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
  24 /*
  25  * Default IO end handler for temporary BJ_IO buffer_heads.
  26  */
  27 static void journal_end_buffer_io_sync(struct buffer_head *bh, int uptodate)
  28 {
  29         BUFFER_TRACE(bh, "");
  30         if (uptodate)
  31                 set_buffer_uptodate(bh);
  32         else
  33                 clear_buffer_uptodate(bh);
  34         unlock_buffer(bh);
  35 }
  36
  37 /*
  38  * When an ext3-ordered file is truncated, it is possible that many pages are
  39  * not successfully freed, because they are attached to a committing transaction.
  40  * After the transaction commits, these pages are left on the LRU, with no
  41  * ->mapping, and with attached buffers.  These pages are trivially reclaimable
  42  * by the VM, but their apparent absence upsets the VM accounting, and it makes
  43  * the numbers in /proc/meminfo look odd.
  44  *
  45  * So here, we have a buffer which has just come off the forget list.  Look to
  46  * see if we can strip all buffers from the backing page.
  47  *
  48  * Called under journal->j_list_lock.  The caller provided us with a ref
  49  * against the buffer, and we drop that here.
  50  */
  51 static void release_buffer_page(struct buffer_head *bh)
  52 {
  53         struct page *page;
  54
  55         if (buffer_dirty(bh))
  56                 goto nope;
  57         if (atomic_read(&bh->b_count) != 1)
  58                 goto nope;
  59         page = bh->b_page;
  60         if (!page)
  61                 goto nope;
  62         if (page->mapping)
  63                 goto nope;
  64
  65         /* OK, it's a truncated page */
  66         if (!trylock_page(page))
  67                 goto nope;
  68
  69         page_cache_get(page);
  70         __brelse(bh);
  71         try_to_free_buffers(page);
  72         unlock_page(page);
  73         page_cache_release(page);
  74         return;
  75
  76 nope:
  77         __brelse(bh);
  78 }
  79
  80 /*
  81  * Decrement reference counter for data buffer. If it has been marked
  82  * 'BH_Freed', release it and the page to which it belongs if possible.
  83  */
  84 static void release_data_buffer(struct buffer_head *bh)
  85 {
  86         if (buffer_freed(bh)) {
  87                 clear_buffer_freed(bh);
  88                 release_buffer_page(bh);
  89         } else
  90                 put_bh(bh);
  91 }
  92
  93 /*
  94  * Try to acquire jbd_lock_bh_state() against the buffer, when j_list_lock is
  95  * held.  For ranking reasons we must trylock.  If we lose, schedule away and
  96  * return 0.  j_list_lock is dropped in this case.
  97  */
  98 static int inverted_lock(journal_t *journal, struct buffer_head *bh)
  99 {
 100         if (!jbd_trylock_bh_state(bh)) {
 101                 spin_unlock(&journal->j_list_lock);
 102                 schedule();
 103                 return 0;
 104         }
 105         return 1;
 106 }
 107
 108 /* Done it all: now write the commit record.  We should have
 109  * cleaned up our previous buffers by now, so if we are in abort
 110  * mode we can now just skip the rest of the journal write
 111  * entirely.
 112  *
 113  * Returns 1 if the journal needs to be aborted or 0 on success
 114  */
 115 static int journal_write_commit_record(journal_t *journal,
 116                                         transaction_t *commit_transaction)
 117 {
 118         struct journal_head *descriptor;
 119         struct buffer_head *bh;
 120         journal_header_t *header;
 121         int ret;
 122         int barrier_done = 0;
 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         if (journal->j_flags & JFS_BARRIER) {
 141                 set_buffer_ordered(bh);
 142                 barrier_done = 1;
 143         }
 144         ret = sync_dirty_buffer(bh);
 145         if (barrier_done)
 146                 clear_buffer_ordered(bh);
 147         /* is it possible for another commit to fail at roughly
 148          * the same time as this one?  If so, we don't want to
 149          * trust the barrier flag in the super, but instead want
 150          * to remember if we sent a barrier request
 151          */
 152         if (ret == -EOPNOTSUPP && barrier_done) {
 153                 char b[BDEVNAME_SIZE];
 154
 155                 printk(KERN_WARNING
 156                         "JBD: barrier-based sync failed on %s - "
 157                         "disabling barriers\n",
 158                         bdevname(journal->j_dev, b));
 159                 spin_lock(&journal->j_state_lock);
 160                 journal->j_flags &= ~JFS_BARRIER;
 161                 spin_unlock(&journal->j_state_lock);
 162
 163                 /* And try again, without the barrier */
 164                 set_buffer_uptodate(bh);
 165                 set_buffer_dirty(bh);
 166                 ret = sync_dirty_buffer(bh);
 167         }
 168         put_bh(bh);             /* One for getblk() */
 169         journal_put_journal_head(descriptor);
 170
 171         return (ret == -EIO);
 172 }
 173
 174 static void journal_do_submit_data(struct buffer_head **wbuf, int bufs,
 175                                    int write_op)
 176 {
 177         int i;
 178
 179         for (i = 0; i < bufs; i++) {
 180                 wbuf[i]->b_end_io = end_buffer_write_sync;
 181                 /* We use-up our safety reference in submit_bh() */
 182                 submit_bh(write_op, wbuf[i]);
 183         }
 184 }
 185
 186 /*
 187  *  Submit all the data buffers to disk
 188  */
 189 static int journal_submit_data_buffers(journal_t *journal,
 190                                        transaction_t *commit_transaction,
 191                                        int write_op)
 192 {
 193         struct journal_head *jh;
 194         struct buffer_head *bh;
 195         int locked;
 196         int bufs = 0;
 197         struct buffer_head **wbuf = journal->j_wbuf;
 198         int err = 0;
 199
 200         /*
 201          * Whenever we unlock the journal and sleep, things can get added
 202          * onto ->t_sync_datalist, so we have to keep looping back to
 203          * write_out_data until we *know* that the list is empty.
