fs/jbd2/commit.c

   1 /*
   2  * linux/fs/jbd2/commit.c
   3  *
   4  * Written by Stephen C. Tweedie <sct@redhat.com>, 1998
   5  *
   6  * Copyright 1998 Red Hat corp --- All Rights Reserved
   7  *
   8  * This file is part of the Linux kernel and is made available under
   9  * the terms of the GNU General Public License, version 2, or at your
  10  * option, any later version, incorporated herein by reference.
  11  *
  12  * Journal commit routines for the generic filesystem journaling code;
  13  * part of the ext2fs journaling system.
  14  */
  15
  16 #include <linux/time.h>
  17 #include <linux/fs.h>
  18 #include <linux/jbd2.h>
  19 #include <linux/errno.h>
  20 #include <linux/slab.h>
  21 #include <linux/mm.h>
  22 #include <linux/pagemap.h>
  23 #include <linux/jiffies.h>
  24 #include <linux/crc32.h>
  25 #include <linux/writeback.h>
  26 #include <linux/backing-dev.h>
  27 #include <linux/bio.h>
  28 #include <linux/blkdev.h>
  29 #include <trace/events/jbd2.h>
  30
  31 /*
  32  * Default IO end handler for temporary BJ_IO buffer_heads.
  33  */
  34 static void journal_end_buffer_io_sync(struct buffer_head *bh, int uptodate)
  35 {
  36         BUFFER_TRACE(bh, "");
  37         if (uptodate)
  38                 set_buffer_uptodate(bh);
  39         else
  40                 clear_buffer_uptodate(bh);
  41         unlock_buffer(bh);
  42 }
  43
  44 /*
  45  * When an ext4 file is truncated, it is possible that some pages are not
  46  * successfully freed, because they are attached to a committing transaction.
  47  * After the transaction commits, these pages are left on the LRU, with no
  48  * ->mapping, and with attached buffers.  These pages are trivially reclaimable
  49  * by the VM, but their apparent absence upsets the VM accounting, and it makes
  50  * the numbers in /proc/meminfo look odd.
  51  *
  52  * So here, we have a buffer which has just come off the forget list.  Look to
  53  * see if we can strip all buffers from the backing page.
  54  *
  55  * Called under lock_journal(), and possibly under journal_datalist_lock.  The
  56  * caller provided us with a ref against the buffer, and we drop that here.
  57  */
  58 static void release_buffer_page(struct buffer_head *bh)
  59 {
  60         struct page *page;
  61
  62         if (buffer_dirty(bh))
  63                 goto nope;
  64         if (atomic_read(&bh->b_count) != 1)
  65                 goto nope;
  66         page = bh->b_page;
  67         if (!page)
  68                 goto nope;
  69         if (page->mapping)
  70                 goto nope;
  71
  72         /* OK, it's a truncated page */
  73         if (!trylock_page(page))
  74                 goto nope;
  75
  76         page_cache_get(page);
  77         __brelse(bh);
  78         try_to_free_buffers(page);
  79         unlock_page(page);
  80         page_cache_release(page);
  81         return;
  82
  83 nope:
  84         __brelse(bh);
  85 }
  86
  87 /*
  88  * Done it all: now submit the commit record.  We should have
  89  * cleaned up our previous buffers by now, so if we are in abort
  90  * mode we can now just skip the rest of the journal write
  91  * entirely.
  92  *
  93  * Returns 1 if the journal needs to be aborted or 0 on success
  94  */
  95 static int journal_submit_commit_record(journal_t *journal,
  96                                         transaction_t *commit_transaction,
  97                                         struct buffer_head **cbh,
  98                                         __u32 crc32_sum)
  99 {
 100         struct journal_head *descriptor;
 101         struct commit_header *tmp;
 102         struct buffer_head *bh;
 103         int ret;
 104         struct timespec now = current_kernel_time();
 105
 106         if (is_journal_aborted(journal))
 107                 return 0;
 108
 109         descriptor = jbd2_journal_get_descriptor_buffer(journal);
 110         if (!descriptor)
 111                 return 1;
 112
 113         bh = jh2bh(descriptor);
 114
 115         tmp = (struct commit_header *)bh->b_data;
 116         tmp->h_magic = cpu_to_be32(JBD2_MAGIC_NUMBER);
 117         tmp->h_blocktype = cpu_to_be32(JBD2_COMMIT_BLOCK);
 118         tmp->h_sequence = cpu_to_be32(commit_transaction->t_tid);
 119         tmp->h_commit_sec = cpu_to_be64(now.tv_sec);
 120         tmp->h_commit_nsec = cpu_to_be32(now.tv_nsec);
 121
 122         if (JBD2_HAS_COMPAT_FEATURE(journal,
 123                                     JBD2_FEATURE_COMPAT_CHECKSUM)) {
 124                 tmp->h_chksum_type      = JBD2_CRC32_CHKSUM;
 125                 tmp->h_chksum_size      = JBD2_CRC32_CHKSUM_SIZE;
 126                 tmp->h_chksum[0]        = cpu_to_be32(crc32_sum);
 127         }
 128
 129         JBUFFER_TRACE(descriptor, "submit commit block");
 130         lock_buffer(bh);
 131         clear_buffer_dirty(bh);
 132         set_buffer_uptodate(bh);
 133         bh->b_end_io = journal_end_buffer_io_sync;
 134
 135         if (journal->j_flags & JBD2_BARRIER &&
 136             !JBD2_HAS_INCOMPAT_FEATURE(journal,
 137                                        JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
 138                 ret = submit_bh(WRITE_SYNC_PLUG | WRITE_BARRIER, bh);
 139                 if (ret == -EOPNOTSUPP) {
 140                         printk(KERN_WARNING
 141                                "JBD2: Disabling barriers on %s, "
 142                                "not supported by device\n", journal->j_devname);
 143                         write_lock(&journal->j_state_lock);
 144                         journal->j_flags &= ~JBD2_BARRIER;
 145                         write_unlock(&journal->j_state_lock);
 146
 147                         /* And try again, without the barrier */
 148                         lock_buffer(bh);
 149                         set_buffer_uptodate(bh);
 150                         clear_buffer_dirty(bh);
 151                         ret = submit_bh(WRITE_SYNC_PLUG, bh);
 152                 }
 153         } else {
 154                 ret = submit_bh(WRITE_SYNC_PLUG, bh);
 155         }
 156         *cbh = bh;
 157         return ret;
 158 }
 159
 160 /*
 161  * This function along with journal_submit_commit_record
 162  * allows to write the commit record asynchronously.
