fs/jbd2/commit.c

   1 /*
   2  * linux/fs/jbd2/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/jbd2.h>
  19 #include <linux/marker.h>
  20 #include <linux/errno.h>
  21 #include <linux/slab.h>
  22 #include <linux/mm.h>
  23 #include <linux/pagemap.h>
  24 #include <linux/jiffies.h>
  25 #include <linux/crc32.h>
  26 #include <linux/writeback.h>
  27 #include <linux/backing-dev.h>
  28 #include <linux/bio.h>
  29
  30 /*
  31  * Default IO end handler for temporary BJ_IO buffer_heads.
  32  */
  33 static void journal_end_buffer_io_sync(struct buffer_head *bh, int uptodate)
  34 {
  35         BUFFER_TRACE(bh, "");
  36         if (uptodate)
  37                 set_buffer_uptodate(bh);
  38         else
  39                 clear_buffer_uptodate(bh);
  40         unlock_buffer(bh);
  41 }
  42
  43 /*
  44  * When an ext4 file is truncated, it is possible that some pages are not
  45  * successfully freed, because they are attached to a committing transaction.
  46  * After the transaction commits, these pages are left on the LRU, with no
  47  * ->mapping, and with attached buffers.  These pages are trivially reclaimable
  48  * by the VM, but their apparent absence upsets the VM accounting, and it makes
  49  * the numbers in /proc/meminfo look odd.
  50  *
  51  * So here, we have a buffer which has just come off the forget list.  Look to
  52  * see if we can strip all buffers from the backing page.
  53  *
  54  * Called under lock_journal(), and possibly under journal_datalist_lock.  The
  55  * caller provided us with a ref against the buffer, and we drop that here.
  56  */
  57 static void release_buffer_page(struct buffer_head *bh)
  58 {
  59         struct page *page;
  60
  61         if (buffer_dirty(bh))
  62                 goto nope;
  63         if (atomic_read(&bh->b_count) != 1)
  64                 goto nope;
  65         page = bh->b_page;
  66         if (!page)
  67                 goto nope;
  68         if (page->mapping)
  69                 goto nope;
  70
  71         /* OK, it's a truncated page */
  72         if (!trylock_page(page))
  73                 goto nope;
  74
  75         page_cache_get(page);
  76         __brelse(bh);
  77         try_to_free_buffers(page);
  78         unlock_page(page);
  79         page_cache_release(page);
  80         return;
  81
  82 nope:
  83         __brelse(bh);
  84 }
  85
  86 /*
  87  * Done it all: now submit the commit record.  We should have
  88  * cleaned up our previous buffers by now, so if we are in abort
  89  * mode we can now just skip the rest of the journal write
  90  * entirely.
  91  *
  92  * Returns 1 if the journal needs to be aborted or 0 on success
  93  */
  94 static int journal_submit_commit_record(journal_t *journal,
  95                                         transaction_t *commit_transaction,
  96                                         struct buffer_head **cbh,
  97                                         __u32 crc32_sum)
  98 {
  99         struct journal_head *descriptor;
 100         struct commit_header *tmp;
 101         struct buffer_head *bh;
 102         int ret;
 103         int barrier_done = 0;
 104         struct timespec now = current_kernel_time();
 105
 106         if (is_journal_aborted(journal))
 107                 return 0;
 108
 109         descriptor = jbd2_journal_get_descriptor_buffer(journal);
 110         if (!descriptor)
 111                 return 1;
 112
 113         bh = jh2bh(descriptor);
 114
 115         tmp = (struct commit_header *)bh->b_data;
 116         tmp->h_magic = cpu_to_be32(JBD2_MAGIC_NUMBER);
 117         tmp->h_blocktype = cpu_to_be32(JBD2_COMMIT_BLOCK);
 118         tmp->h_sequence = cpu_to_be32(commit_transaction->t_tid);
 119         tmp->h_commit_sec = cpu_to_be64(now.tv_sec);
 120         tmp->h_commit_nsec = cpu_to_be32(now.tv_nsec);
 121
 122         if (JBD2_HAS_COMPAT_FEATURE(journal,
 123                                     JBD2_FEATURE_COMPAT_CHECKSUM)) {
 124                 tmp->h_chksum_type      = JBD2_CRC32_CHKSUM;
 125                 tmp->h_chksum_size      = JBD2_CRC32_CHKSUM_SIZE;
 126                 tmp->h_chksum[0]        = cpu_to_be32(crc32_sum);
 127         }
 128
 129         JBUFFER_TRACE(descriptor, "submit commit block");
 130         lock_buffer(bh);
 131         clear_buffer_dirty(bh);
 132         set_buffer_uptodate(bh);
 133         bh->b_end_io = journal_end_buffer_io_sync;
 134
 135         if (journal->j_flags & JBD2_BARRIER &&
 136                 !JBD2_HAS_INCOMPAT_FEATURE(journal,
 137                                          JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
 138                 set_buffer_ordered(bh);
 139                 barrier_done = 1;
 140         }
 141         ret = submit_bh(WRITE, bh);
 142         if (barrier_done)
 143                 clear_buffer_ordered(bh);
 144
 145         /* is it possible for another commit to fail at roughly
 146          * the same time as this one?  If so, we don't want to
 147          * trust the barrier flag in the super, but instead want
 148          * to remember if we sent a barrier request
 149          */
 150         if (ret == -EOPNOTSUPP && barrier_done) {
 151                 printk(KERN_WARNING
 152                        "JBD: barrier-based sync failed on %s - "
 153                        "disabling barriers\n", journal->j_devname);
 154                 spin_lock(&journal->j_state_lock);
 155                 journal->j_flags &= ~JBD2_BARRIER;
 156                 spin_unlock(&journal->j_state_lock);
 157
 158                 /* And try again, without the barrier */
 159                 lock_buffer(bh);
 160                 set_buffer_uptodate(bh);
 161                 clear_buffer_dirty(bh);
 162                 ret = submit_bh(WRITE, bh);
 163         }
 164         *cbh = bh;
 165         return ret;
 166 }
 167
 168 /*
 169  * This function along with journal_submit_commit_record
 170  * allows to write the commit record asynchronously.
