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
 347         if (!JBD2_HAS_INCOMPAT_FEATURE(j, JBD2_FEATURE_INCOMPAT_CSUM_V2))
 348                 return;
 349
 350         sequence = cpu_to_be32(sequence);
 351         addr = kmap_atomic(page);
 352         csum32 = jbd2_chksum(j, j->j_csum_seed, (__u8 *)&sequence,
 353                              sizeof(sequence));
 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         J_ASSERT(commit_transaction->t_state == T_RUNNING);
 431
 432         trace_jbd2_start_commit(journal, commit_transaction);
 433         jbd_debug(1, "JBD2: starting commit of transaction %d\n",
 434                         commit_transaction->t_tid);
 435
 436         write_lock(&journal->j_state_lock);
 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         trace_jbd2_commit_flushing(journal, commit_transaction);
 527         stats.run.rs_flushing = jiffies;
 528         stats.run.rs_locked = jbd2_time_diff(stats.run.rs_locked,
 529                                              stats.run.rs_flushing);
 530
 531         commit_transaction->t_state = T_FLUSH;
 532         journal->j_committing_transaction = commit_transaction;
 533         journal->j_running_transaction = NULL;
 534         start_time = ktime_get();
 535         commit_transaction->t_log_start = journal->j_head;
 536         wake_up(&journal->j_wait_transaction_locked);
 537         write_unlock(&journal->j_state_lock);
 538
 539         jbd_debug(3, "JBD2: commit phase 2\n");
 540
 541         /*
 542          * Now start flushing things to disk, in the order they appear
 543          * on the transaction lists.  Data blocks go first.
 544          */
 545         err = journal_submit_data_buffers(journal, commit_transaction);
 546         if (err)
 547                 jbd2_journal_abort(journal, err);
 548
 549         blk_start_plug(&plug);
 550         jbd2_journal_write_revoke_records(journal, commit_transaction,
 551                                           &log_bufs, WRITE_SYNC);
 552         blk_finish_plug(&plug);
 553
 554         jbd_debug(3, "JBD2: commit phase 2\n");
 555
 556         /*
 557          * Way to go: we have now written out all of the data for a
 558          * transaction!  Now comes the tricky part: we need to write out
 559          * metadata.  Loop over the transaction's entire buffer list:
 560          */
 561         write_lock(&journal->j_state_lock);
 562         commit_transaction->t_state = T_COMMIT;
 563         write_unlock(&journal->j_state_lock);
 564
 565         trace_jbd2_commit_logging(journal, commit_transaction);
 566         stats.run.rs_logging = jiffies;
 567         stats.run.rs_flushing = jbd2_time_diff(stats.run.rs_flushing,
 568                                                stats.run.rs_logging);
 569         stats.run.rs_blocks =
 570                 atomic_read(&commit_transaction->t_outstanding_credits);
 571         stats.run.rs_blocks_logged = 0;
 572
 573         J_ASSERT(commit_transaction->t_nr_buffers <=
 574                  atomic_read(&commit_transaction->t_outstanding_credits));
 575
 576         err = 0;
 577         bufs = 0;
 578         descriptor = NULL;
 579         blk_start_plug(&plug);
 580         while (commit_transaction->t_buffers) {
 581
 582                 /* Find the next buffer to be journaled... */
 583
 584                 jh = commit_transaction->t_buffers;
 585
 586                 /* If we're in abort mode, we just un-journal the buffer and
 587                    release it. */
 588
 589                 if (is_journal_aborted(journal)) {
 590                         clear_buffer_jbddirty(jh2bh(jh));
 591                         JBUFFER_TRACE(jh, "journal is aborting: refile");
 592                         jbd2_buffer_abort_trigger(jh,
 593                                                   jh->b_frozen_data ?
