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_clear_bit();
  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_HAS_INCOMPAT_FEATURE(j, JBD2_FEATURE_INCOMPAT_CSUM_V2))
 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(journal);
 135         if (!bh)
 136                 return 1;
 137
 138         tmp = (struct commit_header *)bh->b_data;
 139         tmp->h_magic = cpu_to_be32(JBD2_MAGIC_NUMBER);
 140         tmp->h_blocktype = cpu_to_be32(JBD2_COMMIT_BLOCK);
 141         tmp->h_sequence = cpu_to_be32(commit_transaction->t_tid);
 142         tmp->h_commit_sec = cpu_to_be64(now.tv_sec);
 143         tmp->h_commit_nsec = cpu_to_be32(now.tv_nsec);
 144
 145         if (JBD2_HAS_COMPAT_FEATURE(journal,
 146                                     JBD2_FEATURE_COMPAT_CHECKSUM)) {
 147                 tmp->h_chksum_type      = JBD2_CRC32_CHKSUM;
 148                 tmp->h_chksum_size      = JBD2_CRC32_CHKSUM_SIZE;
 149                 tmp->h_chksum[0]        = cpu_to_be32(crc32_sum);
 150         }
 151         jbd2_commit_block_csum_set(journal, bh);
 152
 153         BUFFER_TRACE(bh, "submit commit block");
 154         lock_buffer(bh);
 155         clear_buffer_dirty(bh);
 156         set_buffer_uptodate(bh);
 157         bh->b_end_io = journal_end_buffer_io_sync;
 158
 159         if (journal->j_flags & JBD2_BARRIER &&
 160             !JBD2_HAS_INCOMPAT_FEATURE(journal,
 161                                        JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT))
 162                 ret = submit_bh(WRITE_SYNC | WRITE_FLUSH_FUA, bh);
 163         else
 164                 ret = submit_bh(WRITE_SYNC, bh);
 165
 166         *cbh = bh;
 167         return ret;
 168 }
 169
 170 /*
 171  * This function along with journal_submit_commit_record
 172  * allows to write the commit record asynchronously.
 173  */
 174 static int journal_wait_on_commit_record(journal_t *journal,
 175                                          struct buffer_head *bh)
 176 {
 177         int ret = 0;
 178
 179         clear_buffer_dirty(bh);
 180         wait_on_buffer(bh);
 181
 182         if (unlikely(!buffer_uptodate(bh)))
 183                 ret = -EIO;
 184         put_bh(bh);            /* One for getblk() */
 185
 186         return ret;
 187 }
 188
 189 /*
 190  * write the filemap data using writepage() address_space_operations.
 191  * We don't do block allocation here even for delalloc. We don't
 192  * use writepages() because with dealyed allocation we may be doing
 193  * block allocation in writepages().
 194  */
 195 static int journal_submit_inode_data_buffers(struct address_space *mapping)
 196 {
 197         int ret;
 198         struct writeback_control wbc = {
 199                 .sync_mode =  WB_SYNC_ALL,
 200                 .nr_to_write = mapping->nrpages * 2,
 201                 .range_start = 0,
 202                 .range_end = i_size_read(mapping->host),
 203         };
 204
 205         ret = generic_writepages(mapping, &wbc);
 206         return ret;
 207 }
 208
 209 /*
 210  * Submit all the data buffers of inode associated with the transaction to
 211  * disk.
 212  *
 213  * We are in a committing transaction. Therefore no new inode can be added to
 214  * our inode list. We use JI_COMMIT_RUNNING flag to protect inode we currently
 215  * operate on from being released while we write out pages.
 216  */
 217 static int journal_submit_data_buffers(journal_t *journal,
 218                 transaction_t *commit_transaction)
 219 {
 220         struct jbd2_inode *jinode;
 221         int err, ret = 0;
 222         struct address_space *mapping;
 223
 224         spin_lock(&journal->j_list_lock);
 225         list_for_each_entry(jinode, &commit_transaction->t_inode_list, i_list) {
 226                 mapping = jinode->i_vfs_inode->i_mapping;
 227                 set_bit(__JI_COMMIT_RUNNING, &jinode->i_flags);
 228                 spin_unlock(&journal->j_list_lock);
 229                 /*
 230                  * submit the inode data buffers. We use writepage
 231                  * instead of writepages. Because writepages can do
 232                  * block allocation  with delalloc. We need to write
 233                  * only allocated blocks here.
 234                  */
 235                 trace_jbd2_submit_inode_data(jinode->i_vfs_inode);
 236                 err = journal_submit_inode_data_buffers(mapping);
 237                 if (!ret)
 238                         ret = err;
 239                 spin_lock(&journal->j_list_lock);
 240                 J_ASSERT(jinode->i_transaction == commit_transaction);
 241                 clear_bit(__JI_COMMIT_RUNNING, &jinode->i_flags);
 242                 smp_mb__after_clear_bit();
 243                 wake_up_bit(&jinode->i_flags, __JI_COMMIT_RUNNING);
 244         }
 245         spin_unlock(&journal->j_list_lock);
 246         return ret;
 247 }
 248
 249 /*
 250  * Wait for data submitted for writeout, refile inodes to proper
 251  * transaction if needed.