 171  */
 172 static int journal_wait_on_commit_record(journal_t *journal,
 173                                          struct buffer_head *bh)
 174 {
 175         int ret = 0;
 176
 177 retry:
 178         clear_buffer_dirty(bh);
 179         wait_on_buffer(bh);
 180         if (buffer_eopnotsupp(bh) && (journal->j_flags & JBD2_BARRIER)) {
 181                 printk(KERN_WARNING
 182                        "JBD2: wait_on_commit_record: sync failed on %s - "
 183                        "disabling barriers\n", journal->j_devname);
 184                 spin_lock(&journal->j_state_lock);
 185                 journal->j_flags &= ~JBD2_BARRIER;
 186                 spin_unlock(&journal->j_state_lock);
 187
 188                 lock_buffer(bh);
 189                 clear_buffer_dirty(bh);
 190                 set_buffer_uptodate(bh);
 191                 bh->b_end_io = journal_end_buffer_io_sync;
 192
 193                 ret = submit_bh(WRITE_SYNC, bh);
 194                 if (ret) {
 195                         unlock_buffer(bh);
 196                         return ret;
 197                 }
 198                 goto retry;
 199         }
 200
 201         if (unlikely(!buffer_uptodate(bh)))
 202                 ret = -EIO;
 203         put_bh(bh);            /* One for getblk() */
 204         jbd2_journal_put_journal_head(bh2jh(bh));
 205
 206         return ret;
 207 }
 208
 209 /*
 210  * write the filemap data using writepage() address_space_operations.
 211  * We don't do block allocation here even for delalloc. We don't
 212  * use writepages() because with dealyed allocation we may be doing
 213  * block allocation in writepages().
 214  */
 215 static int journal_submit_inode_data_buffers(struct address_space *mapping)
 216 {
 217         int ret;
 218         struct writeback_control wbc = {
 219                 .sync_mode =  WB_SYNC_ALL,
 220                 .nr_to_write = mapping->nrpages * 2,
 221                 .range_start = 0,
 222                 .range_end = i_size_read(mapping->host),
 223                 .for_writepages = 1,
 224         };
 225
 226         ret = generic_writepages(mapping, &wbc);
 227         return ret;
 228 }
 229
 230 /*
 231  * Submit all the data buffers of inode associated with the transaction to
 232  * disk.
 233  *
 234  * We are in a committing transaction. Therefore no new inode can be added to
 235  * our inode list. We use JI_COMMIT_RUNNING flag to protect inode we currently
 236  * operate on from being released while we write out pages.
 237  */
 238 static int journal_submit_data_buffers(journal_t *journal,
 239                 transaction_t *commit_transaction)
 240 {
 241         struct jbd2_inode *jinode;
 242         int err, ret = 0;
 243         struct address_space *mapping;
 244
 245         spin_lock(&journal->j_list_lock);
 246         list_for_each_entry(jinode, &commit_transaction->t_inode_list, i_list) {
 247                 mapping = jinode->i_vfs_inode->i_mapping;
 248                 jinode->i_flags |= JI_COMMIT_RUNNING;
 249                 spin_unlock(&journal->j_list_lock);
 250                 /*
 251                  * submit the inode data buffers. We use writepage
 252                  * instead of writepages. Because writepages can do
 253                  * block allocation  with delalloc. We need to write
 254                  * only allocated blocks here.
 255                  */
 256                 err = journal_submit_inode_data_buffers(mapping);
 257                 if (!ret)
 258                         ret = err;
 259                 spin_lock(&journal->j_list_lock);
 260                 J_ASSERT(jinode->i_transaction == commit_transaction);
 261                 jinode->i_flags &= ~JI_COMMIT_RUNNING;
 262                 wake_up_bit(&jinode->i_flags, __JI_COMMIT_RUNNING);
 263         }
 264         spin_unlock(&journal->j_list_lock);
 265         return ret;
 266 }
 267
 268 /*
 269  * Wait for data submitted for writeout, refile inodes to proper
 270  * transaction if needed.
 271  *
 272  */
 273 static int journal_finish_inode_data_buffers(journal_t *journal,
 274                 transaction_t *commit_transaction)
 275 {
 276         struct jbd2_inode *jinode, *next_i;
 277         int err, ret = 0;
 278
 279         /* For locking, see the comment in journal_submit_data_buffers() */
 280         spin_lock(&journal->j_list_lock);
 281         list_for_each_entry(jinode, &commit_transaction->t_inode_list, i_list) {
 282                 jinode->i_flags |= JI_COMMIT_RUNNING;
 283                 spin_unlock(&journal->j_list_lock);
 284                 err = filemap_fdatawait(jinode->i_vfs_inode->i_mapping);
 285                 if (err) {
 286                         /*
 287                          * Because AS_EIO is cleared by
 288                          * wait_on_page_writeback_range(), set it again so
 289                          * that user process can get -EIO from fsync().