 163  */
 164 static int journal_wait_on_commit_record(journal_t *journal,
 165                                          struct buffer_head *bh)
 166 {
 167         int ret = 0;
 168
 169 retry:
 170         clear_buffer_dirty(bh);
 171         wait_on_buffer(bh);
 172         if (buffer_eopnotsupp(bh) && (journal->j_flags & JBD2_BARRIER)) {
 173                 printk(KERN_WARNING
 174                        "JBD2: %s: disabling barries on %s - not supported "
 175                        "by device\n", __func__, journal->j_devname);
 176                 write_lock(&journal->j_state_lock);
 177                 journal->j_flags &= ~JBD2_BARRIER;
 178                 write_unlock(&journal->j_state_lock);
 179
 180                 lock_buffer(bh);
 181                 clear_buffer_dirty(bh);
 182                 set_buffer_uptodate(bh);
 183                 bh->b_end_io = journal_end_buffer_io_sync;
 184
 185                 ret = submit_bh(WRITE_SYNC_PLUG, bh);
 186                 if (ret) {
 187                         unlock_buffer(bh);
 188                         return ret;
 189                 }
 190                 goto retry;
 191         }
 192
 193         if (unlikely(!buffer_uptodate(bh)))
 194                 ret = -EIO;
 195         put_bh(bh);            /* One for getblk() */
 196         jbd2_journal_put_journal_head(bh2jh(bh));
 197
 198         return ret;
 199 }
 200
 201 /*
 202  * write the filemap data using writepage() address_space_operations.
 203  * We don't do block allocation here even for delalloc. We don't
 204  * use writepages() because with dealyed allocation we may be doing
 205  * block allocation in writepages().
 206  */
 207 static int journal_submit_inode_data_buffers(struct address_space *mapping)
 208 {
 209         int ret;
 210         struct writeback_control wbc = {
 211                 .sync_mode =  WB_SYNC_ALL,
 212                 .nr_to_write = mapping->nrpages * 2,
 213                 .range_start = 0,
 214                 .range_end = i_size_read(mapping->host),
 215         };
 216
 217         ret = generic_writepages(mapping, &wbc);
 218         return ret;
 219 }
 220
 221 /*
 222  * Submit all the data buffers of inode associated with the transaction to
 223  * disk.
 224  *
 225  * We are in a committing transaction. Therefore no new inode can be added to
 226  * our inode list. We use JI_COMMIT_RUNNING flag to protect inode we currently
 227  * operate on from being released while we write out pages.
 228  */
 229 static int journal_submit_data_buffers(journal_t *journal,
 230                 transaction_t *commit_transaction)
 231 {
 232         struct jbd2_inode *jinode;
 233         int err, ret = 0;
 234         struct address_space *mapping;
 235
 236         spin_lock(&journal->j_list_lock);
 237         list_for_each_entry(jinode, &commit_transaction->t_inode_list, i_list) {
 238                 mapping = jinode->i_vfs_inode->i_mapping;
 239                 jinode->i_flags |= JI_COMMIT_RUNNING;
 240                 spin_unlock(&journal->j_list_lock);
 241                 /*
 242                  * submit the inode data buffers. We use writepage
 243                  * instead of writepages. Because writepages can do
 244                  * block allocation  with delalloc. We need to write
 245                  * only allocated blocks here.
 246                  */
 247                 trace_jbd2_submit_inode_data(jinode->i_vfs_inode);
 248                 err = journal_submit_inode_data_buffers(mapping);
 249                 if (!ret)
 250                         ret = err;
 251                 spin_lock(&journal->j_list_lock);
 252                 J_ASSERT(jinode->i_transaction == commit_transaction);
 253                 jinode->i_flags &= ~JI_COMMIT_RUNNING;
 254                 wake_up_bit(&jinode->i_flags, __JI_COMMIT_RUNNING);
 255         }
 256         spin_unlock(&journal->j_list_lock);
 257         return ret;
 258 }
 259
 260 /*
 261  * Wait for data submitted for writeout, refile inodes to proper
 262  * transaction if needed.
 263  *
 264  */
 265 static int journal_finish_inode_data_buffers(journal_t *journal,
 266                 transaction_t *commit_transaction)
 267 {
 268         struct jbd2_inode *jinode, *next_i;
 269         int err, ret = 0;
 270
 271         /* For locking, see the comment in journal_submit_data_buffers() */
 272         spin_lock(&journal->j_list_lock);
 273         list_for_each_entry(jinode, &commit_transaction->t_inode_list, i_list) {
 274                 jinode->i_flags |= JI_COMMIT_RUNNING;
 275                 spin_unlock(&journal->j_list_lock);
 276                 err = filemap_fdatawait(jinode->i_vfs_inode->i_mapping);
 277                 if (err) {
 278                         /*
 279                          * Because AS_EIO is cleared by
 280                          * filemap_fdatawait_range(), set it again so
 281                          * that user process can get -EIO from fsync().
