fs/jbd2/commit.c

   1 // SPDX-License-Identifier: GPL-2.0+
   2 /*
   3  * linux/fs/jbd2/commit.c
   4  *
   5  * Written by Stephen C. Tweedie <sct@redhat.com>, 1998
   6  *
   7  * Copyright 1998 Red Hat corp --- All Rights Reserved
   8  *
   9  * Journal commit routines for the generic filesystem journaling code;
  10  * part of the ext2fs journaling system.
  11  */
  12
  13 #include <linux/time.h>
  14 #include <linux/fs.h>
  15 #include <linux/jbd2.h>
  16 #include <linux/errno.h>
  17 #include <linux/slab.h>
  18 #include <linux/mm.h>
  19 #include <linux/pagemap.h>
  20 #include <linux/jiffies.h>
  21 #include <linux/crc32.h>
  22 #include <linux/writeback.h>
  23 #include <linux/backing-dev.h>
  24 #include <linux/bio.h>
  25 #include <linux/blkdev.h>
  26 #include <linux/bitops.h>
  27 #include <trace/events/jbd2.h>
  28
  29 /*
  30  * IO end handler for temporary buffer_heads handling writes to the journal.
  31  */
  32 static void journal_end_buffer_io_sync(struct buffer_head *bh, int uptodate)
  33 {
  34         struct buffer_head *orig_bh = bh->b_private;
  35
  36         BUFFER_TRACE(bh, "");
  37         if (uptodate)
  38                 set_buffer_uptodate(bh);
  39         else
  40                 clear_buffer_uptodate(bh);
  41         if (orig_bh) {
  42                 clear_bit_unlock(BH_Shadow, &orig_bh->b_state);
  43                 smp_mb__after_atomic();
  44                 wake_up_bit(&orig_bh->b_state, BH_Shadow);
  45         }
  46         unlock_buffer(bh);
  47 }
  48
  49 /*
  50  * When an ext4 file is truncated, it is possible that some pages are not
  51  * successfully freed, because they are attached to a committing transaction.
  52  * After the transaction commits, these pages are left on the LRU, with no
  53  * ->mapping, and with attached buffers.  These pages are trivially reclaimable
  54  * by the VM, but their apparent absence upsets the VM accounting, and it makes
  55  * the numbers in /proc/meminfo look odd.
  56  *
  57  * So here, we have a buffer which has just come off the forget list.  Look to
  58  * see if we can strip all buffers from the backing page.
  59  *
  60  * Called under lock_journal(), and possibly under journal_datalist_lock.  The
  61  * caller provided us with a ref against the buffer, and we drop that here.
  62  */
  63 static void release_buffer_page(struct buffer_head *bh)
  64 {
  65         struct page *page;
  66
  67         if (buffer_dirty(bh))
  68                 goto nope;
  69         if (atomic_read(&bh->b_count) != 1)
  70                 goto nope;
  71         page = bh->b_page;
  72         if (!page)
  73                 goto nope;
  74         if (page->mapping)
  75                 goto nope;
  76
  77         /* OK, it's a truncated page */
  78         if (!trylock_page(page))
  79                 goto nope;
  80
  81         get_page(page);
  82         __brelse(bh);
  83         try_to_free_buffers(page);
  84         unlock_page(page);
  85         put_page(page);
  86         return;
  87
  88 nope:
  89         __brelse(bh);
  90 }
  91
  92 static void jbd2_commit_block_csum_set(journal_t *j, struct buffer_head *bh)
  93 {
  94         struct commit_header *h;
  95         __u32 csum;
  96
  97         if (!jbd2_journal_has_csum_v2or3(j))
  98                 return;
  99
 100         h = (struct commit_header *)(bh->b_data);
 101         h->h_chksum_type = 0;
 102         h->h_chksum_size = 0;
 103         h->h_chksum[0] = 0;
 104         csum = jbd2_chksum(j, j->j_csum_seed, bh->b_data, j->j_blocksize);
 105         h->h_chksum[0] = cpu_to_be32(csum);
 106 }
 107
 108 /*
 109  * Done it all: now submit the commit record.  We should have
 110  * cleaned up our previous buffers by now, so if we are in abort
 111  * mode we can now just skip the rest of the journal write
 112  * entirely.
 113  *
 114  * Returns 1 if the journal needs to be aborted or 0 on success
 115  */
 116 static int journal_submit_commit_record(journal_t *journal,
 117                                         transaction_t *commit_transaction,
 118                                         struct buffer_head **cbh,
 119                                         __u32 crc32_sum)
 120 {
 121         struct commit_header *tmp;
 122         struct buffer_head *bh;
 123         int ret;
 124         struct timespec64 now;
 125
 126         *cbh = NULL;
 127
 128         if (is_journal_aborted(journal))
 129                 return 0;
 130
 131         bh = jbd2_journal_get_descriptor_buffer(commit_transaction,
 132                                                 JBD2_COMMIT_BLOCK);
 133         if (!bh)
 134                 return 1;
 135
 136         tmp = (struct commit_header *)bh->b_data;
 137         ktime_get_coarse_real_ts64(&now);
 138         tmp->h_commit_sec = cpu_to_be64(now.tv_sec);
 139         tmp->h_commit_nsec = cpu_to_be32(now.tv_nsec);
 140
 141         if (jbd2_has_feature_checksum(journal)) {
 142                 tmp->h_chksum_type      = JBD2_CRC32_CHKSUM;
 143                 tmp->h_chksum_size      = JBD2_CRC32_CHKSUM_SIZE;
 144                 tmp->h_chksum[0]        = cpu_to_be32(crc32_sum);
 145         }
 146         jbd2_commit_block_csum_set(journal, bh);
 147
 148         BUFFER_TRACE(bh, "submit commit block");
 149         lock_buffer(bh);
 150         clear_buffer_dirty(bh);
 151         set_buffer_uptodate(bh);
 152         bh->b_end_io = journal_end_buffer_io_sync;
 153
 154         if (journal->j_flags & JBD2_BARRIER &&
 155             !jbd2_has_feature_async_commit(journal))
 156                 ret = submit_bh(REQ_OP_WRITE,
 157                         REQ_SYNC | REQ_PREFLUSH | REQ_FUA, bh);
 158         else
 159                 ret = submit_bh(REQ_OP_WRITE, REQ_SYNC, bh);
 160
 161         *cbh = bh;
 162         return ret;
 163 }
 164
 165 /*
 166  * This function along with journal_submit_commit_record
 167  * allows to write the commit record asynchronously.