 252  *
 253  */
 254 static int journal_finish_inode_data_buffers(journal_t *journal,
 255                 transaction_t *commit_transaction)
 256 {
 257         struct jbd2_inode *jinode, *next_i;
 258         int err, ret = 0;
 259
 260         /* For locking, see the comment in journal_submit_data_buffers() */
 261         spin_lock(&journal->j_list_lock);
 262         list_for_each_entry(jinode, &commit_transaction->t_inode_list, i_list) {
 263                 set_bit(__JI_COMMIT_RUNNING, &jinode->i_flags);
 264                 spin_unlock(&journal->j_list_lock);
 265                 err = filemap_fdatawait(jinode->i_vfs_inode->i_mapping);
 266                 if (err) {
 267                         /*
 268                          * Because AS_EIO is cleared by
 269                          * filemap_fdatawait_range(), set it again so
 270                          * that user process can get -EIO from fsync().
 271                          */
 272                         set_bit(AS_EIO,
 273                                 &jinode->i_vfs_inode->i_mapping->flags);
 274
 275                         if (!ret)
 276                                 ret = err;
 277                 }
 278                 spin_lock(&journal->j_list_lock);
 279                 clear_bit(__JI_COMMIT_RUNNING, &jinode->i_flags);
 280                 smp_mb__after_clear_bit();
 281                 wake_up_bit(&jinode->i_flags, __JI_COMMIT_RUNNING);
 282         }
 283
 284         /* Now refile inode to proper lists */
 285         list_for_each_entry_safe(jinode, next_i,
 286                                  &commit_transaction->t_inode_list, i_list) {
 287                 list_del(&jinode->i_list);
 288                 if (jinode->i_next_transaction) {
 289                         jinode->i_transaction = jinode->i_next_transaction;
 290                         jinode->i_next_transaction = NULL;
 291                         list_add(&jinode->i_list,
 292                                 &jinode->i_transaction->t_inode_list);
 293                 } else {
 294                         jinode->i_transaction = NULL;
 295                 }
 296         }
 297         spin_unlock(&journal->j_list_lock);
 298
 299         return ret;
 300 }
 301
 302 static __u32 jbd2_checksum_data(__u32 crc32_sum, struct buffer_head *bh)
 303 {
 304         struct page *page = bh->b_page;
 305         char *addr;
 306         __u32 checksum;
 307
 308         addr = kmap_atomic(page);
 309         checksum = crc32_be(crc32_sum,
 310                 (void *)(addr + offset_in_page(bh->b_data)), bh->b_size);
 311         kunmap_atomic(addr);
 312
 313         return checksum;
 314 }
 315
 316 static void write_tag_block(int tag_bytes, journal_block_tag_t *tag,
 317                                    unsigned long long block)
 318 {
 319         tag->t_blocknr = cpu_to_be32(block & (u32)~0);
 320         if (tag_bytes > JBD2_TAG_SIZE32)
 321                 tag->t_blocknr_high = cpu_to_be32((block >> 31) >> 1);
 322 }
 323
 324 static void jbd2_descr_block_csum_set(journal_t *j,
 325                                       struct buffer_head *bh)
 326 {
 327         struct jbd2_journal_block_tail *tail;
 328         __u32 csum;
 329
 330         if (!JBD2_HAS_INCOMPAT_FEATURE(j, JBD2_FEATURE_INCOMPAT_CSUM_V2))
 331                 return;
 332
 333         tail = (struct jbd2_journal_block_tail *)(bh->b_data + j->j_blocksize -
 334                         sizeof(struct jbd2_journal_block_tail));
 335         tail->t_checksum = 0;
 336         csum = jbd2_chksum(j, j->j_csum_seed, bh->b_data, j->j_blocksize);
 337         tail->t_checksum = cpu_to_be32(csum);
 338 }
 339
 340 static void jbd2_block_tag_csum_set(journal_t *j, journal_block_tag_t *tag,
 341                                     struct buffer_head *bh, __u32 sequence)
 342 {
 343         struct page *page = bh->b_page;
 344         __u8 *addr;
 345         __u32 csum32;
 346         __be32 seq;
 347
 348         if (!JBD2_HAS_INCOMPAT_FEATURE(j, JBD2_FEATURE_INCOMPAT_CSUM_V2))
 349                 return;
 350
 351         seq = cpu_to_be32(sequence);
 352         addr = kmap_atomic(page);
 353         csum32 = jbd2_chksum(j, j->j_csum_seed, (__u8 *)&seq, sizeof(seq));
 354         csum32 = jbd2_chksum(j, csum32, addr + offset_in_page(bh->b_data),
 355                              bh->b_size);
 356         kunmap_atomic(addr);
 357
 358         /* We only have space to store the lower 16 bits of the crc32c. */
 359         tag->t_checksum = cpu_to_be16(csum32);
 360 }
 361 /*
 362  * jbd2_journal_commit_transaction
 363  *
 364  * The primary function for committing a transaction to the log.  This
 365  * function is called by the journal thread to begin a complete commit.
 366  */
 367 void jbd2_journal_commit_transaction(journal_t *journal)
 368 {
 369         struct transaction_stats_s stats;
 370         transaction_t *commit_transaction;
 371         struct journal_head *jh;
 372         struct buffer_head *descriptor;
 373         struct buffer_head **wbuf = journal->j_wbuf;
 374         int bufs;
 375         int flags;
 376         int err;
 377         unsigned long long blocknr;
 378         ktime_t start_time;
 379         u64 commit_time;
 380         char *tagp = NULL;
 381         journal_header_t *header;
 382         journal_block_tag_t *tag = NULL;
 383         int space_left = 0;
 384         int first_tag = 0;
 385         int tag_flag;
 386         int i;
 387         int tag_bytes = journal_tag_bytes(journal);
 388         struct buffer_head *cbh = NULL; /* For transactional checksums */
 389         __u32 crc32_sum = ~0;
 390         struct blk_plug plug;
 391         /* Tail of the journal */
 392         unsigned long first_block;
 393         tid_t first_tid;
 394         int update_tail;
 395         int csum_size = 0;
 396         LIST_HEAD(io_bufs);
 397         LIST_HEAD(log_bufs);
 398
 399         if (JBD2_HAS_INCOMPAT_FEATURE(journal, JBD2_FEATURE_INCOMPAT_CSUM_V2))
 400                 csum_size = sizeof(struct jbd2_journal_block_tail);
 401
 402         /*
 403          * First job: lock down the current transaction and wait for
 404          * all outstanding updates to complete.
