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 = current_kernel_time64();
 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         tmp->h_commit_sec = cpu_to_be64(now.tv_sec);
 138         tmp->h_commit_nsec = cpu_to_be32(now.tv_nsec);
 139
 140         if (jbd2_has_feature_checksum(journal)) {
 141                 tmp->h_chksum_type      = JBD2_CRC32_CHKSUM;
 142                 tmp->h_chksum_size      = JBD2_CRC32_CHKSUM_SIZE;
 143                 tmp->h_chksum[0]        = cpu_to_be32(crc32_sum);
 144         }
 145         jbd2_commit_block_csum_set(journal, bh);
 146
 147         BUFFER_TRACE(bh, "submit commit block");
 148         lock_buffer(bh);
 149         clear_buffer_dirty(bh);
 150         set_buffer_uptodate(bh);
 151         bh->b_end_io = journal_end_buffer_io_sync;
 152
 153         if (journal->j_flags & JBD2_BARRIER &&
 154             !jbd2_has_feature_async_commit(journal))
 155                 ret = submit_bh(REQ_OP_WRITE,
 156                         REQ_SYNC | REQ_PREFLUSH | REQ_FUA, bh);
 157         else
 158                 ret = submit_bh(REQ_OP_WRITE, REQ_SYNC, bh);
 159
 160         *cbh = bh;
 161         return ret;
 162 }
 163
 164 /*
 165  * This function along with journal_submit_commit_record
 166  * allows to write the commit record asynchronously.
 167  */
 168 static int journal_wait_on_commit_record(journal_t *journal,
 169                                          struct buffer_head *bh)
 170 {
 171         int ret = 0;
 172
 173         clear_buffer_dirty(bh);
 174         wait_on_buffer(bh);
 175
 176         if (unlikely(!buffer_uptodate(bh)))
 177                 ret = -EIO;
 178         put_bh(bh);            /* One for getblk() */
 179
 180         return ret;
 181 }
 182
 183 /*
 184  * write the filemap data using writepage() address_space_operations.
 185  * We don't do block allocation here even for delalloc. We don't
 186  * use writepages() because with dealyed allocation we may be doing
 187  * block allocation in writepages().
 188  */
 189 static int journal_submit_inode_data_buffers(struct address_space *mapping)
 190 {
 191         int ret;
 192         struct writeback_control wbc = {
 193                 .sync_mode =  WB_SYNC_ALL,
 194                 .nr_to_write = mapping->nrpages * 2,
 195                 .range_start = 0,
 196                 .range_end = i_size_read(mapping->host),
 197         };
 198
 199         ret = generic_writepages(mapping, &wbc);
 200         return ret;
 201 }
 202
 203 /*
 204  * Submit all the data buffers of inode associated with the transaction to
 205  * disk.
 206  *
 207  * We are in a committing transaction. Therefore no new inode can be added to
 208  * our inode list. We use JI_COMMIT_RUNNING flag to protect inode we currently
 209  * operate on from being released while we write out pages.
 210  */
 211 static int journal_submit_data_buffers(journal_t *journal,
 212                 transaction_t *commit_transaction)
 213 {
 214         struct jbd2_inode *jinode;
 215         int err, ret = 0;
 216         struct address_space *mapping;
 217
 218         spin_lock(&journal->j_list_lock);
 219         list_for_each_entry(jinode, &commit_transaction->t_inode_list, i_list) {
 220                 if (!(jinode->i_flags & JI_WRITE_DATA))
 221                         continue;
 222                 mapping = jinode->i_vfs_inode->i_mapping;
 223                 jinode->i_flags |= JI_COMMIT_RUNNING;
 224                 spin_unlock(&journal->j_list_lock);
 225                 /*
 226                  * submit the inode data buffers. We use writepage
 227                  * instead of writepages. Because writepages can do
 228                  * block allocation  with delalloc. We need to write
 229                  * only allocated blocks here.
 230                  */
 231                 trace_jbd2_submit_inode_data(jinode->i_vfs_inode);
 232                 err = journal_submit_inode_data_buffers(mapping);
 233                 if (!ret)
 234                         ret = err;
 235                 spin_lock(&journal->j_list_lock);
 236                 J_ASSERT(jinode->i_transaction == commit_transaction);
 237                 jinode->i_flags &= ~JI_COMMIT_RUNNING;
 238                 smp_mb();
 239                 wake_up_bit(&jinode->i_flags, __JI_COMMIT_RUNNING);
 240         }
 241         spin_unlock(&journal->j_list_lock);
 242         return ret;
 243 }
 244
 245 /*
 246  * Wait for data submitted for writeout, refile inodes to proper
 247  * transaction if needed.
