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                 WARN_ON_ONCE(buffer_dirty(bh));
  90                 clear_buffer_freed(bh);
  91                 clear_buffer_mapped(bh);
  92                 clear_buffer_new(bh);
  93                 clear_buffer_req(bh);
  94                 bh->b_bdev = NULL;
  95                 release_buffer_page(bh);
  96         } else
  97                 put_bh(bh);
  98 }
  99
 100 /*
 101  * Try to acquire jbd_lock_bh_state() against the buffer, when j_list_lock is
 102  * held.  For ranking reasons we must trylock.  If we lose, schedule away and
 103  * return 0.  j_list_lock is dropped in this case.
 104  */
 105 static int inverted_lock(journal_t *journal, struct buffer_head *bh)
 106 {
 107         if (!jbd_trylock_bh_state(bh)) {
 108                 spin_unlock(&journal->j_list_lock);
 109                 schedule();
 110                 return 0;
 111         }
 112         return 1;
 113 }
 114
 115 /* Done it all: now write the commit record.  We should have
 116  * cleaned up our previous buffers by now, so if we are in abort
 117  * mode we can now just skip the rest of the journal write
 118  * entirely.
 119  *
 120  * Returns 1 if the journal needs to be aborted or 0 on success
 121  */
 122 static int journal_write_commit_record(journal_t *journal,
 123                                         transaction_t *commit_transaction)
 124 {
 125         struct journal_head *descriptor;
 126         struct buffer_head *bh;
 127         journal_header_t *header;
 128         int ret;
 129
 130         if (is_journal_aborted(journal))
 131                 return 0;
 132
 133         descriptor = journal_get_descriptor_buffer(journal);
 134         if (!descriptor)
 135                 return 1;
 136
 137         bh = jh2bh(descriptor);
 138
 139         header = (journal_header_t *)(bh->b_data);
 140         header->h_magic = cpu_to_be32(JFS_MAGIC_NUMBER);
 141         header->h_blocktype = cpu_to_be32(JFS_COMMIT_BLOCK);
 142         header->h_sequence = cpu_to_be32(commit_transaction->t_tid);
 143
 144         JBUFFER_TRACE(descriptor, "write commit block");
 145         set_buffer_dirty(bh);
 146
 147         if (journal->j_flags & JFS_BARRIER)
 148                 ret = __sync_dirty_buffer(bh, WRITE_SYNC | WRITE_FLUSH_FUA);
 149         else
 150                 ret = sync_dirty_buffer(bh);
 151
 152         put_bh(bh);             /* One for getblk() */
 153         journal_put_journal_head(descriptor);
 154
 155         return (ret == -EIO);
 156 }
 157
 158 static void journal_do_submit_data(struct buffer_head **wbuf, int bufs,
 159                                    int write_op)
 160 {
 161         int i;
 162
 163         for (i = 0; i < bufs; i++) {
 164                 wbuf[i]->b_end_io = end_buffer_write_sync;
 165                 /* We use-up our safety reference in submit_bh() */
 166                 submit_bh(write_op, wbuf[i]);
 167         }
 168 }
 169
 170 /*
 171  *  Submit all the data buffers to disk
 172  */
 173 static int journal_submit_data_buffers(journal_t *journal,
 174                                        transaction_t *commit_transaction,
 175                                        int write_op)
 176 {
 177         struct journal_head *jh;
 178         struct buffer_head *bh;
 179         int locked;
 180         int bufs = 0;
 181         struct buffer_head **wbuf = journal->j_wbuf;
 182         int err = 0;
 183
 184         /*
 185          * Whenever we unlock the journal and sleep, things can get added
 186          * onto ->t_sync_datalist, so we have to keep looping back to
 187          * write_out_data until we *know* that the list is empty.
 188          *
 189          * Cleanup any flushed data buffers from the data list.  Even in
 190          * abort mode, we want to flush this out as soon as possible.
 191          */
 192 write_out_data:
 193         cond_resched();
 194         spin_lock(&journal->j_list_lock);
 195
 196         while (commit_transaction->t_sync_datalist) {
 197                 jh = commit_transaction->t_sync_datalist;
 198                 bh = jh2bh(jh);
 199                 locked = 0;
 200
 201                 /* Get reference just to make sure buffer does not disappear
 202                  * when we are forced to drop various locks */
 203                 get_bh(bh);
 204                 /* If the buffer is dirty, we need to submit IO and hence
 205                  * we need the buffer lock. We try to lock the buffer without
 206                  * blocking. If we fail, we need to drop j_list_lock and do
 207                  * blocking lock_buffer().
