summaryrefslogtreecommitdiffstats
path: root/fs/jbd2/checkpoint.c
Commit message (Collapse)AuthorAgeFilesLines
* jbd2: avoid infinite loop when destroying aborted journalJan Kara2015-07-281-6/+33
| | | | | | | | | | | | | | | | | | | Commit 6f6a6fda2945 "jbd2: fix ocfs2 corrupt when updating journal superblock fails" changed jbd2_cleanup_journal_tail() to return EIO when the journal is aborted. That makes logic in jbd2_log_do_checkpoint() bail out which is fine, except that jbd2_journal_destroy() expects jbd2_log_do_checkpoint() to always make a progress in cleaning the journal. Without it jbd2_journal_destroy() just loops in an infinite loop. Fix jbd2_journal_destroy() to cleanup journal checkpoint lists of jbd2_log_do_checkpoint() fails with error. Reported-by: Eryu Guan <guaneryu@gmail.com> Tested-by: Eryu Guan <guaneryu@gmail.com> Fixes: 6f6a6fda294506dfe0e3e0a253bb2d2923f28f0a Signed-off-by: Jan Kara <jack@suse.com> Signed-off-by: Theodore Ts'o <tytso@mit.edu>
* jbd2: fix ocfs2 corrupt when updating journal superblock failsJoseph Qi2015-06-151-3/+2
| | | | | | | | | | | | | | | | | | | | | | | | | | | | If updating journal superblock fails after journal data has been flushed, the error is omitted and this will mislead the caller as a normal case. In ocfs2, the checkpoint will be treated successfully and the other node can get the lock to update. Since the sb_start is still pointing to the old log block, it will rewrite the journal data during journal recovery by the other node. Thus the new updates will be overwritten and ocfs2 corrupts. So in above case we have to return the error, and ocfs2_commit_cache will take care of the error and prevent the other node to do update first. And only after recovering journal it can do the new updates. The issue discussion mail can be found at: https://oss.oracle.com/pipermail/ocfs2-devel/2015-June/010856.html http://comments.gmane.org/gmane.comp.file-systems.ext4/48841 [ Fixed bug in patch which allowed a non-negative error return from jbd2_cleanup_journal_tail() to leak out of jbd2_fjournal_flush(); this was causing xfstests ext4/306 to fail. -- Ted ] Reported-by: Yiwen Jiang <jiangyiwen@huawei.com> Signed-off-by: Joseph Qi <joseph.qi@huawei.com> Signed-off-by: Theodore Ts'o <tytso@mit.edu> Tested-by: Yiwen Jiang <jiangyiwen@huawei.com> Cc: Junxiao Bi <junxiao.bi@oracle.com> Cc: stable@vger.kernel.org
* jbd2: use GFP_NOFS in jbd2_cleanup_journal_tail()Dmitry Monakhov2015-06-151-1/+1
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | jbd2_cleanup_journal_tail() can be invoked by jbd2__journal_start() So allocations should be done with GFP_NOFS [Full stack trace snipped from 3.10-rh7] [<ffffffff815c4bd4>] dump_stack+0x19/0x1b [<ffffffff8105dba1>] warn_slowpath_common+0x61/0x80 [<ffffffff8105dcca>] warn_slowpath_null+0x1a/0x20 [<ffffffff815c2142>] slab_pre_alloc_hook.isra.31.part.32+0x15/0x17 [<ffffffff8119c045>] kmem_cache_alloc+0x55/0x210 [<ffffffff811477f5>] ? mempool_alloc_slab+0x15/0x20 [<ffffffff811477f5>] mempool_alloc_slab+0x15/0x20 [<ffffffff81147939>] mempool_alloc+0x69/0x170 [<ffffffff815cb69e>] ? _raw_spin_unlock_irq+0xe/0x20 [<ffffffff8109160d>] ? finish_task_switch+0x5d/0x150 [<ffffffff811f1a8e>] bio_alloc_bioset+0x1be/0x2e0 [<ffffffff8127ee49>] blkdev_issue_flush+0x99/0x120 [<ffffffffa019a733>] jbd2_cleanup_journal_tail+0x93/0xa0 [jbd2] -->GFP_KERNEL [<ffffffffa019aca1>] jbd2_log_do_checkpoint+0x221/0x4a0 [jbd2] [<ffffffffa019afc7>] __jbd2_log_wait_for_space+0xa7/0x1e0 [jbd2] [<ffffffffa01952d8>] start_this_handle+0x2d8/0x550 [jbd2] [<ffffffff811b02a9>] ? __memcg_kmem_put_cache+0x29/0x30 [<ffffffff8119c120>] ? kmem_cache_alloc+0x130/0x210 [<ffffffffa019573a>] jbd2__journal_start+0xba/0x190 [jbd2] [<ffffffff811532ce>] ? lru_cache_add+0xe/0x10 [<ffffffffa01c9549>] ? ext4_da_write_begin+0xf9/0x330 [ext4] [<ffffffffa01f2c77>] __ext4_journal_start_sb+0x77/0x160 [ext4] [<ffffffffa01c9549>] ext4_da_write_begin+0xf9/0x330 [ext4] [<ffffffff811446ec>] generic_file_buffered_write_iter+0x10c/0x270 [<ffffffff81146918>] __generic_file_write_iter+0x178/0x390 [<ffffffff81146c6b>] __generic_file_aio_write+0x8b/0xb0 [<ffffffff81146ced>] generic_file_aio_write+0x5d/0xc0 [<ffffffffa01bf289>] ext4_file_write+0xa9/0x450 [ext4] [<ffffffff811c31d9>] ? pipe_read+0x379/0x4f0 [<ffffffff811b93f0>] do_sync_write+0x90/0xe0 [<ffffffff811b9b6d>] vfs_write+0xbd/0x1e0 [<ffffffff811ba5b8>] SyS_write+0x58/0xb0 [<ffffffff815d4799>] system_call_fastpath+0x16/0x1b Signed-off-by: Dmitry Monakhov <dmonakhov@openvz.org> Signed-off-by: Theodore Ts'o <tytso@mit.edu> Cc: stable@vger.kernel.org
* jbd2: simplify calling convention around __jbd2_journal_clean_checkpoint_listJan Kara2014-09-181-32/+24
