2 * linux/fs/jbd/commit.c
4 * Written by Stephen C. Tweedie <sct@redhat.com>, 1998
6 * Copyright 1998 Red Hat corp --- All Rights Reserved
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.
12 * Journal commit routines for the generic filesystem journaling code;
13 * part of the ext2fs journaling system.
16 #include <linux/time.h>
18 #include <linux/jbd.h>
19 #include <linux/errno.h>
20 #include <linux/slab.h>
22 #include <linux/pagemap.h>
25 * Default IO end handler for temporary BJ_IO buffer_heads.
27 static void journal_end_buffer_io_sync(struct buffer_head *bh, int uptodate)
31 set_buffer_uptodate(bh);
33 clear_buffer_uptodate(bh);
38 * When an ext3-ordered file is truncated, it is possible that many pages are
39 * not successfully freed, because they are attached to a committing transaction.
40 * After the transaction commits, these pages are left on the LRU, with no
41 * ->mapping, and with attached buffers. These pages are trivially reclaimable
42 * by the VM, but their apparent absence upsets the VM accounting, and it makes
43 * the numbers in /proc/meminfo look odd.
45 * So here, we have a buffer which has just come off the forget list. Look to
46 * see if we can strip all buffers from the backing page.
48 * Called under journal->j_list_lock. The caller provided us with a ref
49 * against the buffer, and we drop that here.
51 static void release_buffer_page(struct buffer_head *bh)
57 if (atomic_read(&bh->b_count) != 1)
65 /* OK, it's a truncated page */
66 if (!trylock_page(page))
71 try_to_free_buffers(page);
73 page_cache_release(page);
81 * Decrement reference counter for data buffer. If it has been marked
82 * 'BH_Freed', release it and the page to which it belongs if possible.
84 static void release_data_buffer(struct buffer_head *bh)
86 if (buffer_freed(bh)) {
87 clear_buffer_freed(bh);
88 release_buffer_page(bh);
94 * Try to acquire jbd_lock_bh_state() against the buffer, when j_list_lock is
95 * held. For ranking reasons we must trylock. If we lose, schedule away and
96 * return 0. j_list_lock is dropped in this case.
98 static int inverted_lock(journal_t *journal, struct buffer_head *bh)
100 if (!jbd_trylock_bh_state(bh)) {
101 spin_unlock(&journal->j_list_lock);
108 /* Done it all: now write the commit record. We should have
109 * cleaned up our previous buffers by now, so if we are in abort
110 * mode we can now just skip the rest of the journal write
113 * Returns 1 if the journal needs to be aborted or 0 on success
115 static int journal_write_commit_record(journal_t *journal,
116 transaction_t *commit_transaction)
118 struct journal_head *descriptor;
119 struct buffer_head *bh;
120 journal_header_t *header;
122 int barrier_done = 0;
124 if (is_journal_aborted(journal))
127 descriptor = journal_get_descriptor_buffer(journal);
131 bh = jh2bh(descriptor);
133 header = (journal_header_t *)(bh->b_data);
134 header->h_magic = cpu_to_be32(JFS_MAGIC_NUMBER);
135 header->h_blocktype = cpu_to_be32(JFS_COMMIT_BLOCK);
136 header->h_sequence = cpu_to_be32(commit_transaction->t_tid);
138 JBUFFER_TRACE(descriptor, "write commit block");
139 set_buffer_dirty(bh);
140 if (journal->j_flags & JFS_BARRIER) {
141 set_buffer_ordered(bh);
144 ret = sync_dirty_buffer(bh);
146 clear_buffer_ordered(bh);
147 /* is it possible for another commit to fail at roughly
148 * the same time as this one? If so, we don't want to
149 * trust the barrier flag in the super, but instead want
150 * to remember if we sent a barrier request
152 if (ret == -EOPNOTSUPP && barrier_done) {
153 char b[BDEVNAME_SIZE];
156 "JBD: barrier-based sync failed on %s - "
157 "disabling barriers\n",
158 bdevname(journal->j_dev, b));
159 spin_lock(&journal->j_state_lock);
160 journal->j_flags &= ~JFS_BARRIER;
161 spin_unlock(&journal->j_state_lock);
163 /* And try again, without the barrier */
164 set_buffer_uptodate(bh);
165 set_buffer_dirty(bh);
166 ret = sync_dirty_buffer(bh);
168 put_bh(bh); /* One for getblk() */
169 journal_put_journal_head(descriptor);
171 return (ret == -EIO);
174 static void journal_do_submit_data(struct buffer_head **wbuf, int bufs)
178 for (i = 0; i < bufs; i++) {
179 wbuf[i]->b_end_io = end_buffer_write_sync;
180 /* We use-up our safety reference in submit_bh() */
181 submit_bh(WRITE, wbuf[i]);
186 * Submit all the data buffers to disk
188 static int journal_submit_data_buffers(journal_t *journal,
189 transaction_t *commit_transaction)
191 struct journal_head *jh;
192 struct buffer_head *bh;
195 struct buffer_head **wbuf = journal->j_wbuf;
199 * Whenever we unlock the journal and sleep, things can get added
200 * onto ->t_sync_datalist, so we have to keep looping back to
201 * write_out_data until we *know* that the list is empty.