 168  */
 169 static int journal_wait_on_commit_record(journal_t *journal,
 170                                          struct buffer_head *bh)
 171 {
 172         int ret = 0;
 173
 174         clear_buffer_dirty(bh);
 175         wait_on_buffer(bh);
 176
 177         if (unlikely(!buffer_uptodate(bh)))
 178                 ret = -EIO;
 179         put_bh(bh);            /* One for getblk() */
 180
 181         return ret;
 182 }
 183
 184 /*
 185  * write the filemap data using writepage() address_space_operations.
 186  * We don't do block allocation here even for delalloc. We don't
 187  * use writepages() because with dealyed allocation we may be doing
 188  * block allocation in writepages().
 189  */
 190 static int journal_submit_inode_data_buffers(struct address_space *mapping)
 191 {
 192         int ret;
 193         struct writeback_control wbc = {
 194                 .sync_mode =  WB_SYNC_ALL,
 195                 .nr_to_write = mapping->nrpages * 2,
 196                 .range_start = 0,
 197                 .range_end = i_size_read(mapping->host),
 198         };
 199
 200         ret = generic_writepages(mapping, &wbc);
 201         return ret;
 202 }
 203
 204 /*
 205  * Submit all the data buffers of inode associated with the transaction to
 206  * disk.
 207  *
 208  * We are in a committing transaction. Therefore no new inode can be added to
 209  * our inode list. We use JI_COMMIT_RUNNING flag to protect inode we currently
 210  * operate on from being released while we write out pages.
 211  */
 212 static int journal_submit_data_buffers(journal_t *journal,
 213                 transaction_t *commit_transaction)
 214 {
 215         struct jbd2_inode *jinode;
 216         int err, ret = 0;
 217         struct address_space *mapping;
 218
 219         spin_lock(&journal->j_list_lock);
 220         list_for_each_entry(jinode, &commit_transaction->t_inode_list, i_list) {
 221                 if (!(jinode->i_flags & JI_WRITE_DATA))
 222                         continue;
 223                 mapping = jinode->i_vfs_inode->i_mapping;
 224                 jinode->i_flags |= JI_COMMIT_RUNNING;
 225                 spin_unlock(&journal->j_list_lock);
 226                 /*
 227                  * submit the inode data buffers. We use writepage
 228                  * instead of writepages. Because writepages can do
 229                  * block allocation  with delalloc. We need to write
 230                  * only allocated blocks here.
 231                  */
 232                 trace_jbd2_submit_inode_data(jinode->i_vfs_inode);
 233                 err = journal_submit_inode_data_buffers(mapping);
 234                 if (!ret)
 235                         ret = err;
 236                 spin_lock(&journal->j_list_lock);
 237                 J_ASSERT(jinode->i_transaction == commit_transaction);
 238                 jinode->i_flags &= ~JI_COMMIT_RUNNING;
 239                 smp_mb();
 240                 wake_up_bit(&jinode->i_flags, __JI_COMMIT_RUNNING);
 241         }
 242         spin_unlock(&journal->j_list_lock);
 243         return ret;
 244 }
 245
 246 /*
 247  * Wait for data submitted for writeout, refile inodes to proper
 248  * transaction if needed.
 249  *
 250  */
 251 static int journal_finish_inode_data_buffers(journal_t *journal,
 252                 transaction_t *commit_transaction)
 253 {
 254         struct jbd2_inode *jinode, *next_i;
 255         int err, ret = 0;
 256
 257         /* For locking, see the comment in journal_submit_data_buffers() */
 258         spin_lock(&journal->j_list_lock);
 259         list_for_each_entry(jinode, &commit_transaction->t_inode_list, i_list) {
 260                 if (!(jinode->i_flags & JI_WAIT_DATA))
 261                         continue;
 262                 jinode->i_flags |= JI_COMMIT_RUNNING;
 263                 spin_unlock(&journal->j_list_lock);
 264                 err = filemap_fdatawait_keep_errors(
 265                                 jinode->i_vfs_inode->i_mapping);
 266                 if (!ret)
 267                         ret = err;
 268                 spin_lock(&journal->j_list_lock);
 269                 jinode->i_flags &= ~JI_COMMIT_RUNNING;
 270                 smp_mb();
 271                 wake_up_bit(&jinode->i_flags, __JI_COMMIT_RUNNING);
 272         }
 273
 274         /* Now refile inode to proper lists */
 275         list_for_each_entry_safe(jinode, next_i,
 276                                  &commit_transaction->t_inode_list, i_list) {
 277                 list_del(&jinode->i_list);
 278                 if (jinode->i_next_transaction) {
 279                         jinode->i_transaction = jinode->i_next_transaction;
 280                         jinode->i_next_transaction = NULL;
 281                         list_add(&jinode->i_list,
 282                                 &jinode->i_transaction->t_inode_list);
 283                 } else {
 284                         jinode->i_transaction = NULL;
 285                 }
 286         }
 287         spin_unlock(&journal->j_list_lock);
 288
 289         return ret;
 290 }
 291
 292 static __u32 jbd2_checksum_data(__u32 crc32_sum, struct buffer_head *bh)
 293 {
 294         struct page *page = bh->b_page;
 295         char *addr;
 296         __u32 checksum;
 297
 298         addr = kmap_atomic(page);
 299         checksum = crc32_be(crc32_sum,
 300                 (void *)(addr + offset_in_page(bh->b_data)), bh->b_size);
 301         kunmap_atomic(addr);
 302
 303         return checksum;
 304 }
 305
 306 static void write_tag_block(journal_t *j, journal_block_tag_t *tag,
 307                                    unsigned long long block)
 308 {
 309         tag->t_blocknr = cpu_to_be32(block & (u32)~0);
 310         if (jbd2_has_feature_64bit(j))
 311                 tag->t_blocknr_high = cpu_to_be32((block >> 31) >> 1);
 312 }
 313
 314 static void jbd2_block_tag_csum_set(journal_t *j, journal_block_tag_t *tag,
 315                                     struct buffer_head *bh, __u32 sequence)
 316 {
 317         journal_block_tag3_t *tag3 = (journal_block_tag3_t *)tag;
 318         struct page *page = bh->b_page;
 319         __u8 *addr;
 320         __u32 csum32;
 321         __be32 seq;
 322
 323         if (!jbd2_journal_has_csum_v2or3(j))
 324                 return;
 325
 326         seq = cpu_to_be32(sequence);
 327         addr = kmap_atomic(page);
 328         csum32 = jbd2_chksum(j, j->j_csum_seed, (__u8 *)&seq, sizeof(seq));
 329         csum32 = jbd2_chksum(j, csum32, addr + offset_in_page(bh->b_data),
 330                              bh->b_size);
 331         kunmap_atomic(addr);
 332
 333         if (jbd2_has_feature_csum3(j))
 334                 tag3->t_checksum = cpu_to_be32(csum32);
 335         else
 336                 tag->t_checksum = cpu_to_be16(csum32);
 337 }
 338 /*
 339  * jbd2_journal_commit_transaction
 340  *
 341  * The primary function for committing a transaction to the log.  This
 342  * function is called by the journal thread to begin a complete commit.