 405          */
 406
 407         /* Do we need to erase the effects of a prior jbd2_journal_flush? */
 408         if (journal->j_flags & JBD2_FLUSHED) {
 409                 jbd_debug(3, "super block updated\n");
 410                 mutex_lock(&journal->j_checkpoint_mutex);
 411                 /*
 412                  * We hold j_checkpoint_mutex so tail cannot change under us.
 413                  * We don't need any special data guarantees for writing sb
 414                  * since journal is empty and it is ok for write to be
 415                  * flushed only with transaction commit.
 416                  */
 417                 jbd2_journal_update_sb_log_tail(journal,
 418                                                 journal->j_tail_sequence,
 419                                                 journal->j_tail,
 420                                                 WRITE_SYNC);
 421                 mutex_unlock(&journal->j_checkpoint_mutex);
 422         } else {
 423                 jbd_debug(3, "superblock not updated\n");
 424         }
 425
 426         J_ASSERT(journal->j_running_transaction != NULL);
 427         J_ASSERT(journal->j_committing_transaction == NULL);
 428
 429         commit_transaction = journal->j_running_transaction;
 430
 431         trace_jbd2_start_commit(journal, commit_transaction);
 432         jbd_debug(1, "JBD2: starting commit of transaction %d\n",
 433                         commit_transaction->t_tid);
 434
 435         write_lock(&journal->j_state_lock);
 436         J_ASSERT(commit_transaction->t_state == T_RUNNING);
 437         commit_transaction->t_state = T_LOCKED;
 438
 439         trace_jbd2_commit_locking(journal, commit_transaction);
 440         stats.run.rs_wait = commit_transaction->t_max_wait;
 441         stats.run.rs_request_delay = 0;
 442         stats.run.rs_locked = jiffies;
 443         if (commit_transaction->t_requested)
 444                 stats.run.rs_request_delay =
 445                         jbd2_time_diff(commit_transaction->t_requested,
 446                                        stats.run.rs_locked);
 447         stats.run.rs_running = jbd2_time_diff(commit_transaction->t_start,
 448                                               stats.run.rs_locked);
 449
 450         spin_lock(&commit_transaction->t_handle_lock);
 451         while (atomic_read(&commit_transaction->t_updates)) {
 452                 DEFINE_WAIT(wait);
 453
 454                 prepare_to_wait(&journal->j_wait_updates, &wait,
 455                                         TASK_UNINTERRUPTIBLE);
 456                 if (atomic_read(&commit_transaction->t_updates)) {
 457                         spin_unlock(&commit_transaction->t_handle_lock);
 458                         write_unlock(&journal->j_state_lock);
 459                         schedule();
 460                         write_lock(&journal->j_state_lock);
 461                         spin_lock(&commit_transaction->t_handle_lock);
 462                 }
 463                 finish_wait(&journal->j_wait_updates, &wait);
 464         }
 465         spin_unlock(&commit_transaction->t_handle_lock);
 466
 467         J_ASSERT (atomic_read(&commit_transaction->t_outstanding_credits) <=
 468                         journal->j_max_transaction_buffers);
 469
 470         /*
 471          * First thing we are allowed to do is to discard any remaining
 472          * BJ_Reserved buffers.  Note, it is _not_ permissible to assume
 473          * that there are no such buffers: if a large filesystem
 474          * operation like a truncate needs to split itself over multiple
 475          * transactions, then it may try to do a jbd2_journal_restart() while
 476          * there are still BJ_Reserved buffers outstanding.  These must
 477          * be released cleanly from the current transaction.
 478          *
 479          * In this case, the filesystem must still reserve write access
 480          * again before modifying the buffer in the new transaction, but
 481          * we do not require it to remember exactly which old buffers it
 482          * has reserved.  This is consistent with the existing behaviour
 483          * that multiple jbd2_journal_get_write_access() calls to the same
 484          * buffer are perfectly permissible.
 485          */
 486         while (commit_transaction->t_reserved_list) {
 487                 jh = commit_transaction->t_reserved_list;
 488                 JBUFFER_TRACE(jh, "reserved, unused: refile");
 489                 /*
 490                  * A jbd2_journal_get_undo_access()+jbd2_journal_release_buffer() may
 491                  * leave undo-committed data.
 492                  */
 493                 if (jh->b_committed_data) {
 494                         struct buffer_head *bh = jh2bh(jh);
 495
 496                         jbd_lock_bh_state(bh);
 497                         jbd2_free(jh->b_committed_data, bh->b_size);
 498                         jh->b_committed_data = NULL;
 499                         jbd_unlock_bh_state(bh);
 500                 }
 501                 jbd2_journal_refile_buffer(journal, jh);
 502         }
 503
 504         /*
 505          * Now try to drop any written-back buffers from the journal's
 506          * checkpoint lists.  We do this *before* commit because it potentially
 507          * frees some memory
 508          */
 509         spin_lock(&journal->j_list_lock);
 510         __jbd2_journal_clean_checkpoint_list(journal);
 511         spin_unlock(&journal->j_list_lock);
 512
 513         jbd_debug(3, "JBD2: commit phase 1\n");
 514
 515         /*
 516          * Clear revoked flag to reflect there is no revoked buffers
 517          * in the next transaction which is going to be started.