 248  *
 249  */
 250 static int journal_finish_inode_data_buffers(journal_t *journal,
 251                 transaction_t *commit_transaction)
 252 {
 253         struct jbd2_inode *jinode, *next_i;
 254         int err, ret = 0;
 255
 256         /* For locking, see the comment in journal_submit_data_buffers() */
 257         spin_lock(&journal->j_list_lock);
 258         list_for_each_entry(jinode, &commit_transaction->t_inode_list, i_list) {
 259                 if (!(jinode->i_flags & JI_WAIT_DATA))
 260                         continue;
 261                 jinode->i_flags |= JI_COMMIT_RUNNING;
 262                 spin_unlock(&journal->j_list_lock);
 263                 err = filemap_fdatawait_keep_errors(
 264                                 jinode->i_vfs_inode->i_mapping);
 265                 if (!ret)
 266                         ret = err;
 267                 spin_lock(&journal->j_list_lock);
 268                 jinode->i_flags &= ~JI_COMMIT_RUNNING;
 269                 smp_mb();
 270                 wake_up_bit(&jinode->i_flags, __JI_COMMIT_RUNNING);
 271         }
 272
 273         /* Now refile inode to proper lists */
 274         list_for_each_entry_safe(jinode, next_i,
 275                                  &commit_transaction->t_inode_list, i_list) {
 276                 list_del(&jinode->i_list);
 277                 if (jinode->i_next_transaction) {
 278                         jinode->i_transaction = jinode->i_next_transaction;
 279                         jinode->i_next_transaction = NULL;
 280                         list_add(&jinode->i_list,
 281                                 &jinode->i_transaction->t_inode_list);
 282                 } else {
 283                         jinode->i_transaction = NULL;
 284                 }
 285         }
 286         spin_unlock(&journal->j_list_lock);
 287
 288         return ret;
 289 }
 290
 291 static __u32 jbd2_checksum_data(__u32 crc32_sum, struct buffer_head *bh)
 292 {
 293         struct page *page = bh->b_page;
 294         char *addr;
 295         __u32 checksum;
 296
 297         addr = kmap_atomic(page);
 298         checksum = crc32_be(crc32_sum,
 299                 (void *)(addr + offset_in_page(bh->b_data)), bh->b_size);
 300         kunmap_atomic(addr);
 301
 302         return checksum;
 303 }
 304
 305 static void write_tag_block(journal_t *j, journal_block_tag_t *tag,
 306                                    unsigned long long block)
 307 {
 308         tag->t_blocknr = cpu_to_be32(block & (u32)~0);
 309         if (jbd2_has_feature_64bit(j))
 310                 tag->t_blocknr_high = cpu_to_be32((block >> 31) >> 1);
 311 }
 312
 313 static void jbd2_block_tag_csum_set(journal_t *j, journal_block_tag_t *tag,
 314                                     struct buffer_head *bh, __u32 sequence)
 315 {
 316         journal_block_tag3_t *tag3 = (journal_block_tag3_t *)tag;
 317         struct page *page = bh->b_page;
 318         __u8 *addr;
 319         __u32 csum32;
 320         __be32 seq;
 321
 322         if (!jbd2_journal_has_csum_v2or3(j))
 323                 return;
 324
 325         seq = cpu_to_be32(sequence);
 326         addr = kmap_atomic(page);
 327         csum32 = jbd2_chksum(j, j->j_csum_seed, (__u8 *)&seq, sizeof(seq));
 328         csum32 = jbd2_chksum(j, csum32, addr + offset_in_page(bh->b_data),
 329                              bh->b_size);
 330         kunmap_atomic(addr);
 331
 332         if (jbd2_has_feature_csum3(j))
 333                 tag3->t_checksum = cpu_to_be32(csum32);
 334         else
 335                 tag->t_checksum = cpu_to_be16(csum32);
 336 }
 337 /*
 338  * jbd2_journal_commit_transaction
 339  *
 340  * The primary function for committing a transaction to the log.  This
 341  * function is called by the journal thread to begin a complete commit.
 342  */
 343 void jbd2_journal_commit_transaction(journal_t *journal)
 344 {
 345         struct transaction_stats_s stats;
 346         transaction_t *commit_transaction;
 347         struct journal_head *jh;
 348         struct buffer_head *descriptor;
 349         struct buffer_head **wbuf = journal->j_wbuf;
 350         int bufs;
 351         int flags;
 352         int err;
 353         unsigned long long blocknr;
 354         ktime_t start_time;
 355         u64 commit_time;
 356         char *tagp = NULL;
 357         journal_block_tag_t *tag = NULL;
 358         int space_left = 0;
 359         int first_tag = 0;
 360         int tag_flag;
 361         int i;
 362         int tag_bytes = journal_tag_bytes(journal);
 363         struct buffer_head *cbh = NULL; /* For transactional checksums */
 364         __u32 crc32_sum = ~0;
 365         struct blk_plug plug;
 366         /* Tail of the journal */
 367         unsigned long first_block;
 368         tid_t first_tid;
 369         int update_tail;
 370         int csum_size = 0;
 371         LIST_HEAD(io_bufs);
 372         LIST_HEAD(log_bufs);
 373
 374         if (jbd2_journal_has_csum_v2or3(journal))
 375                 csum_size = sizeof(struct jbd2_journal_block_tail);
 376
 377         /*
 378          * First job: lock down the current transaction and wait for
 379          * all outstanding updates to complete.
 380          */
 381
 382         /* Do we need to erase the effects of a prior jbd2_journal_flush? */
 383         if (journal->j_flags & JBD2_FLUSHED) {
 384                 jbd_debug(3, "super block updated\n");
 385                 mutex_lock_io(&journal->j_checkpoint_mutex);
 386                 /*
 387                  * We hold j_checkpoint_mutex so tail cannot change under us.
 388                  * We don't need any special data guarantees for writing sb
 389                  * since journal is empty and it is ok for write to be
 390                  * flushed only with transaction commit.