 208                  */
 209                 if (buffer_dirty(bh)) {
 210                         if (!trylock_buffer(bh)) {
 211                                 BUFFER_TRACE(bh, "needs blocking lock");
 212                                 spin_unlock(&journal->j_list_lock);
 213                                 trace_jbd_do_submit_data(journal,
 214                                                      commit_transaction);
 215                                 /* Write out all data to prevent deadlocks */
 216                                 journal_do_submit_data(wbuf, bufs, write_op);
 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) || bh2jh(bh) != jh
 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                         release_data_buffer(bh);
 237                         continue;
 238                 }
 239                 if (locked && test_clear_buffer_dirty(bh)) {
 240                         BUFFER_TRACE(bh, "needs writeout, adding to array");
 241                         wbuf[bufs++] = bh;
 242                         __journal_file_buffer(jh, commit_transaction,
 243                                                 BJ_Locked);
 244                         jbd_unlock_bh_state(bh);
 245                         if (bufs == journal->j_wbufsize) {
 246                                 spin_unlock(&journal->j_list_lock);
 247                                 trace_jbd_do_submit_data(journal,
 248                                                      commit_transaction);
 249                                 journal_do_submit_data(wbuf, bufs, write_op);
 250                                 bufs = 0;
 251                                 goto write_out_data;
 252                         }
 253                 } else if (!locked && buffer_locked(bh)) {
 254                         __journal_file_buffer(jh, commit_transaction,
 255                                                 BJ_Locked);
 256                         jbd_unlock_bh_state(bh);
 257                         put_bh(bh);
 258                 } else {
 259                         BUFFER_TRACE(bh, "writeout complete: unfile");
 260                         if (unlikely(!buffer_uptodate(bh)))
 261                                 err = -EIO;
 262                         __journal_unfile_buffer(jh);
 263                         jbd_unlock_bh_state(bh);
 264                         if (locked)
 265                                 unlock_buffer(bh);
 266                         release_data_buffer(bh);
 267                 }
 268
 269                 if (need_resched() || spin_needbreak(&journal->j_list_lock)) {
 270                         spin_unlock(&journal->j_list_lock);
 271                         goto write_out_data;
 272                 }
 273         }
 274         spin_unlock(&journal->j_list_lock);
 275         trace_jbd_do_submit_data(journal, commit_transaction);
 276         journal_do_submit_data(wbuf, bufs, write_op);
 277
 278         return err;
 279 }
 280
 281 /*
 282  * journal_commit_transaction
 283  *
 284  * The primary function for committing a transaction to the log.  This
 285  * function is called by the journal thread to begin a complete commit.
 286  */
 287 void journal_commit_transaction(journal_t *journal)
 288 {
 289         transaction_t *commit_transaction;
 290         struct journal_head *jh, *new_jh, *descriptor;
 291         struct buffer_head **wbuf = journal->j_wbuf;
 292         int bufs;
 293         int flags;
 294         int err;
 295         unsigned int blocknr;
 296         ktime_t start_time;
 297         u64 commit_time;
 298         char *tagp = NULL;
 299         journal_header_t *header;
 300         journal_block_tag_t *tag = NULL;
 301         int space_left = 0;
 302         int first_tag = 0;
 303         int tag_flag;
 304         int i;
 305         struct blk_plug plug;
 306
 307         /*
 308          * First job: lock down the current transaction and wait for
 309          * all outstanding updates to complete.
