fs/jbd2/commit.c

   1 /*
   2  * linux/fs/jbd2/commit.c
   3  *
   4  * Written by Stephen C. Tweedie <sct@redhat.com>, 1998
   5  *
   6  * Copyright 1998 Red Hat corp --- All Rights Reserved
   7  *
   8  * This file is part of the Linux kernel and is made available under
   9  * the terms of the GNU General Public License, version 2, or at your
  10  * option, any later version, incorporated herein by reference.
  11  *
  12  * Journal commit routines for the generic filesystem journaling code;
  13  * part of the ext2fs journaling system.
  14  */
  15
  16 #include <linux/time.h>
  17 #include <linux/fs.h>
  18 #include <linux/jbd2.h>
  19 #include <linux/errno.h>
  20 #include <linux/slab.h>
  21 #include <linux/mm.h>
  22 #include <linux/pagemap.h>
  23 #include <linux/jiffies.h>
  24 #include <linux/crc32.h>
  25 #include <linux/writeback.h>
  26 #include <linux/backing-dev.h>
  27 #include <linux/bio.h>
  28 #include <linux/blkdev.h>
  29 #include <linux/bitops.h>
  30 #include <trace/events/jbd2.h>
  31
  32 /*
  33  * IO end handler for temporary buffer_heads handling writes to the journal.
  34  */
  35 static void journal_end_buffer_io_sync(struct buffer_head *bh, int uptodate)
  36 {
  37         struct buffer_head *orig_bh = bh->b_private;
  38
  39         BUFFER_TRACE(bh, "");
  40         if (uptodate)
  41                 set_buffer_uptodate(bh);
  42         else
  43                 clear_buffer_uptodate(bh);
  44         if (orig_bh) {
  45                 clear_bit_unlock(BH_Shadow, &orig_bh->b_state);
  46                 smp_mb__after_atomic();
  47                 wake_up_bit(&orig_bh->b_state, BH_Shadow);
  48         }
  49         unlock_buffer(bh);
  50 }
  51
  52 /*
  53  * When an ext4 file is truncated, it is possible that some pages are not
  54  * successfully freed, because they are attached to a committing transaction.
  55  * After the transaction commits, these pages are left on the LRU, with no
  56  * ->mapping, and with attached buffers.  These pages are trivially reclaimable
  57  * by the VM, but their apparent absence upsets the VM accounting, and it makes
  58  * the numbers in /proc/meminfo look odd.
  59  *
  60  * So here, we have a buffer which has just come off the forget list.  Look to
  61  * see if we can strip all buffers from the backing page.
  62  *
  63  * Called under lock_journal(), and possibly under journal_datalist_lock.  The
  64  * caller provided us with a ref against the buffer, and we drop that here.
  65  */
  66 static void release_buffer_page(struct buffer_head *bh)
  67 {
  68         struct page *page;
  69
  70         if (buffer_dirty(bh))
  71                 goto nope;
  72         if (atomic_read(&bh->b_count) != 1)
  73                 goto nope;
  74         page = bh->b_page;
  75         if (!page)
  76                 goto nope;
  77         if (page->mapping)
  78                 goto nope;
  79
  80         /* OK, it's a truncated page */
  81         if (!trylock_page(page))
  82                 goto nope;
  83
  84         page_cache_get(page);
  85         __brelse(bh);
  86         try_to_free_buffers(page);
  87         unlock_page(page);
  88         page_cache_release(page);
  89         return;
  90
  91 nope:
  92         __brelse(bh);
  93 }
  94
  95 static void jbd2_commit_block_csum_set(journal_t *j, struct buffer_head *bh)
  96 {
  97         struct commit_header *h;
  98         __u32 csum;
  99
 100         if (!jbd2_journal_has_csum_v2or3(j))
 101                 return;
 102
 103         h = (struct commit_header *)(bh->b_data);
 104         h->h_chksum_type = 0;
 105         h->h_chksum_size = 0;
 106         h->h_chksum[0] = 0;
 107         csum = jbd2_chksum(j, j->j_csum_seed, bh->b_data, j->j_blocksize);
 108         h->h_chksum[0] = cpu_to_be32(csum);
 109 }
 110
 111 /*
 112  * Done it all: now submit the commit record.  We should have
 113  * cleaned up our previous buffers by now, so if we are in abort
 114  * mode we can now just skip the rest of the journal write
 115  * entirely.
 116  *
 117  * Returns 1 if the journal needs to be aborted or 0 on success
 118  */
 119 static int journal_submit_commit_record(journal_t *journal,
 120                                         transaction_t *commit_transaction,
 121                                         struct buffer_head **cbh,
 122                                         __u32 crc32_sum)
 123 {
 124         struct commit_header *tmp;
 125         struct buffer_head *bh;
 126         int ret;
 127         struct timespec now = current_kernel_time();
 128
 129         *cbh = NULL;
 130
 131         if (is_journal_aborted(journal))
 132                 return 0;
 133
 134         bh = jbd2_journal_get_descriptor_buffer(commit_transaction,
 135                                                 JBD2_COMMIT_BLOCK);
 136         if (!bh)
 137                 return 1;
 138
 139         tmp = (struct commit_header *)bh->b_data;
 140         tmp->h_commit_sec = cpu_to_be64(now.tv_sec);
 141         tmp->h_commit_nsec = cpu_to_be32(now.tv_nsec);
 142
 143         if (jbd2_has_feature_checksum(journal)) {
 144                 tmp->h_chksum_type      = JBD2_CRC32_CHKSUM;
 145                 tmp->h_chksum_size      = JBD2_CRC32_CHKSUM_SIZE;
 146                 tmp->h_chksum[0]        = cpu_to_be32(crc32_sum);
 147         }
 148         jbd2_commit_block_csum_set(journal, bh);
 149
 150         BUFFER_TRACE(bh, "submit commit block");
 151         lock_buffer(bh);
 152         clear_buffer_dirty(bh);
 153         set_buffer_uptodate(bh);
 154         bh->b_end_io = journal_end_buffer_io_sync;
 155
 156         if (journal->j_flags & JBD2_BARRIER &&
 157             !jbd2_has_feature_async_commit(journal))
 158                 ret = submit_bh(WRITE_SYNC | WRITE_FLUSH_FUA, bh);
 159         else
 160                 ret = submit_bh(WRITE_SYNC, bh);
 161
 162         *cbh = bh;
 163         return ret;
 164 }
 165
 166 /*
 167  * This function along with journal_submit_commit_record
 168  * allows to write the commit record asynchronously.
