Our gluster boxes were hitting a problem where they'd run out of space when
updating the block group cache and therefore wouldn't be able to update the free
space inode. This is a problem because this is how we invalidate the cache and
protect ourselves from errors further down the stack, so if this fails we have
to abort the transaction so we make sure we don't end up with stale free space
cache. Thanks,
Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Chris Mason <clm@fb.com>
Pull btrfs updates from Chris Mason:
"This pull is mostly cleanups and fixes:
- The raid5/6 cleanups from Zhao Lei fixup some long standing warts
in the code and add improvements on top of the scrubbing support
from 3.19.
- Josef has round one of our ENOSPC fixes coming from large btrfs
clusters here at FB.
- Dave Sterba continues a long series of cleanups (thanks Dave), and
Filipe continues hammering on corner cases in fsync and others
This all was held up a little trying to track down a use-after-free in
btrfs raid5/6. It's not clear yet if this is just made easier to
trigger with this pull or if its a new bug from the raid5/6 cleanups.
Dave Sterba is the only one to trigger it so far, but he has a
consistent way to reproduce, so we'll get it nailed shortly"
* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/mason/linux-btrfs: (68 commits)
Btrfs: don't remove extents and xattrs when logging new names
Btrfs: fix fsync data loss after adding hard link to inode
Btrfs: fix BUG_ON in btrfs_orphan_add() when delete unused block group
Btrfs: account for large extents with enospc
Btrfs: don't set and clear delalloc for O_DIRECT writes
Btrfs: only adjust outstanding_extents when we do a short write
btrfs: Fix out-of-space bug
Btrfs: scrub, fix sleep in atomic context
Btrfs: fix scheduler warning when syncing log
Btrfs: Remove unnecessary placeholder in btrfs_err_code
btrfs: cleanup init for list in free-space-cache
btrfs: delete chunk allocation attemp when setting block group ro
btrfs: clear bio reference after submit_one_bio()
Btrfs: fix scrub race leading to use-after-free
Btrfs: add missing cleanup on sysfs init failure
Btrfs: fix race between transaction commit and empty block group removal
btrfs: add more checks to btrfs_read_sys_array
btrfs: cleanup, rename a few variables in btrfs_read_sys_array
btrfs: add checks for sys_chunk_array sizes
btrfs: more superblock checks, lower bounds on devices and sectorsize/nodesize
...
int alloc_chunk is never used in this function, remove it.
Signed-off-by: Zhao Lei <zhaolei@cn.fujitsu.com>
Signed-off-by: David Sterba <dsterba@suse.cz>
This is the 3rd independent patch of a larger project to cleanup btrfs's
internal usage of btrfs_root. Many functions take btrfs_root only to
grab the fs_info struct.
By requiring a root these functions cause programmer overhead. That
these functions can accept any valid root is not obvious until
inspection.
This patch reduces the specificity of such functions to accept the
fs_info directly.
These patches can be applied independently and thus are not being
submitted as a patch series. There should be about 26 patches by the
project's completion. Each patch will cleanup between 1 and 34 functions
apiece. Each patch covers a single file's functions.
This patch affects the following function(s):
1) csum_tree_block
2) csum_dirty_buffer
3) check_tree_block_fsid
4) btrfs_find_tree_block
5) clean_tree_block
Signed-off-by: Daniel Dressler <danieru.dressler@gmail.com>
Signed-off-by: David Sterba <dsterba@suse.cz>
On our gluster boxes we stream large tar balls of backups onto our fses. With
160gb of ram this means we get really large contiguous ranges of dirty data, but
the way our ENOSPC stuff works is that as long as it's contiguous we only hold
metadata reservation for one extent. The problem is we limit our extents to
128mb, so we'll end up with at least 800 extents so our enospc accounting is
quite a bit lower than what we need. To keep track of this make sure we
increase outstanding_extents for every multiple of the max extent size so we can
be sure to have enough reserved metadata space. Thanks,
Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Chris Mason <clm@fb.com>
Below test will fail currently:
mkfs.ext4 -F /dev/sda
btrfs-convert /dev/sda
mount /dev/sda /mnt
btrfs device add -f /dev/sdb /mnt
btrfs balance start -v -dconvert=raid1 -mconvert=raid1 /mnt
The reason is there are some block groups with usage 0, but the whole
disk hasn't free space to allocate new chunk, so we even can't set such
block group readonly. This patch deletes the chunk allocation when
setting block group ro. For META, we already have reserve. But for
SYSTEM, we don't have, so the check_system_chunk is still required.
Signed-off-by: Shaohua Li <shli@fb.com>
Signed-off-by: Chris Mason <clm@fb.com>
"run_most" is not used anymore.
Signed-off-by: Liu Bo <bo.li.liu@oracle.com>
Reviewed-by: Satoru Takeuchi <takeuchi_satoru@jp.fujitsu.com>
Signed-off-by: Chris Mason <clm@fb.com>
1: ref_count is simple than current RBIO_HOLD_BBIO_MAP_BIT flag
to keep btrfs_bio's memory in raid56 recovery implement.
2: free function for bbio will make code clean and flexible, plus
forced data type checking in compile.
Changelog v1->v2:
Rename following by David Sterba's suggestion:
put_btrfs_bio() -> btrfs_put_bio()
get_btrfs_bio() -> btrfs_get_bio()
bbio->ref_count -> bbio->refs
Signed-off-by: Zhao Lei <zhaolei@cn.fujitsu.com>
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Chris Mason <clm@fb.com>
Very often our extent buffer's header generation doesn't match the current
transaction's id or it is also referenced by other trees (snapshots), so
we don't need the corresponding block group cache object. Therefore only
search for it if we are going to use it, so we avoid an unnecessary search
in the block groups rbtree (and acquiring and releasing its spinlock).
Freeing a tree block is performed when COWing or deleting a node/leaf,
which implies we are holding the node/leaf's parent node lock, therefore
reducing the amount of time spent when freeing a tree block helps reducing
the amount of time we are holding the parent node's lock.
For example, for a run of xfstests/generic/083, the block group cache
object was needed only 682 times for a total of 226691 calls to free
a tree block.
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
Currently any time we try to update the block groups on disk we will walk _all_
block groups and check for the ->dirty flag to see if it is set. This function
can get called several times during a commit. So if you have several terabytes
of data you will be a very sad panda as we will loop through _all_ of the block
groups several times, which makes the commit take a while which slows down the
rest of the file system operations.
This patch introduces a dirty list for the block groups that we get added to
when we dirty the block group for the first time. Then we simply update any
block groups that have been dirtied since the last time we called
btrfs_write_dirty_block_groups. This allows us to clean up how we write the
free space cache out so it is much cleaner. Thanks,
Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Chris Mason <clm@fb.com>
When removing a block group we were deleting it from its space_info's
ro_bgs list without the correct protection - the space info's spinlock.
Fix this by doing the list delete while holding the spinlock of the
corresponding space info, which is the correct lock for any operation
on that list.
This issue was introduced in the 3.19 kernel by the following change:
Btrfs: move read only block groups onto their own list V2
commit 633c0aad4c
I ran into a kernel crash while a task was running statfs, which iterates
the space_info->ro_bgs list while holding the space info's spinlock,
and another task was deleting it from the same list, without holding that
spinlock, as part of the block group remove operation (while running the
function btrfs_remove_block_group). This happened often when running the
stress test xfstests/generic/038 I recently made.
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
We shouldn't BUG_ON() if there is corruption. I hit this while testing my block
group patch and the abort worked properly. Thanks,
Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Chris Mason <clm@fb.com>
Finally it's clear that the requested blocksize is always equal to
nodesize, with one exception, the superblock.
Superblock has fixed size regardless of the metadata block size, but
uses the same helpers to initialize sys array/chunk tree and to work
with the chunk items. So it pretends to be an extent_buffer for a
moment, btrfs_read_sys_array is full of special cases, we're adding one
more.
Signed-off-by: David Sterba <dsterba@suse.cz>
It doesn't do anything special, it just calls btrfs_discard_extent(),
so just remove it.
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
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] [<ffffffff8140c7b6>] dump_stack+0x4d/0x66
[112065.254271] [<ffffffff81042984>] warn_slowpath_common+0x7f/0x98
[112065.254567] [<ffffffffa0325990>] ? __btrfs_abort_transaction+0x50/0x10b [btrfs]
[112065.261674] [<ffffffff810429e5>] warn_slowpath_fmt+0x48/0x50
[112065.261922] [<ffffffffa032949e>] ? btrfs_free_path+0x26/0x29 [btrfs]
[112065.262211] [<ffffffffa0325990>] __btrfs_abort_transaction+0x50/0x10b [btrfs]
[112065.262545] [<ffffffffa036b1d6>] btrfs_remove_chunk+0x537/0x58b [btrfs]
[112065.262771] [<ffffffffa033840f>] btrfs_delete_unused_bgs+0x1de/0x21b [btrfs]
[112065.263105] [<ffffffffa0343106>] 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
acce952b02).
Cc: stable@vger.kernel.org # any kernel released after 2011-01-06
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
Always clear a block group's rbnode after removing it from the rbtree to
ensure that any tasks that might be holding a reference on the block group
don't end up accessing stale rbnode left and right child pointers through
next_block_group().
This is a leftover from the change titled:
"Btrfs: fix invalid block group rbtree access after bg is removed"
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
This was written when we didn't do a caching control for the fast free space
cache loading. However we started doing that a long time ago, and there is
still a small window of time that we could be caching the block group the fast
way, so if there is a caching_ctl at all on the block group just return it, the
callers all wait properly for what they want. Thanks,
Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Chris Mason <clm@fb.com>
On block group remove if the corresponding extent map was on the
transaction->pending_chunks list, we were deleting the extent map
from that list, through remove_extent_mapping(), without any
synchronization with chunk allocation (which iterates that list
and adds new elements to it). Fix this by ensure that this is done
while the chunk mutex is held, since that's the mutex that protects
the list in the chunk allocation code path.
This applies on top (depends on) of my previous patch titled:
"Btrfs: fix race between fs trimming and block group remove/allocation"
But the issue in fact was already present before that change, it only
became easier to hit after Josef's 3.18 patch that added automatic
removal of empty block groups.
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
On chunk allocation error (label "error_del_extent"), after adding the
extent map to the tree and to the pending chunks list, we would leave
decrementing the extent map's refcount by 2 instead of 3 (our allocation
+ tree reference + list reference).
Also, on chunk/block group removal, if the block group was on the list
pending_chunks we weren't decrementing the respective list reference.
Detected by 'rmmod btrfs':
[20770.105881] kmem_cache_destroy btrfs_extent_map: Slab cache still has objects
[20770.106127] CPU: 2 PID: 11093 Comm: rmmod Tainted: G W L 3.17.0-rc5-btrfs-next-1+ #1
[20770.106128] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.7.5-0-ge51488c-20140602_164612-nilsson.home.kraxel.org 04/01/2014
[20770.106130] 0000000000000000 ffff8800ba867eb8 ffffffff813e7a13 ffff8800a2e11040
[20770.106132] ffff8800ba867ed0 ffffffff81105d0c 0000000000000000 ffff8800ba867ee0
[20770.106134] ffffffffa035d65e ffff8800ba867ef0 ffffffffa03b0654 ffff8800ba867f78
[20770.106136] Call Trace:
[20770.106142] [<ffffffff813e7a13>] dump_stack+0x45/0x56
[20770.106145] [<ffffffff81105d0c>] kmem_cache_destroy+0x4b/0x90
[20770.106164] [<ffffffffa035d65e>] extent_map_exit+0x1a/0x1c [btrfs]
[20770.106176] [<ffffffffa03b0654>] exit_btrfs_fs+0x27/0x9d3 [btrfs]
[20770.106179] [<ffffffff8109dc97>] SyS_delete_module+0x153/0x1c4
[20770.106182] [<ffffffff8121261b>] ? trace_hardirqs_on_thunk+0x3a/0x3c
[20770.106184] [<ffffffff813ebf52>] system_call_fastpath+0x16/0x1b
This applies on top (depends on) of my previous patch titled:
"Btrfs: fix race between fs trimming and block group remove/allocation"
But the issue in fact was already present before that change, it only
became easier to hit after Josef's 3.18 patch that added automatic
removal of empty block groups.
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
If the transaction handle doesn't have used blocks but has created new block
groups make sure we turn the fs into readonly mode too. This is because the
new block groups didn't get all their metadata persisted into the chunk and
device trees, and therefore if a subsequent transaction starts, allocates
space from the new block groups, writes data or metadata into that space,
commits successfully and then after we unmount and mount the filesystem
again, the same space can be allocated again for a new block group,
resulting in file data or metadata corruption.