 310          */
 311
 312         /* Do we need to erase the effects of a prior journal_flush? */
 313         if (journal->j_flags & JFS_FLUSHED) {
 314                 jbd_debug(3, "super block updated\n");
 315                 journal_update_superblock(journal, 1);
 316         } else {
 317                 jbd_debug(3, "superblock not updated\n");
 318         }
 319
 320         J_ASSERT(journal->j_running_transaction != NULL);
 321         J_ASSERT(journal->j_committing_transaction == NULL);
 322
 323         commit_transaction = journal->j_running_transaction;
 324         J_ASSERT(commit_transaction->t_state == T_RUNNING);
 325
 326         trace_jbd_start_commit(journal, commit_transaction);
 327         jbd_debug(1, "JBD: starting commit of transaction %d\n",
 328                         commit_transaction->t_tid);
 329
 330         spin_lock(&journal->j_state_lock);
 331         commit_transaction->t_state = T_LOCKED;
 332
 333         trace_jbd_commit_locking(journal, commit_transaction);
 334         spin_lock(&commit_transaction->t_handle_lock);
 335         while (commit_transaction->t_updates) {
 336                 DEFINE_WAIT(wait);
 337
 338                 prepare_to_wait(&journal->j_wait_updates, &wait,
 339                                         TASK_UNINTERRUPTIBLE);
 340                 if (commit_transaction->t_updates) {
 341                         spin_unlock(&commit_transaction->t_handle_lock);
 342                         spin_unlock(&journal->j_state_lock);
 343                         schedule();
 344                         spin_lock(&journal->j_state_lock);
 345                         spin_lock(&commit_transaction->t_handle_lock);
 346                 }
 347                 finish_wait(&journal->j_wait_updates, &wait);
 348         }
 349         spin_unlock(&commit_transaction->t_handle_lock);
 350
 351         J_ASSERT (commit_transaction->t_outstanding_credits <=
 352                         journal->j_max_transaction_buffers);
 353
 354         /*
 355          * First thing we are allowed to do is to discard any remaining
 356          * BJ_Reserved buffers.  Note, it is _not_ permissible to assume
 357          * that there are no such buffers: if a large filesystem
 358          * operation like a truncate needs to split itself over multiple
 359          * transactions, then it may try to do a journal_restart() while
 360          * there are still BJ_Reserved buffers outstanding.  These must
 361          * be released cleanly from the current transaction.
 362          *
 363          * In this case, the filesystem must still reserve write access
 364          * again before modifying the buffer in the new transaction, but
 365          * we do not require it to remember exactly which old buffers it
 366          * has reserved.  This is consistent with the existing behaviour
 367          * that multiple journal_get_write_access() calls to the same
 368          * buffer are perfectly permissible.
 369          */
 370         while (commit_transaction->t_reserved_list) {
 371                 jh = commit_transaction->t_reserved_list;
 372                 JBUFFER_TRACE(jh, "reserved, unused: refile");
 373                 /*
 374                  * A journal_get_undo_access()+journal_release_buffer() may
 375                  * leave undo-committed data.
 376                  */
 377                 if (jh->b_committed_data) {
 378                         struct buffer_head *bh = jh2bh(jh);
 379
 380                         jbd_lock_bh_state(bh);
 381                         jbd_free(jh->b_committed_data, bh->b_size);
 382                         jh->b_committed_data = NULL;
 383                         jbd_unlock_bh_state(bh);
 384                 }
 385                 journal_refile_buffer(journal, jh);
 386         }
 387
 388         /*
 389          * Now try to drop any written-back buffers from the journal's
 390          * checkpoint lists.  We do this *before* commit because it potentially
 391          * frees some memory
 392          */
 393         spin_lock(&journal->j_list_lock);
 394         __journal_clean_checkpoint_list(journal);
 395         spin_unlock(&journal->j_list_lock);
 396
 397         jbd_debug (3, "JBD: commit phase 1\n");
 398
 399         /*
 400          * Switch to a new revoke table.
 401          */
 402         journal_switch_revoke_table(journal);
 403
 404         trace_jbd_commit_flushing(journal, commit_transaction);
 405         commit_transaction->t_state = T_FLUSH;
 406         journal->j_committing_transaction = commit_transaction;
 407         journal->j_running_transaction = NULL;
 408         start_time = ktime_get();
 409         commit_transaction->t_log_start = journal->j_head;
 410         wake_up(&journal->j_wait_transaction_locked);
 411         spin_unlock(&journal->j_state_lock);
 412
 413         jbd_debug (3, "JBD: commit phase 2\n");
 414
 415         /*
 416          * Now start flushing things to disk, in the order they appear
 417          * on the transaction lists.  Data blocks go first.
 418          */
 419         blk_start_plug(&plug);
 420         err = journal_submit_data_buffers(journal, commit_transaction,
 421                                           WRITE_SYNC);
 422         blk_finish_plug(&plug);
 423
 424         /*
 425          * Wait for all previously submitted IO to complete.