203 * Cleanup any flushed data buffers from the data list. Even in
204 * abort mode, we want to flush this out as soon as possible.
208 spin_lock(&journal->j_list_lock);
210 while (commit_transaction->t_sync_datalist) {
211 jh = commit_transaction->t_sync_datalist;
215 /* Get reference just to make sure buffer does not disappear
216 * when we are forced to drop various locks */
218 /* If the buffer is dirty, we need to submit IO and hence
219 * we need the buffer lock. We try to lock the buffer without
220 * blocking. If we fail, we need to drop j_list_lock and do
221 * blocking lock_buffer().
223 if (buffer_dirty(bh)) {
224 if (!trylock_buffer(bh)) {
225 BUFFER_TRACE(bh, "needs blocking lock");
226 spin_unlock(&journal->j_list_lock);
227 /* Write out all data to prevent deadlocks */
228 journal_do_submit_data(wbuf, bufs);
231 spin_lock(&journal->j_list_lock);
235 /* We have to get bh_state lock. Again out of order, sigh. */
236 if (!inverted_lock(journal, bh)) {
237 jbd_lock_bh_state(bh);
238 spin_lock(&journal->j_list_lock);
240 /* Someone already cleaned up the buffer? */
242 || jh->b_transaction != commit_transaction
243 || jh->b_jlist != BJ_SyncData) {
244 jbd_unlock_bh_state(bh);
247 BUFFER_TRACE(bh, "already cleaned up");
248 release_data_buffer(bh);
251 if (locked && test_clear_buffer_dirty(bh)) {
252 BUFFER_TRACE(bh, "needs writeout, adding to array");
254 __journal_file_buffer(jh, commit_transaction,
256 jbd_unlock_bh_state(bh);
257 if (bufs == journal->j_wbufsize) {
258 spin_unlock(&journal->j_list_lock);
259 journal_do_submit_data(wbuf, bufs);
263 } else if (!locked && buffer_locked(bh)) {
264 __journal_file_buffer(jh, commit_transaction,
266 jbd_unlock_bh_state(bh);
269 BUFFER_TRACE(bh, "writeout complete: unfile");
270 if (unlikely(!buffer_uptodate(bh)))
272 __journal_unfile_buffer(jh);
273 jbd_unlock_bh_state(bh);
276 journal_remove_journal_head(bh);
277 /* One for our safety reference, other for
278 * journal_remove_journal_head() */
280 release_data_buffer(bh);
283 if (need_resched() || spin_needbreak(&journal->j_list_lock)) {
284 spin_unlock(&journal->j_list_lock);
288 spin_unlock(&journal->j_list_lock);
289 journal_do_submit_data(wbuf, bufs);
295 * journal_commit_transaction
297 * The primary function for committing a transaction to the log. This
298 * function is called by the journal thread to begin a complete commit.
300 void journal_commit_transaction(journal_t *journal)
302 transaction_t *commit_transaction;
303 struct journal_head *jh, *new_jh, *descriptor;
304 struct buffer_head **wbuf = journal->j_wbuf;
308 unsigned long blocknr;
312 journal_header_t *header;
313 journal_block_tag_t *tag = NULL;
320 * First job: lock down the current transaction and wait for
321 * all outstanding updates to complete.
