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 folio *folio;
  66
  67         if (buffer_dirty(bh))
  68                 goto nope;
  69         if (atomic_read(&bh->b_count) != 1)
  70                 goto nope;
  71         folio = bh->b_folio;
  72         if (folio->mapping)
  73                 goto nope;
  74
  75         /* OK, it's a truncated page */
  76         if (!folio_trylock(folio))
  77                 goto nope;
  78
  79         folio_get(folio);
  80         __brelse(bh);
  81         try_to_free_buffers(folio);
  82         folio_unlock(folio);
  83         folio_put(folio);
  84         return;
  85
  86 nope:
  87         __brelse(bh);
  88 }
  89
  90 static void jbd2_commit_block_csum_set(journal_t *j, struct buffer_head *bh)
  91 {
  92         struct commit_header *h;
  93         __u32 csum;
  94
  95         if (!jbd2_journal_has_csum_v2or3(j))
  96                 return;
  97
  98         h = (struct commit_header *)(bh->b_data);
  99         h->h_chksum_type = 0;
 100         h->h_chksum_size = 0;
 101         h->h_chksum[0] = 0;
 102         csum = jbd2_chksum(j, j->j_csum_seed, bh->b_data, j->j_blocksize);
 103         h->h_chksum[0] = cpu_to_be32(csum);
 104 }
 105
 106 /*
 107  * Done it all: now submit the commit record.  We should have
 108  * cleaned up our previous buffers by now, so if we are in abort
 109  * mode we can now just skip the rest of the journal write
 110  * entirely.
 111  *
 112  * Returns 1 if the journal needs to be aborted or 0 on success
 113  */
 114 static int journal_submit_commit_record(journal_t *journal,
 115                                         transaction_t *commit_transaction,
 116                                         struct buffer_head **cbh,
 117                                         __u32 crc32_sum)
 118 {
 119         struct commit_header *tmp;
 120         struct buffer_head *bh;
 121         struct timespec64 now;
 122         blk_opf_t write_flags = REQ_OP_WRITE | REQ_SYNC;
 123
 124         *cbh = NULL;
 125
 126         if (is_journal_aborted(journal))
 127                 return 0;
 128
 129         bh = jbd2_journal_get_descriptor_buffer(commit_transaction,
 130                                                 JBD2_COMMIT_BLOCK);
 131         if (!bh)
 132                 return 1;
 133
 134         tmp = (struct commit_header *)bh->b_data;
 135         ktime_get_coarse_real_ts64(&now);
 136         tmp->h_commit_sec = cpu_to_be64(now.tv_sec);
 137         tmp->h_commit_nsec = cpu_to_be32(now.tv_nsec);
 138
 139         if (jbd2_has_feature_checksum(journal)) {
 140                 tmp->h_chksum_type      = JBD2_CRC32_CHKSUM;
 141                 tmp->h_chksum_size      = JBD2_CRC32_CHKSUM_SIZE;
 142                 tmp->h_chksum[0]        = cpu_to_be32(crc32_sum);
 143         }
 144         jbd2_commit_block_csum_set(journal, bh);
 145
 146         BUFFER_TRACE(bh, "submit commit block");
 147         lock_buffer(bh);
 148         clear_buffer_dirty(bh);
 149         set_buffer_uptodate(bh);
 150         bh->b_end_io = journal_end_buffer_io_sync;
 151
 152         if (journal->j_flags & JBD2_BARRIER &&
 153             !jbd2_has_feature_async_commit(journal))
 154                 write_flags |= REQ_PREFLUSH | REQ_FUA;
 155
 156         submit_bh(write_flags, bh);
 157         *cbh = bh;
 158         return 0;
 159 }
 160
 161 /*
 162  * This function along with journal_submit_commit_record
 163  * allows to write the commit record asynchronously.
 164  */
 165 static int journal_wait_on_commit_record(journal_t *journal,
 166                                          struct buffer_head *bh)
 167 {
 168         int ret = 0;
 169
 170         clear_buffer_dirty(bh);
 171         wait_on_buffer(bh);
 172
 173         if (unlikely(!buffer_uptodate(bh)))
 174                 ret = -EIO;
 175         put_bh(bh);            /* One for getblk() */
 176
 177         return ret;
 178 }
 179
 180 /* Send all the data buffers related to an inode */
 181 int jbd2_submit_inode_data(journal_t *journal, struct jbd2_inode *jinode)
 182 {
 183         if (!jinode || !(jinode->i_flags & JI_WRITE_DATA))
 184                 return 0;
 185
 186         trace_jbd2_submit_inode_data(jinode->i_vfs_inode);
 187         return journal->j_submit_inode_data_buffers(jinode);
 188
 189 }
 190 EXPORT_SYMBOL(jbd2_submit_inode_data);
 191
 192 int jbd2_wait_inode_data(journal_t *journal, struct jbd2_inode *jinode)
 193 {
 194         if (!jinode || !(jinode->i_flags & JI_WAIT_DATA) ||
 195                 !jinode->i_vfs_inode || !jinode->i_vfs_inode->i_mapping)
 196                 return 0;
 197         return filemap_fdatawait_range_keep_errors(
 198                 jinode->i_vfs_inode->i_mapping, jinode->i_dirty_start,
 199                 jinode->i_dirty_end);
 200 }
 201 EXPORT_SYMBOL(jbd2_wait_inode_data);
 202
 203 /*
 204  * Submit all the data buffers of inode associated with the transaction to
 205  * disk.
 206  *
 207  * We are in a committing transaction. Therefore no new inode can be added to
 208  * our inode list. We use JI_COMMIT_RUNNING flag to protect inode we currently
 209  * operate on from being released while we write out pages.