 282                          */
 283                         set_bit(AS_EIO,
 284                                 &jinode->i_vfs_inode->i_mapping->flags);
 285
 286                         if (!ret)
 287                                 ret = err;
 288                 }
 289                 spin_lock(&journal->j_list_lock);
 290                 jinode->i_flags &= ~JI_COMMIT_RUNNING;
 291                 wake_up_bit(&jinode->i_flags, __JI_COMMIT_RUNNING);
 292         }
 293
 294         /* Now refile inode to proper lists */
 295         list_for_each_entry_safe(jinode, next_i,
 296                                  &commit_transaction->t_inode_list, i_list) {
 297                 list_del(&jinode->i_list);
 298                 if (jinode->i_next_transaction) {
 299                         jinode->i_transaction = jinode->i_next_transaction;
 300                         jinode->i_next_transaction = NULL;
 301                         list_add(&jinode->i_list,
 302                                 &jinode->i_transaction->t_inode_list);
 303                 } else {
 304                         jinode->i_transaction = NULL;
 305                 }
 306         }
 307         spin_unlock(&journal->j_list_lock);
 308
 309         return ret;
 310 }
 311
 312 static __u32 jbd2_checksum_data(__u32 crc32_sum, struct buffer_head *bh)
 313 {
 314         struct page *page = bh->b_page;
 315         char *addr;
 316         __u32 checksum;
 317
 318         addr = kmap_atomic(page, KM_USER0);
 319         checksum = crc32_be(crc32_sum,
 320                 (void *)(addr + offset_in_page(bh->b_data)), bh->b_size);
 321         kunmap_atomic(addr, KM_USER0);
 322
 323         return checksum;
 324 }
 325
 326 static void write_tag_block(int tag_bytes, journal_block_tag_t *tag,
 327                                    unsigned long long block)
 328 {
 329         tag->t_blocknr = cpu_to_be32(block & (u32)~0);
 330         if (tag_bytes > JBD2_TAG_SIZE32)
 331                 tag->t_blocknr_high = cpu_to_be32((block >> 31) >> 1);
 332 }
 333
 334 /*
 335  * jbd2_journal_commit_transaction
 336  *
 337  * The primary function for committing a transaction to the log.  This
 338  * function is called by the journal thread to begin a complete commit.
 339  */
 340 void jbd2_journal_commit_transaction(journal_t *journal)
 341 {
 342         struct transaction_stats_s stats;
 343         transaction_t *commit_transaction;
 344         struct journal_head *jh, *new_jh, *descriptor;
 345         struct buffer_head **wbuf = journal->j_wbuf;
 346         int bufs;
 347         int flags;
 348         int err;
 349         unsigned long long blocknr;
 350         ktime_t start_time;
 351         u64 commit_time;
 352         char *tagp = NULL;
 353         journal_header_t *header;
 354         journal_block_tag_t *tag = NULL;
 355         int space_left = 0;
 356         int first_tag = 0;
 357         int tag_flag;
 358         int i, to_free = 0;
 359         int tag_bytes = journal_tag_bytes(journal);
 360         struct buffer_head *cbh = NULL; /* For transactional checksums */
 361         __u32 crc32_sum = ~0;
 362         int write_op = WRITE;
 363
 364         /*
 365          * First job: lock down the current transaction and wait for
 366          * all outstanding updates to complete.
 367          */
 368
 369 #ifdef COMMIT_STATS
 370         spin_lock(&journal->j_list_lock);
 371         summarise_journal_usage(journal);
 372         spin_unlock(&journal->j_list_lock);
 373 #endif
 374
 375         /* Do we need to erase the effects of a prior jbd2_journal_flush? */
 376         if (journal->j_flags & JBD2_FLUSHED) {
 377                 jbd_debug(3, "super block updated\n");
 378                 jbd2_journal_update_superblock(journal, 1);
 379         } else {
 380                 jbd_debug(3, "superblock not updated\n");
 381         }
 382
 383         J_ASSERT(journal->j_running_transaction != NULL);
 384         J_ASSERT(journal->j_committing_transaction == NULL);
 385
 386         commit_transaction = journal->j_running_transaction;
 387         J_ASSERT(commit_transaction->t_state == T_RUNNING);
 388
 389         trace_jbd2_start_commit(journal, commit_transaction);
 390         jbd_debug(1, "JBD: starting commit of transaction %d\n",
 391                         commit_transaction->t_tid);
 392
 393         write_lock(&journal->j_state_lock);
 394         commit_transaction->t_state = T_LOCKED;
 395
 396         /*
 397          * Use plugged writes here, since we want to submit several before
 398          * we unplug the device. We don't do explicit unplugging in here,
 399          * instead we rely on sync_buffer() doing the unplug for us.
 400          */
 401         if (commit_transaction->t_synchronous_commit)
 402                 write_op = WRITE_SYNC_PLUG;
 403         trace_jbd2_commit_locking(journal, commit_transaction);
 404         stats.run.rs_wait = commit_transaction->t_max_wait;
 405         stats.run.rs_locked = jiffies;
 406         stats.run.rs_running = jbd2_time_diff(commit_transaction->t_start,
 407                                               stats.run.rs_locked);
 408
 409         spin_lock(&commit_transaction->t_handle_lock);
 410         while (atomic_read(&commit_transaction->t_updates)) {
 411                 DEFINE_WAIT(wait);
 412
 413                 prepare_to_wait(&journal->j_wait_updates, &wait,
 414                                         TASK_UNINTERRUPTIBLE);
 415                 if (atomic_read(&commit_transaction->t_updates)) {
 416                         spin_unlock(&commit_transaction->t_handle_lock);
 417                         write_unlock(&journal->j_state_lock);
 418                         schedule();
 419                         write_lock(&journal->j_state_lock);
 420                         spin_lock(&commit_transaction->t_handle_lock);
 421                 }
 422                 finish_wait(&journal->j_wait_updates, &wait);
 423         }
 424         spin_unlock(&commit_transaction->t_handle_lock);
 425
 426         J_ASSERT (atomic_read(&commit_transaction->t_outstanding_credits) <=
 427                         journal->j_max_transaction_buffers);
 428
 429         /*
 430          * First thing we are allowed to do is to discard any remaining
 431          * BJ_Reserved buffers.  Note, it is _not_ permissible to assume
 432          * that there are no such buffers: if a large filesystem
 433          * operation like a truncate needs to split itself over multiple
 434          * transactions, then it may try to do a jbd2_journal_restart() while
 435          * there are still BJ_Reserved buffers outstanding.  These must
 436          * be released cleanly from the current transaction.
