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         blk_finish_plug(&plug);
 559
 560         jbd_debug(3, "JBD2: commit phase 2b\n");
 561
 562         /*
 563          * Way to go: we have now written out all of the data for a
 564          * transaction!  Now comes the tricky part: we need to write out
 565          * metadata.  Loop over the transaction's entire buffer list:
 566          */
 567         write_lock(&journal->j_state_lock);
 568         commit_transaction->t_state = T_COMMIT;
 569         write_unlock(&journal->j_state_lock);
 570
 571         trace_jbd2_commit_logging(journal, commit_transaction);
 572         stats.run.rs_logging = jiffies;
 573         stats.run.rs_flushing = jbd2_time_diff(stats.run.rs_flushing,
 574                                                stats.run.rs_logging);
 575         stats.run.rs_blocks =
 576                 atomic_read(&commit_transaction->t_outstanding_credits);
 577         stats.run.rs_blocks_logged = 0;
 578
 579         J_ASSERT(commit_transaction->t_nr_buffers <=
 580                  atomic_read(&commit_transaction->t_outstanding_credits));
 581
 582         err = 0;
 583         bufs = 0;
 584         descriptor = NULL;
 585         blk_start_plug(&plug);
 586         while (commit_transaction->t_buffers) {
 587
 588                 /* Find the next buffer to be journaled... */
 589
 590                 jh = commit_transaction->t_buffers;
 591
 592                 /* If we're in abort mode, we just un-journal the buffer and
 593                    release it. */
 594
 595                 if (is_journal_aborted(journal)) {
 596                         clear_buffer_jbddirty(jh2bh(jh));
 597                         JBUFFER_TRACE(jh, "journal is aborting: refile");
 598                         jbd2_buffer_abort_trigger(jh,
 599                                                   jh->b_frozen_data ?
 600                                                   jh->b_frozen_triggers :
 601                                                   jh->b_triggers);
 602                         jbd2_journal_refile_buffer(journal, jh);
 603                         /* If that was the last one, we need to clean up
 604                          * any descriptor buffers which may have been
 605                          * already allocated, even if we are now
 606                          * aborting. */
 607                         if (!commit_transaction->t_buffers)
 608                                 goto start_journal_io;
 609                         continue;
 610                 }
 611
 612                 /* Make sure we have a descriptor block in which to
 613                    record the metadata buffer. */
 614
 615                 if (!descriptor) {
 616                         J_ASSERT (bufs == 0);
 617
 618                         jbd_debug(4, "JBD2: get descriptor\n");
 619
 620                         descriptor = jbd2_journal_get_descriptor_buffer(journal);
 621                         if (!descriptor) {
 622                                 jbd2_journal_abort(journal, -EIO);
 623                                 continue;
 624                         }
 625
 626                         jbd_debug(4, "JBD2: got buffer %llu (%p)\n",
 627                                 (unsigned long long)descriptor->b_blocknr,
 628                                 descriptor->b_data);
 629                         header = (journal_header_t *)descriptor->b_data;
 630                         header->h_magic     = cpu_to_be32(JBD2_MAGIC_NUMBER);
 631                         header->h_blocktype = cpu_to_be32(JBD2_DESCRIPTOR_BLOCK);
 632                         header->h_sequence  = cpu_to_be32(commit_transaction->t_tid);
 633
 634                         tagp = &descriptor->b_data[sizeof(journal_header_t)];
 635                         space_left = descriptor->b_size -
 636                                                 sizeof(journal_header_t);
 637                         first_tag = 1;
 638                         set_buffer_jwrite(descriptor);
 639                         set_buffer_dirty(descriptor);
 640                         wbuf[bufs++] = descriptor;
 641
 642                         /* Record it so that we can wait for IO
 643                            completion later */
 644                         BUFFER_TRACE(descriptor, "ph3: file as descriptor");
 645                         jbd2_file_log_bh(&log_bufs, descriptor);
 646                 }
 647
 648                 /* Where is the buffer to be written? */
 649
 650                 err = jbd2_journal_next_log_block(journal, &blocknr);
 651                 /* If the block mapping failed, just abandon the buffer
 652                    and repeat this loop: we'll fall into the
 653                    refile-on-abort condition above. */
 654                 if (err) {
 655                         jbd2_journal_abort(journal, err);
 656                         continue;
 657                 }
 658
 659                 /*
 660                  * start_this_handle() uses t_outstanding_credits to determine
 661                  * the free space in the log, but this counter is changed
 662                  * by jbd2_journal_next_log_block() also.