 518          */
 519         jbd2_clear_buffer_revoked_flags(journal);
 520
 521         /*
 522          * Switch to a new revoke table.
 523          */
 524         jbd2_journal_switch_revoke_table(journal);
 525
 526         /*
 527          * Reserved credits cannot be claimed anymore, free them
 528          */
 529         atomic_sub(atomic_read(&journal->j_reserved_credits),
 530                    &commit_transaction->t_outstanding_credits);
 531
 532         trace_jbd2_commit_flushing(journal, commit_transaction);
 533         stats.run.rs_flushing = jiffies;
 534         stats.run.rs_locked = jbd2_time_diff(stats.run.rs_locked,
 535                                              stats.run.rs_flushing);
 536
 537         commit_transaction->t_state = T_FLUSH;
 538         journal->j_committing_transaction = commit_transaction;
 539         journal->j_running_transaction = NULL;
 540         start_time = ktime_get();
 541         commit_transaction->t_log_start = journal->j_head;
 542         wake_up(&journal->j_wait_transaction_locked);
 543         write_unlock(&journal->j_state_lock);
 544
 545         jbd_debug(3, "JBD2: commit phase 2a\n");
 546
 547         /*
 548          * Now start flushing things to disk, in the order they appear
 549          * on the transaction lists.  Data blocks go first.
 550          */
 551         err = journal_submit_data_buffers(journal, commit_transaction);
 552         if (err)
 553                 jbd2_journal_abort(journal, err);
 554
 555         blk_start_plug(&plug);
 556         jbd2_journal_write_revoke_records(journal, commit_transaction,
 557                                           &log_bufs, WRITE_SYNC);
 558
 559         jbd_debug(3, "JBD2: commit phase 2b\n");
 560
 561         /*
 562          * Way to go: we have now written out all of the data for a
 563          * transaction!  Now comes the tricky part: we need to write out
 564          * metadata.  Loop over the transaction's entire buffer list:
 565          */
 566         write_lock(&journal->j_state_lock);
 567         commit_transaction->t_state = T_COMMIT;
 568         write_unlock(&journal->j_state_lock);
 569
 570         trace_jbd2_commit_logging(journal, commit_transaction);
 571         stats.run.rs_logging = jiffies;
 572         stats.run.rs_flushing = jbd2_time_diff(stats.run.rs_flushing,
 573                                                stats.run.rs_logging);
 574         stats.run.rs_blocks =
 575                 atomic_read(&commit_transaction->t_outstanding_credits);
 576         stats.run.rs_blocks_logged = 0;
 577
 578         J_ASSERT(commit_transaction->t_nr_buffers <=
 579                  atomic_read(&commit_transaction->t_outstanding_credits));
 580
 581         err = 0;
 582         bufs = 0;
 583         descriptor = NULL;
 584         while (commit_transaction->t_buffers) {
 585
 586                 /* Find the next buffer to be journaled... */
 587
 588                 jh = commit_transaction->t_buffers;
 589
 590                 /* If we're in abort mode, we just un-journal the buffer and
 591                    release it. */
 592
 593                 if (is_journal_aborted(journal)) {
 594                         clear_buffer_jbddirty(jh2bh(jh));
 595                         JBUFFER_TRACE(jh, "journal is aborting: refile");
 596                         jbd2_buffer_abort_trigger(jh,
 597                                                   jh->b_frozen_data ?
 598                                                   jh->b_frozen_triggers :
 599                                                   jh->b_triggers);
 600                         jbd2_journal_refile_buffer(journal, jh);
 601                         /* If that was the last one, we need to clean up
 602                          * any descriptor buffers which may have been
 603                          * already allocated, even if we are now
 604                          * aborting. */
 605                         if (!commit_transaction->t_buffers)
 606                                 goto start_journal_io;
 607                         continue;
 608                 }
 609
 610                 /* Make sure we have a descriptor block in which to
 611                    record the metadata buffer. */
 612
 613                 if (!descriptor) {
 614                         J_ASSERT (bufs == 0);
 615
 616                         jbd_debug(4, "JBD2: get descriptor\n");
 617
 618                         descriptor = jbd2_journal_get_descriptor_buffer(journal);
 619                         if (!descriptor) {
 620                                 jbd2_journal_abort(journal, -EIO);
 621                                 continue;
 622                         }
 623
 624                         jbd_debug(4, "JBD2: got buffer %llu (%p)\n",
 625                                 (unsigned long long)descriptor->b_blocknr,
 626                                 descriptor->b_data);
 627                         header = (journal_header_t *)descriptor->b_data;
 628                         header->h_magic     = cpu_to_be32(JBD2_MAGIC_NUMBER);
 629                         header->h_blocktype = cpu_to_be32(JBD2_DESCRIPTOR_BLOCK);
 630                         header->h_sequence  = cpu_to_be32(commit_transaction->t_tid);
 631
 632                         tagp = &descriptor->b_data[sizeof(journal_header_t)];
 633                         space_left = descriptor->b_size -
 634                                                 sizeof(journal_header_t);
 635                         first_tag = 1;
 636                         set_buffer_jwrite(descriptor);
 637                         set_buffer_dirty(descriptor);
 638                         wbuf[bufs++] = descriptor;
 639
 640                         /* Record it so that we can wait for IO
 641                            completion later */
 642                         BUFFER_TRACE(descriptor, "ph3: file as descriptor");
 643                         jbd2_file_log_bh(&log_bufs, descriptor);
 644                 }
 645
 646                 /* Where is the buffer to be written? */
 647
 648                 err = jbd2_journal_next_log_block(journal, &blocknr);
 649                 /* If the block mapping failed, just abandon the buffer
 650                    and repeat this loop: we'll fall into the
 651                    refile-on-abort condition above. */
 652                 if (err) {
 653                         jbd2_journal_abort(journal, err);
 654                         continue;
 655                 }
 656
 657                 /*
 658                  * start_this_handle() uses t_outstanding_credits to determine
 659                  * the free space in the log, but this counter is changed
 660                  * by jbd2_journal_next_log_block() also.