 204          *
 205          * Cleanup any flushed data buffers from the data list.  Even in
 206          * abort mode, we want to flush this out as soon as possible.
 207          */
 208 write_out_data:
 209         cond_resched();
 210         spin_lock(&journal->j_list_lock);
 211
 212         while (commit_transaction->t_sync_datalist) {
 213                 jh = commit_transaction->t_sync_datalist;
 214                 bh = jh2bh(jh);
 215                 locked = 0;
 216
 217                 /* Get reference just to make sure buffer does not disappear
 218                  * when we are forced to drop various locks */
 219                 get_bh(bh);
 220                 /* If the buffer is dirty, we need to submit IO and hence
 221                  * we need the buffer lock. We try to lock the buffer without
 222                  * blocking. If we fail, we need to drop j_list_lock and do
 223                  * blocking lock_buffer().
 224                  */
 225                 if (buffer_dirty(bh)) {
 226                         if (!trylock_buffer(bh)) {
 227                                 BUFFER_TRACE(bh, "needs blocking lock");
 228                                 spin_unlock(&journal->j_list_lock);
 229                                 /* Write out all data to prevent deadlocks */
 230                                 journal_do_submit_data(wbuf, bufs, write_op);
 231                                 bufs = 0;
 232                                 lock_buffer(bh);
 233                                 spin_lock(&journal->j_list_lock);
 234                         }
 235                         locked = 1;
 236                 }
 237                 /* We have to get bh_state lock. Again out of order, sigh. */
 238                 if (!inverted_lock(journal, bh)) {
 239                         jbd_lock_bh_state(bh);
 240                         spin_lock(&journal->j_list_lock);
 241                 }
 242                 /* Someone already cleaned up the buffer? */
 243                 if (!buffer_jbd(bh) || bh2jh(bh) != jh
 244                         || jh->b_transaction != commit_transaction
 245                         || jh->b_jlist != BJ_SyncData) {
 246                         jbd_unlock_bh_state(bh);
 247                         if (locked)
 248                                 unlock_buffer(bh);
 249                         BUFFER_TRACE(bh, "already cleaned up");
 250                         release_data_buffer(bh);
 251                         continue;
 252                 }
 253                 if (locked && test_clear_buffer_dirty(bh)) {
 254                         BUFFER_TRACE(bh, "needs writeout, adding to array");
 255                         wbuf[bufs++] = bh;
 256                         __journal_file_buffer(jh, commit_transaction,
 257                                                 BJ_Locked);
 258                         jbd_unlock_bh_state(bh);
 259                         if (bufs == journal->j_wbufsize) {
 260                                 spin_unlock(&journal->j_list_lock);
 261                                 journal_do_submit_data(wbuf, bufs, write_op);
 262                                 bufs = 0;
 263                                 goto write_out_data;
 264                         }
 265                 } else if (!locked && buffer_locked(bh)) {
 266                         __journal_file_buffer(jh, commit_transaction,
 267                                                 BJ_Locked);
 268                         jbd_unlock_bh_state(bh);
 269                         put_bh(bh);
 270                 } else {
 271                         BUFFER_TRACE(bh, "writeout complete: unfile");
 272                         if (unlikely(!buffer_uptodate(bh)))
 273                                 err = -EIO;
 274                         __journal_unfile_buffer(jh);
 275                         jbd_unlock_bh_state(bh);
 276                         if (locked)
 277                                 unlock_buffer(bh);
 278                         journal_remove_journal_head(bh);
 279                         /* One for our safety reference, other for
 280                          * journal_remove_journal_head() */
 281                         put_bh(bh);
 282                         release_data_buffer(bh);
 283                 }
 284
 285                 if (need_resched() || spin_needbreak(&journal->j_list_lock)) {
 286                         spin_unlock(&journal->j_list_lock);
 287                         goto write_out_data;
 288                 }
 289         }
 290         spin_unlock(&journal->j_list_lock);
 291         journal_do_submit_data(wbuf, bufs, write_op);
 292
 293         return err;
 294 }
 295
 296 /*
 297  * journal_commit_transaction
 298  *
 299  * The primary function for committing a transaction to the log.  This
 300  * function is called by the journal thread to begin a complete commit.
