openwrt/staging/blogic.git
10 years agoBtrfs: check for an extent_op on the locked ref
Josef Bacik [Thu, 27 Mar 2014 23:41:34 +0000 (19:41 -0400)]
Btrfs: check for an extent_op on the locked ref

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>
10 years agoBtrfs: do not reset last_snapshot after relocation
Josef Bacik [Thu, 27 Mar 2014 18:56:51 +0000 (14:56 -0400)]
Btrfs: do not reset last_snapshot after relocation

This was done to allow NO_COW to continue to be NO_COW after relocation but it
is not right.  When relocating we will convert blocks to FULL_BACKREF that we
relocate.  We can leave some of these full backref blocks behind if they are not
cow'ed out during the relocation, like if we fail the relocation with ENOSPC and
then just drop the reloc tree.  Then when we go to cow the block again we won't
lookup the extent flags because we won't think there has been a snapshot
recently which means we will do our normal ref drop thing instead of adding back
a tree ref and dropping the shared ref.  This will cause btrfs_free_extent to
blow up because it can't find the ref we are trying to free.  This was found
with my ref verifying tool.  Thanks,

Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Chris Mason <clm@fb.com>
10 years agoBtrfs: fix a crash of clone with inline extents's split
Liu Bo [Mon, 10 Mar 2014 10:56:07 +0000 (18:56 +0800)]
Btrfs: fix a crash of clone with inline extents's split

xfstests's btrfs/035 triggers a BUG_ON, which we use to detect the split
of inline extents in __btrfs_drop_extents().

For inline extents, we cannot duplicate another EXTENT_DATA item, because
it breaks the rule of inline extents, that is, 'start offset' needs to be 0.

We have set limitations for the source inode's compressed inline extents,
because it needs to decompress and recompress.  Now the destination inode's
inline extents also need similar limitations.

With this, xfstests btrfs/035 doesn't run into panic.

Signed-off-by: Liu Bo <bo.li.liu@oracle.com>
Signed-off-by: Chris Mason <clm@fb.com>
10 years agobtrfs: fix uninit variable warning
Chris Mason [Fri, 21 Mar 2014 22:30:44 +0000 (15:30 -0700)]
btrfs: fix uninit variable warning

fs/btrfs/send.c:2926: warning: ‘entry’ may be used uninitialized in this
function

Signed-off-by: Chris Mason <clm@fb.com>
10 years agoBtrfs: take into account total references when doing backref lookup
Josef Bacik [Wed, 19 Mar 2014 17:35:14 +0000 (13:35 -0400)]
Btrfs: take into account total references when doing backref lookup

I added an optimization for large files where we would stop searching for
backrefs once we had looked at the number of references we currently had for
this extent.  This works great most of the time, but for snapshots that point to
this extent and has changes in the original root this assumption falls on it
face.  So keep track of any delayed ref mods made and add in the actual ref
count as reported by the extent item and use that to limit how far down an inode
we'll search for extents.  Thanks,

Reportedy-by: Hugo Mills <hugo@carfax.org.uk>
Signed-off-by: Josef Bacik <jbacik@fb.com>
Reported-by: Hugo Mills <hugo@carfax.org.uk>
Tested-by: Hugo Mills <hugo@carfax.org.uk>
Signed-off-by: Chris Mason <clm@fb.com>
10 years agoBtrfs: part 2, fix incremental send's decision to delay a dir move/rename
Filipe Manana [Wed, 19 Mar 2014 14:20:54 +0000 (14:20 +0000)]
Btrfs: part 2, fix incremental send's decision to delay a dir move/rename

For an incremental send, fix the process of determining whether the directory
inode we're currently processing needs to have its move/rename operation delayed.

We were ignoring the fact that if the inode's new immediate ancestor has a higher
inode number than ours but wasn't renamed/moved, we might still need to delay our
move/rename, because some other ancestor directory higher in the hierarchy might
have an inode number higher than ours *and* was renamed/moved too - in this case
we have to wait for rename/move of that ancestor to happen before our current
directory's rename/move operation.

Simple steps to reproduce this issue:

      $ mkfs.btrfs -f /dev/sdd
      $ mount /dev/sdd /mnt

      $ mkdir -p /mnt/a/x1/x2
      $ mkdir /mnt/a/Z
      $ mkdir -p /mnt/a/x1/x2/x3/x4/x5

      $ btrfs subvolume snapshot -r /mnt /mnt/snap1
      $ btrfs send /mnt/snap1 -f /tmp/base.send

      $ mv /mnt/a/x1/x2/x3 /mnt/a/Z/X33
      $ mv /mnt/a/x1/x2 /mnt/a/Z/X33/x4/x5/X22

      $ btrfs subvolume snapshot -r /mnt /mnt/snap2
      $ btrfs send -p /mnt/snap1 /mnt/snap2 -f /tmp/incremental.send

The incremental send caused the kernel code to enter an infinite loop when
building the path string for directory Z after its references are processed.

A more complex scenario:

      $ mkfs.btrfs -f /dev/sdd
      $ mount /dev/sdd /mnt

      $ mkdir -p /mnt/a/b/c/d
      $ mkdir /mnt/a/b/c/d/e
      $ mkdir /mnt/a/b/c/d/f
      $ mv /mnt/a/b/c/d/e /mnt/a/b/c/d/f/E2
      $ mkdir /mmt/a/b/c/g
      $ mv /mnt/a/b/c/d /mnt/a/b/D2

      $ btrfs subvolume snapshot -r /mnt /mnt/snap1
      $ btrfs send /mnt/snap1 -f /tmp/base.send

      $ mkdir /mnt/a/o
      $ mv /mnt/a/b/c/g /mnt/a/b/D2/f/G2
      $ mv /mnt/a/b/D2 /mnt/a/b/dd
      $ mv /mnt/a/b/c /mnt/a/C2
      $ mv /mnt/a/b/dd/f /mnt/a/o/FF
      $ mv /mnt/a/b /mnt/a/o/FF/E2/BB

      $ btrfs subvolume snapshot -r /mnt /mnt/snap2
      $ btrfs send -p /mnt/snap1 /mnt/snap2 -f /tmp/incremental.send

A test case for xfstests follows.

Signed-off-by: Filipe David Borba Manana <fdmanana@gmail.com>
Signed-off-by: Chris Mason <clm@fb.com>
10 years agoBtrfs: fix incremental send's decision to delay a dir move/rename
Filipe Manana [Sun, 16 Mar 2014 20:37:26 +0000 (20:37 +0000)]
Btrfs: fix incremental send's decision to delay a dir move/rename

It's possible to change the parent/child relationship between directories
in such a way that if a child directory has a higher inode number than
its parent, it doesn't necessarily means the child rename/move operation
can be performed immediately. The parent migth have its own rename/move
operation delayed, therefore in this case the child needs to have its
rename/move operation delayed too, and be performed after its new parent's
rename/move.

Steps to reproduce the issue:

      $ umount /mnt
      $ mkfs.btrfs -f /dev/sdd
      $ mount /dev/sdd /mnt

      $ mkdir /mnt/A
      $ mkdir /mnt/B
      $ mkdir /mnt/C
      $ mv /mnt/C /mnt/A
      $ mv /mnt/B /mnt/A/C
      $ mkdir /mnt/A/C/D

      $ btrfs subvolume snapshot -r /mnt /mnt/snap1
      $ btrfs send /mnt/snap1 -f /tmp/base.send

      $ mv /mnt/A/C/D /mnt/A/D2
      $ mv /mnt/A/C/B /mnt/A/D2/B2
      $ mv /mnt/A/C /mnt/A/D2/B2/C2

      $ btrfs subvolume snapshot -r /mnt /mnt/snap2
      $ btrfs send -p /mnt/snap1 /mnt/snap2 -f /tmp/incremental.send

The incremental send caused the kernel code to enter an infinite loop when
building the path string for directory C after its references are processed.

The necessary conditions here are that C has an inode number higher than both
A and B, and B as an higher inode number higher than A, and D has the highest
inode number, that is:
    inode_number(A) < inode_number(B) < inode_number(C) < inode_number(D)

The same issue could happen if after the first snapshot there's any number
of intermediary parent directories between A2 and B2, and between B2 and C2.

A test case for xfstests follows, covering this simple case and more advanced
ones, with files and hard links created inside the directories.

Signed-off-by: Filipe David Borba Manana <fdmanana@gmail.com>
Signed-off-by: Chris Mason <clm@fb.com>
10 years agoBtrfs: remove unnecessary inode generation lookup in send
Filipe Manana [Tue, 18 Mar 2014 17:56:06 +0000 (17:56 +0000)]
Btrfs: remove unnecessary inode generation lookup in send

No need to search in the send tree for the generation number of the inode,
we already have it in the recorded_ref structure passed to us.

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>
10 years agoBtrfs: fix race when updating existing ref head
Filipe Manana [Fri, 14 Mar 2014 20:55:01 +0000 (20:55 +0000)]
Btrfs: fix race when updating existing ref head

While we update an existing ref head's extent_op, we're not holding
its spinlock, so while we're updating its extent_op contents (key,
flags) we can have a task running __btrfs_run_delayed_refs() that
holds the ref head's lock and sets its extent_op to NULL right after
the task updating the ref head just checked its extent_op was not NULL.

Signed-off-by: Filipe David Borba Manana <fdmanana@gmail.com>
Signed-off-by: Chris Mason <clm@fb.com>
10 years agobtrfs: Add trace for btrfs_workqueue alloc/destroy
Qu Wenruo [Wed, 12 Mar 2014 08:05:33 +0000 (08:05 +0000)]
btrfs: Add trace for btrfs_workqueue alloc/destroy

Since most of the btrfs_workqueue is printed as pointer address,
for easier analysis, add trace for btrfs_workqueue alloc/destroy.
So it is possible to determine the workqueue that a given work belongs
to(by comparing the wq pointer address with alloc trace event).