 210  */
 211 static int journal_submit_data_buffers(journal_t *journal,
 212                 transaction_t *commit_transaction)
 213 {
 214         struct jbd2_inode *jinode;
 215         int err, ret = 0;
 216
 217         spin_lock(&journal->j_list_lock);
 218         list_for_each_entry(jinode, &commit_transaction->t_inode_list, i_list) {
 219                 if (!(jinode->i_flags & JI_WRITE_DATA))
 220                         continue;
 221                 jinode->i_flags |= JI_COMMIT_RUNNING;
 222                 spin_unlock(&journal->j_list_lock);
 223                 /* submit the inode data buffers. */
 224                 trace_jbd2_submit_inode_data(jinode->i_vfs_inode);
 225                 if (journal->j_submit_inode_data_buffers) {
 226                         err = journal->j_submit_inode_data_buffers(jinode);
 227                         if (!ret)
 228                                 ret = err;
 229                 }
 230                 spin_lock(&journal->j_list_lock);
 231                 J_ASSERT(jinode->i_transaction == commit_transaction);
 232                 jinode->i_flags &= ~JI_COMMIT_RUNNING;
 233                 smp_mb();
 234                 wake_up_bit(&jinode->i_flags, __JI_COMMIT_RUNNING);
 235         }
 236         spin_unlock(&journal->j_list_lock);
 237         return ret;
 238 }
 239
 240 int jbd2_journal_finish_inode_data_buffers(struct jbd2_inode *jinode)
 241 {
 242         struct address_space *mapping = jinode->i_vfs_inode->i_mapping;
 243
 244         return filemap_fdatawait_range_keep_errors(mapping,
 245                                                    jinode->i_dirty_start,
 246                                                    jinode->i_dirty_end);
 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                 if (!(jinode->i_flags & JI_WAIT_DATA))
 264                         continue;
 265                 jinode->i_flags |= JI_COMMIT_RUNNING;
 266                 spin_unlock(&journal->j_list_lock);
 267                 /* wait for the inode data buffers writeout. */
 268                 if (journal->j_finish_inode_data_buffers) {
 269                         err = journal->j_finish_inode_data_buffers(jinode);
 270                         if (!ret)
 271                                 ret = err;
 272                 }
 273                 spin_lock(&journal->j_list_lock);
 274                 jinode->i_flags &= ~JI_COMMIT_RUNNING;
 275                 smp_mb();
 276                 wake_up_bit(&jinode->i_flags, __JI_COMMIT_RUNNING);
 277         }
 278
 279         /* Now refile inode to proper lists */
 280         list_for_each_entry_safe(jinode, next_i,
 281                                  &commit_transaction->t_inode_list, i_list) {
 282                 list_del(&jinode->i_list);
 283                 if (jinode->i_next_transaction) {
 284                         jinode->i_transaction = jinode->i_next_transaction;
 285                         jinode->i_next_transaction = NULL;
 286                         list_add(&jinode->i_list,
 287                                 &jinode->i_transaction->t_inode_list);
 288                 } else {
 289                         jinode->i_transaction = NULL;
 290                         jinode->i_dirty_start = 0;
 291                         jinode->i_dirty_end = 0;
 292                 }
 293         }
 294         spin_unlock(&journal->j_list_lock);
 295
 296         return ret;
 297 }
 298
 299 static __u32 jbd2_checksum_data(__u32 crc32_sum, struct buffer_head *bh)
 300 {
 301         struct page *page = bh->b_page;
 302         char *addr;
 303         __u32 checksum;
 304
 305         addr = kmap_atomic(page);
 306         checksum = crc32_be(crc32_sum,
 307                 (void *)(addr + offset_in_page(bh->b_data)), bh->b_size);
 308         kunmap_atomic(addr);
 309
 310         return checksum;
 311 }
 312
 313 static void write_tag_block(journal_t *j, journal_block_tag_t *tag,
 314                                    unsigned long long block)
 315 {
 316         tag->t_blocknr = cpu_to_be32(block & (u32)~0);
 317         if (jbd2_has_feature_64bit(j))
 318                 tag->t_blocknr_high = cpu_to_be32((block >> 31) >> 1);
 319 }
 320
 321 static void jbd2_block_tag_csum_set(journal_t *j, journal_block_tag_t *tag,
 322                                     struct buffer_head *bh, __u32 sequence)
 323 {
 324         journal_block_tag3_t *tag3 = (journal_block_tag3_t *)tag;
 325         struct page *page = bh->b_page;
 326         __u8 *addr;
 327         __u32 csum32;
 328         __be32 seq;
 329
 330         if (!jbd2_journal_has_csum_v2or3(j))
 331                 return;
 332
 333         seq = cpu_to_be32(sequence);
 334         addr = kmap_atomic(page);
 335         csum32 = jbd2_chksum(j, j->j_csum_seed, (__u8 *)&seq, sizeof(seq));
 336         csum32 = jbd2_chksum(j, csum32, addr + offset_in_page(bh->b_data),
 337                              bh->b_size);
 338         kunmap_atomic(addr);
 339
 340         if (jbd2_has_feature_csum3(j))
 341                 tag3->t_checksum = cpu_to_be32(csum32);
 342         else
 343                 tag->t_checksum = cpu_to_be16(csum32);
 344 }
 345 /*
 346  * jbd2_journal_commit_transaction
 347  *
 348  * The primary function for committing a transaction to the log.  This
 349  * function is called by the journal thread to begin a complete commit.
