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
 123         if (is_journal_aborted(journal))
 124                 return 0;
 125
 126         descriptor = journal_get_descriptor_buffer(journal);
 127         if (!descriptor)
 128                 return 1;
 129
 130         bh = jh2bh(descriptor);
 131
 132         header = (journal_header_t *)(bh->b_data);
 133         header->h_magic = cpu_to_be32(JFS_MAGIC_NUMBER);
 134         header->h_blocktype = cpu_to_be32(JFS_COMMIT_BLOCK);
 135         header->h_sequence = cpu_to_be32(commit_transaction->t_tid);
 136
 137         JBUFFER_TRACE(descriptor, "write commit block");
 138         set_buffer_dirty(bh);
 139
 140         if (journal->j_flags & JFS_BARRIER)
 141                 ret = __sync_dirty_buffer(bh, WRITE_SYNC | WRITE_FLUSH_FUA);
 142         else
 143                 ret = sync_dirty_buffer(bh);
 144
 145         put_bh(bh);             /* One for getblk() */
 146         journal_put_journal_head(descriptor);
 147
 148         return (ret == -EIO);
 149 }
 150
 151 static void journal_do_submit_data(struct buffer_head **wbuf, int bufs,
 152                                    int write_op)
 153 {
 154         int i;
 155
 156         for (i = 0; i < bufs; i++) {
 157                 wbuf[i]->b_end_io = end_buffer_write_sync;
 158                 /* We use-up our safety reference in submit_bh() */
 159                 submit_bh(write_op, wbuf[i]);
 160         }
 161 }
 162
 163 /*
 164  *  Submit all the data buffers to disk
 165  */
 166 static int journal_submit_data_buffers(journal_t *journal,
 167                                        transaction_t *commit_transaction,
 168                                        int write_op)
 169 {
 170         struct journal_head *jh;
 171         struct buffer_head *bh;
 172         int locked;
 173         int bufs = 0;
 174         struct buffer_head **wbuf = journal->j_wbuf;
 175         int err = 0;
 176
 177         /*
 178          * Whenever we unlock the journal and sleep, things can get added
 179          * onto ->t_sync_datalist, so we have to keep looping back to
 180          * write_out_data until we *know* that the list is empty.
 181          *
 182          * Cleanup any flushed data buffers from the data list.  Even in
 183          * abort mode, we want to flush this out as soon as possible.
 184          */
 185 write_out_data:
 186         cond_resched();
 187         spin_lock(&journal->j_list_lock);
 188
 189         while (commit_transaction->t_sync_datalist) {
 190                 jh = commit_transaction->t_sync_datalist;
 191                 bh = jh2bh(jh);
 192                 locked = 0;
 193
 194                 /* Get reference just to make sure buffer does not disappear
 195                  * when we are forced to drop various locks */
 196                 get_bh(bh);
 197                 /* If the buffer is dirty, we need to submit IO and hence
 198                  * we need the buffer lock. We try to lock the buffer without
 199                  * blocking. If we fail, we need to drop j_list_lock and do
 200                  * blocking lock_buffer().
 201                  */
 202                 if (buffer_dirty(bh)) {
 203                         if (!trylock_buffer(bh)) {
 204                                 BUFFER_TRACE(bh, "needs blocking lock");
 205                                 spin_unlock(&journal->j_list_lock);
 206                                 /* Write out all data to prevent deadlocks */
 207                                 journal_do_submit_data(wbuf, bufs, write_op);
 208                                 bufs = 0;
 209                                 lock_buffer(bh);
 210                                 spin_lock(&journal->j_list_lock);
 211                         }
 212                         locked = 1;
 213                 }
 214                 /* We have to get bh_state lock. Again out of order, sigh. */
 215                 if (!inverted_lock(journal, bh)) {
 216                         jbd_lock_bh_state(bh);
 217                         spin_lock(&journal->j_list_lock);
 218                 }
 219                 /* Someone already cleaned up the buffer? */
 220                 if (!buffer_jbd(bh) || bh2jh(bh) != jh
 221                         || jh->b_transaction != commit_transaction
 222                         || jh->b_jlist != BJ_SyncData) {
 223                         jbd_unlock_bh_state(bh);
 224                         if (locked)
 225                                 unlock_buffer(bh);
 226                         BUFFER_TRACE(bh, "already cleaned up");
