Commit Graph

456687 Commits

Author SHA1 Message Date
Chris Mason
0f23ae74f5 Revert "Btrfs: device_list_add() should not update list when mounted"
This reverts commit b96de000bc.

This commit is triggering failures to mount by subvolume id in some
configurations.  The main problem is how many different ways this
scanning function is used, both for scanning while mounted and
unmounted.  A proper cleanup is too big for late rcs.

For now, just revert the commit and we'll put a better fix into a later
merge window.

Signed-off-by: Chris Mason <clm@fb.com>
2014-09-18 07:49:05 -07:00
Filipe Manana
125c4cf9f3 Btrfs: set inode's logged_trans/last_log_commit after ranged fsync
When a ranged fsync finishes if there are still extent maps in the modified
list, still set the inode's logged_trans and last_log_commit. This is important
in case an inode is fsync'ed and unlinked in the same transaction, to ensure its
inode ref gets deleted from the log and the respective dentries in its parent
are deleted too from the log (if the parent directory was fsync'ed in the same
transaction).

Instead make btrfs_inode_in_log() return false if the list of modified extent
maps isn't empty.

This is an incremental on top of the v4 version of the patch:

    "Btrfs: fix fsync data loss after a ranged fsync"

which was added to its v5, but didn't make it on time.

Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-09-16 16:12:19 -07:00
Chris Mason
b0d5d10f41 Btrfs: use insert_inode_locked4 for inode creation
Btrfs was inserting inodes into the hash table before we had fully
set the inode up on disk.  This leaves us open to rare races that allow
two different inodes in memory for the same [root, inode] pair.

This patch fixes things by using insert_inode_locked4 to insert an I_NEW
inode and unlock_new_inode when we're ready for the rest of the kernel
to use the inode.

It also makes sure to init the operations pointers on the inode before
going into the error handling paths.

Signed-off-by: Chris Mason <clm@fb.com>
Reported-by: Al Viro <viro@zeniv.linux.org.uk>
2014-09-08 13:56:45 -07:00
Filipe Manana
49dae1bc1c Btrfs: fix fsync data loss after a ranged fsync
While we're doing a full fsync (when the inode has the flag
BTRFS_INODE_NEEDS_FULL_SYNC set) that is ranged too (covers only a
portion of the file), we might have ordered operations that are started
before or while we're logging the inode and that fall outside the fsync
range.

Therefore when a full ranged fsync finishes don't remove every extent
map from the list of modified extent maps - as for some of them, that
fall outside our fsync range, their respective ordered operation hasn't
finished yet, meaning the corresponding file extent item wasn't inserted
into the fs/subvol tree yet and therefore we didn't log it, and we must
let the next fast fsync (one that checks only the modified list) see this
extent map and log a matching file extent item to the log btree and wait
for its ordered operation to finish (if it's still ongoing).

A test case for xfstests follows.

Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-09-08 13:56:43 -07:00
Dan Carpenter
c47ca32d3a Btrfs: kfree()ing ERR_PTRs
The "inherit" in btrfs_ioctl_snap_create_v2() and "vol_args" in
btrfs_ioctl_rm_dev() are ERR_PTRs so we can't call kfree() on them.

These kind of bugs are "One Err Bugs" where there is just one error
label that does everything.  I could set the "inherit = NULL" and keep
the single out label but it ends up being more complicated that way.  It
makes the code simpler to re-order the unwind so it's in the mirror
order of the allocation and introduce some new error labels.

Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-09-08 13:56:42 -07:00
Filipe Manana
dac5705cad Btrfs: fix crash while doing a ranged fsync
While doing a ranged fsync, that is, one whose range doesn't cover the
whole possible file range (0 to LLONG_MAX), we can crash under certain
circumstances with a trace like the following:

