fs/jbd2/commit.c

   1 // SPDX-License-Identifier: GPL-2.0+
   2 /*
   3  * linux/fs/jbd2/commit.c
   4  *
   5  * Written by Stephen C. Tweedie <sct@redhat.com>, 1998
   6  *
   7  * Copyright 1998 Red Hat corp --- All Rights Reserved
   8  *
   9  * Journal commit routines for the generic filesystem journaling code;
  10  * part of the ext2fs journaling system.
  11  */
  12
  13 #include <linux/time.h>
  14 #include <linux/fs.h>
  15 #include <linux/jbd2.h>
  16 #include <linux/errno.h>
  17 #include <linux/slab.h>
  18 #include <linux/mm.h>
  19 #include <linux/pagemap.h>
  20 #include <linux/jiffies.h>
  21 #include <linux/crc32.h>
  22 #include <linux/writeback.h>
  23 #include <linux/backing-dev.h>
  24 #include <linux/bio.h>
  25 #include <linux/blkdev.h>
  26 #include <linux/bitops.h>
  27 #include <trace/events/jbd2.h>
  28
  29 /*
  30  * IO end handler for temporary buffer_heads handling writes to the journal.
  31  */
  32 static void journal_end_buffer_io_sync(struct buffer_head *bh, int uptodate)
  33 {
  34         struct buffer_head *orig_bh = bh->b_private;
  35
  36         BUFFER_TRACE(bh, "");
  37         if (uptodate)
  38                 set_buffer_uptodate(bh);
  39         else
  40                 clear_buffer_uptodate(bh);
  41         if (orig_bh) {
  42                 clear_bit_unlock(BH_Shadow, &orig_bh->b_state);
  43                 smp_mb__after_atomic();
  44                 wake_up_bit(&orig_bh->b_state, BH_Shadow);
  45         }
  46         unlock_buffer(bh);
  47 }
  48
  49 /*
  50  * When an ext4 file is truncated, it is possible that some pages are not
  51  * successfully freed, because they are attached to a committing transaction.
  52  * After the transaction commits, these pages are left on the LRU, with no
  53  * ->mapping, and with attached buffers.  These pages are trivially reclaimable
  54  * by the VM, but their apparent absence upsets the VM accounting, and it makes
  55  * the numbers in /proc/meminfo look odd.
  56  *
  57  * So here, we have a buffer which has just come off the forget list.  Look to
  58  * see if we can strip all buffers from the backing page.
  59  *
  60  * Called under lock_journal(), and possibly under journal_datalist_lock.  The
  61  * caller provided us with a ref against the buffer, and we drop that here.
  62  */
  63 static void release_buffer_page(struct buffer_head *bh)
  64 {
  65         struct page *page;
  66
  67         if (buffer_dirty(bh))
  68                 goto nope;
  69         if (atomic_read(&bh->b_count) != 1)
  70                 goto nope;
  71         page = bh->b_page;
  72         if (!page)
  73                 goto nope;
  74         if (page->mapping)
  75                 goto nope;
  76
  77         /* OK, it's a truncated page */
  78         if (!trylock_page(page))
  79                 goto nope;
  80
  81         get_page(page);
  82         __brelse(bh);
  83         try_to_free_buffers(page);
  84         unlock_page(page);
  85         put_page(page);
  86         return;
  87
  88 nope:
  89         __brelse(bh);
  90 }
  91
  92 static void jbd2_commit_block_csum_set(journal_t *j, struct buffer_head *bh)
  93 {
  94         struct commit_header *h;
  95         __u32 csum;
  96
  97         if (!jbd2_journal_has_csum_v2or3(j))
  98                 return;
  99
 100         h = (struct commit_header *)(bh->b_data);
 101         h->h_chksum_type = 0;
 102         h->h_chksum_size = 0;
 103         h->h_chksum[0] = 0;
 104         csum = jbd2_chksum(j, j->j_csum_seed, bh->b_data, j->j_blocksize);
 105         h->h_chksum[0] = cpu_to_be32(csum);
 106 }
 107
 108 /*
 109  * Done it all: now submit the commit record.  We should have
 110  * cleaned up our previous buffers by now, so if we are in abort
 111  * mode we can now just skip the rest of the journal write
 112  * entirely.
 113  *
 114  * Returns 1 if the journal needs to be aborted or 0 on success
 115  */
 116 static int journal_submit_commit_record(journal_t *journal,
 117                                         transaction_t *commit_transaction,
 118                                         struct buffer_head **cbh,
 119                                         __u32 crc32_sum)
 120 {
 121         struct commit_header *tmp;
 122         struct buffer_head *bh;
 123         int ret;
 124         struct timespec64 now;
 125
 126         *cbh = NULL;
 127
 128         if (is_journal_aborted(journal))
 129                 return 0;
 130
 131         bh = jbd2_journal_get_descriptor_buffer(commit_transaction,
 132                                                 JBD2_COMMIT_BLOCK);
 133         if (!bh)
 134                 return 1;
 135
 136         tmp = (struct commit_header *)bh->b_data;
 137         ktime_get_coarse_real_ts64(&now);
 138         tmp->h_commit_sec = cpu_to_be64(now.tv_sec);
 139         tmp->h_commit_nsec = cpu_to_be32(now.tv_nsec);
 140
 141         if (jbd2_has_feature_checksum(journal)) {
 142                 tmp->h_chksum_type      = JBD2_CRC32_CHKSUM;
 143                 tmp->h_chksum_size      = JBD2_CRC32_CHKSUM_SIZE;
 144                 tmp->h_chksum[0]        = cpu_to_be32(crc32_sum);
 145         }
 146         jbd2_commit_block_csum_set(journal, bh);
 147
 148         BUFFER_TRACE(bh, "submit commit block");
 149         lock_buffer(bh);
 150         clear_buffer_dirty(bh);
 151         set_buffer_uptodate(bh);
 152         bh->b_end_io = journal_end_buffer_io_sync;
 153
 154         if (journal->j_flags & JBD2_BARRIER &&
 155             !jbd2_has_feature_async_commit(journal))
 156                 ret = submit_bh(REQ_OP_WRITE,
 157                         REQ_SYNC | REQ_PREFLUSH | REQ_FUA, bh);
 158         else
 159                 ret = submit_bh(REQ_OP_WRITE, REQ_SYNC, bh);
 160
 161         *cbh = bh;
 162         return ret;
 163 }
 164
 165 /*
 166  * This function along with journal_submit_commit_record
 167  * allows to write the commit record asynchronously.