Example where we don't abort the transaction when we fail to finish the
chunk allocation (add items to the chunk and device trees) and later a
future transaction where the block group is removed fails because it can't
find the chunk item in the chunk tree:
[25230.404300] WARNING: CPU: 0 PID: 7721 at fs/btrfs/super.c:260 __btrfs_abort_transaction+0x50/0xfc [btrfs]()
[25230.404301] BTRFS: Transaction aborted (error -28)
[25230.404302] Modules linked in: btrfs dm_flakey nls_utf8 fuse xor raid6_pq ntfs vfat msdos fat xfs crc32c_generic libcrc32c ext3 jbd ext2 dm_mod nfsd auth_rpcgss oid_registry nfs_acl nfs lockd fscache sunrpc loop psmouse i2c_piix4 i2ccore parport_pc parport processor button pcspkr serio_raw thermal_sys evdev microcode ext4 crc16 jbd2 mbcache sr_mod cdrom ata_generic sg sd_mod crc_t10dif crct10dif_generic crct10dif_common virtio_scsi floppy e1000 ata_piix libata virtio_pci virtio_ring scsi_mod virtio [last unloaded: btrfs]
[25230.404325] CPU: 0 PID: 7721 Comm: xfs_io Not tainted 3.17.0-rc5-btrfs-next-1+ #1
[25230.404326] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.7.5-0-ge51488c-20140602_164612-nilsson.home.kraxel.org 04/01/2014
[25230.404328] 0000000000000000 ffff88004581bb08 ffffffff813e7a13 ffff88004581bb50
[25230.404330] ffff88004581bb40 ffffffff810423aa ffffffffa049386a 00000000ffffffe4
[25230.404332] ffffffffa05214c0 000000000000240c ffff88010fc8f800 ffff88004581bba8
[25230.404334] Call Trace:
[25230.404338] [<ffffffff813e7a13>] dump_stack+0x45/0x56
[25230.404342] [<ffffffff810423aa>] warn_slowpath_common+0x7f/0x98
[25230.404351] [<ffffffffa049386a>] ? __btrfs_abort_transaction+0x50/0xfc [btrfs]
[25230.404353] [<ffffffff8104240b>] warn_slowpath_fmt+0x48/0x50
[25230.404362] [<ffffffffa049386a>] __btrfs_abort_transaction+0x50/0xfc [btrfs]
[25230.404374] [<ffffffffa04a8c43>] btrfs_create_pending_block_groups+0x10c/0x135 [btrfs]
[25230.404387] [<ffffffffa04b77fd>] __btrfs_end_transaction+0x7e/0x2de [btrfs]
[25230.404398] [<ffffffffa04b7a6d>] btrfs_end_transaction+0x10/0x12 [btrfs]
[25230.404408] [<ffffffffa04a3d64>] btrfs_check_data_free_space+0x111/0x1f0 [btrfs]
[25230.404421] [<ffffffffa04c53bd>] __btrfs_buffered_write+0x160/0x48d [btrfs]
[25230.404425] [<ffffffff811a9268>] ? cap_inode_need_killpriv+0x2d/0x37
[25230.404429] [<ffffffff810f6501>] ? get_page+0x1a/0x2b
[25230.404441] [<ffffffffa04c7c95>] btrfs_file_write_iter+0x321/0x42f [btrfs]
[25230.404443] [<ffffffff8110f5d9>] ? handle_mm_fault+0x7f3/0x846
[25230.404446] [<ffffffff813e98c5>] ? mutex_unlock+0x16/0x18
[25230.404449] [<ffffffff81138d68>] new_sync_write+0x7c/0xa0
[25230.404450] [<ffffffff81139401>] vfs_write+0xb0/0x112
[25230.404452] [<ffffffff81139c9d>] SyS_pwrite64+0x66/0x84
[25230.404454] [<ffffffff813ebf52>] system_call_fastpath+0x16/0x1b
[25230.404455] ---[ end trace 5aa5684fdf47ab38 ]---
[25230.404458] BTRFS warning (device sdc): btrfs_create_pending_block_groups:9228: Aborting unused transaction(No space left).
[25288.084814] BTRFS: error (device sdc) in btrfs_free_chunk:2509: errno=-2 No such entry (Failed lookup while freeing chunk.)
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Chris Mason <clm@fb.com>
Our fs trim operation, which is completely transactionless (doesn't start
or joins an existing transaction) consists of visiting all block groups
and then for each one to iterate its free space entries and perform a
discard operation against the space range represented by the free space
entries. However before performing a discard, the corresponding free space
entry is removed from the free space rbtree, and when the discard completes
it is added back to the free space rbtree.
If a block group remove operation happens while the discard is ongoing (or
before it starts and after a free space entry is hidden), we end up not
waiting for the discard to complete, remove the extent map that maps
logical address to physical addresses and the corresponding chunk metadata
from the the chunk and device trees. After that and before the discard
completes, the current running transaction can finish and a new one start,
allowing for new block groups that map to the same physical addresses to
be allocated and written to.
So fix this by keeping the extent map in memory until the discard completes
so that the same physical addresses aren't reused before it completes.
If the physical locations that are under a discard operation end up being
used for a new metadata block group for example, and dirty metadata extents
are written before the discard finishes (the VM might call writepages() of
our btree inode's i_mapping for example, or an fsync log commit happens) we
end up overwriting metadata with zeroes, which leads to errors from fsck
like the following:
checking extents
Check tree block failed, want=833912832, have=0
Check tree block failed, want=833912832, have=0
Check tree block failed, want=833912832, have=0
Check tree block failed, want=833912832, have=0
Check tree block failed, want=833912832, have=0
read block failed check_tree_block
owner ref check failed [833912832 16384]
Errors found in extent allocation tree or chunk allocation
checking free space cache
checking fs roots
Check tree block failed, want=833912832, have=0
Check tree block failed, want=833912832, have=0
Check tree block failed, want=833912832, have=0
Check tree block failed, want=833912832, have=0
Check tree block failed, want=833912832, have=0
read block failed check_tree_block
root 5 root dir 256 error
root 5 inode 260 errors 2001, no inode item, link count wrong
unresolved ref dir 256 index 0 namelen 8 name foobar_3 filetype 1 errors 6, no dir index, no inode ref
root 5 inode 262 errors 2001, no inode item, link count wrong
unresolved ref dir 256 index 0 namelen 8 name foobar_5 filetype 1 errors 6, no dir index, no inode ref
root 5 inode 263 errors 2001, no inode item, link count wrong
(...)
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
There's a race between adding a block group to the list of the unused
block groups and removing an unused block group (cleaner kthread) that
leads to freeing extents that are in use or a crash during transaction
commmit. Basically the cleaner kthread, when executing
btrfs_delete_unused_bgs(), might catch the newly added block group to
the list fs_info->unused_bgs and clear the range representing the whole
group from fs_info->freed_extents[] before the task that added the block
group to the list (running update_block_group()) marked the last freed
extent as dirty in fs_info->freed_extents (pinned_extents).
That is:
CPU 1 CPU 2
btrfs_delete_unused_bgs()
update_block_group()
add block group to
fs_info->unused_bgs
got block group from the list
clear_extent_bits for the whole
block group range in freed_extents[]
set_extent_dirty for the
range covering the freed
extent in freed_extents[]
(fs_info->pinned_extents)
block group deleted, and a new block
group with the same logical address is
created
reserve space from the new block group
for new data or metadata - the reserved
space overlaps the range specified by
CPU 1 for set_extent_dirty()
commit transaction
find all ranges marked as dirty in
fs_info->pinned_extents, clear them
and add them to the free space cache
Alternatively, if CPU 2 doesn't create a new block group with the same
logical address, we get a crash/BUG_ON at transaction commit when unpining
extent ranges because we can't find a block group for the range marked as
dirty by CPU 1. Sample trace:
[ 2163.426462] invalid opcode: 0000 [#1] SMP DEBUG_PAGEALLOC
[ 2163.426640] Modules linked in: btrfs xor raid6_pq dm_thin_pool dm_persistent_data dm_bio_prison dm_bufio crc32c_generic libcrc32c dm_mod nfsd auth_rpc
gss oid_registry nfs_acl nfs lockd fscache sunrpc loop psmouse parport_pc parport i2c_piix4 processor thermal_sys i2ccore evdev button pcspkr microcode serio_raw ext4 crc16 jbd2 mbcache
sg sr_mod cdrom sd_mod crc_t10dif crct10dif_generic crct10dif_common ata_generic virtio_scsi floppy ata_piix libata e1000 scsi_mod virtio_pci virtio_ring virtio
[ 2163.428209] CPU: 0 PID: 11858 Comm: btrfs-transacti Tainted: G W 3.17.0-rc5-btrfs-next-1+ #1
[ 2163.428519] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.7.5-0-ge51488c-20140602_164612-nilsson.home.kraxel.org 04/01/2014
[ 2163.428875] task: ffff88009f2c0650 ti: ffff8801356bc000 task.ti: ffff8801356bc000
[ 2163.429157] RIP: 0010:[<ffffffffa037728e>] [<ffffffffa037728e>] unpin_extent_range.isra.58+0x62/0x192 [btrfs]
[ 2163.429562] RSP: 0018:ffff8801356bfda8 EFLAGS: 00010246
[ 2163.429802] RAX: 0000000000000000 RBX: 0000000000000000 RCX: 0000000000000000
[ 2163.429990] RDX: 0000000041bfffff RSI: 0000000001c00000 RDI: ffff880024307080
[ 2163.430042] RBP: ffff8801356bfde8 R08: 0000000000000068 R09: ffff88003734f118
[ 2163.430042] R10: ffff8801356bfcb8 R11: fffffffffffffb69 R12: ffff8800243070d0
[ 2163.430042] R13: 0000000083c04000 R14: ffff8800751b0f00 R15: ffff880024307000
[ 2163.430042] FS: 0000000000000000(0000) GS:ffff88013f400000(0000) knlGS:0000000000000000
[ 2163.430042] CS: 0010 DS: 0000 ES: 0000 CR0: 000000008005003b
[ 2163.430042] CR2: 00007ff10eb43fc0 CR3: 0000000004cb8000 CR4: 00000000000006f0
[ 2163.430042] Stack:
[ 2163.430042] ffff8800243070d0 0000000083c08000 0000000083c07fff ffff88012d6bc800
[ 2163.430042] ffff8800243070d0 ffff8800751b0f18 ffff8800751b0f00 0000000000000000
[ 2163.430042] ffff8801356bfe18 ffffffffa037a481 0000000083c04000 0000000083c07fff
[ 2163.430042] Call Trace:
[ 2163.430042] [<ffffffffa037a481>] btrfs_finish_extent_commit+0xac/0xbf [btrfs]
[ 2163.430042] [<ffffffffa038c06d>] btrfs_commit_transaction+0x6ee/0x882 [btrfs]
[ 2163.430042] [<ffffffffa03881f1>] transaction_kthread+0xf2/0x1a4 [btrfs]
[ 2163.430042] [<ffffffffa03880ff>] ? btrfs_cleanup_transaction+0x3d8/0x3d8 [btrfs]
[ 2163.430042] [<ffffffff8105966b>] kthread+0xb7/0xbf
[ 2163.430042] [<ffffffff810595b4>] ? __kthread_parkme+0x67/0x67
[ 2163.430042] [<ffffffff813ebeac>] ret_from_fork+0x7c/0xb0
[ 2163.430042] [<ffffffff810595b4>] ? __kthread_parkme+0x67/0x67
So fix this by making update_block_group() first set the range as dirty
in pinned_extents before adding the block group to the unused_bgs list.
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Chris Mason <clm@fb.com>
If we grab a block group, for example in btrfs_trim_fs(), we will be holding
a reference on it but the block group can be removed after we got it (via
btrfs_remove_block_group), which means it will no longer be part of the
rbtree.
However, btrfs_remove_block_group() was only calling rb_erase() which leaves
the block group's rb_node left and right child pointers with the same content
they had before calling rb_erase. This was dangerous because a call to
next_block_group() would access the node's left and right child pointers (via
rb_next), which can be no longer valid.
Fix this by clearing a block group's node after removing it from the tree,
and have next_block_group() do a tree search to get the next block group
instead of using rb_next() if our block group was removed.
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Chris Mason <clm@fb.com>
If right after starting the snapshot creation ioctl we perform a write against a
file followed by a truncate, with both operations increasing the file's size, we
can get a snapshot tree that reflects a state of the source subvolume's tree where
the file truncation happened but the write operation didn't. This leaves a gap
between 2 file extent items of the inode, which makes btrfs' fsck complain about it.
For example, if we perform the following file operations:
$ mkfs.btrfs -f /dev/vdd
$ mount /dev/vdd /mnt
$ xfs_io -f \
-c "pwrite -S 0xaa -b 32K 0 32K" \
-c "fsync" \
-c "pwrite -S 0xbb -b 32770 16K 32770" \
-c "truncate 90123" \
/mnt/foobar
and the snapshot creation ioctl was just called before the second write, we often
can get the following inode items in the snapshot's btree:
item 120 key (257 INODE_ITEM 0) itemoff 7987 itemsize 160
inode generation 146 transid 7 size 90123 block group 0 mode 100600 links 1 uid 0 gid 0 rdev 0 flags 0x0
item 121 key (257 INODE_REF 256) itemoff 7967 itemsize 20
inode ref index 282 namelen 10 name: foobar
item 122 key (257 EXTENT_DATA 0) itemoff 7914 itemsize 53
extent data disk byte 1104855040 nr 32768
extent data offset 0 nr 32768 ram 32768
extent compression 0
item 123 key (257 EXTENT_DATA 53248) itemoff 7861 itemsize 53
extent data disk byte 0 nr 0
extent data offset 0 nr 40960 ram 40960
extent compression 0
There's a file range, corresponding to the interval [32K; ALIGN(16K + 32770, 4096)[
for which there's no file extent item covering it. This is because the file write
and file truncate operations happened both right after the snapshot creation ioctl
called btrfs_start_delalloc_inodes(), which means we didn't start and wait for the
ordered extent that matches the write and, in btrfs_setsize(), we were able to call
btrfs_cont_expand() before being able to commit the current transaction in the
snapshot creation ioctl. So this made it possibe to insert the hole file extent
item in the source subvolume (which represents the region added by the truncate)
right before the transaction commit from the snapshot creation ioctl.
Btrfs' fsck tool complains about such cases with a message like the following:
"root 331 inode 257 errors 100, file extent discount"
>From a user perspective, the expectation when a snapshot is created while those
file operations are being performed is that the snapshot will have a file that
either:
1) is empty
2) only the first write was captured
3) only the 2 writes were captured
4) both writes and the truncation were captured
But never capture a state where only the first write and the truncation were
captured (since the second write was performed before the truncation).
A test case for xfstests follows.
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
Due to ignoring errors returned by clear_extent_bits (at the moment only
-ENOMEM is possible), we can end up freeing an extent that is actually in
use (i.e. return the extent to the free space cache).