 350  */
 351 void jbd2_journal_commit_transaction(journal_t *journal)
 352 {
 353         struct transaction_stats_s stats;
 354         transaction_t *commit_transaction;
 355         struct journal_head *jh;
 356         struct buffer_head *descriptor;
 357         struct buffer_head **wbuf = journal->j_wbuf;
 358         int bufs;
 359         int flags;
 360         int err;
 361         unsigned long long blocknr;
 362         ktime_t start_time;
 363         u64 commit_time;
 364         char *tagp = NULL;
 365         journal_block_tag_t *tag = NULL;
 366         int space_left = 0;
 367         int first_tag = 0;
 368         int tag_flag;
 369         int i;
 370         int tag_bytes = journal_tag_bytes(journal);
 371         struct buffer_head *cbh = NULL; /* For transactional checksums */
 372         __u32 crc32_sum = ~0;
 373         struct blk_plug plug;
 374         /* Tail of the journal */
 375         unsigned long first_block;
 376         tid_t first_tid;
 377         int update_tail;
 378         int csum_size = 0;
 379         LIST_HEAD(io_bufs);
 380         LIST_HEAD(log_bufs);
 381
 382         if (jbd2_journal_has_csum_v2or3(journal))
 383                 csum_size = sizeof(struct jbd2_journal_block_tail);
 384
 385         /*
 386          * First job: lock down the current transaction and wait for
 387          * all outstanding updates to complete.
 388          */
 389
 390         /* Do we need to erase the effects of a prior jbd2_journal_flush? */
 391         if (journal->j_flags & JBD2_FLUSHED) {
 392                 jbd2_debug(3, "super block updated\n");
 393                 mutex_lock_io(&journal->j_checkpoint_mutex);
 394                 /*
 395                  * We hold j_checkpoint_mutex so tail cannot change under us.
 396                  * We don't need any special data guarantees for writing sb
 397                  * since journal is empty and it is ok for write to be
 398                  * flushed only with transaction commit.
 399                  */
 400                 jbd2_journal_update_sb_log_tail(journal,
 401                                                 journal->j_tail_sequence,
 402                                                 journal->j_tail,
 403                                                 REQ_SYNC);
 404                 mutex_unlock(&journal->j_checkpoint_mutex);
 405         } else {
 406                 jbd2_debug(3, "superblock not updated\n");
 407         }
 408
 409         J_ASSERT(journal->j_running_transaction != NULL);
 410         J_ASSERT(journal->j_committing_transaction == NULL);
 411
 412         write_lock(&journal->j_state_lock);
 413         journal->j_flags |= JBD2_FULL_COMMIT_ONGOING;
 414         while (journal->j_flags & JBD2_FAST_COMMIT_ONGOING) {
 415                 DEFINE_WAIT(wait);
 416
 417                 prepare_to_wait(&journal->j_fc_wait, &wait,
 418                                 TASK_UNINTERRUPTIBLE);
 419                 write_unlock(&journal->j_state_lock);
 420                 schedule();
 421                 write_lock(&journal->j_state_lock);
 422                 finish_wait(&journal->j_fc_wait, &wait);
 423                 /*
 424                  * TODO: by blocking fast commits here, we are increasing
 425                  * fsync() latency slightly. Strictly speaking, we don't need
 426                  * to block fast commits until the transaction enters T_FLUSH
 427                  * state. So an optimization is possible where we block new fast
 428                  * commits here and wait for existing ones to complete
 429                  * just before we enter T_FLUSH. That way, the existing fast
 430                  * commits and this full commit can proceed parallely.
 431                  */
 432         }
 433         write_unlock(&journal->j_state_lock);
 434
 435         commit_transaction = journal->j_running_transaction;
 436
 437         trace_jbd2_start_commit(journal, commit_transaction);
 438         jbd2_debug(1, "JBD2: starting commit of transaction %d\n",
 439                         commit_transaction->t_tid);
 440
 441         write_lock(&journal->j_state_lock);
 442         journal->j_fc_off = 0;
 443         J_ASSERT(commit_transaction->t_state == T_RUNNING);
 444         commit_transaction->t_state = T_LOCKED;
 445
 446         trace_jbd2_commit_locking(journal, commit_transaction);
 447         stats.run.rs_wait = commit_transaction->t_max_wait;
 448         stats.run.rs_request_delay = 0;
 449         stats.run.rs_locked = jiffies;
 450         if (commit_transaction->t_requested)
 451                 stats.run.rs_request_delay =
 452                         jbd2_time_diff(commit_transaction->t_requested,
 453                                        stats.run.rs_locked);
 454         stats.run.rs_running = jbd2_time_diff(commit_transaction->t_start,
 455                                               stats.run.rs_locked);
 456
 457         // waits for any t_updates to finish
 458         jbd2_journal_wait_updates(journal);
 459
 460         commit_transaction->t_state = T_SWITCH;
 461
 462         J_ASSERT (atomic_read(&commit_transaction->t_outstanding_credits) <=
 463                         journal->j_max_transaction_buffers);
 464
 465         /*
 466          * First thing we are allowed to do is to discard any remaining
 467          * BJ_Reserved buffers.  Note, it is _not_ permissible to assume
 468          * that there are no such buffers: if a large filesystem
 469          * operation like a truncate needs to split itself over multiple
 470          * transactions, then it may try to do a jbd2_journal_restart() while
 471          * there are still BJ_Reserved buffers outstanding.  These must
 472          * be released cleanly from the current transaction.
 473          *
 474          * In this case, the filesystem must still reserve write access
 475          * again before modifying the buffer in the new transaction, but
 476          * we do not require it to remember exactly which old buffers it
 477          * has reserved.  This is consistent with the existing behaviour
 478          * that multiple jbd2_journal_get_write_access() calls to the same
 479          * buffer are perfectly permissible.
 480          * We use journal->j_state_lock here to serialize processing of
 481          * t_reserved_list with eviction of buffers from journal_unmap_buffer().
 482          */
 483         while (commit_transaction->t_reserved_list) {
 484                 jh = commit_transaction->t_reserved_list;
 485                 JBUFFER_TRACE(jh, "reserved, unused: refile");
 486                 /*
 487                  * A jbd2_journal_get_undo_access()+jbd2_journal_release_buffer() may
 488                  * leave undo-committed data.