 391                  */
 392                 jbd2_journal_update_sb_log_tail(journal,
 393                                                 journal->j_tail_sequence,
 394                                                 journal->j_tail,
 395                                                 REQ_SYNC);
 396                 mutex_unlock(&journal->j_checkpoint_mutex);
 397         } else {
 398                 jbd_debug(3, "superblock not updated\n");
 399         }
 400
 401         J_ASSERT(journal->j_running_transaction != NULL);
 402         J_ASSERT(journal->j_committing_transaction == NULL);
 403
 404         commit_transaction = journal->j_running_transaction;
 405
 406         trace_jbd2_start_commit(journal, commit_transaction);
 407         jbd_debug(1, "JBD2: starting commit of transaction %d\n",
 408                         commit_transaction->t_tid);
 409
 410         write_lock(&journal->j_state_lock);
 411         J_ASSERT(commit_transaction->t_state == T_RUNNING);
 412         commit_transaction->t_state = T_LOCKED;
 413
 414         trace_jbd2_commit_locking(journal, commit_transaction);
 415         stats.run.rs_wait = commit_transaction->t_max_wait;
 416         stats.run.rs_request_delay = 0;
 417         stats.run.rs_locked = jiffies;
 418         if (commit_transaction->t_requested)
 419                 stats.run.rs_request_delay =
 420                         jbd2_time_diff(commit_transaction->t_requested,
 421                                        stats.run.rs_locked);
 422         stats.run.rs_running = jbd2_time_diff(commit_transaction->t_start,
 423                                               stats.run.rs_locked);
 424
 425         spin_lock(&commit_transaction->t_handle_lock);
 426         while (atomic_read(&commit_transaction->t_updates)) {
 427                 DEFINE_WAIT(wait);
 428
 429                 prepare_to_wait(&journal->j_wait_updates, &wait,
 430                                         TASK_UNINTERRUPTIBLE);
 431                 if (atomic_read(&commit_transaction->t_updates)) {
 432                         spin_unlock(&commit_transaction->t_handle_lock);
 433                         write_unlock(&journal->j_state_lock);
 434                         schedule();
 435                         write_lock(&journal->j_state_lock);
 436                         spin_lock(&commit_transaction->t_handle_lock);
 437                 }
 438                 finish_wait(&journal->j_wait_updates, &wait);
 439         }
 440         spin_unlock(&commit_transaction->t_handle_lock);
 441
 442         J_ASSERT (atomic_read(&commit_transaction->t_outstanding_credits) <=
 443                         journal->j_max_transaction_buffers);
 444
 445         /*
 446          * First thing we are allowed to do is to discard any remaining
 447          * BJ_Reserved buffers.  Note, it is _not_ permissible to assume
 448          * that there are no such buffers: if a large filesystem
 449          * operation like a truncate needs to split itself over multiple
 450          * transactions, then it may try to do a jbd2_journal_restart() while
 451          * there are still BJ_Reserved buffers outstanding.  These must
 452          * be released cleanly from the current transaction.
 453          *
 454          * In this case, the filesystem must still reserve write access
 455          * again before modifying the buffer in the new transaction, but
 456          * we do not require it to remember exactly which old buffers it
 457          * has reserved.  This is consistent with the existing behaviour
 458          * that multiple jbd2_journal_get_write_access() calls to the same
 459          * buffer are perfectly permissible.
 460          */
 461         while (commit_transaction->t_reserved_list) {
 462                 jh = commit_transaction->t_reserved_list;
 463                 JBUFFER_TRACE(jh, "reserved, unused: refile");
 464                 /*
 465                  * A jbd2_journal_get_undo_access()+jbd2_journal_release_buffer() may
 466                  * leave undo-committed data.
 467                  */
 468                 if (jh->b_committed_data) {
 469                         struct buffer_head *bh = jh2bh(jh);
 470
 471                         jbd_lock_bh_state(bh);
 472                         jbd2_free(jh->b_committed_data, bh->b_size);
 473                         jh->b_committed_data = NULL;
 474                         jbd_unlock_bh_state(bh);
 475                 }
 476                 jbd2_journal_refile_buffer(journal, jh);
 477         }
 478
 479         /*
 480          * Now try to drop any written-back buffers from the journal's
 481          * checkpoint lists.  We do this *before* commit because it potentially
 482          * frees some memory
 483          */
 484         spin_lock(&journal->j_list_lock);
 485         __jbd2_journal_clean_checkpoint_list(journal, false);
 486         spin_unlock(&journal->j_list_lock);
 487
 488         jbd_debug(3, "JBD2: commit phase 1\n");
 489
 490         /*
 491          * Clear revoked flag to reflect there is no revoked buffers
 492          * in the next transaction which is going to be started.
 493          */
 494         jbd2_clear_buffer_revoked_flags(journal);
 495
 496         /*
 497          * Switch to a new revoke table.
 498          */
 499         jbd2_journal_switch_revoke_table(journal);
 500
 501         /*
 502          * Reserved credits cannot be claimed anymore, free them
 503          */
 504         atomic_sub(atomic_read(&journal->j_reserved_credits),
 505                    &commit_transaction->t_outstanding_credits);
 506
 507         trace_jbd2_commit_flushing(journal, commit_transaction);
 508         stats.run.rs_flushing = jiffies;
 509         stats.run.rs_locked = jbd2_time_diff(stats.run.rs_locked,
 510                                              stats.run.rs_flushing);
 511
 512         commit_transaction->t_state = T_FLUSH;
 513         journal->j_committing_transaction = commit_transaction;
 514         journal->j_running_transaction = NULL;
 515         start_time = ktime_get();
 516         commit_transaction->t_log_start = journal->j_head;
 517         wake_up(&journal->j_wait_transaction_locked);
 518         write_unlock(&journal->j_state_lock);
 519
 520         jbd_debug(3, "JBD2: commit phase 2a\n");
 521
 522         /*
 523          * Now start flushing things to disk, in the order they appear
 524          * on the transaction lists.  Data blocks go first.