325 spin_lock(&journal->j_list_lock);
326 summarise_journal_usage(journal);
327 spin_unlock(&journal->j_list_lock);
330 /* Do we need to erase the effects of a prior journal_flush? */
331 if (journal->j_flags & JFS_FLUSHED) {
332 jbd_debug(3, "super block updated\n");
333 journal_update_superblock(journal, 1);
335 jbd_debug(3, "superblock not updated\n");
338 J_ASSERT(journal->j_running_transaction != NULL);
339 J_ASSERT(journal->j_committing_transaction == NULL);
341 commit_transaction = journal->j_running_transaction;
342 J_ASSERT(commit_transaction->t_state == T_RUNNING);
344 jbd_debug(1, "JBD: starting commit of transaction %d\n",
345 commit_transaction->t_tid);
347 spin_lock(&journal->j_state_lock);
348 commit_transaction->t_state = T_LOCKED;
350 spin_lock(&commit_transaction->t_handle_lock);
351 while (commit_transaction->t_updates) {
354 prepare_to_wait(&journal->j_wait_updates, &wait,
355 TASK_UNINTERRUPTIBLE);
356 if (commit_transaction->t_updates) {
357 spin_unlock(&commit_transaction->t_handle_lock);
358 spin_unlock(&journal->j_state_lock);
360 spin_lock(&journal->j_state_lock);
361 spin_lock(&commit_transaction->t_handle_lock);
363 finish_wait(&journal->j_wait_updates, &wait);
365 spin_unlock(&commit_transaction->t_handle_lock);
367 J_ASSERT (commit_transaction->t_outstanding_credits <=
368 journal->j_max_transaction_buffers);
371 * First thing we are allowed to do is to discard any remaining
372 * BJ_Reserved buffers. Note, it is _not_ permissible to assume
373 * that there are no such buffers: if a large filesystem
374 * operation like a truncate needs to split itself over multiple
375 * transactions, then it may try to do a journal_restart() while
376 * there are still BJ_Reserved buffers outstanding. These must
377 * be released cleanly from the current transaction.
379 * In this case, the filesystem must still reserve write access
380 * again before modifying the buffer in the new transaction, but
381 * we do not require it to remember exactly which old buffers it
382 * has reserved. This is consistent with the existing behaviour
383 * that multiple journal_get_write_access() calls to the same
384 * buffer are perfectly permissable.
386 while (commit_transaction->t_reserved_list) {
387 jh = commit_transaction->t_reserved_list;
388 JBUFFER_TRACE(jh, "reserved, unused: refile");
390 * A journal_get_undo_access()+journal_release_buffer() may
391 * leave undo-committed data.
393 if (jh->b_committed_data) {
394 struct buffer_head *bh = jh2bh(jh);
396 jbd_lock_bh_state(bh);
397 jbd_free(jh->b_committed_data, bh->b_size);
398 jh->b_committed_data = NULL;
399 jbd_unlock_bh_state(bh);
401 journal_refile_buffer(journal, jh);
405 * Now try to drop any written-back buffers from the journal's
406 * checkpoint lists. We do this *before* commit because it potentially
409 spin_lock(&journal->j_list_lock);
410 __journal_clean_checkpoint_list(journal);
411 spin_unlock(&journal->j_list_lock);
413 jbd_debug (3, "JBD: commit phase 1\n");
416 * Switch to a new revoke table.
418 journal_switch_revoke_table(journal);
420 commit_transaction->t_state = T_FLUSH;
421 journal->j_committing_transaction = commit_transaction;
422 journal->j_running_transaction = NULL;
423 start_time = ktime_get();
424 commit_transaction->t_log_start = journal->j_head;
425 wake_up(&journal->j_wait_transaction_locked);
426 spin_unlock(&journal->j_state_lock);
428 jbd_debug (3, "JBD: commit phase 2\n");
431 * Now start flushing things to disk, in the order they appear
432 * on the transaction lists. Data blocks go first.
434 err = journal_submit_data_buffers(journal, commit_transaction);
437 * Wait for all previously submitted IO to complete.