 489                  */
 490                 if (jh->b_committed_data) {
 491                         struct buffer_head *bh = jh2bh(jh);
 492
 493                         spin_lock(&jh->b_state_lock);
 494                         jbd2_free(jh->b_committed_data, bh->b_size);
 495                         jh->b_committed_data = NULL;
 496                         spin_unlock(&jh->b_state_lock);
 497                 }
 498                 jbd2_journal_refile_buffer(journal, jh);
 499         }
 500
 501         write_unlock(&journal->j_state_lock);
 502         /*
 503          * Now try to drop any written-back buffers from the journal's
 504          * checkpoint lists.  We do this *before* commit because it potentially
 505          * frees some memory
 506          */
 507         spin_lock(&journal->j_list_lock);
 508         __jbd2_journal_clean_checkpoint_list(journal, false);
 509         spin_unlock(&journal->j_list_lock);
 510
 511         jbd2_debug(3, "JBD2: commit phase 1\n");
 512
 513         /*
 514          * Clear revoked flag to reflect there is no revoked buffers
 515          * in the next transaction which is going to be started.
 516          */
 517         jbd2_clear_buffer_revoked_flags(journal);
 518
 519         /*
 520          * Switch to a new revoke table.
 521          */
 522         jbd2_journal_switch_revoke_table(journal);
 523
 524         write_lock(&journal->j_state_lock);
 525         /*
 526          * Reserved credits cannot be claimed anymore, free them
 527          */
 528         atomic_sub(atomic_read(&journal->j_reserved_credits),
 529                    &commit_transaction->t_outstanding_credits);
 530
 531         trace_jbd2_commit_flushing(journal, commit_transaction);
 532         stats.run.rs_flushing = jiffies;
 533         stats.run.rs_locked = jbd2_time_diff(stats.run.rs_locked,
 534                                              stats.run.rs_flushing);
 535
 536         commit_transaction->t_state = T_FLUSH;
 537         journal->j_committing_transaction = commit_transaction;
 538         journal->j_running_transaction = NULL;
 539         start_time = ktime_get();
 540         commit_transaction->t_log_start = journal->j_head;
 541         wake_up_all(&journal->j_wait_transaction_locked);
 542         write_unlock(&journal->j_state_lock);
 543
 544         jbd2_debug(3, "JBD2: commit phase 2a\n");
 545
 546         /*
 547          * Now start flushing things to disk, in the order they appear
 548          * on the transaction lists.  Data blocks go first.
 549          */
 550         err = journal_submit_data_buffers(journal, commit_transaction);
 551         if (err)
 552                 jbd2_journal_abort(journal, err);
 553
 554         blk_start_plug(&plug);
 555         jbd2_journal_write_revoke_records(commit_transaction, &log_bufs);
 556
 557         jbd2_debug(3, "JBD2: commit phase 2b\n");
 558
 559         /*
 560          * Way to go: we have now written out all of the data for a
 561          * transaction!  Now comes the tricky part: we need to write out
 562          * metadata.  Loop over the transaction's entire buffer list:
 563          */
 564         write_lock(&journal->j_state_lock);
 565         commit_transaction->t_state = T_COMMIT;
 566         write_unlock(&journal->j_state_lock);
 567
 568         trace_jbd2_commit_logging(journal, commit_transaction);
 569         stats.run.rs_logging = jiffies;
 570         stats.run.rs_flushing = jbd2_time_diff(stats.run.rs_flushing,
 571                                                stats.run.rs_logging);
 572         stats.run.rs_blocks = commit_transaction->t_nr_buffers;
 573         stats.run.rs_blocks_logged = 0;
 574
 575         J_ASSERT(commit_transaction->t_nr_buffers <=
 576                  atomic_read(&commit_transaction->t_outstanding_credits));
 577
 578         err = 0;
 579         bufs = 0;
 580         descriptor = NULL;
 581         while (commit_transaction->t_buffers) {
 582
 583                 /* Find the next buffer to be journaled... */
 584
 585                 jh = commit_transaction->t_buffers;
 586
 587                 /* If we're in abort mode, we just un-journal the buffer and
 588                    release it. */
 589
 590                 if (is_journal_aborted(journal)) {
 591                         clear_buffer_jbddirty(jh2bh(jh));
 592                         JBUFFER_TRACE(jh, "journal is aborting: refile");
 593                         jbd2_buffer_abort_trigger(jh,
 594                                                   jh->b_frozen_data ?