 594                                                   jh->b_frozen_triggers :
 595                                                   jh->b_triggers);
 596                         jbd2_journal_refile_buffer(journal, jh);
 597                         /* If that was the last one, we need to clean up
 598                          * any descriptor buffers which may have been
 599                          * already allocated, even if we are now
 600                          * aborting. */
 601                         if (!commit_transaction->t_buffers)
 602                                 goto start_journal_io;
 603                         continue;
 604                 }
 605
 606                 /* Make sure we have a descriptor block in which to
 607                    record the metadata buffer. */
 608
 609                 if (!descriptor) {
 610                         J_ASSERT (bufs == 0);
 611
 612                         jbd_debug(4, "JBD2: get descriptor\n");
 613
 614                         descriptor = jbd2_journal_get_descriptor_buffer(journal);
 615                         if (!descriptor) {
 616                                 jbd2_journal_abort(journal, -EIO);
 617                                 continue;
 618                         }
 619
 620                         jbd_debug(4, "JBD2: got buffer %llu (%p)\n",
 621                                 (unsigned long long)descriptor->b_blocknr,
 622                                 descriptor->b_data);
 623                         header = (journal_header_t *)descriptor->b_data;
 624                         header->h_magic     = cpu_to_be32(JBD2_MAGIC_NUMBER);
 625                         header->h_blocktype = cpu_to_be32(JBD2_DESCRIPTOR_BLOCK);
 626                         header->h_sequence  = cpu_to_be32(commit_transaction->t_tid);
 627
 628                         tagp = &descriptor->b_data[sizeof(journal_header_t)];
 629                         space_left = descriptor->b_size -
 630                                                 sizeof(journal_header_t);
 631                         first_tag = 1;
 632                         set_buffer_jwrite(descriptor);
 633                         set_buffer_dirty(descriptor);
 634                         wbuf[bufs++] = descriptor;
 635
 636                         /* Record it so that we can wait for IO
 637                            completion later */
 638                         BUFFER_TRACE(descriptor, "ph3: file as descriptor");
 639                         jbd2_file_log_bh(&log_bufs, descriptor);
 640                 }
 641
 642                 /* Where is the buffer to be written? */
 643
 644                 err = jbd2_journal_next_log_block(journal, &blocknr);
 645                 /* If the block mapping failed, just abandon the buffer
 646                    and repeat this loop: we'll fall into the
 647                    refile-on-abort condition above. */
 648                 if (err) {
 649                         jbd2_journal_abort(journal, err);
 650                         continue;
 651                 }
 652
 653                 /*
 654                  * start_this_handle() uses t_outstanding_credits to determine
 655                  * the free space in the log, but this counter is changed
 656                  * by jbd2_journal_next_log_block() also.
 657                  */
 658                 atomic_dec(&commit_transaction->t_outstanding_credits);
 659
 660                 /* Bump b_count to prevent truncate from stumbling over
 661                    the shadowed buffer!  @@@ This can go if we ever get
 662                    rid of the shadow pairing of buffers. */
 663                 atomic_inc(&jh2bh(jh)->b_count);
 664
 665                 /*
 666                  * Make a temporary IO buffer with which to write it out
 667                  * (this will requeue the metadata buffer to BJ_Shadow).