 227                         release_data_buffer(bh);
 228                         continue;
 229                 }
 230                 if (locked && test_clear_buffer_dirty(bh)) {
 231                         BUFFER_TRACE(bh, "needs writeout, adding to array");
 232                         wbuf[bufs++] = bh;
 233                         __journal_file_buffer(jh, commit_transaction,
 234                                                 BJ_Locked);
 235                         jbd_unlock_bh_state(bh);
 236                         if (bufs == journal->j_wbufsize) {
 237                                 spin_unlock(&journal->j_list_lock);
 238                                 journal_do_submit_data(wbuf, bufs, write_op);
 239                                 bufs = 0;
 240                                 goto write_out_data;
 241                         }
 242                 } else if (!locked && buffer_locked(bh)) {
 243                         __journal_file_buffer(jh, commit_transaction,
 244                                                 BJ_Locked);
 245                         jbd_unlock_bh_state(bh);
 246                         put_bh(bh);
 247                 } else {
 248                         BUFFER_TRACE(bh, "writeout complete: unfile");
 249                         if (unlikely(!buffer_uptodate(bh)))
 250                                 err = -EIO;
 251                         __journal_unfile_buffer(jh);
 252                         jbd_unlock_bh_state(bh);
 253                         if (locked)
 254                                 unlock_buffer(bh);
 255                         journal_remove_journal_head(bh);
 256                         /* One for our safety reference, other for
 257                          * journal_remove_journal_head() */
 258                         put_bh(bh);
 259                         release_data_buffer(bh);
 260                 }
 261
 262                 if (need_resched() || spin_needbreak(&journal->j_list_lock)) {
 263                         spin_unlock(&journal->j_list_lock);
 264                         goto write_out_data;
 265                 }
 266         }
 267         spin_unlock(&journal->j_list_lock);
 268         journal_do_submit_data(wbuf, bufs, write_op);
 269
 270         return err;
 271 }
 272
 273 /*
 274  * journal_commit_transaction
 275  *
 276  * The primary function for committing a transaction to the log.  This
 277  * function is called by the journal thread to begin a complete commit.
 278  */
 279 void journal_commit_transaction(journal_t *journal)
 280 {
 281         transaction_t *commit_transaction;
 282         struct journal_head *jh, *new_jh, *descriptor;
 283         struct buffer_head **wbuf = journal->j_wbuf;
 284         int bufs;
 285         int flags;
 286         int err;
 287         unsigned int blocknr;
 288         ktime_t start_time;
 289         u64 commit_time;
 290         char *tagp = NULL;
 291         journal_header_t *header;
 292         journal_block_tag_t *tag = NULL;
 293         int space_left = 0;
 294         int first_tag = 0;
 295         int tag_flag;
 296         int i;
 297         int write_op = WRITE_SYNC;
 298
 299         /*
 300          * First job: lock down the current transaction and wait for
 301          * all outstanding updates to complete.
 302          */
 303
 304 #ifdef COMMIT_STATS
 305         spin_lock(&journal->j_list_lock);
 306         summarise_journal_usage(journal);
 307         spin_unlock(&journal->j_list_lock);
 308 #endif
 309
 310         /* Do we need to erase the effects of a prior journal_flush? */
 311         if (journal->j_flags & JFS_FLUSHED) {
 312                 jbd_debug(3, "super block updated\n");
 313                 journal_update_superblock(journal, 1);
 314         } else {
 315                 jbd_debug(3, "superblock not updated\n");
 316         }
 317
 318         J_ASSERT(journal->j_running_transaction != NULL);
 319         J_ASSERT(journal->j_committing_transaction == NULL);
 320
 321         commit_transaction = journal->j_running_transaction;
 322         J_ASSERT(commit_transaction->t_state == T_RUNNING);
 323
 324         jbd_debug(1, "JBD: starting commit of transaction %d\n",
 325                         commit_transaction->t_tid);
 326
 327         spin_lock(&journal->j_state_lock);
 328         commit_transaction->t_state = T_LOCKED;
 329
 330         /*
 331          * Use plugged writes here, since we want to submit several before
 332          * we unplug the device. We don't do explicit unplugging in here,
 333          * instead we rely on sync_buffer() doing the unplug for us.