 663                  */
 664                 atomic_dec(&commit_transaction->t_outstanding_credits);
 665
 666                 /* Bump b_count to prevent truncate from stumbling over
 667                    the shadowed buffer!  @@@ This can go if we ever get
 668                    rid of the shadow pairing of buffers. */
 669                 atomic_inc(&jh2bh(jh)->b_count);
 670
 671                 /*
 672                  * Make a temporary IO buffer with which to write it out
 673                  * (this will requeue the metadata buffer to BJ_Shadow).
 674                  */
 675                 set_bit(BH_JWrite, &jh2bh(jh)->b_state);
 676                 JBUFFER_TRACE(jh, "ph3: write metadata");
 677                 flags = jbd2_journal_write_metadata_buffer(commit_transaction,
 678                                                 jh, &wbuf[bufs], blocknr);
 679                 if (flags < 0) {
 680                         jbd2_journal_abort(journal, flags);
 681                         continue;
 682                 }
 683                 jbd2_file_log_bh(&io_bufs, wbuf[bufs]);
 684
 685                 /* Record the new block's tag in the current descriptor
 686                    buffer */
 687
 688                 tag_flag = 0;
 689                 if (flags & 1)
 690                         tag_flag |= JBD2_FLAG_ESCAPE;
 691                 if (!first_tag)
 692                         tag_flag |= JBD2_FLAG_SAME_UUID;
 693
 694                 tag = (journal_block_tag_t *) tagp;
 695                 write_tag_block(tag_bytes, tag, jh2bh(jh)->b_blocknr);
 696                 tag->t_flags = cpu_to_be16(tag_flag);
 697                 jbd2_block_tag_csum_set(journal, tag, wbuf[bufs],
 698                                         commit_transaction->t_tid);
 699                 tagp += tag_bytes;
 700                 space_left -= tag_bytes;
 701                 bufs++;
 702
 703                 if (first_tag) {
 704                         memcpy (tagp, journal->j_uuid, 16);
 705                         tagp += 16;
 706                         space_left -= 16;
 707                         first_tag = 0;
 708                 }
 709
 710                 /* If there's no more to do, or if the descriptor is full,
 711                    let the IO rip! */
 712
 713                 if (bufs == journal->j_wbufsize ||
 714                     commit_transaction->t_buffers == NULL ||
 715                     space_left < tag_bytes + 16 + csum_size) {
 716
 717                         jbd_debug(4, "JBD2: Submit %d IOs\n", bufs);
 718
 719                         /* Write an end-of-descriptor marker before
 720                            submitting the IOs.  "tag" still points to
 721                            the last tag we set up. */
 722
 723                         tag->t_flags |= cpu_to_be16(JBD2_FLAG_LAST_TAG);
 724
 725                         jbd2_descr_block_csum_set(journal, descriptor);
 726 start_journal_io:
 727                         for (i = 0; i < bufs; i++) {
 728                                 struct buffer_head *bh = wbuf[i];
 729                                 /*
 730                                  * Compute checksum.
 731                                  */
 732                                 if (JBD2_HAS_COMPAT_FEATURE(journal,
 733                                         JBD2_FEATURE_COMPAT_CHECKSUM)) {
 734                                         crc32_sum =
 735                                             jbd2_checksum_data(crc32_sum, bh);
 736                                 }
 737
 738                                 lock_buffer(bh);
 739                                 clear_buffer_dirty(bh);
 740                                 set_buffer_uptodate(bh);
 741                                 bh->b_end_io = journal_end_buffer_io_sync;
 742                                 submit_bh(WRITE_SYNC, bh);
 743                         }
 744                         cond_resched();
 745                         stats.run.rs_blocks_logged += bufs;
 746
 747                         /* Force a new descriptor to be generated next
 748                            time round the loop. */
 749                         descriptor = NULL;
 750                         bufs = 0;
 751                 }
 752         }
 753
 754         err = journal_finish_inode_data_buffers(journal, commit_transaction);
 755         if (err) {
 756                 printk(KERN_WARNING
 757                         "JBD2: Detected IO errors while flushing file data "
 758                        "on %s\n", journal->j_devname);
 759                 if (journal->j_flags & JBD2_ABORT_ON_SYNCDATA_ERR)
 760                         jbd2_journal_abort(journal, err);
 761                 err = 0;
 762         }
 763
 764         /*
 765          * Get current oldest transaction in the log before we issue flush
 766          * to the filesystem device. After the flush we can be sure that
 767          * blocks of all older transactions are checkpointed to persistent
 768          * storage and we will be safe to update journal start in the
 769          * superblock with the numbers we get here.