 595                                                   jh->b_frozen_triggers :
 596                                                   jh->b_triggers);
 597                         jbd2_journal_refile_buffer(journal, jh);
 598                         /* If that was the last one, we need to clean up
 599                          * any descriptor buffers which may have been
 600                          * already allocated, even if we are now
 601                          * aborting. */
 602                         if (!commit_transaction->t_buffers)
 603                                 goto start_journal_io;
 604                         continue;
 605                 }
 606
 607                 /* Make sure we have a descriptor block in which to
 608                    record the metadata buffer. */
 609
 610                 if (!descriptor) {
 611                         J_ASSERT (bufs == 0);
 612
 613                         jbd2_debug(4, "JBD2: get descriptor\n");
 614
 615                         descriptor = jbd2_journal_get_descriptor_buffer(
 616                                                         commit_transaction,
 617                                                         JBD2_DESCRIPTOR_BLOCK);
 618                         if (!descriptor) {
 619                                 jbd2_journal_abort(journal, -EIO);
 620                                 continue;
 621                         }
 622
 623                         jbd2_debug(4, "JBD2: got buffer %llu (%p)\n",
 624                                 (unsigned long long)descriptor->b_blocknr,
 625                                 descriptor->b_data);
 626                         tagp = &descriptor->b_data[sizeof(journal_header_t)];
 627                         space_left = descriptor->b_size -
 628                                                 sizeof(journal_header_t);
 629                         first_tag = 1;
 630                         set_buffer_jwrite(descriptor);
 631                         set_buffer_dirty(descriptor);
 632                         wbuf[bufs++] = descriptor;
 633
 634                         /* Record it so that we can wait for IO
 635                            completion later */
 636                         BUFFER_TRACE(descriptor, "ph3: file as descriptor");
 637                         jbd2_file_log_bh(&log_bufs, descriptor);
 638                 }
 639
 640                 /* Where is the buffer to be written? */
 641
 642                 err = jbd2_journal_next_log_block(journal, &blocknr);
 643                 /* If the block mapping failed, just abandon the buffer
 644                    and repeat this loop: we'll fall into the
 645                    refile-on-abort condition above. */
 646                 if (err) {
 647                         jbd2_journal_abort(journal, err);
 648                         continue;
 649                 }
 650
 651                 /*
 652                  * start_this_handle() uses t_outstanding_credits to determine
 653                  * the free space in the log.
 654                  */
 655                 atomic_dec(&commit_transaction->t_outstanding_credits);
 656
 657                 /* Bump b_count to prevent truncate from stumbling over
 658                    the shadowed buffer!  @@@ This can go if we ever get
 659                    rid of the shadow pairing of buffers. */
 660                 atomic_inc(&jh2bh(jh)->b_count);
 661
 662                 /*
 663                  * Make a temporary IO buffer with which to write it out
 664                  * (this will requeue the metadata buffer to BJ_Shadow).
 665                  */
 666                 set_bit(BH_JWrite, &jh2bh(jh)->b_state);
 667                 JBUFFER_TRACE(jh, "ph3: write metadata");
 668                 flags = jbd2_journal_write_metadata_buffer(commit_transaction,
 669                                                 jh, &wbuf[bufs], blocknr);
 670                 if (flags < 0) {
 671                         jbd2_journal_abort(journal, flags);
 672                         continue;
 673                 }
 674                 jbd2_file_log_bh(&io_bufs, wbuf[bufs]);
 675
 676                 /* Record the new block's tag in the current descriptor
 677                    buffer */
 678
 679                 tag_flag = 0;
 680                 if (flags & 1)
 681                         tag_flag |= JBD2_FLAG_ESCAPE;
 682                 if (!first_tag)
 683                         tag_flag |= JBD2_FLAG_SAME_UUID;
 684
 685                 tag = (journal_block_tag_t *) tagp;
 686                 write_tag_block(journal, tag, jh2bh(jh)->b_blocknr);
 687                 tag->t_flags = cpu_to_be16(tag_flag);
 688                 jbd2_block_tag_csum_set(journal, tag, wbuf[bufs],
 689                                         commit_transaction->t_tid);
 690                 tagp += tag_bytes;
 691                 space_left -= tag_bytes;
 692                 bufs++;
 693
 694                 if (first_tag) {
 695                         memcpy (tagp, journal->j_uuid, 16);
 696                         tagp += 16;
 697                         space_left -= 16;
 698                         first_tag = 0;
 699                 }
 700
 701                 /* If there's no more to do, or if the descriptor is full,
 702                    let the IO rip! */
 703
 704                 if (bufs == journal->j_wbufsize ||
 705                     commit_transaction->t_buffers == NULL ||
 706                     space_left < tag_bytes + 16 + csum_size) {
 707
 708                         jbd2_debug(4, "JBD2: Submit %d IOs\n", bufs);
 709
 710                         /* Write an end-of-descriptor marker before
 711                            submitting the IOs.  "tag" still points to
 712                            the last tag we set up. */
 713
 714                         tag->t_flags |= cpu_to_be16(JBD2_FLAG_LAST_TAG);
 715 start_journal_io:
 716                         if (descriptor)
 717                                 jbd2_descriptor_block_csum_set(journal,
 718                                                         descriptor);
 719
 720                         for (i = 0; i < bufs; i++) {
 721                                 struct buffer_head *bh = wbuf[i];
 722                                 /*
 723                                  * Compute checksum.
 724                                  */
 725                                 if (jbd2_has_feature_checksum(journal)) {
 726                                         crc32_sum =
 727                                             jbd2_checksum_data(crc32_sum, bh);
 728                                 }
 729
 730                                 lock_buffer(bh);
 731                                 clear_buffer_dirty(bh);
 732                                 set_buffer_uptodate(bh);
 733                                 bh->b_end_io = journal_end_buffer_io_sync;
 734                                 submit_bh(REQ_OP_WRITE | REQ_SYNC, bh);
 735                         }
 736                         cond_resched();
 737
 738                         /* Force a new descriptor to be generated next
 739                            time round the loop. */
 740                         descriptor = NULL;
 741                         bufs = 0;
 742                 }
 743         }
 744
 745         err = journal_finish_inode_data_buffers(journal, commit_transaction);
 746         if (err) {
 747                 printk(KERN_WARNING
 748                         "JBD2: Detected IO errors while flushing file data "
 749                        "on %s\n", journal->j_devname);
 750                 if (journal->j_flags & JBD2_ABORT_ON_SYNCDATA_ERR)
 751                         jbd2_journal_abort(journal, err);
 752                 err = 0;
 753         }
 754
 755         /*
 756          * Get current oldest transaction in the log before we issue flush
 757          * to the filesystem device. After the flush we can be sure that
 758          * blocks of all older transactions are checkpointed to persistent
 759          * storage and we will be safe to update journal start in the
 760          * superblock with the numbers we get here.
