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
  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 successfully 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 journal->j_list_lock.  The caller provided us with a ref
  50  * 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 (!trylock_page(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  * Decrement reference counter for data buffer. If it has been marked
  83  * 'BH_Freed', release it and the page to which it belongs if possible.
  84  */
  85 static void release_data_buffer(struct buffer_head *bh)
  86 {
  87         if (buffer_freed(bh)) {
  88                 clear_buffer_freed(bh);
  89                 release_buffer_page(bh);
  90         } else
  91                 put_bh(bh);
  92 }
  93
  94 /*
  95  * Try to acquire jbd_lock_bh_state() against the buffer, when j_list_lock is
  96  * held.  For ranking reasons we must trylock.  If we lose, schedule away and
  97  * return 0.  j_list_lock is dropped in this case.
  98  */
  99 static int inverted_lock(journal_t *journal, struct buffer_head *bh)
 100 {
 101         if (!jbd_trylock_bh_state(bh)) {
 102                 spin_unlock(&journal->j_list_lock);
 103                 schedule();
 104                 return 0;
 105         }
 106         return 1;
 107 }
 108
 109 /* Done it all: now write the commit record.  We should have
 110  * cleaned up our previous buffers by now, so if we are in abort
 111  * mode we can now just skip the rest of the journal write
 112  * entirely.
 113  *
 114  * Returns 1 if the journal needs to be aborted or 0 on success
 115  */
 116 static int journal_write_commit_record(journal_t *journal,
 117                                         transaction_t *commit_transaction)
 118 {
 119         struct journal_head *descriptor;
 120         struct buffer_head *bh;
 121         journal_header_t *header;
 122         int ret;
 123
 124         if (is_journal_aborted(journal))
 125                 return 0;
 126
 127         descriptor = journal_get_descriptor_buffer(journal);
 128         if (!descriptor)
 129                 return 1;
 130
 131         bh = jh2bh(descriptor);
 132
 133         header = (journal_header_t *)(bh->b_data);
 134         header->h_magic = cpu_to_be32(JFS_MAGIC_NUMBER);
 135         header->h_blocktype = cpu_to_be32(JFS_COMMIT_BLOCK);
 136         header->h_sequence = cpu_to_be32(commit_transaction->t_tid);
 137
 138         JBUFFER_TRACE(descriptor, "write commit block");
 139         set_buffer_dirty(bh);
 140
 141         if (journal->j_flags & JFS_BARRIER)
 142                 ret = __sync_dirty_buffer(bh, WRITE_SYNC | WRITE_FLUSH_FUA);
 143         else
 144                 ret = sync_dirty_buffer(bh);
 145
 146         put_bh(bh);             /* One for getblk() */
 147         journal_put_journal_head(descriptor);
 148
 149         return (ret == -EIO);
 150 }
 151
 152 static void journal_do_submit_data(struct buffer_head **wbuf, int bufs,
 153                                    int write_op)
 154 {
 155         int i;
 156
 157         for (i = 0; i < bufs; i++) {
 158                 wbuf[i]->b_end_io = end_buffer_write_sync;
 159                 /* We use-up our safety reference in submit_bh() */
 160                 submit_bh(write_op, wbuf[i]);
 161         }
 162 }
 163
 164 /*
 165  *  Submit all the data buffers to disk
 166  */
 167 static int journal_submit_data_buffers(journal_t *journal,
 168                                        transaction_t *commit_transaction,
 169                                        int write_op)
 170 {
 171         struct journal_head *jh;
 172         struct buffer_head *bh;
 173         int locked;
 174         int bufs = 0;
 175         struct buffer_head **wbuf = journal->j_wbuf;
 176         int err = 0;
 177
 178         /*
 179          * Whenever we unlock the journal and sleep, things can get added
 180          * onto ->t_sync_datalist, so we have to keep looping back to
 181          * write_out_data until we *know* that the list is empty.
 182          *
 183          * Cleanup any flushed data buffers from the data list.  Even in
 184          * abort mode, we want to flush this out as soon as possible.