 525          */
 526         err = journal_submit_data_buffers(journal, commit_transaction);
 527         if (err)
 528                 jbd2_journal_abort(journal, err);
 529
 530         blk_start_plug(&plug);
 531         jbd2_journal_write_revoke_records(commit_transaction, &log_bufs);
 532
 533         jbd_debug(3, "JBD2: commit phase 2b\n");
 534
 535         /*
 536          * Way to go: we have now written out all of the data for a
 537          * transaction!  Now comes the tricky part: we need to write out
 538          * metadata.  Loop over the transaction's entire buffer list:
 539          */
 540         write_lock(&journal->j_state_lock);
 541         commit_transaction->t_state = T_COMMIT;
 542         write_unlock(&journal->j_state_lock);
 543
 544         trace_jbd2_commit_logging(journal, commit_transaction);
 545         stats.run.rs_logging = jiffies;
 546         stats.run.rs_flushing = jbd2_time_diff(stats.run.rs_flushing,
 547                                                stats.run.rs_logging);
 548         stats.run.rs_blocks =
 549                 atomic_read(&commit_transaction->t_outstanding_credits);
 550         stats.run.rs_blocks_logged = 0;
 551
 552         J_ASSERT(commit_transaction->t_nr_buffers <=
 553                  atomic_read(&commit_transaction->t_outstanding_credits));
 554
 555         err = 0;
 556         bufs = 0;
 557         descriptor = NULL;
 558         while (commit_transaction->t_buffers) {
 559
 560                 /* Find the next buffer to be journaled... */
 561
 562                 jh = commit_transaction->t_buffers;
 563
 564                 /* If we're in abort mode, we just un-journal the buffer and
 565                    release it. */
 566
 567                 if (is_journal_aborted(journal)) {
 568                         clear_buffer_jbddirty(jh2bh(jh));
 569                         JBUFFER_TRACE(jh, "journal is aborting: refile");
 570                         jbd2_buffer_abort_trigger(jh,
 571                                                   jh->b_frozen_data ?
 572                                                   jh->b_frozen_triggers :
 573                                                   jh->b_triggers);
 574                         jbd2_journal_refile_buffer(journal, jh);
 575                         /* If that was the last one, we need to clean up
 576                          * any descriptor buffers which may have been
 577                          * already allocated, even if we are now
 578                          * aborting. */
 579                         if (!commit_transaction->t_buffers)
 580                                 goto start_journal_io;
 581                         continue;
 582                 }
 583
 584                 /* Make sure we have a descriptor block in which to
 585                    record the metadata buffer. */
 586
 587                 if (!descriptor) {
 588                         J_ASSERT (bufs == 0);
 589
 590                         jbd_debug(4, "JBD2: get descriptor\n");
 591
 592                         descriptor = jbd2_journal_get_descriptor_buffer(
 593                                                         commit_transaction,
 594                                                         JBD2_DESCRIPTOR_BLOCK);
 595                         if (!descriptor) {
 596                                 jbd2_journal_abort(journal, -EIO);
 597                                 continue;
 598                         }
 599
 600                         jbd_debug(4, "JBD2: got buffer %llu (%p)\n",
 601                                 (unsigned long long)descriptor->b_blocknr,
 602                                 descriptor->b_data);
 603                         tagp = &descriptor->b_data[sizeof(journal_header_t)];
 604                         space_left = descriptor->b_size -
 605                                                 sizeof(journal_header_t);
 606                         first_tag = 1;
 607                         set_buffer_jwrite(descriptor);
 608                         set_buffer_dirty(descriptor);
 609                         wbuf[bufs++] = descriptor;
 610
 611                         /* Record it so that we can wait for IO
 612                            completion later */
 613                         BUFFER_TRACE(descriptor, "ph3: file as descriptor");
 614                         jbd2_file_log_bh(&log_bufs, descriptor);
 615                 }
 616
 617                 /* Where is the buffer to be written? */
 618
 619                 err = jbd2_journal_next_log_block(journal, &blocknr);
 620                 /* If the block mapping failed, just abandon the buffer
 621                    and repeat this loop: we'll fall into the
 622                    refile-on-abort condition above. */
 623                 if (err) {
 624                         jbd2_journal_abort(journal, err);
 625                         continue;
 626                 }
 627
 628                 /*
 629                  * start_this_handle() uses t_outstanding_credits to determine
 630                  * the free space in the log, but this counter is changed
 631                  * by jbd2_journal_next_log_block() also.
 632                  */
 633                 atomic_dec(&commit_transaction->t_outstanding_credits);
 634
 635                 /* Bump b_count to prevent truncate from stumbling over
 636                    the shadowed buffer!  @@@ This can go if we ever get
 637                    rid of the shadow pairing of buffers. */
 638                 atomic_inc(&jh2bh(jh)->b_count);
 639
 640                 /*
 641                  * Make a temporary IO buffer with which to write it out
 642                  * (this will requeue the metadata buffer to BJ_Shadow).