 770          */
 771         update_tail =
 772                 jbd2_journal_get_log_tail(journal, &first_tid, &first_block);
 773
 774         write_lock(&journal->j_state_lock);
 775         if (update_tail) {
 776                 long freed = first_block - journal->j_tail;
 777
 778                 if (first_block < journal->j_tail)
 779                         freed += journal->j_last - journal->j_first;
 780                 /* Update tail only if we free significant amount of space */
 781                 if (freed < journal->j_maxlen / 4)
 782                         update_tail = 0;
 783         }
 784         J_ASSERT(commit_transaction->t_state == T_COMMIT);
 785         commit_transaction->t_state = T_COMMIT_DFLUSH;
 786         write_unlock(&journal->j_state_lock);
 787
 788         /*
 789          * If the journal is not located on the file system device,
 790          * then we must flush the file system device before we issue
 791          * the commit record
 792          */
 793         if (commit_transaction->t_need_data_flush &&
 794             (journal->j_fs_dev != journal->j_dev) &&
 795             (journal->j_flags & JBD2_BARRIER))
 796                 blkdev_issue_flush(journal->j_fs_dev, GFP_NOFS, NULL);
 797
 798         /* Done it all: now write the commit record asynchronously. */
 799         if (JBD2_HAS_INCOMPAT_FEATURE(journal,
 800                                       JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
 801                 err = journal_submit_commit_record(journal, commit_transaction,
 802                                                  &cbh, crc32_sum);
 803                 if (err)
 804                         __jbd2_journal_abort_hard(journal);
 805         }
 806
 807         blk_finish_plug(&plug);
 808
 809         /* Lo and behold: we have just managed to send a transaction to
 810            the log.  Before we can commit it, wait for the IO so far to
 811            complete.  Control buffers being written are on the
 812            transaction's t_log_list queue, and metadata buffers are on
 813            the io_bufs list.
 814
 815            Wait for the buffers in reverse order.  That way we are
 816            less likely to be woken up until all IOs have completed, and
 817            so we incur less scheduling load.
 818         */
 819
 820         jbd_debug(3, "JBD2: commit phase 3\n");
 821
 822         while (!list_empty(&io_bufs)) {
 823                 struct buffer_head *bh = list_entry(io_bufs.prev,
 824                                                     struct buffer_head,
 825                                                     b_assoc_buffers);
 826
 827                 wait_on_buffer(bh);
 828                 cond_resched();
 829
 830                 if (unlikely(!buffer_uptodate(bh)))
 831                         err = -EIO;
 832                 jbd2_unfile_log_bh(bh);
 833
 834                 /*
 835                  * The list contains temporary buffer heads created by
 836                  * jbd2_journal_write_metadata_buffer().