 343  */
 344 void jbd2_journal_commit_transaction(journal_t *journal)
 345 {
 346         struct transaction_stats_s stats;
 347         transaction_t *commit_transaction;
 348         struct journal_head *jh;
 349         struct buffer_head *descriptor;
 350         struct buffer_head **wbuf = journal->j_wbuf;
 351         int bufs;
 352         int flags;
 353         int err;
 354         unsigned long long blocknr;
 355         ktime_t start_time;
 356         u64 commit_time;
 357         char *tagp = NULL;
 358         journal_block_tag_t *tag = NULL;
 359         int space_left = 0;
 360         int first_tag = 0;
 361         int tag_flag;
 362         int i;
 363         int tag_bytes = journal_tag_bytes(journal);
 364         struct buffer_head *cbh = NULL; /* For transactional checksums */
 365         __u32 crc32_sum = ~0;
 366         struct blk_plug plug;
 367         /* Tail of the journal */
 368         unsigned long first_block;
 369         tid_t first_tid;
 370         int update_tail;
 371         int csum_size = 0;
 372         LIST_HEAD(io_bufs);
 373         LIST_HEAD(log_bufs);
 374
 375         if (jbd2_journal_has_csum_v2or3(journal))
 376                 csum_size = sizeof(struct jbd2_journal_block_tail);
 377
 378         /*
 379          * First job: lock down the current transaction and wait for
 380          * all outstanding updates to complete.
 381          */
 382
 383         /* Do we need to erase the effects of a prior jbd2_journal_flush? */
 384         if (journal->j_flags & JBD2_FLUSHED) {
 385                 jbd_debug(3, "super block updated\n");
 386                 mutex_lock_io(&journal->j_checkpoint_mutex);
 387                 /*
 388                  * We hold j_checkpoint_mutex so tail cannot change under us.
 389                  * We don't need any special data guarantees for writing sb
 390                  * since journal is empty and it is ok for write to be
 391                  * flushed only with transaction commit.
 392                  */
 393                 jbd2_journal_update_sb_log_tail(journal,
 394                                                 journal->j_tail_sequence,
 395                                                 journal->j_tail,
 396                                                 REQ_SYNC);
 397                 mutex_unlock(&journal->j_checkpoint_mutex);
 398         } else {
 399                 jbd_debug(3, "superblock not updated\n");
 400         }
 401
 402         J_ASSERT(journal->j_running_transaction != NULL);
 403         J_ASSERT(journal->j_committing_transaction == NULL);
 404
 405         commit_transaction = journal->j_running_transaction;
 406
 407         trace_jbd2_start_commit(journal, commit_transaction);
 408         jbd_debug(1, "JBD2: starting commit of transaction %d\n",
 409                         commit_transaction->t_tid);
 410
 411         write_lock(&journal->j_state_lock);
 412         J_ASSERT(commit_transaction->t_state == T_RUNNING);
 413         commit_transaction->t_state = T_LOCKED;
 414
 415         trace_jbd2_commit_locking(journal, commit_transaction);
 416         stats.run.rs_wait = commit_transaction->t_max_wait;
 417         stats.run.rs_request_delay = 0;
 418         stats.run.rs_locked = jiffies;
 419         if (commit_transaction->t_requested)
 420                 stats.run.rs_request_delay =
 421                         jbd2_time_diff(commit_transaction->t_requested,
 422                                        stats.run.rs_locked);
 423         stats.run.rs_running = jbd2_time_diff(commit_transaction->t_start,
 424                                               stats.run.rs_locked);
 425
 426         spin_lock(&commit_transaction->t_handle_lock);
 427         while (atomic_read(&commit_transaction->t_updates)) {
 428                 DEFINE_WAIT(wait);
 429
 430                 prepare_to_wait(&journal->j_wait_updates, &wait,
 431                                         TASK_UNINTERRUPTIBLE);
 432                 if (atomic_read(&commit_transaction->t_updates)) {
 433                         spin_unlock(&commit_transaction->t_handle_lock);
 434                         write_unlock(&journal->j_state_lock);
 435                         schedule();
 436                         write_lock(&journal->j_state_lock);
 437                         spin_lock(&commit_transaction->t_handle_lock);
 438                 }
 439                 finish_wait(&journal->j_wait_updates, &wait);
 440         }
 441         spin_unlock(&commit_transaction->t_handle_lock);
 442
 443         J_ASSERT (atomic_read(&commit_transaction->t_outstanding_credits) <=
 444                         journal->j_max_transaction_buffers);
 445
 446         /*
 447          * First thing we are allowed to do is to discard any remaining
 448          * BJ_Reserved buffers.  Note, it is _not_ permissible to assume
 449          * that there are no such buffers: if a large filesystem
 450          * operation like a truncate needs to split itself over multiple
 451          * transactions, then it may try to do a jbd2_journal_restart() while
 452          * there are still BJ_Reserved buffers outstanding.  These must
 453          * be released cleanly from the current transaction.