 661                  */
 662                 atomic_dec(&commit_transaction->t_outstanding_credits);
 663
 664                 /* Bump b_count to prevent truncate from stumbling over
 665                    the shadowed buffer!  @@@ This can go if we ever get
 666                    rid of the shadow pairing of buffers. */
 667                 atomic_inc(&jh2bh(jh)->b_count);
 668
 669                 /*
 670                  * Make a temporary IO buffer with which to write it out
 671                  * (this will requeue the metadata buffer to BJ_Shadow).
 672                  */
 673                 set_bit(BH_JWrite, &jh2bh(jh)->b_state);
 674                 JBUFFER_TRACE(jh, "ph3: write metadata");
 675                 flags = jbd2_journal_write_metadata_buffer(commit_transaction,
 676                                                 jh, &wbuf[bufs], blocknr);
 677                 if (flags < 0) {
 678                         jbd2_journal_abort(journal, flags);
 679                         continue;
 680                 }
 681                 jbd2_file_log_bh(&io_bufs, wbuf[bufs]);
 682
 683                 /* Record the new block's tag in the current descriptor
 684                    buffer */
 685
 686                 tag_flag = 0;
 687                 if (flags & 1)
 688                         tag_flag |= JBD2_FLAG_ESCAPE;
 689                 if (!first_tag)
 690                         tag_flag |= JBD2_FLAG_SAME_UUID;
 691
 692                 tag = (journal_block_tag_t *) tagp;
 693                 write_tag_block(tag_bytes, tag, jh2bh(jh)->b_blocknr);
 694                 tag->t_flags = cpu_to_be16(tag_flag);
 695                 jbd2_block_tag_csum_set(journal, tag, wbuf[bufs],
 696                                         commit_transaction->t_tid);
 697                 tagp += tag_bytes;
 698                 space_left -= tag_bytes;
 699                 bufs++;
 700
 701                 if (first_tag) {
 702                         memcpy (tagp, journal->j_uuid, 16);
 703                         tagp += 16;
 704                         space_left -= 16;
 705                         first_tag = 0;
 706                 }
 707
 708                 /* If there's no more to do, or if the descriptor is full,
 709                    let the IO rip! */
 710
 711                 if (bufs == journal->j_wbufsize ||
 712                     commit_transaction->t_buffers == NULL ||
 713                     space_left < tag_bytes + 16 + csum_size) {
 714
 715                         jbd_debug(4, "JBD2: Submit %d IOs\n", bufs);
 716
 717                         /* Write an end-of-descriptor marker before
 718                            submitting the IOs.  "tag" still points to
 719                            the last tag we set up. */
 720
 721                         tag->t_flags |= cpu_to_be16(JBD2_FLAG_LAST_TAG);
 722
 723                         jbd2_descr_block_csum_set(journal, descriptor);
 724 start_journal_io:
 725                         for (i = 0; i < bufs; i++) {
 726                                 struct buffer_head *bh = wbuf[i];
 727                                 /*
 728                                  * Compute checksum.
 729                                  */
 730                                 if (JBD2_HAS_COMPAT_FEATURE(journal,
 731                                         JBD2_FEATURE_COMPAT_CHECKSUM)) {
 732                                         crc32_sum =
 733                                             jbd2_checksum_data(crc32_sum, bh);
 734                                 }
 735
 736                                 lock_buffer(bh);
 737                                 clear_buffer_dirty(bh);
 738                                 set_buffer_uptodate(bh);
 739                                 bh->b_end_io = journal_end_buffer_io_sync;
 740                                 submit_bh(WRITE_SYNC, bh);
 741                         }
 742                         cond_resched();
 743                         stats.run.rs_blocks_logged += bufs;
 744
 745                         /* Force a new descriptor to be generated next
 746                            time round the loop. */
 747                         descriptor = NULL;
 748                         bufs = 0;
 749                 }
 750         }
 751
 752         err = journal_finish_inode_data_buffers(journal, commit_transaction);
 753         if (err) {
 754                 printk(KERN_WARNING
 755                         "JBD2: Detected IO errors while flushing file data "
 756                        "on %s\n", journal->j_devname);
 757                 if (journal->j_flags & JBD2_ABORT_ON_SYNCDATA_ERR)
 758                         jbd2_journal_abort(journal, err);
 759                 err = 0;
 760         }
 761
 762         /*
 763          * Get current oldest transaction in the log before we issue flush
 764          * to the filesystem device. After the flush we can be sure that
 765          * blocks of all older transactions are checkpointed to persistent
 766          * storage and we will be safe to update journal start in the
 767          * superblock with the numbers we get here.