 168  */
 169 static int journal_wait_on_commit_record(journal_t *journal,
 170                                          struct buffer_head *bh)
 171 {
 172         int ret = 0;
 173
 174         clear_buffer_dirty(bh);
 175         wait_on_buffer(bh);
 176
 177         if (unlikely(!buffer_uptodate(bh)))
 178                 ret = -EIO;
 179         put_bh(bh);            /* One for getblk() */
 180
 181         return ret;
 182 }
 183
 184 /*
 185  * write the filemap data using writepage() address_space_operations.
 186  * We don't do block allocation here even for delalloc. We don't
 187  * use writepages() because with delayed allocation we may be doing
 188  * block allocation in writepages().
 189  */
 190 static int journal_submit_inode_data_buffers(struct address_space *mapping,
 191                 loff_t dirty_start, loff_t dirty_end)
 192 {
 193         int ret;
 194         struct writeback_control wbc = {
 195                 .sync_mode =  WB_SYNC_ALL,
 196                 .nr_to_write = mapping->nrpages * 2,
 197                 .range_start = dirty_start,
 198                 .range_end = dirty_end,
 199         };
 200
 201         ret = generic_writepages(mapping, &wbc);
 202         return ret;
 203 }
 204
 205 /*
 206  * Submit all the data buffers of inode associated with the transaction to
 207  * disk.
 208  *
 209  * We are in a committing transaction. Therefore no new inode can be added to
 210  * our inode list. We use JI_COMMIT_RUNNING flag to protect inode we currently
 211  * operate on from being released while we write out pages.
 212  */
 213 static int journal_submit_data_buffers(journal_t *journal,
 214                 transaction_t *commit_transaction)
 215 {
 216         struct jbd2_inode *jinode;
 217         int err, ret = 0;
 218         struct address_space *mapping;
 219
 220         spin_lock(&journal->j_list_lock);
 221         list_for_each_entry(jinode, &commit_transaction->t_inode_list, i_list) {
 222                 loff_t dirty_start = jinode->i_dirty_start;
 223                 loff_t dirty_end = jinode->i_dirty_end;
 224
 225                 if (!(jinode->i_flags & JI_WRITE_DATA))
 226                         continue;
 227                 mapping = jinode->i_vfs_inode->i_mapping;
 228                 jinode->i_flags |= JI_COMMIT_RUNNING;
 229                 spin_unlock(&journal->j_list_lock);
 230                 /*
 231                  * submit the inode data buffers. We use writepage
 232                  * instead of writepages. Because writepages can do
 233                  * block allocation  with delalloc. We need to write
 234                  * only allocated blocks here.
 235                  */
 236                 trace_jbd2_submit_inode_data(jinode->i_vfs_inode);
 237                 err = journal_submit_inode_data_buffers(mapping, dirty_start,
 238                                 dirty_end);
 239                 if (!ret)
 240                         ret = err;
 241                 spin_lock(&journal->j_list_lock);
 242                 J_ASSERT(jinode->i_transaction == commit_transaction);
 243                 jinode->i_flags &= ~JI_COMMIT_RUNNING;
 244                 smp_mb();
 245                 wake_up_bit(&jinode->i_flags, __JI_COMMIT_RUNNING);
 246         }
 247         spin_unlock(&journal->j_list_lock);
 248         return ret;
 249 }
 250
 251 /*
 252  * Wait for data submitted for writeout, refile inodes to proper
 253  * transaction if needed.
 254  *
 255  */
 256 static int journal_finish_inode_data_buffers(journal_t *journal,
 257                 transaction_t *commit_transaction)
 258 {
 259         struct jbd2_inode *jinode, *next_i;
 260         int err, ret = 0;
 261
 262         /* For locking, see the comment in journal_submit_data_buffers() */
 263         spin_lock(&journal->j_list_lock);
 264         list_for_each_entry(jinode, &commit_transaction->t_inode_list, i_list) {
 265                 loff_t dirty_start = jinode->i_dirty_start;
 266                 loff_t dirty_end = jinode->i_dirty_end;
 267
 268                 if (!(jinode->i_flags & JI_WAIT_DATA))
 269                         continue;
 270                 jinode->i_flags |= JI_COMMIT_RUNNING;
 271                 spin_unlock(&journal->j_list_lock);
 272                 err = filemap_fdatawait_range_keep_errors(
 273                                 jinode->i_vfs_inode->i_mapping, dirty_start,
 274                                 dirty_end);
 275                 if (!ret)
 276                         ret = err;
 277                 spin_lock(&journal->j_list_lock);
 278                 jinode->i_flags &= ~JI_COMMIT_RUNNING;
 279                 smp_mb();
 280                 wake_up_bit(&jinode->i_flags, __JI_COMMIT_RUNNING);
 281         }
 282
 283         /* Now refile inode to proper lists */
 284         list_for_each_entry_safe(jinode, next_i,
 285                                  &commit_transaction->t_inode_list, i_list) {
 286                 list_del(&jinode->i_list);
 287                 if (jinode->i_next_transaction) {
 288                         jinode->i_transaction = jinode->i_next_transaction;
 289                         jinode->i_next_transaction = NULL;
 290                         list_add(&jinode->i_list,
 291                                 &jinode->i_transaction->t_inode_list);
 292                 } else {
 293                         jinode->i_transaction = NULL;
 294                         jinode->i_dirty_start = 0;
 295                         jinode->i_dirty_end = 0;
 296                 }
 297         }
 298         spin_unlock(&journal->j_list_lock);
 299
 300         return ret;
 301 }
 302
 303 static __u32 jbd2_checksum_data(__u32 crc32_sum, struct buffer_head *bh)
 304 {
 305         struct page *page = bh->b_page;
 306         char *addr;
 307         __u32 checksum;
 308
 309         addr = kmap_atomic(page);
 310         checksum = crc32_be(crc32_sum,
 311                 (void *)(addr + offset_in_page(bh->b_data)), bh->b_size);
 312         kunmap_atomic(addr);
 313
 314         return checksum;
 315 }
 316
 317 static void write_tag_block(journal_t *j, journal_block_tag_t *tag,
 318                                    unsigned long long block)
 319 {
 320         tag->t_blocknr = cpu_to_be32(block & (u32)~0);
 321         if (jbd2_has_feature_64bit(j))
 322                 tag->t_blocknr_high = cpu_to_be32((block >> 31) >> 1);
 323 }
 324