 169  */
 170 static int journal_wait_on_commit_record(journal_t *journal,
 171                                          struct buffer_head *bh)
 172 {
 173         int ret = 0;
 174
 175         clear_buffer_dirty(bh);
 176         wait_on_buffer(bh);
 177
 178         if (unlikely(!buffer_uptodate(bh)))
 179                 ret = -EIO;
 180         put_bh(bh);            /* One for getblk() */
 181
 182         return ret;
 183 }
 184
 185 /*
 186  * write the filemap data using writepage() address_space_operations.
 187  * We don't do block allocation here even for delalloc. We don't
 188  * use writepages() because with dealyed allocation we may be doing
 189  * block allocation in writepages().
 190  */
 191 static int journal_submit_inode_data_buffers(struct address_space *mapping)
 192 {
 193         int ret;
 194         struct writeback_control wbc = {
 195                 .sync_mode =  WB_SYNC_ALL,
 196                 .nr_to_write = mapping->nrpages * 2,
 197                 .range_start = 0,
 198                 .range_end = i_size_read(mapping->host),
 199         };
 200
 201         ret = generic_writepages(mapping, &wbc);
 202         return ret;
 203 }
 204
 205 /*
 206  * Submit all the data buffers of inode associated with the transaction to
 207  * disk.
 208  *
 209  * We are in a committing transaction. Therefore no new inode can be added to
 210  * our inode list. We use JI_COMMIT_RUNNING flag to protect inode we currently
 211  * operate on from being released while we write out pages.
 212  */
 213 static int journal_submit_data_buffers(journal_t *journal,
 214                 transaction_t *commit_transaction)
 215 {
 216         struct jbd2_inode *jinode;
 217         int err, ret = 0;
 218         struct address_space *mapping;
 219
 220         spin_lock(&journal->j_list_lock);
 221         list_for_each_entry(jinode, &commit_transaction->t_inode_list, i_list) {
 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                 jinode->i_flags |= JI_COMMIT_RUNNING;
 260                 spin_unlock(&journal->j_list_lock);
 261                 err = filemap_fdatawait(jinode->i_vfs_inode->i_mapping);
 262                 if (err) {
 263                         /*
 264                          * Because AS_EIO is cleared by
 265                          * filemap_fdatawait_range(), set it again so
 266                          * that user process can get -EIO from fsync().
 267                          */
 268                         set_bit(AS_EIO,
 269                                 &jinode->i_vfs_inode->i_mapping->flags);
 270
 271                         if (!ret)
 272                                 ret = err;
 273                 }
 274                 spin_lock(&journal->j_list_lock);
 275                 jinode->i_flags &= ~JI_COMMIT_RUNNING;
 276                 smp_mb();
 277                 wake_up_bit(&jinode->i_flags, __JI_COMMIT_RUNNING);
 278         }
 279
 280         /* Now refile inode to proper lists */
 281         list_for_each_entry_safe(jinode, next_i,
 282                                  &commit_transaction->t_inode_list, i_list) {
 283                 list_del(&jinode->i_list);
 284                 if (jinode->i_next_transaction) {
 285                         jinode->i_transaction = jinode->i_next_transaction;
 286                         jinode->i_next_transaction = NULL;
 287                         list_add(&jinode->i_list,
 288                                 &jinode->i_transaction->t_inode_list);
 289                 } else {
 290                         jinode->i_transaction = NULL;
 291                 }
 292         }
 293         spin_unlock(&journal->j_list_lock);
 294
 295         return ret;
 296 }
 297
 298 static __u32 jbd2_checksum_data(__u32 crc32_sum, struct buffer_head *bh)
 299 {
 300         struct page *page = bh->b_page;
 301         char *addr;
 302         __u32 checksum;
 303
 304         addr = kmap_atomic(page);
 305         checksum = crc32_be(crc32_sum,
 306                 (void *)(addr + offset_in_page(bh->b_data)), bh->b_size);
 307         kunmap_atomic(addr);
 308
 309         return checksum;
 310 }
 311
 312 static void write_tag_block(journal_t *j, journal_block_tag_t *tag,
 313                                    unsigned long long block)
 314 {
 315         tag->t_blocknr = cpu_to_be32(block & (u32)~0);
 316         if (jbd2_has_feature_64bit(j))
 317                 tag->t_blocknr_high = cpu_to_be32((block >> 31) >> 1);
 318 }
 319
 320 static void jbd2_block_tag_csum_set(journal_t *j, journal_block_tag_t *tag,
 321                                     struct buffer_head *bh, __u32 sequence)
 322 {
 323         journal_block_tag3_t *tag3 = (journal_block_tag3_t *)tag;
 324         struct page *page = bh->b_page;
 325         __u8 *addr;
 326         __u32 csum32;
 327         __be32 seq;
 328
 329         if (!jbd2_journal_has_csum_v2or3(j))
 330                 return;
 331
 332         seq = cpu_to_be32(sequence);
 333         addr = kmap_atomic(page);
 334         csum32 = jbd2_chksum(j, j->j_csum_seed, (__u8 *)&seq, sizeof(seq));
 335         csum32 = jbd2_chksum(j, csum32, addr + offset_in_page(bh->b_data),
 336                              bh->b_size);
 337         kunmap_atomic(addr);
 338
 339         if (jbd2_has_feature_csum3(j))
 340                 tag3->t_checksum = cpu_to_be32(csum32);
 341         else
 342                 tag->t_checksum = cpu_to_be16(csum32);
 343 }
 344 /*
 345  * jbd2_journal_commit_transaction
 346  *
 347  * The primary function for committing a transaction to the log.  This
 348  * function is called by the journal thread to begin a complete commit.
