kernel.git/fs/btrfs/disk-io.c, branch linux-3.17.y

Btrfs: fix fs corruption on transaction abort if device supports discard

2015-01-08T18:27:49Z

commit 678886bdc6378c1cbd5072da2c5a3035000214e3 upstream. When we abort a transaction we iterate over all the ranges marked as dirty in fs_info->freed_extents[0] and fs_info->freed_extents[1], clear them from those trees, add them back (unpin) to the free space caches and, if the fs was mounted with "-o discard", perform a discard on those regions. Also, after adding the regions to the free space caches, a fitrim ioctl call can see those ranges in a block group's free space cache and perform a discard on the ranges, so the same issue can happen without "-o discard" as well. This causes corruption, affecting one or multiple btree nodes (in the worst case leaving the fs unmountable) because some of those ranges (the ones in the fs_info->pinned_extents tree) correspond to btree nodes/leafs that are referred by the last committed super block - breaking the rule that anything that was committed by a transaction is untouched until the next transaction commits successfully. I ran into this while running in a loop (for several hours) the fstest that I recently submitted: [PATCH] fstests: add btrfs test to stress chunk allocation/removal and fstrim The corruption always happened when a transaction aborted and then fsck complained like this: _check_btrfs_filesystem: filesystem on /dev/sdc is inconsistent *** fsck.btrfs output *** Check tree block failed, want=94945280, have=0 Check tree block failed, want=94945280, have=0 Check tree block failed, want=94945280, have=0 Check tree block failed, want=94945280, have=0 Check tree block failed, want=94945280, have=0 read block failed check_tree_block Couldn't open file system In this case 94945280 corresponded to the root of a tree. Using frace what I observed was the following sequence of steps happened: 1) transaction N started, fs_info->pinned_extents pointed to fs_info->freed_extents[0]; 2) node/eb 94945280 is created; 3) eb is persisted to disk; 4) transaction N commit starts, fs_info->pinned_extents now points to fs_info->freed_extents[1], and transaction N completes; 5) transaction N + 1 starts; 6) eb is COWed, and btrfs_free_tree_block() called for this eb; 7) eb range (94945280 to 94945280 + 16Kb) is added to fs_info->pinned_extents (fs_info->freed_extents[1]); 8) Something goes wrong in transaction N + 1, like hitting ENOSPC for example, and the transaction is aborted, turning the fs into readonly mode. The stack trace I got for example: [112065.253935] [] dump_stack+0x4d/0x66 [112065.254271] [] warn_slowpath_common+0x7f/0x98 [112065.254567] [] ? __btrfs_abort_transaction+0x50/0x10b [btrfs] [112065.261674] [] warn_slowpath_fmt+0x48/0x50 [112065.261922] [] ? btrfs_free_path+0x26/0x29 [btrfs] [112065.262211] [] __btrfs_abort_transaction+0x50/0x10b [btrfs] [112065.262545] [] btrfs_remove_chunk+0x537/0x58b [btrfs] [112065.262771] [] btrfs_delete_unused_bgs+0x1de/0x21b [btrfs] [112065.263105] [] cleaner_kthread+0x100/0x12f [btrfs] (...) [112065.264493] ---[ end trace dd7903a975a31a08 ]--- [112065.264673] BTRFS: error (device sdc) in btrfs_remove_chunk:2625: errno=-28 No space left [112065.264997] BTRFS info (device sdc): forced readonly 9) The clear kthread sees that the BTRFS_FS_STATE_ERROR bit is set in fs_info->fs_state and calls btrfs_cleanup_transaction(), which in turn calls btrfs_destroy_pinned_extent(); 10) Then btrfs_destroy_pinned_extent() iterates over all the ranges marked as dirty in fs_info->freed_extents[], and for each one it calls discard, if the fs was mounted with "-o discard", and adds the range to the free space cache of the respective block group; 11) btrfs_trim_block_group(), invoked from the fitrim ioctl code path, sees the free space entries and performs a discard; 12) After an umount and mount (or fsck), our eb's location on disk was full of zeroes, and it should have been untouched, because it was marked as dirty in the fs_info->pinned_extents tree, and therefore used by the trees that the last committed superblock points to. Fix this by not performing a discard and not adding the ranges to the free space caches - it's useless from this point since the fs is now in readonly mode and we won't write free space caches to disk anymore (otherwise we would leak space) nor any new superblock. By not adding the ranges to the free space caches, it prevents other code paths from allocating that space and write to it as well, therefore being safer and simpler. This isn't a new problem, as it's been present since 2011 (git commit acce952b0263825da32cf10489413dec78053347). Signed-off-by: Filipe Manana Signed-off-by: Chris Mason Signed-off-by: Greg Kroah-Hartman