The sequence of steps that lead to this:
1) Cleaner thread starts execution and calls btrfs_delete_unused_bgs(), with
the goal of freeing empty block groups;
2) btrfs_delete_unused_bgs() finds an empty block group, joins the current
transaction (or starts a new one if none is running) and attempts to
clear the EXTENT_DIRTY bit for the block group's range from freed_extents[0]
and freed_extents[1] (of which one corresponds to fs_info->pinned_extents);
3) Clearing the EXTENT_DIRTY bit (via clear_extent_bits()) fails with
-ENOMEM, but such error is ignored and btrfs_delete_unused_bgs() proceeds
to delete the block group and the respective chunk, while pinned_extents
remains with that bit set for the whole (or a part of the) range covered
by the block group;
4) Later while the transaction is still running, the chunk ends up being reused
for a new block group (maybe for different purpose, data or metadata), and
extents belonging to the new block group are allocated for file data or btree
nodes/leafs;
5) The current transaction is committed, meaning that we unpinned one or more
extents from the new block group (through btrfs_finish_extent_commit() and
unpin_extent_range()) which are now being used for new file data or new
metadata (through btrfs_finish_extent_commit() and unpin_extent_range()).
And unpinning means we returned the extents to the free space cache of the
new block group, which implies those extents can be used for future allocations
while they're still in use.
Alternatively, we can hit a BUG_ON() when doing a lookup for a block group's cache
object in unpin_extent_range() if a new block group didn't end up being allocated for
the same chunk (step 4 above).
Fix this by not freeing the block group and chunk if we fail to clear the dirty bit.
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
Our gluster boxes were spending lots of time in statfs because our fs'es are
huge. The problem is statfs loops through all of the block groups looking for
read only block groups, and when you have several terabytes worth of data that
ends up being a lot of block groups. Move the read only block groups onto a
read only list and only proces that list in
btrfs_account_ro_block_groups_free_space to reduce the amount of churn. Thanks,
Signed-off-by: Josef Bacik <jbacik@fb.com>
Reviewed-by: Liu Bo <bo.li.liu@oracle.com>
Signed-off-by: Chris Mason <clm@fb.com>
Pull btrfs fixes from Chris Mason:
"Filipe is nailing down some problems with our skinny extent variation,
and Dave's patch fixes endian problems in the new super block checks"
* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/mason/linux-btrfs:
Btrfs: fix race that makes btrfs_lookup_extent_info miss skinny extent items
Btrfs: properly clean up btrfs_end_io_wq_cache
Btrfs: fix invalid leaf slot access in btrfs_lookup_extent()
btrfs: use macro accessors in superblock validation checks
We have a race that can lead us to miss skinny extent items in the function
btrfs_lookup_extent_info() when the skinny metadata feature is enabled.
So basically the sequence of steps is:
1) We search in the extent tree for the skinny extent, which returns > 0
(not found);
2) We check the previous item in the returned leaf for a non-skinny extent,
and we don't find it;
3) Because we didn't find the non-skinny extent in step 2), we release our
path to search the extent tree again, but this time for a non-skinny
extent key;
4) Right after we released our path in step 3), a skinny extent was inserted
in the extent tree (delayed refs were run) - our second extent tree search
will miss it, because it's not looking for a skinny extent;
5) After the second search returned (with ret > 0), we look for any delayed
ref for our extent's bytenr (and we do it while holding a read lock on the
leaf), but we won't find any, as such delayed ref had just run and completed
after we released out path in step 3) before doing the second search.
Fix this by removing completely the path release and re-search logic. This is
safe, because if we seach for a metadata item and we don't find it, we have the
guarantee that the returned leaf is the one where the item would be inserted,
and so path->slots[0] > 0 and path->slots[0] - 1 must be the slot where the
non-skinny extent item is if it exists. The only case where path->slots[0] is
zero is when there are no smaller keys in the tree (i.e. no left siblings for
our leaf), in which case the re-search logic isn't needed as well.
This race has been present since the introduction of skinny metadata (change
3173a18f70).
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Chris Mason <clm@fb.com>
If we couldn't find our extent item, we accessed the current slot
(path->slots[0]) to check if it corresponds to an equivalent skinny
metadata item. However this slot could be beyond our last item in the
leaf (i.e. path->slots[0] >= btrfs_header_nritems(leaf)), in which case
we shouldn't process it.
Since btrfs_lookup_extent() is only used to find extent items for data
extents, fix this by removing completely the logic that looks up for an
equivalent skinny metadata item, since it can not exist.
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
Pull btrfs updates from Chris Mason:
"The largest set of changes here come from Miao Xie. He's cleaning up
and improving read recovery/repair for raid, and has a number of
related fixes.
I've merged another set of fsync fixes from Filipe, and he's also
improved the way we handle metadata write errors to make sure we force
the FS readonly if things go wrong.
Otherwise we have a collection of fixes and cleanups. Dave Sterba
gets a cookie for removing the most lines (thanks Dave)"
* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/mason/linux-btrfs: (139 commits)
btrfs: Fix compile error when CONFIG_SECURITY is not set.
Btrfs: fix compiles when CONFIG_BTRFS_FS_RUN_SANITY_TESTS is off
btrfs: Make btrfs handle security mount options internally to avoid losing security label.
Btrfs: send, don't delay dir move if there's a new parent inode
btrfs: add more superblock checks
Btrfs: fix race in WAIT_SYNC ioctl
Btrfs: be aware of btree inode write errors to avoid fs corruption
Btrfs: remove redundant btrfs_verify_qgroup_counts declaration.
btrfs: fix shadow warning on cmp
Btrfs: fix compilation errors under DEBUG
Btrfs: fix crash of btrfs_release_extent_buffer_page
Btrfs: add missing end_page_writeback on submit_extent_page failure
btrfs: Fix the wrong condition judgment about subset extent map
Btrfs: fix build_backref_tree issue with multiple shared blocks
Btrfs: cleanup error handling in build_backref_tree
btrfs: move checks for DUMMY_ROOT into a helper
btrfs: new define for the inline extent data start
btrfs: kill extent_buffer_page helper
btrfs: drop constant param from btrfs_release_extent_buffer_page
btrfs: hide typecast to definition of BTRFS_SEND_TRANS_STUB
...
While we have a transaction ongoing, the VM might decide at any time
to call btree_inode->i_mapping->a_ops->writepages(), which will start
writeback of dirty pages belonging to btree nodes/leafs. This call
might return an error or the writeback might finish with an error
before we attempt to commit the running transaction. If this happens,
we might have no way of knowing that such error happened when we are
committing the transaction - because the pages might no longer be
marked dirty nor tagged for writeback (if a subsequent modification
to the extent buffer didn't happen before the transaction commit) which
makes filemap_fdata[write|wait]_range unable to find such pages (even
if they're marked with SetPageError).
So if this happens we must abort the transaction, otherwise we commit
a super block with btree roots that point to btree nodes/leafs whose
content on disk is invalid - either garbage or the content of some
node/leaf from a past generation that got cowed or deleted and is no
longer valid (for this later case we end up getting error messages like
"parent transid verify failed on 10826481664 wanted 25748 found 29562"
when reading btree nodes/leafs from disk).
Note that setting and checking AS_EIO/AS_ENOSPC in the btree inode's
i_mapping would not be enough because we need to distinguish between
log tree extents (not fatal) vs non-log tree extents (fatal) and
because the next call to filemap_fdatawait_range() will catch and clear
such errors in the mapping - and that call might be from a log sync and
not from a transaction commit, which means we would not know about the
error at transaction commit time. Also, checking for the eb flag
EXTENT_BUFFER_IOERR at transaction commit time isn't done and would
not be completely reliable, as the eb might be removed from memory and
read back when trying to get it, which clears that flag right before
reading the eb's pages from disk, making us not know about the previous
write error.
Using the new 3 flags for the btree inode also makes us achieve the
goal of AS_EIO/AS_ENOSPC when writepages() returns success, started
writeback for all dirty pages and before filemap_fdatawait_range() is
called, the writeback for all dirty pages had already finished with
errors - because we were not using AS_EIO/AS_ENOSPC,
filemap_fdatawait_range() would return success, as it could not know
that writeback errors happened (the pages were no longer tagged for
writeback).
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
Rename to btrfs_alloc_tree_block as it fits to the alloc/find/free +
_tree_block family. The parameter blocksize was set to the metadata
block size, directly or indirectly.
Signed-off-by: David Sterba <dsterba@suse.cz>
The parent_transid parameter has been unused since its introduction in
ca7a79ad8d ("Pass down the expected generation number when reading
tree blocks"). In reada_tree_block, it was even wrongly set to leafsize.
Transid check is done in the proper read and readahead ignores errors.
Signed-off-by: David Sterba <dsterba@suse.cz>
There are the branch hints that obviously depend on the data being
processed, the CPU predictor will do better job according to the actual
load. It also does not make sense to use the hints in slow paths that do
a lot of other operations like locking, waiting or IO.
Signed-off-by: David Sterba <dsterba@suse.cz>
This is to receive 0a30288da1 ("blk-mq, percpu_ref: implement a
kludge for SCSI blk-mq stall during probe") which implements
__percpu_ref_kill_expedited() to work around SCSI blk-mq stall. The
commit reverted and patches to implement proper fix will be added.
Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Kent Overstreet <kmo@daterainc.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Christoph Hellwig <hch@lst.de>
Trying to reproduce a log enospc bug I hit a panic in the async reclaim code
during log replay. This is because we use fs_info->fs_root as our root for
shrinking and such. Technically we can use whatever root we want, but let's
just not allow async reclaim while we're doing log replay. Thanks,
Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Chris Mason <clm@fb.com>
One problem that has plagued us is that a user will use up all of his space with
data, remove a bunch of that data, and then try to create a bunch of small files
and run out of space. This happens because all the chunks were allocated for
data since the metadata requirements were so low. But now there's a bunch of
empty data block groups and not enough metadata space to do anything. This
patch solves this problem by automatically deleting empty block groups. If we
notice the used count go down to 0 when deleting or on mount notice that a block
group has a used count of 0 then we will queue it to be deleted.
When the cleaner thread runs we will double check to make sure the block group
is still empty and then we will delete it. This patch has the side effect of no
longer having a bunch of BUG_ON()'s in the chunk delete code, which will be
helpful for both this and relocate. Thanks,
Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Chris Mason <clm@fb.com>
There were several problems about chunk mutex usage:
- Lock chunk mutex when updating metadata. It would cause the nested
deadlock because updating metadata might need allocate new chunks
that need acquire chunk mutex. We remove chunk mutex at this case,
because b-tree lock and other lock mechanism can help us.
- ABBA deadlock occured between device_list_mutex and chunk_mutex.
When we update device status, we must acquire device_list_mutex at the
beginning, and then we might get chunk_mutex during the device status
update because we need allocate new chunks for metadata COW. But at
most place, we acquire chunk_mutex at first and then acquire device list
mutex. We need change the lock order.
- Some place we needn't acquire chunk_mutex. For example we needn't get
chunk_mutex when we free a empty seed fs_devices structure.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Chris Mason <clm@fb.com>
One of my tests shows that when we really don't have space to reclaim via
flush_space and also run out of space, this async reclaim work loops on adding
itself into the workqueue and keeps writing something to disk according to
iostat's results, and these writes mainly comes from commit_transaction which
writes super_block. This's unacceptable as it can be bad to disks, especially
memeory storages.
This adds a check to avoid the above situation.
Signed-off-by: Liu Bo <bo.li.liu@oracle.com>
Signed-off-by: Chris Mason <clm@fb.com>
The nodesize and leafsize were never of different values. Unify the
usage and make nodesize the one. Cleanup the redundant checks and
helpers.
Shaves a few bytes from .text:
text data bss dec hex filename
852418 24560 23112 900090 dbbfa btrfs.ko.before
851074 24584 23112 898770 db6d2 btrfs.ko.after
Signed-off-by: David Sterba <dsterba@suse.cz>
Signed-off-by: Chris Mason <clm@fb.com>
btrfs_set_key_type and btrfs_key_type are used inconsistently along with
open coded variants. Other members of btrfs_key are accessed directly
without any helpers anyway.
Signed-off-by: David Sterba <dsterba@suse.cz>
Signed-off-by: Chris Mason <clm@fb.com>
Percpu allocator now supports allocation mask. Add @gfp to
percpu_counter_init() so that !GFP_KERNEL allocation masks can be used
with percpu_counters too.
We could have left percpu_counter_init() alone and added
percpu_counter_init_gfp(); however, the number of users isn't that
high and introducing _gfp variants to all percpu data structures would
be quite ugly, so let's just do the conversion. This is the one with
the most users. Other percpu data structures are a lot easier to
convert.
This patch doesn't make any functional difference.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Jan Kara <jack@suse.cz>
Acked-by: "David S. Miller" <davem@davemloft.net>
Cc: x86@kernel.org
Cc: Jens Axboe <axboe@kernel.dk>
Cc: "Theodore Ts'o" <tytso@mit.edu>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Andrew Morton <akpm@linux-foundation.org>
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 <bo.li.liu@oracle.com>
Signed-off-by: Chris Mason <clm@fb.com>
The original code allocated new chunks by the number of the writable devices
and missing devices to make sure that any RAID levels on a degraded FS continue
to be honored, but it introduced a problem that it stopped us to allocating
new chunks, the steps to reproduce is following:
# mkfs.btrfs -m raid1 -d raid1 -f <dev0> <dev1>
# mkfs.btrfs -f <dev1> //Removing <dev1> from the original fs
# mount -o degraded <dev0> <mnt>
# dd if=/dev/null of=<mnt>/tmpfile bs=1M
It is because we allocate new chunks only on the writable devices, if we take
the number of missing devices into account, and want to allocate new chunks
with higher RAID level, we will fail becaue we don't have enough writable
device. Fix it by ignoring the number of missing devices when allocating
new chunks.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Chris Mason <clm@fb.com>
During its tree walk, btrfs_drop_snapshot() will skip any shared
subtrees it encounters. This is incorrect when we have qgroups
turned on as those subtrees need to have their contents
accounted. In particular, the case we're concerned with is when
removing our snapshot root leaves the subtree with only one root
reference.
In those cases we need to find the last remaining root and add
each extent in the subtree to the corresponding qgroup exclusive
counts.