 426          */
 427         spin_lock(&journal->j_list_lock);
 428         while (commit_transaction->t_locked_list) {
 429                 struct buffer_head *bh;
 430
 431                 jh = commit_transaction->t_locked_list->b_tprev;
 432                 bh = jh2bh(jh);
 433                 get_bh(bh);
 434                 if (buffer_locked(bh)) {
 435                         spin_unlock(&journal->j_list_lock);
 436                         wait_on_buffer(bh);
 437                         spin_lock(&journal->j_list_lock);
 438                 }
 439                 if (unlikely(!buffer_uptodate(bh))) {
 440                         if (!trylock_page(bh->b_page)) {
 441                                 spin_unlock(&journal->j_list_lock);
 442                                 lock_page(bh->b_page);
 443                                 spin_lock(&journal->j_list_lock);
 444                         }
 445                         if (bh->b_page->mapping)
 446                                 set_bit(AS_EIO, &bh->b_page->mapping->flags);
 447
 448                         unlock_page(bh->b_page);
 449                         SetPageError(bh->b_page);
 450                         err = -EIO;
 451                 }
 452                 if (!inverted_lock(journal, bh)) {
 453                         put_bh(bh);
 454                         spin_lock(&journal->j_list_lock);
 455                         continue;
 456                 }
 457                 if (buffer_jbd(bh) && bh2jh(bh) == jh &&
 458                     jh->b_transaction == commit_transaction &&
 459                     jh->b_jlist == BJ_Locked)
 460                         __journal_unfile_buffer(jh);
 461                 jbd_unlock_bh_state(bh);
 462                 release_data_buffer(bh);
 463                 cond_resched_lock(&journal->j_list_lock);
 464         }
 465         spin_unlock(&journal->j_list_lock);
 466
 467         if (err) {
 468                 char b[BDEVNAME_SIZE];
 469
 470                 printk(KERN_WARNING
 471                         "JBD: Detected IO errors while flushing file data "
 472                         "on %s\n", bdevname(journal->j_fs_dev, b));
 473                 if (journal->j_flags & JFS_ABORT_ON_SYNCDATA_ERR)
 474                         journal_abort(journal, err);
 475                 err = 0;
 476         }
 477
 478         blk_start_plug(&plug);
 479
 480         journal_write_revoke_records(journal, commit_transaction, WRITE_SYNC);
 481
 482         /*
 483          * If we found any dirty or locked buffers, then we should have
 484          * looped back up to the write_out_data label.  If there weren't
 485          * any then journal_clean_data_list should have wiped the list
 486          * clean by now, so check that it is in fact empty.
 487          */
 488         J_ASSERT (commit_transaction->t_sync_datalist == NULL);
 489
 490         jbd_debug (3, "JBD: commit phase 3\n");
 491
 492         /*
 493          * Way to go: we have now written out all of the data for a
 494          * transaction!  Now comes the tricky part: we need to write out
 495          * metadata.  Loop over the transaction's entire buffer list:
 496          */
 497         spin_lock(&journal->j_state_lock);
 498         commit_transaction->t_state = T_COMMIT;
 499         spin_unlock(&journal->j_state_lock);
 500
 501         trace_jbd_commit_logging(journal, commit_transaction);
 502         J_ASSERT(commit_transaction->t_nr_buffers <=
 503                  commit_transaction->t_outstanding_credits);
 504
 505         descriptor = NULL;
 506         bufs = 0;
 507         while (commit_transaction->t_buffers) {
 508
 509                 /* Find the next buffer to be journaled... */
 510
 511                 jh = commit_transaction->t_buffers;
 512
 513                 /* If we're in abort mode, we just un-journal the buffer and
 514                    release it. */
 515
 516                 if (is_journal_aborted(journal)) {
 517                         clear_buffer_jbddirty(jh2bh(jh));
 518                         JBUFFER_TRACE(jh, "journal is aborting: refile");
 519                         journal_refile_buffer(journal, jh);
 520                         /* If that was the last one, we need to clean up
 521                          * any descriptor buffers which may have been
 522                          * already allocated, even if we are now
 523                          * aborting. */
 524                         if (!commit_transaction->t_buffers)
 525                                 goto start_journal_io;
 526                         continue;
 527                 }
 528
 529                 /* Make sure we have a descriptor block in which to
 530                    record the metadata buffer. */
 531
 532                 if (!descriptor) {
 533                         struct buffer_head *bh;
 534
 535                         J_ASSERT (bufs == 0);
 536
 537                         jbd_debug(4, "JBD: get descriptor\n");
 538
 539                         descriptor = journal_get_descriptor_buffer(journal);
 540                         if (!descriptor) {
 541                                 journal_abort(journal, -EIO);
 542                                 continue;
 543                         }
 544
 545                         bh = jh2bh(descriptor);
 546                         jbd_debug(4, "JBD: got buffer %llu (%p)\n",
 547                                 (unsigned long long)bh->b_blocknr, bh->b_data);
 548                         header = (journal_header_t *)&bh->b_data[0];
 549                         header->h_magic     = cpu_to_be32(JFS_MAGIC_NUMBER);
 550                         header->h_blocktype = cpu_to_be32(JFS_DESCRIPTOR_BLOCK);
 551                         header->h_sequence  = cpu_to_be32(commit_transaction->t_tid);
 552
 553                         tagp = &bh->b_data[sizeof(journal_header_t)];
 554                         space_left = bh->b_size - sizeof(journal_header_t);
 555                         first_tag = 1;
 556                         set_buffer_jwrite(bh);
 557                         set_buffer_dirty(bh);
 558                         wbuf[bufs++] = bh;
 559
 560                         /* Record it so that we can wait for IO
 561                            completion later */
 562                         BUFFER_TRACE(bh, "ph3: file as descriptor");
 563                         journal_file_buffer(descriptor, commit_transaction,
 564                                         BJ_LogCtl);
 565                 }
 566
 567                 /* Where is the buffer to be written? */
 568
 569                 err = journal_next_log_block(journal, &blocknr);
 570                 /* If the block mapping failed, just abandon the buffer
 571                    and repeat this loop: we'll fall into the
 572                    refile-on-abort condition above. */
 573                 if (err) {
 574                         journal_abort(journal, err);
 575                         continue;
 576                 }
 577
 578                 /*
 579                  * start_this_handle() uses t_outstanding_credits to determine
 580                  * the free space in the log, but this counter is changed
 581                  * by journal_next_log_block() also.