 325 static void jbd2_block_tag_csum_set(journal_t *j, journal_block_tag_t *tag,
 326                                     struct buffer_head *bh, __u32 sequence)
 327 {
 328         journal_block_tag3_t *tag3 = (journal_block_tag3_t *)tag;
 329         struct page *page = bh->b_page;
 330         __u8 *addr;
 331         __u32 csum32;
 332         __be32 seq;
 333
 334         if (!jbd2_journal_has_csum_v2or3(j))
 335                 return;
 336
 337         seq = cpu_to_be32(sequence);
 338         addr = kmap_atomic(page);
 339         csum32 = jbd2_chksum(j, j->j_csum_seed, (__u8 *)&seq, sizeof(seq));
 340         csum32 = jbd2_chksum(j, csum32, addr + offset_in_page(bh->b_data),
 341                              bh->b_size);
 342         kunmap_atomic(addr);
 343
 344         if (jbd2_has_feature_csum3(j))
 345                 tag3->t_checksum = cpu_to_be32(csum32);
 346         else
 347                 tag->t_checksum = cpu_to_be16(csum32);
 348 }
 349 /*
 350  * jbd2_journal_commit_transaction
 351  *
 352  * The primary function for committing a transaction to the log.  This
 353  * function is called by the journal thread to begin a complete commit.
 354  */
 355 void jbd2_journal_commit_transaction(journal_t *journal)
 356 {
 357         struct transaction_stats_s stats;
 358         transaction_t *commit_transaction;
 359         struct journal_head *jh;
 360         struct buffer_head *descriptor;
 361         struct buffer_head **wbuf = journal->j_wbuf;
 362         int bufs;
 363         int flags;
 364         int err;
 365         unsigned long long blocknr;
 366         ktime_t start_time;
 367         u64 commit_time;
 368         char *tagp = NULL;
 369         journal_block_tag_t *tag = NULL;
 370         int space_left = 0;
 371         int first_tag = 0;
 372         int tag_flag;
 373         int i;
 374         int tag_bytes = journal_tag_bytes(journal);
 375         struct buffer_head *cbh = NULL; /* For transactional checksums */
 376         __u32 crc32_sum = ~0;
 377         struct blk_plug plug;
 378         /* Tail of the journal */
 379         unsigned long first_block;
 380         tid_t first_tid;
 381         int update_tail;
 382         int csum_size = 0;
 383         LIST_HEAD(io_bufs);
 384         LIST_HEAD(log_bufs);
 385
 386         if (jbd2_journal_has_csum_v2or3(journal))
 387                 csum_size = sizeof(struct jbd2_journal_block_tail);
 388
 389         /*
 390          * First job: lock down the current transaction and wait for
 391          * all outstanding updates to complete.
 392          */
 393
 394         /* Do we need to erase the effects of a prior jbd2_journal_flush? */
 395         if (journal->j_flags & JBD2_FLUSHED) {
 396                 jbd_debug(3, "super block updated\n");
 397                 mutex_lock_io(&journal->j_checkpoint_mutex);
 398                 /*
 399                  * We hold j_checkpoint_mutex so tail cannot change under us.
 400                  * We don't need any special data guarantees for writing sb
 401                  * since journal is empty and it is ok for write to be
 402                  * flushed only with transaction commit.
 403                  */
 404                 jbd2_journal_update_sb_log_tail(journal,
 405                                                 journal->j_tail_sequence,
 406                                                 journal->j_tail,
 407                                                 REQ_SYNC);
 408                 mutex_unlock(&journal->j_checkpoint_mutex);
 409         } else {
 410                 jbd_debug(3, "superblock not updated\n");
 411         }
 412
 413         J_ASSERT(journal->j_running_transaction != NULL);
 414         J_ASSERT(journal->j_committing_transaction == NULL);
 415
 416         commit_transaction = journal->j_running_transaction;
 417
 418         trace_jbd2_start_commit(journal, commit_transaction);
 419         jbd_debug(1, "JBD2: starting commit of transaction %d\n",
 420                         commit_transaction->t_tid);
 421
 422         write_lock(&journal->j_state_lock);
 423         J_ASSERT(commit_transaction->t_state == T_RUNNING);
 424         commit_transaction->t_state = T_LOCKED;
 425
 426         trace_jbd2_commit_locking(journal, commit_transaction);
 427         stats.run.rs_wait = commit_transaction->t_max_wait;
 428         stats.run.rs_request_delay = 0;
 429         stats.run.rs_locked = jiffies;
 430         if (commit_transaction->t_requested)
 431                 stats.run.rs_request_delay =
 432                         jbd2_time_diff(commit_transaction->t_requested,
 433                                        stats.run.rs_locked);
 434         stats.run.rs_running = jbd2_time_diff(commit_transaction->t_start,
 435                                               stats.run.rs_locked);
 436
 437         spin_lock(&commit_transaction->t_handle_lock);
 438         while (atomic_read(&commit_transaction->t_updates)) {
 439                 DEFINE_WAIT(wait);
 440
 441                 prepare_to_wait(&journal->j_wait_updates, &wait,
 442                                         TASK_UNINTERRUPTIBLE);
 443                 if (atomic_read(&commit_transaction->t_updates)) {
 444                         spin_unlock(&commit_transaction->t_handle_lock);
 445                         write_unlock(&journal->j_state_lock);
 446                         schedule();
 447                         write_lock(&journal->j_state_lock);
 448                         spin_lock(&commit_transaction->t_handle_lock);
 449                 }
 450                 finish_wait(&journal->j_wait_updates, &wait);
 451         }
 452         spin_unlock(&commit_transaction->t_handle_lock);
 453         commit_transaction->t_state = T_SWITCH;
 454
 455         J_ASSERT (atomic_read(&commit_transaction->t_outstanding_credits) <=
 456                         journal->j_max_transaction_buffers);
 457
 458         /*
 459          * First thing we are allowed to do is to discard any remaining
 460          * BJ_Reserved buffers.  Note, it is _not_ permissible to assume
 461          * that there are no such buffers: if a large filesystem
 462          * operation like a truncate needs to split itself over multiple
 463          * transactions, then it may try to do a jbd2_journal_restart() while
 464          * there are still BJ_Reserved buffers outstanding.  These must
 465          * be released cleanly from the current transaction.