 668                  */
 669                 set_bit(BH_JWrite, &jh2bh(jh)->b_state);
 670                 JBUFFER_TRACE(jh, "ph3: write metadata");
 671                 flags = jbd2_journal_write_metadata_buffer(commit_transaction,
 672                                                 jh, &wbuf[bufs], blocknr);
 673                 if (flags < 0) {
 674                         jbd2_journal_abort(journal, flags);
 675                         continue;
 676                 }
 677                 jbd2_file_log_bh(&io_bufs, wbuf[bufs]);
 678
 679                 /* Record the new block's tag in the current descriptor
 680                    buffer */
 681
 682                 tag_flag = 0;
 683                 if (flags & 1)
 684                         tag_flag |= JBD2_FLAG_ESCAPE;
 685                 if (!first_tag)
 686                         tag_flag |= JBD2_FLAG_SAME_UUID;
 687
 688                 tag = (journal_block_tag_t *) tagp;
 689                 write_tag_block(tag_bytes, tag, jh2bh(jh)->b_blocknr);
 690                 tag->t_flags = cpu_to_be16(tag_flag);
 691                 jbd2_block_tag_csum_set(journal, tag, wbuf[bufs],
 692                                         commit_transaction->t_tid);
 693                 tagp += tag_bytes;
 694                 space_left -= tag_bytes;
 695                 bufs++;
 696
 697                 if (first_tag) {
 698                         memcpy (tagp, journal->j_uuid, 16);
 699                         tagp += 16;
 700                         space_left -= 16;
 701                         first_tag = 0;
 702                 }
 703
 704                 /* If there's no more to do, or if the descriptor is full,
 705                    let the IO rip! */
 706
 707                 if (bufs == journal->j_wbufsize ||
 708                     commit_transaction->t_buffers == NULL ||
 709                     space_left < tag_bytes + 16 + csum_size) {
 710
 711                         jbd_debug(4, "JBD2: Submit %d IOs\n", bufs);
 712
 713                         /* Write an end-of-descriptor marker before
 714                            submitting the IOs.  "tag" still points to
 715                            the last tag we set up. */
 716
 717                         tag->t_flags |= cpu_to_be16(JBD2_FLAG_LAST_TAG);
 718
 719                         jbd2_descr_block_csum_set(journal, descriptor);
 720 start_journal_io:
 721                         for (i = 0; i < bufs; i++) {
 722                                 struct buffer_head *bh = wbuf[i];
 723                                 /*
 724                                  * Compute checksum.
 725                                  */
 726                                 if (JBD2_HAS_COMPAT_FEATURE(journal,
 727                                         JBD2_FEATURE_COMPAT_CHECKSUM)) {
 728                                         crc32_sum =
 729                                             jbd2_checksum_data(crc32_sum, bh);
 730                                 }
 731
 732                                 lock_buffer(bh);
 733                                 clear_buffer_dirty(bh);
 734                                 set_buffer_uptodate(bh);
 735                                 bh->b_end_io = journal_end_buffer_io_sync;
 736                                 submit_bh(WRITE_SYNC, bh);
 737                         }
 738                         cond_resched();
 739                         stats.run.rs_blocks_logged += bufs;
 740
 741                         /* Force a new descriptor to be generated next
 742                            time round the loop. */
 743                         descriptor = NULL;
 744                         bufs = 0;
 745                 }
 746         }
 747
 748         err = journal_finish_inode_data_buffers(journal, commit_transaction);
 749         if (err) {
 750                 printk(KERN_WARNING
 751                         "JBD2: Detected IO errors while flushing file data "
 752                        "on %s\n", journal->j_devname);
 753                 if (journal->j_flags & JBD2_ABORT_ON_SYNCDATA_ERR)
 754                         jbd2_journal_abort(journal, err);
 755                 err = 0;
 756         }
 757
 758         /*
 759          * Get current oldest transaction in the log before we issue flush
 760          * to the filesystem device. After the flush we can be sure that
 761          * blocks of all older transactions are checkpointed to persistent
 762          * storage and we will be safe to update journal start in the
 763          * superblock with the numbers we get here.
 764          */
 765         update_tail =
 766                 jbd2_journal_get_log_tail(journal, &first_tid, &first_block);
 767
 768         write_lock(&journal->j_state_lock);
 769         if (update_tail) {
 770                 long freed = first_block - journal->j_tail;
 771
 772                 if (first_block < journal->j_tail)
 773                         freed += journal->j_last - journal->j_first;
 774                 /* Update tail only if we free significant amount of space */
 775                 if (freed < journal->j_maxlen / 4)
 776                         update_tail = 0;
 777         }
 778         J_ASSERT(commit_transaction->t_state == T_COMMIT);
 779         commit_transaction->t_state = T_COMMIT_DFLUSH;
 780         write_unlock(&journal->j_state_lock);
 781
 782         /*
 783          * If the journal is not located on the file system device,
 784          * then we must flush the file system device before we issue
 785          * the commit record
 786          */
 787         if (commit_transaction->t_need_data_flush &&
 788             (journal->j_fs_dev != journal->j_dev) &&
 789             (journal->j_flags & JBD2_BARRIER))
 790                 blkdev_issue_flush(journal->j_fs_dev, GFP_NOFS, NULL);
 791
 792         /* Done it all: now write the commit record asynchronously. */
 793         if (JBD2_HAS_INCOMPAT_FEATURE(journal,
 794                                       JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
 795                 err = journal_submit_commit_record(journal, commit_transaction,
 796                                                  &cbh, crc32_sum);
 797                 if (err)
 798                         __jbd2_journal_abort_hard(journal);
 799         }
 800
 801         blk_finish_plug(&plug);
 802
 803         /* Lo and behold: we have just managed to send a transaction to
 804            the log.  Before we can commit it, wait for the IO so far to
 805            complete.  Control buffers being written are on the
 806            transaction's t_log_list queue, and metadata buffers are on
 807            the io_bufs list.