 334          */
 335         if (commit_transaction->t_synchronous_commit)
 336                 write_op = WRITE_SYNC_PLUG;
 337         spin_lock(&commit_transaction->t_handle_lock);
 338         while (commit_transaction->t_updates) {
 339                 DEFINE_WAIT(wait);
 340
 341                 prepare_to_wait(&journal->j_wait_updates, &wait,
 342                                         TASK_UNINTERRUPTIBLE);
 343                 if (commit_transaction->t_updates) {
 344                         spin_unlock(&commit_transaction->t_handle_lock);
 345                         spin_unlock(&journal->j_state_lock);
 346                         schedule();
 347                         spin_lock(&journal->j_state_lock);
 348                         spin_lock(&commit_transaction->t_handle_lock);
 349                 }
 350                 finish_wait(&journal->j_wait_updates, &wait);
 351         }
 352         spin_unlock(&commit_transaction->t_handle_lock);
 353
 354         J_ASSERT (commit_transaction->t_outstanding_credits <=
 355                         journal->j_max_transaction_buffers);
 356
 357         /*
 358          * First thing we are allowed to do is to discard any remaining
 359          * BJ_Reserved buffers.  Note, it is _not_ permissible to assume
 360          * that there are no such buffers: if a large filesystem
 361          * operation like a truncate needs to split itself over multiple
 362          * transactions, then it may try to do a journal_restart() while
 363          * there are still BJ_Reserved buffers outstanding.  These must
 364          * be released cleanly from the current transaction.
 365          *
 366          * In this case, the filesystem must still reserve write access
 367          * again before modifying the buffer in the new transaction, but
 368          * we do not require it to remember exactly which old buffers it
 369          * has reserved.  This is consistent with the existing behaviour
 370          * that multiple journal_get_write_access() calls to the same
 371          * buffer are perfectly permissable.
 372          */
 373         while (commit_transaction->t_reserved_list) {
 374                 jh = commit_transaction->t_reserved_list;
 375                 JBUFFER_TRACE(jh, "reserved, unused: refile");
 376                 /*
 377                  * A journal_get_undo_access()+journal_release_buffer() may
 378                  * leave undo-committed data.
 379                  */
 380                 if (jh->b_committed_data) {
 381                         struct buffer_head *bh = jh2bh(jh);
 382
 383                         jbd_lock_bh_state(bh);
 384                         jbd_free(jh->b_committed_data, bh->b_size);
 385                         jh->b_committed_data = NULL;
 386                         jbd_unlock_bh_state(bh);
 387                 }
 388                 journal_refile_buffer(journal, jh);
 389         }
 390
 391         /*
 392          * Now try to drop any written-back buffers from the journal's
 393          * checkpoint lists.  We do this *before* commit because it potentially
 394          * frees some memory
 395          */
 396         spin_lock(&journal->j_list_lock);
 397         __journal_clean_checkpoint_list(journal);
 398         spin_unlock(&journal->j_list_lock);
 399
 400         jbd_debug (3, "JBD: commit phase 1\n");
 401
 402         /*
 403          * Switch to a new revoke table.
 404          */
 405         journal_switch_revoke_table(journal);
 406
 407         commit_transaction->t_state = T_FLUSH;
 408         journal->j_committing_transaction = commit_transaction;
 409         journal->j_running_transaction = NULL;
 410         start_time = ktime_get();
 411         commit_transaction->t_log_start = journal->j_head;
 412         wake_up(&journal->j_wait_transaction_locked);
 413         spin_unlock(&journal->j_state_lock);
 414
 415         jbd_debug (3, "JBD: commit phase 2\n");
 416
 417         /*
 418          * Now start flushing things to disk, in the order they appear
 419          * on the transaction lists.  Data blocks go first.
 420          */
 421         err = journal_submit_data_buffers(journal, commit_transaction,
 422                                           write_op);
 423
 424         /*
 425          * Wait for all previously submitted IO to complete.