[41074.641913] invalid opcode: 0000 [#1] SMP DEBUG_PAGEALLOC
(...)
[41074.642692] CPU: 0 PID: 24580 Comm: fsx Not tainted 3.16.0-fdm-btrfs-next-45+ #1
(...)
[41074.643886] RIP: 0010:[<ffffffffa01ecc99>]  [<ffffffffa01ecc99>] btrfs_ordered_update_i_size+0x279/0x2b0 [btrfs]
(...)
[41074.644919] Stack:
(...)
[41074.644919] Call Trace:
[41074.644919]  [<ffffffffa01db531>] btrfs_truncate_inode_items+0x3f1/0xa10 [btrfs]
[41074.644919]  [<ffffffffa01eb54f>] ? btrfs_get_logged_extents+0x4f/0x80 [btrfs]
[41074.644919]  [<ffffffffa02137a9>] btrfs_log_inode+0x2f9/0x970 [btrfs]
[41074.644919]  [<ffffffff81090875>] ? sched_clock_local+0x25/0xa0
[41074.644919]  [<ffffffff8164a55e>] ? mutex_unlock+0xe/0x10
[41074.644919]  [<ffffffff810af51d>] ? trace_hardirqs_on+0xd/0x10
[41074.644919]  [<ffffffffa0214b4f>] btrfs_log_inode_parent+0x1ef/0x560 [btrfs]
[41074.644919]  [<ffffffff811d0c55>] ? dget_parent+0x5/0x180
[41074.644919]  [<ffffffffa0215d11>] btrfs_log_dentry_safe+0x51/0x80 [btrfs]
[41074.644919]  [<ffffffffa01e2d1a>] btrfs_sync_file+0x1ba/0x3e0 [btrfs]
[41074.644919]  [<ffffffff811eda6b>] vfs_fsync_range+0x1b/0x30
(...)

The necessary conditions that lead to such crash are:

* an incremental fsync (when the inode doesn't have the
  BTRFS_INODE_NEEDS_FULL_SYNC flag set) happened for our file and it logged
  a file extent item ending at offset X;

* the file got the flag BTRFS_INODE_NEEDS_FULL_SYNC set in its inode, due
  to a file truncate operation that reduces the file to a size smaller
  than X;

* a ranged fsync call happens (via an msync for example), with a range that
  doesn't cover the whole file and the end of this range, lets call it Y, is
  smaller than X;

* btrfs_log_inode, sees the flag BTRFS_INODE_NEEDS_FULL_SYNC set and
  calls btrfs_truncate_inode_items() to remove all items from the log
  tree that are associated with our file;

* btrfs_truncate_inode_items() removes all of the inode's items, and the lowest
  file extent item it removed is the one ending at offset X, where X > 0 and
  X > Y - before returning, it calls btrfs_ordered_update_i_size() with an offset
  parameter set to X;

* btrfs_ordered_update_i_size() sees that X is greater then the current ordered
  size (btrfs_inode's disk_i_size) and then it assumes there can't be any ongoing
  ordered operation with a range covering the offset X, calling a BUG_ON() if
  such ordered operation exists. This assumption is made because the disk_i_size
  is only increased after the corresponding file extent item is added to the
  btree (btrfs_finish_ordered_io);

* But because our fsync covers only a limited range, such an ordered extent might
  exist, and our fsync callback (btrfs_sync_file) doesn't wait for such ordered
  extent to finish when calling btrfs_wait_ordered_range();

And then by the time btrfs_ordered_update_i_size() is called, via:

   btrfs_sync_file() ->
       btrfs_log_dentry_safe() ->
           btrfs_log_inode_parent() ->
               btrfs_log_inode() ->
                   btrfs_truncate_inode_items() ->
                       btrfs_ordered_update_i_size()

We hit the BUG_ON(), which could never happen if the fsync range covered the whole
possible file range (0 to LLONG_MAX), as we would wait for all ordered extents to
finish before calling btrfs_truncate_inode_items().

So just don't call btrfs_ordered_update_i_size() if we're removing the inode's items
from a log tree, which isn't supposed to change the in memory inode's disk_i_size.

Issue found while running xfstests/generic/127 (happens very rarely for me), more
specifically via the fsx calls that use memory mapped IO (and issue msync calls).

Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-09-02 16:46:05 -07:00
Filipe Manana
d9f85963e3 Btrfs: fix corruption after write/fsync failure + fsync + log recovery
While writing to a file, in inode.c:cow_file_range() (and same applies to
submit_compressed_extents()), after reserving an extent for the file data,
we create a new extent map for the written range and insert it into the
extent map cache. After that, we create an ordered operation, but if it
fails (due to a transient/temporary-ENOMEM), we return without dropping
that extent map, which points to a reserved extent that is freed when we
return. A subsequent incremental fsync (when the btrfs inode doesn't have
the flag BTRFS_INODE_NEEDS_FULL_SYNC) considers this extent map valid and
logs a file extent item based on that extent map, which points to a disk
extent that doesn't contain valid data - it was freed by us earlier, at this
point it might contain any random/garbage data.