 437          *
 438          * In this case, the filesystem must still reserve write access
 439          * again before modifying the buffer in the new transaction, but
 440          * we do not require it to remember exactly which old buffers it
 441          * has reserved.  This is consistent with the existing behaviour
 442          * that multiple jbd2_journal_get_write_access() calls to the same
 443          * buffer are perfectly permissable.
 444          */
 445         while (commit_transaction->t_reserved_list) {
 446                 jh = commit_transaction->t_reserved_list;
 447                 JBUFFER_TRACE(jh, "reserved, unused: refile");
 448                 /*
 449                  * A jbd2_journal_get_undo_access()+jbd2_journal_release_buffer() may
 450                  * leave undo-committed data.
 451                  */
 452                 if (jh->b_committed_data) {
 453                         struct buffer_head *bh = jh2bh(jh);
 454
 455                         jbd_lock_bh_state(bh);
 456                         jbd2_free(jh->b_committed_data, bh->b_size);
 457                         jh->b_committed_data = NULL;
 458                         jbd_unlock_bh_state(bh);
 459                 }
 460                 jbd2_journal_refile_buffer(journal, jh);
 461         }
 462
 463         /*
 464          * Now try to drop any written-back buffers from the journal's
 465          * checkpoint lists.  We do this *before* commit because it potentially
 466          * frees some memory
 467          */
 468         spin_lock(&journal->j_list_lock);
 469         __jbd2_journal_clean_checkpoint_list(journal);
 470         spin_unlock(&journal->j_list_lock);
 471
 472         jbd_debug (3, "JBD: commit phase 1\n");
 473
 474         /*
 475          * Switch to a new revoke table.
 476          */
 477         jbd2_journal_switch_revoke_table(journal);
 478
 479         trace_jbd2_commit_flushing(journal, commit_transaction);
 480         stats.run.rs_flushing = jiffies;
 481         stats.run.rs_locked = jbd2_time_diff(stats.run.rs_locked,
 482                                              stats.run.rs_flushing);
 483
 484         commit_transaction->t_state = T_FLUSH;
 485         journal->j_committing_transaction = commit_transaction;
 486         journal->j_running_transaction = NULL;
 487         start_time = ktime_get();
 488         commit_transaction->t_log_start = journal->j_head;
 489         wake_up(&journal->j_wait_transaction_locked);
 490         write_unlock(&journal->j_state_lock);
 491
 492         jbd_debug (3, "JBD: commit phase 2\n");
 493
 494         /*
 495          * Now start flushing things to disk, in the order they appear
 496          * on the transaction lists.  Data blocks go first.
 497          */
 498         err = journal_submit_data_buffers(journal, commit_transaction);
 499         if (err)
 500                 jbd2_journal_abort(journal, err);
 501
 502         jbd2_journal_write_revoke_records(journal, commit_transaction,
 503                                           write_op);
 504
 505         jbd_debug(3, "JBD: commit phase 2\n");
 506
 507         /*
 508          * Way to go: we have now written out all of the data for a
 509          * transaction!  Now comes the tricky part: we need to write out
 510          * metadata.  Loop over the transaction's entire buffer list:
 511          */
 512         write_lock(&journal->j_state_lock);
 513         commit_transaction->t_state = T_COMMIT;
 514         write_unlock(&journal->j_state_lock);
 515
 516         trace_jbd2_commit_logging(journal, commit_transaction);
 517         stats.run.rs_logging = jiffies;
 518         stats.run.rs_flushing = jbd2_time_diff(stats.run.rs_flushing,
 519                                                stats.run.rs_logging);
 520         stats.run.rs_blocks =
 521                 atomic_read(&commit_transaction->t_outstanding_credits);
 522         stats.run.rs_blocks_logged = 0;
 523
 524         J_ASSERT(commit_transaction->t_nr_buffers <=
 525                  atomic_read(&commit_transaction->t_outstanding_credits));
 526
 527         err = 0;
 528         descriptor = NULL;
 529         bufs = 0;
 530         while (commit_transaction->t_buffers) {
 531
 532                 /* Find the next buffer to be journaled... */
 533
 534                 jh = commit_transaction->t_buffers;
 535
 536                 /* If we're in abort mode, we just un-journal the buffer and
 537                    release it. */
 538
 539                 if (is_journal_aborted(journal)) {
 540                         clear_buffer_jbddirty(jh2bh(jh));
 541                         JBUFFER_TRACE(jh, "journal is aborting: refile");
 542                         jbd2_buffer_abort_trigger(jh,
 543                                                   jh->b_frozen_data ?