 837                  */
 838                 BUFFER_TRACE(bh, "dumping temporary bh");
 839                 __brelse(bh);
 840                 J_ASSERT_BH(bh, atomic_read(&bh->b_count) == 0);
 841                 free_buffer_head(bh);
 842
 843                 /* We also have to refile the corresponding shadowed buffer */
 844                 jh = commit_transaction->t_shadow_list->b_tprev;
 845                 bh = jh2bh(jh);
 846                 clear_buffer_jwrite(bh);
 847                 J_ASSERT_BH(bh, buffer_jbddirty(bh));
 848                 J_ASSERT_BH(bh, !buffer_shadow(bh));
 849
 850                 /* The metadata is now released for reuse, but we need
 851                    to remember it against this transaction so that when
 852                    we finally commit, we can do any checkpointing
 853                    required. */
 854                 JBUFFER_TRACE(jh, "file as BJ_Forget");
 855                 jbd2_journal_file_buffer(jh, commit_transaction, BJ_Forget);
 856                 JBUFFER_TRACE(jh, "brelse shadowed buffer");
 857                 __brelse(bh);
 858         }
 859
 860         J_ASSERT (commit_transaction->t_shadow_list == NULL);
 861
 862         jbd_debug(3, "JBD2: commit phase 4\n");
 863
 864         /* Here we wait for the revoke record and descriptor record buffers */
 865         while (!list_empty(&log_bufs)) {
 866                 struct buffer_head *bh;
 867
 868                 bh = list_entry(log_bufs.prev, struct buffer_head, b_assoc_buffers);
 869                 wait_on_buffer(bh);
 870                 cond_resched();
 871
 872                 if (unlikely(!buffer_uptodate(bh)))
 873                         err = -EIO;
 874
 875                 BUFFER_TRACE(bh, "ph5: control buffer writeout done: unfile");
 876                 clear_buffer_jwrite(bh);
 877                 jbd2_unfile_log_bh(bh);
 878                 __brelse(bh);           /* One for getblk */
 879                 /* AKPM: bforget here */
 880         }
 881
 882         if (err)
 883                 jbd2_journal_abort(journal, err);
 884
 885         jbd_debug(3, "JBD2: commit phase 5\n");
 886         write_lock(&journal->j_state_lock);
 887         J_ASSERT(commit_transaction->t_state == T_COMMIT_DFLUSH);
 888         commit_transaction->t_state = T_COMMIT_JFLUSH;
 889         write_unlock(&journal->j_state_lock);
 890
 891         if (!JBD2_HAS_INCOMPAT_FEATURE(journal,
 892                                        JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
 893                 err = journal_submit_commit_record(journal, commit_transaction,
 894                                                 &cbh, crc32_sum);
 895                 if (err)
 896                         __jbd2_journal_abort_hard(journal);
 897         }
 898         if (cbh)
 899                 err = journal_wait_on_commit_record(journal, cbh);
 900         if (JBD2_HAS_INCOMPAT_FEATURE(journal,
 901                                       JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT) &&
 902             journal->j_flags & JBD2_BARRIER) {
 903                 blkdev_issue_flush(journal->j_dev, GFP_NOFS, NULL);
 904         }
 905
 906         if (err)
 907                 jbd2_journal_abort(journal, err);
 908
 909         /*
 910          * Now disk caches for filesystem device are flushed so we are safe to
 911          * erase checkpointed transactions from the log by updating journal
 912          * superblock.
 913          */
 914         if (update_tail)
 915                 jbd2_update_log_tail(journal, first_tid, first_block);
 916
 917         /* End of a transaction!  Finally, we can do checkpoint
 918            processing: any buffers committed as a result of this
 919            transaction can be removed from any checkpoint list it was on
 920            before. */
 921
 922         jbd_debug(3, "JBD2: commit phase 6\n");
 923
 924         J_ASSERT(list_empty(&commit_transaction->t_inode_list));
 925         J_ASSERT(commit_transaction->t_buffers == NULL);
 926         J_ASSERT(commit_transaction->t_checkpoint_list == NULL);
 927         J_ASSERT(commit_transaction->t_shadow_list == NULL);
 928
 929 restart_loop:
 930         /*
 931          * As there are other places (journal_unmap_buffer()) adding buffers
 932          * to this list we have to be careful and hold the j_list_lock.
 933          */
 934         spin_lock(&journal->j_list_lock);
 935         while (commit_transaction->t_forget) {
 936                 transaction_t *cp_transaction;
 937                 struct buffer_head *bh;
 938                 int try_to_free = 0;
 939
 940                 jh = commit_transaction->t_forget;
 941                 spin_unlock(&journal->j_list_lock);
 942                 bh = jh2bh(jh);
 943                 /*
 944                  * Get a reference so that bh cannot be freed before we are
 945                  * done with it.
 946                  */
 947                 get_bh(bh);
 948                 jbd_lock_bh_state(bh);
 949                 J_ASSERT_JH(jh, jh->b_transaction == commit_transaction);
 950
 951                 /*
 952                  * If there is undo-protected committed data against
 953                  * this buffer, then we can remove it now.  If it is a
 954                  * buffer needing such protection, the old frozen_data
 955                  * field now points to a committed version of the
 956                  * buffer, so rotate that field to the new committed
 957                  * data.
 958                  *
 959                  * Otherwise, we can just throw away the frozen data now.