Signed-off-by: Qu Wenruo <quenruo@cn.fujitsu.com>
Signed-off-by: Chris Mason <clm@fb.com>
10 years agoBtrfs: less fs tree lock contention when using autodefrag
Filipe Manana [Wed, 12 Mar 2014 01:28:24 +0000 (01:28 +0000)]
Btrfs: less fs tree lock contention when using autodefrag

When finding new extents during an autodefrag, don't do so many fs tree
lookups to find an extent with a size smaller then the target treshold.
Instead, after each fs tree forward search immediately unlock upper
levels and process the entire leaf while holding a read lock on the leaf,
since our leaf processing is very fast.
This reduces lock contention, allowing for higher concurrency when other
tasks want to write/update items related to other inodes in the fs tree,
as we're not holding read locks on upper tree levels while processing the
leaf and we do less tree searches.

Test:

    sysbench --test=fileio --file-num=512 --file-total-size=16G \
       --file-test-mode=rndrw --num-threads=32 --file-block-size=32768 \
       --file-rw-ratio=3 --file-io-mode=sync --max-time=1800 \
       --max-requests=10000000000 [prepare|run]

(fileystem mounted with -o autodefrag, averages of 5 runs)

Before this change: 58.852Mb/sec throughtput, read 77.589Gb, written 25.863Gb
After this change:  63.034Mb/sec throughtput, read 83.102Gb, written 27.701Gb

Test machine: quad core intel i5-3570K, 32Gb of RAM, SSD.

Signed-off-by: Filipe David Borba Manana <fdmanana@gmail.com>
Signed-off-by: Chris Mason <clm@fb.com>
10 years agoBtrfs: return EPERM when deleting a default subvolume
Guangyu Sun [Tue, 11 Mar 2014 18:24:18 +0000 (11:24 -0700)]
Btrfs: return EPERM when deleting a default subvolume

The error message is confusing:

 # btrfs sub delete /mnt/mysub/
 Delete subvolume '/mnt/mysub'
 ERROR: cannot delete '/mnt/mysub' - Directory not empty

The error message does not make sense to me: It's not about deleting a
directory but it's a subvolume, and it doesn't matter if the subvolume is
empty or not.

Maybe EPERM or is more appropriate in this case, combined with an explanatory
kernel log message. (e.g. "subvolume with ID 123 cannot be deleted because
it is configured as default subvolume.")

Reported-by: Koen De Wit <koen.de.wit@oracle.com>
Signed-off-by: Guangyu Sun <guangyu.sun@oracle.com>
Reviewed-by: David Sterba <dsterba@suse.cz>
Signed-off-by: Chris Mason <clm@fb.com>
10 years agoBtrfs: add missing kfree in btrfs_destroy_workqueue
Filipe Manana [Tue, 11 Mar 2014 14:31:44 +0000 (14:31 +0000)]
Btrfs: add missing kfree in btrfs_destroy_workqueue

Signed-off-by: Filipe David Borba Manana <fdmanana@gmail.com>
Signed-off-by: Chris Mason <clm@fb.com>
10 years agoBtrfs: cache extent states in defrag code path
Filipe Manana [Tue, 11 Mar 2014 13:56:15 +0000 (13:56 +0000)]
Btrfs: cache extent states in defrag code path

When locking file ranges in the inode's io_tree, cache the first
extent state that belongs to the target range, so that when unlocking
the range we don't need to search in the io_tree again, reducing cpu
time and making and therefore holding the io_tree's lock for a shorter
period.

Signed-off-by: Filipe David Borba Manana <fdmanana@gmail.com>
Signed-off-by: Chris Mason <clm@fb.com>
10 years agoBtrfs: fix deadlock with nested trans handles
Josef Bacik [Fri, 7 Mar 2014 00:01:07 +0000 (19:01 -0500)]
Btrfs: fix deadlock with nested trans handles

Zach found this deadlock that would happen like this

btrfs_end_transaction <- reduce trans->use_count to 0
  btrfs_run_delayed_refs
    btrfs_cow_block
      find_free_extent
btrfs_start_transaction <- increase trans->use_count to 1
          allocate chunk
btrfs_end_transaction <- decrease trans->use_count to 0
  btrfs_run_delayed_refs
    lock tree block we are cowing above ^^

We need to only decrease trans->use_count if it is above 1, otherwise leave it
alone.  This will make nested trans be the only ones who decrease their added
ref, and will let us get rid of the trans->use_count++ hack if we have to commit
the transaction.  Thanks,

cc: stable@vger.kernel.org
Reported-by: Zach Brown <zab@redhat.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
Tested-by: Zach Brown <zab@redhat.com>
Signed-off-by: Chris Mason <clm@fb.com>
10 years agoBtrfs: fix possible empty list access when flushing the delalloc inodes
Miao Xie [Thu, 6 Mar 2014 05:55:03 +0000 (13:55 +0800)]
Btrfs: fix possible empty list access when flushing the delalloc inodes

We didn't have a lock to protect the access to the delalloc inodes list, that is
we might access a empty delalloc inodes list if someone start flushing delalloc
inodes because the delalloc inodes were moved into a other list temporarily.
Fix it by wrapping the access with a lock.

Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agoBtrfs: split the global ordered extents mutex
Miao Xie [Thu, 6 Mar 2014 05:55:02 +0000 (13:55 +0800)]
Btrfs: split the global ordered extents mutex

When we create a snapshot, we just need wait the ordered extents in
the source fs/file root, but because we use the global mutex to protect
this ordered extents list of the source fs/file root to avoid accessing
a empty list, if someone got the mutex to access the ordered extents list
of the other fs/file root, we had to wait.

This patch splits the above global mutex, now every fs/file root has
its own mutex to protect its own list.

Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agoBtrfs: don't flush all delalloc inodes when we doesn't get s_umount lock
Miao Xie [Thu, 6 Mar 2014 05:55:01 +0000 (13:55 +0800)]
Btrfs: don't flush all delalloc inodes when we doesn't get s_umount lock

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>
10 years agoBtrfs: reclaim delalloc metadata more aggressively
Miao Xie [Thu, 6 Mar 2014 05:55:00 +0000 (13:55 +0800)]
Btrfs: reclaim delalloc metadata more aggressively

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>
10 years agoBtrfs: remove unnecessary lock in may_commit_transaction()
Miao Xie [Thu, 6 Mar 2014 05:54:59 +0000 (13:54 +0800)]
Btrfs: remove unnecessary lock in may_commit_transaction()

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>
10 years agoBtrfs: remove the unnecessary flush when preparing the pages
Miao Xie [Thu, 6 Mar 2014 05:54:58 +0000 (13:54 +0800)]
Btrfs: remove the unnecessary flush when preparing the pages

Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agoBtrfs: just do dirty page flush for the inode with compression before direct IO
Miao Xie [Thu, 6 Mar 2014 05:54:57 +0000 (13:54 +0800)]
Btrfs: just do dirty page flush for the inode with compression before direct IO

As the comment in the btrfs_direct_IO says, only the compressed pages need be
flush again to make sure they are on the disk, but the common pages needn't,
so we add a if statement to check if the inode has compressed pages or not,
if no, skip the flush.

And in order to prevent the write ranges from intersecting, we need wait for
the running ordered extents. But the current code waits for them twice, one
is done before the direct IO starts (in btrfs_wait_ordered_range()), the other
is before we get the blocks, it is unnecessary. because we can do the direct
IO without holding i_mutex, it means that the intersected ordered extents may
happen during the direct IO, the first wait can not avoid this problem. So we
use filemap_fdatawrite_range() instead of btrfs_wait_ordered_range() to remove
the first wait.

Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agoBtrfs: wake up the tasks that wait for the io earlier
Miao Xie [Thu, 6 Mar 2014 05:54:56 +0000 (13:54 +0800)]
Btrfs: wake up the tasks that wait for the io earlier

The tasks that wait for the IO_DONE flag just care about the io of the dirty
pages, so it is better to wake up them immediately after all the pages are
written, not the whole process of the io completes.

Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agoBtrfs: fix early enospc due to the race of the two ordered extent wait
Miao Xie [Thu, 6 Mar 2014 05:54:55 +0000 (13:54 +0800)]
Btrfs: fix early enospc due to the race of the two ordered extent wait

btrfs_wait_ordered_roots() moves all the list entries to a new list,
and then deals with them one by one. But if the other task invokes this
function at that time, it would get a empty list. It makes the enospc
error happens more early. Fix it.

Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agoBtrfs: introduce btrfs_{start, end}_nocow_write() for each subvolume
Miao Xie [Thu, 6 Mar 2014 05:38:19 +0000 (13:38 +0800)]
Btrfs: introduce btrfs_{start, end}_nocow_write() for each subvolume

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>
10 years agobtrfs: Add ftrace for btrfs_workqueue
Qu Wenruo [Thu, 6 Mar 2014 04:19:50 +0000 (04:19 +0000)]
btrfs: Add ftrace for btrfs_workqueue

Add ftrace for btrfs_workqueue for further workqueue tunning.
This patch needs to applied after the workqueue replace patchset.

Signed-off-by: Qu Wenruo <quwenruo@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agobtrfs: Cleanup the btrfs_workqueue related function type
Qu Wenruo [Thu, 6 Mar 2014 04:19:50 +0000 (04:19 +0000)]
btrfs: Cleanup the btrfs_workqueue related function type

The new btrfs_workqueue still use open-coded function defition,
this patch will change them into btrfs_func_t type which is much the
same as kernel workqueue.

Signed-off-by: Qu Wenruo <quwenruo@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agoBtrfs: add readahead for send_write
Liu Bo [Wed, 5 Mar 2014 02:07:35 +0000 (10:07 +0800)]
Btrfs: add readahead for send_write

Btrfs send reads data from disk and then writes to a stream via pipe or
a file via flush.

Currently we're going to read each page a time, so every page results
in a disk read, which is not friendly to disks, esp. HDD.  Given that,
the performance can be gained by adding readahead for those pages.