439 spin_lock(&journal->j_list_lock);
440 while (commit_transaction->t_locked_list) {
441 struct buffer_head *bh;
443 jh = commit_transaction->t_locked_list->b_tprev;
446 if (buffer_locked(bh)) {
447 spin_unlock(&journal->j_list_lock);
449 spin_lock(&journal->j_list_lock);
451 if (unlikely(!buffer_uptodate(bh))) {
452 if (!trylock_page(bh->b_page)) {
453 spin_unlock(&journal->j_list_lock);
454 lock_page(bh->b_page);
455 spin_lock(&journal->j_list_lock);
457 if (bh->b_page->mapping)
458 set_bit(AS_EIO, &bh->b_page->mapping->flags);
460 unlock_page(bh->b_page);
461 SetPageError(bh->b_page);
464 if (!inverted_lock(journal, bh)) {
466 spin_lock(&journal->j_list_lock);
469 if (buffer_jbd(bh) && jh->b_jlist == BJ_Locked) {
470 __journal_unfile_buffer(jh);
471 jbd_unlock_bh_state(bh);
472 journal_remove_journal_head(bh);
475 jbd_unlock_bh_state(bh);
477 release_data_buffer(bh);
478 cond_resched_lock(&journal->j_list_lock);
480 spin_unlock(&journal->j_list_lock);
483 char b[BDEVNAME_SIZE];
486 "JBD: Detected IO errors while flushing file data "
487 "on %s\n", bdevname(journal->j_fs_dev, b));
488 if (journal->j_flags & JFS_ABORT_ON_SYNCDATA_ERR)
489 journal_abort(journal, err);
493 journal_write_revoke_records(journal, commit_transaction);
496 * If we found any dirty or locked buffers, then we should have
497 * looped back up to the write_out_data label. If there weren't
498 * any then journal_clean_data_list should have wiped the list
499 * clean by now, so check that it is in fact empty.
501 J_ASSERT (commit_transaction->t_sync_datalist == NULL);
503 jbd_debug (3, "JBD: commit phase 3\n");
506 * Way to go: we have now written out all of the data for a
507 * transaction! Now comes the tricky part: we need to write out
508 * metadata. Loop over the transaction's entire buffer list:
510 spin_lock(&journal->j_state_lock);
511 commit_transaction->t_state = T_COMMIT;
512 spin_unlock(&journal->j_state_lock);
514 J_ASSERT(commit_transaction->t_nr_buffers <=
515 commit_transaction->t_outstanding_credits);
519 while (commit_transaction->t_buffers) {
521 /* Find the next buffer to be journaled... */
523 jh = commit_transaction->t_buffers;
525 /* If we're in abort mode, we just un-journal the buffer and
528 if (is_journal_aborted(journal)) {
529 clear_buffer_jbddirty(jh2bh(jh));
530 JBUFFER_TRACE(jh, "journal is aborting: refile");
531 journal_refile_buffer(journal, jh);
532 /* If that was the last one, we need to clean up
533 * any descriptor buffers which may have been
534 * already allocated, even if we are now
536 if (!commit_transaction->t_buffers)
537 goto start_journal_io;
541 /* Make sure we have a descriptor block in which to
542 record the metadata buffer. */
545 struct buffer_head *bh;
547 J_ASSERT (bufs == 0);
549 jbd_debug(4, "JBD: get descriptor\n");
551 descriptor = journal_get_descriptor_buffer(journal);
553 journal_abort(journal, -EIO);
557 bh = jh2bh(descriptor);
558 jbd_debug(4, "JBD: got buffer %llu (%p)\n",
559 (unsigned long long)bh->b_blocknr, bh->b_data);
560 header = (journal_header_t *)&bh->b_data[0];
561 header->h_magic = cpu_to_be32(JFS_MAGIC_NUMBER);
562 header->h_blocktype = cpu_to_be32(JFS_DESCRIPTOR_BLOCK);
563 header->h_sequence = cpu_to_be32(commit_transaction->t_tid);
565 tagp = &bh->b_data[sizeof(journal_header_t)];
566 space_left = bh->b_size - sizeof(journal_header_t);
568 set_buffer_jwrite(bh);
569 set_buffer_dirty(bh);
572 /* Record it so that we can wait for IO
574 BUFFER_TRACE(bh, "ph3: file as descriptor");
575 journal_file_buffer(descriptor, commit_transaction,
579 /* Where is the buffer to be written? */
581 err = journal_next_log_block(journal, &blocknr);
582 /* If the block mapping failed, just abandon the buffer
583 and repeat this loop: we'll fall into the
584 refile-on-abort condition above. */
586 journal_abort(journal, err);
591 * start_this_handle() uses t_outstanding_credits to determine
592 * the free space in the log, but this counter is changed
593 * by journal_next_log_block() also.