 643                  */
 644                 set_bit(BH_JWrite, &jh2bh(jh)->b_state);
 645                 JBUFFER_TRACE(jh, "ph3: write metadata");
 646                 flags = jbd2_journal_write_metadata_buffer(commit_transaction,
 647                                                 jh, &wbuf[bufs], blocknr);
 648                 if (flags < 0) {
 649                         jbd2_journal_abort(journal, flags);
 650                         continue;
 651                 }
 652                 jbd2_file_log_bh(&io_bufs, wbuf[bufs]);
 653
 654                 /* Record the new block's tag in the current descriptor
 655                    buffer */
 656
 657                 tag_flag = 0;
 658                 if (flags & 1)
 659                         tag_flag |= JBD2_FLAG_ESCAPE;
 660                 if (!first_tag)
 661                         tag_flag |= JBD2_FLAG_SAME_UUID;
 662
 663                 tag = (journal_block_tag_t *) tagp;
 664                 write_tag_block(journal, tag, jh2bh(jh)->b_blocknr);
 665                 tag->t_flags = cpu_to_be16(tag_flag);
 666                 jbd2_block_tag_csum_set(journal, tag, wbuf[bufs],
 667                                         commit_transaction->t_tid);
 668                 tagp += tag_bytes;
 669                 space_left -= tag_bytes;
 670                 bufs++;
 671
 672                 if (first_tag) {
 673                         memcpy (tagp, journal->j_uuid, 16);
 674                         tagp += 16;
 675                         space_left -= 16;
 676                         first_tag = 0;
 677                 }
 678
 679                 /* If there's no more to do, or if the descriptor is full,
 680                    let the IO rip! */
 681
 682                 if (bufs == journal->j_wbufsize ||
 683                     commit_transaction->t_buffers == NULL ||
 684                     space_left < tag_bytes + 16 + csum_size) {
 685
 686                         jbd_debug(4, "JBD2: Submit %d IOs\n", bufs);
 687
 688                         /* Write an end-of-descriptor marker before
 689                            submitting the IOs.  "tag" still points to
 690                            the last tag we set up. */
 691
 692                         tag->t_flags |= cpu_to_be16(JBD2_FLAG_LAST_TAG);
 693
 694                         jbd2_descriptor_block_csum_set(journal, descriptor);
 695 start_journal_io:
 696                         for (i = 0; i < bufs; i++) {
 697                                 struct buffer_head *bh = wbuf[i];
 698                                 /*
 699                                  * Compute checksum.
 700                                  */
 701                                 if (jbd2_has_feature_checksum(journal)) {
 702                                         crc32_sum =
 703                                             jbd2_checksum_data(crc32_sum, bh);
 704                                 }
 705
 706                                 lock_buffer(bh);
 707                                 clear_buffer_dirty(bh);
 708                                 set_buffer_uptodate(bh);
 709                                 bh->b_end_io = journal_end_buffer_io_sync;
 710                                 submit_bh(REQ_OP_WRITE, REQ_SYNC, bh);
 711                         }
 712                         cond_resched();
 713                         stats.run.rs_blocks_logged += bufs;
 714
 715                         /* Force a new descriptor to be generated next
 716                            time round the loop. */
 717                         descriptor = NULL;
 718                         bufs = 0;
 719                 }
 720         }
 721
 722         err = journal_finish_inode_data_buffers(journal, commit_transaction);
 723         if (err) {
 724                 printk(KERN_WARNING
 725                         "JBD2: Detected IO errors while flushing file data "
 726                        "on %s\n", journal->j_devname);
 727                 if (journal->j_flags & JBD2_ABORT_ON_SYNCDATA_ERR)
 728                         jbd2_journal_abort(journal, err);
 729                 err = 0;
 730         }
 731
 732         /*
 733          * Get current oldest transaction in the log before we issue flush
 734          * to the filesystem device. After the flush we can be sure that
 735          * blocks of all older transactions are checkpointed to persistent
 736          * storage and we will be safe to update journal start in the
 737          * superblock with the numbers we get here.
 738          */
 739         update_tail =
 740                 jbd2_journal_get_log_tail(journal, &first_tid, &first_block);
 741
 742         write_lock(&journal->j_state_lock);
 743         if (update_tail) {
 744                 long freed = first_block - journal->j_tail;
 745
 746                 if (first_block < journal->j_tail)
 747                         freed += journal->j_last - journal->j_first;
 748                 /* Update tail only if we free significant amount of space */
 749                 if (freed < journal->j_maxlen / 4)
 750                         update_tail = 0;
 751         }
 752         J_ASSERT(commit_transaction->t_state == T_COMMIT);
 753         commit_transaction->t_state = T_COMMIT_DFLUSH;
 754         write_unlock(&journal->j_state_lock);
 755
 756         /*
 757          * If the journal is not located on the file system device,
 758          * then we must flush the file system device before we issue
 759          * the commit record
 760          */
 761         if (commit_transaction->t_need_data_flush &&
 762             (journal->j_fs_dev != journal->j_dev) &&
 763             (journal->j_flags & JBD2_BARRIER))
 764                 blkdev_issue_flush(journal->j_fs_dev, GFP_NOFS, NULL);
 765
 766         /* Done it all: now write the commit record asynchronously. */
 767         if (jbd2_has_feature_async_commit(journal)) {
 768                 err = journal_submit_commit_record(journal, commit_transaction,
 769                                                  &cbh, crc32_sum);
 770                 if (err)
 771                         __jbd2_journal_abort_hard(journal);
 772         }
 773
 774         blk_finish_plug(&plug);
 775
 776         /* Lo and behold: we have just managed to send a transaction to
 777            the log.  Before we can commit it, wait for the IO so far to
 778            complete.  Control buffers being written are on the
 779            transaction's t_log_list queue, and metadata buffers are on
 780            the io_bufs list.