 544                                                   jh->b_frozen_triggers :
 545                                                   jh->b_triggers);
 546                         jbd2_journal_refile_buffer(journal, jh);
 547                         /* If that was the last one, we need to clean up
 548                          * any descriptor buffers which may have been
 549                          * already allocated, even if we are now
 550                          * aborting. */
 551                         if (!commit_transaction->t_buffers)
 552                                 goto start_journal_io;
 553                         continue;
 554                 }
 555
 556                 /* Make sure we have a descriptor block in which to
 557                    record the metadata buffer. */
 558
 559                 if (!descriptor) {
 560                         struct buffer_head *bh;
 561
 562                         J_ASSERT (bufs == 0);
 563
 564                         jbd_debug(4, "JBD: get descriptor\n");
 565
 566                         descriptor = jbd2_journal_get_descriptor_buffer(journal);
 567                         if (!descriptor) {
 568                                 jbd2_journal_abort(journal, -EIO);
 569                                 continue;
 570                         }
 571
 572                         bh = jh2bh(descriptor);
 573                         jbd_debug(4, "JBD: got buffer %llu (%p)\n",
 574                                 (unsigned long long)bh->b_blocknr, bh->b_data);
 575                         header = (journal_header_t *)&bh->b_data[0];
 576                         header->h_magic     = cpu_to_be32(JBD2_MAGIC_NUMBER);
 577                         header->h_blocktype = cpu_to_be32(JBD2_DESCRIPTOR_BLOCK);
 578                         header->h_sequence  = cpu_to_be32(commit_transaction->t_tid);
 579
 580                         tagp = &bh->b_data[sizeof(journal_header_t)];
 581                         space_left = bh->b_size - sizeof(journal_header_t);
 582                         first_tag = 1;
 583                         set_buffer_jwrite(bh);
 584                         set_buffer_dirty(bh);
 585                         wbuf[bufs++] = bh;
 586
 587                         /* Record it so that we can wait for IO
 588                            completion later */
 589                         BUFFER_TRACE(bh, "ph3: file as descriptor");
 590                         jbd2_journal_file_buffer(descriptor, commit_transaction,
 591                                         BJ_LogCtl);
 592                 }
 593
 594                 /* Where is the buffer to be written? */
 595
 596                 err = jbd2_journal_next_log_block(journal, &blocknr);
 597                 /* If the block mapping failed, just abandon the buffer
 598                    and repeat this loop: we'll fall into the
 599                    refile-on-abort condition above. */
 600                 if (err) {
 601                         jbd2_journal_abort(journal, err);
 602                         continue;
 603                 }
 604
 605                 /*
 606                  * start_this_handle() uses t_outstanding_credits to determine
 607                  * the free space in the log, but this counter is changed
 608                  * by jbd2_journal_next_log_block() also.
 609                  */
 610                 atomic_dec(&commit_transaction->t_outstanding_credits);
 611
 612                 /* Bump b_count to prevent truncate from stumbling over
 613                    the shadowed buffer!  @@@ This can go if we ever get
 614                    rid of the BJ_IO/BJ_Shadow pairing of buffers. */
 615                 atomic_inc(&jh2bh(jh)->b_count);
 616
 617                 /* Make a temporary IO buffer with which to write it out
 618                    (this will requeue both the metadata buffer and the
 619                    temporary IO buffer). new_bh goes on BJ_IO*/
 620
 621                 set_bit(BH_JWrite, &jh2bh(jh)->b_state);
 622                 /*
 623                  * akpm: jbd2_journal_write_metadata_buffer() sets
 624                  * new_bh->b_transaction to commit_transaction.
 625                  * We need to clean this up before we release new_bh
 626                  * (which is of type BJ_IO)
 627                  */
 628                 JBUFFER_TRACE(jh, "ph3: write metadata");
 629                 flags = jbd2_journal_write_metadata_buffer(commit_transaction,
 630                                                       jh, &new_jh, blocknr);
 631                 if (flags < 0) {
 632                         jbd2_journal_abort(journal, flags);
 633                         continue;
 634                 }
 635                 set_bit(BH_JWrite, &jh2bh(new_jh)->b_state);
 636                 wbuf[bufs++] = jh2bh(new_jh);
 637
 638                 /* Record the new block's tag in the current descriptor
 639                    buffer */
 640
 641                 tag_flag = 0;
 642                 if (flags & 1)
 643                         tag_flag |= JBD2_FLAG_ESCAPE;
 644                 if (!first_tag)
 645                         tag_flag |= JBD2_FLAG_SAME_UUID;
 646
 647                 tag = (journal_block_tag_t *) tagp;
 648                 write_tag_block(tag_bytes, tag, jh2bh(jh)->b_blocknr);
 649                 tag->t_flags = cpu_to_be32(tag_flag);
 650                 tagp += tag_bytes;
 651                 space_left -= tag_bytes;
 652
 653                 if (first_tag) {
 654                         memcpy (tagp, journal->j_uuid, 16);
 655                         tagp += 16;
 656                         space_left -= 16;
 657                         first_tag = 0;
 658                 }
 659
 660                 /* If there's no more to do, or if the descriptor is full,
 661                    let the IO rip! */
 662
 663                 if (bufs == journal->j_wbufsize ||
 664                     commit_transaction->t_buffers == NULL ||
 665                     space_left < tag_bytes + 16) {
 666
 667                         jbd_debug(4, "JBD: Submit %d IOs\n", bufs);
 668
 669                         /* Write an end-of-descriptor marker before
 670                            submitting the IOs.  "tag" still points to
 671                            the last tag we set up. */
 672
 673                         tag->t_flags |= cpu_to_be32(JBD2_FLAG_LAST_TAG);
 674
 675 start_journal_io:
 676                         for (i = 0; i < bufs; i++) {
 677                                 struct buffer_head *bh = wbuf[i];
 678                                 /*
 679                                  * Compute checksum.