Here is a quick test:
$ btrfs subvolume create send
$ xfs_io -f -c "pwrite 0 1G" send/foobar
$ btrfs subvolume snap -r send ro
$ time "btrfs send ro -f /dev/null"

           w/o             w
real    1m37.527s       0m9.097s
user    0m0.122s        0m0.086s
sys     0m53.191s       0m12.857s

Signed-off-by: Liu Bo <bo.li.liu@oracle.com>
Reviewed-by: David Sterba <dsterba@suse.cz>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agoBtrfs: share the same code for __record_{new,deleted}_ref
Liu Bo [Mon, 3 Mar 2014 13:31:03 +0000 (21:31 +0800)]
Btrfs: share the same code for __record_{new,deleted}_ref

This has no functional change, only picks out the same part of two functions,
and makes it shared.

Signed-off-by: Liu Bo <bo.li.liu@oracle.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agoBtrfs: avoid unnecessary utimes update in incremental send
Filipe Manana [Mon, 3 Mar 2014 12:28:40 +0000 (12:28 +0000)]
Btrfs: avoid unnecessary utimes update in incremental send

When we're finishing processing of an inode, if we're dealing with a
directory inode that has a pending move/rename operation, we don't
need to send a utimes update instruction to the send stream, as we'll
do it later after doing the move/rename operation. Therefore we save
some time here building paths and doing btree lookups.

Signed-off-by: Filipe David Borba Manana <fdmanana@gmail.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agoBtrfs: make defrag not fragment files when using prealloc extents
Filipe Manana [Sat, 1 Mar 2014 10:57:03 +0000 (10:57 +0000)]
Btrfs: make defrag not fragment files when using prealloc extents

When using prealloc extents, a file defragment operation may actually
fragment the file and increase the amount of data space used by the file.
This change fixes that behaviour.

Example:

$ mkfs.btrfs -f /dev/sdb3
$ mount /dev/sdb3 /mnt
$ cd /mnt
$ xfs_io -f -c 'falloc 0 1048576' foobar && sync
$ xfs_io -c 'pwrite -S 0xff -b 100000 5000 100000' foobar
$ xfs_io -c 'pwrite -S 0xac -b 100000 200000 100000' foobar
$ xfs_io -c 'pwrite -S 0xe1 -b 100000 900000 100000' foobar && sync

Before defragmenting the file:

$ btrfs filesystem df /mnt
Data, single: total=8.00MiB, used=1.25MiB
System, DUP: total=8.00MiB, used=16.00KiB
System, single: total=4.00MiB, used=0.00
Metadata, DUP: total=1.00GiB, used=112.00KiB
Metadata, single: total=8.00MiB, used=0.00

$ btrfs-debug-tree /dev/sdb3
(...)
item 6 key (257 EXTENT_DATA 0) itemoff 15810 itemsize 53
prealloc data disk byte 12845056 nr 1048576
prealloc data offset 0 nr 4096
item 7 key (257 EXTENT_DATA 4096) itemoff 15757 itemsize 53
extent data disk byte 12845056 nr 1048576
extent data offset 4096 nr 102400 ram 1048576
extent compression 0
item 8 key (257 EXTENT_DATA 106496) itemoff 15704 itemsize 53
prealloc data disk byte 12845056 nr 1048576
prealloc data offset 106496 nr 90112
item 9 key (257 EXTENT_DATA 196608) itemoff 15651 itemsize 53
extent data disk byte 12845056 nr 1048576
extent data offset 196608 nr 106496 ram 1048576
extent compression 0
item 10 key (257 EXTENT_DATA 303104) itemoff 15598 itemsize 53
prealloc data disk byte 12845056 nr 1048576
prealloc data offset 303104 nr 593920
item 11 key (257 EXTENT_DATA 897024) itemoff 15545 itemsize 53
extent data disk byte 12845056 nr 1048576
extent data offset 897024 nr 106496 ram 1048576
extent compression 0
item 12 key (257 EXTENT_DATA 1003520) itemoff 15492 itemsize 53
prealloc data disk byte 12845056 nr 1048576
prealloc data offset 1003520 nr 45056
(...)

Now defragmenting the file results in more data space used than before:

$ btrfs filesystem defragment -f foobar && sync
$ btrfs filesystem df /mnt
Data, single: total=8.00MiB, used=1.55MiB
System, DUP: total=8.00MiB, used=16.00KiB
System, single: total=4.00MiB, used=0.00
Metadata, DUP: total=1.00GiB, used=112.00KiB
Metadata, single: total=8.00MiB, used=0.00

And the corresponding file extent items are now no longer perfectly sequential
as before, and we're now needlessly using more space from data block groups:

$ btrfs-debug-tree /dev/sdb3
(...)
item 6 key (257 EXTENT_DATA 0) itemoff 15810 itemsize 53
extent data disk byte 12845056 nr 1048576
extent data offset 0 nr 4096 ram 1048576
extent compression 0
item 7 key (257 EXTENT_DATA 4096) itemoff 15757 itemsize 53
extent data disk byte 13893632 nr 102400
extent data offset 0 nr 102400 ram 102400
extent compression 0
item 8 key (257 EXTENT_DATA 106496) itemoff 15704 itemsize 53
extent data disk byte 12845056 nr 1048576
extent data offset 106496 nr 90112 ram 1048576
extent compression 0
item 9 key (257 EXTENT_DATA 196608) itemoff 15651 itemsize 53
extent data disk byte 13996032 nr 106496
extent data offset 0 nr 106496 ram 106496
extent compression 0
item 10 key (257 EXTENT_DATA 303104) itemoff 15598 itemsize 53
prealloc data disk byte 12845056 nr 1048576
prealloc data offset 303104 nr 593920
item 11 key (257 EXTENT_DATA 897024) itemoff 15545 itemsize 53
extent data disk byte 14102528 nr 106496
extent data offset 0 nr 106496 ram 106496
extent compression 0
item 12 key (257 EXTENT_DATA 1003520) itemoff 15492 itemsize 53
extent data disk byte 12845056 nr 1048576
extent data offset 1003520 nr 45056 ram 1048576
extent compression 0
(...)

With this change, the above example will no longer cause allocation of new data
space nor change the sequentiality of the file extents, that is, defragment will
be effectless, leaving all extent items pointing to the extent starting at disk
byte 12845056.

In a 20Gb filesystem I had, mounted with the autodefrag option and 20 files of
400Mb each, initially consisting of a single prealloc extent of 400Mb, having
random writes happening at a low rate, lead to a total of over ~17Gb of data
space used, not far from eventually reaching an ENOSPC state.

Signed-off-by: Filipe David Borba Manana <fdmanana@gmail.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agoBtrfs: correctly flush data on defrag when compression is enabled
Filipe Manana [Sat, 1 Mar 2014 10:55:54 +0000 (10:55 +0000)]
Btrfs: correctly flush data on defrag when compression is enabled

When the defrag flag BTRFS_DEFRAG_RANGE_START_IO is set and compression
enabled, we weren't flushing completely, as writing compressed extents
is a 2 steps process, one to compress the data and another one to write
the compressed data to disk.

Signed-off-by: Filipe David Borba Manana <fdmanana@gmail.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agobtrfs: Cleanup the "_struct" suffix in btrfs_workequeue
Qu Wenruo [Fri, 28 Feb 2014 02:46:19 +0000 (10:46 +0800)]
btrfs: Cleanup the "_struct" suffix in btrfs_workequeue

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>
10 years agobtrfs: Cleanup the old btrfs_worker.
Qu Wenruo [Fri, 28 Feb 2014 02:46:18 +0000 (10:46 +0800)]
btrfs: Cleanup the old btrfs_worker.

Since all the btrfs_worker is replaced with the newly created
btrfs_workqueue, the old codes can be easily remove.

Signed-off-by: Quwenruo <quwenruo@cn.fujitsu.com>
Tested-by: David Sterba <dsterba@suse.cz>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agobtrfs: Replace fs_info->scrub_* workqueue with btrfs_workqueue.
Qu Wenruo [Fri, 28 Feb 2014 02:46:17 +0000 (10:46 +0800)]
btrfs: Replace fs_info->scrub_* workqueue with btrfs_workqueue.

Replace the fs_info->scrub_* 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>
10 years agobtrfs: Replace fs_info->qgroup_rescan_worker workqueue with btrfs_workqueue.
Qu Wenruo [Fri, 28 Feb 2014 02:46:16 +0000 (10:46 +0800)]
btrfs: Replace fs_info->qgroup_rescan_worker workqueue with btrfs_workqueue.

Replace the fs_info->qgroup_rescan_worker 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>
10 years agobtrfs: Replace fs_info->delayed_workers workqueue with btrfs_workqueue.
Qu Wenruo [Fri, 28 Feb 2014 02:46:15 +0000 (10:46 +0800)]
btrfs: Replace fs_info->delayed_workers workqueue with btrfs_workqueue.

Replace the fs_info->delayed_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>
10 years agobtrfs: Replace fs_info->fixup_workers workqueue with btrfs_workqueue.
Qu Wenruo [Fri, 28 Feb 2014 02:46:14 +0000 (10:46 +0800)]
btrfs: Replace fs_info->fixup_workers workqueue with btrfs_workqueue.

Replace the fs_info->fixup_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>
10 years agobtrfs: Replace fs_info->readahead_workers workqueue with btrfs_workqueue.
Qu Wenruo [Fri, 28 Feb 2014 02:46:13 +0000 (10:46 +0800)]
btrfs: Replace fs_info->readahead_workers workqueue with btrfs_workqueue.

Replace the fs_info->readahead_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>
10 years agobtrfs: Replace fs_info->cache_workers workqueue with btrfs_workqueue.
Qu Wenruo [Fri, 28 Feb 2014 02:46:12 +0000 (10:46 +0800)]
btrfs: Replace fs_info->cache_workers workqueue with btrfs_workqueue.