 301  */
 302 void journal_commit_transaction(journal_t *journal)
 303 {
 304         transaction_t *commit_transaction;
 305         struct journal_head *jh, *new_jh, *descriptor;
 306         struct buffer_head **wbuf = journal->j_wbuf;
 307         int bufs;
 308         int flags;
 309         int err;
 310         unsigned int blocknr;
 311         ktime_t start_time;
 312         u64 commit_time;
 313         char *tagp = NULL;
 314         journal_header_t *header;
 315         journal_block_tag_t *tag = NULL;
 316         int space_left = 0;
 317         int first_tag = 0;
 318         int tag_flag;
 319         int i;
 320         int write_op = WRITE;
 321
 322         /*
 323          * First job: lock down the current transaction and wait for
 324          * all outstanding updates to complete.
 325          */
 326
 327 #ifdef COMMIT_STATS
 328         spin_lock(&journal->j_list_lock);
 329         summarise_journal_usage(journal);
 330         spin_unlock(&journal->j_list_lock);
 331 #endif
 332
 333         /* Do we need to erase the effects of a prior journal_flush? */
 334         if (journal->j_flags & JFS_FLUSHED) {
 335                 jbd_debug(3, "super block updated\n");
 336                 journal_update_superblock(journal, 1);
 337         } else {
 338                 jbd_debug(3, "superblock not updated\n");
 339         }
 340
 341         J_ASSERT(journal->j_running_transaction != NULL);
 342         J_ASSERT(journal->j_committing_transaction == NULL);
 343
 344         commit_transaction = journal->j_running_transaction;
 345         J_ASSERT(commit_transaction->t_state == T_RUNNING);
 346
 347         jbd_debug(1, "JBD: starting commit of transaction %d\n",
 348                         commit_transaction->t_tid);
 349
 350         spin_lock(&journal->j_state_lock);
 351         commit_transaction->t_state = T_LOCKED;
 352
 353         /*
 354          * Use plugged writes here, since we want to submit several before
 355          * we unplug the device. We don't do explicit unplugging in here,
 356          * instead we rely on sync_buffer() doing the unplug for us.
 357          */
 358         if (commit_transaction->t_synchronous_commit)
 359                 write_op = WRITE_SYNC_PLUG;
 360         spin_lock(&commit_transaction->t_handle_lock);
 361         while (commit_transaction->t_updates) {
 362                 DEFINE_WAIT(wait);
 363
 364                 prepare_to_wait(&journal->j_wait_updates, &wait,
 365                                         TASK_UNINTERRUPTIBLE);
 366                 if (commit_transaction->t_updates) {
 367                         spin_unlock(&commit_transaction->t_handle_lock);
 368                         spin_unlock(&journal->j_state_lock);
 369                         schedule();
 370                         spin_lock(&journal->j_state_lock);
 371                         spin_lock(&commit_transaction->t_handle_lock);
 372                 }
 373                 finish_wait(&journal->j_wait_updates, &wait);
 374         }
 375         spin_unlock(&commit_transaction->t_handle_lock);
 376
 377         J_ASSERT (commit_transaction->t_outstanding_credits <=
 378                         journal->j_max_transaction_buffers);
 379
 380         /*
 381          * First thing we are allowed to do is to discard any remaining
 382          * BJ_Reserved buffers.  Note, it is _not_ permissible to assume
 383          * that there are no such buffers: if a large filesystem
 384          * operation like a truncate needs to split itself over multiple
 385          * transactions, then it may try to do a journal_restart() while
 386          * there are still BJ_Reserved buffers outstanding.  These must
 387          * be released cleanly from the current transaction.
 388          *
 389          * In this case, the filesystem must still reserve write access
 390          * again before modifying the buffer in the new transaction, but
 391          * we do not require it to remember exactly which old buffers it
 392          * has reserved.  This is consistent with the existing behaviour
 393          * that multiple journal_get_write_access() calls to the same
 394          * buffer are perfectly permissable.
 395          */
 396         while (commit_transaction->t_reserved_list) {
 397                 jh = commit_transaction->t_reserved_list;
 398                 JBUFFER_TRACE(jh, "reserved, unused: refile");
 399                 /*
 400                  * A journal_get_undo_access()+journal_release_buffer() may
 401                  * leave undo-committed data.