 290                          */
 291                         set_bit(AS_EIO,
 292                                 &jinode->i_vfs_inode->i_mapping->flags);
 293
 294                         if (!ret)
 295                                 ret = err;
 296                 }
 297                 spin_lock(&journal->j_list_lock);
 298                 jinode->i_flags &= ~JI_COMMIT_RUNNING;
 299                 wake_up_bit(&jinode->i_flags, __JI_COMMIT_RUNNING);
 300         }
 301
 302         /* Now refile inode to proper lists */
 303         list_for_each_entry_safe(jinode, next_i,
 304                                  &commit_transaction->t_inode_list, i_list) {
 305                 list_del(&jinode->i_list);
 306                 if (jinode->i_next_transaction) {
 307                         jinode->i_transaction = jinode->i_next_transaction;
 308                         jinode->i_next_transaction = NULL;
 309                         list_add(&jinode->i_list,
 310                                 &jinode->i_transaction->t_inode_list);
 311                 } else {
 312                         jinode->i_transaction = NULL;
 313                 }
 314         }
 315         spin_unlock(&journal->j_list_lock);
 316
 317         return ret;
 318 }
 319
 320 static __u32 jbd2_checksum_data(__u32 crc32_sum, struct buffer_head *bh)
 321 {
 322         struct page *page = bh->b_page;
 323         char *addr;
 324         __u32 checksum;
 325
 326         addr = kmap_atomic(page, KM_USER0);
 327         checksum = crc32_be(crc32_sum,
 328                 (void *)(addr + offset_in_page(bh->b_data)), bh->b_size);
 329         kunmap_atomic(addr, KM_USER0);
 330
 331         return checksum;
 332 }
 333
 334 static void write_tag_block(int tag_bytes, journal_block_tag_t *tag,
 335                                    unsigned long long block)
 336 {
 337         tag->t_blocknr = cpu_to_be32(block & (u32)~0);
 338         if (tag_bytes > JBD2_TAG_SIZE32)
 339                 tag->t_blocknr_high = cpu_to_be32((block >> 31) >> 1);
 340 }
 341
 342 /*
 343  * jbd2_journal_commit_transaction
 344  *
 345  * The primary function for committing a transaction to the log.  This
 346  * function is called by the journal thread to begin a complete commit.
 347  */
 348 void jbd2_journal_commit_transaction(journal_t *journal)
 349 {
 350         struct transaction_stats_s stats;
 351         transaction_t *commit_transaction;
 352         struct journal_head *jh, *new_jh, *descriptor;
 353         struct buffer_head **wbuf = journal->j_wbuf;
 354         int bufs;
 355         int flags;
 356         int err;
 357         unsigned long long blocknr;
 358         char *tagp = NULL;
 359         journal_header_t *header;
 360         journal_block_tag_t *tag = NULL;
 361         int space_left = 0;
 362         int first_tag = 0;
 363         int tag_flag;
 364         int i;
 365         int tag_bytes = journal_tag_bytes(journal);
 366         struct buffer_head *cbh = NULL; /* For transactional checksums */
 367         __u32 crc32_sum = ~0;
 368
 369         /*
 370          * First job: lock down the current transaction and wait for
 371          * all outstanding updates to complete.
 372          */
 373
 374 #ifdef COMMIT_STATS
 375         spin_lock(&journal->j_list_lock);
 376         summarise_journal_usage(journal);
 377         spin_unlock(&journal->j_list_lock);
 378 #endif
 379
 380         /* Do we need to erase the effects of a prior jbd2_journal_flush? */
 381         if (journal->j_flags & JBD2_FLUSHED) {
 382                 jbd_debug(3, "super block updated\n");
 383                 jbd2_journal_update_superblock(journal, 1);
 384         } else {
 385                 jbd_debug(3, "superblock not updated\n");
 386         }
 387
 388         J_ASSERT(journal->j_running_transaction != NULL);
 389         J_ASSERT(journal->j_committing_transaction == NULL);
 390
 391         commit_transaction = journal->j_running_transaction;
 392         J_ASSERT(commit_transaction->t_state == T_RUNNING);
 393
 394         trace_mark(jbd2_start_commit, "dev %s transaction %d",
 395                    journal->j_devname, commit_transaction->t_tid);
 396         jbd_debug(1, "JBD: starting commit of transaction %d\n",
 397                         commit_transaction->t_tid);
 398
 399         spin_lock(&journal->j_state_lock);
 400         commit_transaction->t_state = T_LOCKED;
 401
 402         stats.u.run.rs_wait = commit_transaction->t_max_wait;
 403         stats.u.run.rs_locked = jiffies;
 404         stats.u.run.rs_running = jbd2_time_diff(commit_transaction->t_start,
 405                                                 stats.u.run.rs_locked);
 406
 407         spin_lock(&commit_transaction->t_handle_lock);
 408         while (commit_transaction->t_updates) {
 409                 DEFINE_WAIT(wait);
 410
 411                 prepare_to_wait(&journal->j_wait_updates, &wait,
 412                                         TASK_UNINTERRUPTIBLE);
 413                 if (commit_transaction->t_updates) {
 414                         spin_unlock(&commit_transaction->t_handle_lock);
 415                         spin_unlock(&journal->j_state_lock);
 416                         schedule();
 417                         spin_lock(&journal->j_state_lock);
 418                         spin_lock(&commit_transaction->t_handle_lock);
 419                 }
 420                 finish_wait(&journal->j_wait_updates, &wait);
 421         }
 422         spin_unlock(&commit_transaction->t_handle_lock);
 423
 424         J_ASSERT (commit_transaction->t_outstanding_credits <=
 425                         journal->j_max_transaction_buffers);
 426
 427         /*
 428          * First thing we are allowed to do is to discard any remaining
 429          * BJ_Reserved buffers.  Note, it is _not_ permissible to assume
 430          * that there are no such buffers: if a large filesystem
 431          * operation like a truncate needs to split itself over multiple
 432          * transactions, then it may try to do a jbd2_journal_restart() while
 433          * there are still BJ_Reserved buffers outstanding.  These must
 434          * be released cleanly from the current transaction.