 808
 809            Wait for the buffers in reverse order.  That way we are
 810            less likely to be woken up until all IOs have completed, and
 811            so we incur less scheduling load.
 812         */
 813
 814         jbd_debug(3, "JBD2: commit phase 3\n");
 815
 816         while (!list_empty(&io_bufs)) {
 817                 struct buffer_head *bh = list_entry(io_bufs.prev,
 818                                                     struct buffer_head,
 819                                                     b_assoc_buffers);
 820
 821                 wait_on_buffer(bh);
 822                 cond_resched();
 823
 824                 if (unlikely(!buffer_uptodate(bh)))
 825                         err = -EIO;
 826                 jbd2_unfile_log_bh(bh);
 827
 828                 /*
 829                  * The list contains temporary buffer heads created by
 830                  * jbd2_journal_write_metadata_buffer().
 831                  */
 832                 BUFFER_TRACE(bh, "dumping temporary bh");
 833                 __brelse(bh);
 834                 J_ASSERT_BH(bh, atomic_read(&bh->b_count) == 0);
 835                 free_buffer_head(bh);
 836
 837                 /* We also have to refile the corresponding shadowed buffer */
 838                 jh = commit_transaction->t_shadow_list->b_tprev;
 839                 bh = jh2bh(jh);
 840                 clear_buffer_jwrite(bh);
 841                 J_ASSERT_BH(bh, buffer_jbddirty(bh));
 842                 J_ASSERT_BH(bh, !buffer_shadow(bh));
 843
 844                 /* The metadata is now released for reuse, but we need
 845                    to remember it against this transaction so that when
 846                    we finally commit, we can do any checkpointing
 847                    required. */
 848                 JBUFFER_TRACE(jh, "file as BJ_Forget");
 849                 jbd2_journal_file_buffer(jh, commit_transaction, BJ_Forget);
 850                 JBUFFER_TRACE(jh, "brelse shadowed buffer");
 851                 __brelse(bh);
 852         }
 853
 854         J_ASSERT (commit_transaction->t_shadow_list == NULL);
 855
 856         jbd_debug(3, "JBD2: commit phase 4\n");
 857
 858         /* Here we wait for the revoke record and descriptor record buffers */
 859         while (!list_empty(&log_bufs)) {
 860                 struct buffer_head *bh;
 861
 862                 bh = list_entry(log_bufs.prev, struct buffer_head, b_assoc_buffers);
 863                 wait_on_buffer(bh);
 864                 cond_resched();
 865
 866                 if (unlikely(!buffer_uptodate(bh)))
 867                         err = -EIO;
 868
 869                 BUFFER_TRACE(bh, "ph5: control buffer writeout done: unfile");
 870                 clear_buffer_jwrite(bh);
 871                 jbd2_unfile_log_bh(bh);
 872                 __brelse(bh);           /* One for getblk */
 873                 /* AKPM: bforget here */
 874         }
 875
 876         if (err)
 877                 jbd2_journal_abort(journal, err);
 878
 879         jbd_debug(3, "JBD2: commit phase 5\n");
 880         write_lock(&journal->j_state_lock);
 881         J_ASSERT(commit_transaction->t_state == T_COMMIT_DFLUSH);
 882         commit_transaction->t_state = T_COMMIT_JFLUSH;
 883         write_unlock(&journal->j_state_lock);
 884
 885         if (!JBD2_HAS_INCOMPAT_FEATURE(journal,
 886                                        JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
 887                 err = journal_submit_commit_record(journal, commit_transaction,
 888                                                 &cbh, crc32_sum);
 889                 if (err)
 890                         __jbd2_journal_abort_hard(journal);
 891         }
 892         if (cbh)
 893                 err = journal_wait_on_commit_record(journal, cbh);
 894         if (JBD2_HAS_INCOMPAT_FEATURE(journal,
 895                                       JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT) &&
 896             journal->j_flags & JBD2_BARRIER) {
 897                 blkdev_issue_flush(journal->j_dev, GFP_NOFS, NULL);
 898         }
 899
 900         if (err)
 901                 jbd2_journal_abort(journal, err);
 902
 903         /*
 904          * Now disk caches for filesystem device are flushed so we are safe to
 905          * erase checkpointed transactions from the log by updating journal
 906          * superblock.