This patch implements the shared subtree walk and a new qgroup
operation, BTRFS_QGROUP_OPER_SUB_SUBTREE. When an operation of
this type is encountered during qgroup accounting, we search for
any root references to that extent and in the case that we find
only one reference left, we go ahead and do the math on it's
exclusive counts.
Signed-off-by: Mark Fasheh <mfasheh@suse.de>
Reviewed-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Chris Mason <clm@fb.com>
Before I extended the no_quota arg to btrfs_dec/inc_ref because I didn't
understand how snapshot delete was using it and assumed that we needed the
quota operations there. With Mark's work this has turned out to be not the
case, we _always_ need to use no_quota for btrfs_dec/inc_ref, so just drop the
argument and make __btrfs_mod_ref call it's process function with no_quota set
always. Thanks,
Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Chris Mason <clm@fb.com>
This percpu counter @total_bytes_pinned is introduced to skip unnecessary
operations of 'commit transaction', it accounts for those space we may free
but are stuck in delayed refs.
And we zero out @space_info->total_bytes_pinned every transaction period so
we have a better idea of how much space we'll actually free up by committing
this transaction. However, we do the 'zero out' part a little earlier, before
we actually unpin space, so we end up returning ENOSPC when we actually have
free space that's just unpinned from committing transaction.
xfstests/generic/074 complained then.
This fixes it by actually accounting the percpu pinned number when 'unpin',
and since it's protected by space_info->lock, the race is gone now.
Signed-off-by: Liu Bo <bo.li.liu@oracle.com>
Reviewed-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Chris Mason <clm@fb.com>
When we mounted the filesystem after the crash, we got the following
message:
BTRFS error (device xxx): block group xxxx has wrong amount of free space
BTRFS error (device xxx): failed to load free space cache for block group xxx
It is because we didn't update the metadata of the allocated space (in extent
tree) until the file data was written into the disk. During this time, there was
no information about the allocated spaces in either the extent tree nor the
free space cache. when we wrote out the free space cache at this time (commit
transaction), those spaces were lost. In fact, only the free space that is
used to store the file data had this problem, the others didn't because
the metadata of them is updated in the same transaction context.
There are many methods which can fix the above problem
- track the allocated space, and write it out when we write out the free
space cache
- account the size of the allocated space that is used to store the file
data, if the size is not zero, don't write out the free space cache.
The first one is complex and may make the performance drop down.
This patch chose the second method, we use a per-block-group variant to
account the size of that allocated space. Besides that, we also introduce
a per-block-group read-write semaphore to avoid the race between
the allocation and the free space cache write out.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Chris Mason <clm@fb.com>
We are currently allocating space_info objects in an array when we
allocate space_info. When a user does something like:
# btrfs balance start -mconvert=raid1 -dconvert=raid1 /mnt
# btrfs balance start -mconvert=single -dconvert=single /mnt -f
# btrfs balance start -mconvert=raid1 -dconvert=raid1 /
We can end up with memory corruption since the kobject hasn't
been reinitialized properly and the name pointer was left set.
The rationale behind allocating them statically was to avoid
creating a separate kobject container that just contained the
raid type. It used the index in the array to determine the index.
Ultimately, though, this wastes more memory than it saves in all
but the most complex scenarios and introduces kobject lifetime
questions.
This patch allocates the kobjects dynamically instead. Note that
we also remove the kobject_get/put of the parent kobject since
kobject_add and kobject_del do that internally.
Signed-off-by: Jeff Mahoney <jeffm@suse.com>
Reported-by: David Sterba <dsterba@suse.cz>
Signed-off-by: Chris Mason <clm@fb.com>
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 <clm@fb.com>
I've noticed an extra line after "use no compression", but search
revealed much more in messages of more critical levels and rare errors.
Signed-off-by: David Sterba <dsterba@suse.cz>
Signed-off-by: Chris Mason <clm@fb.com>
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 <jbacik@fb.com>
Signed-off-by: Chris Mason <clm@fb.com>
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 <jbacik@fb.com>
Signed-off-by: Chris Mason <clm@fb.com>
Before applying this patch, the task had to reclaim the metadata space
by itself if the metadata space was not enough. And When the task started
the space reclamation, all the other tasks which wanted to reserve the
metadata space were blocked. At some cases, they would be blocked for
a long time, it made the performance fluctuate wildly.
So we introduce the background metadata space reclamation, when the space
is about to be exhausted, we insert a reclaim work into the workqueue, the
worker of the workqueue helps us to reclaim the reserved space at the
background. By this way, the tasks needn't reclaim the space by themselves at
most cases, and even if the tasks have to reclaim the space or are blocked
for the space reclamation, they will get enough space more quickly.
Here is my test result(Tested by compilebench):
Memory: 2GB
CPU: 2Cores * 1CPU
Partition: 40GB(SSD)
Test command:
# compilebench -D <mnt> -m
Without this patch:
intial create total runs 30 avg 54.36 MB/s (user 0.52s sys 2.44s)
compile total runs 30 avg 123.72 MB/s (user 0.13s sys 1.17s)
read compiled tree total runs 3 avg 81.15 MB/s (user 0.74s sys 4.89s)
delete compiled tree total runs 30 avg 5.32 seconds (user 0.35s sys 4.37s)
With this patch:
intial create total runs 30 avg 59.80 MB/s (user 0.52s sys 2.53s)
compile total runs 30 avg 151.44 MB/s (user 0.13s sys 1.11s)
read compiled tree total runs 3 avg 83.25 MB/s (user 0.76s sys 4.91s)
delete compiled tree total runs 30 avg 5.29 seconds (user 0.34s sys 4.34s)
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Chris Mason <clm@fb.com>
If we had to retry on the profiles seqlock (due to a concurrent write), we
would set bits on the input flags that corresponded both to the current
profile and to previous values of the profile.
Signed-off-by: Filipe David Borba Manana <fdmanana@gmail.com>
Signed-off-by: Chris Mason <clm@fb.com>
If skinny metadata is enabled and our first tree search fails to find a
skinny extent item, we may repeat a tree search for a "fat" extent item
(if the previous item in the leaf is not the "fat" extent we're looking
for). However we were not setting the new key's objectid to the right
value, as we previously used the same key variable to peek at the previous
item in the leaf, which has a different objectid. So just set the right
objectid to avoid modifying/deleting a wrong item if we repeat the tree
search.
Signed-off-by: Filipe David Borba Manana <fdmanana@gmail.com>
Signed-off-by: Chris Mason <clm@fb.com>
I'm not sure why we weren't aborting here in the first place, it is obviously a
bad time from the fact that we print the leaf and yell loudly about it. Fix
this up, otherwise we panic because our path could be pointing into oblivion.
Thanks,
Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Chris Mason <clm@fb.com>
When encountering memory pressure, testers have run into the following
lockdep warning. It was caused by __link_block_group calling kobject_add
with the groups_sem held. kobject_add calls kvasprintf with GFP_KERNEL,
which gets us into reclaim context. The kobject doesn't actually need
to be added under the lock -- it just needs to ensure that it's only
added for the first block group to be linked.
=========================================================
[ INFO: possible irq lock inversion dependency detected ]
3.14.0-rc8-default #1 Not tainted
---------------------------------------------------------
kswapd0/169 just changed the state of lock:
(&delayed_node->mutex){+.+.-.}, at: [<ffffffffa018baea>] __btrfs_release_delayed_node+0x3a/0x200 [btrfs]
but this lock took another, RECLAIM_FS-unsafe lock in the past:
(&found->groups_sem){+++++.}
and interrupts could create inverse lock ordering between them.
other info that might help us debug this:
Possible interrupt unsafe locking scenario:
CPU0 CPU1
---- ----
lock(&found->groups_sem);
local_irq_disable();
lock(&delayed_node->mutex);
lock(&found->groups_sem);
<Interrupt>
lock(&delayed_node->mutex);
*** DEADLOCK ***
2 locks held by kswapd0/169:
#0: (shrinker_rwsem){++++..}, at: [<ffffffff81159e8a>] shrink_slab+0x3a/0x160
#1: (&type->s_umount_key#27){++++..}, at: [<ffffffff811bac6f>] grab_super_passive+0x3f/0x90
Signed-off-by: Jeff Mahoney <jeffm@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
Lets try this again. We can deadlock the box if we send on a box and try to
write onto the same fs with the app that is trying to listen to the send pipe.
This is because the writer could get stuck waiting for a transaction commit
which is being blocked by the send. So fix this by making sure looking at the
commit roots is always going to be consistent. We do this by keeping track of
which roots need to have their commit roots swapped during commit, and then
taking the commit_root_sem and swapping them all at once. Then make sure we
take a read lock on the commit_root_sem in cases where we search the commit root
to make sure we're always looking at a consistent view of the commit roots.
Previously we had problems with this because we would swap a fs tree commit root
and then swap the extent tree commit root independently which would cause the
backref walking code to screw up sometimes. With this patch we no longer
deadlock and pass all the weird send/receive corner cases. Thanks,
Reportedy-by: Hugo Mills <hugo@carfax.org.uk>
Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Chris Mason <clm@fb.com>
We could have possibly added an extent_op to the locked_ref while we dropped
locked_ref->lock, so check for this case as well and loop around. Otherwise we
could lose flag updates which would lead to extent tree corruption. Thanks,
cc: stable@vger.kernel.org
Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Chris Mason <clm@fb.com>
We needn't flush all delalloc inodes when we doesn't get s_umount lock,
or we would make the tasks wait for a long time.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
generic/074 in xfstests failed sometimes because of the enospc error,
the reason of this problem is that we just reclaimed the space we need
from the reserved space for delalloc, and then tried to reserve the space,
but if some task did no-flush reservation between the above reclamation
and reservation,
Task1 Task2
shrink_delalloc()
reclaim 1 block
(The space that can
be reserved now is 1
block)
do no-flush reservation
reserve 1 block
(The space that can
be reserved now is 0
block)
reserving 1 block failed
the reservation of Task1 failed, but in fact, there was enough space to
reserve if we could reclaim more space before.
Fix this problem by the aggressive reclamation of the reserved delalloc
metadata space.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
The reason is:
- The per-cpu counter has its own lock to protect itself.
- Here we needn't get a exact value.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
If the snapshot creation happened after the nocow write but before the dirty
data flush, we would fail to flush the dirty data because of no space.
So we must keep track of when those nocow write operations start and when they
end, if there are nocow writers, the snapshot creators must wait. In order
to implement this function, I introduce btrfs_{start, end}_nocow_write(),
which is similar to mnt_{want,drop}_write().
These two functions are only used for nocow file write operations.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
Since the "_struct" suffix is mainly used for distinguish the differnt
btrfs_work between the original and the newly created one,
there is no need using the suffix since all btrfs_workers are changed
into btrfs_workqueue.
Also this patch fixed some codes whose code style is changed due to the
too long "_struct" suffix.
Signed-off-by: Qu Wenruo <quwenruo@cn.fujitsu.com>
Tested-by: David Sterba <dsterba@suse.cz>
Signed-off-by: Josef Bacik <jbacik@fb.com>
Replace the fs_info->cache_workers with the newly created
btrfs_workqueue.
Signed-off-by: Qu Wenruo <quwenruo@cn.fujitsu.com>
Tested-by: David Sterba <dsterba@suse.cz>
Signed-off-by: Josef Bacik <jbacik@fb.com>
A user reported a 100% cpu hang with my new delayed ref code. Turns out I
forgot to increase the count check when we can't run a delayed ref because of
the tree mod log. If we can't run any delayed refs during this there is no
point in continuing to look, and we need to break out. Thanks,
Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Chris Mason <clm@fb.com>
We allocate the free space from the former block group, not the current
one, so should use the former one to output the trace information.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Chris Mason <clm@fb.com>
used_block_group is just used for the space cluster which doesn't
belong to the current block group, the other place needn't use it.
Or the logic of code seems unclear.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Chris Mason <clm@fb.com>
After the change titled "Btrfs: add support for inode properties", if
btrfs was built-in the kernel (i.e. not as a module), it would cause a
kernel panic, as reported recently by Fengguang:
[ 2.024722] BUG: unable to handle kernel NULL pointer dereference at (null)
[ 2.027814] IP: [<ffffffff81501594>] crc32c+0xc/0x6b
[ 2.028684] PGD 0
[ 2.028684] Oops: 0000 [#1] SMP
[ 2.028684] Modules linked in:
[ 2.028684] CPU: 0 PID: 1 Comm: swapper/0 Not tainted 3.13.0-rc7-04795-ga7b57c2 #1
[ 2.028684] Hardware name: Bochs Bochs, BIOS Bochs 01/01/2011
[ 2.028684] task: ffff88000edba100 ti: ffff88000edd6000 task.ti: ffff88000edd6000
[ 2.028684] RIP: 0010:[<ffffffff81501594>] [<ffffffff81501594>] crc32c+0xc/0x6b
[ 2.028684] RSP: 0000:ffff88000edd7e58 EFLAGS: 00010246
[ 2.028684] RAX: 0000000000000000 RBX: ffffffff82295550 RCX: 0000000000000000
[ 2.028684] RDX: 0000000000000011 RSI: ffffffff81efe393 RDI: 00000000fffffffe
[ 2.028684] RBP: ffff88000edd7e60 R08: 0000000000000003 R09: 0000000000015d20
[ 2.028684] R10: ffffffff81ef225e R11: ffffffff811b0222 R12: ffffffffffffffff
[ 2.028684] R13: 0000000000000239 R14: 0000000000000000 R15: 0000000000000000
[ 2.028684] FS: 0000000000000000(0000) GS:ffff88000fa00000(0000) knlGS:0000000000000000
[ 2.028684] CS: 0010 DS: 0000 ES: 0000 CR0: 000000008005003b
[ 2.028684] CR2: 0000000000000000 CR3: 000000000220c000 CR4: 00000000000006f0
[ 2.028684] Stack:
[ 2.028684] ffffffff82295550 ffff88000edd7e80 ffffffff8238af62 ffffffff8238ac05
[ 2.028684] 0000000000000000 ffff88000edd7e98 ffffffff8238ac0f ffffffff8238ac05
[ 2.028684] ffff88000edd7f08 ffffffff810002ba ffff88000edd7f00 ffffffff810e2404
[ 2.028684] Call Trace:
[ 2.028684] [<ffffffff8238af62>] btrfs_props_init+0x4f/0x96
[ 2.028684] [<ffffffff8238ac05>] ? ftrace_define_fields_btrfs_space_reservation+0x145/0x145
[ 2.028684] [<ffffffff8238ac0f>] init_btrfs_fs+0xa/0xf0
[ 2.028684] [<ffffffff8238ac05>] ? ftrace_define_fields_btrfs_space_reservation+0x145/0x145
[ 2.028684] [<ffffffff810002ba>] do_one_initcall+0xa4/0x13a
[ 2.028684] [<ffffffff810e2404>] ? parse_args+0x25f/0x33d
[ 2.028684] [<ffffffff8234cf75>] kernel_init_freeable+0x1aa/0x230
[ 2.028684] [<ffffffff8234c785>] ? do_early_param+0x88/0x88
[ 2.028684] [<ffffffff819f61b5>] ? rest_init+0x89/0x89
[ 2.028684] [<ffffffff819f61c3>] kernel_init+0xe/0x109
The issue here is that the initialization function of btrfs (super.c:init_btrfs_fs)
started using crc32c (from lib/libcrc32c.c). But when it needs to call crc32c (as
part of the properties initialization routine), the libcrc32c is not yet initialized,
so crc32c derreferenced a NULL pointer (lib/libcrc32c.c:tfm), causing the kernel
panic on boot.