 582                  */
 583                 commit_transaction->t_outstanding_credits--;
 584
 585                 /* Bump b_count to prevent truncate from stumbling over
 586                    the shadowed buffer!  @@@ This can go if we ever get
 587                    rid of the BJ_IO/BJ_Shadow pairing of buffers. */
 588                 get_bh(jh2bh(jh));
 589
 590                 /* Make a temporary IO buffer with which to write it out
 591                    (this will requeue both the metadata buffer and the
 592                    temporary IO buffer). new_bh goes on BJ_IO*/
 593
 594                 set_buffer_jwrite(jh2bh(jh));
 595                 /*
 596                  * akpm: journal_write_metadata_buffer() sets
 597                  * new_bh->b_transaction to commit_transaction.
 598                  * We need to clean this up before we release new_bh
 599                  * (which is of type BJ_IO)
 600                  */
 601                 JBUFFER_TRACE(jh, "ph3: write metadata");
 602                 flags = journal_write_metadata_buffer(commit_transaction,
 603                                                       jh, &new_jh, blocknr);
 604                 set_buffer_jwrite(jh2bh(new_jh));
 605                 wbuf[bufs++] = jh2bh(new_jh);
 606
 607                 /* Record the new block's tag in the current descriptor
 608                    buffer */
 609
 610                 tag_flag = 0;
 611                 if (flags & 1)
 612                         tag_flag |= JFS_FLAG_ESCAPE;
 613                 if (!first_tag)
 614                         tag_flag |= JFS_FLAG_SAME_UUID;
 615
 616                 tag = (journal_block_tag_t *) tagp;
 617                 tag->t_blocknr = cpu_to_be32(jh2bh(jh)->b_blocknr);
 618                 tag->t_flags = cpu_to_be32(tag_flag);
 619                 tagp += sizeof(journal_block_tag_t);
 620                 space_left -= sizeof(journal_block_tag_t);
 621
 622                 if (first_tag) {
 623                         memcpy (tagp, journal->j_uuid, 16);
 624                         tagp += 16;
 625                         space_left -= 16;
 626                         first_tag = 0;
 627                 }
 628
 629                 /* If there's no more to do, or if the descriptor is full,
 630                    let the IO rip! */
 631
 632                 if (bufs == journal->j_wbufsize ||
 633                     commit_transaction->t_buffers == NULL ||
 634                     space_left < sizeof(journal_block_tag_t) + 16) {
 635
 636                         jbd_debug(4, "JBD: Submit %d IOs\n", bufs);
 637
 638                         /* Write an end-of-descriptor marker before
 639                            submitting the IOs.  "tag" still points to
 640                            the last tag we set up. */
 641
 642                         tag->t_flags |= cpu_to_be32(JFS_FLAG_LAST_TAG);
 643
 644 start_journal_io:
 645                         for (i = 0; i < bufs; i++) {
 646                                 struct buffer_head *bh = wbuf[i];
 647                                 lock_buffer(bh);
 648                                 clear_buffer_dirty(bh);
 649                                 set_buffer_uptodate(bh);
 650                                 bh->b_end_io = journal_end_buffer_io_sync;
 651                                 submit_bh(WRITE_SYNC, bh);
 652                         }
 653                         cond_resched();
 654
 655                         /* Force a new descriptor to be generated next
 656                            time round the loop. */
 657                         descriptor = NULL;
 658                         bufs = 0;
 659                 }
 660         }
 661
 662         blk_finish_plug(&plug);
 663
 664         /* Lo and behold: we have just managed to send a transaction to
 665            the log.  Before we can commit it, wait for the IO so far to
 666            complete.  Control buffers being written are on the
 667            transaction's t_log_list queue, and metadata buffers are on
 668            the t_iobuf_list queue.