 349  */
 350 void jbd2_journal_commit_transaction(journal_t *journal)
 351 {
 352         struct transaction_stats_s stats;
 353         transaction_t *commit_transaction;
 354         struct journal_head *jh;
 355         struct buffer_head *descriptor;
 356         struct buffer_head **wbuf = journal->j_wbuf;
 357         int bufs;
 358         int flags;
 359         int err;
 360         unsigned long long blocknr;
 361         ktime_t start_time;
 362         u64 commit_time;
 363         char *tagp = NULL;
 364         journal_block_tag_t *tag = NULL;
 365         int space_left = 0;
 366         int first_tag = 0;
 367         int tag_flag;
 368         int i;
 369         int tag_bytes = journal_tag_bytes(journal);
 370         struct buffer_head *cbh = NULL; /* For transactional checksums */
 371         __u32 crc32_sum = ~0;
 372         struct blk_plug plug;
 373         /* Tail of the journal */
 374         unsigned long first_block;
 375         tid_t first_tid;
 376         int update_tail;
 377         int csum_size = 0;
 378         LIST_HEAD(io_bufs);
 379         LIST_HEAD(log_bufs);
 380
 381         if (jbd2_journal_has_csum_v2or3(journal))
 382                 csum_size = sizeof(struct jbd2_journal_block_tail);
 383
 384         /*
 385          * First job: lock down the current transaction and wait for
 386          * all outstanding updates to complete.
 387          */
 388
 389         /* Do we need to erase the effects of a prior jbd2_journal_flush? */
 390         if (journal->j_flags & JBD2_FLUSHED) {
 391                 jbd_debug(3, "super block updated\n");
 392                 mutex_lock(&journal->j_checkpoint_mutex);
 393                 /*
 394                  * We hold j_checkpoint_mutex so tail cannot change under us.
 395                  * We don't need any special data guarantees for writing sb
 396                  * since journal is empty and it is ok for write to be
 397                  * flushed only with transaction commit.
 398                  */
 399                 jbd2_journal_update_sb_log_tail(journal,
 400                                                 journal->j_tail_sequence,
 401                                                 journal->j_tail,
 402                                                 WRITE_SYNC);
 403                 mutex_unlock(&journal->j_checkpoint_mutex);
 404         } else {
 405                 jbd_debug(3, "superblock not updated\n");
 406         }
 407
 408         J_ASSERT(journal->j_running_transaction != NULL);
 409         J_ASSERT(journal->j_committing_transaction == NULL);
 410
 411         commit_transaction = journal->j_running_transaction;
 412
 413         trace_jbd2_start_commit(journal, commit_transaction);
 414         jbd_debug(1, "JBD2: starting commit of transaction %d\n",
 415                         commit_transaction->t_tid);
 416
 417         write_lock(&journal->j_state_lock);
 418         J_ASSERT(commit_transaction->t_state == T_RUNNING);
 419         commit_transaction->t_state = T_LOCKED;
 420
 421         trace_jbd2_commit_locking(journal, commit_transaction);
 422         stats.run.rs_wait = commit_transaction->t_max_wait;
 423         stats.run.rs_request_delay = 0;
 424         stats.run.rs_locked = jiffies;
 425         if (commit_transaction->t_requested)
 426                 stats.run.rs_request_delay =
 427                         jbd2_time_diff(commit_transaction->t_requested,
 428                                        stats.run.rs_locked);
 429         stats.run.rs_running = jbd2_time_diff(commit_transaction->t_start,
 430                                               stats.run.rs_locked);
 431
 432         spin_lock(&commit_transaction->t_handle_lock);
 433         while (atomic_read(&commit_transaction->t_updates)) {
 434                 DEFINE_WAIT(wait);
 435
 436                 prepare_to_wait(&journal->j_wait_updates, &wait,
 437                                         TASK_UNINTERRUPTIBLE);
 438                 if (atomic_read(&commit_transaction->t_updates)) {
 439                         spin_unlock(&commit_transaction->t_handle_lock);
 440                         write_unlock(&journal->j_state_lock);
 441                         schedule();
 442                         write_lock(&journal->j_state_lock);
 443                         spin_lock(&commit_transaction->t_handle_lock);
 444                 }
 445                 finish_wait(&journal->j_wait_updates, &wait);
 446         }
 447         spin_unlock(&commit_transaction->t_handle_lock);
 448
 449         J_ASSERT (atomic_read(&commit_transaction->t_outstanding_credits) <=
 450                         journal->j_max_transaction_buffers);
 451
 452         /*
 453          * First thing we are allowed to do is to discard any remaining
 454          * BJ_Reserved buffers.  Note, it is _not_ permissible to assume
 455          * that there are no such buffers: if a large filesystem
 456          * operation like a truncate needs to split itself over multiple
 457          * transactions, then it may try to do a jbd2_journal_restart() while
 458          * there are still BJ_Reserved buffers outstanding.  These must
 459          * be released cleanly from the current transaction.
 460          *
 461          * In this case, the filesystem must still reserve write access
 462          * again before modifying the buffer in the new transaction, but
 463          * we do not require it to remember exactly which old buffers it
 464          * has reserved.  This is consistent with the existing behaviour
 465          * that multiple jbd2_journal_get_write_access() calls to the same
 466          * buffer are perfectly permissible.