The approach to fix this is to use crypto component directly to use its crc32c (which
is basically what lib/libcrc32c.c is, a wrapper around crypto). This is what ext4 is
doing as well, it uses crypto directly to get crc32c functionality.
Verified this works both when btrfs is built-in and when it's loadable kernel module.
Signed-off-by: Filipe David Borba Manana <fdmanana@gmail.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Chris Mason <clm@fb.com>
On one of our gluster clusters we noticed some pretty big lag spikes. This
turned out to be because our transaction commit was taking like 3 minutes to
complete. This is because we have like 30 gigs of metadata, so our global
reserve would end up being the max which is like 512 mb. So our throttling code
would allow a ridiculous amount of delayed refs to build up and then they'd all
get run at transaction commit time, and for a cold mounted file system that
could take up to 3 minutes to run. So fix the throttling to be based on both
the size of the global reserve and how long it takes us to run delayed refs.
This patch tracks the time it takes to run delayed refs and then only allows 1
seconds worth of outstanding delayed refs at a time. This way it will auto-tune
itself from cold cache up to when everything is in memory and it no longer has
to go to disk. This makes our transaction commits take much less time to run.
Thanks,
Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Chris Mason <clm@fb.com>
Currently we have two rb-trees, one for delayed ref heads and one for all of the
delayed refs, including the delayed ref heads. When we process the delayed refs
we have to hold onto the delayed ref lock for all of the selecting and merging
and such, which results in quite a bit of lock contention. This was solved by
having a waitqueue and only one flusher at a time, however this hurts if we get
a lot of delayed refs queued up.
So instead just have an rb tree for the delayed ref heads, and then attach the
delayed ref updates to an rb tree that is per delayed ref head. Then we only
need to take the delayed ref lock when adding new delayed refs and when
selecting a delayed ref head to process, all the rest of the time we deal with a
per delayed ref head lock which will be much less contentious.
The locking rules for this get a little more complicated since we have to lock
up to 3 things to properly process delayed refs, but I will address that problem
later. For now this passes all of xfstests and my overnight stress tests.
Thanks,
Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Chris Mason <clm@fb.com>
We may return early in btrfs_drop_snapshot(), we shouldn't
call btrfs_std_err() for this case, fix it.
Cc: stable@vger.kernel.org
Signed-off-by: Wang Shilong <wangsl.fnst@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Chris Mason <clm@fb.com>
@full is not protected within global_rsv.lock, so we may think global_rsv
is already full but in fact it's not, so we miss the opportunity to return
free space to global_rsv directly when we release other block_rsvs.
Signed-off-by: Liu Bo <bo.li.liu@oracle.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Chris Mason <clm@fb.com>
We can starve out the transaction commit with a bunch of caching threads all
running at the same time. This is because we will only drop the
extent_commit_sem if we need_resched(), which isn't likely to happen since we
will be reading a lot from the disk so have already schedule()'ed plenty. Alex
observed that he could starve out a transaction commit for up to a minute with
32 caching threads all running at once. This will allow us to drop the
extent_commit_sem to allow the transaction commit to swap the commit_root out
and then all the cachers will start back up. Here is an explanation provided by
Igno
So, just to fill in what happens in this loop:
mutex_unlock(&caching_ctl->mutex);
cond_resched();
goto again;
where 'again:' takes caching_ctl->mutex and fs_info->extent_commit_sem
again:
again:
mutex_lock(&caching_ctl->mutex);
/* need to make sure the commit_root doesn't disappear */
down_read(&fs_info->extent_commit_sem);
So, if I'm reading the code correct, there can be a fair amount of
concurrency here: there may be multiple 'caching kthreads' per filesystem
active, while there's one fs_info->extent_commit_sem per filesystem
AFAICS.
So, what happens if there are a lot of CPUs all busy holding the
->extent_commit_sem rwsem read-locked and a writer arrives? They'd all
rush to try to release the fs_info->extent_commit_sem, and they'd block in
the down_read() because there's a writer waiting.
So there's a guarantee of forward progress. This should answer akpm's
concern I think.
Thanks,
Acked-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Signed-off-by: Chris Mason <clm@fb.com>
Convert all applicable cases of printk and pr_* to the btrfs_* macros.
Fix all uses of the BTRFS prefix.
Signed-off-by: Frank Holton <fholton@gmail.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Chris Mason <clm@fb.com>
We met the following oops when doing space balance:
kobject (ffff88081b590278): tried to init an initialized object, something is seriously wrong.
...
Call Trace:
[<ffffffff81937262>] dump_stack+0x49/0x5f
[<ffffffff8137d259>] kobject_init+0x89/0xa0
[<ffffffff8137d36a>] kobject_init_and_add+0x2a/0x70
[<ffffffffa009bd79>] ? clear_extent_bit+0x199/0x470 [btrfs]
[<ffffffffa005e82c>] __link_block_group+0xfc/0x120 [btrfs]
[<ffffffffa006b9db>] btrfs_make_block_group+0x24b/0x370 [btrfs]
[<ffffffffa00a899b>] __btrfs_alloc_chunk+0x54b/0x7e0 [btrfs]
[<ffffffffa00a8c6f>] btrfs_alloc_chunk+0x3f/0x50 [btrfs]
[<ffffffffa0060123>] do_chunk_alloc+0x363/0x440 [btrfs]
[<ffffffffa00633d4>] btrfs_check_data_free_space+0x104/0x310 [btrfs]
[<ffffffffa0069f4d>] btrfs_write_dirty_block_groups+0x48d/0x600 [btrfs]
[<ffffffffa007aad4>] commit_cowonly_roots+0x184/0x250 [btrfs]
...
Steps to reproduce:
# mkfs.btrfs -f <dev>
# mount -o nospace_cache <dev> <mnt>
# btrfs balance start <mnt>
# dd if=/dev/zero of=<mnt>/tmpfile bs=1M count=1
The reason of this problem is that we initialized the raid kobject when we added
a block group into a empty raid list. As we know, when we mounted a btrfs filesystem,
the raid list was empty, we would initialize the raid kobject when we added the first
block group. But if there was not data stored in the block group, the block group
would be freed when doing balance, and the raid list would be empty. And then if we
allocated a new block group and added it into the raid list, we would initialize
the raid kobject again, the oops happened.
Fix this problem by initializing the raid kobject just when mounting the fs.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Reported-by: Wang Shilong <wangsl.fnst@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Chris Mason <clm@fb.com>
This patch fixes the following warnings:
fs/btrfs/extent-tree.c:6201:12: sparse: symbol 'get_raid_name' was not declared. Should it be static?
fs/btrfs/extent-tree.c:8430:9: error: format not a string literal and no format arguments [-Werror=format-security] get_raid_name(index));
Signed-off-by: Jeff Mahoney <jeffm@suse.com>
Reviewed-by: Kees Cook <keescook@chromium.org>
Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Chris Mason <clm@fb.com>
The variable found_uncached_bg in find_free_extent is not used since commit
285ff5af6c
(Btrfs: remove the ideal caching code)
Signed-off-by: Valentina Giusti <valentina.giusti@microon.de>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Signed-off-by: Chris Mason <clm@fb.com>
While trying to debug ENOSPC issues, it's helpful to understand what the
kernel's view of the available space is. We export this information
via ioctl, but sysfs files are more easily used.
Signed-off-by: Jeff Mahoney <jeffm@suse.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Signed-off-by: Chris Mason <clm@fb.com>
The way how we process delayed refs is
1) get a bunch of head refs,
2) pick up one head ref,
3) go one node back for any delayed ref updates.
The head ref is also linked in the same rbtree as the delayed ref is,
so in 1) stage, we have to walk one by one including not only head refs, but
delayed refs.
When we have a great number of delayed refs pending to process,
this'll cost time a lot.
Here we introduce a head ref specific rbtree, it only has head refs, so troubles
go away.
Signed-off-by: Liu Bo <bo.li.liu@oracle.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Signed-off-by: Chris Mason <clm@fb.com>
Currently extent-tree.c:btrfs_lookup_extent_info() can miss the lookup
of skinny extent items. This can happen when the execution flow is the
following:
* We do an extent tree lookup and fail to find a skinny extent item;
* As a result, we attempt to see if a non-skinny extent item exists,
either by looking at previous item in the leaf or by doing another
full extent tree search;
* We have a transaction and then we check for a matching delayed ref
head in the transaction's delayed refs rbtree;
* We find such delayed ref head and then we try to lock it with a
call to mutex_trylock();
* The lock was contended so we jump to the label "again", which repeats
the extent tree search but for a non-skinny extent item, because we set
previously metadata variable to 0 and the search key to look for a
non-skinny extent-item;
* After the jump (and after releasing the transaction's delayed refs
lock), a skinny extent item might have been added to the extent tree
but we will miss it because metadata is set to 0 and the search key
is set for a non-skinny extent-item.
The fix here is to not reset metadata to 0 and to jump to the initial search
key setup if the delayed ref head is contended, instead of jumping directly
to the extent tree search label ("again").
This issue was found while investigating the issue reported at Bugzilla 64961.
David Sterba suspected this function was missing extent items, and that
this could be caused by the last change to this function, which was made
in the following patch:
[PATCH] Btrfs: optimize btrfs_lookup_extent_info()
(commit 74be951087)
But in fact this issue already existed before, because after failing to find
a skinny extent item, the code set the search key for a non-skinny extent
item, and on contention of a matching delayed ref head it would not search
the extent tree for a skinny extent item anymore.
Signed-off-by: Filipe David Borba Manana <fdmanana@gmail.com>
Reviewed-by: Liu Bo <bo.li.liu@oracle.com>
Signed-off-by: Chris Mason <clm@fb.com>
rename the function -- btrfs_start_all_delalloc_inodes(), and make its
name be compatible to btrfs_wait_ordered_roots(), since they are always
used at the same place.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Signed-off-by: Chris Mason <chris.mason@fusionio.com>
It is very likely that there are lots of ordered extents in the filesytem,
if we wait for the completion of all of them when we want to reclaim some
space for the metadata space reservation, we would be blocked for a long
time. The performance would drop down suddenly for a long time.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Signed-off-by: Chris Mason <chris.mason@fusionio.com>
It was very likely that there were lots of async delalloc pages in the
filesystem, if we waited until all the pages were flushed, we would be
blocked for a long time, and the performance would also drop down.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Signed-off-by: Chris Mason <chris.mason@fusionio.com>
In shrink_delalloc(), what we need reclaim is the metadata space, so
flushing pages by to_reclaim is not reasonable, it is very likely that
the pages we flush are not enough. And then we had to invoke the flush
function for several times, at the worst, we need call flush_space for
several times. It wasted time.
We improve this problem by converting the metadata space size we need
reserve to the delalloc bytes, By this way, we can flush the pages
by a reasonable number.
(Now we use a fixed number to do conversion, it is not flexible, maybe
we can find a good way to improve it in the future.)
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Signed-off-by: Chris Mason <chris.mason@fusionio.com>
This patch picked up the code that was used to calculate the number of
the items for which we need reserve space, and we will use it in the next
patch.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Signed-off-by: Chris Mason <chris.mason@fusionio.com>
Fix spacing issues detected via checkpatch.pl in accordance with the
kernel style guidelines.
Signed-off-by: Dulshani Gunawardhana <dulshani.gunawardhana89@gmail.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Signed-off-by: Chris Mason <chris.mason@fusionio.com>
Use WARN_ON()'s return value in place of WARN_ON(1) for cleaner source
code that outputs a more descriptive warnings. Also fix the styling
warning of redundant braces that came up as a result of this fix.
Signed-off-by: Dulshani Gunawardhana <dulshani.gunawardhana89@gmail.com>
Reviewed-by: Zach Brown <zab@redhat.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Signed-off-by: Chris Mason <chris.mason@fusionio.com>
After running space balance on a new fs, the fs check program outputed the
following warning message:
free space inode generation (0) did not match free space cache generation (20)
Steps to reproduce:
# mkfs.btrfs -f <dev>
# mount <dev> <mnt>
# btrfs balance start <mnt>
# umount <mnt>
# btrfs check <dev>
It was because there was no data space after the space balance, and the free
space write out task didn't try to allocate a new data chunk for the free space
inode when doing the reservation. So the data space reservation failed, and in
order to tell the free space loader that this free space inode could not be
trusted, the generation of the free space inode wasn't updated. Then the check
program found this problem and outputed the above message.