 669
 670            Wait for the buffers in reverse order.  That way we are
 671            less likely to be woken up until all IOs have completed, and
 672            so we incur less scheduling load.
 673         */
 674
 675         jbd_debug(3, "JBD: commit phase 4\n");
 676
 677         /*
 678          * akpm: these are BJ_IO, and j_list_lock is not needed.
 679          * See __journal_try_to_free_buffer.
 680          */
 681 wait_for_iobuf:
 682         while (commit_transaction->t_iobuf_list != NULL) {
 683                 struct buffer_head *bh;
 684
 685                 jh = commit_transaction->t_iobuf_list->b_tprev;
 686                 bh = jh2bh(jh);
 687                 if (buffer_locked(bh)) {
 688                         wait_on_buffer(bh);
 689                         goto wait_for_iobuf;
 690                 }
 691                 if (cond_resched())
 692                         goto wait_for_iobuf;
 693
 694                 if (unlikely(!buffer_uptodate(bh)))
 695                         err = -EIO;
 696
 697                 clear_buffer_jwrite(bh);
 698
 699                 JBUFFER_TRACE(jh, "ph4: unfile after journal write");
 700                 journal_unfile_buffer(journal, jh);
 701
 702                 /*
 703                  * ->t_iobuf_list should contain only dummy buffer_heads
 704                  * which were created by journal_write_metadata_buffer().
 705                  */
 706                 BUFFER_TRACE(bh, "dumping temporary bh");
 707                 journal_put_journal_head(jh);
 708                 __brelse(bh);
 709                 J_ASSERT_BH(bh, atomic_read(&bh->b_count) == 0);
 710                 free_buffer_head(bh);
 711
 712                 /* We also have to unlock and free the corresponding
 713                    shadowed buffer */
 714                 jh = commit_transaction->t_shadow_list->b_tprev;
 715                 bh = jh2bh(jh);
 716                 clear_buffer_jwrite(bh);
 717                 J_ASSERT_BH(bh, buffer_jbddirty(bh));
 718
 719                 /* The metadata is now released for reuse, but we need
 720                    to remember it against this transaction so that when
 721                    we finally commit, we can do any checkpointing
 722                    required. */
 723                 JBUFFER_TRACE(jh, "file as BJ_Forget");
 724                 journal_file_buffer(jh, commit_transaction, BJ_Forget);
 725                 /*
 726                  * Wake up any transactions which were waiting for this
 727                  * IO to complete. The barrier must be here so that changes
 728                  * by journal_file_buffer() take effect before wake_up_bit()
 729                  * does the waitqueue check.
 730                  */
 731                 smp_mb();
 732                 wake_up_bit(&bh->b_state, BH_Unshadow);
 733                 JBUFFER_TRACE(jh, "brelse shadowed buffer");
 734                 __brelse(bh);
 735         }
 736
 737         J_ASSERT (commit_transaction->t_shadow_list == NULL);
 738
 739         jbd_debug(3, "JBD: commit phase 5\n");
 740
 741         /* Here we wait for the revoke record and descriptor record buffers */
 742  wait_for_ctlbuf:
 743         while (commit_transaction->t_log_list != NULL) {
 744                 struct buffer_head *bh;
 745
 746                 jh = commit_transaction->t_log_list->b_tprev;
 747                 bh = jh2bh(jh);
 748                 if (buffer_locked(bh)) {
 749                         wait_on_buffer(bh);
 750                         goto wait_for_ctlbuf;
 751                 }
 752                 if (cond_resched())
 753                         goto wait_for_ctlbuf;
 754
 755                 if (unlikely(!buffer_uptodate(bh)))
 756                         err = -EIO;
 757
 758                 BUFFER_TRACE(bh, "ph5: control buffer writeout done: unfile");
 759                 clear_buffer_jwrite(bh);
 760                 journal_unfile_buffer(journal, jh);
 761                 journal_put_journal_head(jh);
 762                 __brelse(bh);           /* One for getblk */
 763                 /* AKPM: bforget here */
 764         }
 765
 766         if (err)
 767                 journal_abort(journal, err);
 768
 769         jbd_debug(3, "JBD: commit phase 6\n");
 770
 771         /* All metadata is written, now write commit record and do cleanup */
 772         spin_lock(&journal->j_state_lock);
 773         J_ASSERT(commit_transaction->t_state == T_COMMIT);
 774         commit_transaction->t_state = T_COMMIT_RECORD;
 775         spin_unlock(&journal->j_state_lock);
 776
 777         if (journal_write_commit_record(journal, commit_transaction))
 778                 err = -EIO;
 779
 780         if (err)
 781                 journal_abort(journal, err);
 782
 783         /* End of a transaction!  Finally, we can do checkpoint
 784            processing: any buffers committed as a result of this
 785            transaction can be removed from any checkpoint list it was on
 786            before. */
 787
 788         jbd_debug(3, "JBD: commit phase 7\n");
 789
 790         J_ASSERT(commit_transaction->t_sync_datalist == NULL);
 791         J_ASSERT(commit_transaction->t_buffers == NULL);
 792         J_ASSERT(commit_transaction->t_checkpoint_list == NULL);
 793         J_ASSERT(commit_transaction->t_iobuf_list == NULL);
 794         J_ASSERT(commit_transaction->t_shadow_list == NULL);
 795         J_ASSERT(commit_transaction->t_log_list == NULL);
 796
 797 restart_loop:
 798         /*
 799          * As there are other places (journal_unmap_buffer()) adding buffers
 800          * to this list we have to be careful and hold the j_list_lock.