Therefore, if we reach an error condition when cowing a file range after
we added the new extent map to the cache, drop it from the cache before
returning.

Some sequence of steps that lead to this:

    $ mkfs.btrfs -f /dev/sdd
    $ mount -o commit=9999 /dev/sdd /mnt
    $ cd /mnt

    $ xfs_io -f -c "pwrite -S 0x01 -b 4096 0 4096" -c "fsync" foo
    $ xfs_io -c "pwrite -S 0x02 -b 4096 4096 4096"
    $ sync

    $ od -t x1 foo
    0000000 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01
    *
    0010000 02 02 02 02 02 02 02 02 02 02 02 02 02 02 02 02
    *
    0020000

    $ xfs_io -c "pwrite -S 0xa1 -b 4096 0 4096" foo

    # Now this write + fsync fail with -ENOMEM, which was returned by
    # btrfs_add_ordered_extent() in inode.c:cow_file_range().
    $ xfs_io -c "pwrite -S 0xff -b 4096 4096 4096" foo
    $ xfs_io -c "fsync" foo
    fsync: Cannot allocate memory

    # Now do a new write + fsync, which will succeed. Our previous
    # -ENOMEM was a transient/temporary error.
    $ xfs_io -c "pwrite -S 0xee -b 4096 16384 4096" foo
    $ xfs_io -c "fsync" foo

    # Our file content (in page cache) is now:
    $ od -t x1 foo
    0000000 a1 a1 a1 a1 a1 a1 a1 a1 a1 a1 a1 a1 a1 a1 a1 a1
    *
    0010000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff
    *
    0020000 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
    *
    0040000 ee ee ee ee ee ee ee ee ee ee ee ee ee ee ee ee
    *
    0050000

    # Now reboot the machine, and mount the fs, so that fsync log replay
    # takes place.

    # The file content is now weird, in particular the first 8Kb, which
    # do not match our data before nor after the sync command above.
    $ od -t x1 foo
    0000000 ee ee ee ee ee ee ee ee ee ee ee ee ee ee ee ee
    *
    0010000 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01
    *
    0020000 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
    *
    0040000 ee ee ee ee ee ee ee ee ee ee ee ee ee ee ee ee
    *
    0050000

    # In fact these first 4Kb are a duplicate of the last 4kb block.
    # The last write got an extent map/file extent item that points to
    # the same disk extent that we got in the write+fsync that failed
    # with the -ENOMEM error. btrfs-debug-tree and btrfsck allow us to
    # verify that:

    $ btrfs-debug-tree /dev/sdd
    (...)
	item 6 key (257 EXTENT_DATA 0) itemoff 15819 itemsize 53
		extent data disk byte 12582912 nr 8192
		extent data offset 0 nr 8192 ram 8192
	item 7 key (257 EXTENT_DATA 8192) itemoff 15766 itemsize 53
		extent data disk byte 0 nr 0
		extent data offset 0 nr 8192 ram 8192
	item 8 key (257 EXTENT_DATA 16384) itemoff 15713 itemsize 53
		extent data disk byte 12582912 nr 4096
		extent data offset 0 nr 4096 ram 4096

    $ umount /dev/sdd
    $ btrfsck /dev/sdd
    Checking filesystem on /dev/sdd
    UUID: db5e60e1-050d-41e6-8c7f-3d742dea5d8f
    checking extents
    extent item 12582912 has multiple extent items
    ref mismatch on [12582912 4096] extent item 1, found 2
    Backref bytes do not match extent backref, bytenr=12582912, ref bytes=4096, backref bytes=8192
    backpointer mismatch on [12582912 4096]
    Errors found in extent allocation tree or chunk allocation
    checking free space cache
    checking fs roots
    root 5 inode 257 errors 1000, some csum missing
    found 131074 bytes used err is 1
    total csum bytes: 4
    total tree bytes: 131072
    total fs tree bytes: 32768
    total extent tree bytes: 16384
    btree space waste bytes: 123404
    file data blocks allocated: 274432
     referenced 274432
    Btrfs v3.14.1-96-gcc7fd5a-dirty

Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-09-02 16:46:05 -07:00
Chris Mason
e9512d72e8 Btrfs: fix autodefrag with compression
The autodefrag code skips defrag when two extents are adjacent.  But one
big advantage for autodefrag is cutting down on the number of small
extents, even when they are adjacent.  This commit changes it to defrag
all small extents.