 185          */
 186 write_out_data:
 187         cond_resched();
 188         spin_lock(&journal->j_list_lock);
 189
 190         while (commit_transaction->t_sync_datalist) {
 191                 jh = commit_transaction->t_sync_datalist;
 192                 bh = jh2bh(jh);
 193                 locked = 0;
 194
 195                 /* Get reference just to make sure buffer does not disappear
 196                  * when we are forced to drop various locks */
 197                 get_bh(bh);
 198                 /* If the buffer is dirty, we need to submit IO and hence
 199                  * we need the buffer lock. We try to lock the buffer without
 200                  * blocking. If we fail, we need to drop j_list_lock and do
 201                  * blocking lock_buffer().
 202                  */
 203                 if (buffer_dirty(bh)) {
 204                         if (!trylock_buffer(bh)) {
 205                                 BUFFER_TRACE(bh, "needs blocking lock");
 206                                 spin_unlock(&journal->j_list_lock);
 207                                 /* Write out all data to prevent deadlocks */
 208                                 journal_do_submit_data(wbuf, bufs, write_op);
 209                                 bufs = 0;
 210                                 lock_buffer(bh);
 211                                 spin_lock(&journal->j_list_lock);
 212                         }
 213                         locked = 1;
 214                 }
 215                 /* We have to get bh_state lock. Again out of order, sigh. */
 216                 if (!inverted_lock(journal, bh)) {
 217                         jbd_lock_bh_state(bh);
 218                         spin_lock(&journal->j_list_lock);
 219                 }
 220                 /* Someone already cleaned up the buffer? */
 221                 if (!buffer_jbd(bh) || bh2jh(bh) != jh
 222                         || jh->b_transaction != commit_transaction
 223                         || jh->b_jlist != BJ_SyncData) {
 224                         jbd_unlock_bh_state(bh);
 225                         if (locked)
 226                                 unlock_buffer(bh);
 227                         BUFFER_TRACE(bh, "already cleaned up");
 228                         release_data_buffer(bh);
 229                         continue;
 230                 }
 231                 if (locked && test_clear_buffer_dirty(bh)) {
 232                         BUFFER_TRACE(bh, "needs writeout, adding to array");
 233                         wbuf[bufs++] = bh;
 234                         __journal_file_buffer(jh, commit_transaction,
 235                                                 BJ_Locked);
 236                         jbd_unlock_bh_state(bh);
 237                         if (bufs == journal->j_wbufsize) {
 238                                 spin_unlock(&journal->j_list_lock);
 239                                 journal_do_submit_data(wbuf, bufs, write_op);
 240                                 bufs = 0;
 241                                 goto write_out_data;
 242                         }
 243                 } else if (!locked && buffer_locked(bh)) {
 244                         __journal_file_buffer(jh, commit_transaction,
 245                                                 BJ_Locked);
 246                         jbd_unlock_bh_state(bh);
 247                         put_bh(bh);
 248                 } else {
 249                         BUFFER_TRACE(bh, "writeout complete: unfile");
 250                         if (unlikely(!buffer_uptodate(bh)))
 251                                 err = -EIO;
 252                         __journal_unfile_buffer(jh);
 253                         jbd_unlock_bh_state(bh);
 254                         if (locked)
 255                                 unlock_buffer(bh);
 256                         journal_remove_journal_head(bh);
 257                         /* One for our safety reference, other for
 258                          * journal_remove_journal_head() */
 259                         put_bh(bh);
 260                         release_data_buffer(bh);
 261                 }
 262
 263                 if (need_resched() || spin_needbreak(&journal->j_list_lock)) {
 264                         spin_unlock(&journal->j_list_lock);
 265                         goto write_out_data;
 266                 }
 267         }
 268         spin_unlock(&journal->j_list_lock);
 269         journal_do_submit_data(wbuf, bufs, write_op);
 270
 271         return err;
 272 }
 273
 274 /*
 275  * journal_commit_transaction
 276  *
 277  * The primary function for committing a transaction to the log.  This
 278  * function is called by the journal thread to begin a complete commit.
 279  */
 280 void journal_commit_transaction(journal_t *journal)
 281 {
 282         transaction_t *commit_transaction;
 283         struct journal_head *jh, *new_jh, *descriptor;
 284         struct buffer_head **wbuf = journal->j_wbuf;
 285         int bufs;
 286         int flags;
 287         int err;
 288         unsigned int blocknr;
 289         ktime_t start_time;
 290         u64 commit_time;
 291         char *tagp = NULL;
 292         journal_header_t *header;
 293         journal_block_tag_t *tag = NULL;
 294         int space_left = 0;
 295         int first_tag = 0;
 296         int tag_flag;
 297         int i;
 298         struct blk_plug plug;
 299
 300         /*
 301          * First job: lock down the current transaction and wait for
 302          * all outstanding updates to complete.