 466          *
 467          * In this case, the filesystem must still reserve write access
 468          * again before modifying the buffer in the new transaction, but
 469          * we do not require it to remember exactly which old buffers it
 470          * has reserved.  This is consistent with the existing behaviour
 471          * that multiple jbd2_journal_get_write_access() calls to the same
 472          * buffer are perfectly permissible.
 473          * We use journal->j_state_lock here to serialize processing of
 474          * t_reserved_list with eviction of buffers from journal_unmap_buffer().
 475          */
 476         while (commit_transaction->t_reserved_list) {
 477                 jh = commit_transaction->t_reserved_list;
 478                 JBUFFER_TRACE(jh, "reserved, unused: refile");
 479                 /*
 480                  * A jbd2_journal_get_undo_access()+jbd2_journal_release_buffer() may
 481                  * leave undo-committed data.
 482                  */
 483                 if (jh->b_committed_data) {
 484                         struct buffer_head *bh = jh2bh(jh);
 485
 486                         jbd_lock_bh_state(bh);
 487                         jbd2_free(jh->b_committed_data, bh->b_size);
 488                         jh->b_committed_data = NULL;
 489                         jbd_unlock_bh_state(bh);
 490                 }
 491                 jbd2_journal_refile_buffer(journal, jh);
 492         }
 493
 494         write_unlock(&journal->j_state_lock);
 495         /*
 496          * Now try to drop any written-back buffers from the journal's
 497          * checkpoint lists.  We do this *before* commit because it potentially
 498          * frees some memory
 499          */
 500         spin_lock(&journal->j_list_lock);
 501         __jbd2_journal_clean_checkpoint_list(journal, false);
 502         spin_unlock(&journal->j_list_lock);
 503
 504         jbd_debug(3, "JBD2: commit phase 1\n");
 505
 506         /*
 507          * Clear revoked flag to reflect there is no revoked buffers
 508          * in the next transaction which is going to be started.
 509          */
 510         jbd2_clear_buffer_revoked_flags(journal);
 511
 512         /*
 513          * Switch to a new revoke table.
 514          */
 515         jbd2_journal_switch_revoke_table(journal);
 516
 517         write_lock(&journal->j_state_lock);
 518         /*
 519          * Reserved credits cannot be claimed anymore, free them
 520          */
 521         atomic_sub(atomic_read(&journal->j_reserved_credits),
 522                    &commit_transaction->t_outstanding_credits);
 523
 524         trace_jbd2_commit_flushing(journal, commit_transaction);
 525         stats.run.rs_flushing = jiffies;
 526         stats.run.rs_locked = jbd2_time_diff(stats.run.rs_locked,
 527                                              stats.run.rs_flushing);
 528
 529         commit_transaction->t_state = T_FLUSH;
 530         journal->j_committing_transaction = commit_transaction;
 531         journal->j_running_transaction = NULL;
 532         start_time = ktime_get();
 533         commit_transaction->t_log_start = journal->j_head;
 534         wake_up_all(&journal->j_wait_transaction_locked);
 535         write_unlock(&journal->j_state_lock);
 536
 537         jbd_debug(3, "JBD2: commit phase 2a\n");
 538
 539         /*
 540          * Now start flushing things to disk, in the order they appear
 541          * on the transaction lists.  Data blocks go first.
 542          */
 543         err = journal_submit_data_buffers(journal, commit_transaction);
 544         if (err)
 545                 jbd2_journal_abort(journal, err);
 546
 547         blk_start_plug(&plug);
 548         jbd2_journal_write_revoke_records(commit_transaction, &log_bufs);
 549
 550         jbd_debug(3, "JBD2: commit phase 2b\n");
 551
 552         /*
 553          * Way to go: we have now written out all of the data for a
 554          * transaction!  Now comes the tricky part: we need to write out
 555          * metadata.  Loop over the transaction's entire buffer list:
 556          */
 557         write_lock(&journal->j_state_lock);
 558         commit_transaction->t_state = T_COMMIT;
 559         write_unlock(&journal->j_state_lock);
 560
 561         trace_jbd2_commit_logging(journal, commit_transaction);
 562         stats.run.rs_logging = jiffies;
 563         stats.run.rs_flushing = jbd2_time_diff(stats.run.rs_flushing,
 564                                                stats.run.rs_logging);
 565         stats.run.rs_blocks =
 566                 atomic_read(&commit_transaction->t_outstanding_credits);
 567         stats.run.rs_blocks_logged = 0;
 568
 569         J_ASSERT(commit_transaction->t_nr_buffers <=
 570                  atomic_read(&commit_transaction->t_outstanding_credits));
 571
 572         err = 0;
 573         bufs = 0;
 574         descriptor = NULL;
 575         while (commit_transaction->t_buffers) {
 576
 577                 /* Find the next buffer to be journaled... */
 578
 579                 jh = commit_transaction->t_buffers;
 580
 581                 /* If we're in abort mode, we just un-journal the buffer and
 582                    release it. */
 583
 584                 if (is_journal_aborted(journal)) {
 585                         clear_buffer_jbddirty(jh2bh(jh));
 586                         JBUFFER_TRACE(jh, "journal is aborting: refile");
 587                         jbd2_buffer_abort_trigger(jh,
 588                                                   jh->b_frozen_data ?