| | | | | | | | | | __jbd2_journal_clean_checkpoint_list() returns number of buffers it freed but noone was using the value so just stop doing that. This also allows for simplifying the calling convention for journal_clean_once_cp_list(). Signed-off-by: Jan Kara <jack@suse.cz> Signed-off-by: Theodore Ts'o <tytso@mit.edu>
* jbd2: avoid pointless scanning of checkpoint listsJan Kara2014-09-181-14/+18
| | | | | | | | | | | | | | | | | | Yuanhan has reported that when he is running fsync(2) heavy workload creating new files over ramdisk, significant amount of time is spent in __jbd2_journal_clean_checkpoint_list() trying to clean old transactions (but they cannot be cleaned up because flusher hasn't yet checkpointed those buffers). The workload can be generated by: fs_mark -d /fs/ram0/1 -D 2 -N 2560 -n 1000000 -L 1 -S 1 -s 4096 Reduce the amount of scanning by stopping to scan the transaction list once we find a transaction that cannot be checkpointed. Note that this way of cleaning is still enough to keep freeing space in the journal after fully checkpointed transactions. Reported-and-tested-by: Yuanhan Liu <yuanhan.liu@linux.intel.com> Signed-off-by: Jan Kara <jack@suse.cz> Signed-off-by: Theodore Ts'o <tytso@mit.edu>
* jbd2: jbd2_log_wait_for_space improve error detetcionDmitry Monakhov2014-09-161-2/+4
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | If EIO happens after we have dropped j_state_lock, we won't notice that the journal has been aborted. So it is reasonable to move this check after we have grabbed the j_checkpoint_mutex and re-grabbed the j_state_lock. This patch helps to prevent false positive complain after EIO. #DMESG: __jbd2_log_wait_for_space: needed 8448 blocks and only had 8386 space available __jbd2_log_wait_for_space: no way to get more journal space in ram1-8 ------------[ cut here ]------------ WARNING: CPU: 15 PID: 6739 at fs/jbd2/checkpoint.c:168 __jbd2_log_wait_for_space+0x188/0x200() Modules linked in: brd iTCO_wdt lpc_ich mfd_core igb ptp dm_mirror dm_region_hash dm_log dm_mod CPU: 15 PID: 6739 Comm: fsstress Tainted: G W 3.17.0-rc2-00429-g684de57 #139 Hardware name: Intel Corporation W2600CR/W2600CR, BIOS SE5C600.86B.99.99.x028.061320111235 06/13/2011 00000000000000a8 ffff88077aaab878 ffffffff815c1a8c 00000000000000a8 0000000000000000 ffff88077aaab8b8 ffffffff8106ce8c ffff88077aaab898 ffff8807c57e6000 ffff8807c57e6028 0000000000002100 ffff8807c57e62f0 Call Trace: [<ffffffff815c1a8c>] dump_stack+0x51/0x6d [<ffffffff8106ce8c>] warn_slowpath_common+0x8c/0xc0 [<ffffffff8106ceda>] warn_slowpath_null+0x1a/0x20 [<ffffffff812419f8>] __jbd2_log_wait_for_space+0x188/0x200 [<ffffffff8123be9a>] start_this_handle+0x4da/0x7b0 [<ffffffff810990e5>] ? local_clock+0x25/0x30 [<ffffffff810aba87>] ? lockdep_init_map+0xe7/0x180 [<ffffffff8123c5bc>] jbd2__journal_start+0xdc/0x1d0 [<ffffffff811f2414>] ? __ext4_new_inode+0x7f4/0x1330 [<ffffffff81222a38>] __ext4_journal_start_sb+0xf8/0x110 [<ffffffff811f2414>] __ext4_new_inode+0x7f4/0x1330 [<ffffffff810ac359>] ? lock_release_holdtime+0x29/0x190 [<ffffffff812025bb>] ext4_create+0x8b/0x150 [<ffffffff8117fe3b>] vfs_create+0x7b/0xb0 [<ffffffff8118097b>] do_last+0x7db/0xcf0 [<ffffffff8117e31d>] ? inode_permission+0x4d/0x50 [<ffffffff811845d2>] path_openat+0x242/0x590 [<ffffffff81191a76>] ? __alloc_fd+0x36/0x140 [<ffffffff81184a6a>] do_filp_open+0x4a/0xb0 [<ffffffff81191b61>] ? __alloc_fd+0x121/0x140 [<ffffffff81172f20>] do_sys_open+0x170/0x220 [<ffffffff8117300e>] SyS_open+0x1e/0x20 [<ffffffff811715d6>] SyS_creat+0x16/0x20 [<ffffffff815c7e12>] system_call_fastpath+0x16/0x1b ---[ end trace cd71c831f82059db ]--- Signed-off-by: Dmitry Monakhov <dmonakhov@openvz.org> Signed-off-by: Theodore Ts'o <tytso@mit.edu>
* jbd2: optimize jbd2_log_do_checkpoint() a bitJan Kara2014-09-041-3/+4
| | | | | | | | | | | When we discover written out buffer in transaction checkpoint list we don't have to recheck validity of a transaction. Either this is the last buffer in a transaction - and then we are done - or this isn't and then we can just take another buffer from the checkpoint list without dropping j_list_lock. Signed-off-by: Jan Kara <jack@suse.cz> Signed-off-by: Theodore Ts'o <tytso@mit.edu>
* jbd2: don't call get_bh() before calling __jbd2_journal_remove_checkpoint()Theodore Ts'o2014-09-041-14/+5
| | | | | | | | | | | The __jbd2_journal_remove_checkpoint() doesn't require an elevated b_count; indeed, until the jh structure gets released by the call to jbd2_journal_put_journal_head(), the bh's b_count is elevated by virtue of the existence of the jh structure. Suggested-by: Jan Kara <jack@suse.cz> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Theodore Ts'o <tytso@mit.edu>
* jbd2: fold __wait_cp_io into jbd2_log_do_checkpoint()Theodore Ts'o2014-09-011-56/+31
| | | | | | | __wait_cp_io() is only called by jbd2_log_do_checkpoint(). Fold it in to make it a bit easier to understand. Signed-off-by: Theodore Ts'o <tytso@mit.edu>