 402                  */
 403                 if (jh->b_committed_data) {
 404                         struct buffer_head *bh = jh2bh(jh);
 405
 406                         jbd_lock_bh_state(bh);
 407                         jbd_free(jh->b_committed_data, bh->b_size);
 408                         jh->b_committed_data = NULL;
 409                         jbd_unlock_bh_state(bh);
 410                 }
 411                 journal_refile_buffer(journal, jh);
 412         }
 413
 414         /*
 415          * Now try to drop any written-back buffers from the journal's
 416          * checkpoint lists.  We do this *before* commit because it potentially
 417          * frees some memory
 418          */
 419         spin_lock(&journal->j_list_lock);
 420         __journal_clean_checkpoint_list(journal);
 421         spin_unlock(&journal->j_list_lock);
 422
 423         jbd_debug (3, "JBD: commit phase 1\n");
 424
 425         /*
 426          * Switch to a new revoke table.
 427          */
 428         journal_switch_revoke_table(journal);
 429
 430         commit_transaction->t_state = T_FLUSH;
 431         journal->j_committing_transaction = commit_transaction;
 432         journal->j_running_transaction = NULL;
 433         start_time = ktime_get();
 434         commit_transaction->t_log_start = journal->j_head;
 435         wake_up(&journal->j_wait_transaction_locked);
 436         spin_unlock(&journal->j_state_lock);
 437
 438         jbd_debug (3, "JBD: commit phase 2\n");
 439
 440         /*
 441          * Now start flushing things to disk, in the order they appear
 442          * on the transaction lists.  Data blocks go first.
 443          */
 444         err = journal_submit_data_buffers(journal, commit_transaction,
 445                                           write_op);
 446
 447         /*
 448          * Wait for all previously submitted IO to complete.
 449          */
 450         spin_lock(&journal->j_list_lock);
 451         while (commit_transaction->t_locked_list) {
 452                 struct buffer_head *bh;
 453
 454                 jh = commit_transaction->t_locked_list->b_tprev;
 455                 bh = jh2bh(jh);
 456                 get_bh(bh);
 457                 if (buffer_locked(bh)) {
 458                         spin_unlock(&journal->j_list_lock);
 459                         wait_on_buffer(bh);
 460                         spin_lock(&journal->j_list_lock);
 461                 }
 462                 if (unlikely(!buffer_uptodate(bh))) {
 463                         if (!trylock_page(bh->b_page)) {
 464                                 spin_unlock(&journal->j_list_lock);
 465                                 lock_page(bh->b_page);
 466                                 spin_lock(&journal->j_list_lock);
 467                         }
 468                         if (bh->b_page->mapping)
 469                                 set_bit(AS_EIO, &bh->b_page->mapping->flags);
 470
 471                         unlock_page(bh->b_page);
 472                         SetPageError(bh->b_page);
 473                         err = -EIO;
 474                 }
 475                 if (!inverted_lock(journal, bh)) {
 476                         put_bh(bh);
 477                         spin_lock(&journal->j_list_lock);
 478                         continue;
 479                 }
 480                 if (buffer_jbd(bh) && bh2jh(bh) == jh &&
 481                     jh->b_transaction == commit_transaction &&
 482                     jh->b_jlist == BJ_Locked) {
 483                         __journal_unfile_buffer(jh);
 484                         jbd_unlock_bh_state(bh);
 485                         journal_remove_journal_head(bh);
 486                         put_bh(bh);
 487                 } else {
 488                         jbd_unlock_bh_state(bh);
 489                 }
 490                 release_data_buffer(bh);
 491                 cond_resched_lock(&journal->j_list_lock);
 492         }
 493         spin_unlock(&journal->j_list_lock);
 494
 495         if (err) {
 496                 char b[BDEVNAME_SIZE];
 497
 498                 printk(KERN_WARNING
 499                         "JBD: Detected IO errors while flushing file data "
 500                         "on %s\n", bdevname(journal->j_fs_dev, b));
 501                 if (journal->j_flags & JFS_ABORT_ON_SYNCDATA_ERR)
 502                         journal_abort(journal, err);
 503                 err = 0;
 504         }
 505
 506         journal_write_revoke_records(journal, commit_transaction, write_op);
 507
 508         /*
 509          * If we found any dirty or locked buffers, then we should have
 510          * looped back up to the write_out_data label.  If there weren't
 511          * any then journal_clean_data_list should have wiped the list
 512          * clean by now, so check that it is in fact empty.