 589                                                   jh->b_frozen_triggers :
 590                                                   jh->b_triggers);
 591                         jbd2_journal_refile_buffer(journal, jh);
 592                         /* If that was the last one, we need to clean up
 593                          * any descriptor buffers which may have been
 594                          * already allocated, even if we are now
 595                          * aborting. */
 596                         if (!commit_transaction->t_buffers)
 597                                 goto start_journal_io;
 598                         continue;
 599                 }
 600
 601                 /* Make sure we have a descriptor block in which to
 602                    record the metadata buffer. */
 603
 604                 if (!descriptor) {
 605                         J_ASSERT (bufs == 0);
 606
 607                         jbd_debug(4, "JBD2: get descriptor\n");
 608
 609                         descriptor = jbd2_journal_get_descriptor_buffer(
 610                                                         commit_transaction,
 611                                                         JBD2_DESCRIPTOR_BLOCK);
 612                         if (!descriptor) {
 613                                 jbd2_journal_abort(journal, -EIO);
 614                                 continue;
 615                         }
 616
 617                         jbd_debug(4, "JBD2: got buffer %llu (%p)\n",
 618                                 (unsigned long long)descriptor->b_blocknr,
 619                                 descriptor->b_data);
 620                         tagp = &descriptor->b_data[sizeof(journal_header_t)];
 621                         space_left = descriptor->b_size -
 622                                                 sizeof(journal_header_t);
 623                         first_tag = 1;
 624                         set_buffer_jwrite(descriptor);
 625                         set_buffer_dirty(descriptor);
 626                         wbuf[bufs++] = descriptor;
 627
 628                         /* Record it so that we can wait for IO
 629                            completion later */
 630                         BUFFER_TRACE(descriptor, "ph3: file as descriptor");
 631                         jbd2_file_log_bh(&log_bufs, descriptor);
 632                 }
 633
 634                 /* Where is the buffer to be written? */
 635
 636                 err = jbd2_journal_next_log_block(journal, &blocknr);
 637                 /* If the block mapping failed, just abandon the buffer
 638                    and repeat this loop: we'll fall into the
 639                    refile-on-abort condition above. */
 640                 if (err) {
 641                         jbd2_journal_abort(journal, err);
 642                         continue;
 643                 }
 644
 645                 /*
 646                  * start_this_handle() uses t_outstanding_credits to determine
 647                  * the free space in the log, but this counter is changed
 648                  * by jbd2_journal_next_log_block() also.
 649                  */
 650                 atomic_dec(&commit_transaction->t_outstanding_credits);
 651
 652                 /* Bump b_count to prevent truncate from stumbling over
 653                    the shadowed buffer!  @@@ This can go if we ever get
 654                    rid of the shadow pairing of buffers. */
 655                 atomic_inc(&jh2bh(jh)->b_count);
 656
 657                 /*
 658                  * Make a temporary IO buffer with which to write it out
 659                  * (this will requeue the metadata buffer to BJ_Shadow).
 660                  */
 661                 set_bit(BH_JWrite, &jh2bh(jh)->b_state);
 662                 JBUFFER_TRACE(jh, "ph3: write metadata");
 663                 flags = jbd2_journal_write_metadata_buffer(commit_transaction,
 664                                                 jh, &wbuf[bufs], blocknr);
 665                 if (flags < 0) {
 666                         jbd2_journal_abort(journal, flags);
 667                         continue;
 668                 }
 669                 jbd2_file_log_bh(&io_bufs, wbuf[bufs]);
 670
 671                 /* Record the new block's tag in the current descriptor
 672                    buffer */
 673
 674                 tag_flag = 0;
 675                 if (flags & 1)
 676                         tag_flag |= JBD2_FLAG_ESCAPE;
 677                 if (!first_tag)
 678                         tag_flag |= JBD2_FLAG_SAME_UUID;
 679
 680                 tag = (journal_block_tag_t *) tagp;
 681                 write_tag_block(journal, tag, jh2bh(jh)->b_blocknr);
 682                 tag->t_flags = cpu_to_be16(tag_flag);
 683                 jbd2_block_tag_csum_set(journal, tag, wbuf[bufs],
 684                                         commit_transaction->t_tid);
 685                 tagp += tag_bytes;
 686                 space_left -= tag_bytes;
 687                 bufs++;
 688
 689                 if (first_tag) {
 690                         memcpy (tagp, journal->j_uuid, 16);
 691                         tagp += 16;
 692                         space_left -= 16;
 693                         first_tag = 0;
 694                 }
 695
 696                 /* If there's no more to do, or if the descriptor is full,
 697                    let the IO rip! */
 698
 699                 if (bufs == journal->j_wbufsize ||
 700                     commit_transaction->t_buffers == NULL ||
 701                     space_left < tag_bytes + 16 + csum_size) {
 702
 703                         jbd_debug(4, "JBD2: Submit %d IOs\n", bufs);
 704
 705                         /* Write an end-of-descriptor marker before
 706                            submitting the IOs.  "tag" still points to
 707                            the last tag we set up. */
 708
 709                         tag->t_flags |= cpu_to_be16(JBD2_FLAG_LAST_TAG);
 710 start_journal_io:
 711                         if (descriptor)
 712                                 jbd2_descriptor_block_csum_set(journal,
 713                                                         descriptor);
 714
 715                         for (i = 0; i < bufs; i++) {
 716                                 struct buffer_head *bh = wbuf[i];
 717                                 /*
 718                                  * Compute checksum.
 719                                  */
 720                                 if (jbd2_has_feature_checksum(journal)) {
 721                                         crc32_sum =
 722                                             jbd2_checksum_data(crc32_sum, bh);
 723                                 }
 724
 725                                 lock_buffer(bh);
 726                                 clear_buffer_dirty(bh);
 727                                 set_buffer_uptodate(bh);
 728                                 bh->b_end_io = journal_end_buffer_io_sync;
 729                                 submit_bh(REQ_OP_WRITE, REQ_SYNC, bh);
 730                         }
 731                         cond_resched();
 732
 733                         /* Force a new descriptor to be generated next
 734                            time round the loop. */
 735                         descriptor = NULL;
 736                         bufs = 0;
 737                 }
 738         }
 739
 740         err = journal_finish_inode_data_buffers(journal, commit_transaction);
 741         if (err) {
 742                 printk(KERN_WARNING
 743                         "JBD2: Detected IO errors while flushing file data "
 744                        "on %s\n", journal->j_devname);
 745                 if (journal->j_flags & JBD2_ABORT_ON_SYNCDATA_ERR)
 746                         jbd2_journal_abort(journal, err);
 747                 err = 0;
 748         }
 749
 750         /*
 751          * Get current oldest transaction in the log before we issue flush
 752          * to the filesystem device. After the flush we can be sure that
 753          * blocks of all older transactions are checkpointed to persistent
 754          * storage and we will be safe to update journal start in the
 755          * superblock with the numbers we get here.