 680                                  */
 681                                 if (JBD2_HAS_COMPAT_FEATURE(journal,
 682                                         JBD2_FEATURE_COMPAT_CHECKSUM)) {
 683                                         crc32_sum =
 684                                             jbd2_checksum_data(crc32_sum, bh);
 685                                 }
 686
 687                                 lock_buffer(bh);
 688                                 clear_buffer_dirty(bh);
 689                                 set_buffer_uptodate(bh);
 690                                 bh->b_end_io = journal_end_buffer_io_sync;
 691                                 submit_bh(write_op, bh);
 692                         }
 693                         cond_resched();
 694                         stats.run.rs_blocks_logged += bufs;
 695
 696                         /* Force a new descriptor to be generated next
 697                            time round the loop. */
 698                         descriptor = NULL;
 699                         bufs = 0;
 700                 }
 701         }
 702
 703         write_lock(&journal->j_state_lock);
 704         J_ASSERT(commit_transaction->t_state == T_COMMIT);
 705         commit_transaction->t_state = T_COMMIT_DFLUSH;
 706         write_unlock(&journal->j_state_lock);
 707
 708         /*
 709          * If the journal is not located on the file system device,
 710          * then we must flush the file system device before we issue
 711          * the commit record
 712          */
 713         if (commit_transaction->t_need_data_flush &&
 714             (journal->j_fs_dev != journal->j_dev) &&
 715             (journal->j_flags & JBD2_BARRIER))
 716                 blkdev_issue_flush(journal->j_fs_dev, GFP_KERNEL, NULL,
 717                         BLKDEV_IFL_WAIT);
 718
 719         /* Done it all: now write the commit record asynchronously. */
 720         if (JBD2_HAS_INCOMPAT_FEATURE(journal,
 721                                       JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
 722                 err = journal_submit_commit_record(journal, commit_transaction,
 723                                                  &cbh, crc32_sum);
 724                 if (err)
 725                         __jbd2_journal_abort_hard(journal);
 726                 if (journal->j_flags & JBD2_BARRIER)
 727                         blkdev_issue_flush(journal->j_dev, GFP_KERNEL, NULL,
 728                                 BLKDEV_IFL_WAIT);
 729         }
 730
 731         err = journal_finish_inode_data_buffers(journal, commit_transaction);
 732         if (err) {
 733                 printk(KERN_WARNING
 734                         "JBD2: Detected IO errors while flushing file data "
 735                        "on %s\n", journal->j_devname);
 736                 if (journal->j_flags & JBD2_ABORT_ON_SYNCDATA_ERR)
 737                         jbd2_journal_abort(journal, err);
 738                 err = 0;
 739         }
 740
 741         /* Lo and behold: we have just managed to send a transaction to
 742            the log.  Before we can commit it, wait for the IO so far to
 743            complete.  Control buffers being written are on the
 744            transaction's t_log_list queue, and metadata buffers are on
 745            the t_iobuf_list queue.
 746
 747            Wait for the buffers in reverse order.  That way we are
 748            less likely to be woken up until all IOs have completed, and
 749            so we incur less scheduling load.
 750         */
 751
 752         jbd_debug(3, "JBD: commit phase 3\n");
 753
 754         /*
 755          * akpm: these are BJ_IO, and j_list_lock is not needed.
 756          * See __journal_try_to_free_buffer.
 757          */
 758 wait_for_iobuf:
 759         while (commit_transaction->t_iobuf_list != NULL) {
 760                 struct buffer_head *bh;
 761
 762                 jh = commit_transaction->t_iobuf_list->b_tprev;
 763                 bh = jh2bh(jh);
 764                 if (buffer_locked(bh)) {
 765                         wait_on_buffer(bh);
 766                         goto wait_for_iobuf;
 767                 }
 768                 if (cond_resched())
 769                         goto wait_for_iobuf;
 770
 771                 if (unlikely(!buffer_uptodate(bh)))
 772                         err = -EIO;
 773
 774                 clear_buffer_jwrite(bh);
 775
 776                 JBUFFER_TRACE(jh, "ph4: unfile after journal write");
 777                 jbd2_journal_unfile_buffer(journal, jh);
 778
 779                 /*
 780                  * ->t_iobuf_list should contain only dummy buffer_heads
 781                  * which were created by jbd2_journal_write_metadata_buffer().
 782                  */
 783                 BUFFER_TRACE(bh, "dumping temporary bh");
 784                 jbd2_journal_put_journal_head(jh);
 785                 __brelse(bh);
 786                 J_ASSERT_BH(bh, atomic_read(&bh->b_count) == 0);
 787                 free_buffer_head(bh);
 788
 789                 /* We also have to unlock and free the corresponding
 790                    shadowed buffer */
 791                 jh = commit_transaction->t_shadow_list->b_tprev;
 792                 bh = jh2bh(jh);
 793                 clear_bit(BH_JWrite, &bh->b_state);
 794                 J_ASSERT_BH(bh, buffer_jbddirty(bh));
 795
 796                 /* The metadata is now released for reuse, but we need
 797                    to remember it against this transaction so that when
 798                    we finally commit, we can do any checkpointing
 799                    required. */
 800                 JBUFFER_TRACE(jh, "file as BJ_Forget");
 801                 jbd2_journal_file_buffer(jh, commit_transaction, BJ_Forget);
 802                 /* Wake up any transactions which were waiting for this
 803                    IO to complete */
 804                 wake_up_bit(&bh->b_state, BH_Unshadow);
 805                 JBUFFER_TRACE(jh, "brelse shadowed buffer");
 806                 __brelse(bh);
 807         }
 808
 809         J_ASSERT (commit_transaction->t_shadow_list == NULL);
 810
 811         jbd_debug(3, "JBD: commit phase 4\n");
 812
 813         /* Here we wait for the revoke record and descriptor record buffers */
 814  wait_for_ctlbuf:
 815         while (commit_transaction->t_log_list != NULL) {
 816                 struct buffer_head *bh;
 817
 818                 jh = commit_transaction->t_log_list->b_tprev;
 819                 bh = jh2bh(jh);
 820                 if (buffer_locked(bh)) {
 821                         wait_on_buffer(bh);
 822                         goto wait_for_ctlbuf;
 823                 }
 824                 if (cond_resched())
 825                         goto wait_for_ctlbuf;
 826
 827                 if (unlikely(!buffer_uptodate(bh)))
 828                         err = -EIO;
 829
 830                 BUFFER_TRACE(bh, "ph5: control buffer writeout done: unfile");
 831                 clear_buffer_jwrite(bh);
 832                 jbd2_journal_unfile_buffer(journal, jh);
 833                 jbd2_journal_put_journal_head(jh);
 834                 __brelse(bh);           /* One for getblk */
 835                 /* AKPM: bforget here */
 836         }
 837
 838         if (err)
 839                 jbd2_journal_abort(journal, err);
 840
 841         jbd_debug(3, "JBD: commit phase 5\n");