But in fact, it is safe that we try to allocate a new data chunk when we find
the data space is not enough. The patch fixes the above problem by this way.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Signed-off-by: Chris Mason <chris.mason@fusionio.com>
Instead of doing another extent tree search if the first search failed
to find a metadata item, check if the previous item in the leaf is an
extent item and use it if it is, otherwise do the second tree search
for an extent item.
Signed-off-by: Filipe David Borba Manana <fdmanana@gmail.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Signed-off-by: Chris Mason <chris.mason@fusionio.com>
When debugging ENOSPC issues, it's nice to be able to see which
reservations failed as well as the ones which succeeded.
Signed-off-by: Jeff Mahoney <jeffm@suse.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Signed-off-by: Chris Mason <chris.mason@fusionio.com>
fs/btrfs/compat.h only contained trivial macro wrappers of drop_nlink()
and inc_nlink(). This doesn't belong in mainline.
Signed-off-by: Zach Brown <zab@redhat.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Signed-off-by: Chris Mason <chris.mason@fusionio.com>
When we fail to add a reference after a non-inline insertion by some reasons,
eg. ENOSPC, we'll abort the transaction, but we don't return this error to
the caller who has to walk around again to find something wrong, that's
unnecessary.
Also fixup other error paths to keep it simple.
Signed-off-by: Liu Bo <bo.li.liu@oracle.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Signed-off-by: Chris Mason <chris.mason@fusionio.com>
While trying to track down a reserved space leak I noticed a few places where we
won't properly clean up reserved space if we have an error, this patch fixes
those up. Thanks,
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Signed-off-by: Chris Mason <chris.mason@fusionio.com>
In trying to track down where we were leaking reserved space I noticed our
reserve extent tracepoints are a little off. First we were saying that the
reserved space had been alloced in btrfs_reserve_extent, which isn't the case,
this needs to be triggered when we actually allocate the space when we run the
delayed ref. We were also missing a few places where we should have been
tracing the btrfs_reserve_extent_free tracepoint. With these in place I was
able to put together where we were leaking reserved space. Thanks,
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Signed-off-by: Chris Mason <chris.mason@fusionio.com>
In extent-tree.c:btrfs_write_dirty_block_groups(), if the call to
write_one_cache_group() failed, we would return without putting
the block group first.
Signed-off-by: Filipe David Borba Manana <fdmanana@gmail.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Signed-off-by: Chris Mason <chris.mason@fusionio.com>
Not used for anything, and removing it avoids caller's need to
allocate a path structure.
Signed-off-by: Filipe David Borba Manana <fdmanana@gmail.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Signed-off-by: Chris Mason <chris.mason@fusionio.com>
This isn't used for anything anymore, just remove it.
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Signed-off-by: Chris Mason <chris.mason@fusionio.com>
This is a left over of how we used to wait for ordered extents, which was to
grab the inode and then run filemap flush on it. However if we have an ordered
extent then we already are holding a ref on the inode, and we just use
btrfs_start_ordered_extent anyway, so there is no reason to have an extra ref on
the inode to start work on the ordered extent. Thanks,
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Signed-off-by: Chris Mason <chris.mason@fusionio.com>
This reverts commit 70afa3998c. It is causing
performance issues and wasn't actually correct. There were problems with the
way we flushed delalloc and that was the real cause of the early enospc.
Thanks,
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Signed-off-by: Chris Mason <chris.mason@fusionio.com>
By the current code, if the requested size is very large, and all the extents
in the free space cache are small, we will waste lots of the cpu time to cut
the requested size in half and search the cache again and again until it gets
down to the size the allocator can return. In fact, we can know the max extent
size in the cache after the first search, so we needn't cut the size in half
repeatedly, and just use the max extent size directly. This way can save
lots of cpu time and make the performance grow up when there are only fragments
in the free space cache.
According to my test, if there are only 4KB free space extents in the fs,
and the total size of those extents are 256MB, we can reduce the execute
time of the following test from 5.4s to 1.4s.
dd if=/dev/zero of=<testfile> bs=1MB count=1 oflag=sync
Changelog v2 -> v3:
- fix the problem that we skip the block group with the space which is
less than we need.
Changelog v1 -> v2:
- address the problem that we return a wrong start position when searching
the free space in a bitmap.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Signed-off-by: Chris Mason <chris.mason@fusionio.com>
A user was reporting weird warnings from btrfs_put_delayed_ref() and I noticed
that we were doing this list_del_init() on our head ref outside of
delayed_refs->lock. This is a problem if we have people still on the list, we
could end up modifying old pointers and such. Fix this by removing us from the
list before we do our run_delayed_ref on our head ref. Thanks,
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Signed-off-by: Chris Mason <chris.mason@fusionio.com>
u64 is "unsigned long long" on all architectures now, so there's no need to
cast it when formatting it using the "ll" length modifier.
Signed-off-by: Geert Uytterhoeven <geert@linux-m68k.org>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Signed-off-by: Chris Mason <chris.mason@fusionio.com>
This tree is not created by mkfs.btrfs. Therefore when a filesystem
is mounted writable and the UUID tree does not exist, this tree is
created if required. The tree is also added to the fs_info structure
and initialized, but this commit does not yet read or write UUID tree
elements.
Signed-off-by: Stefan Behrens <sbehrens@giantdisaster.de>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Signed-off-by: Chris Mason <chris.mason@fusionio.com>
I noticed while looking at a deadlock that we are always starting a transaction
in cow_file_range(). This isn't really needed since we only need a transaction
if we are doing an inline extent, or if the allocator needs to allocate a chunk.
So push down all the transaction start stuff to be closer to where we actually
need a transaction in all of these cases. This will hopefully reduce our write
latency when we are committing often. Thanks,
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Signed-off-by: Chris Mason <chris.mason@fusionio.com>
In extent-tree.c:do_chunk_alloc(), early on we returned 0 (success)
when the target space was full and when chunk allocation is needed.
However, later on in that same function we return ENOSPC if
btrfs_alloc_chunk() fails (and chunk allocation was needed) and
set the space's full flag.
This was inconsistent, as -ENOSPC should be returned if the space
is full and a chunk allocation needs to performed. If the space is
full but no chunk allocation is needed, just return 0 (success).
Signed-off-by: Filipe David Borba Manana <fdmanana@gmail.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Signed-off-by: Chris Mason <chris.mason@fusionio.com>
Alex Lyakas reported a bug where wait_block_group_cache_progress() would wait
forever if a drive failed. This is because we just bail out if there is an
error while trying to cache a block group, we don't update anybody who may be
waiting. So this introduces a new enum for the cache state in case of error and
makes everybody bail out if we have an error. Alex tested and verified this
patch fixed his problem. This fixes bz 59431. Thanks,
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Signed-off-by: Chris Mason <chris.mason@fusionio.com>
I was hitting the BUG_ON() at the end of merge_reloc_roots() because we were
aborting the transaction at some point previously and then getting an error when
we tried to drop the reloc root. I fixed btrfs_drop_snapshot to re-add us to
the dead roots list if we failed, but this isn't the right thing to do for reloc
roots since it uses root->root_list for it's own stuff in order to know what
needs to be cleaned up. So fix btrfs_drop_snapshot to only do the re-add if we
aren't dropping for reloc, and handle errors from merge_reloc_root() by dropping
the reloc root we are processing since it won't be on the list of roots to
cleanup. With this patch my reproducer no longer panics. Thanks,
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Signed-off-by: Chris Mason <chris.mason@fusionio.com>
This shows exactly how btrfs processes the delayed refs onto disks,
which is very helpful on understanding delayed ref mechanism and
debugging related bugs.
Signed-off-by: Liu Bo <bo.li.liu@oracle.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Signed-off-by: Chris Mason <chris.mason@fusionio.com>
btrfs_space_info->block_groups.
The current code uses integer literals to index
btrfs_space_info->block_groups[] array. Instead use corresponding
enums from 'enum btrfs_raid_types'.
Signed-off-by: chandan <chandan@linux.vnet.ibm.com>
Reviewed-by: David Sterba <dsterba@suse.cz>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Signed-off-by: Chris Mason <chris.mason@fusionio.com>
So to cache free space, we iterate every extent item to gather free space info.
When we have say 10,000 non-inline extent refs(such as BTRFS_EXTENT_DATA_REF),
it takes quite a long time, and since inline extent refs and non-inline ones have
same objectid in their keys, we can just re-search the tree with the next address
to skip non-inline references.
(This is found by dedup feature because dedup extents can end up with many
non-inline extent refs.)
Signed-off-by: Liu Bo <bo.li.liu@oracle.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Signed-off-by: Chris Mason <chris.mason@fusionio.com>
I recently did some ENOSPC testing that involved filling the disk
while create and removing snapshots in a loop. During the test cycle,
I ran into an ENOSPC when trying to remove a snapshot, leaving the fs
stuck in ENOSPC even after a umount/mount cycle.
This patch allow subvolume removal to fall back onto the global
block reservation in order to succeed when it would have failed
otherwise.
Signed-off-by: Jeff Mahoney <jeffm@suse.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Signed-off-by: Chris Mason <chris.mason@fusionio.com>
If we're looking for a metadata item in the tree and the
search fails with return value of 1, and the slot doesn't
point to the first item in the leaf, check if the previous
item in the leaf corresponds to an extent item for the same
object id - if it does, then don't do another tree search
to get it.
This optimization is already done by btrfs-progs.
V2: updated commit message.
Signed-off-by: Filipe David Borba Manana <fdmanana@gmail.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Signed-off-by: Chris Mason <chris.mason@fusionio.com>
We've been seeing spurious complaints out of lockdep because the lock class name
changes. This is happening because when we drop a snapshot we will lock a block
before we've read it in, which sets the lockdep class to whatever the default
is. Then once we read the thing in we reset the lockdep class to what it is
supposed to be, which blows lockdeps' mind. This patch should fix the problem,
it appears to be the only place where we do this sort of thing. Thanks,
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Signed-off-by: Chris Mason <chris.mason@fusionio.com>
If we stop dropping a root for whatever reason we need to add it back to the
dead root list so that we will re-start the dropping next transaction commit.
The other case this happens is if we recover a drop because we will add a root
without adding it to the fs radix tree, so we can leak it's root and commit root
extent buffer, adding this to the dead root list makes this cleanup happen.
Thanks,
Cc: stable@vger.kernel.org
Reported-by: Alex Lyakas <alex.btrfs@zadarastorage.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
We aren't setting path->locks[level] when we resume a snapshot deletion which
means we won't unlock the buffer when we free the path. This causes deadlocks
if we happen to re-allocate the block before we've evicted the extent buffer
from cache. Thanks,
Cc: stable@vger.kernel.org
Reported-by: Alex Lyakas <alex.btrfs@zadarastorage.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Alex pointed out a problem and fix that exists in the drop one snapshot at a
time patch. If we decide we need to exit for whatever reason (umount for
example) we will just exit the snapshot dropping without updating the drop
progress. So the next time we go to resume we will BUG_ON() because we can't
find the extent we left off at because we never updated it. This patch fixes
the problem.
Cc: stable@vger.kernel.org
Reported-by: Alex Lyakas <alex.btrfs@zadarastorage.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
When adjusting the enospc rules for relocation I ran into a deadlock because we
were relocating the only system chunk and that forced us to try and allocate a
new system chunk while holding locks in the chunk tree, which caused us to
deadlock. To fix this I've moved all of the dev extent addition and chunk
addition out to the delayed chunk completion stuff. We still keep the in-memory
stuff which makes sure everything is consistent.
One change I had to make was to search the commit root of the device tree to
find a free dev extent, and hold onto any chunk em's that we allocated in that
transaction so we do not allocate the same dev extent twice. This has the side
effect of fixing a bug with balance that has been there ever since balance
existed. Basically you can free a block group and it's dev extent and then
immediately allocate that dev extent for a new block group and write stuff to
that dev extent, all within the same transaction. So if you happen to crash
during a balance you could come back to a completely broken file system. This
patch should keep these sort of things from happening in the future since we
won't be able to allocate free'd dev extents until after the transaction
commits. This has passed all of the xfstests and my super annoying stress test
followed by a balance. Thanks,
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
We always just try and reserve data space when we write, but if we are out of
space but have prealloc'ed extents we should still successfully write. This
patch will try and see if we can write to prealloc'ed space and if we can go
ahead and allow the write to continue. With this patch we now pass xfstests
generic/274. Thanks,
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
try_to_writeback_inodes_sb_nr returns 1 if writeback is already underway, which
is completely fraking useless for us as we need to make sure pages are actually
written before we go and check if there are ordered extents. So replace this
with an open coding of try_to_writeback_inodes_sb_nr minus the writeback
underway check so that we are sure to actually have flushed some dirty pages out
and will have ordered extents to use. With this patch xfstests generic/273 now
passes. Thanks,
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
There are all of these checks in the ENOSPC code to see if committing the
transaction would free up enough space to make the allocation. This is because
early on we just committed the transaction and hoped and prayed, which resulted
in cases where it took _forever_ to get an ENOSPC when we really were out of
space. So we check space_info->bytes_pinned, except this isn't completely true
because it doesn't account for space we may free but are stuck in delayed refs.