 303          */
 304
 305 #ifdef COMMIT_STATS
 306         spin_lock(&journal->j_list_lock);
 307         summarise_journal_usage(journal);
 308         spin_unlock(&journal->j_list_lock);
 309 #endif
 310
 311         /* Do we need to erase the effects of a prior journal_flush? */
 312         if (journal->j_flags & JFS_FLUSHED) {
 313                 jbd_debug(3, "super block updated\n");
 314                 journal_update_superblock(journal, 1);
 315         } else {
 316                 jbd_debug(3, "superblock not updated\n");
 317         }
 318
 319         J_ASSERT(journal->j_running_transaction != NULL);
 320         J_ASSERT(journal->j_committing_transaction == NULL);
 321
 322         commit_transaction = journal->j_running_transaction;
 323         J_ASSERT(commit_transaction->t_state == T_RUNNING);
 324
 325         jbd_debug(1, "JBD: starting commit of transaction %d\n",
 326                         commit_transaction->t_tid);
 327
 328         spin_lock(&journal->j_state_lock);
 329         commit_transaction->t_state = T_LOCKED;
 330
 331         spin_lock(&commit_transaction->t_handle_lock);
 332         while (commit_transaction->t_updates) {
 333                 DEFINE_WAIT(wait);
 334
 335                 prepare_to_wait(&journal->j_wait_updates, &wait,
 336                                         TASK_UNINTERRUPTIBLE);
 337                 if (commit_transaction->t_updates) {
 338                         spin_unlock(&commit_transaction->t_handle_lock);
 339                         spin_unlock(&journal->j_state_lock);
 340                         schedule();
 341                         spin_lock(&journal->j_state_lock);
 342                         spin_lock(&commit_transaction->t_handle_lock);
 343                 }
 344                 finish_wait(&journal->j_wait_updates, &wait);
 345         }
 346         spin_unlock(&commit_transaction->t_handle_lock);
 347
 348         J_ASSERT (commit_transaction->t_outstanding_credits <=
 349                         journal->j_max_transaction_buffers);
 350
 351         /*
 352          * First thing we are allowed to do is to discard any remaining
 353          * BJ_Reserved buffers.  Note, it is _not_ permissible to assume
 354          * that there are no such buffers: if a large filesystem
 355          * operation like a truncate needs to split itself over multiple
 356          * transactions, then it may try to do a journal_restart() while
 357          * there are still BJ_Reserved buffers outstanding.  These must
 358          * be released cleanly from the current transaction.
 359          *
 360          * In this case, the filesystem must still reserve write access
 361          * again before modifying the buffer in the new transaction, but
 362          * we do not require it to remember exactly which old buffers it
 363          * has reserved.  This is consistent with the existing behaviour
 364          * that multiple journal_get_write_access() calls to the same
 365          * buffer are perfectly permissible.
 366          */
 367         while (commit_transaction->t_reserved_list) {
 368                 jh = commit_transaction->t_reserved_list;
 369                 JBUFFER_TRACE(jh, "reserved, unused: refile");
 370                 /*
 371                  * A journal_get_undo_access()+journal_release_buffer() may
 372                  * leave undo-committed data.
 373                  */
 374                 if (jh->b_committed_data) {
 375                         struct buffer_head *bh = jh2bh(jh);
 376
 377                         jbd_lock_bh_state(bh);
 378                         jbd_free(jh->b_committed_data, bh->b_size);
 379                         jh->b_committed_data = NULL;
 380                         jbd_unlock_bh_state(bh);
 381                 }
 382                 journal_refile_buffer(journal, jh);
 383         }
 384
 385         /*
 386          * Now try to drop any written-back buffers from the journal's
 387          * checkpoint lists.  We do this *before* commit because it potentially
 388          * frees some memory
 389          */
 390         spin_lock(&journal->j_list_lock);
 391         __journal_clean_checkpoint_list(journal);
 392         spin_unlock(&journal->j_list_lock);
 393
 394         jbd_debug (3, "JBD: commit phase 1\n");
 395
 396         /*
 397          * Switch to a new revoke table.