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>
10 years agobtrfs: Replace fs_info->rmw_workers workqueue with btrfs_workqueue.
Qu Wenruo [Fri, 28 Feb 2014 02:46:11 +0000 (10:46 +0800)]
btrfs: Replace fs_info->rmw_workers workqueue with btrfs_workqueue.

Replace the fs_info->rmw_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>
10 years agobtrfs: Replace fs_info->endio_* workqueue with btrfs_workqueue.
Qu Wenruo [Fri, 28 Feb 2014 02:46:10 +0000 (10:46 +0800)]
btrfs: Replace fs_info->endio_* workqueue with btrfs_workqueue.

Replace the fs_info->endio_* workqueues 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>
10 years agobtrfs: Replace fs_info->flush_workers with btrfs_workqueue.
Qu Wenruo [Fri, 28 Feb 2014 02:46:09 +0000 (10:46 +0800)]
btrfs: Replace fs_info->flush_workers with btrfs_workqueue.

Replace the fs_info->submit_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>
10 years agobtrfs: Replace fs_info->submit_workers with btrfs_workqueue.
Qu Wenruo [Fri, 28 Feb 2014 02:46:08 +0000 (10:46 +0800)]
btrfs: Replace fs_info->submit_workers with btrfs_workqueue.

Much like the fs_info->workers, replace the fs_info->submit_workers
use the same 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>
10 years agobtrfs: Replace fs_info->delalloc_workers with btrfs_workqueue
Qu Wenruo [Fri, 28 Feb 2014 02:46:07 +0000 (10:46 +0800)]
btrfs: Replace fs_info->delalloc_workers with btrfs_workqueue

Much like the fs_info->workers, replace the fs_info->delalloc_workers
use the same 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>
10 years agobtrfs: Replace fs_info->workers with btrfs_workqueue.
Qu Wenruo [Fri, 28 Feb 2014 02:46:06 +0000 (10:46 +0800)]
btrfs: Replace fs_info->workers with btrfs_workqueue.

Use the newly created btrfs_workqueue_struct to replace the original
fs_info->workers

Signed-off-by: Qu Wenruo <quwenruo@cn.fujitsu.com>
Tested-by: David Sterba <dsterba@suse.cz>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agobtrfs: Add threshold workqueue based on kernel workqueue
Qu Wenruo [Fri, 28 Feb 2014 02:46:05 +0000 (10:46 +0800)]
btrfs: Add threshold workqueue based on kernel workqueue

The original btrfs_workers has thresholding functions to dynamically
create or destroy kthreads.

Though there is no such function in kernel workqueue because the worker
is not created manually, we can still use the workqueue_set_max_active
to simulated the behavior, mainly to achieve a better HDD performance by
setting a high threshold on submit_workers.
(Sadly, no resource can be saved)

So in this patch, extra workqueue pending counters are introduced to
dynamically change the max active of each btrfs_workqueue_struct, hoping
to restore the behavior of the original thresholding function.

Also, workqueue_set_max_active use a mutex to protect workqueue_struct,
which is not meant to be called too frequently, so a new interval
mechanism is applied, that will only call workqueue_set_max_active after
a count of work is queued. Hoping to balance both the random and
sequence performance on HDD.

Signed-off-by: Qu Wenruo <quwenruo@cn.fujitsu.com>
Tested-by: David Sterba <dsterba@suse.cz>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agobtrfs: Add high priority workqueue support for btrfs_workqueue_struct
Qu Wenruo [Fri, 28 Feb 2014 02:46:04 +0000 (10:46 +0800)]
btrfs: Add high priority workqueue support for btrfs_workqueue_struct

Add high priority function to btrfs_workqueue.

This is implemented by embedding a btrfs_workqueue into a
btrfs_workqueue and use some helper functions to differ the normal
priority wq and high priority wq.
So the high priority wq is completely independent from the normal
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>
10 years agobtrfs: Added btrfs_workqueue_struct implemented ordered execution based on kernel...
Qu Wenruo [Fri, 28 Feb 2014 02:46:03 +0000 (10:46 +0800)]
btrfs: Added btrfs_workqueue_struct implemented ordered execution based on kernel workqueue

Use kernel workqueue to implement a new btrfs_workqueue_struct, which
has the ordering execution feature like the btrfs_worker.

The func is executed in a concurrency way, and the
ordred_func/ordered_free is executed in the sequence them are queued
after the corresponding func is done.

The new btrfs_workqueue works much like the original one, one workqueue
for normal work and a list for ordered work.
When a work is queued, ordered work will be added to the list and helper
function will be queued into the workqueue.
The helper function will execute a normal work and then check and execute as many
ordered work as possible in the sequence they were queued.

At this patch, high priority work queue or thresholding is not added yet.
The high priority feature and thresholding will be added in the following patches.

Signed-off-by: Qu Wenruo <quwenruo@cn.fujitsu.com>
Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Tested-by: David Sterba <dsterba@suse.cz>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agobtrfs: Cleanup the unused struct async_sched.
Qu Wenruo [Fri, 28 Feb 2014 02:46:02 +0000 (10:46 +0800)]
btrfs: Cleanup the unused struct async_sched.

The struct async_sched is not used by any codes and can be removed.

Signed-off-by: Qu Wenruo <quwenruo@cn.fujitsu.com>
Reviewed-by: Josef Bacik <jbacik@fusionio.com>
Tested-by: David Sterba <dsterba@suse.cz>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agoBtrfs: skip search tree for REG files
Liu Bo [Thu, 27 Feb 2014 09:29:01 +0000 (17:29 +0800)]
Btrfs: skip search tree for REG files

It is really unnecessary to search tree again for @gen, @mode and @rdev
in the case of REG inodes' creation, as we've got btrfs_inode_item in sctx,
and @gen, @mode and @rdev can easily be fetched.

Signed-off-by: Liu Bo <bo.li.liu@oracle.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agoBtrfs: fix preallocate vs double nocow write
Miao Xie [Thu, 27 Feb 2014 05:58:05 +0000 (13:58 +0800)]
Btrfs: fix preallocate vs double nocow write

We can not release the reserved metadata space for the first write if we
find the write position is pre-allocated. Because the kernel might write
the data on the disk before we do the second write but after the can-nocow
check, if we release the space for the first write, we might fail to update
the metadata because of no space.

Fix this problem by end nocow write if there is dirty data in the range whose
space is pre-allocated.

Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agoBtrfs: fix wrong lock range and write size in check_can_nocow()
Miao Xie [Thu, 27 Feb 2014 05:58:04 +0000 (13:58 +0800)]
Btrfs: fix wrong lock range and write size in check_can_nocow()

The write range may not be sector-aligned, for example:

       |--------|--------| <- write range, sector-unaligned, size: 2blocks
  |--------|--------|--------|  <- correct lock range, size: 3blocks

But according to the old code, we used the size of write range to calculate
the lock range directly, not considered the offset, we would get a wrong lock
range:

       |--------|--------| <- write range, sector-unaligned, size: 2blocks
  |--------|--------| <- wrong lock range, size: 2blocks

And besides that, the old code also had the same problem when calculating
the real write size. Correct them.

Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agobtrfs: send: simplify allocation code in fs_path_ensure_buf
David Sterba [Tue, 25 Feb 2014 18:33:08 +0000 (19:33 +0100)]
btrfs: send: simplify allocation code in fs_path_ensure_buf

Signed-off-by: David Sterba <dsterba@suse.cz>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agobtrfs: send: fix old buffer length in fs_path_ensure_buf
David Sterba [Tue, 25 Feb 2014 18:32:59 +0000 (19:32 +0100)]
btrfs: send: fix old buffer length in fs_path_ensure_buf

In "btrfs: send: lower memory requirements in common case" the code to
save the old_buf_len was incorrectly moved to a wrong place and broke
the original logic.

Reported-by: Filipe David Manana <fdmanana@gmail.com>
Signed-off-by: David Sterba <dsterba@suse.cz>
Reviewed-by: Filipe David Manana <fdmanana@gmail.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agoBtrfs: more efficient btrfs_drop_extent_cache
Filipe Manana [Tue, 25 Feb 2014 14:15:13 +0000 (14:15 +0000)]
Btrfs: more efficient btrfs_drop_extent_cache

While droping extent map structures from the extent cache that cover our
target range, we would remove each extent map structure from the red black
tree and then add either 1 or 2 new extent map structures if the former
extent map covered sections outside our target range.

This change simply attempts to replace the existing extent map structure
with a new one that covers the subsection we're not interested in, instead
of doing a red black remove operation followed by an insertion operation.

The number of elements in an inode's extent map tree can get very high for large
files under random writes. For example, while running the following test:

    sysbench --test=fileio --file-num=1 --file-total-size=10G \
        --file-test-mode=rndrw --num-threads=32 --file-block-size=32768 \
        --max-requests=500000 --file-rw-ratio=2 [prepare|run]

I captured the following histogram capturing the number of extent_map items
in the red black tree while that test was running:

    Count: 122462
    Range:  1.000 - 172231.000; Mean: 96415.831; Median: 101855.000; Stddev: 49700.981
    Percentiles:  90th: 160120.000; 95th: 166335.000; 99th: 171070.000
       1.000 -    5.231:   452 |
       5.231 -  187.392:    87 |
     187.392 -  585.911:   206 |
     585.911 - 1827.438:   623 |
    1827.438 - 5695.245:  1962 #
    5695.245 - 17744.861:  6204 ####
   17744.861 - 55283.764: 21115 ############
   55283.764 - 172231.000: 91813 #####################################################

Benchmark:

    sysbench --test=fileio --file-num=1 --file-total-size=10G --file-test-mode=rndwr \
        --num-threads=64 --file-block-size=32768 --max-requests=0 --max-time=60 \
        --file-io-mode=sync --file-fsync-freq=0 [prepare|run]

Before this change: 122.1Mb/sec
After this change:  125.07Mb/sec
(averages of 5 test runs)

Test machine: quad core intel i5-3570K, 32Gb of ram, SSD

Signed-off-by: Filipe David Borba Manana <fdmanana@gmail.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agoBtrfs: more efficient split extent state insertion
Filipe Manana [Wed, 12 Feb 2014 15:05:53 +0000 (15:05 +0000)]
Btrfs: more efficient split extent state insertion

When we split an extent state there's no need to start the rbtree search
from the root node - we can start it from the original extent state node,
since we would end up in its subtree if we do the search starting at the
root node anyway.