So tests like xfstests 226 would fail because we wouldn't commit the transaction
to free up the data space. So instead add a percpu counter that will be a
little fuzzier, it will add bytes as soon as we try to free up the space, and
remove any space it doesn't actually free up when we get around to doing the
actual free. We then 0 out this counter every transaction period so we have a
better idea of how much space we will actually free up by committing this
transaction. With this patch we now pass xfstests 226. Thanks,
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Dave has this fs_mark script that can make btrfs abort with sufficient amount of
ram. This is because with more ram we can keep more dirty metadata in cache
which in a round about way makes for many more pending delayed refs. What
happens is we end up not throttling the transaction enough so when we go to
commit the transaction when we've completely filled the file system we'll
abort() because we use all of the space in the global reserve and we still have
delayed refs to run. To fix this we need to make the delayed ref flushing and
the transaction throttling dependant upon the number of delayed refs that we
have instead of how much reserved space is left in the global reserve. With
this patch we not only stop aborting transactions but we also get a smoother run
speed with fs_mark and it makes us about 10% faster. Thanks,
Reported-by: David Sterba <dsterba@suse.cz>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
I hit a deadlock because we aborted when flushing delayed refs but didn't wake
any of the other flushers up and so everybody was just sleeping forever. This
should fix the problem. Thanks,
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
With non-mixed block groups we replay the logs before we're allowed to do any
writes, so we get away with not pinning/removing the data extents until right
when we replay them. However with mixed block groups we allocate out of the
same pool, so we could easily allocate a metadata block that was logged in our
tree log. To deal with this we just need to notice that we have mixed block
groups and do the normal excluding/removal dance during the pin stage of the log
replay and that way we don't allocate metadata blocks from areas we have logged
data extents. With this patch we now pass xfstests generic/311 with mixed
block groups turned on. Thanks,
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Dave pointed out a problem where if you filled up a file system as much as
possible you couldn't remove any files. The whole unlink reservation thing is
convoluted because it tries to guess if it's going to add space to unlink
something or not, and has all these odd uncommented cases where it simply does
not try. So to fix this I've added a way to conditionally steal from the global
reserve if we can't make our normal reservation. If we have more than half the
space in the global reserve free we will go ahead and steal from the global
reserve. With this patch Dave's reproducer now works and I can rm all the files
on the file system. Thanks,
Reported-by: David Sterba <dsterba@suse.cz>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
The reason we introduce per-subvolume ordered extent list is the same
as the per-subvolume delalloc inode list.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
When we create a snapshot, we need flush all delalloc inodes in the
fs, just flushing the inodes in the source tree is OK. So we introduce
per-subvolume delalloc inode list.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
The grab/put funtions will be used in the next patch, which need grab
the root object and ensure it is not freed. We use reference counter
instead of the srcu lock is to aovid blocking the memory reclaim task,
which invokes synchronize_srcu().
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
There are several functions whose code is similar, such as
btrfs_find_last_root()
btrfs_read_fs_root_no_radix()
Besides that, some functions are invoked twice, it is unnecessary,
for example, we are sure that all roots which is found in
btrfs_find_orphan_roots()
have their orphan items, so it is unnecessary to check the orphan
item again.
So cleanup it.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
The snapshot/subvolume deletion might spend lots of time, it would make
the remount task wait for a long time. This patch improve this problem,
we will break the deletion if the fs is remounted to be R/O. It will make
the users happy.
Cc: David Sterba <dsterba@suse.cz>
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
The quota_tree was set up to use the empty_block_rsv before
which would be problematic when the filesystem is filled up
and ENOSPC happens during internal operations while the quota
tree is updated and COWed (when the btrfs_qgroup_info_item
items) are written. In fact, use_block_rsv() which is used
in btrfs_cow_block() falls back to the global_block_rsv in
this case. But just in order to make it more clear what is
happening, change it to explicitly use the global_block_rsv.
Signed-off-by: Stefan Behrens <sbehrens@giantdisaster.de>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Before applying this patch, we reserved the space for the global reserve
by the minimum unit if we found it is empty, it was unreasonable and
inefficient, because if the global reserve space was depleted, it implied
that the size of the global reserve was too small. In this case, we shoud
update the global reserve and fill it.
Cc: Tsutomu Itoh <t-itoh@jp.fujitsu.com>
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
If the type of the space we need is different with the global reserve, we
can not steal the space from the global reserve, because we can not allocate
the space from the free space cache that the global reserve points to.
Cc: Tsutomu Itoh <t-itoh@jp.fujitsu.com>
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
It is very likely that there are lots of subvolumes/snapshots in the filesystem,
so if we use global block reservation to do inode cache truncation, we may hog
all the free space that is reserved in global rsv. So it is better that we do
the free space reservation for inode cache truncation by ourselves.
Cc: Tsutomu Itoh <t-itoh@jp.fujitsu.com>
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Chris hit a bug where we weren't finding extent records when running extent ops.
This is because we use the delayed_ref_head when running the extent op, which
means we can't use the ->type checks to see if we are metadata. We also lose
the level of the metadata we are working on. So to fix this we can just check
the ->is_data section of the extent_op, and we can store the level of the buffer
we were modifying in the extent_op. Thanks,
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
The variable was named 'data' in btrfs_reserve_extent and that's the
only function that actually uses it to let btrfs_get_alloc_profile know
what profile we want. Then it's passed down as u64 flags.
Signed-off-by: David Sterba <dsterba@suse.cz>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Big patch, but all it does is add statics to functions which
are in fact static, then remove the associated dead-code fallout.
removed functions:
btrfs_iref_to_path()
__btrfs_lookup_delayed_deletion_item()
__btrfs_search_delayed_insertion_item()
__btrfs_search_delayed_deletion_item()
find_eb_for_page()
btrfs_find_block_group()
range_straddles_pages()
extent_range_uptodate()
btrfs_file_extent_length()
btrfs_scrub_cancel_devid()
btrfs_start_transaction_lflush()
btrfs_print_tree() is left because it is used for debugging.
btrfs_start_transaction_lflush() and btrfs_reada_detach() are
left for symmetry.
ulist.c functions are left, another patch will take care of those.
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
So everybody who got hit by my fsync bug will still continue to hit this
BUG_ON() in the free space cache, which is pretty heavy handed. So I took a
file system that had this bug and fixed up all the BUG_ON()'s and leaks that
popped up when I tried to mount a broken file system like this. With this patch
we just fail to mount instead of panicing. Thanks,
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
When running the 208th of xfstests, the fs returned the enospc
error when there was lots of free space in the disk.
By bisect debug, we found it was introduced by commit 96f1bb5777.
This commit makes the space check for the global reservation in
can_overcommit() be inconsistent with should_alloc_chunk().
can_overcommit() requires that the free space is 2 times the size
of the global reservation, or we can't do overcommit. And instead,
we need reclaim some reserved space, and if we still don't have
enough free space, we need allocate a new chunk. But unfortunately,
should_alloc_chunk() just requires that the free space is 1 time
the size of the global reservation, that is we would not try to
allocate a new chunk if the free space size is in the middle of
these two requires, and just return the enospc error. Fix it.
Cc: Jim Schutt <jaschut@sandia.gov>
Cc: Josef Bacik <jbacik@fusionio.com>
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Sequence numbers for delayed refs have been introduced in the first version
of the qgroup patch set. To solve the problem of find_all_roots on a busy
file system, the tree mod log was introduced. The sequence numbers for that
were simply shared between those two users.
However, at one point in qgroup's quota accounting, there's a statement
accessing the previous sequence number, that's still just doing (seq - 1)
just as it would have to in the very first version.
To satisfy that requirement, this patch makes the sequence number counter 64
bit and splits it into a major part (used for qgroup sequence number
counting) and a minor part (incremented for each tree modification in the
log). This enables us to go exactly one major step backwards, as required
for qgroups, while still incrementing the sequence counter for tree mod log
insertions to keep track of their order. Keeping them in a single variable
means there's no need to change all the code dealing with comparisons of two
sequence numbers.
The sequence number is reset to 0 on commit (not new in this patch), which
ensures we won't overflow the two 32 bit counters.
Without this fix, the qgroup tracking can occasionally go wrong and WARN_ONs
from the tree mod log code may happen.
Signed-off-by: Jan Schmidt <list.btrfs@jan-o-sch.net>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
We kept leaking extent buffers when mounting a broken file system and it turns
out it's because not everybody uses read_tree_block properly. You need to check
and make sure the extent_buffer is uptodate before you use it. This patch fixes
everybody who calls read_tree_block directly to make sure they check that it is
uptodate and free it and return an error if it is not. With this we no longer
leak EB's when things go horribly wrong. Thanks,
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
If we fail to load block groups halfway through we can leave extent_state's on
the excluded tree. This is because we just lookup the supers and add them to
the excluded tree regardless of which block group we are looking at currently.
This is a problem because we remove the excluded extents for the range of the
block group only, so if we don't ever load a block group for one of the excluded
extents we won't ever free it. This fixes the problem by only adding excluded
extents if it falls in the block group range we care about. With this patch
we're no longer leaking space when we fail to read all of the block groups.
Thanks,
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
So I noticed there is an infinite loop in the slow caching code. If we return 1
when we hit the end of the tree, so we could end up caching the last block group
the slow way and suddenly we're looping forever because we just keep
re-searching and trying again. Fix this by only doing btrfs_next_leaf() if we
don't need_resched(). Thanks,
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Argument 'trans' became unnecessary from setup_inline_extent_backref()
that called btrfs_extend_item().
Signed-off-by: Tsutomu Itoh <t-itoh@jp.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Argument 'trans' is not used in btrfs_extend_item().
Signed-off-by: Tsutomu Itoh <t-itoh@jp.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
If argument 'trans' is unnecessary in the function where
fixup_low_keys() is called, 'trans' is deleted.
Signed-off-by: Tsutomu Itoh <t-itoh@jp.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Dave was hitting a lockdep warning because we're now properly taking the ordered
operations mutex in the ordered wait stuff. This is because some cases we will
have a trans handle when we are flushing delalloc space, but we can't wait on
ordered extents because we could potentially deadlock, so fix this by not doing
the wait if we have a trans handle. Thanks
Reported-and-tested-by: David Sterba <dsterba@suse.cz>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
I noticed that we will add a block group to the space info before we add it to
the block group cache rb tree, so we could potentially allocate from the block
group before it's able to be searched for. I don't think this is too much of
a problem, the race window is microscopic, but just in case move the tree
insertion to above the space info linking. This makes it easier to adjust the
error handling as well, so we can remove a couple of BUG_ON(ret)'s and have real
error handling setup for these scenarios. Thanks,
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
With more than one btrfs volume mounted, it can be very difficult to find
out which volume is hitting an error. btrfs_error() will print this, but
it is currently rigged as more of a fatal error handler, while many of
the printk()s are currently for debugging and yet-unhandled cases.
This patch just changes the functions where the device information is
already available. Some cases remain where the root or fs_info is not
passed to the function emitting the error.
This may introduce some confusion with volumes backed by multiple devices
emitting errors referring to the primary device in the set instead of the
one on which the error occurred.
Use btrfs_printk(fs_info, format, ...) rather than writing the device
string every time, and introduce macro wrappers ala XFS for brevity.
Since the function already cannot be used for continuations, print a
newline as part of the btrfs_printk() message rather than at each caller.
Signed-off-by: Simon Kirby <sim@hostway.ca>
Reviewed-by: David Sterba <dsterba@suse.cz>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Each time pick one dead root from the list and let the caller know if
it's needed to continue. This should improve responsiveness during
umount and balance which at some point waits for cleaning all currently
queued dead roots.
A new dead root is added to the end of the list, so the snapshots
disappear in the order of deletion.
The snapshot cleaning work is now done only from the cleaner thread and the
others wake it if needed.
Signed-off-by: David Sterba <dsterba@suse.cz>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
We currently store the first key of the tree block inside the reference for the
tree block in the extent tree. This takes up quite a bit of space. Make a new
key type for metadata which holds the level as the offset and completely removes
storing the btrfs_tree_block_info inside the extent ref. This reduces the size
from 51 bytes to 33 bytes per extent reference for each tree block. In practice
this results in a 30-35% decrease in the size of our extent tree, which means we
COW less and can keep more of the extent tree in memory which makes our heavy
metadata operations go much faster. This is not an automatic format change, you
must enable it at mkfs time or with btrfstune. This patch deals with having
metadata stored as either the old format or the new format so it is easy to
convert. Thanks,
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Pull btrfs fixes from Chris Mason:
"We've had a busy two weeks of bug fixing. The biggest patches in here
are some long standing early-enospc problems (Josef) and a very old
race where compression and mmap combine forces to lose writes (me).
I'm fairly sure the mmap bug goes all the way back to the introduction
of the compression code, which is proof that fsx doesn't trigger every
possible mmap corner after all.
I'm sure you'll notice one of these is from this morning, it's a small
and isolated use-after-free fix in our scrub error reporting. I
double checked it here."
* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/mason/linux-btrfs:
Btrfs: don't drop path when printing out tree errors in scrub
Btrfs: fix wrong return value of btrfs_lookup_csum()
Btrfs: fix wrong reservation of csums
Btrfs: fix double free in the btrfs_qgroup_account_ref()
Btrfs: limit the global reserve to 512mb
Btrfs: hold the ordered operations mutex when waiting on ordered extents
Btrfs: fix space accounting for unlink and rename
Btrfs: fix space leak when we fail to reserve metadata space
Btrfs: fix EIO from btrfs send in is_extent_unchanged for punched holes
Btrfs: fix race between mmap writes and compression
Btrfs: fix memory leak in btrfs_create_tree()
Btrfs: fix locking on ROOT_REPLACE operations in tree mod log
Btrfs: fix missing qgroup reservation before fallocating
Btrfs: handle a bogus chunk tree nicely
Btrfs: update to use fs_state bit
A user reported a problem where he was getting early ENOSPC with hundreds of
gigs of free data space and 6 gigs of free metadata space. This is because the
global block reserve was taking up the entire free metadata space. This is
ridiculous, we have infrastructure in place to throttle if we start using too
much of the global reserve, so instead of letting it get this huge just limit it
to 512mb so that users can still get work done. This allowed the user to
complete his rsync without issues. Thanks
Cc: stable@vger.kernel.org
Reported-and-tested-by: Stefan Priebe <s.priebe@profihost.ag>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Dave reported a warning when running xfstest 275. We have been leaking delalloc
metadata space when our reservations fail. This is because we were improperly
calculating how much space to free for our checksum reservations. The problem
is we would sometimes free up space that had already been freed in another
thread and we would end up with negative usage for the delalloc space. This
patch fixes the problem by calculating how much space the other threads would
have already freed, and then calculate how much space we need to free had we not
done the reservation at all, and then freeing any excess space. This makes
xfstests 275 no longer have leaked space. Thanks
Cc: stable@vger.kernel.org
Reported-by: David Sterba <dsterba@suse.cz>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
If you restore a btrfs-image file system and try to mount that file system we'll
panic. That's because btrfs-image restores and just makes one big chunk to
envelope the whole disk, since they are really only meant to be messed with by
our btrfs-progs. So fix up btrfs_rmap_block and the callers of it for mount so
that we no longer panic but instead just return an error and fail to mount.