* jbd2: fold __process_buffer() into jbd2_log_do_checkpoint()Theodore Ts'o2014-09-011-111/+84
| | | | | | | | | | | | | | | | | | __process_buffer() is only called by jbd2_log_do_checkpoint(), and it had a very complex locking protocol where it would be called with the j_list_lock, and sometimes exit with the lock held (if the return code was 0), or release the lock. This was confusing both to humans and to smatch (which erronously complained that the lock was taken twice). Folding __process_buffer() to the caller allows us to simplify the control flow, making the resulting function easier to read and reason about, and dropping the compiled size of fs/jbd2/checkpoint.c by 150 bytes (over 4% of the text size). Signed-off-by: Theodore Ts'o <tytso@mit.edu> Reviewed-by: Jan Kara <jack@suse.cz>
* jbd2: drop checkpoint mutex when waiting in __jbd2_log_wait_for_space()Paul Gortmaker2013-06-121-0/+8
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | While trying to debug an an issue under extreme I/O loading on preempt-rt kernels, the following backtrace was observed via SysRQ output: rm D ffff8802203afbc0 4600 4878 4748 0x00000000 ffff8802217bfb78 0000000000000082 ffff88021fc2bb80 ffff88021fc2bb80 ffff88021fc2bb80 ffff8802217bffd8 ffff8802217bffd8 ffff8802217bffd8 ffff88021f1d4c80 ffff88021fc2bb80 ffff8802217bfb88 ffff88022437b000 Call Trace: [<ffffffff8172dc34>] schedule+0x24/0x70 [<ffffffff81225b5d>] jbd2_log_wait_commit+0xbd/0x140 [<ffffffff81060390>] ? __init_waitqueue_head+0x50/0x50 [<ffffffff81223635>] jbd2_log_do_checkpoint+0xf5/0x520 [<ffffffff81223b09>] __jbd2_log_wait_for_space+0xa9/0x1f0 [<ffffffff8121dc40>] start_this_handle.isra.10+0x2e0/0x530 [<ffffffff81060390>] ? __init_waitqueue_head+0x50/0x50 [<ffffffff8121e0a3>] jbd2__journal_start+0xc3/0x110 [<ffffffff811de7ce>] ? ext4_rmdir+0x6e/0x230 [<ffffffff8121e0fe>] jbd2_journal_start+0xe/0x10 [<ffffffff811f308b>] ext4_journal_start_sb+0x5b/0x160 [<ffffffff811de7ce>] ext4_rmdir+0x6e/0x230 [<ffffffff811435c5>] vfs_rmdir+0xd5/0x140 [<ffffffff8114370f>] do_rmdir+0xdf/0x120 [<ffffffff8105c6b4>] ? task_work_run+0x44/0x80 [<ffffffff81002889>] ? do_notify_resume+0x89/0x100 [<ffffffff817361ae>] ? int_signal+0x12/0x17 [<ffffffff81145d85>] sys_unlinkat+0x25/0x40 [<ffffffff81735f22>] system_call_fastpath+0x16/0x1b What is interesting here, is that we call log_wait_commit, from within wait_for_space, but we are still holding the checkpoint_mutex as it surrounds mostly the whole of wait_for_space. And then, as we are waiting, journal_commit_transaction can run, and if the JBD2_FLUSHED bit is set, then we will also try to take the same checkpoint_mutex. It seems that we need to drop the checkpoint_mutex while sitting in jbd2_log_wait_commit, if we want to guarantee that progress can be made by jbd2_journal_commit_transaction(). There does not seem to be anything preempt-rt specific about this, other then perhaps increasing the odds of it happening. Signed-off-by: Paul Gortmaker <paul.gortmaker@windriver.com> Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
* jbd2: remove unused waitqueuesJan Kara2013-06-041-4/+0
| | | | | | | | j_wait_logspace and j_wait_checkpoint are unused. Remove them. Reviewed-by: Zheng Liu <wenqing.lz@taobao.com> Signed-off-by: Jan Kara <jack@suse.cz> Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
* jbd2: cleanup needed free block estimates when starting a transactionJan Kara2013-06-041-4/+4
| | | | | | | | | | | | | | | | | | | __jbd2_log_space_left() and jbd_space_needed() were kind of odd. jbd_space_needed() accounted also credits needed for currently committing transaction while it didn't account for credits needed for control blocks. __jbd2_log_space_left() then accounted for control blocks as a fraction of free space. Since results of these two functions are always only compared against each other, this works correct but is somewhat strange. Move the estimates so that jbd_space_needed() returns number of blocks needed for a transaction including control blocks and __jbd2_log_space_left() returns free space in the journal (with the committing transaction already subtracted). Rename functions to jbd2_log_space_left() and jbd2_space_needed() while we are changing them. Reviewed-by: Zheng Liu <wenqing.lz@taobao.com> Signed-off-by: Jan Kara <jack@suse.cz> Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
* jbd2: remove journal_head from descriptor buffersJan Kara2013-06-041-1/+0
| | | | | | | | | Similarly as for metadata buffers, also log descriptor buffers don't really need the journal head. So strip it and remove BJ_LogCtl list. Reviewed-by: Zheng Liu <wenqing.lz@taobao.com> Signed-off-by: Jan Kara <jack@suse.cz> Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
* jbd2: don't create journal_head for temporary journal buffersJan Kara2013-06-041-1/+0