 781
 782            Wait for the buffers in reverse order.  That way we are
 783            less likely to be woken up until all IOs have completed, and
 784            so we incur less scheduling load.
 785         */
 786
 787         jbd_debug(3, "JBD2: commit phase 3\n");
 788
 789         while (!list_empty(&io_bufs)) {
 790                 struct buffer_head *bh = list_entry(io_bufs.prev,
 791                                                     struct buffer_head,
 792                                                     b_assoc_buffers);
 793
 794                 wait_on_buffer(bh);
 795                 cond_resched();
 796
 797                 if (unlikely(!buffer_uptodate(bh)))
 798                         err = -EIO;
 799                 jbd2_unfile_log_bh(bh);
 800
 801                 /*
 802                  * The list contains temporary buffer heads created by
 803                  * jbd2_journal_write_metadata_buffer().
 804                  */
 805                 BUFFER_TRACE(bh, "dumping temporary bh");
 806                 __brelse(bh);
 807                 J_ASSERT_BH(bh, atomic_read(&bh->b_count) == 0);
 808                 free_buffer_head(bh);
 809
 810                 /* We also have to refile the corresponding shadowed buffer */
 811                 jh = commit_transaction->t_shadow_list->b_tprev;
 812                 bh = jh2bh(jh);
 813                 clear_buffer_jwrite(bh);
 814                 J_ASSERT_BH(bh, buffer_jbddirty(bh));
 815                 J_ASSERT_BH(bh, !buffer_shadow(bh));
 816
 817                 /* The metadata is now released for reuse, but we need
 818                    to remember it against this transaction so that when
 819                    we finally commit, we can do any checkpointing
 820                    required. */
 821                 JBUFFER_TRACE(jh, "file as BJ_Forget");
 822                 jbd2_journal_file_buffer(jh, commit_transaction, BJ_Forget);
 823                 JBUFFER_TRACE(jh, "brelse shadowed buffer");
 824                 __brelse(bh);
 825         }
 826
 827         J_ASSERT (commit_transaction->t_shadow_list == NULL);
 828
 829         jbd_debug(3, "JBD2: commit phase 4\n");
 830
 831         /* Here we wait for the revoke record and descriptor record buffers */
 832         while (!list_empty(&log_bufs)) {
 833                 struct buffer_head *bh;
 834
 835                 bh = list_entry(log_bufs.prev, struct buffer_head, b_assoc_buffers);
 836                 wait_on_buffer(bh);
 837                 cond_resched();
 838
 839                 if (unlikely(!buffer_uptodate(bh)))
 840                         err = -EIO;
 841
 842                 BUFFER_TRACE(bh, "ph5: control buffer writeout done: unfile");
 843                 clear_buffer_jwrite(bh);
 844                 jbd2_unfile_log_bh(bh);
 845                 __brelse(bh);           /* One for getblk */
 846                 /* AKPM: bforget here */
 847         }
 848
 849         if (err)
 850                 jbd2_journal_abort(journal, err);
 851
 852         jbd_debug(3, "JBD2: commit phase 5\n");
 853         write_lock(&journal->j_state_lock);
 854         J_ASSERT(commit_transaction->t_state == T_COMMIT_DFLUSH);
 855         commit_transaction->t_state = T_COMMIT_JFLUSH;
 856         write_unlock(&journal->j_state_lock);
 857
 858         if (!jbd2_has_feature_async_commit(journal)) {
 859                 err = journal_submit_commit_record(journal, commit_transaction,
 860                                                 &cbh, crc32_sum);
 861                 if (err)
 862                         __jbd2_journal_abort_hard(journal);
 863         }
 864         if (cbh)
 865                 err = journal_wait_on_commit_record(journal, cbh);
 866         if (jbd2_has_feature_async_commit(journal) &&
 867             journal->j_flags & JBD2_BARRIER) {
 868                 blkdev_issue_flush(journal->j_dev, GFP_NOFS, NULL);
 869         }
 870
 871         if (err)
 872                 jbd2_journal_abort(journal, err);
 873
 874         /*
 875          * Now disk caches for filesystem device are flushed so we are safe to
 876          * erase checkpointed transactions from the log by updating journal
 877          * superblock.
 878          */
 879         if (update_tail)
 880                 jbd2_update_log_tail(journal, first_tid, first_block);
 881
 882         /* End of a transaction!  Finally, we can do checkpoint
 883            processing: any buffers committed as a result of this
 884            transaction can be removed from any checkpoint list it was on
 885            before. */
 886
 887         jbd_debug(3, "JBD2: commit phase 6\n");
 888
 889         J_ASSERT(list_empty(&commit_transaction->t_inode_list));
 890         J_ASSERT(commit_transaction->t_buffers == NULL);
 891         J_ASSERT(commit_transaction->t_checkpoint_list == NULL);
 892         J_ASSERT(commit_transaction->t_shadow_list == NULL);
 893
 894 restart_loop:
 895         /*
 896          * As there are other places (journal_unmap_buffer()) adding buffers
 897          * to this list we have to be careful and hold the j_list_lock.