 768          */
 769         update_tail =
 770                 jbd2_journal_get_log_tail(journal, &first_tid, &first_block);
 771
 772         write_lock(&journal->j_state_lock);
 773         if (update_tail) {
 774                 long freed = first_block - journal->j_tail;
 775
 776                 if (first_block < journal->j_tail)
 777                         freed += journal->j_last - journal->j_first;
 778                 /* Update tail only if we free significant amount of space */
 779                 if (freed < journal->j_maxlen / 4)
 780                         update_tail = 0;
 781         }
 782         J_ASSERT(commit_transaction->t_state == T_COMMIT);
 783         commit_transaction->t_state = T_COMMIT_DFLUSH;
 784         write_unlock(&journal->j_state_lock);
 785
 786         /*
 787          * If the journal is not located on the file system device,
 788          * then we must flush the file system device before we issue
 789          * the commit record
 790          */
 791         if (commit_transaction->t_need_data_flush &&
 792             (journal->j_fs_dev != journal->j_dev) &&
 793             (journal->j_flags & JBD2_BARRIER))
 794                 blkdev_issue_flush(journal->j_fs_dev, GFP_NOFS, NULL);
 795
 796         /* Done it all: now write the commit record asynchronously. */
 797         if (JBD2_HAS_INCOMPAT_FEATURE(journal,
 798                                       JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
 799                 err = journal_submit_commit_record(journal, commit_transaction,
 800                                                  &cbh, crc32_sum);
 801                 if (err)
 802                         __jbd2_journal_abort_hard(journal);
 803         }
 804
 805         blk_finish_plug(&plug);
 806
 807         /* Lo and behold: we have just managed to send a transaction to
 808            the log.  Before we can commit it, wait for the IO so far to
 809            complete.  Control buffers being written are on the
 810            transaction's t_log_list queue, and metadata buffers are on
 811            the io_bufs list.
 812
 813            Wait for the buffers in reverse order.  That way we are
 814            less likely to be woken up until all IOs have completed, and
 815            so we incur less scheduling load.
 816         */
 817
 818         jbd_debug(3, "JBD2: commit phase 3\n");
 819
 820         while (!list_empty(&io_bufs)) {
 821                 struct buffer_head *bh = list_entry(io_bufs.prev,
 822                                                     struct buffer_head,
 823                                                     b_assoc_buffers);
 824
 825                 wait_on_buffer(bh);
 826                 cond_resched();
 827
 828                 if (unlikely(!buffer_uptodate(bh)))
 829                         err = -EIO;
 830                 jbd2_unfile_log_bh(bh);
 831
 832                 /*
 833                  * The list contains temporary buffer heads created by
 834                  * jbd2_journal_write_metadata_buffer().
 835                  */
 836                 BUFFER_TRACE(bh, "dumping temporary bh");
 837                 __brelse(bh);
 838                 J_ASSERT_BH(bh, atomic_read(&bh->b_count) == 0);
 839                 free_buffer_head(bh);
 840
 841                 /* We also have to refile the corresponding shadowed buffer */
 842                 jh = commit_transaction->t_shadow_list->b_tprev;
 843                 bh = jh2bh(jh);
 844                 clear_buffer_jwrite(bh);
 845                 J_ASSERT_BH(bh, buffer_jbddirty(bh));
 846                 J_ASSERT_BH(bh, !buffer_shadow(bh));
 847
 848                 /* The metadata is now released for reuse, but we need
 849                    to remember it against this transaction so that when
 850                    we finally commit, we can do any checkpointing
 851                    required. */
 852                 JBUFFER_TRACE(jh, "file as BJ_Forget");
 853                 jbd2_journal_file_buffer(jh, commit_transaction, BJ_Forget);
 854                 JBUFFER_TRACE(jh, "brelse shadowed buffer");
 855                 __brelse(bh);
 856         }
 857
 858         J_ASSERT (commit_transaction->t_shadow_list == NULL);
 859
 860         jbd_debug(3, "JBD2: commit phase 4\n");
 861
 862         /* Here we wait for the revoke record and descriptor record buffers */
 863         while (!list_empty(&log_bufs)) {
 864                 struct buffer_head *bh;
 865
 866                 bh = list_entry(log_bufs.prev, struct buffer_head, b_assoc_buffers);
 867                 wait_on_buffer(bh);
 868                 cond_resched();
 869
 870                 if (unlikely(!buffer_uptodate(bh)))
 871                         err = -EIO;
 872
 873                 BUFFER_TRACE(bh, "ph5: control buffer writeout done: unfile");
 874                 clear_buffer_jwrite(bh);
 875                 jbd2_unfile_log_bh(bh);
 876                 __brelse(bh);           /* One for getblk */
 877                 /* AKPM: bforget here */
 878         }
 879
 880         if (err)
 881                 jbd2_journal_abort(journal, err);
 882
 883         jbd_debug(3, "JBD2: commit phase 5\n");
 884         write_lock(&journal->j_state_lock);
 885         J_ASSERT(commit_transaction->t_state == T_COMMIT_DFLUSH);
 886         commit_transaction->t_state = T_COMMIT_JFLUSH;
 887         write_unlock(&journal->j_state_lock);
 888
 889         if (!JBD2_HAS_INCOMPAT_FEATURE(journal,
 890                                        JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
 891                 err = journal_submit_commit_record(journal, commit_transaction,
 892                                                 &cbh, crc32_sum);
 893                 if (err)
 894                         __jbd2_journal_abort_hard(journal);
 895         }
 896         if (cbh)
 897                 err = journal_wait_on_commit_record(journal, cbh);
 898         if (JBD2_HAS_INCOMPAT_FEATURE(journal,
 899                                       JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT) &&
 900             journal->j_flags & JBD2_BARRIER) {
 901                 blkdev_issue_flush(journal->j_dev, GFP_NOFS, NULL);
 902         }
 903
 904         if (err)
 905                 jbd2_journal_abort(journal, err);
 906
 907         /*
 908          * Now disk caches for filesystem device are flushed so we are safe to
 909          * erase checkpointed transactions from the log by updating journal
 910          * superblock.