 907          */
 908         if (update_tail)
 909                 jbd2_update_log_tail(journal, first_tid, first_block);
 910
 911         /* End of a transaction!  Finally, we can do checkpoint
 912            processing: any buffers committed as a result of this
 913            transaction can be removed from any checkpoint list it was on
 914            before. */
 915
 916         jbd_debug(3, "JBD2: commit phase 6\n");
 917
 918         J_ASSERT(list_empty(&commit_transaction->t_inode_list));
 919         J_ASSERT(commit_transaction->t_buffers == NULL);
 920         J_ASSERT(commit_transaction->t_checkpoint_list == NULL);
 921         J_ASSERT(commit_transaction->t_shadow_list == NULL);
 922
 923 restart_loop:
 924         /*
 925          * As there are other places (journal_unmap_buffer()) adding buffers
 926          * to this list we have to be careful and hold the j_list_lock.
 927          */
 928         spin_lock(&journal->j_list_lock);
 929         while (commit_transaction->t_forget) {
 930                 transaction_t *cp_transaction;
 931                 struct buffer_head *bh;
 932                 int try_to_free = 0;
 933
 934                 jh = commit_transaction->t_forget;
 935                 spin_unlock(&journal->j_list_lock);
 936                 bh = jh2bh(jh);
 937                 /*
 938                  * Get a reference so that bh cannot be freed before we are
 939                  * done with it.
 940                  */
 941                 get_bh(bh);
 942                 jbd_lock_bh_state(bh);
 943                 J_ASSERT_JH(jh, jh->b_transaction == commit_transaction);
 944
 945                 /*
 946                  * If there is undo-protected committed data against
 947                  * this buffer, then we can remove it now.  If it is a
 948                  * buffer needing such protection, the old frozen_data
 949                  * field now points to a committed version of the
 950                  * buffer, so rotate that field to the new committed
 951                  * data.
 952                  *
 953                  * Otherwise, we can just throw away the frozen data now.
 954                  *
 955                  * We also know that the frozen data has already fired
 956                  * its triggers if they exist, so we can clear that too.
 957                  */
 958                 if (jh->b_committed_data) {
 959                         jbd2_free(jh->b_committed_data, bh->b_size);
 960                         jh->b_committed_data = NULL;
 961                         if (jh->b_frozen_data) {
 962                                 jh->b_committed_data = jh->b_frozen_data;
 963                                 jh->b_frozen_data = NULL;
 964                                 jh->b_frozen_triggers = NULL;
 965                         }
 966                 } else if (jh->b_frozen_data) {
 967                         jbd2_free(jh->b_frozen_data, bh->b_size);
 968                         jh->b_frozen_data = NULL;
 969                         jh->b_frozen_triggers = NULL;
 970                 }
 971
 972                 spin_lock(&journal->j_list_lock);
 973                 cp_transaction = jh->b_cp_transaction;
 974                 if (cp_transaction) {
 975                         JBUFFER_TRACE(jh, "remove from old cp transaction");
 976                         cp_transaction->t_chp_stats.cs_dropped++;
 977                         __jbd2_journal_remove_checkpoint(jh);
 978                 }
 979
 980                 /* Only re-checkpoint the buffer_head if it is marked
 981                  * dirty.  If the buffer was added to the BJ_Forget list
 982                  * by jbd2_journal_forget, it may no longer be dirty and
 983                  * there's no point in keeping a checkpoint record for
 984                  * it. */
 985
 986                 /*
 987                 * A buffer which has been freed while still being journaled by
 988                 * a previous transaction.