 513          */
 514         J_ASSERT (commit_transaction->t_sync_datalist == NULL);
 515
 516         jbd_debug (3, "JBD: commit phase 3\n");
 517
 518         /*
 519          * Way to go: we have now written out all of the data for a
 520          * transaction!  Now comes the tricky part: we need to write out
 521          * metadata.  Loop over the transaction's entire buffer list:
 522          */
 523         spin_lock(&journal->j_state_lock);
 524         commit_transaction->t_state = T_COMMIT;
 525         spin_unlock(&journal->j_state_lock);
 526
 527         J_ASSERT(commit_transaction->t_nr_buffers <=
 528                  commit_transaction->t_outstanding_credits);
 529
 530         descriptor = NULL;
 531         bufs = 0;
 532         while (commit_transaction->t_buffers) {
 533
 534                 /* Find the next buffer to be journaled... */
 535
 536                 jh = commit_transaction->t_buffers;
 537
 538                 /* If we're in abort mode, we just un-journal the buffer and
 539                    release it. */
 540
 541                 if (is_journal_aborted(journal)) {
 542                         clear_buffer_jbddirty(jh2bh(jh));
 543                         JBUFFER_TRACE(jh, "journal is aborting: refile");
 544                         journal_refile_buffer(journal, jh);
 545                         /* If that was the last one, we need to clean up
 546                          * any descriptor buffers which may have been
 547                          * already allocated, even if we are now
 548                          * aborting. */
 549                         if (!commit_transaction->t_buffers)
 550                                 goto start_journal_io;
 551                         continue;
 552                 }
 553
 554                 /* Make sure we have a descriptor block in which to
 555                    record the metadata buffer. */
 556
 557                 if (!descriptor) {
 558                         struct buffer_head *bh;
 559
 560                         J_ASSERT (bufs == 0);
 561
 562                         jbd_debug(4, "JBD: get descriptor\n");
 563
 564                         descriptor = journal_get_descriptor_buffer(journal);
 565                         if (!descriptor) {
 566                                 journal_abort(journal, -EIO);
 567                                 continue;
 568                         }
 569
 570                         bh = jh2bh(descriptor);
 571                         jbd_debug(4, "JBD: got buffer %llu (%p)\n",
 572                                 (unsigned long long)bh->b_blocknr, bh->b_data);
 573                         header = (journal_header_t *)&bh->b_data[0];
 574                         header->h_magic     = cpu_to_be32(JFS_MAGIC_NUMBER);
 575                         header->h_blocktype = cpu_to_be32(JFS_DESCRIPTOR_BLOCK);
 576                         header->h_sequence  = cpu_to_be32(commit_transaction->t_tid);
 577
 578                         tagp = &bh->b_data[sizeof(journal_header_t)];
 579                         space_left = bh->b_size - sizeof(journal_header_t);
 580                         first_tag = 1;
 581                         set_buffer_jwrite(bh);
 582                         set_buffer_dirty(bh);
 583                         wbuf[bufs++] = bh;
 584
 585                         /* Record it so that we can wait for IO
 586                            completion later */
 587                         BUFFER_TRACE(bh, "ph3: file as descriptor");
 588                         journal_file_buffer(descriptor, commit_transaction,
 589                                         BJ_LogCtl);
 590                 }
 591
 592                 /* Where is the buffer to be written? */
 593
 594                 err = journal_next_log_block(journal, &blocknr);
 595                 /* If the block mapping failed, just abandon the buffer
 596                    and repeat this loop: we'll fall into the
 597                    refile-on-abort condition above. */
 598                 if (err) {
 599                         journal_abort(journal, err);
 600                         continue;
 601                 }
 602
 603                 /*
 604                  * start_this_handle() uses t_outstanding_credits to determine
 605                  * the free space in the log, but this counter is changed
 606                  * by journal_next_log_block() also.
 607                  */
 608                 commit_transaction->t_outstanding_credits--;
 609
 610                 /* Bump b_count to prevent truncate from stumbling over
 611                    the shadowed buffer!  @@@ This can go if we ever get
 612                    rid of the BJ_IO/BJ_Shadow pairing of buffers. */
 613                 atomic_inc(&jh2bh(jh)->b_count);
 614
 615                 /* Make a temporary IO buffer with which to write it out
 616                    (this will requeue both the metadata buffer and the
 617                    temporary IO buffer). new_bh goes on BJ_IO*/
 618
 619                 set_bit(BH_JWrite, &jh2bh(jh)->b_state);
 620                 /*
 621                  * akpm: journal_write_metadata_buffer() sets
 622                  * new_bh->b_transaction to commit_transaction.