Btrfs: make sure logged extents complete in the current transaction V3

2015-01-08T18:27:49Z

commit 50d9aa99bd35c77200e0e3dd7a72274f8304701f upstream. Liu Bo pointed out that my previous fix would lose the generation update in the scenario I described. It is actually much worse than that, we could lose the entire extent if we lose power right after the transaction commits. Consider the following write extent 0-4k log extent in log tree commit transaction < power fail happens here ordered extent completes We would lose the 0-4k extent because it hasn't updated the actual fs tree, and the transaction commit will reset the log so it isn't replayed. If we lose power before the transaction commit we are save, otherwise we are not. Fix this by keeping track of all extents we logged in this transaction. Then when we go to commit the transaction make sure we wait for all of those ordered extents to complete before proceeding. This will make sure that if we lose power after the transaction commit we still have our data. This also fixes the problem of the improperly updated extent generation. Thanks, Signed-off-by: Josef Bacik Signed-off-by: Chris Mason Signed-off-by: Greg Kroah-Hartman

Btrfs: fix task hang under heavy compressed write

2014-08-24T14:17:02Z

This has been reported and discussed for a long time, and this hang occurs in both 3.15 and 3.16. Btrfs now migrates to use kernel workqueue, but it introduces this hang problem. Btrfs has a kind of work queued as an ordered way, which means that its ordered_func() must be processed in the way of FIFO, so it usually looks like -- normal_work_helper(arg) work = container_of(arg, struct btrfs_work, normal_work); work->func() <---- (we name it work X) for ordered_work in wq->ordered_list ordered_work->ordered_func() ordered_work->ordered_free() The hang is a rare case, first when we find free space, we get an uncached block group, then we go to read its free space cache inode for free space information, so it will file a readahead request btrfs_readpages() for page that is not in page cache __do_readpage() submit_extent_page() btrfs_submit_bio_hook() btrfs_bio_wq_end_io() submit_bio() end_workqueue_bio() <--(ret by the 1st endio) queue a work(named work Y) for the 2nd also the real endio() So the hang occurs when work Y's work_struct and work X's work_struct happens to share the same address. A bit more explanation, A,B,C -- struct btrfs_work arg -- struct work_struct kthread: worker_thread() pick up a work_struct from @worklist process_one_work(arg) worker->current_work = arg; <-- arg is A->normal_work worker->current_func(arg) normal_work_helper(arg) A = container_of(arg, struct btrfs_work, normal_work); A->func() A->ordered_func() A->ordered_free() <-- A gets freed B->ordered_func() submit_compressed_extents() find_free_extent() load_free_space_inode() ... <-- (the above readhead stack) end_workqueue_bio() btrfs_queue_work(work C) B->ordered_free() As if work A has a high priority in wq->ordered_list and there are more ordered works queued after it, such as B->ordered_func(), its memory could have been freed before normal_work_helper() returns, which means that kernel workqueue code worker_thread() still has worker->current_work pointer to be work A->normal_work's, ie. arg's address. Meanwhile, work C is allocated after work A is freed, work C->normal_work and work A->normal_work are likely to share the same address(I confirmed this with ftrace output, so I'm not just guessing, it's rare though). When another kthread picks up work C->normal_work to process, and finds our kthread is processing it(see find_worker_executing_work()), it'll think work C as a collision and skip then, which ends up nobody processing work C. So the situation is that our kthread is waiting forever on work C. Besides, there're other cases that can lead to deadlock, but the real problem is that all btrfs workqueue shares one work->func, -- normal_work_helper, so this makes each workqueue to have its own helper function, but only a wraper pf normal_work_helper. With this patch, I no long hit the above hang. Signed-off-by: Liu Bo Signed-off-by: Chris Mason