| | | | | | | | | | | | | | | When writing metadata to the journal, we create temporary buffer heads for that task. We also attach journal heads to these buffer heads but the only purpose of the journal heads is to keep buffers linked in transaction's BJ_IO list. We remove the need for journal heads by reusing buffer_head's b_assoc_buffers list for that purpose. Also since BJ_IO list is just a temporary list for transaction commit, we use a private list in jbd2_journal_commit_transaction() for that thus removing BJ_IO list from transaction completely. Reviewed-by: Zheng Liu <wenqing.lz@taobao.com> Signed-off-by: Jan Kara <jack@suse.cz> Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
* jbd2: remove bh_state lock from checkpointing codeJan Kara2012-03-131-52/+7
| | | | | | | | | | | | | | | | All accesses to checkpointing entries in journal_head are protected by j_list_lock. Thus __jbd2_journal_remove_checkpoint() doesn't really need bh_state lock. Also the only part of journal head that the rest of checkpointing code needs to check is jh->b_transaction which is safe to read under j_list_lock. So we can safely remove bh_state lock from all of checkpointing code which makes it considerably prettier. Signed-off-by: Jan Kara <jack@suse.cz> Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
* jbd2: fix BH_JWrite setting in checkpointing codeJan Kara2012-03-131-2/+0
| | | | | | | | | | BH_JWrite bit should be set when buffer is written to the journal. So checkpointing shouldn't set this bit when writing out buffer. This didn't cause any observable bug since BH_JWrite bit is used only for debugging purposes but it's good to have this consistent. Signed-off-by: Jan Kara <jack@suse.cz> Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
* jbd2: issue cache flush after checkpointing even with internal journalJan Kara2012-03-131-63/+12
| | | | | | | | | | | | | | | | | | | | | | When we reach jbd2_cleanup_journal_tail(), there is no guarantee that checkpointed buffers are on a stable storage - especially if buffers were written out by jbd2_log_do_checkpoint(), they are likely to be only in disk's caches. Thus when we update journal superblock effectively removing old transaction from journal, this write of superblock can get to stable storage before those checkpointed buffers which can result in filesystem corruption after a crash. Thus we must unconditionally issue a cache flush before we update journal superblock in these cases. A similar problem can also occur if journal superblock is written only in disk's caches, other transaction starts reusing space of the transaction cleaned from the log and power failure happens. Subsequent journal replay would still try to replay the old transaction but some of it's blocks may be already overwritten by the new transaction. For this reason we must use WRITE_FUA when updating log tail and we must first write new log tail to disk and update in-memory information only after that. Signed-off-by: Jan Kara <jack@suse.cz> Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
* jbd2: split updating of journal superblock and marking journal emptyJan Kara2012-03-131-1/+1
| | | | | | | | | | | There are three case of updating journal superblock. In the first case, we want to mark journal as empty (setting s_sequence to 0), in the second case we want to update log tail, in the third case we want to update s_errno. Split these cases into separate functions. It makes the code slightly more straightforward and later patches will make the distinction even more important. Signed-off-by: Jan Kara <jack@suse.cz> Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
* jbd2: allocate transaction from separate slab cacheYongqiang Yang2012-02-201-1/+1
| | | | | | | | | | | There is normally only a handful of these active at any one time, but putting them in a separate slab cache makes debugging memory corruption problems easier. Manish Katiyar also wanted this make it easier to test memory failure scenarios in the jbd2 layer. Cc: Manish Katiyar <mkatiyar@gmail.com> Signed-off-by: Yongqiang Yang <xiaoqiangnk@gmail.com> Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
* jbd2: add drop_transaction/update_superblock_end tracepointsSeiji Aguchi2012-02-201-0/+2
| | | | | | | | | | | This patch adds trace_jbd2_drop_transaction and trace_jbd2_update_superblock_end because there are similar tracepoints in jbd and they are needed in jbd2 as well. Reviewed-by: Lukas Czerner <lczerner@redhat.com> Signed-off-by: Seiji Aguchi <seiji.aguchi@hds.com> Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
* treewide: Fix comment and string typo 'bufer'Paul Bolle2011-12-061-1/+1
| | | | | Signed-off-by: Paul Bolle <pebolle@tiscali.nl> Signed-off-by: Jiri Kosina <jkosina@suse.cz>
* jbd2: use WRITE_SYNC in journal checkpointTao Ma2011-06-271-1/+4