 467          */
 468         while (commit_transaction->t_reserved_list) {
 469                 jh = commit_transaction->t_reserved_list;
 470                 JBUFFER_TRACE(jh, "reserved, unused: refile");
 471                 /*
 472                  * A jbd2_journal_get_undo_access()+jbd2_journal_release_buffer() may
 473                  * leave undo-committed data.
 474                  */
 475                 if (jh->b_committed_data) {
 476                         struct buffer_head *bh = jh2bh(jh);
 477
 478                         jbd_lock_bh_state(bh);
 479                         jbd2_free(jh->b_committed_data, bh->b_size);
 480                         jh->b_committed_data = NULL;
 481                         jbd_unlock_bh_state(bh);
 482                 }
 483                 jbd2_journal_refile_buffer(journal, jh);
 484         }
 485
 486         /*
 487          * Now try to drop any written-back buffers from the journal's
 488          * checkpoint lists.  We do this *before* commit because it potentially
 489          * frees some memory
 490          */
 491         spin_lock(&journal->j_list_lock);
 492         __jbd2_journal_clean_checkpoint_list(journal, false);
 493         spin_unlock(&journal->j_list_lock);
 494
 495         jbd_debug(3, "JBD2: commit phase 1\n");
 496
 497         /*
 498          * Clear revoked flag to reflect there is no revoked buffers
 499          * in the next transaction which is going to be started.
 500          */
 501         jbd2_clear_buffer_revoked_flags(journal);
 502
 503         /*
 504          * Switch to a new revoke table.
 505          */
 506         jbd2_journal_switch_revoke_table(journal);
 507
 508         /*
 509          * Reserved credits cannot be claimed anymore, free them
 510          */
 511         atomic_sub(atomic_read(&journal->j_reserved_credits),
 512                    &commit_transaction->t_outstanding_credits);
 513
 514         trace_jbd2_commit_flushing(journal, commit_transaction);
 515         stats.run.rs_flushing = jiffies;
 516         stats.run.rs_locked = jbd2_time_diff(stats.run.rs_locked,
 517                                              stats.run.rs_flushing);
 518
 519         commit_transaction->t_state = T_FLUSH;
 520         journal->j_committing_transaction = commit_transaction;
 521         journal->j_running_transaction = NULL;
 522         start_time = ktime_get();
 523         commit_transaction->t_log_start = journal->j_head;
 524         wake_up(&journal->j_wait_transaction_locked);
 525         write_unlock(&journal->j_state_lock);
 526
 527         jbd_debug(3, "JBD2: commit phase 2a\n");
 528
 529         /*
 530          * Now start flushing things to disk, in the order they appear
 531          * on the transaction lists.  Data blocks go first.
 532          */
 533         err = journal_submit_data_buffers(journal, commit_transaction);
 534         if (err)
 535                 jbd2_journal_abort(journal, err);
 536
 537         blk_start_plug(&plug);
 538         jbd2_journal_write_revoke_records(commit_transaction, &log_bufs);
 539
 540         jbd_debug(3, "JBD2: commit phase 2b\n");
 541
 542         /*
 543          * Way to go: we have now written out all of the data for a
 544          * transaction!  Now comes the tricky part: we need to write out
 545          * metadata.  Loop over the transaction's entire buffer list:
 546          */
 547         write_lock(&journal->j_state_lock);
 548         commit_transaction->t_state = T_COMMIT;
 549         write_unlock(&journal->j_state_lock);
 550
 551         trace_jbd2_commit_logging(journal, commit_transaction);
 552         stats.run.rs_logging = jiffies;
 553         stats.run.rs_flushing = jbd2_time_diff(stats.run.rs_flushing,
 554                                                stats.run.rs_logging);
 555         stats.run.rs_blocks =
 556                 atomic_read(&commit_transaction->t_outstanding_credits);
 557         stats.run.rs_blocks_logged = 0;
 558
 559         J_ASSERT(commit_transaction->t_nr_buffers <=
 560                  atomic_read(&commit_transaction->t_outstanding_credits));
 561
 562         err = 0;
 563         bufs = 0;
 564         descriptor = NULL;
 565         while (commit_transaction->t_buffers) {
 566
 567                 /* Find the next buffer to be journaled... */
 568
 569                 jh = commit_transaction->t_buffers;
 570
 571                 /* If we're in abort mode, we just un-journal the buffer and
 572                    release it. */
 573
 574                 if (is_journal_aborted(journal)) {
 575                         clear_buffer_jbddirty(jh2bh(jh));
 576                         JBUFFER_TRACE(jh, "journal is aborting: refile");
 577                         jbd2_buffer_abort_trigger(jh,
 578                                                   jh->b_frozen_data ?