 911          */
 912         if (update_tail)
 913                 jbd2_update_log_tail(journal, first_tid, first_block);
 914
 915         /* End of a transaction!  Finally, we can do checkpoint
 916            processing: any buffers committed as a result of this
 917            transaction can be removed from any checkpoint list it was on
 918            before. */
 919
 920         jbd_debug(3, "JBD2: commit phase 6\n");
 921
 922         J_ASSERT(list_empty(&commit_transaction->t_inode_list));
 923         J_ASSERT(commit_transaction->t_buffers == NULL);
 924         J_ASSERT(commit_transaction->t_checkpoint_list == NULL);
 925         J_ASSERT(commit_transaction->t_shadow_list == NULL);
 926
 927 restart_loop:
 928         /*
 929          * As there are other places (journal_unmap_buffer()) adding buffers
 930          * to this list we have to be careful and hold the j_list_lock.
 931          */
 932         spin_lock(&journal->j_list_lock);
 933         while (commit_transaction->t_forget) {
 934                 transaction_t *cp_transaction;
 935                 struct buffer_head *bh;
 936                 int try_to_free = 0;
 937
 938                 jh = commit_transaction->t_forget;
 939                 spin_unlock(&journal->j_list_lock);
 940                 bh = jh2bh(jh);
 941                 /*
 942                  * Get a reference so that bh cannot be freed before we are
 943                  * done with it.
 944                  */
 945                 get_bh(bh);
 946                 jbd_lock_bh_state(bh);
 947                 J_ASSERT_JH(jh, jh->b_transaction == commit_transaction);
 948
 949                 /*
 950                  * If there is undo-protected committed data against
 951                  * this buffer, then we can remove it now.  If it is a
 952                  * buffer needing such protection, the old frozen_data
 953                  * field now points to a committed version of the
 954                  * buffer, so rotate that field to the new committed
 955                  * data.
 956                  *
 957                  * Otherwise, we can just throw away the frozen data now.
 958                  *
 959                  * We also know that the frozen data has already fired
 960                  * its triggers if they exist, so we can clear that too.
 961                  */
 962                 if (jh->b_committed_data) {
 963                         jbd2_free(jh->b_committed_data, bh->b_size);
 964                         jh->b_committed_data = NULL;
 965                         if (jh->b_frozen_data) {
 966                                 jh->b_committed_data = jh->b_frozen_data;
 967                                 jh->b_frozen_data = NULL;
 968                                 jh->b_frozen_triggers = NULL;
 969                         }
 970                 } else if (jh->b_frozen_data) {
 971                         jbd2_free(jh->b_frozen_data, bh->b_size);
 972                         jh->b_frozen_data = NULL;
 973                         jh->b_frozen_triggers = NULL;
 974                 }
 975
 976                 spin_lock(&journal->j_list_lock);
 977                 cp_transaction = jh->b_cp_transaction;
 978                 if (cp_transaction) {
 979                         JBUFFER_TRACE(jh, "remove from old cp transaction");
 980                         cp_transaction->t_chp_stats.cs_dropped++;
 981                         __jbd2_journal_remove_checkpoint(jh);
 982                 }
 983
 984                 /* Only re-checkpoint the buffer_head if it is marked
 985                  * dirty.  If the buffer was added to the BJ_Forget list
 986                  * by jbd2_journal_forget, it may no longer be dirty and
 987                  * there's no point in keeping a checkpoint record for
 988                  * it. */
 989
 990                 /*
 991                 * A buffer which has been freed while still being journaled by
 992                 * a previous transaction.
 993                 */
 994                 if (buffer_freed(bh)) {
 995                         /*
 996                          * If the running transaction is the one containing
 997                          * "add to orphan" operation (b_next_transaction !=
 998                          * NULL), we have to wait for that transaction to
 999                          * commit before we can really get rid of the buffer.
1000                          * So just clear b_modified to not confuse transaction
1001                          * credit accounting and refile the buffer to
1002                          * BJ_Forget of the running transaction. If the just
1003                          * committed transaction contains "add to orphan"
1004                          * operation, we can completely invalidate the buffer
1005                          * now. We are rather through in that since the
1006                          * buffer may be still accessible when blocksize <
1007                          * pagesize and it is attached to the last partial
1008                          * page.
1009                          */
1010                         jh->b_modified = 0;
1011                         if (!jh->b_next_transaction) {
1012                                 clear_buffer_freed(bh);
1013                                 clear_buffer_jbddirty(bh);
1014                                 clear_buffer_mapped(bh);
1015                                 clear_buffer_new(bh);
1016                                 clear_buffer_req(bh);
1017                                 bh->b_bdev = NULL;
1018                         }
1019                 }
1020
1021                 if (buffer_jbddirty(bh)) {
1022                         JBUFFER_TRACE(jh, "add to new checkpointing trans");
1023                         __jbd2_journal_insert_checkpoint(jh, commit_transaction);
1024                         if (is_journal_aborted(journal))
1025                                 clear_buffer_jbddirty(bh);
1026                 } else {
1027                         J_ASSERT_BH(bh, !buffer_dirty(bh));
1028                         /*
1029                          * The buffer on BJ_Forget list and not jbddirty means
1030                          * it has been freed by this transaction and hence it
1031                          * could not have been reallocated until this
1032                          * transaction has committed. *BUT* it could be
1033                          * reallocated once we have written all the data to
1034                          * disk and before we process the buffer on BJ_Forget
1035                          * list.