Signed-off-by: Filipe David Borba Manana <fdmanana@gmail.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agoBtrfs: remove unneeded field / smaller extent_map structure
Filipe Manana [Tue, 25 Feb 2014 14:15:12 +0000 (14:15 +0000)]
Btrfs: remove unneeded field / smaller extent_map structure

We don't need to have an unsigned int field in the extent_map struct
to tell us whether the extent map is in the inode's extent_map tree or
not. We can use the rb_node struct field and the RB_CLEAR_NODE and
RB_EMPTY_NODE macros to achieve the same task.

This reduces sizeof(struct extent_map) from 152 bytes to 144 bytes (on a
64 bits system).

Signed-off-by: Filipe David Borba Manana <fdmanana@gmail.com>
Reviewed-by: David Sterba <dsterba@suse.cz>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agoBtrfs: skip locking when searching commit root
Wang Shilong [Thu, 13 Feb 2014 03:19:47 +0000 (11:19 +0800)]
Btrfs: skip locking when searching commit root

We won't change commit root, skip locking dance with commit root
when walking backrefs, this can speed up btrfs send operations.

Signed-off-by: Wang Shilong <wangsl.fnst@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agoBtrfs: wake up @scrub_pause_wait as much as we can
Wang Shilong [Wed, 19 Feb 2014 11:24:19 +0000 (19:24 +0800)]
Btrfs: wake up @scrub_pause_wait as much as we can

check if @scrubs_running=@scrubs_paused condition inside wait_event()
is not an atomic operation which means we may inc/dec @scrub_running/
paused at any time. Let's wake up @scrub_pause_wait as much as we can
to let commit transaction blocked less.

An example below:

Thread1 Thread2
|->scrub_blocked_if_needed() |->scrub_pending_trans_workers_inc
  |->increase @scrub_paused
                                       |->increase @scrub_running
  |->wake up scrub_pause_wait list
                                       |->scrub blocked
                                       |->increase @scrub_paused

Thread3 is commiting transaction which is blocked at btrfs_scrub_pause().
So after Thread2 increase @scrub_paused, we meet the condition
@scrub_paused=@scrub_running, but transaction will be still blocked until
another calling to wake up @scrub_pause_wait.

Signed-off-by: Wang Shilong <wangsl.fnst@cn.fujitsu.com>
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agoBtrfs: cancel scrub on transaction abortion
Wang Shilong [Wed, 19 Feb 2014 11:24:18 +0000 (19:24 +0800)]
Btrfs: cancel scrub on transaction abortion

If we fail to commit transaction, we'd better
cancel scrub operations.

Suggested-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Wang Shilong <wangsl.fnst@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agoBtrfs: device_replace: fix deadlock for nocow case
Wang Shilong [Wed, 19 Feb 2014 11:24:17 +0000 (19:24 +0800)]
Btrfs: device_replace: fix deadlock for nocow case

commit cb7ab02156e4 cause a following deadlock found by
xfstests,btrfs/011:

Thread1 is commiting transaction which is blocked at
btrfs_scrub_pause().

Thread2 is calling btrfs_file_aio_write() which has held
inode's @i_mutex and commit transaction(blocked because
Thread1 is committing transaction).

Thread3 is copy_nocow_page worker which will also try to
hold inode @i_mutex, so thread3 will wait Thread1 finished.

Thread4 is waiting pending workers finished which will wait
Thread3 finished. So the problem is like this:

Thread1--->Thread4--->Thread3--->Thread2---->Thread1

Deadlock happens! we fix it by letting Thread1 go firstly,
which means we won't block transaction commit while we are
waiting pending workers finished.

Reported-by: Qu Wenruo <quwenruo@cn.fujitsu.com>
Signed-off-by: Wang Shilong <wangsl.fnst@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agoBtrfs: fix a possible deadlock between scrub and transaction committing
Wang Shilong [Wed, 19 Feb 2014 11:24:16 +0000 (19:24 +0800)]
Btrfs: fix a possible deadlock between scrub and transaction committing

btrfs_scrub_continue() will be called when cleaning up transaction.However,
this can only be called if btrfs_scrub_pause() is called before.

Signed-off-by: Wang Shilong <wangsl.fnst@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agobtrfs: Use PTR_ERR_OR_ZERO
Sachin Kamat [Mon, 17 Feb 2014 03:43:57 +0000 (09:13 +0530)]
btrfs: Use PTR_ERR_OR_ZERO

PTR_RET is deprecated. Use PTR_ERR_OR_ZERO instead. While at it
also include missing err.h header.

Signed-off-by: Sachin Kamat <sachin.kamat@linaro.org>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agoBtrfs: fix send issuing outdated paths for utimes, chown and chmod
Filipe Manana [Fri, 21 Feb 2014 00:01:32 +0000 (00:01 +0000)]
Btrfs: fix send issuing outdated paths for utimes, chown and chmod

When doing an incremental send, if we had a directory pending a move/rename
operation and none of its parents, except for the immediate parent, were
pending a move/rename, after processing the directory's references, we would
be issuing utimes, chown and chmod intructions against am outdated path - a
path which matched the one in the parent root.

This change also simplifies a bit the code that deals with building a path
for a directory which has a move/rename operation delayed.

Steps to reproduce:

    $ mkfs.btrfs -f /dev/sdb3
    $ mount /dev/sdb3 /mnt/btrfs
    $ mkdir -p /mnt/btrfs/a/b/c/d/e
    $ mkdir /mnt/btrfs/a/b/c/f
    $ chmod 0777 /mnt/btrfs/a/b/c/d/e
    $ btrfs subvolume snapshot -r /mnt/btrfs /mnt/btrfs/snap1
    $ btrfs send /mnt/btrfs/snap1 -f /tmp/base.send
    $ mv /mnt/btrfs/a/b/c/f /mnt/btrfs/a/b/f2
    $ mv /mnt/btrfs/a/b/c/d/e /mnt/btrfs/a/b/f2/e2
    $ mv /mnt/btrfs/a/b/c /mnt/btrfs/a/b/c2
    $ mv /mnt/btrfs/a/b/c2/d /mnt/btrfs/a/b/c2/d2
    $ chmod 0700 /mnt/btrfs/a/b/f2/e2
    $ btrfs subvolume snapshot -r /mnt/btrfs /mnt/btrfs/snap2
    $ btrfs send -p /mnt/btrfs/snap1 /mnt/btrfs/snap2 -f /tmp/incremental.send

    $ umount /mnt/btrfs
    $ mkfs.btrfs -f /dev/sdb3
    $ mount /dev/sdb3 /mnt/btrfs
    $ btrfs receive /mnt/btrfs -f /tmp/base.send
    $ btrfs receive /mnt/btrfs -f /tmp/incremental.send

The second btrfs receive command failed with:

    ERROR: chmod a/b/c/d/e failed. No such file or directory

A test case for xfstests follows.

Signed-off-by: Filipe David Borba Manana <fdmanana@gmail.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agoBtrfs: correctly determine if blocks are shared in btrfs_compare_trees
Filipe Manana [Thu, 20 Feb 2014 21:15:25 +0000 (21:15 +0000)]
Btrfs: correctly determine if blocks are shared in btrfs_compare_trees

Just comparing the pointers (logical disk addresses) of the btree nodes is
not completely bullet proof, we have to check if their generation numbers
match too.

It is guaranteed that a COW operation will result in a block with a different
logical disk address than the original block's address, but over time we can
reuse that former logical disk address.

For example, creating a 2Gb filesystem on a loop device, and having a script
running in a loop always updating the access timestamp of a file, resulted in
the same logical disk address being reused for the same fs btree block in about
only 4 minutes.

This could make us skip entire subtrees when doing an incremental send (which
is currently the only user of btrfs_compare_trees). However the odds of getting
2 blocks at the same tree level, with the same logical disk address, equal first
slot keys and different generations, should hopefully be very low.

Signed-off-by: Filipe David Borba Manana <fdmanana@gmail.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agoBtrfs: fix send attempting to rmdir non-empty directories
Filipe Manana [Wed, 19 Feb 2014 14:31:44 +0000 (14:31 +0000)]
Btrfs: fix send attempting to rmdir non-empty directories

The incremental send algorithm assumed that it was possible to issue
a directory remove (rmdir) if the the inode number it was currently
processing was greater than (or equal) to any inode that referenced
the directory's inode. This wasn't a valid assumption because any such
inode might be a child directory that is pending a move/rename operation,
because it was moved into a directory that has a higher inode number and
was moved/renamed too - in other words, the case the following commit
addressed:

    9f03740a956d7ac6a1b8f8c455da6fa5cae11c22
    (Btrfs: fix infinite path build loops in incremental send)

This made an incremental send issue an rmdir operation before the
target directory was actually empty, which made btrfs receive fail.
Therefore it needs to wait for all pending child directory inodes to
be moved/renamed before sending an rmdir operation.