 435          *
 436          * In this case, the filesystem must still reserve write access
 437          * again before modifying the buffer in the new transaction, but
 438          * we do not require it to remember exactly which old buffers it
 439          * has reserved.  This is consistent with the existing behaviour
 440          * that multiple jbd2_journal_get_write_access() calls to the same
 441          * buffer are perfectly permissable.
 442          */
 443         while (commit_transaction->t_reserved_list) {
 444                 jh = commit_transaction->t_reserved_list;
 445                 JBUFFER_TRACE(jh, "reserved, unused: refile");
 446                 /*
 447                  * A jbd2_journal_get_undo_access()+jbd2_journal_release_buffer() may
 448                  * leave undo-committed data.
 449                  */
 450                 if (jh->b_committed_data) {
 451                         struct buffer_head *bh = jh2bh(jh);
 452
 453                         jbd_lock_bh_state(bh);
 454                         jbd2_free(jh->b_committed_data, bh->b_size);
 455                         jh->b_committed_data = NULL;
 456                         jbd_unlock_bh_state(bh);
 457                 }
 458                 jbd2_journal_refile_buffer(journal, jh);
 459         }
 460
 461         /*
 462          * Now try to drop any written-back buffers from the journal's
 463          * checkpoint lists.  We do this *before* commit because it potentially
 464          * frees some memory
 465          */
 466         spin_lock(&journal->j_list_lock);
 467         __jbd2_journal_clean_checkpoint_list(journal);
 468         spin_unlock(&journal->j_list_lock);
 469
 470         jbd_debug (3, "JBD: commit phase 1\n");
 471
 472         /*
 473          * Switch to a new revoke table.
 474          */
 475         jbd2_journal_switch_revoke_table(journal);
 476
 477         stats.u.run.rs_flushing = jiffies;
 478         stats.u.run.rs_locked = jbd2_time_diff(stats.u.run.rs_locked,
 479                                                stats.u.run.rs_flushing);
 480
 481         commit_transaction->t_state = T_FLUSH;
 482         journal->j_committing_transaction = commit_transaction;
 483         journal->j_running_transaction = NULL;
 484         commit_transaction->t_log_start = journal->j_head;
 485         wake_up(&journal->j_wait_transaction_locked);
 486         spin_unlock(&journal->j_state_lock);
 487
 488         jbd_debug (3, "JBD: commit phase 2\n");
 489
 490         /*
 491          * Now start flushing things to disk, in the order they appear
 492          * on the transaction lists.  Data blocks go first.
 493          */
 494         err = journal_submit_data_buffers(journal, commit_transaction);
 495         if (err)
 496                 jbd2_journal_abort(journal, err);
 497
 498         jbd2_journal_write_revoke_records(journal, commit_transaction);
 499
 500         jbd_debug(3, "JBD: commit phase 2\n");
 501
 502         /*
 503          * Way to go: we have now written out all of the data for a
 504          * transaction!  Now comes the tricky part: we need to write out
 505          * metadata.  Loop over the transaction's entire buffer list:
 506          */
 507         spin_lock(&journal->j_state_lock);
 508         commit_transaction->t_state = T_COMMIT;
 509         spin_unlock(&journal->j_state_lock);
 510
 511         stats.u.run.rs_logging = jiffies;
 512         stats.u.run.rs_flushing = jbd2_time_diff(stats.u.run.rs_flushing,
 513                                                  stats.u.run.rs_logging);
 514         stats.u.run.rs_blocks = commit_transaction->t_outstanding_credits;
 515         stats.u.run.rs_blocks_logged = 0;
 516
 517         J_ASSERT(commit_transaction->t_nr_buffers <=
 518                  commit_transaction->t_outstanding_credits);
 519
 520         err = 0;
 521         descriptor = NULL;
 522         bufs = 0;
 523         while (commit_transaction->t_buffers) {
 524
 525                 /* Find the next buffer to be journaled... */
 526
 527                 jh = commit_transaction->t_buffers;
 528
 529                 /* If we're in abort mode, we just un-journal the buffer and
 530                    release it. */
 531
 532                 if (is_journal_aborted(journal)) {
 533                         clear_buffer_jbddirty(jh2bh(jh));
 534                         JBUFFER_TRACE(jh, "journal is aborting: refile");
 535                         jbd2_journal_refile_buffer(journal, jh);
 536                         /* If that was the last one, we need to clean up
 537                          * any descriptor buffers which may have been
 538                          * already allocated, even if we are now
 539                          * aborting. */
 540                         if (!commit_transaction->t_buffers)
 541                                 goto start_journal_io;
 542                         continue;
 543                 }
 544
 545                 /* Make sure we have a descriptor block in which to
 546                    record the metadata buffer. */
 547
 548                 if (!descriptor) {
 549                         struct buffer_head *bh;
 550
 551                         J_ASSERT (bufs == 0);
 552
 553                         jbd_debug(4, "JBD: get descriptor\n");
 554
 555                         descriptor = jbd2_journal_get_descriptor_buffer(journal);