 579                                                   jh->b_frozen_triggers :
 580                                                   jh->b_triggers);
 581                         jbd2_journal_refile_buffer(journal, jh);
 582                         /* If that was the last one, we need to clean up
 583                          * any descriptor buffers which may have been
 584                          * already allocated, even if we are now
 585                          * aborting. */
 586                         if (!commit_transaction->t_buffers)
 587                                 goto start_journal_io;
 588                         continue;
 589                 }
 590
 591                 /* Make sure we have a descriptor block in which to
 592                    record the metadata buffer. */
 593
 594                 if (!descriptor) {
 595                         J_ASSERT (bufs == 0);
 596
 597                         jbd_debug(4, "JBD2: get descriptor\n");
 598
 599                         descriptor = jbd2_journal_get_descriptor_buffer(
 600                                                         commit_transaction,
 601                                                         JBD2_DESCRIPTOR_BLOCK);
 602                         if (!descriptor) {
 603                                 jbd2_journal_abort(journal, -EIO);
 604                                 continue;
 605                         }
 606
 607                         jbd_debug(4, "JBD2: got buffer %llu (%p)\n",
 608                                 (unsigned long long)descriptor->b_blocknr,
 609                                 descriptor->b_data);
 610                         tagp = &descriptor->b_data[sizeof(journal_header_t)];
 611                         space_left = descriptor->b_size -
 612                                                 sizeof(journal_header_t);
 613                         first_tag = 1;
 614                         set_buffer_jwrite(descriptor);
 615                         set_buffer_dirty(descriptor);
 616                         wbuf[bufs++] = descriptor;
 617
 618                         /* Record it so that we can wait for IO
 619                            completion later */
 620                         BUFFER_TRACE(descriptor, "ph3: file as descriptor");
 621                         jbd2_file_log_bh(&log_bufs, descriptor);
 622                 }
 623
 624                 /* Where is the buffer to be written? */
 625
 626                 err = jbd2_journal_next_log_block(journal, &blocknr);
 627                 /* If the block mapping failed, just abandon the buffer
 628                    and repeat this loop: we'll fall into the
 629                    refile-on-abort condition above. */
 630                 if (err) {
 631                         jbd2_journal_abort(journal, err);
 632                         continue;
 633                 }
 634
 635                 /*
 636                  * start_this_handle() uses t_outstanding_credits to determine
 637                  * the free space in the log, but this counter is changed
 638                  * by jbd2_journal_next_log_block() also.
 639                  */
 640                 atomic_dec(&commit_transaction->t_outstanding_credits);
 641
 642                 /* Bump b_count to prevent truncate from stumbling over
 643                    the shadowed buffer!  @@@ This can go if we ever get
 644                    rid of the shadow pairing of buffers. */
 645                 atomic_inc(&jh2bh(jh)->b_count);
 646
 647                 /*
 648                  * Make a temporary IO buffer with which to write it out
 649                  * (this will requeue the metadata buffer to BJ_Shadow).
 650                  */
 651                 set_bit(BH_JWrite, &jh2bh(jh)->b_state);
 652                 JBUFFER_TRACE(jh, "ph3: write metadata");
 653                 flags = jbd2_journal_write_metadata_buffer(commit_transaction,
 654                                                 jh, &wbuf[bufs], blocknr);
 655                 if (flags < 0) {
 656                         jbd2_journal_abort(journal, flags);
 657                         continue;
 658                 }
 659                 jbd2_file_log_bh(&io_bufs, wbuf[bufs]);
 660
 661                 /* Record the new block's tag in the current descriptor
 662                    buffer */
 663
 664                 tag_flag = 0;
 665                 if (flags & 1)
 666                         tag_flag |= JBD2_FLAG_ESCAPE;
 667                 if (!first_tag)
 668                         tag_flag |= JBD2_FLAG_SAME_UUID;
 669
 670                 tag = (journal_block_tag_t *) tagp;
 671                 write_tag_block(journal, tag, jh2bh(jh)->b_blocknr);
 672                 tag->t_flags = cpu_to_be16(tag_flag);
 673                 jbd2_block_tag_csum_set(journal, tag, wbuf[bufs],
 674                                         commit_transaction->t_tid);
 675                 tagp += tag_bytes;
 676                 space_left -= tag_bytes;
 677                 bufs++;
 678
 679                 if (first_tag) {
 680                         memcpy (tagp, journal->j_uuid, 16);
 681                         tagp += 16;
 682                         space_left -= 16;
 683                         first_tag = 0;
 684                 }
 685
 686                 /* If there's no more to do, or if the descriptor is full,
 687                    let the IO rip! */
 688
 689                 if (bufs == journal->j_wbufsize ||
 690                     commit_transaction->t_buffers == NULL ||
 691                     space_left < tag_bytes + 16 + csum_size) {
 692
 693                         jbd_debug(4, "JBD2: Submit %d IOs\n", bufs);
 694
 695                         /* Write an end-of-descriptor marker before
 696                            submitting the IOs.  "tag" still points to
 697                            the last tag we set up. */
 698
 699                         tag->t_flags |= cpu_to_be16(JBD2_FLAG_LAST_TAG);
 700
 701                         jbd2_descriptor_block_csum_set(journal, descriptor);
 702 start_journal_io:
 703                         for (i = 0; i < bufs; i++) {
 704                                 struct buffer_head *bh = wbuf[i];
 705                                 /*
 706                                  * Compute checksum.
 707                                  */
 708                                 if (jbd2_has_feature_checksum(journal)) {
 709                                         crc32_sum =
 710                                             jbd2_checksum_data(crc32_sum, bh);
 711                                 }
 712
 713                                 lock_buffer(bh);
 714                                 clear_buffer_dirty(bh);
 715                                 set_buffer_uptodate(bh);
 716                                 bh->b_end_io = journal_end_buffer_io_sync;
 717                                 submit_bh(WRITE_SYNC, bh);
 718                         }
 719                         cond_resched();
 720                         stats.run.rs_blocks_logged += bufs;
 721
 722                         /* Force a new descriptor to be generated next
 723                            time round the loop. */
 724                         descriptor = NULL;
 725                         bufs = 0;
 726                 }
 727         }
 728
 729         err = journal_finish_inode_data_buffers(journal, commit_transaction);
 730         if (err) {
 731                 printk(KERN_WARNING
 732                         "JBD2: Detected IO errors while flushing file data "
 733                        "on %s\n", journal->j_devname);
 734                 if (journal->j_flags & JBD2_ABORT_ON_SYNCDATA_ERR)
 735                         jbd2_journal_abort(journal, err);
 736                 err = 0;
 737         }
 738
 739         /*
 740          * Get current oldest transaction in the log before we issue flush
 741          * to the filesystem device. After the flush we can be sure that
 742          * blocks of all older transactions are checkpointed to persistent
 743          * storage and we will be safe to update journal start in the
 744          * superblock with the numbers we get here.