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | In journal checkpoint, we write the buffer and wait for its finish. But in cfq, the async queue has a very low priority, and in our test, if there are too many sync queues and every queue is filled up with requests, the write request will be delayed for quite a long time and all the tasks which are waiting for journal space will end with errors like: INFO: task attr_set:3816 blocked for more than 120 seconds. "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message. attr_set D ffff880028393480 0 3816 1 0x00000000 ffff8802073fbae8 0000000000000086 ffff8802140847c8 ffff8800283934e8 ffff8802073fb9d8 ffffffff8103e456 ffff8802140847b8 ffff8801ed728080 ffff8801db4bc080 ffff8801ed728450 ffff880028393480 0000000000000002 Call Trace: [<ffffffff8103e456>] ? __dequeue_entity+0x33/0x38 [<ffffffff8103caad>] ? need_resched+0x23/0x2d [<ffffffff814006a6>] ? thread_return+0xa2/0xbc [<ffffffffa01f6224>] ? jbd2_journal_dirty_metadata+0x116/0x126 [jbd2] [<ffffffffa01f6224>] ? jbd2_journal_dirty_metadata+0x116/0x126 [jbd2] [<ffffffff81400d31>] __mutex_lock_common+0x14e/0x1a9 [<ffffffffa021dbfb>] ? brelse+0x13/0x15 [ext4] [<ffffffff81400ddb>] __mutex_lock_slowpath+0x19/0x1b [<ffffffff81400b2d>] mutex_lock+0x1b/0x32 [<ffffffffa01f927b>] __jbd2_journal_insert_checkpoint+0xe3/0x20c [jbd2] [<ffffffffa01f547b>] start_this_handle+0x438/0x527 [jbd2] [<ffffffff8106f491>] ? autoremove_wake_function+0x0/0x3e [<ffffffffa01f560b>] jbd2_journal_start+0xa1/0xcc [jbd2] [<ffffffffa02353be>] ext4_journal_start_sb+0x57/0x81 [ext4] [<ffffffffa024a314>] ext4_xattr_set+0x6c/0xe3 [ext4] [<ffffffffa024aaff>] ext4_xattr_user_set+0x42/0x4b [ext4] [<ffffffff81145adb>] generic_setxattr+0x6b/0x76 [<ffffffff81146ac0>] __vfs_setxattr_noperm+0x47/0xc0 [<ffffffff81146bb8>] vfs_setxattr+0x7f/0x9a [<ffffffff81146c88>] setxattr+0xb5/0xe8 [<ffffffff81137467>] ? do_filp_open+0x571/0xa6e [<ffffffff81146d26>] sys_fsetxattr+0x6b/0x91 [<ffffffff81002d32>] system_call_fastpath+0x16/0x1b So this patch tries to use WRITE_SYNC in __flush_batch so that the request will be moved into sync queue and handled by cfq timely. We also use the new plug, sot that all the WRITE_SYNC requests can be given as a whole when we unplug it. Signed-off-by: Tao Ma <boyu.mt@taobao.com> Signed-off-by: "Theodore Ts'o" <tytso@mit.edu> Cc: Jan Kara <jack@suse.cz> Reported-by: Robin Dong <sanbai@taobao.com>
* jbd2: Fix oops in jbd2_journal_remove_journal_head()Jan Kara2011-06-131-12/+16
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | jbd2_journal_remove_journal_head() can oops when trying to access journal_head returned by bh2jh(). This is caused for example by the following race: TASK1 TASK2 jbd2_journal_commit_transaction() ... processing t_forget list __jbd2_journal_refile_buffer(jh); if (!jh->b_transaction) { jbd_unlock_bh_state(bh); jbd2_journal_try_to_free_buffers() jbd2_journal_grab_journal_head(bh) jbd_lock_bh_state(bh) __journal_try_to_free_buffer() jbd2_journal_put_journal_head(jh) jbd2_journal_remove_journal_head(bh); jbd2_journal_put_journal_head() in TASK2 sees that b_jcount == 0 and buffer is not part of any transaction and thus frees journal_head before TASK1 gets to doing so. Note that even buffer_head can be released by try_to_free_buffers() after jbd2_journal_put_journal_head() which adds even larger opportunity for oops (but I didn't see this happen in reality). Fix the problem by making transactions hold their own journal_head reference (in b_jcount). That way we don't have to remove journal_head explicitely via jbd2_journal_remove_journal_head() and instead just remove journal_head when b_jcount drops to zero. The result of this is that [__]jbd2_journal_refile_buffer(), [__]jbd2_journal_unfile_buffer(), and __jdb2_journal_remove_checkpoint() can free journal_head which needs modification of a few callers. Also we have to be careful because once journal_head is removed, buffer_head might be freed as well. So we have to get our own buffer_head reference where it matters. Signed-off-by: Jan Kara <jack@suse.cz> Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
* Merge branch 'next' into upstream-mergeTheodore Ts'o2010-10-271-0/+10
|\ | | | | | | | | | | | | Conflicts: fs/ext4/inode.c fs/ext4/mballoc.c include/trace/events/ext4.h
| * jbd2: Add sanity check for attempts to start handle during umountTheodore Ts'o2010-10-271-0/+10
| | | | | | | | | | | | | | | | | | An attempt to modify the file system during the call to jbd2_destroy_journal() can lead to a system lockup. So add some checking to make it much more obvious when this happens to and to determine where the offending code is located. Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
* | block: remove BLKDEV_IFL_WAITChristoph Hellwig2010-09-161-2/+1