 556                         if (!descriptor) {
 557                                 jbd2_journal_abort(journal, -EIO);
 558                                 continue;
 559                         }
 560
 561                         bh = jh2bh(descriptor);
 562                         jbd_debug(4, "JBD: got buffer %llu (%p)\n",
 563                                 (unsigned long long)bh->b_blocknr, bh->b_data);
 564                         header = (journal_header_t *)&bh->b_data[0];
 565                         header->h_magic     = cpu_to_be32(JBD2_MAGIC_NUMBER);
 566                         header->h_blocktype = cpu_to_be32(JBD2_DESCRIPTOR_BLOCK);
 567                         header->h_sequence  = cpu_to_be32(commit_transaction->t_tid);
 568
 569                         tagp = &bh->b_data[sizeof(journal_header_t)];
 570                         space_left = bh->b_size - sizeof(journal_header_t);
 571                         first_tag = 1;
 572                         set_buffer_jwrite(bh);
 573                         set_buffer_dirty(bh);
 574                         wbuf[bufs++] = bh;
 575
 576                         /* Record it so that we can wait for IO
 577                            completion later */
 578                         BUFFER_TRACE(bh, "ph3: file as descriptor");
 579                         jbd2_journal_file_buffer(descriptor, commit_transaction,
 580                                         BJ_LogCtl);
 581                 }
 582
 583                 /* Where is the buffer to be written? */
 584
 585                 err = jbd2_journal_next_log_block(journal, &blocknr);
 586                 /* If the block mapping failed, just abandon the buffer
 587                    and repeat this loop: we'll fall into the
 588                    refile-on-abort condition above. */
 589                 if (err) {
 590                         jbd2_journal_abort(journal, err);
 591                         continue;
 592                 }
 593
 594                 /*
 595                  * start_this_handle() uses t_outstanding_credits to determine
 596                  * the free space in the log, but this counter is changed
 597                  * by jbd2_journal_next_log_block() also.
 598                  */
 599                 commit_transaction->t_outstanding_credits--;
 600
 601                 /* Bump b_count to prevent truncate from stumbling over
 602                    the shadowed buffer!  @@@ This can go if we ever get
 603                    rid of the BJ_IO/BJ_Shadow pairing of buffers. */
 604                 atomic_inc(&jh2bh(jh)->b_count);
 605
 606                 /* Make a temporary IO buffer with which to write it out
 607                    (this will requeue both the metadata buffer and the
 608                    temporary IO buffer). new_bh goes on BJ_IO*/
 609
 610                 set_bit(BH_JWrite, &jh2bh(jh)->b_state);
 611                 /*
 612                  * akpm: jbd2_journal_write_metadata_buffer() sets
 613                  * new_bh->b_transaction to commit_transaction.
 614                  * We need to clean this up before we release new_bh
 615                  * (which is of type BJ_IO)
 616                  */
 617                 JBUFFER_TRACE(jh, "ph3: write metadata");
 618                 flags = jbd2_journal_write_metadata_buffer(commit_transaction,
 619                                                       jh, &new_jh, blocknr);
 620                 set_bit(BH_JWrite, &jh2bh(new_jh)->b_state);
 621                 wbuf[bufs++] = jh2bh(new_jh);
 622
 623                 /* Record the new block's tag in the current descriptor
 624                    buffer */
 625
 626                 tag_flag = 0;
 627                 if (flags & 1)
 628                         tag_flag |= JBD2_FLAG_ESCAPE;
 629                 if (!first_tag)
 630                         tag_flag |= JBD2_FLAG_SAME_UUID;
 631
 632                 tag = (journal_block_tag_t *) tagp;
 633                 write_tag_block(tag_bytes, tag, jh2bh(jh)->b_blocknr);
 634                 tag->t_flags = cpu_to_be32(tag_flag);
 635                 tagp += tag_bytes;
 636                 space_left -= tag_bytes;
 637
 638                 if (first_tag) {
 639                         memcpy (tagp, journal->j_uuid, 16);
 640                         tagp += 16;
 641                         space_left -= 16;
 642                         first_tag = 0;
 643                 }
 644
 645                 /* If there's no more to do, or if the descriptor is full,
 646                    let the IO rip! */
 647
 648                 if (bufs == journal->j_wbufsize ||
 649                     commit_transaction->t_buffers == NULL ||
 650                     space_left < tag_bytes + 16) {
 651
 652                         jbd_debug(4, "JBD: Submit %d IOs\n", bufs);
 653
 654                         /* Write an end-of-descriptor marker before
 655                            submitting the IOs.  "tag" still points to
 656                            the last tag we set up. */
 657
 658                         tag->t_flags |= cpu_to_be32(JBD2_FLAG_LAST_TAG);
 659
 660 start_journal_io:
 661                         for (i = 0; i < bufs; i++) {
 662                                 struct buffer_head *bh = wbuf[i];
 663                                 /*
 664                                  * Compute checksum.