 745          */
 746         update_tail =
 747                 jbd2_journal_get_log_tail(journal, &first_tid, &first_block);
 748
 749         write_lock(&journal->j_state_lock);
 750         if (update_tail) {
 751                 long freed = first_block - journal->j_tail;
 752
 753                 if (first_block < journal->j_tail)
 754                         freed += journal->j_last - journal->j_first;
 755                 /* Update tail only if we free significant amount of space */
 756                 if (freed < journal->j_maxlen / 4)
 757                         update_tail = 0;
 758         }
 759         J_ASSERT(commit_transaction->t_state == T_COMMIT);
 760         commit_transaction->t_state = T_COMMIT_DFLUSH;
 761         write_unlock(&journal->j_state_lock);
 762
 763         /*
 764          * If the journal is not located on the file system device,
 765          * then we must flush the file system device before we issue
 766          * the commit record
 767          */
 768         if (commit_transaction->t_need_data_flush &&
 769             (journal->j_fs_dev != journal->j_dev) &&
 770             (journal->j_flags & JBD2_BARRIER))
 771                 blkdev_issue_flush(journal->j_fs_dev, GFP_NOFS, NULL);
 772
 773         /* Done it all: now write the commit record asynchronously. */
 774         if (jbd2_has_feature_async_commit(journal)) {
 775                 err = journal_submit_commit_record(journal, commit_transaction,
 776                                                  &cbh, crc32_sum);
 777                 if (err)
 778                         __jbd2_journal_abort_hard(journal);
 779         }
 780
 781         blk_finish_plug(&plug);
 782
 783         /* Lo and behold: we have just managed to send a transaction to
 784            the log.  Before we can commit it, wait for the IO so far to
 785            complete.  Control buffers being written are on the
 786            transaction's t_log_list queue, and metadata buffers are on
 787            the io_bufs list.
 788
 789            Wait for the buffers in reverse order.  That way we are
 790            less likely to be woken up until all IOs have completed, and
 791            so we incur less scheduling load.
 792         */
 793
 794         jbd_debug(3, "JBD2: commit phase 3\n");
 795
 796         while (!list_empty(&io_bufs)) {
 797                 struct buffer_head *bh = list_entry(io_bufs.prev,
 798                                                     struct buffer_head,
 799                                                     b_assoc_buffers);
 800
 801                 wait_on_buffer(bh);
 802                 cond_resched();
 803
 804                 if (unlikely(!buffer_uptodate(bh)))
 805                         err = -EIO;
 806                 jbd2_unfile_log_bh(bh);
 807
 808                 /*
 809                  * The list contains temporary buffer heads created by
 810                  * jbd2_journal_write_metadata_buffer().
 811                  */
 812                 BUFFER_TRACE(bh, "dumping temporary bh");
 813                 __brelse(bh);
 814                 J_ASSERT_BH(bh, atomic_read(&bh->b_count) == 0);
 815                 free_buffer_head(bh);
 816
 817                 /* We also have to refile the corresponding shadowed buffer */
 818                 jh = commit_transaction->t_shadow_list->b_tprev;
 819                 bh = jh2bh(jh);
 820                 clear_buffer_jwrite(bh);
 821                 J_ASSERT_BH(bh, buffer_jbddirty(bh));
 822                 J_ASSERT_BH(bh, !buffer_shadow(bh));
 823
 824                 /* The metadata is now released for reuse, but we need
 825                    to remember it against this transaction so that when
 826                    we finally commit, we can do any checkpointing
 827                    required. */
 828                 JBUFFER_TRACE(jh, "file as BJ_Forget");
 829                 jbd2_journal_file_buffer(jh, commit_transaction, BJ_Forget);
 830                 JBUFFER_TRACE(jh, "brelse shadowed buffer");
 831                 __brelse(bh);
 832         }
 833
 834         J_ASSERT (commit_transaction->t_shadow_list == NULL);
 835
 836         jbd_debug(3, "JBD2: commit phase 4\n");
 837
 838         /* Here we wait for the revoke record and descriptor record buffers */
 839         while (!list_empty(&log_bufs)) {
 840                 struct buffer_head *bh;
 841
 842                 bh = list_entry(log_bufs.prev, struct buffer_head, b_assoc_buffers);
 843                 wait_on_buffer(bh);
 844                 cond_resched();
 845
 846                 if (unlikely(!buffer_uptodate(bh)))
 847                         err = -EIO;
 848
 849                 BUFFER_TRACE(bh, "ph5: control buffer writeout done: unfile");
 850                 clear_buffer_jwrite(bh);
 851                 jbd2_unfile_log_bh(bh);
 852                 __brelse(bh);           /* One for getblk */
 853                 /* AKPM: bforget here */
 854         }
 855
 856         if (err)
 857                 jbd2_journal_abort(journal, err);
 858
 859         jbd_debug(3, "JBD2: commit phase 5\n");
 860         write_lock(&journal->j_state_lock);
 861         J_ASSERT(commit_transaction->t_state == T_COMMIT_DFLUSH);
 862         commit_transaction->t_state = T_COMMIT_JFLUSH;
 863         write_unlock(&journal->j_state_lock);
 864
 865         if (!jbd2_has_feature_async_commit(journal)) {
 866                 err = journal_submit_commit_record(journal, commit_transaction,
 867                                                 &cbh, crc32_sum);
 868                 if (err)
 869                         __jbd2_journal_abort_hard(journal);
 870         }
 871         if (cbh)
 872                 err = journal_wait_on_commit_record(journal, cbh);
 873         if (jbd2_has_feature_async_commit(journal) &&
 874             journal->j_flags & JBD2_BARRIER) {
 875                 blkdev_issue_flush(journal->j_dev, GFP_NOFS, NULL);
 876         }
 877
 878         if (err)
 879                 jbd2_journal_abort(journal, err);
 880
 881         /*
 882          * Now disk caches for filesystem device are flushed so we are safe to
 883          * erase checkpointed transactions from the log by updating journal
 884          * superblock.