 398          */
 399         journal_switch_revoke_table(journal);
 400
 401         commit_transaction->t_state = T_FLUSH;
 402         journal->j_committing_transaction = commit_transaction;
 403         journal->j_running_transaction = NULL;
 404         start_time = ktime_get();
 405         commit_transaction->t_log_start = journal->j_head;
 406         wake_up(&journal->j_wait_transaction_locked);
 407         spin_unlock(&journal->j_state_lock);
 408
 409         jbd_debug (3, "JBD: commit phase 2\n");
 410
 411         /*
 412          * Now start flushing things to disk, in the order they appear
 413          * on the transaction lists.  Data blocks go first.
 414          */
 415         blk_start_plug(&plug);
 416         err = journal_submit_data_buffers(journal, commit_transaction,
 417                                           WRITE_SYNC);
 418         blk_finish_plug(&plug);
 419
 420         /*
 421          * Wait for all previously submitted IO to complete.
 422          */
 423         spin_lock(&journal->j_list_lock);
 424         while (commit_transaction->t_locked_list) {
 425                 struct buffer_head *bh;
 426
 427                 jh = commit_transaction->t_locked_list->b_tprev;
 428                 bh = jh2bh(jh);
 429                 get_bh(bh);
 430                 if (buffer_locked(bh)) {
 431                         spin_unlock(&journal->j_list_lock);
 432                         wait_on_buffer(bh);
 433                         spin_lock(&journal->j_list_lock);
 434                 }
 435                 if (unlikely(!buffer_uptodate(bh))) {
 436                         if (!trylock_page(bh->b_page)) {
 437                                 spin_unlock(&journal->j_list_lock);
 438                                 lock_page(bh->b_page);
 439                                 spin_lock(&journal->j_list_lock);
 440                         }
 441                         if (bh->b_page->mapping)
 442                                 set_bit(AS_EIO, &bh->b_page->mapping->flags);
 443
 444                         unlock_page(bh->b_page);
 445                         SetPageError(bh->b_page);
 446                         err = -EIO;
 447                 }
 448                 if (!inverted_lock(journal, bh)) {
 449                         put_bh(bh);
 450                         spin_lock(&journal->j_list_lock);
 451                         continue;
 452                 }
 453                 if (buffer_jbd(bh) && bh2jh(bh) == jh &&
 454                     jh->b_transaction == commit_transaction &&
 455                     jh->b_jlist == BJ_Locked) {
 456                         __journal_unfile_buffer(jh);
 457                         jbd_unlock_bh_state(bh);
 458                         journal_remove_journal_head(bh);
 459                         put_bh(bh);
 460                 } else {
 461                         jbd_unlock_bh_state(bh);
 462                 }
 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         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                 /* Wake up any transactions which were waiting for this
 726                    IO to complete */
 727                 wake_up_bit(&bh->b_state, BH_Unshadow);
 728                 JBUFFER_TRACE(jh, "brelse shadowed buffer");
 729                 __brelse(bh);
 730         }
 731
 732         J_ASSERT (commit_transaction->t_shadow_list == NULL);
 733
 734         jbd_debug(3, "JBD: commit phase 5\n");
 735
 736         /* Here we wait for the revoke record and descriptor record buffers */
 737  wait_for_ctlbuf:
 738         while (commit_transaction->t_log_list != NULL) {
 739                 struct buffer_head *bh;
 740
 741                 jh = commit_transaction->t_log_list->b_tprev;
 742                 bh = jh2bh(jh);
 743                 if (buffer_locked(bh)) {
 744                         wait_on_buffer(bh);
 745                         goto wait_for_ctlbuf;
 746                 }
 747                 if (cond_resched())
 748                         goto wait_for_ctlbuf;
 749
 750                 if (unlikely(!buffer_uptodate(bh)))
 751                         err = -EIO;
 752
 753                 BUFFER_TRACE(bh, "ph5: control buffer writeout done: unfile");
 754                 clear_buffer_jwrite(bh);
 755                 journal_unfile_buffer(journal, jh);
 756                 journal_put_journal_head(jh);
 757                 __brelse(bh);           /* One for getblk */
 758                 /* AKPM: bforget here */
 759         }
 760
 761         if (err)
 762                 journal_abort(journal, err);
 763
 764         jbd_debug(3, "JBD: commit phase 6\n");
 765
 766         /* All metadata is written, now write commit record and do cleanup */
 767         spin_lock(&journal->j_state_lock);