 801          */
 802         spin_lock(&journal->j_list_lock);
 803         while (commit_transaction->t_forget) {
 804                 transaction_t *cp_transaction;
 805                 struct buffer_head *bh;
 806                 int try_to_free = 0;
 807
 808                 jh = commit_transaction->t_forget;
 809                 spin_unlock(&journal->j_list_lock);
 810                 bh = jh2bh(jh);
 811                 /*
 812                  * Get a reference so that bh cannot be freed before we are
 813                  * done with it.
 814                  */
 815                 get_bh(bh);
 816                 jbd_lock_bh_state(bh);
 817                 J_ASSERT_JH(jh, jh->b_transaction == commit_transaction ||
 818                         jh->b_transaction == journal->j_running_transaction);
 819
 820                 /*
 821                  * If there is undo-protected committed data against
 822                  * this buffer, then we can remove it now.  If it is a
 823                  * buffer needing such protection, the old frozen_data
 824                  * field now points to a committed version of the
 825                  * buffer, so rotate that field to the new committed
 826                  * data.
 827                  *
 828                  * Otherwise, we can just throw away the frozen data now.
 829                  */
 830                 if (jh->b_committed_data) {
 831                         jbd_free(jh->b_committed_data, bh->b_size);
 832                         jh->b_committed_data = NULL;
 833                         if (jh->b_frozen_data) {
 834                                 jh->b_committed_data = jh->b_frozen_data;
 835                                 jh->b_frozen_data = NULL;
 836                         }
 837                 } else if (jh->b_frozen_data) {
 838                         jbd_free(jh->b_frozen_data, bh->b_size);
 839                         jh->b_frozen_data = NULL;
 840                 }
 841
 842                 spin_lock(&journal->j_list_lock);
 843                 cp_transaction = jh->b_cp_transaction;
 844                 if (cp_transaction) {
 845                         JBUFFER_TRACE(jh, "remove from old cp transaction");
 846                         __journal_remove_checkpoint(jh);
 847                 }
 848
 849                 /* Only re-checkpoint the buffer_head if it is marked
 850                  * dirty.  If the buffer was added to the BJ_Forget list
 851                  * by journal_forget, it may no longer be dirty and
 852                  * there's no point in keeping a checkpoint record for
 853                  * it. */
 854
 855                 /*
 856                  * A buffer which has been freed while still being journaled by
 857                  * a previous transaction.
 858                  */
 859                 if (buffer_freed(bh)) {
 860                         /*
 861                          * If the running transaction is the one containing
 862                          * "add to orphan" operation (b_next_transaction !=
 863                          * NULL), we have to wait for that transaction to
 864                          * commit before we can really get rid of the buffer.
 865                          * So just clear b_modified to not confuse transaction
 866                          * credit accounting and refile the buffer to
 867                          * BJ_Forget of the running transaction. If the just
 868                          * committed transaction contains "add to orphan"
 869                          * operation, we can completely invalidate the buffer
 870                          * now. We are rather throughout in that since the
 871                          * buffer may be still accessible when blocksize <
 872                          * pagesize and it is attached to the last partial
 873                          * page.