 989                 */
 990                 if (buffer_freed(bh)) {
 991                         /*
 992                          * If the running transaction is the one containing
 993                          * "add to orphan" operation (b_next_transaction !=
 994                          * NULL), we have to wait for that transaction to
 995                          * commit before we can really get rid of the buffer.
 996                          * So just clear b_modified to not confuse transaction
 997                          * credit accounting and refile the buffer to
 998                          * BJ_Forget of the running transaction. If the just
 999                          * committed transaction contains "add to orphan"
1000                          * operation, we can completely invalidate the buffer
1001                          * now. We are rather through in that since the
1002                          * buffer may be still accessible when blocksize <
1003                          * pagesize and it is attached to the last partial
1004                          * page.
1005                          */
1006                         jh->b_modified = 0;
1007                         if (!jh->b_next_transaction) {
1008                                 clear_buffer_freed(bh);
1009                                 clear_buffer_jbddirty(bh);
1010                                 clear_buffer_mapped(bh);
1011                                 clear_buffer_new(bh);
1012                                 clear_buffer_req(bh);
1013                                 bh->b_bdev = NULL;
1014                         }
1015                 }
1016
1017                 if (buffer_jbddirty(bh)) {
1018                         JBUFFER_TRACE(jh, "add to new checkpointing trans");
1019                         __jbd2_journal_insert_checkpoint(jh, commit_transaction);
1020                         if (is_journal_aborted(journal))
1021                                 clear_buffer_jbddirty(bh);
1022                 } else {
1023                         J_ASSERT_BH(bh, !buffer_dirty(bh));
1024                         /*
1025                          * The buffer on BJ_Forget list and not jbddirty means
1026                          * it has been freed by this transaction and hence it
1027                          * could not have been reallocated until this
1028                          * transaction has committed. *BUT* it could be
1029                          * reallocated once we have written all the data to
1030                          * disk and before we process the buffer on BJ_Forget
1031                          * list.
1032                          */
1033                         if (!jh->b_next_transaction)
1034                                 try_to_free = 1;
1035                 }
1036                 JBUFFER_TRACE(jh, "refile or unfile buffer");
1037                 __jbd2_journal_refile_buffer(jh);
1038                 jbd_unlock_bh_state(bh);
1039                 if (try_to_free)
1040                         release_buffer_page(bh);        /* Drops bh reference */
1041                 else
1042                         __brelse(bh);
1043                 cond_resched_lock(&journal->j_list_lock);
1044         }
1045         spin_unlock(&journal->j_list_lock);
1046         /*
1047          * This is a bit sleazy.  We use j_list_lock to protect transition
1048          * of a transaction into T_FINISHED state and calling
1049          * __jbd2_journal_drop_transaction(). Otherwise we could race with
1050          * other checkpointing code processing the transaction...
1051          */
1052         write_lock(&journal->j_state_lock);
1053         spin_lock(&journal->j_list_lock);
1054         /*
1055          * Now recheck if some buffers did not get attached to the transaction
1056          * while the lock was dropped...