 842         write_lock(&journal->j_state_lock);
 843         J_ASSERT(commit_transaction->t_state == T_COMMIT_DFLUSH);
 844         commit_transaction->t_state = T_COMMIT_JFLUSH;
 845         write_unlock(&journal->j_state_lock);
 846
 847         if (!JBD2_HAS_INCOMPAT_FEATURE(journal,
 848                                        JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
 849                 err = journal_submit_commit_record(journal, commit_transaction,
 850                                                 &cbh, crc32_sum);
 851                 if (err)
 852                         __jbd2_journal_abort_hard(journal);
 853         }
 854         if (!err && !is_journal_aborted(journal))
 855                 err = journal_wait_on_commit_record(journal, cbh);
 856
 857         if (err)
 858                 jbd2_journal_abort(journal, err);
 859
 860         /* End of a transaction!  Finally, we can do checkpoint
 861            processing: any buffers committed as a result of this
 862            transaction can be removed from any checkpoint list it was on
 863            before. */
 864
 865         jbd_debug(3, "JBD: commit phase 6\n");
 866
 867         J_ASSERT(list_empty(&commit_transaction->t_inode_list));
 868         J_ASSERT(commit_transaction->t_buffers == NULL);
 869         J_ASSERT(commit_transaction->t_checkpoint_list == NULL);
 870         J_ASSERT(commit_transaction->t_iobuf_list == NULL);
 871         J_ASSERT(commit_transaction->t_shadow_list == NULL);
 872         J_ASSERT(commit_transaction->t_log_list == NULL);
 873
 874 restart_loop:
 875         /*
 876          * As there are other places (journal_unmap_buffer()) adding buffers
 877          * to this list we have to be careful and hold the j_list_lock.
 878          */
 879         spin_lock(&journal->j_list_lock);
 880         while (commit_transaction->t_forget) {
 881                 transaction_t *cp_transaction;
 882                 struct buffer_head *bh;
 883
 884                 jh = commit_transaction->t_forget;
 885                 spin_unlock(&journal->j_list_lock);
 886                 bh = jh2bh(jh);
 887                 jbd_lock_bh_state(bh);
 888                 J_ASSERT_JH(jh, jh->b_transaction == commit_transaction);
 889
 890                 /*
 891                  * If there is undo-protected committed data against
 892                  * this buffer, then we can remove it now.  If it is a
 893                  * buffer needing such protection, the old frozen_data
 894                  * field now points to a committed version of the
 895                  * buffer, so rotate that field to the new committed
 896                  * data.
 897                  *
 898                  * Otherwise, we can just throw away the frozen data now.
 899                  *
 900                  * We also know that the frozen data has already fired
 901                  * its triggers if they exist, so we can clear that too.
 902                  */
 903                 if (jh->b_committed_data) {
 904                         jbd2_free(jh->b_committed_data, bh->b_size);
 905                         jh->b_committed_data = NULL;
 906                         if (jh->b_frozen_data) {
 907                                 jh->b_committed_data = jh->b_frozen_data;
 908                                 jh->b_frozen_data = NULL;
 909                                 jh->b_frozen_triggers = NULL;
 910                         }
 911                 } else if (jh->b_frozen_data) {
 912                         jbd2_free(jh->b_frozen_data, bh->b_size);
 913                         jh->b_frozen_data = NULL;
 914                         jh->b_frozen_triggers = NULL;
 915                 }
 916
 917                 spin_lock(&journal->j_list_lock);
 918                 cp_transaction = jh->b_cp_transaction;
 919                 if (cp_transaction) {
 920                         JBUFFER_TRACE(jh, "remove from old cp transaction");
 921                         cp_transaction->t_chp_stats.cs_dropped++;
 922                         __jbd2_journal_remove_checkpoint(jh);
 923                 }
 924
 925                 /* Only re-checkpoint the buffer_head if it is marked
 926                  * dirty.  If the buffer was added to the BJ_Forget list
 927                  * by jbd2_journal_forget, it may no longer be dirty and
 928                  * there's no point in keeping a checkpoint record for
 929                  * it. */
 930
 931                 /* A buffer which has been freed while still being
 932                  * journaled by a previous transaction may end up still
 933                  * being dirty here, but we want to avoid writing back
 934                  * that buffer in the future after the "add to orphan"
 935                  * operation been committed,  That's not only a performance
 936                  * gain, it also stops aliasing problems if the buffer is
 937                  * left behind for writeback and gets reallocated for another
 938                  * use in a different page. */
 939                 if (buffer_freed(bh) && !jh->b_next_transaction) {
 940                         clear_buffer_freed(bh);
 941                         clear_buffer_jbddirty(bh);
 942                 }
 943
 944                 if (buffer_jbddirty(bh)) {
 945                         JBUFFER_TRACE(jh, "add to new checkpointing trans");
 946                         __jbd2_journal_insert_checkpoint(jh, commit_transaction);
 947                         if (is_journal_aborted(journal))
 948                                 clear_buffer_jbddirty(bh);
 949                         JBUFFER_TRACE(jh, "refile for checkpoint writeback");
 950                         __jbd2_journal_refile_buffer(jh);
 951                         jbd_unlock_bh_state(bh);
 952                 } else {
 953                         J_ASSERT_BH(bh, !buffer_dirty(bh));
 954                         /* The buffer on BJ_Forget list and not jbddirty means
 955                          * it has been freed by this transaction and hence it
 956                          * could not have been reallocated until this
 957                          * transaction has committed. *BUT* it could be
 958                          * reallocated once we have written all the data to
 959                          * disk and before we process the buffer on BJ_Forget
 960                          * list. */
 961                         JBUFFER_TRACE(jh, "refile or unfile freed buffer");
 962                         __jbd2_journal_refile_buffer(jh);
 963                         if (!jh->b_transaction) {
 964                                 jbd_unlock_bh_state(bh);
 965                                  /* needs a brelse */
 966                                 jbd2_journal_remove_journal_head(bh);
 967                                 release_buffer_page(bh);
 968                         } else
 969                                 jbd_unlock_bh_state(bh);
 970                 }
 971                 cond_resched_lock(&journal->j_list_lock);
 972         }
 973         spin_unlock(&journal->j_list_lock);
 974         /*
 975          * This is a bit sleazy.  We use j_list_lock to protect transition
 976          * of a transaction into T_FINISHED state and calling
 977          * __jbd2_journal_drop_transaction(). Otherwise we could race with
 978          * other checkpointing code processing the transaction...
