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         /* Do we need to erase the effects of a prior journal_flush? */
 306         if (journal->j_flags & JFS_FLUSHED) {
 307                 jbd_debug(3, "super block updated\n");
 308                 journal_update_superblock(journal, 1);
 309         } else {
 310                 jbd_debug(3, "superblock not updated\n");
 311         }
 312
 313         J_ASSERT(journal->j_running_transaction != NULL);
 314         J_ASSERT(journal->j_committing_transaction == NULL);
 315
 316         commit_transaction = journal->j_running_transaction;
 317         J_ASSERT(commit_transaction->t_state == T_RUNNING);
 318
 319         jbd_debug(1, "JBD: starting commit of transaction %d\n",
 320                         commit_transaction->t_tid);
 321
 322         spin_lock(&journal->j_state_lock);
 323         commit_transaction->t_state = T_LOCKED;
 324
 325         spin_lock(&commit_transaction->t_handle_lock);
 326         while (commit_transaction->t_updates) {
 327                 DEFINE_WAIT(wait);
 328
 329                 prepare_to_wait(&journal->j_wait_updates, &wait,
 330                                         TASK_UNINTERRUPTIBLE);
 331                 if (commit_transaction->t_updates) {
 332                         spin_unlock(&commit_transaction->t_handle_lock);
 333                         spin_unlock(&journal->j_state_lock);
 334                         schedule();
 335                         spin_lock(&journal->j_state_lock);
 336                         spin_lock(&commit_transaction->t_handle_lock);
 337                 }
 338                 finish_wait(&journal->j_wait_updates, &wait);
 339         }
 340         spin_unlock(&commit_transaction->t_handle_lock);
 341
 342         J_ASSERT (commit_transaction->t_outstanding_credits <=
 343                         journal->j_max_transaction_buffers);
 344
 345         /*
 346          * First thing we are allowed to do is to discard any remaining
 347          * BJ_Reserved buffers.  Note, it is _not_ permissible to assume
 348          * that there are no such buffers: if a large filesystem
 349          * operation like a truncate needs to split itself over multiple
 350          * transactions, then it may try to do a journal_restart() while
 351          * there are still BJ_Reserved buffers outstanding.  These must
 352          * be released cleanly from the current transaction.
 353          *
 354          * In this case, the filesystem must still reserve write access
 355          * again before modifying the buffer in the new transaction, but
 356          * we do not require it to remember exactly which old buffers it
 357          * has reserved.  This is consistent with the existing behaviour
 358          * that multiple journal_get_write_access() calls to the same
 359          * buffer are perfectly permissible.
 360          */
 361         while (commit_transaction->t_reserved_list) {
 362                 jh = commit_transaction->t_reserved_list;
 363                 JBUFFER_TRACE(jh, "reserved, unused: refile");
 364                 /*
 365                  * A journal_get_undo_access()+journal_release_buffer() may
 366                  * leave undo-committed data.
 367                  */
 368                 if (jh->b_committed_data) {
 369                         struct buffer_head *bh = jh2bh(jh);
 370
 371                         jbd_lock_bh_state(bh);
 372                         jbd_free(jh->b_committed_data, bh->b_size);
 373                         jh->b_committed_data = NULL;
 374                         jbd_unlock_bh_state(bh);
 375                 }
 376                 journal_refile_buffer(journal, jh);
 377         }
 378
 379         /*
 380          * Now try to drop any written-back buffers from the journal's
 381          * checkpoint lists.  We do this *before* commit because it potentially
 382          * frees some memory
 383          */
 384         spin_lock(&journal->j_list_lock);
 385         __journal_clean_checkpoint_list(journal);
 386         spin_unlock(&journal->j_list_lock);
 387
 388         jbd_debug (3, "JBD: commit phase 1\n");
 389
 390         /*
 391          * Switch to a new revoke table.