595 commit_transaction->t_outstanding_credits--;
597 /* Bump b_count to prevent truncate from stumbling over
598 the shadowed buffer! @@@ This can go if we ever get
599 rid of the BJ_IO/BJ_Shadow pairing of buffers. */
600 atomic_inc(&jh2bh(jh)->b_count);
602 /* Make a temporary IO buffer with which to write it out
603 (this will requeue both the metadata buffer and the
604 temporary IO buffer). new_bh goes on BJ_IO*/
606 set_bit(BH_JWrite, &jh2bh(jh)->b_state);
608 * akpm: journal_write_metadata_buffer() sets
609 * new_bh->b_transaction to commit_transaction.
610 * We need to clean this up before we release new_bh
611 * (which is of type BJ_IO)
613 JBUFFER_TRACE(jh, "ph3: write metadata");
614 flags = journal_write_metadata_buffer(commit_transaction,
615 jh, &new_jh, blocknr);
616 set_bit(BH_JWrite, &jh2bh(new_jh)->b_state);
617 wbuf[bufs++] = jh2bh(new_jh);
619 /* Record the new block's tag in the current descriptor
624 tag_flag |= JFS_FLAG_ESCAPE;
626 tag_flag |= JFS_FLAG_SAME_UUID;
628 tag = (journal_block_tag_t *) tagp;
629 tag->t_blocknr = cpu_to_be32(jh2bh(jh)->b_blocknr);
630 tag->t_flags = cpu_to_be32(tag_flag);
631 tagp += sizeof(journal_block_tag_t);
632 space_left -= sizeof(journal_block_tag_t);
635 memcpy (tagp, journal->j_uuid, 16);
641 /* If there's no more to do, or if the descriptor is full,
644 if (bufs == journal->j_wbufsize ||
645 commit_transaction->t_buffers == NULL ||
646 space_left < sizeof(journal_block_tag_t) + 16) {
648 jbd_debug(4, "JBD: Submit %d IOs\n", bufs);
650 /* Write an end-of-descriptor marker before
651 submitting the IOs. "tag" still points to
652 the last tag we set up. */
654 tag->t_flags |= cpu_to_be32(JFS_FLAG_LAST_TAG);
657 for (i = 0; i < bufs; i++) {
658 struct buffer_head *bh = wbuf[i];
660 clear_buffer_dirty(bh);
661 set_buffer_uptodate(bh);
662 bh->b_end_io = journal_end_buffer_io_sync;
663 submit_bh(WRITE, bh);
667 /* Force a new descriptor to be generated next
668 time round the loop. */
674 /* Lo and behold: we have just managed to send a transaction to
675 the log. Before we can commit it, wait for the IO so far to
676 complete. Control buffers being written are on the
677 transaction's t_log_list queue, and metadata buffers are on
678 the t_iobuf_list queue.
680 Wait for the buffers in reverse order. That way we are
681 less likely to be woken up until all IOs have completed, and
682 so we incur less scheduling load.
685 jbd_debug(3, "JBD: commit phase 4\n");
688 * akpm: these are BJ_IO, and j_list_lock is not needed.
689 * See __journal_try_to_free_buffer.
692 while (commit_transaction->t_iobuf_list != NULL) {
693 struct buffer_head *bh;
695 jh = commit_transaction->t_iobuf_list->b_tprev;
697 if (buffer_locked(bh)) {
704 if (unlikely(!buffer_uptodate(bh)))
707 clear_buffer_jwrite(bh);
709 JBUFFER_TRACE(jh, "ph4: unfile after journal write");
710 journal_unfile_buffer(journal, jh);
713 * ->t_iobuf_list should contain only dummy buffer_heads
714 * which were created by journal_write_metadata_buffer().