 623                  * We need to clean this up before we release new_bh
 624                  * (which is of type BJ_IO)
 625                  */
 626                 JBUFFER_TRACE(jh, "ph3: write metadata");
 627                 flags = journal_write_metadata_buffer(commit_transaction,
 628                                                       jh, &new_jh, blocknr);
 629                 set_bit(BH_JWrite, &jh2bh(new_jh)->b_state);
 630                 wbuf[bufs++] = jh2bh(new_jh);
 631
 632                 /* Record the new block's tag in the current descriptor
 633                    buffer */
 634
 635                 tag_flag = 0;
 636                 if (flags & 1)
 637                         tag_flag |= JFS_FLAG_ESCAPE;
 638                 if (!first_tag)
 639                         tag_flag |= JFS_FLAG_SAME_UUID;
 640
 641                 tag = (journal_block_tag_t *) tagp;
 642                 tag->t_blocknr = cpu_to_be32(jh2bh(jh)->b_blocknr);
 643                 tag->t_flags = cpu_to_be32(tag_flag);
 644                 tagp += sizeof(journal_block_tag_t);
 645                 space_left -= sizeof(journal_block_tag_t);
 646
 647                 if (first_tag) {
 648                         memcpy (tagp, journal->j_uuid, 16);
 649                         tagp += 16;
 650                         space_left -= 16;
 651                         first_tag = 0;
 652                 }
 653
 654                 /* If there's no more to do, or if the descriptor is full,
 655                    let the IO rip! */
 656
 657                 if (bufs == journal->j_wbufsize ||
 658                     commit_transaction->t_buffers == NULL ||
 659                     space_left < sizeof(journal_block_tag_t) + 16) {
 660
 661                         jbd_debug(4, "JBD: Submit %d IOs\n", bufs);
 662
 663                         /* Write an end-of-descriptor marker before
 664                            submitting the IOs.  "tag" still points to
 665                            the last tag we set up. */
 666
 667                         tag->t_flags |= cpu_to_be32(JFS_FLAG_LAST_TAG);
 668
 669 start_journal_io:
 670                         for (i = 0; i < bufs; i++) {
 671                                 struct buffer_head *bh = wbuf[i];
 672                                 lock_buffer(bh);
 673                                 clear_buffer_dirty(bh);
 674                                 set_buffer_uptodate(bh);
 675                                 bh->b_end_io = journal_end_buffer_io_sync;
 676                                 submit_bh(write_op, bh);
 677                         }
 678                         cond_resched();
 679
 680                         /* Force a new descriptor to be generated next
 681                            time round the loop. */
 682                         descriptor = NULL;
 683                         bufs = 0;
 684                 }
 685         }
 686
 687         /* Lo and behold: we have just managed to send a transaction to
 688            the log.  Before we can commit it, wait for the IO so far to
 689            complete.  Control buffers being written are on the
 690            transaction's t_log_list queue, and metadata buffers are on
 691            the t_iobuf_list queue.
 692
 693            Wait for the buffers in reverse order.  That way we are
 694            less likely to be woken up until all IOs have completed, and
 695            so we incur less scheduling load.
 696         */
 697
 698         jbd_debug(3, "JBD: commit phase 4\n");
 699
 700         /*
 701          * akpm: these are BJ_IO, and j_list_lock is not needed.
 702          * See __journal_try_to_free_buffer.
 703          */
 704 wait_for_iobuf:
 705         while (commit_transaction->t_iobuf_list != NULL) {
 706                 struct buffer_head *bh;
 707
 708                 jh = commit_transaction->t_iobuf_list->b_tprev;
 709                 bh = jh2bh(jh);
 710                 if (buffer_locked(bh)) {
 711                         wait_on_buffer(bh);
 712                         goto wait_for_iobuf;
 713                 }
 714                 if (cond_resched())
 715                         goto wait_for_iobuf;
 716
 717                 if (unlikely(!buffer_uptodate(bh)))
 718                         err = -EIO;
 719
 720                 clear_buffer_jwrite(bh);
 721
 722                 JBUFFER_TRACE(jh, "ph4: unfile after journal write");
 723                 journal_unfile_buffer(journal, jh);
 724
 725                 /*
 726                  * ->t_iobuf_list should contain only dummy buffer_heads
 727                  * which were created by journal_write_metadata_buffer().