 756          */
 757         update_tail =
 758                 jbd2_journal_get_log_tail(journal, &first_tid, &first_block);
 759
 760         write_lock(&journal->j_state_lock);
 761         if (update_tail) {
 762                 long freed = first_block - journal->j_tail;
 763
 764                 if (first_block < journal->j_tail)
 765                         freed += journal->j_last - journal->j_first;
 766                 /* Update tail only if we free significant amount of space */
 767                 if (freed < journal->j_maxlen / 4)
 768                         update_tail = 0;
 769         }
 770         J_ASSERT(commit_transaction->t_state == T_COMMIT);
 771         commit_transaction->t_state = T_COMMIT_DFLUSH;
 772         write_unlock(&journal->j_state_lock);
 773
 774         /*
 775          * If the journal is not located on the file system device,
 776          * then we must flush the file system device before we issue
 777          * the commit record
 778          */
 779         if (commit_transaction->t_need_data_flush &&
 780             (journal->j_fs_dev != journal->j_dev) &&
 781             (journal->j_flags & JBD2_BARRIER))
 782                 blkdev_issue_flush(journal->j_fs_dev, GFP_NOFS, NULL);
 783
 784         /* Done it all: now write the commit record asynchronously. */
 785         if (jbd2_has_feature_async_commit(journal)) {
 786                 err = journal_submit_commit_record(journal, commit_transaction,
 787                                                  &cbh, crc32_sum);
 788                 if (err)
 789                         jbd2_journal_abort(journal, err);
 790         }
 791
 792         blk_finish_plug(&plug);
 793
 794         /* Lo and behold: we have just managed to send a transaction to
 795            the log.  Before we can commit it, wait for the IO so far to
 796            complete.  Control buffers being written are on the
 797            transaction's t_log_list queue, and metadata buffers are on
 798            the io_bufs list.
 799
 800            Wait for the buffers in reverse order.  That way we are
 801            less likely to be woken up until all IOs have completed, and
 802            so we incur less scheduling load.
 803         */
 804
 805         jbd_debug(3, "JBD2: commit phase 3\n");
 806
 807         while (!list_empty(&io_bufs)) {
 808                 struct buffer_head *bh = list_entry(io_bufs.prev,
 809                                                     struct buffer_head,
 810                                                     b_assoc_buffers);
 811
 812                 wait_on_buffer(bh);
 813                 cond_resched();
 814
 815                 if (unlikely(!buffer_uptodate(bh)))
 816                         err = -EIO;
 817                 jbd2_unfile_log_bh(bh);
 818                 stats.run.rs_blocks_logged++;
 819
 820                 /*
 821                  * The list contains temporary buffer heads created by
 822                  * jbd2_journal_write_metadata_buffer().
 823                  */
 824                 BUFFER_TRACE(bh, "dumping temporary bh");
 825                 __brelse(bh);
 826                 J_ASSERT_BH(bh, atomic_read(&bh->b_count) == 0);
 827                 free_buffer_head(bh);
 828
 829                 /* We also have to refile the corresponding shadowed buffer */
 830                 jh = commit_transaction->t_shadow_list->b_tprev;
 831                 bh = jh2bh(jh);
 832                 clear_buffer_jwrite(bh);
 833                 J_ASSERT_BH(bh, buffer_jbddirty(bh));
 834                 J_ASSERT_BH(bh, !buffer_shadow(bh));
 835
 836                 /* The metadata is now released for reuse, but we need
 837                    to remember it against this transaction so that when
 838                    we finally commit, we can do any checkpointing
 839                    required. */
 840                 JBUFFER_TRACE(jh, "file as BJ_Forget");
 841                 jbd2_journal_file_buffer(jh, commit_transaction, BJ_Forget);
 842                 JBUFFER_TRACE(jh, "brelse shadowed buffer");
 843                 __brelse(bh);
 844         }
 845
 846         J_ASSERT (commit_transaction->t_shadow_list == NULL);
 847
 848         jbd_debug(3, "JBD2: commit phase 4\n");
 849
 850         /* Here we wait for the revoke record and descriptor record buffers */
 851         while (!list_empty(&log_bufs)) {
 852                 struct buffer_head *bh;
 853
 854                 bh = list_entry(log_bufs.prev, struct buffer_head, b_assoc_buffers);
 855                 wait_on_buffer(bh);
 856                 cond_resched();
 857
 858                 if (unlikely(!buffer_uptodate(bh)))
 859                         err = -EIO;
 860
 861                 BUFFER_TRACE(bh, "ph5: control buffer writeout done: unfile");
 862                 clear_buffer_jwrite(bh);
 863                 jbd2_unfile_log_bh(bh);
 864                 stats.run.rs_blocks_logged++;
 865                 __brelse(bh);           /* One for getblk */
 866                 /* AKPM: bforget here */
 867         }
 868
 869         if (err)
 870                 jbd2_journal_abort(journal, err);
 871
 872         jbd_debug(3, "JBD2: commit phase 5\n");
 873         write_lock(&journal->j_state_lock);
 874         J_ASSERT(commit_transaction->t_state == T_COMMIT_DFLUSH);
 875         commit_transaction->t_state = T_COMMIT_JFLUSH;
 876         write_unlock(&journal->j_state_lock);
 877
 878         if (!jbd2_has_feature_async_commit(journal)) {
 879                 err = journal_submit_commit_record(journal, commit_transaction,
 880                                                 &cbh, crc32_sum);
 881                 if (err)
 882                         jbd2_journal_abort(journal, err);
 883         }
 884         if (cbh)
 885                 err = journal_wait_on_commit_record(journal, cbh);
 886         stats.run.rs_blocks_logged++;
 887         if (jbd2_has_feature_async_commit(journal) &&
 888             journal->j_flags & JBD2_BARRIER) {
 889                 blkdev_issue_flush(journal->j_dev, GFP_NOFS, NULL);
 890         }
 891
 892         if (err)
 893                 jbd2_journal_abort(journal, err);
 894
 895         /*
 896          * Now disk caches for filesystem device are flushed so we are safe to
 897          * erase checkpointed transactions from the log by updating journal
 898          * superblock.