 874                          */
 875                         jh->b_modified = 0;
 876                         if (!jh->b_next_transaction) {
 877                                 clear_buffer_freed(bh);
 878                                 clear_buffer_jbddirty(bh);
 879                                 clear_buffer_mapped(bh);
 880                                 clear_buffer_new(bh);
 881                                 clear_buffer_req(bh);
 882                                 bh->b_bdev = NULL;
 883                         }
 884                 }
 885
 886                 if (buffer_jbddirty(bh)) {
 887                         JBUFFER_TRACE(jh, "add to new checkpointing trans");
 888                         __journal_insert_checkpoint(jh, commit_transaction);
 889                         if (is_journal_aborted(journal))
 890                                 clear_buffer_jbddirty(bh);
 891                 } else {
 892                         J_ASSERT_BH(bh, !buffer_dirty(bh));
 893                         /*
 894                          * The buffer on BJ_Forget list and not jbddirty means
 895                          * it has been freed by this transaction and hence it
 896                          * could not have been reallocated until this
 897                          * transaction has committed. *BUT* it could be
 898                          * reallocated once we have written all the data to
 899                          * disk and before we process the buffer on BJ_Forget
 900                          * list.
 901                          */
 902                         if (!jh->b_next_transaction)
 903                                 try_to_free = 1;
 904                 }
 905                 JBUFFER_TRACE(jh, "refile or unfile freed buffer");
 906                 __journal_refile_buffer(jh);
 907                 jbd_unlock_bh_state(bh);
 908                 if (try_to_free)
 909                         release_buffer_page(bh);
 910                 else
 911                         __brelse(bh);
 912                 cond_resched_lock(&journal->j_list_lock);
 913         }
 914         spin_unlock(&journal->j_list_lock);
 915         /*
 916          * This is a bit sleazy.  We use j_list_lock to protect transition
 917          * of a transaction into T_FINISHED state and calling
 918          * __journal_drop_transaction(). Otherwise we could race with
 919          * other checkpointing code processing the transaction...
 920          */
 921         spin_lock(&journal->j_state_lock);
 922         spin_lock(&journal->j_list_lock);
 923         /*
 924          * Now recheck if some buffers did not get attached to the transaction
 925          * while the lock was dropped...
 926          */
 927         if (commit_transaction->t_forget) {
 928                 spin_unlock(&journal->j_list_lock);
 929                 spin_unlock(&journal->j_state_lock);
 930                 goto restart_loop;
 931         }
 932
 933         /* Done with this transaction! */
 934
 935         jbd_debug(3, "JBD: commit phase 8\n");
 936
 937         J_ASSERT(commit_transaction->t_state == T_COMMIT_RECORD);
 938
 939         commit_transaction->t_state = T_FINISHED;
 940         J_ASSERT(commit_transaction == journal->j_committing_transaction);
 941         journal->j_commit_sequence = commit_transaction->t_tid;
 942         journal->j_committing_transaction = NULL;
 943         commit_time = ktime_to_ns(ktime_sub(ktime_get(), start_time));
 944
 945         /*
 946          * weight the commit time higher than the average time so we don't
 947          * react too strongly to vast changes in commit time
 948          */
 949         if (likely(journal->j_average_commit_time))
 950                 journal->j_average_commit_time = (commit_time*3 +
 951                                 journal->j_average_commit_time) / 4;
 952         else
 953                 journal->j_average_commit_time = commit_time;
 954
 955         spin_unlock(&journal->j_state_lock);
 956
 957         if (commit_transaction->t_checkpoint_list == NULL &&
 958             commit_transaction->t_checkpoint_io_list == NULL) {
 959                 __journal_drop_transaction(journal, commit_transaction);
 960         } else {
 961                 if (journal->j_checkpoint_transactions == NULL) {
 962                         journal->j_checkpoint_transactions = commit_transaction;
 963                         commit_transaction->t_cpnext = commit_transaction;
 964                         commit_transaction->t_cpprev = commit_transaction;
 965                 } else {
 966                         commit_transaction->t_cpnext =
 967                                 journal->j_checkpoint_transactions;
 968                         commit_transaction->t_cpprev =
 969                                 commit_transaction->t_cpnext->t_cpprev;
 970                         commit_transaction->t_cpnext->t_cpprev =
 971                                 commit_transaction;
 972                         commit_transaction->t_cpprev->t_cpnext =
 973                                 commit_transaction;
 974                 }
 975         }
 976         spin_unlock(&journal->j_list_lock);
 977
 978         trace_jbd_end_commit(journal, commit_transaction);
 979         jbd_debug(1, "JBD: commit %d complete, head %d\n",
 980                   journal->j_commit_sequence, journal->j_tail_sequence);
 981
 982         wake_up(&journal->j_wait_done_commit);
 983 }