 392          */
 393         journal_switch_revoke_table(journal);
 394
 395         commit_transaction->t_state = T_FLUSH;
 396         journal->j_committing_transaction = commit_transaction;
 397         journal->j_running_transaction = NULL;
 398         start_time = ktime_get();
 399         commit_transaction->t_log_start = journal->j_head;
 400         wake_up(&journal->j_wait_transaction_locked);
 401         spin_unlock(&journal->j_state_lock);
 402
 403         jbd_debug (3, "JBD: commit phase 2\n");
 404
 405         /*
 406          * Now start flushing things to disk, in the order they appear
 407          * on the transaction lists.  Data blocks go first.
 408          */
 409         blk_start_plug(&plug);
 410         err = journal_submit_data_buffers(journal, commit_transaction,
 411                                           WRITE_SYNC);
 412         blk_finish_plug(&plug);
 413
 414         /*
 415          * Wait for all previously submitted IO to complete.
 416          */
 417         spin_lock(&journal->j_list_lock);
 418         while (commit_transaction->t_locked_list) {
 419                 struct buffer_head *bh;
 420
 421                 jh = commit_transaction->t_locked_list->b_tprev;
 422                 bh = jh2bh(jh);
 423                 get_bh(bh);
 424                 if (buffer_locked(bh)) {
 425                         spin_unlock(&journal->j_list_lock);
 426                         wait_on_buffer(bh);
 427                         spin_lock(&journal->j_list_lock);
 428                 }
 429                 if (unlikely(!buffer_uptodate(bh))) {
 430                         if (!trylock_page(bh->b_page)) {
 431                                 spin_unlock(&journal->j_list_lock);
 432                                 lock_page(bh->b_page);
 433                                 spin_lock(&journal->j_list_lock);
 434                         }
 435                         if (bh->b_page->mapping)
 436                                 set_bit(AS_EIO, &bh->b_page->mapping->flags);
 437
 438                         unlock_page(bh->b_page);
 439                         SetPageError(bh->b_page);
 440                         err = -EIO;
 441                 }
 442                 if (!inverted_lock(journal, bh)) {
 443                         put_bh(bh);
 444                         spin_lock(&journal->j_list_lock);
 445                         continue;
 446                 }
 447                 if (buffer_jbd(bh) && bh2jh(bh) == jh &&
 448                     jh->b_transaction == commit_transaction &&
 449                     jh->b_jlist == BJ_Locked) {
 450                         __journal_unfile_buffer(jh);
 451                         jbd_unlock_bh_state(bh);
 452                         journal_remove_journal_head(bh);
 453                         put_bh(bh);
 454                 } else {
 455                         jbd_unlock_bh_state(bh);
 456                 }
 457                 release_data_buffer(bh);
 458                 cond_resched_lock(&journal->j_list_lock);
 459         }
 460         spin_unlock(&journal->j_list_lock);
 461
 462         if (err) {
 463                 char b[BDEVNAME_SIZE];
 464
 465                 printk(KERN_WARNING
 466                         "JBD: Detected IO errors while flushing file data "
 467                         "on %s\n", bdevname(journal->j_fs_dev, b));
 468                 if (journal->j_flags & JFS_ABORT_ON_SYNCDATA_ERR)
 469                         journal_abort(journal, err);
 470                 err = 0;
 471         }
 472
 473         blk_start_plug(&plug);
 474
 475         journal_write_revoke_records(journal, commit_transaction, WRITE_SYNC);
 476
 477         /*
 478          * If we found any dirty or locked buffers, then we should have
 479          * looped back up to the write_out_data label.  If there weren't
 480          * any then journal_clean_data_list should have wiped the list
 481          * clean by now, so check that it is in fact empty.