Simple steps to reproduce this issue:

    $ mkfs.btrfs -f /dev/sdb3
    $ mount /dev/sdb3 /mnt/btrfs
    $ mkdir -p /mnt/btrfs/a/b/c/x
    $ mkdir /mnt/btrfs/a/b/y
    $ btrfs subvolume snapshot -r /mnt/btrfs /mnt/btrfs/snap1
    $ btrfs send /mnt/btrfs/snap1 -f /tmp/base.send
    $ mv /mnt/btrfs/a/b/y /mnt/btrfs/a/b/YY
    $ mv /mnt/btrfs/a/b/c/x /mnt/btrfs/a/b/YY
    $ rmdir /mnt/btrfs/a/b/c
    $ btrfs subvolume snapshot -r /mnt/btrfs /mnt/btrfs/snap2
    $ btrfs send -p /mnt/btrfs/snap1 /mnt/btrfs/snap2 -f /tmp/incremental.send

    $ umount /mnt/btrfs
    $ mkfs.btrfs -f /dev/sdb3
    $ mount /dev/sdb3 /mnt/btrfs
    $ btrfs receive /mnt/btrfs -f /tmp/base.send
    $ btrfs receive /mnt/btrfs -f /tmp/incremental.send

The second btrfs receive command failed with:

    ERROR: rmdir o259-6-0 failed. Directory not empty

A test case for xfstests follows.

Signed-off-by: Filipe David Borba Manana <fdmanana@gmail.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agoBtrfs: send, don't send rmdir for same target multiple times
Filipe Manana [Sun, 16 Feb 2014 21:01:39 +0000 (21:01 +0000)]
Btrfs: send, don't send rmdir for same target multiple times

When doing an incremental send, if we delete a directory that has N > 1
hardlinks for the same file and that file has the highest inode number
inside the directory contents, an incremental send would send N times an
rmdir operation against the directory. This made the btrfs receive command
fail on the second rmdir instruction, as the target directory didn't exist
anymore.

Steps to reproduce the issue:

    $ mkfs.btrfs -f /dev/sdb3
    $ mount /dev/sdb3 /mnt/btrfs
    $ mkdir -p /mnt/btrfs/a/b/c
    $ echo 'ola mundo' > /mnt/btrfs/a/b/c/foo.txt
    $ ln /mnt/btrfs/a/b/c/foo.txt /mnt/btrfs/a/b/c/bar.txt
    $ btrfs subvolume snapshot -r /mnt/btrfs /mnt/btrfs/snap1
    $ btrfs send /mnt/btrfs/snap1 -f /tmp/base.send
    $ rm -f /mnt/btrfs/a/b/c/foo.txt
    $ rm -f /mnt/btrfs/a/b/c/bar.txt
    $ rmdir /mnt/btrfs/a/b/c
    $ btrfs subvolume snapshot -r /mnt/btrfs /mnt/btrfs/snap2
    $ btrfs send -p /mnt/btrfs/snap1 /mnt/btrfs/snap2 -f /tmp/incremental.send

    $ umount /mnt/btrfs
    $ mkfs.btrfs -f /dev/sdb3
    $ mount /dev/sdb3 /mnt/btrfs
    $ btrfs receive /mnt/btrfs -f /tmp/base.send
    $ btrfs receive /mnt/btrfs -f /tmp/incremental.send

The second btrfs receive command failed with:

    ERROR: rmdir o259-6-0 failed. No such file or directory

A test case for xfstests follows.

Signed-off-by: Filipe David Borba Manana <fdmanana@gmail.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agoBtrfs: incremental send, fix invalid path after dir rename
Filipe Manana [Sun, 16 Feb 2014 13:43:11 +0000 (13:43 +0000)]
Btrfs: incremental send, fix invalid path after dir rename

This fixes yet one more case not caught by the commit titled:

   Btrfs: fix infinite path build loops in incremental send

In this case, even before the initial full send, we have a directory
which is a child of a directory with a higher inode number. Then we
perform the initial send, and after we rename both the child and the
parent, without moving them around. After doing these 2 renames, an
incremental send sent a rename instruction for the child directory
which contained an invalid "from" path (referenced the parent's old
name, not the new one), which made the btrfs receive command fail.

Steps to reproduce:

    $ mkfs.btrfs -f /dev/sdb3
    $ mount /dev/sdb3 /mnt/btrfs
    $ mkdir -p /mnt/btrfs/a/b
    $ mkdir /mnt/btrfs/d
    $ mkdir /mnt/btrfs/a/b/c
    $ mv /mnt/btrfs/d /mnt/btrfs/a/b/c
    $ btrfs subvolume snapshot -r /mnt/btrfs /mnt/btrfs/snap1
    $ btrfs send /mnt/btrfs/snap1 -f /tmp/base.send
    $ mv /mnt/btrfs/a/b/c /mnt/btrfs/a/b/x
    $ mv /mnt/btrfs/a/b/x/d /mnt/btrfs/a/b/x/y
    $ btrfs subvolume snapshot -r /mnt/btrfs /mnt/btrfs/snap2
    $ btrfs send -p /mnt/btrfs/snap1 /mnt/btrfs/snap2 -f /tmp/incremental.send

    $ umout /mnt/btrfs
    $ mkfs.btrfs -f /dev/sdb3
    $ mount /dev/sdb3 /mnt/btrfs
    $ btrfs receive /mnt/btrfs -f /tmp/base.send
    $ btrfs receive /mnt/btrfs -f /tmp/incremental.send

The second btrfs receive command failed with:
  "ERROR: rename a/b/c/d -> a/b/x/y failed. No such file or directory"

A test case for xfstests follows.

Signed-off-by: Filipe David Borba Manana <fdmanana@gmail.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agoBtrfs: don't insert useless holes when punching beyond the inode's size
Filipe Manana [Sat, 15 Feb 2014 15:55:58 +0000 (15:55 +0000)]
Btrfs: don't insert useless holes when punching beyond the inode's size

If we punch beyond the size of an inode, we'll correctly remove any prealloc extents,
but we'll also insert file extent items representing holes (disk bytenr == 0) that start
with a key offset that lies beyond the inode's size and are not contiguous with the last
file extent item.

Example:

  $XFS_IO_PROG -f -c "truncate 118811" $SCRATCH_MNT/foo
  $XFS_IO_PROG -c "fpunch 582007 864596" $SCRATCH_MNT/foo
  $XFS_IO_PROG -c "pwrite -S 0x0d -b 39987 92267 39987" $SCRATCH_MNT/foo

btrfs-debug-tree output:

  item 4 key (257 INODE_ITEM 0) itemoff 15885 itemsize 160
inode generation 6 transid 6 size 132254 block group 0 mode 100600 links 1
  item 5 key (257 INODE_REF 256) itemoff 15872 itemsize 13
inode ref index 2 namelen 3 name: foo
  item 6 key (257 EXTENT_DATA 0) itemoff 15819 itemsize 53
extent data disk byte 0 nr 0 gen 6
extent data offset 0 nr 90112 ram 122880
extent compression 0
  item 7 key (257 EXTENT_DATA 90112) itemoff 15766 itemsize 53
extent data disk byte 12845056 nr 4096 gen 6
extent data offset 0 nr 45056 ram 45056
extent compression 2
  item 8 key (257 EXTENT_DATA 585728) itemoff 15713 itemsize 53
extent data disk byte 0 nr 0 gen 6
extent data offset 0 nr 860160 ram 860160
extent compression 0

The last extent item, which represents a hole, is useless as it lies beyond the inode's
size.

Signed-off-by: Filipe David Borba Manana <fdmanana@gmail.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agoBtrfs: cleanup delayed-ref.c:find_ref_head()
Filipe Manana [Wed, 12 Feb 2014 15:07:53 +0000 (15:07 +0000)]
Btrfs: cleanup delayed-ref.c:find_ref_head()

The argument last wasn't used, all callers supplied a NULL value
for it. Also removed unnecessary intermediate storage of the result
of key comparisons.

Signed-off-by: Filipe David Borba Manana <fdmanana@gmail.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agoBtrfs: remove unnecessary ref heads rb tree search
Filipe Manana [Wed, 12 Feb 2014 15:07:52 +0000 (15:07 +0000)]
Btrfs: remove unnecessary ref heads rb tree search

When we didn't find the exact ref head we were looking for, if
return_bigger != 0 we set a new search key to match either the
next node after the last one we found or the first one in the
ref heads rb tree, and then did another full tree search. For both
cases this ended up being pointless as we would end up returning
an entry we already had before repeating the search.

Signed-off-by: Filipe David Borba Manana <fdmanana@gmail.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agobtrfs: wake up transaction thread upon remount
Justin Maggard [Thu, 20 Feb 2014 16:48:07 +0000 (08:48 -0800)]
btrfs: wake up transaction thread upon remount

Now that we can adjust the commit interval with a remount, we need
to wake up the transaction thread or else he will continue to sleep
until the previous transaction interval has elapsed before waking
up.  So, if we go from a large commit interval to something smaller,
the transaction thread will not wake up until the large interval has
expired.  This also causes the cleaner thread to stay sleeping, since
it gets woken up by the transaction thread.

Fix it by simply waking up the transaction thread during a remount.

Signed-off-by: Justin Maggard <jmaggard10@gmail.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agoBtrfs: stop joining the log transaction if sync log fails
Miao Xie [Thu, 20 Feb 2014 10:08:57 +0000 (18:08 +0800)]
Btrfs: stop joining the log transaction if sync log fails

If the log sync fails, there is something wrong in the log tree, we
should not continue to join the log transaction and log the metadata.
What we should do is to do a full commit.

This patch fixes this problem by setting ->last_trans_log_full_commit
to the current transaction id, it will tell the tasks not to join
the log transaction, and do a full commit.

Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agoBtrfs: just wait or commit our own log sub-transaction
Miao Xie [Thu, 20 Feb 2014 10:08:59 +0000 (18:08 +0800)]
Btrfs: just wait or commit our own log sub-transaction

We might commit the log sub-transaction which didn't contain the metadata we
logged. It was because we didn't record the log transid and just select
the current log sub-transaction to commit, but the right one might be
committed by the other task already. Actually, we needn't do anything
and it is safe that we go back directly in this case.

This patch improves the log sync by the above idea. We record the transid
of the log sub-transaction in which we log the metadata, and the transid
of the log sub-transaction we have committed. If the committed transid
is >= the transid we record when logging the metadata, we just go back.

Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agoBtrfs: fix skipped error handle when log sync failed
Miao Xie [Thu, 20 Feb 2014 10:08:58 +0000 (18:08 +0800)]
Btrfs: fix skipped error handle when log sync failed

It is possible that many tasks sync the log tree at the same time, but
only one task can do the sync work, the others will wait for it. But those
wait tasks didn't get the result of the log sync, and returned 0 when they
ended the wait. It caused those tasks skipped the error handle, and the
serious problem was they told the users the file sync succeeded but in
fact they failed.

This patch fixes this problem by introducing a log context structure,
we insert it into the a global list. When the sync fails, we will set
the error number of every log context in the list, then the waiting tasks
get the error number of the log context and handle the error if need.

Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agoBtrfs: use signed integer instead of unsigned long integer for log transid
Miao Xie [Thu, 20 Feb 2014 10:08:56 +0000 (18:08 +0800)]
Btrfs: use signed integer instead of unsigned long integer for log transid

The log trans id is initialized to be 0 every time we create a log tree,
and the log tree need be re-created after a new transaction is started,
it means the log trans id is unlikely to be a huge number, so we can use
signed integer instead of unsigned long integer to save a bit space.

Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agoBtrfs: remove unnecessary memory barrier in btrfs_sync_log()
Miao Xie [Thu, 20 Feb 2014 10:08:55 +0000 (18:08 +0800)]
Btrfs: remove unnecessary memory barrier in btrfs_sync_log()

Mutex unlock implies certain memory barriers to make sure all the memory
operation completes before the unlock, and the next mutex lock implies memory
barriers to make sure the all the memory happens after the lock. So it is
a full memory barrier(smp_mb), we needn't add memory barriers. Remove them.

Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agoBtrfs: don't start the log transaction if the log tree init fails
Miao Xie [Thu, 20 Feb 2014 10:08:53 +0000 (18:08 +0800)]
Btrfs: don't start the log transaction if the log tree init fails

The old code would start the log transaction even the log tree init
failed, it was unnecessary. Fix it.

Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agoBtrfs: fix the skipped transaction commit during the file sync
Miao Xie [Thu, 20 Feb 2014 10:08:52 +0000 (18:08 +0800)]
Btrfs: fix the skipped transaction commit during the file sync

We may abort the wait earlier if ->last_trans_log_full_commit was set to
the current transaction id, at this case, we need commit the current
transaction instead of the log sub-transaction. But the current code
didn't tell the caller to do it (return 0, not -EAGAIN). Fix it.

Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agoBtrfs: use ACCESS_ONCE to prevent the optimize accesses to ->last_trans_log_full_commit
Miao Xie [Thu, 20 Feb 2014 10:08:51 +0000 (18:08 +0800)]
Btrfs: use ACCESS_ONCE to prevent the optimize accesses to ->last_trans_log_full_commit

->last_trans_log_full_commit may be changed by the other tasks without lock,
so we need prevent the compiler from the optimize access just like
tmp = fs_info->last_trans_log_full_commit
if (tmp == ...)
...

<do something>

if (tmp == ...)
...

In fact, we need get the new value of ->last_trans_log_full_commit during
the second access. Fix it by ACCESS_ONCE().

Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agoBtrfs: avoid warning bomb of btrfs_invalidate_inodes
Liu Bo [Mon, 10 Feb 2014 09:37:25 +0000 (17:37 +0800)]
Btrfs: avoid warning bomb of btrfs_invalidate_inodes

So after transaction is aborted, we need to cleanup inode resources by
calling btrfs_invalidate_inodes(), and btrfs_invalidate_inodes() hopes
roots' refs to be zero in old times and sets a WARN_ON(), however, this
is not always true within cleaning up transaction, so we get to detect
transaction abortion and not warn at all.

Signed-off-by: Liu Bo <bo.li.liu@oracle.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agoBtrfs: fix possible deadlock in btrfs_cleanup_transaction
Liu Bo [Mon, 10 Feb 2014 09:07:16 +0000 (17:07 +0800)]
Btrfs: fix possible deadlock in btrfs_cleanup_transaction

[13654.480669] ======================================================
[13654.480905] [ INFO: possible circular locking dependency detected ]
[13654.481003] 3.12.0+ #4 Tainted: G        W  O
[13654.481060] -------------------------------------------------------
[13654.481060] btrfs-transacti/9347 is trying to acquire lock:
[13654.481060]  (&(&root->ordered_extent_lock)->rlock){+.+...}, at: [<ffffffffa02d30a1>] btrfs_cleanup_transaction+0x271/0x570 [btrfs]
[13654.481060] but task is already holding lock:
[13654.481060]  (&(&fs_info->ordered_root_lock)->rlock){+.+...}, at: [<ffffffffa02d3015>] btrfs_cleanup_transaction+0x1e5/0x570 [btrfs]
[13654.481060] which lock already depends on the new lock.

[13654.481060] the existing dependency chain (in reverse order) is:
[13654.481060] -> #1 (&(&fs_info->ordered_root_lock)->rlock){+.+...}:
[13654.481060]        [<ffffffff810c4103>] lock_acquire+0x93/0x130
[13654.481060]        [<ffffffff81689991>] _raw_spin_lock+0x41/0x50
[13654.481060]        [<ffffffffa02f011b>] __btrfs_add_ordered_extent+0x39b/0x450 [btrfs]
[13654.481060]        [<ffffffffa02f0202>] btrfs_add_ordered_extent+0x32/0x40 [btrfs]
[13654.481060]        [<ffffffffa02df6aa>] run_delalloc_nocow+0x78a/0x9d0 [btrfs]
[13654.481060]        [<ffffffffa02dfc0d>] run_delalloc_range+0x31d/0x390 [btrfs]
[13654.481060]        [<ffffffffa02f7c00>] __extent_writepage+0x310/0x780 [btrfs]
[13654.481060]        [<ffffffffa02f830a>] extent_write_cache_pages.isra.29.constprop.48+0x29a/0x410 [btrfs]
[13654.481060]        [<ffffffffa02f879d>] extent_writepages+0x4d/0x70 [btrfs]
[13654.481060]        [<ffffffffa02d9f68>] btrfs_writepages+0x28/0x30 [btrfs]
[13654.481060]        [<ffffffff8114be91>] do_writepages+0x21/0x50
[13654.481060]        [<ffffffff81140d49>] __filemap_fdatawrite_range+0x59/0x60
[13654.481060]        [<ffffffff81140e13>] filemap_fdatawrite_range+0x13/0x20
[13654.481060]        [<ffffffffa02f1db9>] btrfs_wait_ordered_range+0x49/0x140 [btrfs]
[13654.481060]        [<ffffffffa0318fe2>] __btrfs_write_out_cache+0x682/0x8b0 [btrfs]
[13654.481060]        [<ffffffffa031952d>] btrfs_write_out_cache+0x8d/0xe0 [btrfs]
[13654.481060]        [<ffffffffa02c7083>] btrfs_write_dirty_block_groups+0x593/0x680 [btrfs]
[13654.481060]        [<ffffffffa0345307>] commit_cowonly_roots+0x14b/0x20d [btrfs]
[13654.481060]        [<ffffffffa02d7c1a>] btrfs_commit_transaction+0x43a/0x9d0 [btrfs]
[13654.481060]        [<ffffffffa030061a>] btrfs_create_uuid_tree+0x5a/0x100 [btrfs]
[13654.481060]        [<ffffffffa02d5a8a>] open_ctree+0x21da/0x2210 [btrfs]
[13654.481060]        [<ffffffffa02ab6fe>] btrfs_mount+0x68e/0x870 [btrfs]
[13654.481060]        [<ffffffff811b2409>] mount_fs+0x39/0x1b0
[13654.481060]        [<ffffffff811cd653>] vfs_kern_mount+0x63/0xf0
[13654.481060]        [<ffffffff811cfcce>] do_mount+0x23e/0xa90
[13654.481060]        [<ffffffff811d05a3>] SyS_mount+0x83/0xc0
[13654.481060]        [<ffffffff81692b52>] system_call_fastpath+0x16/0x1b
[13654.481060] -> #0 (&(&root->ordered_extent_lock)->rlock){+.+...}:
[13654.481060]        [<ffffffff810c340a>] __lock_acquire+0x150a/0x1a70
[13654.481060]        [<ffffffff810c4103>] lock_acquire+0x93/0x130
[13654.481060]        [<ffffffff81689991>] _raw_spin_lock+0x41/0x50
[13654.481060]        [<ffffffffa02d30a1>] btrfs_cleanup_transaction+0x271/0x570 [btrfs]
[13654.481060]        [<ffffffffa02d35ce>] transaction_kthread+0x22e/0x270 [btrfs]
[13654.481060]        [<ffffffff81079efa>] kthread+0xea/0xf0
[13654.481060]        [<ffffffff81692aac>] ret_from_fork+0x7c/0xb0
[13654.481060] other info that might help us debug this:

[13654.481060]  Possible unsafe locking scenario:

[13654.481060]        CPU0                    CPU1
[13654.481060]        ----                    ----
[13654.481060]   lock(&(&fs_info->ordered_root_lock)->rlock);
[13654.481060]  lock(&(&root->ordered_extent_lock)->rlock);
[13654.481060]  lock(&(&fs_info->ordered_root_lock)->rlock);
[13654.481060]   lock(&(&root->ordered_extent_lock)->rlock);
[13654.481060]
 *** DEADLOCK ***
[...]

======================================================

btrfs_destroy_all_ordered_extents()
gets &fs_info->ordered_root_lock __BEFORE__ acquiring &root->ordered_extent_lock,
while btrfs_[add,remove]_ordered_extent()
acquires &fs_info->ordered_root_lock __AFTER__ getting &root->ordered_extent_lock.

This patch fixes the above problem.