 728                  */
 729                 BUFFER_TRACE(bh, "dumping temporary bh");
 730                 journal_put_journal_head(jh);
 731                 __brelse(bh);
 732                 J_ASSERT_BH(bh, atomic_read(&bh->b_count) == 0);
 733                 free_buffer_head(bh);
 734
 735                 /* We also have to unlock and free the corresponding
 736                    shadowed buffer */
 737                 jh = commit_transaction->t_shadow_list->b_tprev;
 738                 bh = jh2bh(jh);
 739                 clear_bit(BH_JWrite, &bh->b_state);
 740                 J_ASSERT_BH(bh, buffer_jbddirty(bh));
 741
 742                 /* The metadata is now released for reuse, but we need
 743                    to remember it against this transaction so that when
 744                    we finally commit, we can do any checkpointing
 745                    required. */
 746                 JBUFFER_TRACE(jh, "file as BJ_Forget");
 747                 journal_file_buffer(jh, commit_transaction, BJ_Forget);
 748                 /* Wake up any transactions which were waiting for this
 749                    IO to complete */
 750                 wake_up_bit(&bh->b_state, BH_Unshadow);
 751                 JBUFFER_TRACE(jh, "brelse shadowed buffer");
 752                 __brelse(bh);
 753         }
 754
 755         J_ASSERT (commit_transaction->t_shadow_list == NULL);
 756
 757         jbd_debug(3, "JBD: commit phase 5\n");
 758
 759         /* Here we wait for the revoke record and descriptor record buffers */
 760  wait_for_ctlbuf:
 761         while (commit_transaction->t_log_list != NULL) {
 762                 struct buffer_head *bh;
 763
 764                 jh = commit_transaction->t_log_list->b_tprev;
 765                 bh = jh2bh(jh);
 766                 if (buffer_locked(bh)) {
 767                         wait_on_buffer(bh);
 768                         goto wait_for_ctlbuf;
 769                 }
 770                 if (cond_resched())
 771                         goto wait_for_ctlbuf;
 772
 773                 if (unlikely(!buffer_uptodate(bh)))
 774                         err = -EIO;
 775
 776                 BUFFER_TRACE(bh, "ph5: control buffer writeout done: unfile");
 777                 clear_buffer_jwrite(bh);
 778                 journal_unfile_buffer(journal, jh);
 779                 journal_put_journal_head(jh);
 780                 __brelse(bh);           /* One for getblk */
 781                 /* AKPM: bforget here */
 782         }
 783
 784         if (err)
 785                 journal_abort(journal, err);
 786
 787         jbd_debug(3, "JBD: commit phase 6\n");
 788
 789         /* All metadata is written, now write commit record and do cleanup */
 790         spin_lock(&journal->j_state_lock);
 791         J_ASSERT(commit_transaction->t_state == T_COMMIT);
 792         commit_transaction->t_state = T_COMMIT_RECORD;
 793         spin_unlock(&journal->j_state_lock);
 794
 795         if (journal_write_commit_record(journal, commit_transaction))
 796                 err = -EIO;
 797
 798         if (err)
 799                 journal_abort(journal, err);
 800
 801         /* End of a transaction!  Finally, we can do checkpoint
 802            processing: any buffers committed as a result of this
 803            transaction can be removed from any checkpoint list it was on
 804            before. */
 805
 806         jbd_debug(3, "JBD: commit phase 7\n");
 807
 808         J_ASSERT(commit_transaction->t_sync_datalist == NULL);
 809         J_ASSERT(commit_transaction->t_buffers == NULL);
 810         J_ASSERT(commit_transaction->t_checkpoint_list == NULL);
 811         J_ASSERT(commit_transaction->t_iobuf_list == NULL);
 812         J_ASSERT(commit_transaction->t_shadow_list == NULL);
 813         J_ASSERT(commit_transaction->t_log_list == NULL);
 814
 815 restart_loop:
 816         /*
 817          * As there are other places (journal_unmap_buffer()) adding buffers
 818          * to this list we have to be careful and hold the j_list_lock.
 819          */
 820         spin_lock(&journal->j_list_lock);
 821         while (commit_transaction->t_forget) {
 822                 transaction_t *cp_transaction;
 823                 struct buffer_head *bh;
 824
 825                 jh = commit_transaction->t_forget;
 826                 spin_unlock(&journal->j_list_lock);
 827                 bh = jh2bh(jh);
 828                 jbd_lock_bh_state(bh);
 829                 J_ASSERT_JH(jh, jh->b_transaction == commit_transaction ||
 830                         jh->b_transaction == journal->j_running_transaction);
 831
 832                 /*
 833                  * If there is undo-protected committed data against
 834                  * this buffer, then we can remove it now.  If it is a
 835                  * buffer needing such protection, the old frozen_data
 836                  * field now points to a committed version of the
 837                  * buffer, so rotate that field to the new committed
 838                  * data.
 839                  *
 840                  * Otherwise, we can just throw away the frozen data now.