|/ | | | | | | | | | | | | | All the blkdev_issue_* helpers can only sanely be used for synchronous caller. To issue cache flushes or barriers asynchronously the caller needs to set up a bio by itself with a completion callback to move the asynchronous state machine ahead. So drop the BLKDEV_IFL_WAIT flag that is always specified when calling blkdev_issue_* and also remove the now unused flags argument to blkdev_issue_flush and blkdev_issue_zeroout. For blkdev_issue_discard we need to keep it for the secure discard flag, which gains a more descriptive name and loses the bitops vs flag confusion. Signed-off-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <jaxboe@fusionio.com>
* remove SWRITE* I/O typesChristoph Hellwig2010-08-181-1/+3
| | | | | | | | | | | | | | | | | These flags aren't real I/O types, but tell ll_rw_block to always lock the buffer instead of giving up on a failed trylock. Instead add a new write_dirty_buffer helper that implements this semantic and use it from the existing SWRITE* callers. Note that the ll_rw_block code had a bug where it didn't promote WRITE_SYNC_PLUG properly, which this patch fixes. In the ufs code clean up the helper that used to call ll_rw_block to mirror sync_dirty_buffer, which is the function it implements for compound buffers. Signed-off-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
* jbd2: Change j_state_lock to be a rwlock_tTheodore Ts'o2010-08-031-8/+8
| | | | | | | | Lockstat reports have shown that j_state_lock is a major source of lock contention, especially on systems with more than 4 CPU cores. So change it to be a read/write spinlock. Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
* jbd2: Use atomic variables to avoid taking t_handle_lock in jbd2_journal_stopTheodore Ts'o2010-08-021-1/+1
| | | | | | | | By using an atomic_t for t_updates and t_outstanding credits, this should allow us to not need to take transaction t_handle_lock in jbd2_journal_stop(). Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
* blkdev: generalize flags for blkdev_issue_fn functionsDmitry Monakhov2010-04-281-1/+2
| | | | | | | | The patch just convert all blkdev_issue_xxx function to common set of flags. Wait/allocation semantics preserved. Signed-off-by: Dmitry Monakhov <dmonakhov@openvz.org> Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
* ext4: Add new tracepoint for jbd2_cleanup_journal_tailTheodore Ts'o2009-12-231-0/+1
| | | | Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
* ext4, jbd2: Add barriers for file systems with exernal journalsTheodore Ts'o2009-12-231-0/+15
| | | | | | | | | | | | | | | | | | | | This is a bit complicated because we are trying to optimize when we send barriers to the fs data disk. We could just throw in an extra barrier to the data disk whenever we send a barrier to the journal disk, but that's not always strictly necessary. We only need to send a barrier during a commit when there are data blocks which are must be written out due to an inode written in ordered mode, or if fsync() depends on the commit to force data blocks to disk. Finally, before we drop transactions from the beginning of the journal during a checkpoint operation, we need to guarantee that any blocks that were flushed out to the data disk are firmly on the rust platter before we drop the transaction from the journal. Thanks to Oleg Drokin for pointing out this flaw in ext3/ext4. Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
* jbd2: Use tracepoints for history fileTheodore Ts'o2009-09-301-0/+7
| | | | | | | | | The /proc/fs/jbd2/<dev>/history was maintained manually; by using tracepoints, we can get all of the existing functionality of the /proc file plus extra capabilities thanks to the ftrace infrastructure. We save memory as a bonus. Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
* jbd2: convert instrumentation from markers to tracepointsTheodore Ts'o2009-06-171-3/+2
| | | | Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
* jbd2: Call journal commit callback without holding j_list_lockAneesh Kumar K.V2008-11-061-1/+1
| | | | | | | | | Avoid freeing the transaction in __jbd2_journal_drop_transaction() so the journal commit callback can run without holding j_list_lock, to avoid lock contention on this spinlock. Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com> Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
* jbd2: Remove a large array of bh's from the stack of the checkpoint routineTheodore Ts'o2008-11-051-13/+9
| | | | | | | | jbd2_log_do_checkpoint()n is one of the kernel's largest stack users. Move the array of buffer head's from the stack of jbd2_log_do_checkpoint() to the in-core journal structure. Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
* jbd2: don't give up looking for space so easily in __jbd2_log_wait_for_spaceTheodore Ts'o2008-11-061-7/+25