 768         J_ASSERT(commit_transaction->t_state == T_COMMIT);
 769         commit_transaction->t_state = T_COMMIT_RECORD;
 770         spin_unlock(&journal->j_state_lock);
 771
 772         if (journal_write_commit_record(journal, commit_transaction))
 773                 err = -EIO;
 774
 775         if (err)
 776                 journal_abort(journal, err);
 777
 778         /* End of a transaction!  Finally, we can do checkpoint
 779            processing: any buffers committed as a result of this
 780            transaction can be removed from any checkpoint list it was on
 781            before. */
 782
 783         jbd_debug(3, "JBD: commit phase 7\n");
 784
 785         J_ASSERT(commit_transaction->t_sync_datalist == NULL);
 786         J_ASSERT(commit_transaction->t_buffers == NULL);
 787         J_ASSERT(commit_transaction->t_checkpoint_list == NULL);
 788         J_ASSERT(commit_transaction->t_iobuf_list == NULL);
 789         J_ASSERT(commit_transaction->t_shadow_list == NULL);
 790         J_ASSERT(commit_transaction->t_log_list == NULL);
 791
 792 restart_loop:
 793         /*
 794          * As there are other places (journal_unmap_buffer()) adding buffers
 795          * to this list we have to be careful and hold the j_list_lock.
 796          */
 797         spin_lock(&journal->j_list_lock);
 798         while (commit_transaction->t_forget) {
 799                 transaction_t *cp_transaction;
 800                 struct buffer_head *bh;
 801
 802                 jh = commit_transaction->t_forget;
 803                 spin_unlock(&journal->j_list_lock);
 804                 bh = jh2bh(jh);
 805                 jbd_lock_bh_state(bh);
 806                 J_ASSERT_JH(jh, jh->b_transaction == commit_transaction ||
 807                         jh->b_transaction == journal->j_running_transaction);
 808
 809                 /*
 810                  * If there is undo-protected committed data against
 811                  * this buffer, then we can remove it now.  If it is a
 812                  * buffer needing such protection, the old frozen_data
 813                  * field now points to a committed version of the
 814                  * buffer, so rotate that field to the new committed
 815                  * data.
 816                  *
 817                  * Otherwise, we can just throw away the frozen data now.
 818                  */
 819                 if (jh->b_committed_data) {
 820                         jbd_free(jh->b_committed_data, bh->b_size);
 821                         jh->b_committed_data = NULL;
 822                         if (jh->b_frozen_data) {
 823                                 jh->b_committed_data = jh->b_frozen_data;
 824                                 jh->b_frozen_data = NULL;
 825                         }
 826                 } else if (jh->b_frozen_data) {
 827                         jbd_free(jh->b_frozen_data, bh->b_size);
 828                         jh->b_frozen_data = NULL;
 829                 }
 830
 831                 spin_lock(&journal->j_list_lock);
 832                 cp_transaction = jh->b_cp_transaction;
 833                 if (cp_transaction) {
 834                         JBUFFER_TRACE(jh, "remove from old cp transaction");
 835                         __journal_remove_checkpoint(jh);
 836                 }
 837
 838                 /* Only re-checkpoint the buffer_head if it is marked
 839                  * dirty.  If the buffer was added to the BJ_Forget list
 840                  * by journal_forget, it may no longer be dirty and
 841                  * there's no point in keeping a checkpoint record for
 842                  * it. */
 843
 844                 /* A buffer which has been freed while still being
 845                  * journaled by a previous transaction may end up still
 846                  * being dirty here, but we want to avoid writing back
 847                  * that buffer in the future after the "add to orphan"
 848                  * operation been committed,  That's not only a performance
 849                  * gain, it also stops aliasing problems if the buffer is
 850                  * left behind for writeback and gets reallocated for another
 851                  * use in a different page. */
 852                 if (buffer_freed(bh) && !jh->b_next_transaction) {
 853                         clear_buffer_freed(bh);
 854                         clear_buffer_jbddirty(bh);
 855                 }
 856
 857                 if (buffer_jbddirty(bh)) {
 858                         JBUFFER_TRACE(jh, "add to new checkpointing trans");
 859                         __journal_insert_checkpoint(jh, commit_transaction);
 860                         if (is_journal_aborted(journal))