 426          */
 427         spin_lock(&journal->j_list_lock);
 428         while (commit_transaction->t_locked_list) {
 429                 struct buffer_head *bh;
 430
 431                 jh = commit_transaction->t_locked_list->b_tprev;
 432                 bh = jh2bh(jh);
 433                 get_bh(bh);
 434                 if (buffer_locked(bh)) {
 435                         spin_unlock(&journal->j_list_lock);
 436                         wait_on_buffer(bh);
 437                         spin_lock(&journal->j_list_lock);
 438                 }
 439                 if (unlikely(!buffer_uptodate(bh))) {
 440                         if (!trylock_page(bh->b_page)) {
 441                                 spin_unlock(&journal->j_list_lock);
 442                                 lock_page(bh->b_page);
 443                                 spin_lock(&journal->j_list_lock);
 444                         }
 445                         if (bh->b_page->mapping)
 446                                 set_bit(AS_EIO, &bh->b_page->mapping->flags);
 447
 448                         unlock_page(bh->b_page);
 449                         SetPageError(bh->b_page);
 450                         err = -EIO;
 451                 }
 452                 if (!inverted_lock(journal, bh)) {
 453                         put_bh(bh);
 454                         spin_lock(&journal->j_list_lock);
 455                         continue;
 456                 }
 457                 if (buffer_jbd(bh) && bh2jh(bh) == jh &&
 458                     jh->b_transaction == commit_transaction &&
 459                     jh->b_jlist == BJ_Locked) {
 460                         __journal_unfile_buffer(jh);
 461                         jbd_unlock_bh_state(bh);
 462                         journal_remove_journal_head(bh);
 463                         put_bh(bh);
 464                 } else {
 465                         jbd_unlock_bh_state(bh);
 466                 }
 467                 release_data_buffer(bh);
 468                 cond_resched_lock(&journal->j_list_lock);
 469         }
 470         spin_unlock(&journal->j_list_lock);
 471
 472         if (err) {
 473                 char b[BDEVNAME_SIZE];
 474
 475                 printk(KERN_WARNING
 476                         "JBD: Detected IO errors while flushing file data "
 477                         "on %s\n", bdevname(journal->j_fs_dev, b));
 478                 if (journal->j_flags & JFS_ABORT_ON_SYNCDATA_ERR)
 479                         journal_abort(journal, err);
 480                 err = 0;
 481         }
 482
 483         journal_write_revoke_records(journal, commit_transaction, write_op);
 484
 485         /*
 486          * If we found any dirty or locked buffers, then we should have
 487          * looped back up to the write_out_data label.  If there weren't
 488          * any then journal_clean_data_list should have wiped the list
 489          * clean by now, so check that it is in fact empty.
 490          */
 491         J_ASSERT (commit_transaction->t_sync_datalist == NULL);
 492
 493         jbd_debug (3, "JBD: commit phase 3\n");
 494
 495         /*
 496          * Way to go: we have now written out all of the data for a
 497          * transaction!  Now comes the tricky part: we need to write out
 498          * metadata.  Loop over the transaction's entire buffer list:
 499          */
 500         spin_lock(&journal->j_state_lock);
 501         commit_transaction->t_state = T_COMMIT;
 502         spin_unlock(&journal->j_state_lock);
 503
 504         J_ASSERT(commit_transaction->t_nr_buffers <=
 505                  commit_transaction->t_outstanding_credits);
 506
 507         descriptor = NULL;
 508         bufs = 0;
 509         while (commit_transaction->t_buffers) {
 510
 511                 /* Find the next buffer to be journaled... */
 512
 513                 jh = commit_transaction->t_buffers;
 514
 515                 /* If we're in abort mode, we just un-journal the buffer and
 516                    release it. */
 517
 518                 if (is_journal_aborted(journal)) {
 519                         clear_buffer_jbddirty(jh2bh(jh));
 520                         JBUFFER_TRACE(jh, "journal is aborting: refile");
 521                         journal_refile_buffer(journal, jh);
 522                         /* If that was the last one, we need to clean up
 523                          * any descriptor buffers which may have been
 524                          * already allocated, even if we are now
 525                          * aborting. */
 526                         if (!commit_transaction->t_buffers)
 527                                 goto start_journal_io;
 528                         continue;
 529                 }
 530
 531                 /* Make sure we have a descriptor block in which to
 532                    record the metadata buffer. */
 533
 534                 if (!descriptor) {
 535                         struct buffer_head *bh;
 536
 537                         J_ASSERT (bufs == 0);
 538
 539                         jbd_debug(4, "JBD: get descriptor\n");
 540
 541                         descriptor = journal_get_descriptor_buffer(journal);
 542                         if (!descriptor) {
 543                                 journal_abort(journal, -EIO);
 544                                 continue;
 545                         }
 546
 547                         bh = jh2bh(descriptor);
 548                         jbd_debug(4, "JBD: got buffer %llu (%p)\n",
 549                                 (unsigned long long)bh->b_blocknr, bh->b_data);
 550                         header = (journal_header_t *)&bh->b_data[0];
 551                         header->h_magic     = cpu_to_be32(JFS_MAGIC_NUMBER);
 552                         header->h_blocktype = cpu_to_be32(JFS_DESCRIPTOR_BLOCK);
 553                         header->h_sequence  = cpu_to_be32(commit_transaction->t_tid);
 554
 555                         tagp = &bh->b_data[sizeof(journal_header_t)];
 556                         space_left = bh->b_size - sizeof(journal_header_t);
 557                         first_tag = 1;
 558                         set_buffer_jwrite(bh);
 559                         set_buffer_dirty(bh);
 560                         wbuf[bufs++] = bh;
 561
 562                         /* Record it so that we can wait for IO
 563                            completion later */
 564                         BUFFER_TRACE(bh, "ph3: file as descriptor");
 565                         journal_file_buffer(descriptor, commit_transaction,
 566                                         BJ_LogCtl);
 567                 }
 568
 569                 /* Where is the buffer to be written? */
 570
 571                 err = journal_next_log_block(journal, &blocknr);
 572                 /* If the block mapping failed, just abandon the buffer
 573                    and repeat this loop: we'll fall into the
 574                    refile-on-abort condition above. */
 575                 if (err) {
 576                         journal_abort(journal, err);
 577                         continue;
 578                 }
 579
 580                 /*
 581                  * start_this_handle() uses t_outstanding_credits to determine
 582                  * the free space in the log, but this counter is changed
 583                  * by journal_next_log_block() also.