 898          */
 899         spin_lock(&journal->j_list_lock);
 900         while (commit_transaction->t_forget) {
 901                 transaction_t *cp_transaction;
 902                 struct buffer_head *bh;
 903                 int try_to_free = 0;
 904
 905                 jh = commit_transaction->t_forget;
 906                 spin_unlock(&journal->j_list_lock);
 907                 bh = jh2bh(jh);
 908                 /*
 909                  * Get a reference so that bh cannot be freed before we are
 910                  * done with it.
 911                  */
 912                 get_bh(bh);
 913                 jbd_lock_bh_state(bh);
 914                 J_ASSERT_JH(jh, jh->b_transaction == commit_transaction);
 915
 916                 /*
 917                  * If there is undo-protected committed data against
 918                  * this buffer, then we can remove it now.  If it is a
 919                  * buffer needing such protection, the old frozen_data
 920                  * field now points to a committed version of the
 921                  * buffer, so rotate that field to the new committed
 922                  * data.
 923                  *
 924                  * Otherwise, we can just throw away the frozen data now.
 925                  *
 926                  * We also know that the frozen data has already fired
 927                  * its triggers if they exist, so we can clear that too.
 928                  */
 929                 if (jh->b_committed_data) {
 930                         jbd2_free(jh->b_committed_data, bh->b_size);
 931                         jh->b_committed_data = NULL;
 932                         if (jh->b_frozen_data) {
 933                                 jh->b_committed_data = jh->b_frozen_data;
 934                                 jh->b_frozen_data = NULL;
 935                                 jh->b_frozen_triggers = NULL;
 936                         }
 937                 } else if (jh->b_frozen_data) {
 938                         jbd2_free(jh->b_frozen_data, bh->b_size);
 939                         jh->b_frozen_data = NULL;
 940                         jh->b_frozen_triggers = NULL;
 941                 }
 942
 943                 spin_lock(&journal->j_list_lock);
 944                 cp_transaction = jh->b_cp_transaction;
 945                 if (cp_transaction) {
 946                         JBUFFER_TRACE(jh, "remove from old cp transaction");
 947                         cp_transaction->t_chp_stats.cs_dropped++;
 948                         __jbd2_journal_remove_checkpoint(jh);
 949                 }
 950
 951                 /* Only re-checkpoint the buffer_head if it is marked
 952                  * dirty.  If the buffer was added to the BJ_Forget list
 953                  * by jbd2_journal_forget, it may no longer be dirty and
 954                  * there's no point in keeping a checkpoint record for
 955                  * it. */
 956
 957                 /*
 958                 * A buffer which has been freed while still being journaled by
 959                 * a previous transaction.
 960                 */
 961                 if (buffer_freed(bh)) {
 962                         /*
 963                          * If the running transaction is the one containing
 964                          * "add to orphan" operation (b_next_transaction !=
 965                          * NULL), we have to wait for that transaction to
 966                          * commit before we can really get rid of the buffer.
 967                          * So just clear b_modified to not confuse transaction
 968                          * credit accounting and refile the buffer to
 969                          * BJ_Forget of the running transaction. If the just
 970                          * committed transaction contains "add to orphan"
 971                          * operation, we can completely invalidate the buffer
 972                          * now. We are rather through in that since the
 973                          * buffer may be still accessible when blocksize <
 974                          * pagesize and it is attached to the last partial
 975                          * page.
 976                          */
 977                         jh->b_modified = 0;
 978                         if (!jh->b_next_transaction) {
 979                                 clear_buffer_freed(bh);
 980                                 clear_buffer_jbddirty(bh);
 981                                 clear_buffer_mapped(bh);
 982                                 clear_buffer_new(bh);
 983                                 clear_buffer_req(bh);
 984                                 bh->b_bdev = NULL;
 985                         }
 986                 }
 987
 988                 if (buffer_jbddirty(bh)) {
 989                         JBUFFER_TRACE(jh, "add to new checkpointing trans");
 990                         __jbd2_journal_insert_checkpoint(jh, commit_transaction);
 991                         if (is_journal_aborted(journal))
 992                                 clear_buffer_jbddirty(bh);
 993                 } else {
 994                         J_ASSERT_BH(bh, !buffer_dirty(bh));
 995                         /*
 996                          * The buffer on BJ_Forget list and not jbddirty means
 997                          * it has been freed by this transaction and hence it
 998                          * could not have been reallocated until this
 999                          * transaction has committed. *BUT* it could be
1000                          * reallocated once we have written all the data to
1001                          * disk and before we process the buffer on BJ_Forget
1002                          * list.
1003                          */
1004                         if (!jh->b_next_transaction)
1005                                 try_to_free = 1;
1006                 }
1007                 JBUFFER_TRACE(jh, "refile or unfile buffer");
1008                 __jbd2_journal_refile_buffer(jh);
1009                 jbd_unlock_bh_state(bh);
1010                 if (try_to_free)
1011                         release_buffer_page(bh);        /* Drops bh reference */
1012                 else
1013                         __brelse(bh);
1014                 cond_resched_lock(&journal->j_list_lock);
1015         }
1016         spin_unlock(&journal->j_list_lock);
1017         /*
1018          * This is a bit sleazy.  We use j_list_lock to protect transition
1019          * of a transaction into T_FINISHED state and calling
1020          * __jbd2_journal_drop_transaction(). Otherwise we could race with
1021          * other checkpointing code processing the transaction...