| | | | | | | | | | | | | | | | | | | | | | Commit 23f8b79e introducd a regression because it assumed that if there were no transactions ready to be checkpointed, that no progress could be made on making space available in the journal, and so the journal should be aborted. This assumption is false; it could be the case that simply calling jbd2_cleanup_journal_tail() will recover the necessary space, or, for small journals, the currently committing transaction could be responsible for chewing up the required space in the log, so we need to wait for the currently committing transaction to finish before trying to force a checkpoint operation. This patch fixes a bug reported by Mihai Harpau at: https://bugzilla.redhat.com/show_bug.cgi?id=469582 This patch fixes a bug reported by François Valenduc at: http://bugzilla.kernel.org/show_bug.cgi?id=11840 Signed-off-by: "Theodore Ts'o" <tytso@mit.edu> Cc: Duane Griffin <duaneg@dghda.com> Cc: Toshiyuki Okajima <toshi.okajima@jp.fujitsu.com>
* jbd2: fix error handling for checkpoint ioHidehiro Kawai2008-10-101-12/+37
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | When a checkpointing IO fails, current JBD2 code doesn't check the error and continue journaling. This means latest metadata can be lost from both the journal and filesystem. This patch leaves the failed metadata blocks in the journal space and aborts journaling in the case of jbd2_log_do_checkpoint(). To achieve this, we need to do: 1. don't remove the failed buffer from the checkpoint list where in the case of __try_to_free_cp_buf() because it may be released or overwritten by a later transaction 2. jbd2_log_do_checkpoint() is the last chance, remove the failed buffer from the checkpoint list and abort the journal 3. when checkpointing fails, don't update the journal super block to prevent the journaled contents from being cleaned. For safety, don't update j_tail and j_tail_sequence either 4. when checkpointing fails, notify this error to the ext4 layer so that ext4 don't clear the needs_recovery flag, otherwise the journaled contents are ignored and cleaned in the recovery phase 5. if the recovery fails, keep the needs_recovery flag 6. prevent jbd2_cleanup_journal_tail() from being called between __jbd2_journal_drop_transaction() and jbd2_journal_abort() (a possible race issue between jbd2_log_do_checkpoint()s called by jbd2_journal_flush() and __jbd2_log_wait_for_space()) Signed-off-by: Hidehiro Kawai <hidehiro.kawai.ez@hitachi.com> Signed-off-by: Theodore Ts'o <tytso@mit.edu>
* ext4: Add debugging markers that can be used by systemtapTheodore Ts'o2008-10-051-0/+3
| | | | | | | This debugging markers are designed to debug problems such as the random filesystem latency problems reported by Arjan. Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
* jbd2: abort instead of waiting for nonexistent transactionDuane Griffin2008-10-081-2/+17
| | | | | | | | | | | | | | | | | | | | The __jbd2_log_wait_for_space function sits in a loop checkpointing transactions until there is sufficient space free in the journal. However, if there are no transactions to be processed (e.g. because the free space calculation is wrong due to a corrupted filesystem) it will never progress. Check for space being required when no transactions are outstanding and abort the journal instead of endlessly looping. This patch fixes the bug reported by Sami Liedes at: http://bugzilla.kernel.org/show_bug.cgi?id=10976 Signed-off-by: Duane Griffin <duaneg@dghda.com> Cc: Sami Liedes <sliedes@cc.hut.fi> Cc: <linux-ext4@vger.kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
* jbd2: Remove data=ordered mode support using jbd buffer headsJan Kara2008-07-111-1/+0
| | | | Signed-off-by: Jan Kara <jack@suse.cz>
* spinlock: lockbreak cleanupNick Piggin2008-01-301-1/+2
| | | | | | | | | | | | | | | | | | | | The break_lock data structure and code for spinlocks is quite nasty. Not only does it double the size of a spinlock but it changes locking to a potentially less optimal trylock. Put all of that under CONFIG_GENERIC_LOCKBREAK, and introduce a __raw_spin_is_contended that uses the lock data itself to determine whether there are waiters on the lock, to be used if CONFIG_GENERIC_LOCKBREAK is not set. Rename need_lockbreak to spin_needbreak, make it use spin_is_contended to decouple it from the spinlock implementation, and make it typesafe (rwlocks do not have any need_lockbreak sites -- why do they even get bloated up with that break_lock then?). Signed-off-by: Nick Piggin <npiggin@suse.de> Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
* jbd2: jbd2 stats through procfsJohann Lombardi2008-01-281-2/+8