 665                                  */
 666                                 if (JBD2_HAS_COMPAT_FEATURE(journal,
 667                                         JBD2_FEATURE_COMPAT_CHECKSUM)) {
 668                                         crc32_sum =
 669                                             jbd2_checksum_data(crc32_sum, bh);
 670                                 }
 671
 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, bh);
 677                         }
 678                         cond_resched();
 679                         stats.u.run.rs_blocks_logged += bufs;
 680
 681                         /* Force a new descriptor to be generated next
 682                            time round the loop. */
 683                         descriptor = NULL;
 684                         bufs = 0;
 685                 }
 686         }
 687
 688         /* Done it all: now write the commit record asynchronously. */
 689
 690         if (JBD2_HAS_INCOMPAT_FEATURE(journal,
 691                 JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
 692                 err = journal_submit_commit_record(journal, commit_transaction,
 693                                                  &cbh, crc32_sum);
 694                 if (err)
 695                         __jbd2_journal_abort_hard(journal);
 696         }
 697
 698         /*
 699          * This is the right place to wait for data buffers both for ASYNC
 700          * and !ASYNC commit. If commit is ASYNC, we need to wait only after
 701          * the commit block went to disk (which happens above). If commit is
 702          * SYNC, we need to wait for data buffers before we start writing
 703          * commit block, which happens below in such setting.
 704          */
 705         err = journal_finish_inode_data_buffers(journal, commit_transaction);
 706         if (err) {
 707                 printk(KERN_WARNING
 708                         "JBD2: Detected IO errors while flushing file data "
 709                        "on %s\n", journal->j_devname);
 710                 if (journal->j_flags & JBD2_ABORT_ON_SYNCDATA_ERR)
 711                         jbd2_journal_abort(journal, err);
 712                 err = 0;
 713         }
 714
 715         /* Lo and behold: we have just managed to send a transaction to
 716            the log.  Before we can commit it, wait for the IO so far to
 717            complete.  Control buffers being written are on the
 718            transaction's t_log_list queue, and metadata buffers are on
 719            the t_iobuf_list queue.
 720
 721            Wait for the buffers in reverse order.  That way we are
 722            less likely to be woken up until all IOs have completed, and
 723            so we incur less scheduling load.
 724         */
 725
 726         jbd_debug(3, "JBD: commit phase 3\n");
 727
 728         /*
 729          * akpm: these are BJ_IO, and j_list_lock is not needed.
 730          * See __journal_try_to_free_buffer.
 731          */
 732 wait_for_iobuf:
 733         while (commit_transaction->t_iobuf_list != NULL) {
 734                 struct buffer_head *bh;
 735
 736                 jh = commit_transaction->t_iobuf_list->b_tprev;
 737                 bh = jh2bh(jh);
 738                 if (buffer_locked(bh)) {
 739                         wait_on_buffer(bh);
 740                         goto wait_for_iobuf;
 741                 }
 742                 if (cond_resched())
 743                         goto wait_for_iobuf;
 744
 745                 if (unlikely(!buffer_uptodate(bh)))
 746                         err = -EIO;
 747
 748                 clear_buffer_jwrite(bh);
 749
 750                 JBUFFER_TRACE(jh, "ph4: unfile after journal write");
 751                 jbd2_journal_unfile_buffer(journal, jh);
 752
 753                 /*
 754                  * ->t_iobuf_list should contain only dummy buffer_heads
 755                  * which were created by jbd2_journal_write_metadata_buffer().
 756                  */
 757                 BUFFER_TRACE(bh, "dumping temporary bh");
 758                 jbd2_journal_put_journal_head(jh);
 759                 __brelse(bh);
 760                 J_ASSERT_BH(bh, atomic_read(&bh->b_count) == 0);
 761                 free_buffer_head(bh);
 762
 763                 /* We also have to unlock and free the corresponding
 764                    shadowed buffer */
 765                 jh = commit_transaction->t_shadow_list->b_tprev;
 766                 bh = jh2bh(jh);
 767                 clear_bit(BH_JWrite, &bh->b_state);
 768                 J_ASSERT_BH(bh, buffer_jbddirty(bh));
 769
 770                 /* The metadata is now released for reuse, but we need
 771                    to remember it against this transaction so that when
 772                    we finally commit, we can do any checkpointing
 773                    required. */
 774                 JBUFFER_TRACE(jh, "file as BJ_Forget");
 775                 jbd2_journal_file_buffer(jh, commit_transaction, BJ_Forget);
 776                 /* Wake up any transactions which were waiting for this
 777                    IO to complete */
 778                 wake_up_bit(&bh->b_state, BH_Unshadow);
 779                 JBUFFER_TRACE(jh, "brelse shadowed buffer");
 780                 __brelse(bh);
 781         }
 782
 783         J_ASSERT (commit_transaction->t_shadow_list == NULL);
 784
 785         jbd_debug(3, "JBD: commit phase 4\n");
 786
 787         /* Here we wait for the revoke record and descriptor record buffers */
 788  wait_for_ctlbuf:
 789         while (commit_transaction->t_log_list != NULL) {
 790                 struct buffer_head *bh;
 791
 792                 jh = commit_transaction->t_log_list->b_tprev;
 793                 bh = jh2bh(jh);
 794                 if (buffer_locked(bh)) {
 795                         wait_on_buffer(bh);
 796                         goto wait_for_ctlbuf;
 797                 }
 798                 if (cond_resched())
 799                         goto wait_for_ctlbuf;
 800
 801                 if (unlikely(!buffer_uptodate(bh)))
 802                         err = -EIO;
 803
 804                 BUFFER_TRACE(bh, "ph5: control buffer writeout done: unfile");
 805                 clear_buffer_jwrite(bh);
 806                 jbd2_journal_unfile_buffer(journal, jh);
 807                 jbd2_journal_put_journal_head(jh);
 808                 __brelse(bh);           /* One for getblk */
 809                 /* AKPM: bforget here */
 810         }
 811
 812         if (err)
 813                 jbd2_journal_abort(journal, err);
 814
 815         jbd_debug(3, "JBD: commit phase 5\n");
 816
 817         if (!JBD2_HAS_INCOMPAT_FEATURE(journal,
 818                 JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