 482          */
 483         J_ASSERT (commit_transaction->t_sync_datalist == NULL);
 484
 485         jbd_debug (3, "JBD: commit phase 3\n");
 486
 487         /*
 488          * Way to go: we have now written out all of the data for a
 489          * transaction!  Now comes the tricky part: we need to write out
 490          * metadata.  Loop over the transaction's entire buffer list:
 491          */
 492         spin_lock(&journal->j_state_lock);
 493         commit_transaction->t_state = T_COMMIT;
 494         spin_unlock(&journal->j_state_lock);
 495
 496         J_ASSERT(commit_transaction->t_nr_buffers <=
 497                  commit_transaction->t_outstanding_credits);
 498
 499         descriptor = NULL;
 500         bufs = 0;
 501         while (commit_transaction->t_buffers) {
 502
 503                 /* Find the next buffer to be journaled... */
 504
 505                 jh = commit_transaction->t_buffers;
 506
 507                 /* If we're in abort mode, we just un-journal the buffer and
 508                    release it. */
 509
 510                 if (is_journal_aborted(journal)) {
 511                         clear_buffer_jbddirty(jh2bh(jh));
 512                         JBUFFER_TRACE(jh, "journal is aborting: refile");
 513                         journal_refile_buffer(journal, jh);
 514                         /* If that was the last one, we need to clean up
 515                          * any descriptor buffers which may have been
 516                          * already allocated, even if we are now
 517                          * aborting. */
 518                         if (!commit_transaction->t_buffers)
 519                                 goto start_journal_io;
 520                         continue;
 521                 }
 522
 523                 /* Make sure we have a descriptor block in which to
 524                    record the metadata buffer. */
 525
 526                 if (!descriptor) {
 527                         struct buffer_head *bh;
 528
 529                         J_ASSERT (bufs == 0);
 530
 531                         jbd_debug(4, "JBD: get descriptor\n");
 532
 533                         descriptor = journal_get_descriptor_buffer(journal);
 534                         if (!descriptor) {
 535                                 journal_abort(journal, -EIO);
 536                                 continue;
 537                         }
 538
 539                         bh = jh2bh(descriptor);
 540                         jbd_debug(4, "JBD: got buffer %llu (%p)\n",
 541                                 (unsigned long long)bh->b_blocknr, bh->b_data);
 542                         header = (journal_header_t *)&bh->b_data[0];
 543                         header->h_magic     = cpu_to_be32(JFS_MAGIC_NUMBER);
 544                         header->h_blocktype = cpu_to_be32(JFS_DESCRIPTOR_BLOCK);
 545                         header->h_sequence  = cpu_to_be32(commit_transaction->t_tid);
 546
 547                         tagp = &bh->b_data[sizeof(journal_header_t)];
 548                         space_left = bh->b_size - sizeof(journal_header_t);
 549                         first_tag = 1;
 550                         set_buffer_jwrite(bh);
 551                         set_buffer_dirty(bh);
 552                         wbuf[bufs++] = bh;
 553
 554                         /* Record it so that we can wait for IO
 555                            completion later */
 556                         BUFFER_TRACE(bh, "ph3: file as descriptor");
 557                         journal_file_buffer(descriptor, commit_transaction,
 558                                         BJ_LogCtl);
 559                 }
 560
 561                 /* Where is the buffer to be written? */
 562
 563                 err = journal_next_log_block(journal, &blocknr);
 564                 /* If the block mapping failed, just abandon the buffer
 565                    and repeat this loop: we'll fall into the
 566                    refile-on-abort condition above. */
 567                 if (err) {
 568                         journal_abort(journal, err);
 569                         continue;
 570                 }
 571
 572                 /*
 573                  * start_this_handle() uses t_outstanding_credits to determine
 574                  * the free space in the log, but this counter is changed
 575                  * by journal_next_log_block() also.
 576                  */
 577                 commit_transaction->t_outstanding_credits--;
 578
 579                 /* Bump b_count to prevent truncate from stumbling over
 580                    the shadowed buffer!  @@@ This can go if we ever get
 581                    rid of the BJ_IO/BJ_Shadow pairing of buffers. */
 582                 get_bh(jh2bh(jh));
 583
 584                 /* Make a temporary IO buffer with which to write it out
 585                    (this will requeue both the metadata buffer and the
 586                    temporary IO buffer). new_bh goes on BJ_IO*/
 587
 588                 set_buffer_jwrite(jh2bh(jh));
 589                 /*
 590                  * akpm: journal_write_metadata_buffer() sets
 591                  * new_bh->b_transaction to commit_transaction.