1022          */
1023         write_lock(&journal->j_state_lock);
1024         spin_lock(&journal->j_list_lock);
1025         /*
1026          * Now recheck if some buffers did not get attached to the transaction
1027          * while the lock was dropped...
1028          */
1029         if (commit_transaction->t_forget) {
1030                 spin_unlock(&journal->j_list_lock);
1031                 write_unlock(&journal->j_state_lock);
1032                 goto restart_loop;
1033         }
1034
1035         /* Add the transaction to the checkpoint list
1036          * __journal_remove_checkpoint() can not destroy transaction
1037          * under us because it is not marked as T_FINISHED yet */
1038         if (journal->j_checkpoint_transactions == NULL) {
1039                 journal->j_checkpoint_transactions = commit_transaction;
1040                 commit_transaction->t_cpnext = commit_transaction;
1041                 commit_transaction->t_cpprev = commit_transaction;
1042         } else {
1043                 commit_transaction->t_cpnext =
1044                         journal->j_checkpoint_transactions;
1045                 commit_transaction->t_cpprev =
1046                         commit_transaction->t_cpnext->t_cpprev;
1047                 commit_transaction->t_cpnext->t_cpprev =
1048                         commit_transaction;
1049                 commit_transaction->t_cpprev->t_cpnext =
1050                                 commit_transaction;
1051         }
1052         spin_unlock(&journal->j_list_lock);
1053
1054         /* Done with this transaction! */
1055
1056         jbd_debug(3, "JBD2: commit phase 7\n");
1057
1058         J_ASSERT(commit_transaction->t_state == T_COMMIT_JFLUSH);
1059
1060         commit_transaction->t_start = jiffies;
1061         stats.run.rs_logging = jbd2_time_diff(stats.run.rs_logging,
1062                                               commit_transaction->t_start);
1063
1064         /*
1065          * File the transaction statistics
1066          */
1067         stats.ts_tid = commit_transaction->t_tid;
1068         stats.run.rs_handle_count =
1069                 atomic_read(&commit_transaction->t_handle_count);
1070         trace_jbd2_run_stats(journal->j_fs_dev->bd_dev,
1071                              commit_transaction->t_tid, &stats.run);
1072         stats.ts_requested = (commit_transaction->t_requested) ? 1 : 0;
1073
1074         commit_transaction->t_state = T_COMMIT_CALLBACK;
1075         J_ASSERT(commit_transaction == journal->j_committing_transaction);
1076         journal->j_commit_sequence = commit_transaction->t_tid;
1077         journal->j_committing_transaction = NULL;
1078         commit_time = ktime_to_ns(ktime_sub(ktime_get(), start_time));
1079
1080         /*
1081          * weight the commit time higher than the average time so we don't
1082          * react too strongly to vast changes in the commit time
1083          */
1084         if (likely(journal->j_average_commit_time))
1085                 journal->j_average_commit_time = (commit_time +
1086                                 journal->j_average_commit_time*3) / 4;
1087         else
1088                 journal->j_average_commit_time = commit_time;
1089
1090         write_unlock(&journal->j_state_lock);
1091
1092         if (journal->j_commit_callback)
1093                 journal->j_commit_callback(journal, commit_transaction);
1094
1095         trace_jbd2_end_commit(journal, commit_transaction);
1096         jbd_debug(1, "JBD2: commit %d complete, head %d\n",
1097                   journal->j_commit_sequence, journal->j_tail_sequence);
1098
1099         write_lock(&journal->j_state_lock);
1100         spin_lock(&journal->j_list_lock);
1101         commit_transaction->t_state = T_FINISHED;
1102         /* Check if the transaction can be dropped now that we are finished */
1103         if (commit_transaction->t_checkpoint_list == NULL &&
1104             commit_transaction->t_checkpoint_io_list == NULL) {
1105                 __jbd2_journal_drop_transaction(journal, commit_transaction);
1106                 jbd2_journal_free_transaction(commit_transaction);
1107         }
1108         spin_unlock(&journal->j_list_lock);
1109         write_unlock(&journal->j_state_lock);
1110         wake_up(&journal->j_wait_done_commit);
1111
1112         /*
1113          * Calculate overall stats
1114          */
1115         spin_lock(&journal->j_history_lock);
1116         journal->j_stats.ts_tid++;
1117         journal->j_stats.ts_requested += stats.ts_requested;
1118         journal->j_stats.run.rs_wait += stats.run.rs_wait;
1119         journal->j_stats.run.rs_request_delay += stats.run.rs_request_delay;
1120         journal->j_stats.run.rs_running += stats.run.rs_running;
1121         journal->j_stats.run.rs_locked += stats.run.rs_locked;
1122         journal->j_stats.run.rs_flushing += stats.run.rs_flushing;
1123         journal->j_stats.run.rs_logging += stats.run.rs_logging;
1124         journal->j_stats.run.rs_handle_count += stats.run.rs_handle_count;
1125         journal->j_stats.run.rs_blocks += stats.run.rs_blocks;
1126         journal->j_stats.run.rs_blocks_logged += stats.run.rs_blocks_logged;
1127         spin_unlock(&journal->j_history_lock);
1128 }