 454          *
 455          * In this case, the filesystem must still reserve write access
 456          * again before modifying the buffer in the new transaction, but
 457          * we do not require it to remember exactly which old buffers it
 458          * has reserved.  This is consistent with the existing behaviour
 459          * that multiple jbd2_journal_get_write_access() calls to the same
 460          * buffer are perfectly permissible.
 461          */
 462         while (commit_transaction->t_reserved_list) {
 463                 jh = commit_transaction->t_reserved_list;
 464                 JBUFFER_TRACE(jh, "reserved, unused: refile");
 465                 /*
 466                  * A jbd2_journal_get_undo_access()+jbd2_journal_release_buffer() may
 467                  * leave undo-committed data.
 468                  */
 469                 if (jh->b_committed_data) {
 470                         struct buffer_head *bh = jh2bh(jh);
 471
 472                         jbd_lock_bh_state(bh);
 473                         jbd2_free(jh->b_committed_data, bh->b_size);
 474                         jh->b_committed_data = NULL;
 475                         jbd_unlock_bh_state(bh);
 476                 }
 477                 jbd2_journal_refile_buffer(journal, jh);
 478         }
 479
 480         /*
 481          * Now try to drop any written-back buffers from the journal's
 482          * checkpoint lists.  We do this *before* commit because it potentially
 483          * frees some memory
 484          */
 485         spin_lock(&journal->j_list_lock);
 486         __jbd2_journal_clean_checkpoint_list(journal, false);
 487         spin_unlock(&journal->j_list_lock);
 488
 489         jbd_debug(3, "JBD2: commit phase 1\n");
 490
 491         /*
 492          * Clear revoked flag to reflect there is no revoked buffers
 493          * in the next transaction which is going to be started.
 494          */
 495         jbd2_clear_buffer_revoked_flags(journal);
 496
 497         /*
 498          * Switch to a new revoke table.
 499          */
 500         jbd2_journal_switch_revoke_table(journal);
 501
 502         /*
 503          * Reserved credits cannot be claimed anymore, free them
 504          */
 505         atomic_sub(atomic_read(&journal->j_reserved_credits),
 506                    &commit_transaction->t_outstanding_credits);
 507
 508         trace_jbd2_commit_flushing(journal, commit_transaction);
 509         stats.run.rs_flushing = jiffies;
 510         stats.run.rs_locked = jbd2_time_diff(stats.run.rs_locked,
 511                                              stats.run.rs_flushing);
 512
 513         commit_transaction->t_state = T_FLUSH;
 514         journal->j_committing_transaction = commit_transaction;
 515         journal->j_running_transaction = NULL;
 516         start_time = ktime_get();
 517         commit_transaction->t_log_start = journal->j_head;
 518         wake_up(&journal->j_wait_transaction_locked);
 519         write_unlock(&journal->j_state_lock);
 520
 521         jbd_debug(3, "JBD2: commit phase 2a\n");
 522
 523         /*
 524          * Now start flushing things to disk, in the order they appear
 525          * on the transaction lists.  Data blocks go first.
 526          */
 527         err = journal_submit_data_buffers(journal, commit_transaction);
 528         if (err)
 529                 jbd2_journal_abort(journal, err);
 530
 531         blk_start_plug(&plug);
 532         jbd2_journal_write_revoke_records(commit_transaction, &log_bufs);
 533
 534         jbd_debug(3, "JBD2: commit phase 2b\n");
 535
 536         /*
 537          * Way to go: we have now written out all of the data for a
 538          * transaction!  Now comes the tricky part: we need to write out
 539          * metadata.  Loop over the transaction's entire buffer list:
 540          */
 541         write_lock(&journal->j_state_lock);
 542         commit_transaction->t_state = T_COMMIT;
 543         write_unlock(&journal->j_state_lock);
 544
 545         trace_jbd2_commit_logging(journal, commit_transaction);
 546         stats.run.rs_logging = jiffies;
 547         stats.run.rs_flushing = jbd2_time_diff(stats.run.rs_flushing,
 548                                                stats.run.rs_logging);
 549         stats.run.rs_blocks =
 550                 atomic_read(&commit_transaction->t_outstanding_credits);
 551         stats.run.rs_blocks_logged = 0;
 552
 553         J_ASSERT(commit_transaction->t_nr_buffers <=
 554                  atomic_read(&commit_transaction->t_outstanding_credits));
 555
 556         err = 0;
 557         bufs = 0;
 558         descriptor = NULL;
 559         while (commit_transaction->t_buffers) {
 560
 561                 /* Find the next buffer to be journaled... */
 562
 563                 jh = commit_transaction->t_buffers;
 564
 565                 /* If we're in abort mode, we just un-journal the buffer and
 566                    release it. */
 567
 568                 if (is_journal_aborted(journal)) {
 569                         clear_buffer_jbddirty(jh2bh(jh));
 570                         JBUFFER_TRACE(jh, "journal is aborting: refile");
 571                         jbd2_buffer_abort_trigger(jh,
 572                                                   jh->b_frozen_data ?