 761          */
 762         update_tail =
 763                 jbd2_journal_get_log_tail(journal, &first_tid, &first_block);
 764
 765         write_lock(&journal->j_state_lock);
 766         if (update_tail) {
 767                 long freed = first_block - journal->j_tail;
 768
 769                 if (first_block < journal->j_tail)
 770                         freed += journal->j_last - journal->j_first;
 771                 /* Update tail only if we free significant amount of space */
 772                 if (freed < jbd2_journal_get_max_txn_bufs(journal))
 773                         update_tail = 0;
 774         }
 775         J_ASSERT(commit_transaction->t_state == T_COMMIT);
 776         commit_transaction->t_state = T_COMMIT_DFLUSH;
 777         write_unlock(&journal->j_state_lock);
 778
 779         /*
 780          * If the journal is not located on the file system device,
 781          * then we must flush the file system device before we issue
 782          * the commit record
 783          */
 784         if (commit_transaction->t_need_data_flush &&
 785             (journal->j_fs_dev != journal->j_dev) &&
 786             (journal->j_flags & JBD2_BARRIER))
 787                 blkdev_issue_flush(journal->j_fs_dev);
 788
 789         /* Done it all: now write the commit record asynchronously. */
 790         if (jbd2_has_feature_async_commit(journal)) {
 791                 err = journal_submit_commit_record(journal, commit_transaction,
 792                                                  &cbh, crc32_sum);
 793                 if (err)
 794                         jbd2_journal_abort(journal, err);
 795         }
 796
 797         blk_finish_plug(&plug);
 798
 799         /* Lo and behold: we have just managed to send a transaction to
 800            the log.  Before we can commit it, wait for the IO so far to
 801            complete.  Control buffers being written are on the
 802            transaction's t_log_list queue, and metadata buffers are on
 803            the io_bufs list.
 804
 805            Wait for the buffers in reverse order.  That way we are
 806            less likely to be woken up until all IOs have completed, and
 807            so we incur less scheduling load.
 808         */
 809
 810         jbd2_debug(3, "JBD2: commit phase 3\n");
 811
 812         while (!list_empty(&io_bufs)) {
 813                 struct buffer_head *bh = list_entry(io_bufs.prev,
 814                                                     struct buffer_head,
 815                                                     b_assoc_buffers);
 816
 817                 wait_on_buffer(bh);
 818                 cond_resched();
 819
 820                 if (unlikely(!buffer_uptodate(bh)))
 821                         err = -EIO;
 822                 jbd2_unfile_log_bh(bh);
 823                 stats.run.rs_blocks_logged++;
 824
 825                 /*
 826                  * The list contains temporary buffer heads created by
 827                  * jbd2_journal_write_metadata_buffer().
 828                  */
 829                 BUFFER_TRACE(bh, "dumping temporary bh");
 830                 __brelse(bh);
 831                 J_ASSERT_BH(bh, atomic_read(&bh->b_count) == 0);
 832                 free_buffer_head(bh);
 833
 834                 /* We also have to refile the corresponding shadowed buffer */
 835                 jh = commit_transaction->t_shadow_list->b_tprev;
 836                 bh = jh2bh(jh);
 837                 clear_buffer_jwrite(bh);
 838                 J_ASSERT_BH(bh, buffer_jbddirty(bh));
 839                 J_ASSERT_BH(bh, !buffer_shadow(bh));
 840
 841                 /* The metadata is now released for reuse, but we need
 842                    to remember it against this transaction so that when
 843                    we finally commit, we can do any checkpointing
 844                    required. */
 845                 JBUFFER_TRACE(jh, "file as BJ_Forget");
 846                 jbd2_journal_file_buffer(jh, commit_transaction, BJ_Forget);
 847                 JBUFFER_TRACE(jh, "brelse shadowed buffer");
 848                 __brelse(bh);
 849         }
 850
 851         J_ASSERT (commit_transaction->t_shadow_list == NULL);
 852
 853         jbd2_debug(3, "JBD2: commit phase 4\n");
 854
 855         /* Here we wait for the revoke record and descriptor record buffers */
 856         while (!list_empty(&log_bufs)) {
 857                 struct buffer_head *bh;
 858
 859                 bh = list_entry(log_bufs.prev, struct buffer_head, b_assoc_buffers);
 860                 wait_on_buffer(bh);
 861                 cond_resched();
 862
 863                 if (unlikely(!buffer_uptodate(bh)))
 864                         err = -EIO;
 865
 866                 BUFFER_TRACE(bh, "ph5: control buffer writeout done: unfile");
 867                 clear_buffer_jwrite(bh);
 868                 jbd2_unfile_log_bh(bh);
 869                 stats.run.rs_blocks_logged++;
 870                 __brelse(bh);           /* One for getblk */
 871                 /* AKPM: bforget here */
 872         }
 873
 874         if (err)
 875                 jbd2_journal_abort(journal, err);
 876
 877         jbd2_debug(3, "JBD2: commit phase 5\n");
 878         write_lock(&journal->j_state_lock);
 879         J_ASSERT(commit_transaction->t_state == T_COMMIT_DFLUSH);
 880         commit_transaction->t_state = T_COMMIT_JFLUSH;
 881         write_unlock(&journal->j_state_lock);
 882
 883         if (!jbd2_has_feature_async_commit(journal)) {
 884                 err = journal_submit_commit_record(journal, commit_transaction,
 885                                                 &cbh, crc32_sum);
 886                 if (err)
 887                         jbd2_journal_abort(journal, err);
 888         }
 889         if (cbh)
 890                 err = journal_wait_on_commit_record(journal, cbh);
 891         stats.run.rs_blocks_logged++;
 892         if (jbd2_has_feature_async_commit(journal) &&
 893             journal->j_flags & JBD2_BARRIER) {
 894                 blkdev_issue_flush(journal->j_dev);
 895         }
 896
 897         if (err)
 898                 jbd2_journal_abort(journal, err);
 899
 900         WARN_ON_ONCE(
 901                 atomic_read(&commit_transaction->t_outstanding_credits) < 0);
 902
 903         /*
 904          * Now disk caches for filesystem device are flushed so we are safe to
 905          * erase checkpointed transactions from the log by updating journal
 906          * superblock.