1057          */
1058         if (commit_transaction->t_forget) {
1059                 spin_unlock(&journal->j_list_lock);
1060                 write_unlock(&journal->j_state_lock);
1061                 goto restart_loop;
1062         }
1063
1064         /* Done with this transaction! */
1065
1066         jbd_debug(3, "JBD2: commit phase 7\n");
1067
1068         J_ASSERT(commit_transaction->t_state == T_COMMIT_JFLUSH);
1069
1070         commit_transaction->t_start = jiffies;
1071         stats.run.rs_logging = jbd2_time_diff(stats.run.rs_logging,
1072                                               commit_transaction->t_start);
1073
1074         /*
1075          * File the transaction statistics
1076          */
1077         stats.ts_tid = commit_transaction->t_tid;
1078         stats.run.rs_handle_count =
1079                 atomic_read(&commit_transaction->t_handle_count);
1080         trace_jbd2_run_stats(journal->j_fs_dev->bd_dev,
1081                              commit_transaction->t_tid, &stats.run);
1082
1083         /*
1084          * Calculate overall stats
1085          */
1086         spin_lock(&journal->j_history_lock);
1087         journal->j_stats.ts_tid++;
1088         if (commit_transaction->t_requested)
1089                 journal->j_stats.ts_requested++;
1090         journal->j_stats.run.rs_wait += stats.run.rs_wait;
1091         journal->j_stats.run.rs_request_delay += stats.run.rs_request_delay;
1092         journal->j_stats.run.rs_running += stats.run.rs_running;
1093         journal->j_stats.run.rs_locked += stats.run.rs_locked;
1094         journal->j_stats.run.rs_flushing += stats.run.rs_flushing;
1095         journal->j_stats.run.rs_logging += stats.run.rs_logging;
1096         journal->j_stats.run.rs_handle_count += stats.run.rs_handle_count;
1097         journal->j_stats.run.rs_blocks += stats.run.rs_blocks;
1098         journal->j_stats.run.rs_blocks_logged += stats.run.rs_blocks_logged;
1099         spin_unlock(&journal->j_history_lock);
1100
1101         commit_transaction->t_state = T_COMMIT_CALLBACK;
1102         J_ASSERT(commit_transaction == journal->j_committing_transaction);
1103         journal->j_commit_sequence = commit_transaction->t_tid;
1104         journal->j_committing_transaction = NULL;
1105         commit_time = ktime_to_ns(ktime_sub(ktime_get(), start_time));
1106
1107         /*
1108          * weight the commit time higher than the average time so we don't
1109          * react too strongly to vast changes in the commit time
1110          */
1111         if (likely(journal->j_average_commit_time))
1112                 journal->j_average_commit_time = (commit_time +
1113                                 journal->j_average_commit_time*3) / 4;
1114         else
1115                 journal->j_average_commit_time = commit_time;
1116
1117         write_unlock(&journal->j_state_lock);
1118
1119         if (journal->j_checkpoint_transactions == NULL) {
1120                 journal->j_checkpoint_transactions = commit_transaction;
1121                 commit_transaction->t_cpnext = commit_transaction;
1122                 commit_transaction->t_cpprev = commit_transaction;
1123         } else {
1124                 commit_transaction->t_cpnext =
1125                         journal->j_checkpoint_transactions;
1126                 commit_transaction->t_cpprev =
1127                         commit_transaction->t_cpnext->t_cpprev;
1128                 commit_transaction->t_cpnext->t_cpprev =
1129                         commit_transaction;
1130                 commit_transaction->t_cpprev->t_cpnext =
1131                                 commit_transaction;
1132         }
1133         spin_unlock(&journal->j_list_lock);
1134         /* Drop all spin_locks because commit_callback may be block.
1135          * __journal_remove_checkpoint() can not destroy transaction
1136          * under us because it is not marked as T_FINISHED yet */
1137         if (journal->j_commit_callback)
1138                 journal->j_commit_callback(journal, commit_transaction);
1139
1140         trace_jbd2_end_commit(journal, commit_transaction);
1141         jbd_debug(1, "JBD2: commit %d complete, head %d\n",
1142                   journal->j_commit_sequence, journal->j_tail_sequence);
1143
1144         write_lock(&journal->j_state_lock);
1145         spin_lock(&journal->j_list_lock);
1146         commit_transaction->t_state = T_FINISHED;
1147         /* Recheck checkpoint lists after j_list_lock was dropped */
1148         if (commit_transaction->t_checkpoint_list == NULL &&
1149             commit_transaction->t_checkpoint_io_list == NULL) {
1150                 __jbd2_journal_drop_transaction(journal, commit_transaction);
1151                 jbd2_journal_free_transaction(commit_transaction);
1152         }
1153         spin_unlock(&journal->j_list_lock);
1154         write_unlock(&journal->j_state_lock);
1155         wake_up(&journal->j_wait_done_commit);
1156 }