Thanks,
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Signed-off-by: Chris Mason <chris.mason@fusionio.com>
Pull btrfs fixes from Chris Mason:
"Eric's rcu barrier patch fixes a long standing problem with our
unmount code hanging on to devices in workqueue helpers. Liu Bo
nailed down a difficult assertion for in-memory extent mappings."
* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/mason/linux-btrfs:
Btrfs: fix warning of free_extent_map
Btrfs: fix warning when creating snapshots
Btrfs: return as soon as possible when edquot happens
Btrfs: return EIO if we have extent tree corruption
btrfs: use rcu_barrier() to wait for bdev puts at unmount
Btrfs: remove btrfs_try_spin_lock
Btrfs: get better concurrency for snapshot-aware defrag work
The callers of lookup_inline_extent_info all handle getting an error back
properly, so return an error if we have corruption instead of being a jerk and
panicing. Still WARN_ON() since this is kind of crucial and I've been seeing it
a bit too much recently for my taste, I think we're doing something wrong
somewhere. Thanks,
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Signed-off-by: Chris Mason <chris.mason@fusionio.com>
Pull btrfs update from Chris Mason:
"The biggest feature in the pull is the new (and still experimental)
raid56 code that David Woodhouse started long ago. I'm still working
on the parity logging setup that will avoid inconsistent parity after
a crash, so this is only for testing right now. But, I'd really like
to get it out to a broader audience to hammer out any performance
issues or other problems.
scrub does not yet correct errors on raid5/6 either.
Josef has another pass at fsync performance. The big change here is
to combine waiting for metadata with waiting for data, which is a big
latency win. It is also step one toward using atomics from the
hardware during a commit.
Mark Fasheh has a new way to use btrfs send/receive to send only the
metadata changes. SUSE is using this to make snapper more efficient
at finding changes between snapshosts.
Snapshot-aware defrag is also included.
Otherwise we have a large number of fixes and cleanups. Eric Sandeen
wins the award for removing the most lines, and I'm hoping we steal
this idea from XFS over and over again."
* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/mason/linux-btrfs: (118 commits)
btrfs: fixup/remove module.h usage as required
Btrfs: delete inline extents when we find them during logging
btrfs: try harder to allocate raid56 stripe cache
Btrfs: cleanup to make the function btrfs_delalloc_reserve_metadata more logic
Btrfs: don't call btrfs_qgroup_free if just btrfs_qgroup_reserve fails
Btrfs: remove reduplicate check about root in the function btrfs_clean_quota_tree
Btrfs: return ENOMEM rather than use BUG_ON when btrfs_alloc_path fails
Btrfs: fix missing deleted items in btrfs_clean_quota_tree
btrfs: use only inline_pages from extent buffer
Btrfs: fix wrong reserved space when deleting a snapshot/subvolume
Btrfs: fix wrong reserved space in qgroup during snap/subv creation
Btrfs: remove unnecessary dget_parent/dput when creating the pending snapshot
btrfs: remove a printk from scan_one_device
Btrfs: fix NULL pointer after aborting a transaction
Btrfs: fix memory leak of log roots
Btrfs: copy everything if we've created an inline extent
btrfs: cleanup for open-coded alignment
Btrfs: do not change inode flags in rename
Btrfs: use reserved space for creating a snapshot
clear chunk_alloc flag on retryable failure
...
The original code is a little confusing and not clear, The right
way to deal with the kernel code like this:
[...]
if (ret)
goto out;
[...]
So i move the common clean_up code to the place labeled with
out_fail, this will be easier to maintain.
Signed-off-by: Wang Shilong <wangsl-fnst@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
commit eb6b88d92c leads into another bug.
If it is just because qgroup_reserve fails, the function btrfs_qgroup_free
should not be called, otherwise, it will cause the wrong quota accounting.
Signed-off-by: Wang Shilong <wangsl-fnst@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
There are two problems in the space reservation of the snapshot/
subvolume creation.
- don't reserve the space for the root item insertion
- the space which is reserved in the qgroup is different with
the free space reservation. we need reserve free space for
7 items, but in qgroup reservation, we need reserve space only
for 3 items.
So we implement new metadata reservation functions for the
snapshot/subvolume creation.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Though most of the btrfs codes are using ALIGN macro for page alignment,
there are still some codes using open-coded alignment like the
following:
------
u64 mask = ((u64)root->stripesize - 1);
u64 ret = (val + mask) & ~mask;
------
Or even hidden one:
------
num_bytes = (end - start + blocksize) & ~(blocksize - 1);
------
Sometimes these open-coded alignment is not so easy to understand for
newbie like me.
This commit changes the open-coded alignment to the ALIGN macro for a
better readability.
Also there is a previous patch from David Sterba with similar changes,
but the patch is for 3.2 kernel and seems not merged.
http://www.spinics.net/lists/linux-btrfs/msg12747.html
Cc: David Sterba <dave@jikos.cz>
Signed-off-by: Qu Wenruo <quwenruo@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
I've experienced filesystem freezes with permanent spikes in the active
process count for quite a while, particularly on filesystems whose
available raw space has already been fully allocated to chunks.
While looking into this, I found a pretty obvious error in
do_chunk_alloc: it sets space_info->chunk_alloc, but if
btrfs_alloc_chunk returns an error other than ENOSPC, it returns leaving
that flag set, which causes any other threads waiting for
space_info->chunk_alloc to become zero to spin indefinitely.
I haven't double-checked that this patch fixes the failure I've observed
fully (it's not exactly trivial to trigger), but it surely is a bug and
the fix is trivial, so... Please put it in :-)
What I saw in that function also happens to explain why in some cases I
see filesystems allocate a huge number of chunks that remain unused
(leading to the scenario above, of not having more chunks to allocate).
It happens for data and metadata, but not necessarily both. I'm
guessing some thread sets the force_alloc flag on the corresponding
space_info, and then several threads trying to get disk space end up
attempting to allocate a new chunk concurrently. All of them will see
the force_alloc flag and bump their local copy of force up to the level
they see first, and they won't clear it even if another thread succeeds
in allocating a chunk, thus clearing the force flag. Then each thread
that observed the force flag will, on its turn, force the allocation of
a new chunk. And any threads that come in while it does that will see
the force flag still set and pick it up, and so on. This sounds like a
problem to me, but... what should the correct behavior be? Clear
force_flag once we copy it to a local force? Reset force to the
incoming value on every loop? Set the flag to our incoming force if we
have it at first, clear our local flag, and move it from the space_info
when we determined that we are the thread that's going to perform the
allocation?
btrfs: clear chunk_alloc flag on retryable failure
From: Alexandre Oliva <oliva@gnu.org>
If btrfs_alloc_chunk fails with e.g. ENOMEM, we exit do_chunk_alloc
without clearing chunk_alloc in space_info. As a result, any further
calls to do_chunk_alloc on that filesystem will start busy-waiting for
chunk_alloc to be cleared, but it never will be. This patch adjusts
do_chunk_alloc so that it clears this flag in case of an error.
Signed-off-by: Alexandre Oliva <oliva@gnu.org>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Pull trivial tree from Jiri Kosina:
"Assorted tiny fixes queued in trivial tree"
* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/jikos/trivial: (22 commits)
DocBook: update EXPORT_SYMBOL entry to point at export.h
Documentation: update top level 00-INDEX file with new additions
ARM: at91/ide: remove unsused at91-ide Kconfig entry
percpu_counter.h: comment code for better readability
x86, efi: fix comment typo in head_32.S
IB: cxgb3: delay freeing mem untill entirely done with it
net: mvneta: remove unneeded version.h include
time: x86: report_lost_ticks doesn't exist any more
pcmcia: avoid static analysis complaint about use-after-free
fs/jfs: Fix typo in comment : 'how may' -> 'how many'
of: add missing documentation for of_platform_populate()
btrfs: remove unnecessary cur_trans set before goto loop in join_transaction
sound: soc: Fix typo in sound/codecs
treewide: Fix typo in various drivers
btrfs: fix comment typos
Update ibmvscsi module name in Kconfig.
powerpc: fix typo (utilties -> utilities)
of: fix spelling mistake in comment
h8300: Fix home page URL in h8300/README
xtensa: Fix home page URL in Kconfig
...
Very large fallocate requests are cpu bound and result in extents with a
repeating pattern of ever decreasing size:
$ time fallocate -l 1T file
real 0m13.039s
( an excerpt of the extents from btrfs-debug-tree: )
prealloc data disk byte 1536292564992 nr 397312
prealloc data disk byte 1536292962304 nr 196608
prealloc data disk byte 1536293158912 nr 98304
prealloc data disk byte 1536293257216 nr 49152
prealloc data disk byte 1536293306368 nr 24576
prealloc data disk byte 1536293330944 nr 12288
prealloc data disk byte 1536293343232 nr 8192
prealloc data disk byte 1536293351424 nr 4096
prealloc data disk byte 1536293355520 nr 4096
prealloc data disk byte 1536293359616 nr 4096
The excessive cpu use comes from __btrfs_prealloc_file_range() trying to
allocate the entire remaining size after each extent is allocated.
btrfs_reserve_extent() repeatedly cuts this requested size in half until
it gets down to the size that the allocators can return. We limit the
problem for now by capping each reservation at 256 meg.
The small extents come from a masking bug when decreasing the requested
reservation size. The high 32bits are cleared and the remaining low
bits might happen to reserve a small size. Fix this by using
round_down() which properly casts the mask.
After these fixes huge fallocate requests are fast and result in nice
large extents:
$ time fallocate -l 1T file
real 0m0.082s
prealloc data disk byte 1112425889792 nr 268435456
prealloc data disk byte 1112694325248 nr 268435456
prealloc data disk byte 1112962760704 nr 268435456
Reported-by: Eric Sandeen <sandeen@redhat.com>
Signed-off-by: Zach Brown <zab@redhat.com>
Signed-off-by: Chris Mason <chris.mason@fusionio.com>
The warning in use_block_rsv is not useful for users and may fill
the logs unnecessarily.
Signed-off-by: David Sterba <dsterba@suse.cz>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
The deadlock problem happened when running fsstress(a test program in LTP).
Steps to reproduce:
# mkfs.btrfs -b 100M <partition>
# mount <partition> <mnt>
# <Path>/fsstress -p 3 -n 10000000 -d <mnt>
The reason is:
btrfs_direct_IO()
|->do_direct_IO()
|->get_page()
|->get_blocks()
| |->btrfs_delalloc_resereve_space()
| |->btrfs_add_ordered_extent() ------- Add a new ordered extent
|->dio_send_cur_page(page0) -------------- We didn't submit bio here
|->get_page()
|->get_blocks()
|->btrfs_delalloc_resereve_space()
|->flush_space()
|->btrfs_start_ordered_extent()
|->wait_event() ---------- Wait the completion of
the ordered extent that is
mentioned above
But because we didn't submit the bio that is mentioned above, the ordered
extent can not complete, we would wait for its completion forever.
There are two methods which can fix this deadlock problem:
1. submit the bio before we invoke get_blocks()
2. reserve the space before we do dio
Though the 1st is the simplest way, we need modify the code of VFS, and it
is likely to break contiguous requests, and introduce performance regression
for the other filesystems.
So we have to choose the 2nd way.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Cc: Josef Bacik <jbacik@fusionio.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Sometimes xfstest 83 will fail to remount the scratch device because we've
gotten ourselves so full that we cannot cleanup the orphan items. In this
case check to see if we're doing the orphan cleanup and if we are allow us
to steal our reservation from the global block rsv. With this patch I've
not been able to reproduce the failed mount problem. Thanks,
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
People have been complaining about random ENOSPC errors that will clear up
after a umount or just a given amount of time. Chris was able to reproduce
this with stress.sh and lots of processes and so was I. Basically the
overcommit stuff would really let us get out of hand, in my tests I saw up
to 30 gigs of outstanding reservations with only 2 gigs total of metadata
space. This usually worked out fine but with so much outstanding
reservation the flushing stuff short circuits to make sure we don't hang
forever flushing when we really need ENOSPC. Plus we allocate chunks in
order to alleviate the pressure, but this doesn't actually help us since we
only use the non-allocated area in our over commit logic.
So instead of basing overcommit on the amount of non-allocated space,
instead just do it based on how much total space we have, and then limit it
to the non-allocated space in case we are short on space to spill over into.
This allows us to have the same performance as well as no longer giving
random ENOSPC. Thanks,
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Because of how little we allocate chunks now we can get really tight on
metadata space before we will allocate a new chunk. This resulted in being
unable to add device extents when allocating a new metadata chunk as we did
not have enough space. This is because we were allowed to overcommit too
much metadata without actually making sure we had enough space to make
allocations. The idea behind overcommit is that we are allowed to say "sure
you can have that reservation" when most of the free space is occupied by
reservations, not actual allocations. But in this case where a majority of
the total space is in use by actual allocations we can screw ourselves by
not being able to make real allocations when it matters. So make sure we
have enough real space for our global reserve, and if not then don't allow
overcommitting. Thanks,
Reported-and-tested-by: Jim Schutt <jaschut@sandia.gov>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
There is no lock to protect
fs_info->avail_{data, metadata, system}_alloc_bits,
it may introduce some problem, such as the wrong profile
information, so we add a seqlock to protect them.
Signed-off-by: Zhao Lei <zhaolei@cn.fujitsu.com>
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
fs_info->delalloc_bytes is accessed very frequently, so use percpu
counter instead of the u64 variant for it to reduce the lock
contention.
This patch also fixed the problem that we access the variant
without the lock protection.At worst, we would not flush the
delalloc inodes, and just return ENOSPC error when we still have
some free space in the fs.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