 899          */
 900         if (update_tail)
 901                 jbd2_update_log_tail(journal, first_tid, first_block);
 902
 903         /* End of a transaction!  Finally, we can do checkpoint
 904            processing: any buffers committed as a result of this
 905            transaction can be removed from any checkpoint list it was on
 906            before. */
 907
 908         jbd_debug(3, "JBD2: commit phase 6\n");
 909
 910         J_ASSERT(list_empty(&commit_transaction->t_inode_list));
 911         J_ASSERT(commit_transaction->t_buffers == NULL);
 912         J_ASSERT(commit_transaction->t_checkpoint_list == NULL);
 913         J_ASSERT(commit_transaction->t_shadow_list == NULL);
 914
 915 restart_loop:
 916         /*
 917          * As there are other places (journal_unmap_buffer()) adding buffers
 918          * to this list we have to be careful and hold the j_list_lock.
 919          */
 920         spin_lock(&journal->j_list_lock);
 921         while (commit_transaction->t_forget) {
 922                 transaction_t *cp_transaction;
 923                 struct buffer_head *bh;
 924                 int try_to_free = 0;
 925
 926                 jh = commit_transaction->t_forget;
 927                 spin_unlock(&journal->j_list_lock);
 928                 bh = jh2bh(jh);
 929                 /*
 930                  * Get a reference so that bh cannot be freed before we are
 931                  * done with it.
 932                  */
 933                 get_bh(bh);
 934                 jbd_lock_bh_state(bh);
 935                 J_ASSERT_JH(jh, jh->b_transaction == commit_transaction);
 936
 937                 /*
 938                  * If there is undo-protected committed data against
 939                  * this buffer, then we can remove it now.  If it is a
 940                  * buffer needing such protection, the old frozen_data
 941                  * field now points to a committed version of the
 942                  * buffer, so rotate that field to the new committed
 943                  * data.
 944                  *
 945                  * Otherwise, we can just throw away the frozen data now.
 946                  *
 947                  * We also know that the frozen data has already fired
 948                  * its triggers if they exist, so we can clear that too.
 949                  */
 950                 if (jh->b_committed_data) {
 951                         jbd2_free(jh->b_committed_data, bh->b_size);
 952                         jh->b_committed_data = NULL;
 953                         if (jh->b_frozen_data) {
 954                                 jh->b_committed_data = jh->b_frozen_data;
 955                                 jh->b_frozen_data = NULL;
 956                                 jh->b_frozen_triggers = NULL;
 957                         }
 958                 } else if (jh->b_frozen_data) {
 959                         jbd2_free(jh->b_frozen_data, bh->b_size);
 960                         jh->b_frozen_data = NULL;
 961                         jh->b_frozen_triggers = NULL;
 962                 }
 963
 964                 spin_lock(&journal->j_list_lock);
 965                 cp_transaction = jh->b_cp_transaction;
 966                 if (cp_transaction) {
 967                         JBUFFER_TRACE(jh, "remove from old cp transaction");
 968                         cp_transaction->t_chp_stats.cs_dropped++;
 969                         __jbd2_journal_remove_checkpoint(jh);
 970                 }
 971
 972                 /* Only re-checkpoint the buffer_head if it is marked
 973                  * dirty.  If the buffer was added to the BJ_Forget list
 974                  * by jbd2_journal_forget, it may no longer be dirty and
 975                  * there's no point in keeping a checkpoint record for
 976                  * it. */
 977
 978                 /*
 979                  * A buffer which has been freed while still being journaled
 980                  * by a previous transaction, refile the buffer to BJ_Forget of
 981                  * the running transaction. If the just committed transaction
 982                  * contains "add to orphan" operation, we can completely
 983                  * invalidate the buffer now. We are rather through in that
 984                  * since the buffer may be still accessible when blocksize <
 985                  * pagesize and it is attached to the last partial page.
 986                  */
 987                 if (buffer_freed(bh) && !jh->b_next_transaction) {
 988                         struct address_space *mapping;
 989
 990                         clear_buffer_freed(bh);
 991                         clear_buffer_jbddirty(bh);
 992
 993                         /*
 994                          * Block device buffers need to stay mapped all the
 995                          * time, so it is enough to clear buffer_jbddirty and
 996                          * buffer_freed bits. For the file mapping buffers (i.e.
 997                          * journalled data) we need to unmap buffer and clear
 998                          * more bits. We also need to be careful about the check
 999                          * because the data page mapping can get cleared under
1000                          * our hands. Note that if mapping == NULL, we don't
1001                          * need to make buffer unmapped because the page is
1002                          * already detached from the mapping and buffers cannot
1003                          * get reused.
1004                          */
1005                         mapping = READ_ONCE(bh->b_page->mapping);
1006                         if (mapping && !sb_is_blkdev_sb(mapping->host->i_sb)) {
1007                                 clear_buffer_mapped(bh);
1008                                 clear_buffer_new(bh);
1009                                 clear_buffer_req(bh);
1010                                 bh->b_bdev = NULL;
1011                         }
1012                 }
1013
1014                 if (buffer_jbddirty(bh)) {
1015                         JBUFFER_TRACE(jh, "add to new checkpointing trans");
1016                         __jbd2_journal_insert_checkpoint(jh, commit_transaction);
1017                         if (is_journal_aborted(journal))
1018                                 clear_buffer_jbddirty(bh);
1019                 } else {
1020                         J_ASSERT_BH(bh, !buffer_dirty(bh));
1021                         /*
1022                          * The buffer on BJ_Forget list and not jbddirty means
1023                          * it has been freed by this transaction and hence it
1024                          * could not have been reallocated until this
1025                          * transaction has committed. *BUT* it could be
1026                          * reallocated once we have written all the data to
1027                          * disk and before we process the buffer on BJ_Forget
1028                          * list.