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | The patch below updates the jbd stats patch to 2.6.20/jbd2. The initial patch was posted by Alex Tomas in December 2005 (http://marc.info/?l=linux-ext4&m=113538565128617&w=2). It provides statistics via procfs such as transaction lifetime and size. Sometimes, investigating performance problems, i find useful to have stats from jbd about transaction's lifetime, size, etc. here is a patch for review and inclusion probably. for example, stats after creation of 3M files in htree directory: [root@bob ~]# cat /proc/fs/jbd/sda/history R/C tid wait run lock flush log hndls block inlog ctime write drop close R 261 8260 2720 0 0 750 9892 8170 8187 C 259 750 0 4885 1 R 262 20 2200 10 0 770 9836 8170 8187 R 263 30 2200 10 0 3070 9812 8170 8187 R 264 0 5000 10 0 1340 0 0 0 C 261 8240 3212 4957 0 R 265 8260 1470 0 0 4640 9854 8170 8187 R 266 0 5000 10 0 1460 0 0 0 C 262 8210 2989 4868 0 R 267 8230 1490 10 0 4440 9875 8171 8188 R 268 0 5000 10 0 1260 0 0 0 C 263 7710 2937 4908 0 R 269 7730 1470 10 0 3330 9841 8170 8187 R 270 0 5000 10 0 830 0 0 0 C 265 8140 3234 4898 0 C 267 720 0 4849 1 R 271 8630 2740 20 0 740 9819 8170 8187 C 269 800 0 4214 1 R 272 40 2170 10 0 830 9716 8170 8187 R 273 40 2280 0 0 3530 9799 8170 8187 R 274 0 5000 10 0 990 0 0 0 where, R - line for transaction's life from T_RUNNING to T_FINISHED C - line for transaction's checkpointing tid - transaction's id wait - for how long we were waiting for new transaction to start (the longest period journal_start() took in this transaction) run - real transaction's lifetime (from T_RUNNING to T_LOCKED lock - how long we were waiting for all handles to close (time the transaction was in T_LOCKED) flush - how long it took to flush all data (data=ordered) log - how long it took to write the transaction to the log hndls - how many handles got to the transaction block - how many blocks got to the transaction inlog - how many blocks are written to the log (block + descriptors) ctime - how long it took to checkpoint the transaction write - how many blocks have been written during checkpointing drop - how many blocks have been dropped during checkpointing close - how many running transactions have been closed to checkpoint this one all times are in msec. [root@bob ~]# cat /proc/fs/jbd/sda/info 280 transaction, each upto 8192 blocks average: 1633ms waiting for transaction 3616ms running transaction 5ms transaction was being locked 1ms flushing data (in ordered mode) 1799ms logging transaction 11781 handles per transaction 5629 blocks per transaction 5641 logged blocks per transaction Signed-off-by: Johann Lombardi <johann.lombardi@bull.net> Signed-off-by: Mariusz Kozlowski <m.kozlowski@tuxland.pl> Signed-off-by: Mingming Cao <cmm@us.ibm.com> Signed-off-by: Eric Sandeen <sandeen@redhat.com>
* jbd2: Fix assertion failure in fs/jbd2/checkpoint.cJan Kara2008-01-281-6/+6
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | Before we start committing a transaction, we call __journal_clean_checkpoint_list() to cleanup transaction's written-back buffers. If this call happens to remove all of them (and there were already some buffers), __journal_remove_checkpoint() will decide to free the transaction because it isn't (yet) a committing transaction and soon we fail some assertion - the transaction really isn't ready to be freed :). We change the check in __journal_remove_checkpoint() to free only a transaction in T_FINISHED state. The locking there is subtle though (as everywhere in JBD ;(). We use j_list_lock to protect the check and a subsequent call to __journal_drop_transaction() and do the same in the end of journal_commit_transaction() which is the only place where a transaction can get to T_FINISHED state. Probably I'm too paranoid here and such locking is not really necessary - checkpoint lists are processed only from log_do_checkpoint() where a transaction must be already committed to be processed or from __journal_clean_checkpoint_list() where kjournald itself calls it and thus transaction cannot change state either. Better be safe if something changes in future... Signed-off-by: Jan Kara <jack@suse.cz> Cc: <linux-ext4@vger.kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
* fix file specification in commentsUwe Kleine-König2007-05-091-1/+1
| | | | | | | Many files include the filename at the beginning, serveral used a wrong one. Signed-off-by: Uwe Kleine-König <ukleinek@informatik.uni-freiburg.de> Signed-off-by: Adrian Bunk <bunk@stusta.de>
* [PATCH] jbd2: rename jbd2 symbols to avoid duplication of jbd symbolsMingming Cao2006-10-111-27/+27
| | | | | | | | | | Mingming Cao originally did this work, and Shaggy reproduced it using some scripts from her. Signed-off-by: Mingming Cao <cmm@us.ibm.com> Signed-off-by: Dave Kleikamp <shaggy@austin.ibm.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
* [PATCH] jbd2: initial copy of files from jbdDave Kleikamp2006-10-111-0/+697
This is a simple copy of the files in fs/jbd to fs/jbd2 and /usr/incude/linux/[ext4_]jbd.h to /usr/include/[ext4_]jbd2.h Signed-off-by: Dave Kleikamp <shaggy@austin.ibm.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
OpenPOWER on IntegriCloud