 573                                                   jh->b_frozen_triggers :
 574                                                   jh->b_triggers);
 575                         jbd2_journal_refile_buffer(journal, jh);
 576                         /* If that was the last one, we need to clean up
 577                          * any descriptor buffers which may have been
 578                          * already allocated, even if we are now
 579                          * aborting. */
 580                         if (!commit_transaction->t_buffers)
 581                                 goto start_journal_io;
 582                         continue;
 583                 }
 584
 585                 /* Make sure we have a descriptor block in which to
 586                    record the metadata buffer. */
 587
 588                 if (!descriptor) {
 589                         J_ASSERT (bufs == 0);
 590
 591                         jbd_debug(4, "JBD2: get descriptor\n");
 592
 593                         descriptor = jbd2_journal_get_descriptor_buffer(
 594                                                         commit_transaction,
 595                                                         JBD2_DESCRIPTOR_BLOCK);
 596                         if (!descriptor) {
 597                                 jbd2_journal_abort(journal, -EIO);
 598                                 continue;
 599                         }
 600
 601                         jbd_debug(4, "JBD2: got buffer %llu (%p)\n",
 602                                 (unsigned long long)descriptor->b_blocknr,
 603                                 descriptor->b_data);
 604                         tagp = &descriptor->b_data[sizeof(journal_header_t)];
 605                         space_left = descriptor->b_size -
 606                                                 sizeof(journal_header_t);
 607                         first_tag = 1;
 608                         set_buffer_jwrite(descriptor);
 609                         set_buffer_dirty(descriptor);
 610                         wbuf[bufs++] = descriptor;
 611
 612                         /* Record it so that we can wait for IO
 613                            completion later */
 614                         BUFFER_TRACE(descriptor, "ph3: file as descriptor");
 615                         jbd2_file_log_bh(&log_bufs, descriptor);
 616                 }
 617
 618                 /* Where is the buffer to be written? */
 619
 620                 err = jbd2_journal_next_log_block(journal, &blocknr);
 621                 /* If the block mapping failed, just abandon the buffer
 622                    and repeat this loop: we'll fall into the
 623                    refile-on-abort condition above. */
 624                 if (err) {
 625                         jbd2_journal_abort(journal, err);
 626                         continue;
 627                 }
 628
 629                 /*
 630                  * start_this_handle() uses t_outstanding_credits to determine
 631                  * the free space in the log, but this counter is changed
 632                  * by jbd2_journal_next_log_block() also.
 633                  */
 634                 atomic_dec(&commit_transaction->t_outstanding_credits);
 635
 636                 /* Bump b_count to prevent truncate from stumbling over
 637                    the shadowed buffer!  @@@ This can go if we ever get
 638                    rid of the shadow pairing of buffers. */
 639                 atomic_inc(&jh2bh(jh)->b_count);
 640
 641                 /*
 642                  * Make a temporary IO buffer with which to write it out
 643                  * (this will requeue the metadata buffer to BJ_Shadow).
 644                  */
 645                 set_bit(BH_JWrite, &jh2bh(jh)->b_state);
 646                 JBUFFER_TRACE(jh, "ph3: write metadata");
 647                 flags = jbd2_journal_write_metadata_buffer(commit_transaction,
 648                                                 jh, &wbuf[bufs], blocknr);
 649                 if (flags < 0) {
 650                         jbd2_journal_abort(journal, flags);
 651                         continue;
 652                 }
 653                 jbd2_file_log_bh(&io_bufs, wbuf[bufs]);
 654
 655                 /* Record the new block's tag in the current descriptor
 656                    buffer */
 657
 658                 tag_flag = 0;
 659                 if (flags & 1)
 660                         tag_flag |= JBD2_FLAG_ESCAPE;
 661                 if (!first_tag)
 662                         tag_flag |= JBD2_FLAG_SAME_UUID;
 663
 664                 tag = (journal_block_tag_t *) tagp;
 665                 write_tag_block(journal, tag, jh2bh(jh)->b_blocknr);
 666                 tag->t_flags = cpu_to_be16(tag_flag);
 667                 jbd2_block_tag_csum_set(journal, tag, wbuf[bufs],
 668                                         commit_transaction->t_tid);
 669                 tagp += tag_bytes;
 670                 space_left -= tag_bytes;
 671                 bufs++;
 672
 673                 if (first_tag) {
 674                         memcpy (tagp, journal->j_uuid, 16);
 675                         tagp += 16;
 676                         space_left -= 16;
 677                         first_tag = 0;
 678                 }
 679
 680                 /* If there's no more to do, or if the descriptor is full,
 681                    let the IO rip! */
 682
 683                 if (bufs == journal->j_wbufsize ||
 684                     commit_transaction->t_buffers == NULL ||
 685                     space_left < tag_bytes + 16 + csum_size) {
 686
 687                         jbd_debug(4, "JBD2: Submit %d IOs\n", bufs);
 688
 689                         /* Write an end-of-descriptor marker before
 690                            submitting the IOs.  "tag" still points to
 691                            the last tag we set up. */
 692
 693                         tag->t_flags |= cpu_to_be16(JBD2_FLAG_LAST_TAG);
 694
 695                         jbd2_descriptor_block_csum_set(journal, descriptor);
 696 start_journal_io:
 697                         for (i = 0; i < bufs; i++) {
 698                                 struct buffer_head *bh = wbuf[i];
 699                                 /*
 700                                  * Compute checksum.
 701                                  */
 702                                 if (jbd2_has_feature_checksum(journal)) {
 703                                         crc32_sum =
 704                                             jbd2_checksum_data(crc32_sum, bh);
 705                                 }
 706
 707                                 lock_buffer(bh);
 708                                 clear_buffer_dirty(bh);
 709                                 set_buffer_uptodate(bh);
 710                                 bh->b_end_io = journal_end_buffer_io_sync;
 711                                 submit_bh(REQ_OP_WRITE, REQ_SYNC, bh);
 712                         }
 713                         cond_resched();
 714                         stats.run.rs_blocks_logged += bufs;
 715
 716                         /* Force a new descriptor to be generated next
 717                            time round the loop. */
 718                         descriptor = NULL;
 719                         bufs = 0;
 720                 }
 721         }
 722
 723         err = journal_finish_inode_data_buffers(journal, commit_transaction);
 724         if (err) {
 725                 printk(KERN_WARNING
 726                         "JBD2: Detected IO errors while flushing file data "
 727                        "on %s\n", journal->j_devname);
 728                 if (journal->j_flags & JBD2_ABORT_ON_SYNCDATA_ERR)
 729                         jbd2_journal_abort(journal, err);
 730                 err = 0;
 731         }
 732
 733         /*
 734          * Get current oldest transaction in the log before we issue flush
 735          * to the filesystem device. After the flush we can be sure that
 736          * blocks of all older transactions are checkpointed to persistent
 737          * storage and we will be safe to update journal start in the
 738          * superblock with the numbers we get here.