1029                          */
1030                         if (!jh->b_next_transaction)
1031                                 try_to_free = 1;
1032                 }
1033                 JBUFFER_TRACE(jh, "refile or unfile buffer");
1034                 __jbd2_journal_refile_buffer(jh);
1035                 jbd_unlock_bh_state(bh);
1036                 if (try_to_free)
1037                         release_buffer_page(bh);        /* Drops bh reference */
1038                 else
1039                         __brelse(bh);
1040                 cond_resched_lock(&journal->j_list_lock);
1041         }
1042         spin_unlock(&journal->j_list_lock);
1043         /*
1044          * This is a bit sleazy.  We use j_list_lock to protect transition
1045          * of a transaction into T_FINISHED state and calling
1046          * __jbd2_journal_drop_transaction(). Otherwise we could race with
1047          * other checkpointing code processing the transaction...
1048          */
1049         write_lock(&journal->j_state_lock);
1050         spin_lock(&journal->j_list_lock);
1051         /*
1052          * Now recheck if some buffers did not get attached to the transaction
1053          * while the lock was dropped...
1054          */
1055         if (commit_transaction->t_forget) {
1056                 spin_unlock(&journal->j_list_lock);
1057                 write_unlock(&journal->j_state_lock);
1058                 goto restart_loop;
1059         }
1060
1061         /* Add the transaction to the checkpoint list
1062          * __journal_remove_checkpoint() can not destroy transaction
1063          * under us because it is not marked as T_FINISHED yet */
1064         if (journal->j_checkpoint_transactions == NULL) {
1065                 journal->j_checkpoint_transactions = commit_transaction;
1066                 commit_transaction->t_cpnext = commit_transaction;
1067                 commit_transaction->t_cpprev = commit_transaction;
1068         } else {
1069                 commit_transaction->t_cpnext =
1070                         journal->j_checkpoint_transactions;
1071                 commit_transaction->t_cpprev =
1072                         commit_transaction->t_cpnext->t_cpprev;
1073                 commit_transaction->t_cpnext->t_cpprev =
1074                         commit_transaction;
1075                 commit_transaction->t_cpprev->t_cpnext =
1076                                 commit_transaction;
1077         }
1078         spin_unlock(&journal->j_list_lock);
1079
1080         /* Done with this transaction! */
1081
1082         jbd_debug(3, "JBD2: commit phase 7\n");
1083
1084         J_ASSERT(commit_transaction->t_state == T_COMMIT_JFLUSH);
1085
1086         commit_transaction->t_start = jiffies;
1087         stats.run.rs_logging = jbd2_time_diff(stats.run.rs_logging,
1088                                               commit_transaction->t_start);
1089
1090         /*
1091          * File the transaction statistics
1092          */
1093         stats.ts_tid = commit_transaction->t_tid;
1094         stats.run.rs_handle_count =
1095                 atomic_read(&commit_transaction->t_handle_count);
1096         trace_jbd2_run_stats(journal->j_fs_dev->bd_dev,
1097                              commit_transaction->t_tid, &stats.run);
1098         stats.ts_requested = (commit_transaction->t_requested) ? 1 : 0;
1099
1100         commit_transaction->t_state = T_COMMIT_CALLBACK;
1101         J_ASSERT(commit_transaction == journal->j_committing_transaction);
1102         journal->j_commit_sequence = commit_transaction->t_tid;
1103         journal->j_committing_transaction = NULL;
1104         commit_time = ktime_to_ns(ktime_sub(ktime_get(), start_time));
1105
1106         /*
1107          * weight the commit time higher than the average time so we don't
1108          * react too strongly to vast changes in the commit time
1109          */
1110         if (likely(journal->j_average_commit_time))
1111                 journal->j_average_commit_time = (commit_time +
1112                                 journal->j_average_commit_time*3) / 4;
1113         else
1114                 journal->j_average_commit_time = commit_time;
1115
1116         write_unlock(&journal->j_state_lock);
1117
1118         if (journal->j_commit_callback)
1119                 journal->j_commit_callback(journal, commit_transaction);
1120
1121         trace_jbd2_end_commit(journal, commit_transaction);
1122         jbd_debug(1, "JBD2: commit %d complete, head %d\n",
1123                   journal->j_commit_sequence, journal->j_tail_sequence);
1124
1125         write_lock(&journal->j_state_lock);
1126         spin_lock(&journal->j_list_lock);
1127         commit_transaction->t_state = T_FINISHED;
1128         /* Check if the transaction can be dropped now that we are finished */
1129         if (commit_transaction->t_checkpoint_list == NULL &&
1130             commit_transaction->t_checkpoint_io_list == NULL) {
1131                 __jbd2_journal_drop_transaction(journal, commit_transaction);
1132                 jbd2_journal_free_transaction(commit_transaction);
1133         }
1134         spin_unlock(&journal->j_list_lock);
1135         write_unlock(&journal->j_state_lock);
1136         wake_up(&journal->j_wait_done_commit);
1137
1138         /*
1139          * Calculate overall stats
1140          */
1141         spin_lock(&journal->j_history_lock);
1142         journal->j_stats.ts_tid++;
1143         journal->j_stats.ts_requested += stats.ts_requested;
1144         journal->j_stats.run.rs_wait += stats.run.rs_wait;
1145         journal->j_stats.run.rs_request_delay += stats.run.rs_request_delay;
1146         journal->j_stats.run.rs_running += stats.run.rs_running;
1147         journal->j_stats.run.rs_locked += stats.run.rs_locked;
1148         journal->j_stats.run.rs_flushing += stats.run.rs_flushing;
1149         journal->j_stats.run.rs_logging += stats.run.rs_logging;
1150         journal->j_stats.run.rs_handle_count += stats.run.rs_handle_count;
1151         journal->j_stats.run.rs_blocks += stats.run.rs_blocks;
1152         journal->j_stats.run.rs_blocks_logged += stats.run.rs_blocks_logged;
1153         spin_unlock(&journal->j_history_lock);
1154 }