Btrfs: Fix wrong device size when we are resizing the device

2014-08-19T15:52:18Z

total_bytes of device is just a in-memory variant which is used to record the size of the device, and it might be changed before we resize a device, if the resize operation fails, it will be fallbacked. But some code used it to update on-disk metadata of the device, it would cause the problem that on-disk metadata of the devices was not consistent. We should use the other variant named disk_total_bytes to update the on-disk metadata of device, because that variant is updated only when the resize operation is successful. Fix it. Signed-off-by: Miao Xie Signed-off-by: Chris Mason

btrfs: disable strict file flushes for renames and truncates

2014-08-15T14:43:42Z

Truncates and renames are often used to replace old versions of a file with new versions. Applications often expect this to be an atomic replacement, even if they haven't done anything to make sure the new version is fully on disk. Btrfs has strict flushing in place to make sure that renaming over an old file with a new file will fully flush out the new file before allowing the transaction commit with the rename to complete. This ordering means the commit code needs to be able to lock file pages, and there are a few paths in the filesystem where we will try to end a transaction with the page lock held. It's rare, but these things can deadlock. This patch removes the ordered flushes and switches to a best effort filemap_flush like ext4 uses. It's not perfect, but it should fix the deadlocks. Signed-off-by: Chris Mason

Btrfs: fix race between balance recovery and root deletion

2014-07-03T14:04:04Z

Balance recovery is called when RW mounting or remounting from RO to RW, it is called to finish roots merging. When doing balance recovery, relocation root's corresponding fs root(whose root refs is 0) might be destroyed by cleaner thread, this will make btrfs fail to mount. Fix this problem by holding @cleaner_mutex when doing balance recovery. Signed-off-by: Wang Shilong Signed-off-by: Chris Mason

btrfs: only unlock block in verify_parent_transid if we locked it

2014-06-28T20:48:47Z

This is a regression from my patch a26e8c9f75b0bfd8cccc9e8f110737b136eb5994, we need to only unlock the block if we were the one who locked it. Otherwise this will trip BUG_ON()'s in locking.c Thanks, cc: stable@vger.kernel.org Signed-off-by: Josef Bacik Signed-off-by: Chris Mason

Btrfs: async delayed refs

2014-06-10T00:20:58Z

Delayed extent operations are triggered during transaction commits. The goal is to queue up a healthly batch of changes to the extent allocation tree and run through them in bulk. This farms them off to async helper threads. The goal is to have the bulk of the delayed operations being done in the background, but this is also important to limit our stack footprint. Signed-off-by: Chris Mason

Btrfs: add sanity tests for new qgroup accounting code

2014-06-10T00:20:49Z

This exercises the various parts of the new qgroup accounting code. We do some basic stuff and do some things with the shared refs to make sure all that code works. I had to add a bunch of infrastructure because I needed to be able to insert items into a fake tree without having to do all the hard work myself, hopefully this will be usefull in the future. Thanks, Signed-off-by: Josef Bacik Signed-off-by: Chris Mason

Btrfs: rework qgroup accounting

2014-06-10T00:20:48Z

Currently qgroups account for space by intercepting delayed ref updates to fs trees. It does this by adding sequence numbers to delayed ref updates so that it can figure out how the tree looked before the update so we can adjust the counters properly. The problem with this is that it does not allow delayed refs to be merged, so if you say are defragging an extent with 5k snapshots pointing to it we will thrash the delayed ref lock because we need to go back and manually merge these things together. Instead we want to process quota changes when we know they are going to happen, like when we first allocate an extent, we free a reference for an extent, we add new references etc. This patch accomplishes this by only adding qgroup operations for real ref changes. We only modify the sequence number when we need to lookup roots for bytenrs, this reduces the amount of churn on the sequence number and allows us to merge delayed refs as we add them most of the time. This patch encompasses a bunch of architectural changes 1) qgroup ref operations: instead of tracking qgroup operations through the delayed refs we simply add new ref operations whenever we notice that we need to when we've modified the refs themselves. 2) tree mod seq: we no longer have this separation of major/minor counters. this makes the sequence number stuff much more sane and we can remove some locking that was needed to protect the counter. 3) delayed ref seq: we now read the tree mod seq number and use that as our sequence. This means each new delayed ref doesn't have it's own unique sequence number, rather whenever we go to lookup backrefs we inc the sequence number so we can make sure to keep any new operations from screwing up our world view at that given point. This allows us to merge delayed refs during runtime. With all of these changes the delayed ref stuff is a little saner and the qgroup accounting stuff no longer goes negative in some cases like it was before. Thanks, Signed-off-by: Josef Bacik Signed-off-by: Chris Mason