 739          */
 740         update_tail =
 741                 jbd2_journal_get_log_tail(journal, &first_tid, &first_block);
 742
 743         write_lock(&journal->j_state_lock);
 744         if (update_tail) {
 745                 long freed = first_block - journal->j_tail;
 746
 747                 if (first_block < journal->j_tail)
 748                         freed += journal->j_last - journal->j_first;
 749                 /* Update tail only if we free significant amount of space */
 750                 if (freed < journal->j_maxlen / 4)
 751                         update_tail = 0;
 752         }
 753         J_ASSERT(commit_transaction->t_state == T_COMMIT);
 754         commit_transaction->t_state = T_COMMIT_DFLUSH;
 755         write_unlock(&journal->j_state_lock);
 756
 757         /*
 758          * If the journal is not located on the file system device,
 759          * then we must flush the file system device before we issue
 760          * the commit record
 761          */
 762         if (commit_transaction->t_need_data_flush &&
 763             (journal->j_fs_dev != journal->j_dev) &&
 764             (journal->j_flags & JBD2_BARRIER))
 765                 blkdev_issue_flush(journal->j_fs_dev, GFP_NOFS, NULL);
 766
 767         /* Done it all: now write the commit record asynchronously. */
 768         if (jbd2_has_feature_async_commit(journal)) {
 769                 err = journal_submit_commit_record(journal, commit_transaction,
 770                                                  &cbh, crc32_sum);
 771                 if (err)
 772                         __jbd2_journal_abort_hard(journal);
 773         }
 774
 775         blk_finish_plug(&plug);
 776
 777         /* Lo and behold: we have just managed to send a transaction to
 778            the log.  Before we can commit it, wait for the IO so far to
 779            complete.  Control buffers being written are on the
 780            transaction's t_log_list queue, and metadata buffers are on
 781            the io_bufs list.
 782
 783            Wait for the buffers in reverse order.  That way we are
 784            less likely to be woken up until all IOs have completed, and
 785            so we incur less scheduling load.
 786         */
 787
 788         jbd_debug(3, "JBD2: commit phase 3\n");
 789
 790         while (!list_empty(&io_bufs)) {
 791                 struct buffer_head *bh = list_entry(io_bufs.prev,
 792                                                     struct buffer_head,
 793                                                     b_assoc_buffers);
 794
 795                 wait_on_buffer(bh);
 796                 cond_resched();
 797
 798                 if (unlikely(!buffer_uptodate(bh)))
 799                         err = -EIO;
 800                 jbd2_unfile_log_bh(bh);
 801
 802                 /*
 803                  * The list contains temporary buffer heads created by
 804                  * jbd2_journal_write_metadata_buffer().
 805                  */
 806                 BUFFER_TRACE(bh, "dumping temporary bh");
 807                 __brelse(bh);
 808                 J_ASSERT_BH(bh, atomic_read(&bh->b_count) == 0);
 809                 free_buffer_head(bh);
 810
 811                 /* We also have to refile the corresponding shadowed buffer */
 812                 jh = commit_transaction->t_shadow_list->b_tprev;
 813                 bh = jh2bh(jh);
 814                 clear_buffer_jwrite(bh);
 815                 J_ASSERT_BH(bh, buffer_jbddirty(bh));
 816                 J_ASSERT_BH(bh, !buffer_shadow(bh));
 817
 818                 /* The metadata is now released for reuse, but we need
 819                    to remember it against this transaction so that when
 820                    we finally commit, we can do any checkpointing
 821                    required. */
 822                 JBUFFER_TRACE(jh, "file as BJ_Forget");
 823                 jbd2_journal_file_buffer(jh, commit_transaction, BJ_Forget);
 824                 JBUFFER_TRACE(jh, "brelse shadowed buffer");
 825                 __brelse(bh);
 826         }
 827
 828         J_ASSERT (commit_transaction->t_shadow_list == NULL);
 829
 830         jbd_debug(3, "JBD2: commit phase 4\n");
 831
 832         /* Here we wait for the revoke record and descriptor record buffers */
 833         while (!list_empty(&log_bufs)) {
 834                 struct buffer_head *bh;
 835
 836                 bh = list_entry(log_bufs.prev, struct buffer_head, b_assoc_buffers);
 837                 wait_on_buffer(bh);
 838                 cond_resched();
 839
 840                 if (unlikely(!buffer_uptodate(bh)))
 841                         err = -EIO;
 842
 843                 BUFFER_TRACE(bh, "ph5: control buffer writeout done: unfile");
 844                 clear_buffer_jwrite(bh);
 845                 jbd2_unfile_log_bh(bh);
 846                 __brelse(bh);           /* One for getblk */
 847                 /* AKPM: bforget here */
 848         }
 849
 850         if (err)
 851                 jbd2_journal_abort(journal, err);
 852
 853         jbd_debug(3, "JBD2: commit phase 5\n");
 854         write_lock(&journal->j_state_lock);
 855         J_ASSERT(commit_transaction->t_state == T_COMMIT_DFLUSH);
 856         commit_transaction->t_state = T_COMMIT_JFLUSH;
 857         write_unlock(&journal->j_state_lock);
 858
 859         if (!jbd2_has_feature_async_commit(journal)) {
 860                 err = journal_submit_commit_record(journal, commit_transaction,
 861                                                 &cbh, crc32_sum);
 862                 if (err)
 863                         __jbd2_journal_abort_hard(journal);
 864         }
 865         if (cbh)
 866                 err = journal_wait_on_commit_record(journal, cbh);
 867         if (jbd2_has_feature_async_commit(journal) &&
 868             journal->j_flags & JBD2_BARRIER) {
 869                 blkdev_issue_flush(journal->j_dev, GFP_NOFS, NULL);
 870         }
 871
 872         if (err)
 873                 jbd2_journal_abort(journal, err);
 874
 875         /*
 876          * Now disk caches for filesystem device are flushed so we are safe to
 877          * erase checkpointed transactions from the log by updating journal
 878          * superblock.