 907          */
 908         if (update_tail)
 909                 jbd2_update_log_tail(journal, first_tid, first_block);
 910
 911         /* End of a transaction!  Finally, we can do checkpoint
 912            processing: any buffers committed as a result of this
 913            transaction can be removed from any checkpoint list it was on
 914            before. */
 915
 916         jbd2_debug(3, "JBD2: commit phase 6\n");
 917
 918         J_ASSERT(list_empty(&commit_transaction->t_inode_list));
 919         J_ASSERT(commit_transaction->t_buffers == NULL);
 920         J_ASSERT(commit_transaction->t_checkpoint_list == NULL);
 921         J_ASSERT(commit_transaction->t_shadow_list == NULL);
 922
 923 restart_loop:
 924         /*
 925          * As there are other places (journal_unmap_buffer()) adding buffers
 926          * to this list we have to be careful and hold the j_list_lock.
 927          */
 928         spin_lock(&journal->j_list_lock);
 929         while (commit_transaction->t_forget) {
 930                 transaction_t *cp_transaction;
 931                 struct buffer_head *bh;
 932                 int try_to_free = 0;
 933                 bool drop_ref;
 934
 935                 jh = commit_transaction->t_forget;
 936                 spin_unlock(&journal->j_list_lock);
 937                 bh = jh2bh(jh);
 938                 /*
 939                  * Get a reference so that bh cannot be freed before we are
 940                  * done with it.
 941                  */
 942                 get_bh(bh);
 943                 spin_lock(&jh->b_state_lock);
 944                 J_ASSERT_JH(jh, jh->b_transaction == commit_transaction);
 945
 946                 /*
 947                  * If there is undo-protected committed data against
 948                  * this buffer, then we can remove it now.  If it is a
 949                  * buffer needing such protection, the old frozen_data
 950                  * field now points to a committed version of the
 951                  * buffer, so rotate that field to the new committed
 952                  * data.
 953                  *
 954                  * Otherwise, we can just throw away the frozen data now.
 955                  *
 956                  * We also know that the frozen data has already fired
 957                  * its triggers if they exist, so we can clear that too.
 958                  */
 959                 if (jh->b_committed_data) {
 960                         jbd2_free(jh->b_committed_data, bh->b_size);
 961                         jh->b_committed_data = NULL;
 962                         if (jh->b_frozen_data) {
 963                                 jh->b_committed_data = jh->b_frozen_data;
 964                                 jh->b_frozen_data = NULL;
 965                                 jh->b_frozen_triggers = NULL;
 966                         }
 967                 } else if (jh->b_frozen_data) {
 968                         jbd2_free(jh->b_frozen_data, bh->b_size);
 969                         jh->b_frozen_data = NULL;
 970                         jh->b_frozen_triggers = NULL;
 971                 }
 972
 973                 spin_lock(&journal->j_list_lock);
 974                 cp_transaction = jh->b_cp_transaction;
 975                 if (cp_transaction) {
 976                         JBUFFER_TRACE(jh, "remove from old cp transaction");
 977                         cp_transaction->t_chp_stats.cs_dropped++;
 978                         __jbd2_journal_remove_checkpoint(jh);
 979                 }
 980
 981                 /* Only re-checkpoint the buffer_head if it is marked
 982                  * dirty.  If the buffer was added to the BJ_Forget list
 983                  * by jbd2_journal_forget, it may no longer be dirty and
 984                  * there's no point in keeping a checkpoint record for
 985                  * it. */
 986
 987                 /*
 988                  * A buffer which has been freed while still being journaled
 989                  * by a previous transaction, refile the buffer to BJ_Forget of
 990                  * the running transaction. If the just committed transaction
 991                  * contains "add to orphan" operation, we can completely
 992                  * invalidate the buffer now. We are rather through in that
 993                  * since the buffer may be still accessible when blocksize <
 994                  * pagesize and it is attached to the last partial page.
 995                  */
 996                 if (buffer_freed(bh) && !jh->b_next_transaction) {
 997                         struct address_space *mapping;
 998
 999                         clear_buffer_freed(bh);
1000                         clear_buffer_jbddirty(bh);
1001
1002                         /*
1003                          * Block device buffers need to stay mapped all the
1004                          * time, so it is enough to clear buffer_jbddirty and
1005                          * buffer_freed bits. For the file mapping buffers (i.e.
1006                          * journalled data) we need to unmap buffer and clear
1007                          * more bits. We also need to be careful about the check
1008                          * because the data page mapping can get cleared under
1009                          * our hands. Note that if mapping == NULL, we don't
1010                          * need to make buffer unmapped because the page is
1011                          * already detached from the mapping and buffers cannot
1012                          * get reused.
1013                          */
1014                         mapping = READ_ONCE(bh->b_folio->mapping);
1015                         if (mapping && !sb_is_blkdev_sb(mapping->host->i_sb)) {
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                 drop_ref = __jbd2_journal_refile_buffer(jh);
1044                 spin_unlock(&jh->b_state_lock);
1045                 if (drop_ref)
1046                         jbd2_journal_put_journal_head(jh);
1047                 if (try_to_free)
1048                         release_buffer_page(bh);        /* Drops bh reference */
1049                 else
1050                         __brelse(bh);
1051                 cond_resched_lock(&journal->j_list_lock);
1052         }
1053         spin_unlock(&journal->j_list_lock);
1054         /*
1055          * This is a bit sleazy.  We use j_list_lock to protect transition
1056          * of a transaction into T_FINISHED state and calling
1057          * __jbd2_journal_drop_transaction(). Otherwise we could race with
1058          * other checkpointing code processing the transaction...