 819                 err = journal_submit_commit_record(journal, commit_transaction,
 820                                                 &cbh, crc32_sum);
 821                 if (err)
 822                         __jbd2_journal_abort_hard(journal);
 823         }
 824         if (!err && !is_journal_aborted(journal))
 825                 err = journal_wait_on_commit_record(journal, cbh);
 826
 827         if (err)
 828                 jbd2_journal_abort(journal, err);
 829
 830         /* End of a transaction!  Finally, we can do checkpoint
 831            processing: any buffers committed as a result of this
 832            transaction can be removed from any checkpoint list it was on
 833            before. */
 834
 835         jbd_debug(3, "JBD: commit phase 6\n");
 836
 837         J_ASSERT(list_empty(&commit_transaction->t_inode_list));
 838         J_ASSERT(commit_transaction->t_buffers == NULL);
 839         J_ASSERT(commit_transaction->t_checkpoint_list == NULL);
 840         J_ASSERT(commit_transaction->t_iobuf_list == NULL);
 841         J_ASSERT(commit_transaction->t_shadow_list == NULL);
 842         J_ASSERT(commit_transaction->t_log_list == NULL);
 843
 844 restart_loop:
 845         /*
 846          * As there are other places (journal_unmap_buffer()) adding buffers
 847          * to this list we have to be careful and hold the j_list_lock.
 848          */
 849         spin_lock(&journal->j_list_lock);
 850         while (commit_transaction->t_forget) {
 851                 transaction_t *cp_transaction;
 852                 struct buffer_head *bh;
 853
 854                 jh = commit_transaction->t_forget;
 855                 spin_unlock(&journal->j_list_lock);
 856                 bh = jh2bh(jh);
 857                 jbd_lock_bh_state(bh);
 858                 J_ASSERT_JH(jh, jh->b_transaction == commit_transaction ||
 859                         jh->b_transaction == journal->j_running_transaction);
 860
 861                 /*
 862                  * If there is undo-protected committed data against
 863                  * this buffer, then we can remove it now.  If it is a
 864                  * buffer needing such protection, the old frozen_data
 865                  * field now points to a committed version of the
 866                  * buffer, so rotate that field to the new committed
 867                  * data.
 868                  *
 869                  * Otherwise, we can just throw away the frozen data now.
 870                  */
 871                 if (jh->b_committed_data) {
 872                         jbd2_free(jh->b_committed_data, bh->b_size);
 873                         jh->b_committed_data = NULL;
 874                         if (jh->b_frozen_data) {
 875                                 jh->b_committed_data = jh->b_frozen_data;
 876                                 jh->b_frozen_data = NULL;
 877                         }
 878                 } else if (jh->b_frozen_data) {
 879                         jbd2_free(jh->b_frozen_data, bh->b_size);
 880                         jh->b_frozen_data = NULL;
 881                 }
 882
 883                 spin_lock(&journal->j_list_lock);
 884                 cp_transaction = jh->b_cp_transaction;
 885                 if (cp_transaction) {
 886                         JBUFFER_TRACE(jh, "remove from old cp transaction");
 887                         cp_transaction->t_chp_stats.cs_dropped++;
 888                         __jbd2_journal_remove_checkpoint(jh);
 889                 }
 890
 891                 /* Only re-checkpoint the buffer_head if it is marked
 892                  * dirty.  If the buffer was added to the BJ_Forget list
 893                  * by jbd2_journal_forget, it may no longer be dirty and
 894                  * there's no point in keeping a checkpoint record for
 895                  * it. */
 896
 897                 /* A buffer which has been freed while still being
 898                  * journaled by a previous transaction may end up still
 899                  * being dirty here, but we want to avoid writing back
 900                  * that buffer in the future now that the last use has
 901                  * been committed.  That's not only a performance gain,
 902                  * it also stops aliasing problems if the buffer is left
 903                  * behind for writeback and gets reallocated for another
 904                  * use in a different page. */
 905                 if (buffer_freed(bh)) {
 906                         clear_buffer_freed(bh);
 907                         clear_buffer_jbddirty(bh);
 908                 }
 909
 910                 if (buffer_jbddirty(bh)) {
 911                         JBUFFER_TRACE(jh, "add to new checkpointing trans");
 912                         __jbd2_journal_insert_checkpoint(jh, commit_transaction);
 913                         if (is_journal_aborted(journal))
 914                                 clear_buffer_jbddirty(bh);
 915                         JBUFFER_TRACE(jh, "refile for checkpoint writeback");
 916                         __jbd2_journal_refile_buffer(jh);
 917                         jbd_unlock_bh_state(bh);
 918                 } else {
 919                         J_ASSERT_BH(bh, !buffer_dirty(bh));
 920                         /* The buffer on BJ_Forget list and not jbddirty means
 921                          * it has been freed by this transaction and hence it
 922                          * could not have been reallocated until this
 923                          * transaction has committed. *BUT* it could be
 924                          * reallocated once we have written all the data to
 925                          * disk and before we process the buffer on BJ_Forget
 926                          * list. */
 927                         JBUFFER_TRACE(jh, "refile or unfile freed buffer");
 928                         __jbd2_journal_refile_buffer(jh);
 929                         if (!jh->b_transaction) {
 930                                 jbd_unlock_bh_state(bh);
 931                                  /* needs a brelse */
 932                                 jbd2_journal_remove_journal_head(bh);
 933                                 release_buffer_page(bh);
 934                         } else
 935                                 jbd_unlock_bh_state(bh);
 936                 }
 937                 cond_resched_lock(&journal->j_list_lock);
 938         }
 939         spin_unlock(&journal->j_list_lock);
 940         /*
 941          * This is a bit sleazy.  We use j_list_lock to protect transition
 942          * of a transaction into T_FINISHED state and calling
 943          * __jbd2_journal_drop_transaction(). Otherwise we could race with
 944          * other checkpointing code processing the transaction...