716 BUFFER_TRACE(bh, "dumping temporary bh");
717 journal_put_journal_head(jh);
719 J_ASSERT_BH(bh, atomic_read(&bh->b_count) == 0);
720 free_buffer_head(bh);
722 /* We also have to unlock and free the corresponding
724 jh = commit_transaction->t_shadow_list->b_tprev;
726 clear_bit(BH_JWrite, &bh->b_state);
727 J_ASSERT_BH(bh, buffer_jbddirty(bh));
729 /* The metadata is now released for reuse, but we need
730 to remember it against this transaction so that when
731 we finally commit, we can do any checkpointing
733 JBUFFER_TRACE(jh, "file as BJ_Forget");
734 journal_file_buffer(jh, commit_transaction, BJ_Forget);
735 /* Wake up any transactions which were waiting for this
737 wake_up_bit(&bh->b_state, BH_Unshadow);
738 JBUFFER_TRACE(jh, "brelse shadowed buffer");
742 J_ASSERT (commit_transaction->t_shadow_list == NULL);
744 jbd_debug(3, "JBD: commit phase 5\n");
746 /* Here we wait for the revoke record and descriptor record buffers */
748 while (commit_transaction->t_log_list != NULL) {
749 struct buffer_head *bh;
751 jh = commit_transaction->t_log_list->b_tprev;
753 if (buffer_locked(bh)) {
755 goto wait_for_ctlbuf;
758 goto wait_for_ctlbuf;
760 if (unlikely(!buffer_uptodate(bh)))
763 BUFFER_TRACE(bh, "ph5: control buffer writeout done: unfile");
764 clear_buffer_jwrite(bh);
765 journal_unfile_buffer(journal, jh);
766 journal_put_journal_head(jh);
767 __brelse(bh); /* One for getblk */
768 /* AKPM: bforget here */
772 journal_abort(journal, err);
774 jbd_debug(3, "JBD: commit phase 6\n");
776 if (journal_write_commit_record(journal, commit_transaction))
780 journal_abort(journal, err);
782 /* End of a transaction! Finally, we can do checkpoint
783 processing: any buffers committed as a result of this
784 transaction can be removed from any checkpoint list it was on
787 jbd_debug(3, "JBD: commit phase 7\n");
789 J_ASSERT(commit_transaction->t_sync_datalist == NULL);
790 J_ASSERT(commit_transaction->t_buffers == NULL);
791 J_ASSERT(commit_transaction->t_checkpoint_list == NULL);
792 J_ASSERT(commit_transaction->t_iobuf_list == NULL);
793 J_ASSERT(commit_transaction->t_shadow_list == NULL);
794 J_ASSERT(commit_transaction->t_log_list == NULL);
798 * As there are other places (journal_unmap_buffer()) adding buffers
799 * to this list we have to be careful and hold the j_list_lock.
801 spin_lock(&journal->j_list_lock);
802 while (commit_transaction->t_forget) {
803 transaction_t *cp_transaction;
804 struct buffer_head *bh;
806 jh = commit_transaction->t_forget;
807 spin_unlock(&journal->j_list_lock);
809 jbd_lock_bh_state(bh);
810 J_ASSERT_JH(jh, jh->b_transaction == commit_transaction ||
811 jh->b_transaction == journal->j_running_transaction);
814 * If there is undo-protected committed data against
815 * this buffer, then we can remove it now. If it is a
816 * buffer needing such protection, the old frozen_data
817 * field now points to a committed version of the
818 * buffer, so rotate that field to the new committed
821 * Otherwise, we can just throw away the frozen data now.