 861                                 clear_buffer_jbddirty(bh);
 862                         JBUFFER_TRACE(jh, "refile for checkpoint writeback");
 863                         __journal_refile_buffer(jh);
 864                         jbd_unlock_bh_state(bh);
 865                 } else {
 866                         J_ASSERT_BH(bh, !buffer_dirty(bh));
 867                         /* The buffer on BJ_Forget list and not jbddirty means
 868                          * it has been freed by this transaction and hence it
 869                          * could not have been reallocated until this
 870                          * transaction has committed. *BUT* it could be
 871                          * reallocated once we have written all the data to
 872                          * disk and before we process the buffer on BJ_Forget
 873                          * list. */
 874                         JBUFFER_TRACE(jh, "refile or unfile freed buffer");
 875                         __journal_refile_buffer(jh);
 876                         if (!jh->b_transaction) {
 877                                 jbd_unlock_bh_state(bh);
 878                                  /* needs a brelse */
 879                                 journal_remove_journal_head(bh);
 880                                 release_buffer_page(bh);
 881                         } else
 882                                 jbd_unlock_bh_state(bh);
 883                 }
 884                 cond_resched_lock(&journal->j_list_lock);
 885         }
 886         spin_unlock(&journal->j_list_lock);
 887         /*
 888          * This is a bit sleazy.  We use j_list_lock to protect transition
 889          * of a transaction into T_FINISHED state and calling
 890          * __journal_drop_transaction(). Otherwise we could race with
 891          * other checkpointing code processing the transaction...
 892          */
 893         spin_lock(&journal->j_state_lock);
 894         spin_lock(&journal->j_list_lock);
 895         /*
 896          * Now recheck if some buffers did not get attached to the transaction
 897          * while the lock was dropped...
 898          */
 899         if (commit_transaction->t_forget) {
 900                 spin_unlock(&journal->j_list_lock);
 901                 spin_unlock(&journal->j_state_lock);
 902                 goto restart_loop;
 903         }
 904
 905         /* Done with this transaction! */
 906
 907         jbd_debug(3, "JBD: commit phase 8\n");
 908
 909         J_ASSERT(commit_transaction->t_state == T_COMMIT_RECORD);
 910
 911         commit_transaction->t_state = T_FINISHED;
 912         J_ASSERT(commit_transaction == journal->j_committing_transaction);
 913         journal->j_commit_sequence = commit_transaction->t_tid;
 914         journal->j_committing_transaction = NULL;
 915         commit_time = ktime_to_ns(ktime_sub(ktime_get(), start_time));
 916
 917         /*
 918          * weight the commit time higher than the average time so we don't
 919          * react too strongly to vast changes in commit time
 920          */
 921         if (likely(journal->j_average_commit_time))
 922                 journal->j_average_commit_time = (commit_time*3 +
 923                                 journal->j_average_commit_time) / 4;
 924         else
 925                 journal->j_average_commit_time = commit_time;
 926
 927         spin_unlock(&journal->j_state_lock);
 928
 929         if (commit_transaction->t_checkpoint_list == NULL &&
 930             commit_transaction->t_checkpoint_io_list == NULL) {
 931                 __journal_drop_transaction(journal, commit_transaction);
 932         } else {
 933                 if (journal->j_checkpoint_transactions == NULL) {
 934                         journal->j_checkpoint_transactions = commit_transaction;
 935                         commit_transaction->t_cpnext = commit_transaction;
 936                         commit_transaction->t_cpprev = commit_transaction;
 937                 } else {
 938                         commit_transaction->t_cpnext =
 939                                 journal->j_checkpoint_transactions;
 940                         commit_transaction->t_cpprev =
 941                                 commit_transaction->t_cpnext->t_cpprev;
 942                         commit_transaction->t_cpnext->t_cpprev =
 943                                 commit_transaction;
 944                         commit_transaction->t_cpprev->t_cpnext =
 945                                 commit_transaction;
 946                 }
 947         }
 948         spin_unlock(&journal->j_list_lock);
 949
 950         jbd_debug(1, "JBD: commit %d complete, head %d\n",
 951                   journal->j_commit_sequence, journal->j_tail_sequence);
 952
 953         wake_up(&journal->j_wait_done_commit);
 954 }