 592                  * We need to clean this up before we release new_bh
 593                  * (which is of type BJ_IO)
 594                  */
 595                 JBUFFER_TRACE(jh, "ph3: write metadata");
 596                 flags = journal_write_metadata_buffer(commit_transaction,
 597                                                       jh, &new_jh, blocknr);
 598                 set_buffer_jwrite(jh2bh(new_jh));
 599                 wbuf[bufs++] = jh2bh(new_jh);
 600
 601                 /* Record the new block's tag in the current descriptor
 602                    buffer */
 603
 604                 tag_flag = 0;
 605                 if (flags & 1)
 606                         tag_flag |= JFS_FLAG_ESCAPE;
 607                 if (!first_tag)
 608                         tag_flag |= JFS_FLAG_SAME_UUID;
 609
 610                 tag = (journal_block_tag_t *) tagp;
 611                 tag->t_blocknr = cpu_to_be32(jh2bh(jh)->b_blocknr);
 612                 tag->t_flags = cpu_to_be32(tag_flag);
 613                 tagp += sizeof(journal_block_tag_t);
 614                 space_left -= sizeof(journal_block_tag_t);
 615
 616                 if (first_tag) {
 617                         memcpy (tagp, journal->j_uuid, 16);
 618                         tagp += 16;
 619                         space_left -= 16;
 620                         first_tag = 0;
 621                 }
 622
 623                 /* If there's no more to do, or if the descriptor is full,
 624                    let the IO rip! */
 625
 626                 if (bufs == journal->j_wbufsize ||
 627                     commit_transaction->t_buffers == NULL ||
 628                     space_left < sizeof(journal_block_tag_t) + 16) {
 629
 630                         jbd_debug(4, "JBD: Submit %d IOs\n", bufs);
 631
 632                         /* Write an end-of-descriptor marker before
 633                            submitting the IOs.  "tag" still points to
 634                            the last tag we set up. */
 635
 636                         tag->t_flags |= cpu_to_be32(JFS_FLAG_LAST_TAG);
 637
 638 start_journal_io:
 639                         for (i = 0; i < bufs; i++) {
 640                                 struct buffer_head *bh = wbuf[i];
 641                                 lock_buffer(bh);
 642                                 clear_buffer_dirty(bh);
 643                                 set_buffer_uptodate(bh);
 644                                 bh->b_end_io = journal_end_buffer_io_sync;
 645                                 submit_bh(WRITE_SYNC, bh);
 646                         }
 647                         cond_resched();
 648
 649                         /* Force a new descriptor to be generated next
 650                            time round the loop. */
 651                         descriptor = NULL;
 652                         bufs = 0;
 653                 }
 654         }
 655
 656         blk_finish_plug(&plug);
 657
 658         /* Lo and behold: we have just managed to send a transaction to
 659            the log.  Before we can commit it, wait for the IO so far to
 660            complete.  Control buffers being written are on the
 661            transaction's t_log_list queue, and metadata buffers are on
 662            the t_iobuf_list queue.
 663
 664            Wait for the buffers in reverse order.  That way we are
 665            less likely to be woken up until all IOs have completed, and
 666            so we incur less scheduling load.
 667         */
 668
 669         jbd_debug(3, "JBD: commit phase 4\n");
 670
 671         /*
 672          * akpm: these are BJ_IO, and j_list_lock is not needed.
 673          * See __journal_try_to_free_buffer.