1059          */
1060         write_lock(&journal->j_state_lock);
1061         spin_lock(&journal->j_list_lock);
1062         /*
1063          * Now recheck if some buffers did not get attached to the transaction
1064          * while the lock was dropped...
1065          */
1066         if (commit_transaction->t_forget) {
1067                 spin_unlock(&journal->j_list_lock);
1068                 write_unlock(&journal->j_state_lock);
1069                 goto restart_loop;
1070         }
1071
1072         /* Add the transaction to the checkpoint list
1073          * __journal_remove_checkpoint() can not destroy transaction
1074          * under us because it is not marked as T_FINISHED yet */
1075         if (journal->j_checkpoint_transactions == NULL) {
1076                 journal->j_checkpoint_transactions = commit_transaction;
1077                 commit_transaction->t_cpnext = commit_transaction;
1078                 commit_transaction->t_cpprev = commit_transaction;
1079         } else {
1080                 commit_transaction->t_cpnext =
1081                         journal->j_checkpoint_transactions;
1082                 commit_transaction->t_cpprev =
1083                         commit_transaction->t_cpnext->t_cpprev;
1084                 commit_transaction->t_cpnext->t_cpprev =
1085                         commit_transaction;
1086                 commit_transaction->t_cpprev->t_cpnext =
1087                                 commit_transaction;
1088         }
1089         spin_unlock(&journal->j_list_lock);
1090
1091         /* Done with this transaction! */
1092
1093         jbd2_debug(3, "JBD2: commit phase 7\n");
1094
1095         J_ASSERT(commit_transaction->t_state == T_COMMIT_JFLUSH);
1096
1097         commit_transaction->t_start = jiffies;
1098         stats.run.rs_logging = jbd2_time_diff(stats.run.rs_logging,
1099                                               commit_transaction->t_start);
1100
1101         /*
1102          * File the transaction statistics
1103          */
1104         stats.ts_tid = commit_transaction->t_tid;
1105         stats.run.rs_handle_count =
1106                 atomic_read(&commit_transaction->t_handle_count);
1107         trace_jbd2_run_stats(journal->j_fs_dev->bd_dev,
1108                              commit_transaction->t_tid, &stats.run);
1109         stats.ts_requested = (commit_transaction->t_requested) ? 1 : 0;
1110
1111         commit_transaction->t_state = T_COMMIT_CALLBACK;
1112         J_ASSERT(commit_transaction == journal->j_committing_transaction);
1113         journal->j_commit_sequence = commit_transaction->t_tid;
1114         journal->j_committing_transaction = NULL;
1115         commit_time = ktime_to_ns(ktime_sub(ktime_get(), start_time));
1116
1117         /*
1118          * weight the commit time higher than the average time so we don't
1119          * react too strongly to vast changes in the commit time
1120          */
1121         if (likely(journal->j_average_commit_time))
1122                 journal->j_average_commit_time = (commit_time +
1123                                 journal->j_average_commit_time*3) / 4;
1124         else
1125                 journal->j_average_commit_time = commit_time;
1126
1127         write_unlock(&journal->j_state_lock);
1128
1129         if (journal->j_commit_callback)
1130                 journal->j_commit_callback(journal, commit_transaction);
1131         if (journal->j_fc_cleanup_callback)
1132                 journal->j_fc_cleanup_callback(journal, 1, commit_transaction->t_tid);
1133
1134         trace_jbd2_end_commit(journal, commit_transaction);
1135         jbd2_debug(1, "JBD2: commit %d complete, head %d\n",
1136                   journal->j_commit_sequence, journal->j_tail_sequence);
1137
1138         write_lock(&journal->j_state_lock);
1139         journal->j_flags &= ~JBD2_FULL_COMMIT_ONGOING;
1140         journal->j_flags &= ~JBD2_FAST_COMMIT_ONGOING;
1141         spin_lock(&journal->j_list_lock);
1142         commit_transaction->t_state = T_FINISHED;
1143         /* Check if the transaction can be dropped now that we are finished */
1144         if (commit_transaction->t_checkpoint_list == NULL) {
1145                 __jbd2_journal_drop_transaction(journal, commit_transaction);
1146                 jbd2_journal_free_transaction(commit_transaction);
1147         }
1148         spin_unlock(&journal->j_list_lock);
1149         write_unlock(&journal->j_state_lock);
1150         wake_up(&journal->j_wait_done_commit);
1151         wake_up(&journal->j_fc_wait);
1152
1153         /*
1154          * Calculate overall stats
1155          */
1156         spin_lock(&journal->j_history_lock);
1157         journal->j_stats.ts_tid++;
1158         journal->j_stats.ts_requested += stats.ts_requested;
1159         journal->j_stats.run.rs_wait += stats.run.rs_wait;
1160         journal->j_stats.run.rs_request_delay += stats.run.rs_request_delay;
1161         journal->j_stats.run.rs_running += stats.run.rs_running;
1162         journal->j_stats.run.rs_locked += stats.run.rs_locked;
1163         journal->j_stats.run.rs_flushing += stats.run.rs_flushing;
1164         journal->j_stats.run.rs_logging += stats.run.rs_logging;
1165         journal->j_stats.run.rs_handle_count += stats.run.rs_handle_count;
1166         journal->j_stats.run.rs_blocks += stats.run.rs_blocks;
1167         journal->j_stats.run.rs_blocks_logged += stats.run.rs_blocks_logged;
1168         spin_unlock(&journal->j_history_lock);
1169 }