tag->t_flags = cpu_to_be32(tag_flag);
 620                 tagp += sizeof(journal_block_tag_t);
 621                 space_left -= sizeof(journal_block_tag_t);
 622
 623                 if (first_tag) {
 624                         memcpy (tagp, journal->j_uuid, 16);
 625                         tagp += 16;
 626                         space_left -= 16;
 627                         first_tag = 0;
 628                 }
 629
 630                 /* If there's no more to do, or if the descriptor is full,
 631                    let the IO rip! */
 632
 633                 if (bufs == journal->j_wbufsize ||
 634                     commit_transaction->t_buffers == NULL ||
 635                     space_left < sizeof(journal_block_tag_t) + 16) {
 636
 637                         jbd_debug(4, "JBD: Submit %d IOs\n", bufs);
 638
 639                         /* Write an end-of-descriptor marker before
 640                            submitting the IOs.  "tag" still points to
 641                            the last tag we set up. */
 642
 643                         tag->t_flags |= cpu_to_be32(JFS_FLAG_LAST_TAG);
 644
 645 start_journal_io:
 646                         for (i = 0; i < bufs; i++) {
 647                                 struct buffer_head *bh = wbuf[i];
 648                                 lock_buffer(bh);
 649                                 clear_buffer_dirty(bh);
 650                                 set_buffer_uptodate(bh);
 651                                 bh->b_end_io = journal_end_buffer_io_sync;
 652                                 submit_bh(WRITE_SYNC, bh);
 653                         }
 654                         cond_resched();
 655
 656                         /* Force a new descriptor to be generated next
 657                            time round the loop. */
 658                         descriptor = NULL;
 659                         bufs = 0;
 660                 }
 661         }
 662
 663         blk_finish_plug(&plug);
 664
 665         /* Lo and behold: we have just managed to send a transaction to
 666            the log.  Before we can commit it, wait for the IO so far to
 667            complete.  Control buffers being written are on the
 668            transaction's t_log_list queue, and metadata buffers are on
 669            the t_iobuf_list queue.
 670
 671            Wait for the buffers in reverse order.  That way we are
 672            less likely to be woken up until all IOs have completed, and
 673            so we incur less scheduling load.
 674         */
 675
 676         jbd_debug(3, "JBD: commit phase 4\n");
 677
 678         /*
 679          * akpm: these are BJ_IO, and j_list_lock is not needed.
 680          * See __journal_try_to_free_buffer.
 681          */
 682 wait_for_iobuf:
 683         while (commit_transaction->t_iobuf_list != NULL) {
 684                 struct buffer_head *bh;
 685
 686                 jh = commit_transaction->t_iobuf_list->b_tprev;
 687                 bh = jh2bh(jh);
 688                 if (buffer_locked(bh)) {
 689                         wait_on_buffer(bh);
 690                         goto wait_for_iobuf;
 691                 }
 692                 if (cond_resched())
 693                         goto wait_for_iobuf;
 694
 695                 if (unlikely(!buffer_uptodate(bh)))
 696                         err = -EIO;
 697
 698                 clear_buffer_jwrite(bh);
 699
 700                 JBUFFER_TRACE(jh, "ph4: unfile after journal write");
 701                 journal_unfile_buffer(journal, jh);
 702
 703                 /*
 704                  * ->t_iobuf_list should contain only dummy buffer_heads
 705                  * which were created by journal_write_metadata_buffer().
 706                  */
 707                 BUFFER_TRACE(bh, "dumping temporary bh");
 708                 journal_put_journal_head(jh);
 709                 __brelse(bh);
 710                 J_ASSERT_BH(bh, atomic_read(&bh->b_count) == 0);
 711                 free_buffer_head(bh);
 712
 713                 /* We also have to unlock and free the corresponding
 714                    shadowed buffer */
 715                 jh = commit_transaction->t_shadow_list->b_tprev;
 716                 bh = jh2bh(jh);
 717                 clear_buffer_jwrite(bh);
 718                 J_ASSERT_BH(bh, buffer_jbddirty(bh));
 719
 720                 /* The metadata is now released for reuse, but we need
 721                    to remember it against this transaction so that when
 722                    we finally commit, we can do any checkpointing
 723                    required. */
 724                 JBUFFER_TRACE(jh, "file as BJ_Forget");
 725                 journal_file_buffer(jh, commit_transaction, BJ_Forget);
 726                 /*
 727                  * Wake up any transactions which were waiting for this
 728                  * IO to complete. The barrier must be here so that changes
 729                  * by journal_file_buffer() take effect before wake_up_bit()
 730                  * does the waitqueue check.