 841                  */
 842                 if (jh->b_committed_data) {
 843                         jbd_free(jh->b_committed_data, bh->b_size);
 844                         jh->b_committed_data = NULL;
 845                         if (jh->b_frozen_data) {
 846                                 jh->b_committed_data = jh->b_frozen_data;
 847                                 jh->b_frozen_data = NULL;
 848                         }
 849                 } else if (jh->b_frozen_data) {
 850                         jbd_free(jh->b_frozen_data, bh->b_size);
 851                         jh->b_frozen_data = NULL;
 852                 }
 853
 854                 spin_lock(&journal->j_list_lock);
 855                 cp_transaction = jh->b_cp_transaction;
 856                 if (cp_transaction) {
 857                         JBUFFER_TRACE(jh, "remove from old cp transaction");
 858                         __journal_remove_checkpoint(jh);
 859                 }
 860
 861                 /* Only re-checkpoint the buffer_head if it is marked
 862                  * dirty.  If the buffer was added to the BJ_Forget list
 863                  * by journal_forget, it may no longer be dirty and
 864                  * there's no point in keeping a checkpoint record for
 865                  * it. */
 866
 867                 /* A buffer which has been freed while still being
 868                  * journaled by a previous transaction may end up still
 869                  * being dirty here, but we want to avoid writing back
 870                  * that buffer in the future after the "add to orphan"
 871                  * operation been committed,  That's not only a performance
 872                  * gain, it also stops aliasing problems if the buffer is
 873                  * left behind for writeback and gets reallocated for another
 874                  * use in a different page. */
 875                 if (buffer_freed(bh) && !jh->b_next_transaction) {
 876                         clear_buffer_freed(bh);
 877                         clear_buffer_jbddirty(bh);
 878                 }
 879
 880                 if (buffer_jbddirty(bh)) {
 881                         JBUFFER_TRACE(jh, "add to new checkpointing trans");
 882                         __journal_insert_checkpoint(jh, commit_transaction);
 883                         if (is_journal_aborted(journal))
 884                                 clear_buffer_jbddirty(bh);
 885                         JBUFFER_TRACE(jh, "refile for checkpoint writeback");
 886                         __journal_refile_buffer(jh);
 887                         jbd_unlock_bh_state(bh);
 888                 } else {
 889                         J_ASSERT_BH(bh, !buffer_dirty(bh));
 890                         /* The buffer on BJ_Forget list and not jbddirty means
 891                          * it has been freed by this transaction and hence it
 892                          * could not have been reallocated until this
 893                          * transaction has committed. *BUT* it could be
 894                          * reallocated once we have written all the data to
 895                          * disk and before we process the buffer on BJ_Forget
 896                          * list. */
 897                         JBUFFER_TRACE(jh, "refile or unfile freed buffer");
 898                         __journal_refile_buffer(jh);
 899                         if (!jh->b_transaction) {
 900                                 jbd_unlock_bh_state(bh);
 901                                  /* needs a brelse */
 902                                 journal_remove_journal_head(bh);
 903                                 release_buffer_page(bh);
 904                         } else
 905                                 jbd_unlock_bh_state(bh);
 906                 }
 907                 cond_resched_lock(&journal->j_list_lock);
 908         }
 909         spin_unlock(&journal->j_list_lock);
 910         /*
 911          * This is a bit sleazy.  We use j_list_lock to protect transition
 912          * of a transaction into T_FINISHED state and calling
 913          * __journal_drop_transaction(). Otherwise we could race with
 914          * other checkpointing code processing the transaction...
 915          */
 916         spin_lock(&journal->j_state_lock);
 917         spin_lock(&journal->j_list_lock);
 918         /*
 919          * Now recheck if some buffers did not get attached to the transaction
 920          * while the lock was dropped...
 921          */
 922         if (commit_transaction->t_forget) {
 923                 spin_unlock(&journal->j_list_lock);
 924                 spin_unlock(&journal->j_state_lock);
 925                 goto restart_loop;
 926         }
 927
 928         /* Done with this transaction! */
 929
 930         jbd_debug(3, "JBD: commit phase 8\n");
 931
 932         J_ASSERT(commit_transaction->t_state == T_COMMIT_RECORD);
 933
 934         commit_transaction->t_state = T_FINISHED;
 935         J_ASSERT(commit_transaction == journal->j_committing_transaction);
 936         journal->j_commit_sequence = commit_transaction->t_tid;
 937         journal->j_committing_transaction = NULL;
 938         commit_time = ktime_to_ns(ktime_sub(ktime_get(), start_time));
 939
 940         /*
 941          * weight the commit time higher than the average time so we don't
 942          * react too strongly to vast changes in commit time
 943          */
 944         if (likely(journal->j_average_commit_time))
 945                 journal->j_average_commit_time = (commit_time*3 +
 946                                 journal->j_average_commit_time) / 4;
 947         else
 948                 journal->j_average_commit_time = commit_time;
 949
 950         spin_unlock(&journal->j_state_lock);
 951
 952         if (commit_transaction->t_checkpoint_list == NULL &&
 953             commit_transaction->t_checkpoint_io_list == NULL) {
 954                 __journal_drop_transaction(journal, commit_transaction);
 955         } else {
 956                 if (journal->j_checkpoint_transactions == NULL) {
 957                         journal->j_checkpoint_transactions = commit_transaction;
 958                         commit_transaction->t_cpnext = commit_transaction;
 959                         commit_transaction->t_cpprev = commit_transaction;
 960                 } else {
 961                         commit_transaction->t_cpnext =
 962                                 journal->j_checkpoint_transactions;
 963                         commit_transaction->t_cpprev =
 964                                 commit_transaction->t_cpnext->t_cpprev;
 965                         commit_transaction->t_cpnext->t_cpprev =
 966                                 commit_transaction;
 967                         commit_transaction->t_cpprev->t_cpnext =
 968                                 commit_transaction;
 969                 }
 970         }
 971         spin_unlock(&journal->j_list_lock);
 972
 973         jbd_debug(1, "JBD: commit %d complete, head %d\n",
 974                   journal->j_commit_sequence, journal->j_tail_sequence);
 975
 976         wake_up(&journal->j_wait_done_commit);
 977 }