 674          */
 675 wait_for_iobuf:
 676         while (commit_transaction->t_iobuf_list != NULL) {
 677                 struct buffer_head *bh;
 678
 679                 jh = commit_transaction->t_iobuf_list->b_tprev;
 680                 bh = jh2bh(jh);
 681                 if (buffer_locked(bh)) {
 682                         wait_on_buffer(bh);
 683                         goto wait_for_iobuf;
 684                 }
 685                 if (cond_resched())
 686                         goto wait_for_iobuf;
 687
 688                 if (unlikely(!buffer_uptodate(bh)))
 689                         err = -EIO;
 690
 691                 clear_buffer_jwrite(bh);
 692
 693                 JBUFFER_TRACE(jh, "ph4: unfile after journal write");
 694                 journal_unfile_buffer(journal, jh);
 695
 696                 /*
 697                  * ->t_iobuf_list should contain only dummy buffer_heads
 698                  * which were created by journal_write_metadata_buffer().
 699                  */
 700                 BUFFER_TRACE(bh, "dumping temporary bh");
 701                 journal_put_journal_head(jh);
 702                 __brelse(bh);
 703                 J_ASSERT_BH(bh, atomic_read(&bh->b_count) == 0);
 704                 free_buffer_head(bh);
 705
 706                 /* We also have to unlock and free the corresponding
 707                    shadowed buffer */
 708                 jh = commit_transaction->t_shadow_list->b_tprev;
 709                 bh = jh2bh(jh);
 710                 clear_buffer_jwrite(bh);
 711                 J_ASSERT_BH(bh, buffer_jbddirty(bh));
 712
 713                 /* The metadata is now released for reuse, but we need
 714                    to remember it against this transaction so that when
 715                    we finally commit, we can do any checkpointing
 716                    required. */
 717                 JBUFFER_TRACE(jh, "file as BJ_Forget");
 718                 journal_file_buffer(jh, commit_transaction, BJ_Forget);
 719                 /*
 720                  * Wake up any transactions which were waiting for this
 721                  * IO to complete. The barrier must be here so that changes
 722                  * by journal_file_buffer() take effect before wake_up_bit()
 723                  * does the waitqueue check.
 724                  */
 725                 smp_mb();
 726                 wake_up_bit(&bh->b_state, BH_Unshadow);
 727                 JBUFFER_TRACE(jh, "brelse shadowed buffer");
 728                 __brelse(bh);
 729         }
 730
 731         J_ASSERT (commit_transaction->t_shadow_list == NULL);
 732
 733         jbd_debug(3, "JBD: commit phase 5\n");
 734
 735         /* Here we wait for the revoke record and descriptor record buffers */
 736  wait_for_ctlbuf:
 737         while (commit_transaction->t_log_list != NULL) {
 738                 struct buffer_head *bh;
 739
 740                 jh = commit_transaction->t_log_list->b_tprev;
 741                 bh = jh2bh(jh);
 742                 if (buffer_locked(bh)) {
 743                         wait_on_buffer(bh);
 744                         goto wait_for_ctlbuf;
 745                 }
 746                 if (cond_resched())
 747                         goto wait_for_ctlbuf;
 748
 749                 if (unlikely(!buffer_uptodate(bh)))
 750                         err = -EIO;
 751
 752                 BUFFER_TRACE(bh, "ph5: control buffer writeout done: unfile");
 753                 clear_buffer_jwrite(bh);
 754                 journal_unfile_buffer(journal, jh);
 755                 journal_put_journal_head(jh);
 756                 __brelse(bh);           /* One for getblk */
 757                 /* AKPM: bforget here */
 758         }
 759
 760         if (err)
 761                 journal_abort(journal, err);
 762
 763         jbd_debug(3, "JBD: commit phase 6\n");
 764
 765         /* All metadata is written, now write commit record and do cleanup */
 766         spin_lock(&journal->j_state_lock);
 767         J_ASSERT(commit_transaction->t_state == T_COMMIT);
 768         commit_transaction->t_state = T_COMMIT_RECORD;
 769         spin_unlock(&journal->j_state_lock);
 770
 771         if (journal_write_commit_record(journal, commit_transaction))
 772                 err = -EIO;
 773
 774         if (err)
 775                 journal_abort(journal, err);
 776
 777         /* End of a transaction!  Finally, we can do checkpoint
 778            processing: any buffers committed as a result of this
 779            transaction can be removed from any checkpoint list it was on
 780            before. */
 781
 782         jbd_debug(3, "JBD: commit phase 7\n");
 783
 784         J_ASSERT(commit_transaction->t_sync_datalist == NULL);
 785         J_ASSERT(commit_transaction->t_buffers == NULL);
 786         J_ASSERT(commit_transaction->t_checkpoint_list == NULL);
 787         J_ASSERT(commit_transaction->t_iobuf_list == NULL);
 788         J_ASSERT(commit_transaction->t_shadow_list == NULL);
 789         J_ASSERT(commit_transaction->t_log_list == NULL);
 790
 791 restart_loop:
 792         /*
 793          * As there are other places (journal_unmap_buffer()) adding buffers
 794          * to this list we have to be careful and hold the j_list_lock.