 731                  */
 732                 smp_mb();
 733                 wake_up_bit(&bh->b_state, BH_Unshadow);
 734                 JBUFFER_TRACE(jh, "brelse shadowed buffer");
 735                 __brelse(bh);
 736         }
 737
 738         J_ASSERT (commit_transaction->t_shadow_list == NULL);
 739
 740         jbd_debug(3, "JBD: commit phase 5\n");
 741
 742         /* Here we wait for the revoke record and descriptor record buffers */
 743  wait_for_ctlbuf:
 744         while (commit_transaction->t_log_list != NULL) {
 745                 struct buffer_head *bh;
 746
 747                 jh = commit_transaction->t_log_list->b_tprev;
 748                 bh = jh2bh(jh);
 749                 if (buffer_locked(bh)) {
 750                         wait_on_buffer(bh);
 751                         goto wait_for_ctlbuf;
 752                 }
 753                 if (cond_resched())
 754                         goto wait_for_ctlbuf;
 755
 756                 if (unlikely(!buffer_uptodate(bh)))
 757                         err = -EIO;
 758
 759                 BUFFER_TRACE(bh, "ph5: control buffer writeout done: unfile");
 760                 clear_buffer_jwrite(bh);
 761                 journal_unfile_buffer(journal, jh);
 762                 journal_put_journal_head(jh);
 763                 __brelse(bh);           /* One for getblk */
 764                 /* AKPM: bforget here */
 765         }
 766
 767         if (err)
 768                 journal_abort(journal, err);
 769
 770         jbd_debug(3, "JBD: commit phase 6\n");
 771
 772         /* All metadata is written, now write commit record and do cleanup */
 773         spin_lock(&journal->j_state_lock);
 774         J_ASSERT(commit_transaction->t_state == T_COMMIT);
 775         commit_transaction->t_state = T_COMMIT_RECORD;
 776         spin_unlock(&journal->j_state_lock);
 777
 778         if (journal_write_commit_record(journal, commit_transaction))
 779                 err = -EIO;
 780
 781         if (err)
 782                 journal_abort(journal, err);
 783
 784         /* End of a transaction!  Finally, we can do checkpoint
 785            processing: any buffers committed as a result of this
 786            transaction can be removed from any checkpoint list it was on
 787            before. */
 788
 789         jbd_debug(3, "JBD: commit phase 7\n");
 790
 791         J_ASSERT(commit_transaction->t_sync_datalist == NULL);
 792         J_ASSERT(commit_transaction->t_buffers == NULL);
 793         J_ASSERT(commit_transaction->t_checkpoint_list == NULL);
 794         J_ASSERT(commit_transaction->t_iobuf_list == NULL);
 795         J_ASSERT(commit_transaction->t_shadow_list == NULL);
 796         J_ASSERT(commit_transaction->t_log_list == NULL);
 797
 798 restart_loop:
 799         /*
 800          * As there are other places (journal_unmap_buffer()) adding buffers
 801          * to this list we have to be careful and hold the j_list_lock.
 802          */
 803         spin_lock(&journal->j_list_lock);
 804         while (commit_transaction->t_forget) {
 805                 transaction_t *cp_transaction;
 806                 struct buffer_head *bh;
 807                 int try_to_free = 0;
 808
 809                 jh = commit_transaction->t_forget;
 810                 spin_unlock(&journal->j_list_lock);
 811                 bh = jh2bh(jh);
 812                 /*
 813                  * Get a reference so that bh cannot be freed before we are
 814                  * done with it.
 815                  */
 816                 get_bh(bh);
 817                 jbd_lock_bh_state(bh);
 818                 J_ASSERT_JH(jh, jh->b_transaction == commit_transaction ||
 819                         jh->b_transaction == journal->j_running_transaction);
 820
 821                 /*
 822                  * If there is undo-protected committed data against
 823                  * this buffer, then we can remove it now.  If it is a
 824                  * buffer needing such protection, the old frozen_data
 825                  * field now points to a committed version of the
 826                  * buffer, so rotate that field to the new committed
 827                  * data.
 828                  *
 829                  * Otherwise, we can just throw away the frozen data now.