823 if (jh->b_committed_data) {
824 jbd_free(jh->b_committed_data, bh->b_size);
825 jh->b_committed_data = NULL;
826 if (jh->b_frozen_data) {
827 jh->b_committed_data = jh->b_frozen_data;
828 jh->b_frozen_data = NULL;
830 } else if (jh->b_frozen_data) {
831 jbd_free(jh->b_frozen_data, bh->b_size);
832 jh->b_frozen_data = NULL;
835 spin_lock(&journal->j_list_lock);
836 cp_transaction = jh->b_cp_transaction;
837 if (cp_transaction) {
838 JBUFFER_TRACE(jh, "remove from old cp transaction");
839 __journal_remove_checkpoint(jh);
842 /* Only re-checkpoint the buffer_head if it is marked
843 * dirty. If the buffer was added to the BJ_Forget list
844 * by journal_forget, it may no longer be dirty and
845 * there's no point in keeping a checkpoint record for
848 /* A buffer which has been freed while still being
849 * journaled by a previous transaction may end up still
850 * being dirty here, but we want to avoid writing back
851 * that buffer in the future now that the last use has
852 * been committed. That's not only a performance gain,
853 * it also stops aliasing problems if the buffer is left
854 * behind for writeback and gets reallocated for another
855 * use in a different page. */
856 if (buffer_freed(bh)) {
857 clear_buffer_freed(bh);
858 clear_buffer_jbddirty(bh);
861 if (buffer_jbddirty(bh)) {
862 JBUFFER_TRACE(jh, "add to new checkpointing trans");
863 __journal_insert_checkpoint(jh, commit_transaction);
864 if (is_journal_aborted(journal))
865 clear_buffer_jbddirty(bh);
866 JBUFFER_TRACE(jh, "refile for checkpoint writeback");
867 __journal_refile_buffer(jh);
868 jbd_unlock_bh_state(bh);
870 J_ASSERT_BH(bh, !buffer_dirty(bh));
871 /* The buffer on BJ_Forget list and not jbddirty means
872 * it has been freed by this transaction and hence it
873 * could not have been reallocated until this
874 * transaction has committed. *BUT* it could be
875 * reallocated once we have written all the data to
876 * disk and before we process the buffer on BJ_Forget
878 JBUFFER_TRACE(jh, "refile or unfile freed buffer");
879 __journal_refile_buffer(jh);
880 if (!jh->b_transaction) {
881 jbd_unlock_bh_state(bh);
883 journal_remove_journal_head(bh);
884 release_buffer_page(bh);
886 jbd_unlock_bh_state(bh);
888 cond_resched_lock(&journal->j_list_lock);
890 spin_unlock(&journal->j_list_lock);
892 * This is a bit sleazy. We use j_list_lock to protect transition
893 * of a transaction into T_FINISHED state and calling
894 * __journal_drop_transaction(). Otherwise we could race with
895 * other checkpointing code processing the transaction...
897 spin_lock(&journal->j_state_lock);
898 spin_lock(&journal->j_list_lock);
900 * Now recheck if some buffers did not get attached to the transaction
901 * while the lock was dropped...
903 if (commit_transaction->t_forget) {
904 spin_unlock(&journal->j_list_lock);
905 spin_unlock(&journal->j_state_lock);
909 /* Done with this transaction! */
911 jbd_debug(3, "JBD: commit phase 8\n");
913 J_ASSERT(commit_transaction->t_state == T_COMMIT);
915 commit_transaction->t_state = T_FINISHED;
916 J_ASSERT(commit_transaction == journal->j_committing_transaction);
917 journal->j_commit_sequence = commit_transaction->t_tid;
918 journal->j_committing_transaction = NULL;
919 commit_time = ktime_to_ns(ktime_sub(ktime_get(), start_time));
922 * weight the commit time higher than the average time so we don't
923 * react too strongly to vast changes in commit time
925 if (likely(journal->j_average_commit_time))
926 journal->j_average_commit_time = (commit_time*3 +
927 journal->j_average_commit_time) / 4;
929 journal->j_average_commit_time = commit_time;
931 spin_unlock(&journal->j_state_lock);
933 if (commit_transaction->t_checkpoint_list == NULL &&
934 commit_transaction->t_checkpoint_io_list == NULL) {
935 __journal_drop_transaction(journal, commit_transaction);
937 if (journal->j_checkpoint_transactions == NULL) {
938 journal->j_checkpoint_transactions = commit_transaction;
939 commit_transaction->t_cpnext = commit_transaction;
940 commit_transaction->t_cpprev = commit_transaction;
942 commit_transaction->t_cpnext =
943 journal->j_checkpoint_transactions;
944 commit_transaction->t_cpprev =
945 commit_transaction->t_cpnext->t_cpprev;
946 commit_transaction->t_cpnext->t_cpprev =
948 commit_transaction->t_cpprev->t_cpnext =
952 spin_unlock(&journal->j_list_lock);
954 jbd_debug(1, "JBD: commit %d complete, head %d\n",
955 journal->j_commit_sequence, journal->j_tail_sequence);
957 wake_up(&journal->j_wait_done_commit);