 960                  *
 961                  * We also know that the frozen data has already fired
 962                  * its triggers if they exist, so we can clear that too.
 963                  */
 964                 if (jh->b_committed_data) {
 965                         jbd2_free(jh->b_committed_data, bh->b_size);
 966                         jh->b_committed_data = NULL;
 967                         if (jh->b_frozen_data) {
 968                                 jh->b_committed_data = jh->b_frozen_data;
 969                                 jh->b_frozen_data = NULL;
 970                                 jh->b_frozen_triggers = NULL;
 971                         }
 972                 } else if (jh->b_frozen_data) {
 973                         jbd2_free(jh->b_frozen_data, bh->b_size);
 974                         jh->b_frozen_data = NULL;
 975                         jh->b_frozen_triggers = NULL;
 976                 }
 977
 978                 spin_lock(&journal->j_list_lock);
 979                 cp_transaction = jh->b_cp_transaction;
 980                 if (cp_transaction) {
 981                         JBUFFER_TRACE(jh, "remove from old cp transaction");
 982                         cp_transaction->t_chp_stats.cs_dropped++;
 983                         __jbd2_journal_remove_checkpoint(jh);
 984                 }
 985
 986                 /* Only re-checkpoint the buffer_head if it is marked
 987                  * dirty.  If the buffer was added to the BJ_Forget list
 988                  * by jbd2_journal_forget, it may no longer be dirty and
 989                  * there's no point in keeping a checkpoint record for
 990                  * it. */
 991
 992                 /*
 993                 * A buffer which has been freed while still being journaled by
 994                 * a previous transaction.
 995                 */
 996                 if (buffer_freed(bh)) {
 997                         /*
 998                          * If the running transaction is the one containing
 999                          * "add to orphan" operation (b_next_transaction !=
1000                          * NULL), we have to wait for that transaction to
1001                          * commit before we can really get rid of the buffer.
1002                          * So just clear b_modified to not confuse transaction
1003                          * credit accounting and refile the buffer to
1004                          * BJ_Forget of the running transaction. If the just
1005                          * committed transaction contains "add to orphan"
1006                          * operation, we can completely invalidate the buffer
1007                          * now. We are rather through in that since the
1008                          * buffer may be still accessible when blocksize <
1009                          * pagesize and it is attached to the last partial
1010                          * page.
1011                          */
1012                         jh->b_modified = 0;
1013                         if (!jh->b_next_transaction) {
1014                                 clear_buffer_freed(bh);
1015                                 clear_buffer_jbddirty(bh);
1016                                 clear_buffer_mapped(bh);
1017                                 clear_buffer_new(bh);
1018                                 clear_buffer_req(bh);
1019                                 bh->b_bdev = NULL;
1020                         }
1021                 }
1022
1023                 if (buffer_jbddirty(bh)) {
1024                         JBUFFER_TRACE(jh, "add to new checkpointing trans");
1025                         __jbd2_journal_insert_checkpoint(jh, commit_transaction);
1026                         if (is_journal_aborted(journal))
1027                                 clear_buffer_jbddirty(bh);
1028                 } else {
1029                         J_ASSERT_BH(bh, !buffer_dirty(bh));
1030                         /*
1031                          * The buffer on BJ_Forget list and not jbddirty means
1032                          * it has been freed by this transaction and hence it
1033                          * could not have been reallocated until this
1034                          * transaction has committed. *BUT* it could be
1035                          * reallocated once we have written all the data to
1036                          * disk and before we process the buffer on BJ_Forget
1037                          * list.
1038                          */
1039                         if (!jh->b_next_transaction)
1040                                 try_to_free = 1;
1041                 }
1042                 JBUFFER_TRACE(jh, "refile or unfile buffer");
1043                 __jbd2_journal_refile_buffer(jh);
1044                 jbd_unlock_bh_state(bh);
1045                 if (try_to_free)
1046                         release_buffer_page(bh);        /* Drops bh reference */
1047                 else
1048                         __brelse(bh);
1049                 cond_resched_lock(&journal->j_list_lock);
1050         }
1051         spin_unlock(&journal->j_list_lock);
1052         /*
1053          * This is a bit sleazy.  We use j_list_lock to protect transition
1054          * of a transaction into T_FINISHED state and calling
1055          * __jbd2_journal_drop_transaction(). Otherwise we could race with
1056          * other checkpointing code processing the transaction...