 979          */
 980         write_lock(&journal->j_state_lock);
 981         spin_lock(&journal->j_list_lock);
 982         /*
 983          * Now recheck if some buffers did not get attached to the transaction
 984          * while the lock was dropped...
 985          */
 986         if (commit_transaction->t_forget) {
 987                 spin_unlock(&journal->j_list_lock);
 988                 write_unlock(&journal->j_state_lock);
 989                 goto restart_loop;
 990         }
 991
 992         /* Done with this transaction! */
 993
 994         jbd_debug(3, "JBD: commit phase 7\n");
 995
 996         J_ASSERT(commit_transaction->t_state == T_COMMIT_JFLUSH);
 997
 998         commit_transaction->t_start = jiffies;
 999         stats.run.rs_logging = jbd2_time_diff(stats.run.rs_logging,
1000                                               commit_transaction->t_start);
1001
1002         /*
1003          * File the transaction statistics
1004          */
1005         stats.ts_tid = commit_transaction->t_tid;
1006         stats.run.rs_handle_count =
1007                 atomic_read(&commit_transaction->t_handle_count);
1008         trace_jbd2_run_stats(journal->j_fs_dev->bd_dev,
1009                              commit_transaction->t_tid, &stats.run);
1010
1011         /*
1012          * Calculate overall stats
1013          */
1014         spin_lock(&journal->j_history_lock);
1015         journal->j_stats.ts_tid++;
1016         journal->j_stats.run.rs_wait += stats.run.rs_wait;
1017         journal->j_stats.run.rs_running += stats.run.rs_running;
1018         journal->j_stats.run.rs_locked += stats.run.rs_locked;
1019         journal->j_stats.run.rs_flushing += stats.run.rs_flushing;
1020         journal->j_stats.run.rs_logging += stats.run.rs_logging;
1021         journal->j_stats.run.rs_handle_count += stats.run.rs_handle_count;
1022         journal->j_stats.run.rs_blocks += stats.run.rs_blocks;
1023         journal->j_stats.run.rs_blocks_logged += stats.run.rs_blocks_logged;
1024         spin_unlock(&journal->j_history_lock);
1025
1026         commit_transaction->t_state = T_FINISHED;
1027         J_ASSERT(commit_transaction == journal->j_committing_transaction);
1028         journal->j_commit_sequence = commit_transaction->t_tid;
1029         journal->j_committing_transaction = NULL;
1030         commit_time = ktime_to_ns(ktime_sub(ktime_get(), start_time));
1031
1032         /*
1033          * weight the commit time higher than the average time so we don't
1034          * react too strongly to vast changes in the commit time
1035          */
1036         if (likely(journal->j_average_commit_time))
1037                 journal->j_average_commit_time = (commit_time +
1038                                 journal->j_average_commit_time*3) / 4;
1039         else
1040                 journal->j_average_commit_time = commit_time;
1041         write_unlock(&journal->j_state_lock);
1042
1043         if (commit_transaction->t_checkpoint_list == NULL &&
1044             commit_transaction->t_checkpoint_io_list == NULL) {
1045                 __jbd2_journal_drop_transaction(journal, commit_transaction);
1046                 to_free = 1;
1047         } else {
1048                 if (journal->j_checkpoint_transactions == NULL) {
1049                         journal->j_checkpoint_transactions = commit_transaction;
1050                         commit_transaction->t_cpnext = commit_transaction;
1051                         commit_transaction->t_cpprev = commit_transaction;
1052                 } else {
1053                         commit_transaction->t_cpnext =
1054                                 journal->j_checkpoint_transactions;
1055                         commit_transaction->t_cpprev =
1056                                 commit_transaction->t_cpnext->t_cpprev;
1057                         commit_transaction->t_cpnext->t_cpprev =
1058                                 commit_transaction;
1059                         commit_transaction->t_cpprev->t_cpnext =
1060                                 commit_transaction;
1061                 }
1062         }
1063         spin_unlock(&journal->j_list_lock);
1064
1065         if (journal->j_commit_callback)
1066                 journal->j_commit_callback(journal, commit_transaction);
1067
1068         trace_jbd2_end_commit(journal, commit_transaction);
1069         jbd_debug(1, "JBD: commit %d complete, head %d\n",
1070                   journal->j_commit_sequence, journal->j_tail_sequence);
1071         if (to_free)
1072                 kfree(commit_transaction);
1073
1074         wake_up(&journal->j_wait_done_commit);
1075 }