 879          */
 880         if (update_tail)
 881                 jbd2_update_log_tail(journal, first_tid, first_block);
 882
 883         /* End of a transaction!  Finally, we can do checkpoint
 884            processing: any buffers committed as a result of this
 885            transaction can be removed from any checkpoint list it was on
 886            before. */
 887
 888         jbd_debug(3, "JBD2: commit phase 6\n");
 889
 890         J_ASSERT(list_empty(&commit_transaction->t_inode_list));
 891         J_ASSERT(commit_transaction->t_buffers == NULL);
 892         J_ASSERT(commit_transaction->t_checkpoint_list == NULL);
 893         J_ASSERT(commit_transaction->t_shadow_list == NULL);
 894
 895 restart_loop:
 896         /*
 897          * As there are other places (journal_unmap_buffer()) adding buffers
 898          * to this list we have to be careful and hold the j_list_lock.
 899          */
 900         spin_lock(&journal->j_list_lock);
 901         while (commit_transaction->t_forget) {
 902                 transaction_t *cp_transaction;
 903                 struct buffer_head *bh;
 904                 int try_to_free = 0;
 905
 906                 jh = commit_transaction->t_forget;
 907                 spin_unlock(&journal->j_list_lock);
 908                 bh = jh2bh(jh);
 909                 /*
 910                  * Get a reference so that bh cannot be freed before we are
 911                  * done with it.
 912                  */
 913                 get_bh(bh);
 914                 jbd_lock_bh_state(bh);
 915                 J_ASSERT_JH(jh, jh->b_transaction == commit_transaction);
 916
 917                 /*
 918                  * If there is undo-protected committed data against
 919                  * this buffer, then we can remove it now.  If it is a
 920                  * buffer needing such protection, the old frozen_data
 921                  * field now points to a committed version of the
 922                  * buffer, so rotate that field to the new committed
 923                  * data.
 924                  *
 925                  * Otherwise, we can just throw away the frozen data now.
 926                  *
 927                  * We also know that the frozen data has already fired
 928                  * its triggers if they exist, so we can clear that too.
 929                  */
 930                 if (jh->b_committed_data) {
 931                         jbd2_free(jh->b_committed_data, bh->b_size);
 932                         jh->b_committed_data = NULL;
 933                         if (jh->b_frozen_data) {
 934                                 jh->b_committed_data = jh->b_frozen_data;
 935                                 jh->b_frozen_data = NULL;
 936                                 jh->b_frozen_triggers = NULL;
 937                         }
 938                 } else if (jh->b_frozen_data) {
 939                         jbd2_free(jh->b_frozen_data, bh->b_size);
 940                         jh->b_frozen_data = NULL;
 941                         jh->b_frozen_triggers = NULL;
 942                 }
 943
 944                 spin_lock(&journal->j_list_lock);
 945                 cp_transaction = jh->b_cp_transaction;
 946                 if (cp_transaction) {
 947                         JBUFFER_TRACE(jh, "remove from old cp transaction");
 948                         cp_transaction->t_chp_stats.cs_dropped++;
 949                         __jbd2_journal_remove_checkpoint(jh);
 950                 }
 951
 952                 /* Only re-checkpoint the buffer_head if it is marked
 953                  * dirty.  If the buffer was added to the BJ_Forget list
 954                  * by jbd2_journal_forget, it may no longer be dirty and
 955                  * there's no point in keeping a checkpoint record for
 956                  * it. */
 957
 958                 /*
 959                 * A buffer which has been freed while still being journaled by
 960                 * a previous transaction.
 961                 */
 962                 if (buffer_freed(bh)) {
 963                         /*
 964                          * If the running transaction is the one containing
 965                          * "add to orphan" operation (b_next_transaction !=
 966                          * NULL), we have to wait for that transaction to
 967                          * commit before we can really get rid of the buffer.
 968                          * So just clear b_modified to not confuse transaction
 969                          * credit accounting and refile the buffer to
 970                          * BJ_Forget of the running transaction. If the just
 971                          * committed transaction contains "add to orphan"
 972                          * operation, we can completely invalidate the buffer
 973                          * now. We are rather through in that since the
 974                          * buffer may be still accessible when blocksize <
 975                          * pagesize and it is attached to the last partial
 976                          * page.
 977                          */
 978                         jh->b_modified = 0;
 979                         if (!jh->b_next_transaction) {
 980                                 clear_buffer_freed(bh);
 981                                 clear_buffer_jbddirty(bh);
 982                                 clear_buffer_mapped(bh);
 983                                 clear_buffer_new(bh);
 984                                 clear_buffer_req(bh);
 985                                 bh->b_bdev = NULL;
 986                         }
 987                 }
 988
 989                 if (buffer_jbddirty(bh)) {
 990                         JBUFFER_TRACE(jh, "add to new checkpointing trans");
 991                         __jbd2_journal_insert_checkpoint(jh, commit_transaction);
 992                         if (is_journal_aborted(journal))
 993                                 clear_buffer_jbddirty(bh);
 994                 } else {
 995                         J_ASSERT_BH(bh, !buffer_dirty(bh));
 996                         /*
 997                          * The buffer on BJ_Forget list and not jbddirty means
 998                          * it has been freed by this transaction and hence it
 999                          * could not have been reallocated until this
1000                          * transaction has committed. *BUT* it could be
1001                          * reallocated once we have written all the data to
1002                          * disk and before we process the buffer on BJ_Forget
1003                          * list.