 830                  */
 831                 if (jh->b_committed_data) {
 832                         jbd_free(jh->b_committed_data, bh->b_size);
 833                         jh->b_committed_data = NULL;
 834                         if (jh->b_frozen_data) {
 835                                 jh->b_committed_data = jh->b_frozen_data;
 836                                 jh->b_frozen_data = NULL;
 837                         }
 838                 } else if (jh->b_frozen_data) {
 839                         jbd_free(jh->b_frozen_data, bh->b_size);
 840                         jh->b_frozen_data = NULL;
 841                 }
 842
 843                 spin_lock(&journal->j_list_lock);
 844                 cp_transaction = jh->b_cp_transaction;
 845                 if (cp_transaction) {
 846                         JBUFFER_TRACE(jh, "remove from old cp transaction");
 847                         __journal_remove_checkpoint(jh);
 848                 }
 849
 850                 /* Only re-checkpoint the buffer_head if it is marked
 851                  * dirty.  If the buffer was added to the BJ_Forget list
 852                  * by journal_forget, it may no longer be dirty and
 853                  * there's no point in keeping a checkpoint record for
 854                  * it. */
 855
 856                 /* A buffer which has been freed while still being
 857                  * journaled by a previous transaction may end up still
 858                  * being dirty here, but we want to avoid writing back
 859                  * that buffer in the future after the "add to orphan"
 860                  * operation been committed,  That's not only a performance
 861                  * gain, it also stops aliasing problems if the buffer is
 862                  * left behind for writeback and gets reallocated for another
 863                  * use in a different page. */
 864                 if (buffer_freed(bh) && !jh->b_next_transaction) {
 865                         clear_buffer_freed(bh);
 866                         clear_buffer_jbddirty(bh);
 867                 }
 868
 869                 if (buffer_jbddirty(bh)) {
 870                         JBUFFER_TRACE(jh, "add to new checkpointing trans");
 871                         __journal_insert_checkpoint(jh, commit_transaction);
 872                         if (is_journal_aborted(journal))
 873                                 clear_buffer_jbddirty(bh);
 874                 } else {
 875                         J_ASSERT_BH(bh, !buffer_dirty(bh));
 876                         /*
 877                          * The buffer on BJ_Forget list and not jbddirty means
 878                          * it has been freed by this transaction and hence it
 879                          * could not have been reallocated until this
 880                          * transaction has committed. *BUT* it could be
 881                          * reallocated once we have written all the data to
 882                          * disk and before we process the buffer on BJ_Forget
 883                          * list.
 884                          */
 885                         if (!jh->b_next_transaction)
 886                                 try_to_free = 1;
 887                 }
 888                 JBUFFER_TRACE(jh, "refile or unfile freed buffer");
 889                 __journal_refile_buffer(jh);
 890                 jbd_unlock_bh_state(bh);
 891                 if (try_to_free)
 892                         release_buffer_page(bh);
 893                 else
 894                         __brelse(bh);
 895                 cond_resched_lock(&journal->j_list_lock);
 896         }
 897         spin_unlock(&journal->j_list_lock);
 898         /*
 899          * This is a bit sleazy.  We use j_list_lock to protect transition
 900          * of a transaction into T_FINISHED state and calling
 901          * __journal_drop_transaction(). Otherwise we could race with
 902          * other checkpointing code processing the transaction...
 903          */
 904         spin_lock(&journal->j_state_lock);
 905         spin_lock(&journal->j_list_lock);
 906         /*
 907          * Now recheck if some buffers did not get attached to the transaction
 908          * while the lock was dropped...
 909          */
 910         if (commit_transaction->t_forget) {
 911                 spin_unlock(&journal->j_list_lock);
 912                 spin_unlock(&journal->j_state_lock);
 913                 goto restart_loop;
 914         }
 915
 916         /* Done with this transaction! */
 917
 918         jbd_debug(3, "JBD: commit phase 8\n");
 919
 920         J_ASSERT(commit_transaction->t_state == T_COMMIT_RECORD);
 921
 922         commit_transaction->t_state = T_FINISHED;
 923         J_ASSERT(commit_transaction == journal->j_committing_transaction);
 924         journal->j_commit_sequence = commit_transaction->t_tid;
 925         journal->j_committing_transaction = NULL;
 926         commit_time = ktime_to_ns(ktime_sub(ktime_get(), start_time));
 927
 928         /*
 929          * weight the commit time higher than the average time so we don't
 930          * react too strongly to vast changes in commit time
 931          */
 932         if (likely(journal->j_average_commit_time))
 933                 journal->j_average_commit_time = (commit_time*3 +
 934                                 journal->j_average_commit_time) / 4;
 935         else
 936                 journal->j_average_commit_time = commit_time;
 937
 938         spin_unlock(&journal->j_state_lock);
 939
 940         if (commit_transaction->t_checkpoint_list == NULL &&
 941             commit_transaction->t_checkpoint_io_list == NULL) {
 942                 __journal_drop_transaction(journal, commit_transaction);
 943         } else {
 944                 if (journal->j_checkpoint_transactions == NULL) {
 945                         journal->j_checkpoint_transactions = commit_transaction;
 946                         commit_transaction->t_cpnext = commit_transaction;
 947                         commit_transaction->t_cpprev = commit_transaction;
 948                 } else {
 949                         commit_transaction->t_cpnext =
 950                                 journal->j_checkpoint_transactions;
 951                         commit_transaction->t_cpprev =
 952                                 commit_transaction->t_cpnext->t_cpprev;
 953                         commit_transaction->t_cpnext->t_cpprev =
 954                                 commit_transaction;
 955                         commit_transaction->t_cpprev->t_cpnext =
 956                                 commit_transaction;
 957                 }
 958         }
 959         spin_unlock(&journal->j_list_lock);
 960
 961         trace_jbd_end_commit(journal, commit_transaction);
 962         jbd_debug(1, "JBD: commit %d complete, head %d\n",
 963                   journal->j_commit_sequence, journal->j_tail_sequence);
 964
 965         wake_up(&journal->j_wait_done_commit);
 966 }