1057          */
1058         write_lock(&journal->j_state_lock);
1059         spin_lock(&journal->j_list_lock);
1060         /*
1061          * Now recheck if some buffers did not get attached to the transaction
1062          * while the lock was dropped...
1063          */
1064         if (commit_transaction->t_forget) {
1065                 spin_unlock(&journal->j_list_lock);
1066                 write_unlock(&journal->j_state_lock);
1067                 goto restart_loop;
1068         }
1069
1070         /* Done with this transaction! */
1071
1072         jbd_debug(3, "JBD2: commit phase 7\n");
1073
1074         J_ASSERT(commit_transaction->t_state == T_COMMIT_JFLUSH);
1075
1076         commit_transaction->t_start = jiffies;
1077         stats.run.rs_logging = jbd2_time_diff(stats.run.rs_logging,
1078                                               commit_transaction->t_start);
1079
1080         /*
1081          * File the transaction statistics
1082          */
1083         stats.ts_tid = commit_transaction->t_tid;
1084         stats.run.rs_handle_count =
1085                 atomic_read(&commit_transaction->t_handle_count);
1086         trace_jbd2_run_stats(journal->j_fs_dev->bd_dev,
1087                              commit_transaction->t_tid, &stats.run);
1088
1089         /*
1090          * Calculate overall stats
1091          */
1092         spin_lock(&journal->j_history_lock);
1093         journal->j_stats.ts_tid++;
1094         if (commit_transaction->t_requested)
1095                 journal->j_stats.ts_requested++;
1096         journal->j_stats.run.rs_wait += stats.run.rs_wait;
1097         journal->j_stats.run.rs_request_delay += stats.run.rs_request_delay;
1098         journal->j_stats.run.rs_running += stats.run.rs_running;
1099         journal->j_stats.run.rs_locked += stats.run.rs_locked;
1100         journal->j_stats.run.rs_flushing += stats.run.rs_flushing;
1101         journal->j_stats.run.rs_logging += stats.run.rs_logging;
1102         journal->j_stats.run.rs_handle_count += stats.run.rs_handle_count;
1103         journal->j_stats.run.rs_blocks += stats.run.rs_blocks;
1104         journal->j_stats.run.rs_blocks_logged += stats.run.rs_blocks_logged;
1105         spin_unlock(&journal->j_history_lock);
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_checkpoint_transactions == NULL) {
1126                 journal->j_checkpoint_transactions = commit_transaction;
1127                 commit_transaction->t_cpnext = commit_transaction;
1128                 commit_transaction->t_cpprev = commit_transaction;
1129         } else {
1130                 commit_transaction->t_cpnext =
1131                         journal->j_checkpoint_transactions;
1132                 commit_transaction->t_cpprev =
1133                         commit_transaction->t_cpnext->t_cpprev;
1134                 commit_transaction->t_cpnext->t_cpprev =
1135                         commit_transaction;
1136                 commit_transaction->t_cpprev->t_cpnext =
1137                                 commit_transaction;
1138         }
1139         spin_unlock(&journal->j_list_lock);
1140         /* Drop all spin_locks because commit_callback may be block.
1141          * __journal_remove_checkpoint() can not destroy transaction
1142          * under us because it is not marked as T_FINISHED yet */
1143         if (journal->j_commit_callback)
1144                 journal->j_commit_callback(journal, commit_transaction);
1145
1146         trace_jbd2_end_commit(journal, commit_transaction);
1147         jbd_debug(1, "JBD2: commit %d complete, head %d\n",
1148                   journal->j_commit_sequence, journal->j_tail_sequence);
1149
1150         write_lock(&journal->j_state_lock);
1151         spin_lock(&journal->j_list_lock);
1152         commit_transaction->t_state = T_FINISHED;
1153         /* Recheck checkpoint lists after j_list_lock was dropped */
1154         if (commit_transaction->t_checkpoint_list == NULL &&
1155             commit_transaction->t_checkpoint_io_list == NULL) {
1156                 __jbd2_journal_drop_transaction(journal, commit_transaction);
1157                 jbd2_journal_free_transaction(commit_transaction);
1158         }
1159         spin_unlock(&journal->j_list_lock);
1160         write_unlock(&journal->j_state_lock);
1161         wake_up(&journal->j_wait_done_commit);
1162 }