 795          */
 796         spin_lock(&journal->j_list_lock);
 797         while (commit_transaction->t_forget) {
 798                 transaction_t *cp_transaction;
 799                 struct buffer_head *bh;
 800
 801                 jh = commit_transaction->t_forget;
 802                 spin_unlock(&journal->j_list_lock);
 803                 bh = jh2bh(jh);
 804                 jbd_lock_bh_state(bh);
 805                 J_ASSERT_JH(jh, jh->b_transaction == commit_transaction ||
 806                         jh->b_transaction == journal->j_running_transaction);
 807
 808                 /*
 809                  * If there is undo-protected committed data against
 810                  * this buffer, then we can remove it now.  If it is a
 811                  * buffer needing such protection, the old frozen_data
 812                  * field now points to a committed version of the
 813                  * buffer, so rotate that field to the new committed
 814                  * data.
 815                  *
 816                  * Otherwise, we can just throw away the frozen data now.
 817                  */
 818                 if (jh->b_committed_data) {
 819                         jbd_free(jh->b_committed_data, bh->b_size);
 820                         jh->b_committed_data = NULL;
 821                         if (jh->b_frozen_data) {
 822                                 jh->b_committed_data = jh->b_frozen_data;
 823                                 jh->b_frozen_data = NULL;
 824                         }
 825                 } else if (jh->b_frozen_data) {
 826                         jbd_free(jh->b_frozen_data, bh->b_size);
 827                         jh->b_frozen_data = NULL;
 828                 }
 829
 830                 spin_lock(&journal->j_list_lock);
 831                 cp_transaction = jh->b_cp_transaction;
 832                 if (cp_transaction) {
 833                         JBUFFER_TRACE(jh, "remove from old cp transaction");
 834                         __journal_remove_checkpoint(jh);
 835                 }
 836
 837                 /* Only re-checkpoint the buffer_head if it is marked
 838                  * dirty.  If the buffer was added to the BJ_Forget list
 839                  * by journal_forget, it may no longer be dirty and
 840                  * there's no point in keeping a checkpoint record for
 841                  * it. */
 842
 843                 /* A buffer which has been freed while still being
 844                  * journaled by a previous transaction may end up still
 845                  * being dirty here, but we want to avoid writing back
 846                  * that buffer in the future after the "add to orphan"
 847                  * operation been committed,  That's not only a performance
 848                  * gain, it also stops aliasing problems if the buffer is
 849                  * left behind for writeback and gets reallocated for another
 850                  * use in a different page. */
 851                 if (buffer_freed(bh) && !jh->b_next_transaction) {
 852                         clear_buffer_freed(bh);
 853                         clear_buffer_jbddirty(bh);
 854                 }
 855
 856                 if (buffer_jbddirty(bh)) {
 857                         JBUFFER_TRACE(jh, "add to new checkpointing trans");
 858                         __journal_insert_checkpoint(jh, commit_transaction);
 859                         if (is_journal_aborted(journal))
 860                                 clear_buffer_jbddirty(bh);
 861                         JBUFFER_TRACE(jh, "refile for checkpoint writeback");
 862                         __journal_refile_buffer(jh);
 863                         jbd_unlock_bh_state(bh);
 864                 } else {
 865                         J_ASSERT_BH(bh, !buffer_dirty(bh));
 866                         /* The buffer on BJ_Forget list and not jbddirty means
 867                          * it has been freed by this transaction and hence it
 868                          * could not have been reallocated until this
 869                          * transaction has committed. *BUT* it could be