 945          */
 946         spin_lock(&journal->j_state_lock);
 947         spin_lock(&journal->j_list_lock);
 948         /*
 949          * Now recheck if some buffers did not get attached to the transaction
 950          * while the lock was dropped...
 951          */
 952         if (commit_transaction->t_forget) {
 953                 spin_unlock(&journal->j_list_lock);
 954                 spin_unlock(&journal->j_state_lock);
 955                 goto restart_loop;
 956         }
 957
 958         /* Done with this transaction! */
 959
 960         jbd_debug(3, "JBD: commit phase 7\n");
 961
 962         J_ASSERT(commit_transaction->t_state == T_COMMIT);
 963
 964         commit_transaction->t_start = jiffies;
 965         stats.u.run.rs_logging = jbd2_time_diff(stats.u.run.rs_logging,
 966                                                 commit_transaction->t_start);
 967
 968         /*
 969          * File the transaction for history
 970          */
 971         stats.ts_type = JBD2_STATS_RUN;
 972         stats.ts_tid = commit_transaction->t_tid;
 973         stats.u.run.rs_handle_count = commit_transaction->t_handle_count;
 974         spin_lock(&journal->j_history_lock);
 975         memcpy(journal->j_history + journal->j_history_cur, &stats,
 976                         sizeof(stats));
 977         if (++journal->j_history_cur == journal->j_history_max)
 978                 journal->j_history_cur = 0;
 979
 980         /*
 981          * Calculate overall stats
 982          */
 983         journal->j_stats.ts_tid++;
 984         journal->j_stats.u.run.rs_wait += stats.u.run.rs_wait;
 985         journal->j_stats.u.run.rs_running += stats.u.run.rs_running;
 986         journal->j_stats.u.run.rs_locked += stats.u.run.rs_locked;
 987         journal->j_stats.u.run.rs_flushing += stats.u.run.rs_flushing;
 988         journal->j_stats.u.run.rs_logging += stats.u.run.rs_logging;
 989         journal->j_stats.u.run.rs_handle_count += stats.u.run.rs_handle_count;
 990         journal->j_stats.u.run.rs_blocks += stats.u.run.rs_blocks;
 991         journal->j_stats.u.run.rs_blocks_logged += stats.u.run.rs_blocks_logged;
 992         spin_unlock(&journal->j_history_lock);
 993
 994         commit_transaction->t_state = T_FINISHED;
 995         J_ASSERT(commit_transaction == journal->j_committing_transaction);
 996         journal->j_commit_sequence = commit_transaction->t_tid;
 997         journal->j_committing_transaction = NULL;
 998         spin_unlock(&journal->j_state_lock);
 999
1000         if (journal->j_commit_callback)
1001                 journal->j_commit_callback(journal, commit_transaction);
1002
1003         if (commit_transaction->t_checkpoint_list == NULL &&
1004             commit_transaction->t_checkpoint_io_list == NULL) {
1005                 __jbd2_journal_drop_transaction(journal, commit_transaction);
1006         } else {
1007                 if (journal->j_checkpoint_transactions == NULL) {
1008                         journal->j_checkpoint_transactions = commit_transaction;
1009                         commit_transaction->t_cpnext = commit_transaction;
1010                         commit_transaction->t_cpprev = commit_transaction;
1011                 } else {
1012                         commit_transaction->t_cpnext =
1013                                 journal->j_checkpoint_transactions;
1014                         commit_transaction->t_cpprev =
1015                                 commit_transaction->t_cpnext->t_cpprev;
1016                         commit_transaction->t_cpnext->t_cpprev =
1017                                 commit_transaction;
1018                         commit_transaction->t_cpprev->t_cpnext =
1019                                 commit_transaction;
1020                 }
1021         }
1022         spin_unlock(&journal->j_list_lock);
1023
1024         trace_mark(jbd2_end_commit, "dev %s transaction %d head %d",
1025                    journal->j_devname, journal->j_commit_sequence,
1026                    journal->j_tail_sequence);
1027         jbd_debug(1, "JBD: commit %d complete, head %d\n",
1028                   journal->j_commit_sequence, journal->j_tail_sequence);
1029
1030         wake_up(&journal->j_wait_done_commit);
1031 }