1004                          */
1005                         if (!jh->b_next_transaction)
1006                                 try_to_free = 1;
1007                 }
1008                 JBUFFER_TRACE(jh, "refile or unfile buffer");
1009                 __jbd2_journal_refile_buffer(jh);
1010                 jbd_unlock_bh_state(bh);
1011                 if (try_to_free)
1012                         release_buffer_page(bh);        /* Drops bh reference */
1013                 else
1014                         __brelse(bh);
1015                 cond_resched_lock(&journal->j_list_lock);
1016         }
1017         spin_unlock(&journal->j_list_lock);
1018         /*
1019          * This is a bit sleazy.  We use j_list_lock to protect transition
1020          * of a transaction into T_FINISHED state and calling
1021          * __jbd2_journal_drop_transaction(). Otherwise we could race with
1022          * other checkpointing code processing the transaction...
1023          */
1024         write_lock(&journal->j_state_lock);
1025         spin_lock(&journal->j_list_lock);
1026         /*
1027          * Now recheck if some buffers did not get attached to the transaction
1028          * while the lock was dropped...
1029          */
1030         if (commit_transaction->t_forget) {
1031                 spin_unlock(&journal->j_list_lock);
1032                 write_unlock(&journal->j_state_lock);
1033                 goto restart_loop;
1034         }
1035
1036         /* Add the transaction to the checkpoint list
1037          * __journal_remove_checkpoint() can not destroy transaction
1038          * under us because it is not marked as T_FINISHED yet */
1039         if (journal->j_checkpoint_transactions == NULL) {
1040                 journal->j_checkpoint_transactions = commit_transaction;
1041                 commit_transaction->t_cpnext = commit_transaction;
1042                 commit_transaction->t_cpprev = commit_transaction;
1043         } else {
1044                 commit_transaction->t_cpnext =
1045                         journal->j_checkpoint_transactions;
1046                 commit_transaction->t_cpprev =
1047                         commit_transaction->t_cpnext->t_cpprev;
1048                 commit_transaction->t_cpnext->t_cpprev =
1049                         commit_transaction;
1050                 commit_transaction->t_cpprev->t_cpnext =
1051                                 commit_transaction;
1052         }
1053         spin_unlock(&journal->j_list_lock);
1054
1055         /* Done with this transaction! */
1056
1057         jbd_debug(3, "JBD2: commit phase 7\n");
1058
1059         J_ASSERT(commit_transaction->t_state == T_COMMIT_JFLUSH);
1060
1061         commit_transaction->t_start = jiffies;
1062         stats.run.rs_logging = jbd2_time_diff(stats.run.rs_logging,
1063                                               commit_transaction->t_start);
1064
1065         /*
1066          * File the transaction statistics
1067          */
1068         stats.ts_tid = commit_transaction->t_tid;
1069         stats.run.rs_handle_count =
1070                 atomic_read(&commit_transaction->t_handle_count);
1071         trace_jbd2_run_stats(journal->j_fs_dev->bd_dev,
1072                              commit_transaction->t_tid, &stats.run);
1073         stats.ts_requested = (commit_transaction->t_requested) ? 1 : 0;
1074
1075         commit_transaction->t_state = T_COMMIT_CALLBACK;
1076         J_ASSERT(commit_transaction == journal->j_committing_transaction);
1077         journal->j_commit_sequence = commit_transaction->t_tid;
1078         journal->j_committing_transaction = NULL;
1079         commit_time = ktime_to_ns(ktime_sub(ktime_get(), start_time));
1080
1081         /*
1082          * weight the commit time higher than the average time so we don't
1083          * react too strongly to vast changes in the commit time
1084          */
1085         if (likely(journal->j_average_commit_time))
1086                 journal->j_average_commit_time = (commit_time +
1087                                 journal->j_average_commit_time*3) / 4;
1088         else
1089                 journal->j_average_commit_time = commit_time;
1090
1091         write_unlock(&journal->j_state_lock);
1092
1093         if (journal->j_commit_callback)
1094                 journal->j_commit_callback(journal, commit_transaction);
1095
1096         trace_jbd2_end_commit(journal, commit_transaction);
1097         jbd_debug(1, "JBD2: commit %d complete, head %d\n",
1098                   journal->j_commit_sequence, journal->j_tail_sequence);
1099
1100         write_lock(&journal->j_state_lock);
1101         spin_lock(&journal->j_list_lock);
1102         commit_transaction->t_state = T_FINISHED;
1103         /* Check if the transaction can be dropped now that we are finished */
1104         if (commit_transaction->t_checkpoint_list == NULL &&
1105             commit_transaction->t_checkpoint_io_list == NULL) {
1106                 __jbd2_journal_drop_transaction(journal, commit_transaction);
1107                 jbd2_journal_free_transaction(commit_transaction);
1108         }
1109         spin_unlock(&journal->j_list_lock);
1110         write_unlock(&journal->j_state_lock);
1111         wake_up(&journal->j_wait_done_commit);
1112
1113         /*
1114          * Calculate overall stats
1115          */
1116         spin_lock(&journal->j_history_lock);
1117         journal->j_stats.ts_tid++;
1118         journal->j_stats.ts_requested += stats.ts_requested;
1119         journal->j_stats.run.rs_wait += stats.run.rs_wait;
1120         journal->j_stats.run.rs_request_delay += stats.run.rs_request_delay;
1121         journal->j_stats.run.rs_running += stats.run.rs_running;
1122         journal->j_stats.run.rs_locked += stats.run.rs_locked;
1123         journal->j_stats.run.rs_flushing += stats.run.rs_flushing;
1124         journal->j_stats.run.rs_logging += stats.run.rs_logging;
1125         journal->j_stats.run.rs_handle_count += stats.run.rs_handle_count;
1126         journal->j_stats.run.rs_blocks += stats.run.rs_blocks;
1127         journal->j_stats.run.rs_blocks_logged += stats.run.rs_blocks_logged;
1128         spin_unlock(&journal->j_history_lock);
1129 }