 885          */
 886         if (update_tail)
 887                 jbd2_update_log_tail(journal, first_tid, first_block);
 888
 889         /* End of a transaction!  Finally, we can do checkpoint
 890            processing: any buffers committed as a result of this
 891            transaction can be removed from any checkpoint list it was on
 892            before. */
 893
 894         jbd_debug(3, "JBD2: commit phase 6\n");
 895
 896         J_ASSERT(list_empty(&commit_transaction->t_inode_list));
 897         J_ASSERT(commit_transaction->t_buffers == NULL);
 898         J_ASSERT(commit_transaction->t_checkpoint_list == NULL);
 899         J_ASSERT(commit_transaction->t_shadow_list == NULL);
 900
 901 restart_loop:
 902         /*
 903          * As there are other places (journal_unmap_buffer()) adding buffers
 904          * to this list we have to be careful and hold the j_list_lock.
 905          */
 906         spin_lock(&journal->j_list_lock);
 907         while (commit_transaction->t_forget) {
 908                 transaction_t *cp_transaction;
 909                 struct buffer_head *bh;
 910                 int try_to_free = 0;
 911
 912                 jh = commit_transaction->t_forget;
 913                 spin_unlock(&journal->j_list_lock);
 914                 bh = jh2bh(jh);
 915                 /*
 916                  * Get a reference so that bh cannot be freed before we are
 917                  * done with it.
 918                  */
 919                 get_bh(bh);
 920                 jbd_lock_bh_state(bh);
 921                 J_ASSERT_JH(jh, jh->b_transaction == commit_transaction);
 922
 923                 /*
 924                  * If there is undo-protected committed data against
 925                  * this buffer, then we can remove it now.  If it is a
 926                  * buffer needing such protection, the old frozen_data
 927                  * field now points to a committed version of the
 928                  * buffer, so rotate that field to the new committed
 929                  * data.
 930                  *
 931                  * Otherwise, we can just throw away the frozen data now.
 932                  *
 933                  * We also know that the frozen data has already fired
 934                  * its triggers if they exist, so we can clear that too.
 935                  */
 936                 if (jh->b_committed_data) {
 937                         jbd2_free(jh->b_committed_data, bh->b_size);
 938                         jh->b_committed_data = NULL;
 939                         if (jh->b_frozen_data) {
 940                                 jh->b_committed_data = jh->b_frozen_data;
 941                                 jh->b_frozen_data = NULL;
 942                                 jh->b_frozen_triggers = NULL;
 943                         }
 944                 } else if (jh->b_frozen_data) {
 945                         jbd2_free(jh->b_frozen_data, bh->b_size);
 946                         jh->b_frozen_data = NULL;
 947                         jh->b_frozen_triggers = NULL;
 948                 }
 949
 950                 spin_lock(&journal->j_list_lock);
 951                 cp_transaction = jh->b_cp_transaction;
 952                 if (cp_transaction) {
 953                         JBUFFER_TRACE(jh, "remove from old cp transaction");
 954                         cp_transaction->t_chp_stats.cs_dropped++;
 955                         __jbd2_journal_remove_checkpoint(jh);
 956                 }
 957
 958                 /* Only re-checkpoint the buffer_head if it is marked
 959                  * dirty.  If the buffer was added to the BJ_Forget list
 960                  * by jbd2_journal_forget, it may no longer be dirty and
 961                  * there's no point in keeping a checkpoint record for
 962                  * it. */
 963
 964                 /*
 965                 * A buffer which has been freed while still being journaled by
 966                 * a previous transaction.
 967                 */
 968                 if (buffer_freed(bh)) {
 969                         /*
 970                          * If the running transaction is the one containing
 971                          * "add to orphan" operation (b_next_transaction !=
 972                          * NULL), we have to wait for that transaction to
 973                          * commit before we can really get rid of the buffer.
 974                          * So just clear b_modified to not confuse transaction
 975                          * credit accounting and refile the buffer to
 976                          * BJ_Forget of the running transaction. If the just
 977                          * committed transaction contains "add to orphan"
 978                          * operation, we can completely invalidate the buffer
 979                          * now. We are rather through in that since the
 980                          * buffer may be still accessible when blocksize <
 981                          * pagesize and it is attached to the last partial
 982                          * page.
 983                          */
 984                         jh->b_modified = 0;
 985                         if (!jh->b_next_transaction) {
 986                                 clear_buffer_freed(bh);
 987                                 clear_buffer_jbddirty(bh);
 988                                 clear_buffer_mapped(bh);
 989                                 clear_buffer_new(bh);
 990                                 clear_buffer_req(bh);
 991                                 bh->b_bdev = NULL;
 992                         }
 993                 }
 994
 995                 if (buffer_jbddirty(bh)) {
 996                         JBUFFER_TRACE(jh, "add to new checkpointing trans");
 997                         __jbd2_journal_insert_checkpoint(jh, commit_transaction);
 998                         if (is_journal_aborted(journal))
 999                                 clear_buffer_jbddirty(bh);
1000                 } else {
1001                         J_ASSERT_BH(bh, !buffer_dirty(bh));
1002                         /*
1003                          * The buffer on BJ_Forget list and not jbddirty means
1004                          * it has been freed by this transaction and hence it
1005                          * could not have been reallocated until this
1006                          * transaction has committed. *BUT* it could be
1007                          * reallocated once we have written all the data to
1008                          * disk and before we process the buffer on BJ_Forget
1009                          * list.