Signed-off-by: Liu Bo <bo.li.liu@oracle.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agoBtrfs: faster/more efficient insertion of file extent items
Filipe David Borba Manana [Sun, 9 Feb 2014 23:45:12 +0000 (23:45 +0000)]
Btrfs: faster/more efficient insertion of file extent items

This is an extension to my previous commit titled:

  "Btrfs: faster file extent item replace operations"
  (hash 1acae57b161ef1282f565ef907f72aeed0eb71d9)

Instead of inserting the new file extent item if we deleted existing
file extent items covering our target file range, also allow to insert
the new file extent item if we didn't find any existing items to delete
and replace_extent != 0, since in this case our caller would do another
tree search to insert the new file extent item anyway, therefore just
combine the two tree searches into a single one, saving cpu time, reducing
lock contention and reducing btree node/leaf COW operations.

This covers the case where applications keep doing tail append writes to
files, which for example is the case of Apache CouchDB (its database and
view index files are always open with O_APPEND).

Signed-off-by: Filipe David Borba Manana <fdmanana@gmail.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agobtrfs: always choose work from prio_head first
Stanislaw Gruszka [Sat, 8 Feb 2014 22:18:43 +0000 (23:18 +0100)]
btrfs: always choose work from prio_head first

In case we do not refill, we can overwrite cur pointer from prio_head
by one from not prioritized head, what looks as something that was
not intended.

This change make we always take works from prio_head first until it's
not empty.

Signed-off-by: Stanislaw Gruszka <stf_xl@wp.pl>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agoRevert "Btrfs: remove transaction from btrfs send"
Wang Shilong [Sat, 8 Feb 2014 15:46:36 +0000 (23:46 +0800)]
Revert "Btrfs: remove transaction from btrfs send"

This reverts commit 41ce9970a8a6a362ae8df145f7a03d789e9ef9d2.
Previously i was thinking we can use readonly root's commit root
safely while it is not true, readonly root may be cowed with the
following cases.

1.snapshot send root will cow source root.
2.balance,device operations will also cow readonly send root
to relocate.

So i have two ideas to make us safe to use commit root.

-->approach 1:
make it protected by transaction and end transaction properly and we research
next item from root node(see btrfs_search_slot_for_read()).

-->approach 2:
add another counter to local root structure to sync snapshot with send.
and add a global counter to sync send with exclusive device operations.

So with approach 2, send can use commit root safely, because we make sure
send root can not be cowed during send. Unfortunately, it make codes *ugly*
and more complex to maintain.

To make snapshot and send exclusively, device operations and send operation
exclusively with each other is a little confusing for common users.

So why not drop into previous way.

Cc: Josef Bacik <jbacik@fb.com>
Signed-off-by: Wang Shilong <wangsl.fnst@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agoBtrfs: skip readonly root for snapshot-aware defragment
Wang Shilong [Sat, 8 Feb 2014 15:46:35 +0000 (23:46 +0800)]
Btrfs: skip readonly root for snapshot-aware defragment

Btrfs send is assuming readonly root won't change, let's skip readonly root.

Signed-off-by: Wang Shilong <wangsl.fnst@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agoBtrfs: switch to btrfs_previous_extent_item()
Wang Shilong [Thu, 6 Feb 2014 12:02:29 +0000 (20:02 +0800)]
Btrfs: switch to btrfs_previous_extent_item()

Since we have introduced btrfs_previous_extent_item() to search previous
extent item, just switch into it.

Signed-off-by: Wang Shilong <wangsl.fnst@cn.fujitsu.com>
Reviewed-by: Filipe Manana <fdmanana@gmail.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agoBtrfs: skip submitting barrier for missing device
Hidetoshi Seto [Wed, 5 Feb 2014 07:34:38 +0000 (16:34 +0900)]
Btrfs: skip submitting barrier for missing device

I got an error on v3.13:
 BTRFS error (device sdf1) in write_all_supers:3378: errno=-5 IO failure (errors while submitting device barriers.)

how to reproduce:
  > mkfs.btrfs -f -d raid1 /dev/sdf1 /dev/sdf2
  > wipefs -a /dev/sdf2
  > mount -o degraded /dev/sdf1 /mnt
  > btrfs balance start -f -sconvert=single -mconvert=single -dconvert=single /mnt

The reason of the error is that barrier_all_devices() failed to submit
barrier to the missing device.  However it is clear that we cannot do
anything on missing device, and also it is not necessary to care chunks
on the missing device.

This patch stops sending/waiting barrier if device is missing.

Signed-off-by: Hidetoshi Seto <seto.hidetoshi@jp.fujitsu.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agoBtrfs: unlock extent and pages on error in cow_file_range
Josef Bacik [Fri, 7 Feb 2014 17:21:23 +0000 (12:21 -0500)]
Btrfs: unlock extent and pages on error in cow_file_range

When I converted the BUG_ON() for the free_space_cache_inode in cow_file_range I
made it so we just return an error instead of unlocking all of our various
stuff.  This is a mistake and causes us to hang when we run into this.  This
patch fixes this problem.  Thanks,

Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agoBtrfs: balance delayed inode updates
Josef Bacik [Thu, 6 Feb 2014 21:06:06 +0000 (16:06 -0500)]
Btrfs: balance delayed inode updates

While trying to reproduce a delayed ref problem I noticed the box kept falling
over using all 80gb of my ram with btrfs_inode's and btrfs_delayed_node's.
Turns out this is because we only throttle delayed inode updates in
btrfs_dirty_inode, which doesn't actually get called that often, especially when
all you are doing is creating a bunch of files.  So balance delayed inode
updates everytime we create a new inode.  With this patch we no longer use up
all of our ram with delayed inode updates.  Thanks,

Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agobtrfs: add simple debugfs interface
David Sterba [Wed, 5 Feb 2014 14:36:18 +0000 (15:36 +0100)]
btrfs: add simple debugfs interface

Help during debugging to export various interesting infromation and
tunables without the need of extra mount options or ioctls.

Usage:
* declare your variable in sysfs.h, and include where you need it
* define the variable in sysfs.c and make it visible via
  debugfs_create_TYPE

Depends on CONFIG_DEBUG_FS.

Signed-off-by: David Sterba <dsterba@suse.cz>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agobtrfs: send: lower memory requirements in common case
David Sterba [Wed, 5 Feb 2014 15:17:34 +0000 (16:17 +0100)]
btrfs: send: lower memory requirements in common case

The fs_path structure uses an inline buffer and falls back to a chain of
allocations, but vmalloc is not necessary because PATH_MAX fits into
PAGE_SIZE.

The size of fs_path has been reduced to 256 bytes from PAGE_SIZE,
usually 4k. Experimental measurements show that most paths on a single
filesystem do not exceed 200 bytes, and these get stored into the inline
buffer directly, which is now 230 bytes. Longer paths are kmalloced when
needed.

Signed-off-by: David Sterba <dsterba@suse.cz>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agoBtrfs: make some tree searches in send.c more efficient
Filipe David Borba Manana [Wed, 5 Feb 2014 16:48:56 +0000 (16:48 +0000)]
Btrfs: make some tree searches in send.c more efficient

We have this pattern where we do search for a contiguous group of
items in a tree and everytime we find an item, we process it, then
we release our path, increment the offset of the search key, do
another full tree search and repeat these steps until a tree search
can't find more items we're interested in.

Instead of doing these full tree searches after processing each item,
just process the next item/slot in our leaf and don't release the path.
Since all these trees are read only and we always use the commit root
for a search and skip node/leaf locks, we're not affecting concurrency
on the trees.

Signed-off-by: Filipe David Borba Manana <fdmanana@gmail.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agoBtrfs: use right extent item position in send when finding extent clones
Filipe David Borba Manana [Wed, 5 Feb 2014 16:48:55 +0000 (16:48 +0000)]
Btrfs: use right extent item position in send when finding extent clones

This was a leftover from the commit:

   74dd17fbe3d65829e75d84f00a9525b2ace93998
   (Btrfs: fix btrfs send for inline items and compression)

Signed-off-by: Filipe David Borba Manana <fdmanana@gmail.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agobtrfs: send: remove BUG_ON from name_cache_delete
David Sterba [Mon, 3 Feb 2014 18:24:40 +0000 (19:24 +0100)]
btrfs: send: remove BUG_ON from name_cache_delete

If cleaning the name cache fails, we could try to proceed at the cost of
some memory leak. This is not expected to happen often.

Signed-off-by: David Sterba <dsterba@suse.cz>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agobtrfs: send: remove BUG from process_all_refs
David Sterba [Mon, 3 Feb 2014 18:24:19 +0000 (19:24 +0100)]
btrfs: send: remove BUG from process_all_refs

There are only 2 static callers, the BUG would normally be never
reached, but let's be nice.

Signed-off-by: David Sterba <dsterba@suse.cz>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agobtrfs: send: squeeze bitfilelds in fs_path
David Sterba [Mon, 3 Feb 2014 18:23:47 +0000 (19:23 +0100)]
btrfs: send: squeeze bitfilelds in fs_path

We know that buf_len is at most PATH_MAX, 4k, and can merge it with the
reversed member. This saves 3 bytes in favor of inline_buf.

Signed-off-by: David Sterba <dsterba@suse.cz>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agobtrfs: send: remove virtual_mem member from fs_path
David Sterba [Mon, 3 Feb 2014 18:23:33 +0000 (19:23 +0100)]
btrfs: send: remove virtual_mem member from fs_path

We don't need to keep track of that, it's available via is_vmalloc_addr.

Signed-off-by: David Sterba <dsterba@suse.cz>
Signed-off-by: Josef Bacik <jbacik@fb.com>
10 years agobtrfs: send: remove prepared member from fs_path
David Sterba [Mon, 3 Feb 2014 18:23:19 +0000 (19:23 +0100)]
btrfs: send: remove prepared member from fs_path

The member is used only to return value back from
fs_path_prepare_for_add, we can do it locally and save 8 bytes for the
inline_buf path.

Signed-off-by: David Sterba <dsterba@suse.cz>
Signed-off-by: Josef Bacik <jbacik@fb.com>