 870                          * reallocated once we have written all the data to
 871                          * disk and before we process the buffer on BJ_Forget
 872                          * list. */
 873                         JBUFFER_TRACE(jh, "refile or unfile freed buffer");
 874                         __journal_refile_buffer(jh);
 875                         if (!jh->b_transaction) {
 876                                 jbd_unlock_bh_state(bh);
 877                                  /* needs a brelse */
 878                                 journal_remove_journal_head(bh);
 879                                 release_buffer_page(bh);
 880                         } else
 881                                 jbd_unlock_bh_state(bh);
 882                 }
 883                 cond_resched_lock(&journal->j_list_lock);
 884         }
 885         spin_unlock(&journal->j_list_lock);
 886         /*
 887          * This is a bit sleazy.  We use j_list_lock to protect transition
 888          * of a transaction into T_FINISHED state and calling
 889          * __journal_drop_transaction(). Otherwise we could race with
 890          * other checkpointing code processing the transaction...
 891          */
 892         spin_lock(&journal->j_state_lock);
 893         spin_lock(&journal->j_list_lock);
 894         /*
 895          * Now recheck if some buffers did not get attached to the transaction
 896          * while the lock was dropped...
 897          */
 898         if (commit_transaction->t_forget) {
 899                 spin_unlock(&journal->j_list_lock);
 900                 spin_unlock(&journal->j_state_lock);
 901                 goto restart_loop;
 902         }
 903
 904         /* Done with this transaction! */
 905
 906         jbd_debug(3, "JBD: commit phase 8\n");
 907
 908         J_ASSERT(commit_transaction->t_state == T_COMMIT_RECORD);
 909
 910         commit_transaction->t_state = T_FINISHED;
 911         J_ASSERT(commit_transaction == journal->j_committing_transaction);
 912         journal->j_commit_sequence = commit_transaction->t_tid;
 913         journal->j_committing_transaction = NULL;
 914         commit_time = ktime_to_ns(ktime_sub(ktime_get(), start_time));
 915
 916         /*
 917          * weight the commit time higher than the average time so we don't
 918          * react too strongly to vast changes in commit time
 919          */
 920         if (likely(journal->j_average_commit_time))
 921                 journal->j_average_commit_time = (commit_time*3 +
 922                                 journal->j_average_commit_time) / 4;
 923         else
 924                 journal->j_average_commit_time = commit_time;
 925
 926         spin_unlock(&journal->j_state_lock);
 927
 928         if (commit_transaction->t_checkpoint_list == NULL &&
 929             commit_transaction->t_checkpoint_io_list == NULL) {
 930                 __journal_drop_transaction(journal, commit_transaction);
 931         } else {
 932                 if (journal->j_checkpoint_transactions == NULL) {
 933                         journal->j_checkpoint_transactions = commit_transaction;
 934                         commit_transaction->t_cpnext = commit_transaction;
 935                         commit_transaction->t_cpprev = commit_transaction;
 936                 } else {
 937                         commit_transaction->t_cpnext =
 938                                 journal->j_checkpoint_transactions;
 939                         commit_transaction->t_cpprev =
 940                                 commit_transaction->t_cpnext->t_cpprev;
 941                         commit_transaction->t_cpnext->t_cpprev =
 942                                 commit_transaction;
 943                         commit_transaction->t_cpprev->t_cpnext =
 944                                 commit_transaction;
 945                 }
 946         }
 947         spin_unlock(&journal->j_list_lock);
 948
 949         jbd_debug(1, "JBD: commit %d complete, head %d\n",
 950                   journal->j_commit_sequence, journal->j_tail_sequence);
 951
 952         wake_up(&journal->j_wait_done_commit);
 953 }