 584                  */
 585                 commit_transaction->t_outstanding_credits--;
 586
 587                 /* Bump b_count to prevent truncate from stumbling over
 588                    the shadowed buffer!  @@@ This can go if we ever get
 589                    rid of the BJ_IO/BJ_Shadow pairing of buffers. */
 590                 atomic_inc(&jh2bh(jh)->b_count);
 591
 592                 /* Make a temporary IO buffer with which to write it out
 593                    (this will requeue both the metadata buffer and the
 594                    temporary IO buffer). new_bh goes on BJ_IO*/
 595
 596                 set_bit(BH_JWrite, &jh2bh(jh)->b_state);
 597                 /*
 598                  * akpm: journal_write_metadata_buffer() sets
 599                  * new_bh->b_transaction to commit_transaction.
 600                  * We need to clean this up before we release new_bh
 601                  * (which is of type BJ_IO)
 602                  */
 603                 JBUFFER_TRACE(jh, "ph3: write metadata");
 604                 flags = journal_write_metadata_buffer(commit_transaction,
 605                                                       jh, &new_jh, blocknr);
 606                 set_bit(BH_JWrite, &jh2bh(new_jh)->b_state);
 607                 wbuf[bufs++] = jh2bh(new_jh);
 608
 609                 /* Record the new block's tag in the current descriptor
 610                    buffer */
 611
 612                 tag_flag = 0;
 613                 if (flags & 1)
 614                         tag_flag |= JFS_FLAG_ESCAPE;
 615                 if (!first_tag)
 616                         tag_flag |= JFS_FLAG_SAME_UUID;
 617
 618                 tag = (journal_block_tag_t *) tagp;
 619                 tag->t_blocknr = cpu_to_be32(jh2bh(jh)->b_blocknr);
 620                 tag->t_flags = cpu_to_be32(tag_flag);
 621                 tagp += sizeof(journal_block_tag_t);
 622                 space_left -= sizeof(journal_block_tag_t);
 623
 624                 if (first_tag) {
 625                         memcpy (tagp, journal->j_uuid, 16);
 626                         tagp += 16;
 627                         space_left -= 16;
 628                         first_tag = 0;
 629                 }
 630
 631                 /* If there's no more to do, or if the descriptor is full,
 632                    let the IO rip! */
 633
 634                 if (bufs == journal->j_wbufsize ||
 635                     commit_transaction->t_buffers == NULL ||
 636                     space_left < sizeof(journal_block_tag_t) + 16) {
 637
 638                         jbd_debug(4, "JBD: Submit %d IOs\n", bufs);
 639
 640                         /* Write an end-of-descriptor marker before
 641                            submitting the IOs.  "tag" still points to
 642                            the last tag we set up. */
 643
 644                         tag->t_flags |= cpu_to_be32(JFS_FLAG_LAST_TAG);
 645
 646 start_journal_io:
 647                         for (i = 0; i < bufs; i++) {
 648                                 struct buffer_head *bh = wbuf[i];
 649                                 lock_buffer(bh);
 650                                 clear_buffer_dirty(bh);
 651                                 set_buffer_uptodate(bh);
 652                                 bh->b_end_io = journal_end_buffer_io_sync;
 653                                 submit_bh(write_op, bh);
 654                         }
 655                         cond_resched();
 656
 657                         /* Force a new descriptor to be generated next
 658                            time round the loop. */
 659                         descriptor = NULL;
 660                         bufs = 0;
 661                 }
 662         }
 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_bit(BH_JWrite, &bh->b_state);
 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 }