1010                          */
1011                         if (!jh->b_next_transaction)
1012                                 try_to_free = 1;
1013                 }
1014                 JBUFFER_TRACE(jh, "refile or unfile buffer");
1015                 __jbd2_journal_refile_buffer(jh);
1016                 jbd_unlock_bh_state(bh);
1017                 if (try_to_free)
1018                         release_buffer_page(bh);        /* Drops bh reference */
1019                 else
1020                         __brelse(bh);
1021                 cond_resched_lock(&journal->j_list_lock);
1022         }
1023         spin_unlock(&journal->j_list_lock);
1024         /*
1025          * This is a bit sleazy.  We use j_list_lock to protect transition
1026          * of a transaction into T_FINISHED state and calling
1027          * __jbd2_journal_drop_transaction(). Otherwise we could race with
1028          * other checkpointing code processing the transaction...
1029          */
1030         write_lock(&journal->j_state_lock);
1031         spin_lock(&journal->j_list_lock);
1032         /*
1033          * Now recheck if some buffers did not get attached to the transaction
1034          * while the lock was dropped...
1035          */
1036         if (commit_transaction->t_forget) {
1037                 spin_unlock(&journal->j_list_lock);
1038                 write_unlock(&journal->j_state_lock);
1039                 goto restart_loop;
1040         }
1041
1042         /* Add the transaction to the checkpoint list
1043          * __journal_remove_checkpoint() can not destroy transaction
1044          * under us because it is not marked as T_FINISHED yet */
1045         if (journal->j_checkpoint_transactions == NULL) {
1046                 journal->j_checkpoint_transactions = commit_transaction;
1047                 commit_transaction->t_cpnext = commit_transaction;
1048                 commit_transaction->t_cpprev = commit_transaction;
1049         } else {
1050                 commit_transaction->t_cpnext =
1051                         journal->j_checkpoint_transactions;
1052                 commit_transaction->t_cpprev =
1053                         commit_transaction->t_cpnext->t_cpprev;
1054                 commit_transaction->t_cpnext->t_cpprev =
1055                         commit_transaction;
1056                 commit_transaction->t_cpprev->t_cpnext =
1057                                 commit_transaction;
1058         }
1059         spin_unlock(&journal->j_list_lock);
1060
1061         /* Done with this transaction! */
1062
1063         jbd_debug(3, "JBD2: commit phase 7\n");
1064
1065         J_ASSERT(commit_transaction->t_state == T_COMMIT_JFLUSH);
1066
1067         commit_transaction->t_start = jiffies;
1068         stats.run.rs_logging = jbd2_time_diff(stats.run.rs_logging,
1069                                               commit_transaction->t_start);
1070
1071         /*
1072          * File the transaction statistics
1073          */
1074         stats.ts_tid = commit_transaction->t_tid;
1075         stats.run.rs_handle_count =
1076                 atomic_read(&commit_transaction->t_handle_count);
1077         trace_jbd2_run_stats(journal->j_fs_dev->bd_dev,
1078                              commit_transaction->t_tid, &stats.run);
1079         stats.ts_requested = (commit_transaction->t_requested) ? 1 : 0;
1080
1081         commit_transaction->t_state = T_COMMIT_CALLBACK;
1082         J_ASSERT(commit_transaction == journal->j_committing_transaction);
1083         journal->j_commit_sequence = commit_transaction->t_tid;
1084         journal->j_committing_transaction = NULL;
1085         commit_time = ktime_to_ns(ktime_sub(ktime_get(), start_time));
1086
1087         /*
1088          * weight the commit time higher than the average time so we don't
1089          * react too strongly to vast changes in the commit time
1090          */
1091         if (likely(journal->j_average_commit_time))
1092                 journal->j_average_commit_time = (commit_time +
1093                                 journal->j_average_commit_time*3) / 4;
1094         else
1095                 journal->j_average_commit_time = commit_time;
1096
1097         write_unlock(&journal->j_state_lock);
1098
1099         if (journal->j_commit_callback)
1100                 journal->j_commit_callback(journal, commit_transaction);
1101
1102         trace_jbd2_end_commit(journal, commit_transaction);
1103         jbd_debug(1, "JBD2: commit %d complete, head %d\n",
1104                   journal->j_commit_sequence, journal->j_tail_sequence);
1105
1106         write_lock(&journal->j_state_lock);
1107         spin_lock(&journal->j_list_lock);
1108         commit_transaction->t_state = T_FINISHED;
1109         /* Check if the transaction can be dropped now that we are finished */
1110         if (commit_transaction->t_checkpoint_list == NULL &&
1111             commit_transaction->t_checkpoint_io_list == NULL) {
1112                 __jbd2_journal_drop_transaction(journal, commit_transaction);
1113                 jbd2_journal_free_transaction(commit_transaction);
1114         }
1115         spin_unlock(&journal->j_list_lock);
1116         write_unlock(&journal->j_state_lock);
1117         wake_up(&journal->j_wait_done_commit);
1118
1119         /*
1120          * Calculate overall stats
1121          */
1122         spin_lock(&journal->j_history_lock);
1123         journal->j_stats.ts_tid++;
1124         journal->j_stats.ts_requested += stats.ts_requested;
1125         journal->j_stats.run.rs_wait += stats.run.rs_wait;
1126         journal->j_stats.run.rs_request_delay += stats.run.rs_request_delay;
1127         journal->j_stats.run.rs_running += stats.run.rs_running;
1128         journal->j_stats.run.rs_locked += stats.run.rs_locked;
1129         journal->j_stats.run.rs_flushing += stats.run.rs_flushing;
1130         journal->j_stats.run.rs_logging += stats.run.rs_logging;
1131         journal->j_stats.run.rs_handle_count += stats.run.rs_handle_count;
1132         journal->j_stats.run.rs_blocks += stats.run.rs_blocks;
1133         journal->j_stats.run.rs_blocks_logged += stats.run.rs_blocks_logged;
1134         spin_unlock(&journal->j_history_lock);
1135 }