1036                          */
1037                         if (!jh->b_next_transaction)
1038                                 try_to_free = 1;
1039                 }
1040                 JBUFFER_TRACE(jh, "refile or unfile buffer");
1041                 __jbd2_journal_refile_buffer(jh);
1042                 jbd_unlock_bh_state(bh);
1043                 if (try_to_free)
1044                         release_buffer_page(bh);        /* Drops bh reference */
1045                 else
1046                         __brelse(bh);
1047                 cond_resched_lock(&journal->j_list_lock);
1048         }
1049         spin_unlock(&journal->j_list_lock);
1050         /*
1051          * This is a bit sleazy.  We use j_list_lock to protect transition
1052          * of a transaction into T_FINISHED state and calling
1053          * __jbd2_journal_drop_transaction(). Otherwise we could race with
1054          * other checkpointing code processing the transaction...
1055          */
1056         write_lock(&journal->j_state_lock);
1057         spin_lock(&journal->j_list_lock);
1058         /*
1059          * Now recheck if some buffers did not get attached to the transaction
1060          * while the lock was dropped...
1061          */
1062         if (commit_transaction->t_forget) {
1063                 spin_unlock(&journal->j_list_lock);
1064                 write_unlock(&journal->j_state_lock);
1065                 goto restart_loop;
1066         }
1067
1068         /* Add the transaction to the checkpoint list
1069          * __journal_remove_checkpoint() can not destroy transaction
1070          * under us because it is not marked as T_FINISHED yet */
1071         if (journal->j_checkpoint_transactions == NULL) {
1072                 journal->j_checkpoint_transactions = commit_transaction;
1073                 commit_transaction->t_cpnext = commit_transaction;
1074                 commit_transaction->t_cpprev = commit_transaction;
1075         } else {
1076                 commit_transaction->t_cpnext =
1077                         journal->j_checkpoint_transactions;
1078                 commit_transaction->t_cpprev =
1079                         commit_transaction->t_cpnext->t_cpprev;
1080                 commit_transaction->t_cpnext->t_cpprev =
1081                         commit_transaction;
1082                 commit_transaction->t_cpprev->t_cpnext =
1083                                 commit_transaction;
1084         }
1085         spin_unlock(&journal->j_list_lock);
1086
1087         /* Done with this transaction! */
1088
1089         jbd_debug(3, "JBD2: commit phase 7\n");
1090
1091         J_ASSERT(commit_transaction->t_state == T_COMMIT_JFLUSH);
1092
1093         commit_transaction->t_start = jiffies;
1094         stats.run.rs_logging = jbd2_time_diff(stats.run.rs_logging,
1095                                               commit_transaction->t_start);
1096
1097         /*
1098          * File the transaction statistics
1099          */
1100         stats.ts_tid = commit_transaction->t_tid;
1101         stats.run.rs_handle_count =
1102                 atomic_read(&commit_transaction->t_handle_count);
1103         trace_jbd2_run_stats(journal->j_fs_dev->bd_dev,
1104                              commit_transaction->t_tid, &stats.run);
1105         stats.ts_requested = (commit_transaction->t_requested) ? 1 : 0;
1106
1107         commit_transaction->t_state = T_COMMIT_CALLBACK;
1108         J_ASSERT(commit_transaction == journal->j_committing_transaction);
1109         journal->j_commit_sequence = commit_transaction->t_tid;
1110         journal->j_committing_transaction = NULL;
1111         commit_time = ktime_to_ns(ktime_sub(ktime_get(), start_time));
1112
1113         /*
1114          * weight the commit time higher than the average time so we don't
1115          * react too strongly to vast changes in the commit time
1116          */
1117         if (likely(journal->j_average_commit_time))
1118                 journal->j_average_commit_time = (commit_time +
1119                                 journal->j_average_commit_time*3) / 4;
1120         else
1121                 journal->j_average_commit_time = commit_time;
1122
1123         write_unlock(&journal->j_state_lock);
1124
1125         if (journal->j_commit_callback)
1126                 journal->j_commit_callback(journal, commit_transaction);
1127
1128         trace_jbd2_end_commit(journal, commit_transaction);
1129         jbd_debug(1, "JBD2: commit %d complete, head %d\n",
1130                   journal->j_commit_sequence, journal->j_tail_sequence);
1131
1132         write_lock(&journal->j_state_lock);
1133         spin_lock(&journal->j_list_lock);
1134         commit_transaction->t_state = T_FINISHED;
1135         /* Check if the transaction can be dropped now that we are finished */
1136         if (commit_transaction->t_checkpoint_list == NULL &&
1137             commit_transaction->t_checkpoint_io_list == NULL) {
1138                 __jbd2_journal_drop_transaction(journal, commit_transaction);
1139                 jbd2_journal_free_transaction(commit_transaction);
1140         }
1141         spin_unlock(&journal->j_list_lock);
1142         write_unlock(&journal->j_state_lock);
1143         wake_up(&journal->j_wait_done_commit);
1144
1145         /*
1146          * Calculate overall stats
1147          */
1148         spin_lock(&journal->j_history_lock);
1149         journal->j_stats.ts_tid++;
1150         journal->j_stats.ts_requested += stats.ts_requested;
1151         journal->j_stats.run.rs_wait += stats.run.rs_wait;
1152         journal->j_stats.run.rs_request_delay += stats.run.rs_request_delay;
1153         journal->j_stats.run.rs_running += stats.run.rs_running;
1154         journal->j_stats.run.rs_locked += stats.run.rs_locked;
1155         journal->j_stats.run.rs_flushing += stats.run.rs_flushing;
1156         journal->j_stats.run.rs_logging += stats.run.rs_logging;
1157         journal->j_stats.run.rs_handle_count += stats.run.rs_handle_count;
1158         journal->j_stats.run.rs_blocks += stats.run.rs_blocks;
1159         journal->j_stats.run.rs_blocks_logged += stats.run.rs_blocks_logged;
1160         spin_unlock(&journal->j_history_lock);
1161 }