Signed-off-by: Chris Mason <clm@fb.com>
2014-08-27 08:45:37 -07:00
Liu Bo
9e0af23764 Btrfs: fix task hang under heavy compressed write
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>
2014-08-24 07:17:02 -07:00
Chris Mason
f6dc45c7a9 Btrfs: fix filemap_flush call in btrfs_file_release
We should only be flushing on close if the file was flagged as needing
it during truncate.  I broke this with my ordered data vs transaction
commit deadlock fix.

Thanks to Miao Xie for catching this.

Signed-off-by: Chris Mason <clm@fb.com>
Reported-by: Miao Xie <miaox@cn.fujitsu.com>
Reported-by: Fengguang Wu <fengguang.wu@intel.com>
2014-08-21 07:55:31 -07:00
Liu Bo
38c1c2e44b Btrfs: fix crash on endio of reading corrupted block
The crash is

------------[ cut here ]------------
kernel BUG at fs/btrfs/extent_io.c:2124!
[...]
Workqueue: btrfs-endio normal_work_helper [btrfs]
RIP: 0010:[<ffffffffa02d6055>]  [<ffffffffa02d6055>] end_bio_extent_readpage+0xb45/0xcd0 [btrfs]

This is in fact a regression.

It is because we forgot to increase @offset properly in reading corrupted block,
so that the @offset remains, and this leads to checksum errors while reading
left blocks queued up in the same bio, and then ends up with hiting the above
BUG_ON.

Reported-by: Chris Murphy <lists@colorremedies.com>
Signed-off-by: Liu Bo <bo.li.liu@oracle.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-21 07:55:30 -07:00
Eric Sandeen
a3c108950d btrfs: fix leak in qgroup_subtree_accounting() error path
Coverity pointed this out; in the newly added
qgroup_subtree_accounting(), if btrfs_find_all_roots()
returns an error, we leak at least the parents pointer,
and possibly the roots pointer, depending on what failure
occurs.

If btrfs_find_all_roots() returns an error, we need to
free up all allocations before we return.  "roots" is
initialized to NULL, so it should be safe to free
it unconditionally (ulist_free() handles that case).

Cc: Mark Fasheh <mfasheh@suse.de>
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Reviewed-by: Mark Fasheh <mfasheh@suse.de>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-21 07:55:29 -07:00
Qu Wenruo
51f395ad40 btrfs: Use right extent length when inserting overlap extent map.
When current btrfs finds that a new extent map is going to be insereted
but failed with -EEXIST, it will try again to insert the extent map
but with the length of sectorsize.
This is OK if we don't enable 'no-holes' feature since all extent space
is continuous, we will not go into the not found->insert routine.

But if we enable 'no-holes' feature, it will make things out of control.
e.g. in 4K sectorsize, we pass the following args to btrfs_get_extent():
btrfs_get_extent() args: start:  27874 len 4100
28672		  27874		28672	27874+4100	32768
                    |-----------------------|
|---------hole--------------------|---------data----------|

1) not found and insert
Since no extent map containing the range, btrfs_get_extent() will go
into the not_found and insert routine, which will try to insert the
extent map (27874, 27847 + 4100).

2) first overlap
But it overlaps with (28672, 32768) extent, so -EEXIST will be returned
by add_extent_mapping().

3) retry but still overlap
After catching the -EEXIST, then btrfs_get_extent() will try insert it
again but with 4K length, which still overlaps, so -EEXIST will be
returned.

This makes the following patch fail to punch hole.
d77815461f btrfs: Avoid trucating page or punching hole in a already existed hole.

This patch will use the right length, which is the (exsisting->start -
em->start) to insert, making the above patch works in 'no-holes' mode.
Also, some small code style problems in above patch is fixed too.

Reported-by: Filipe David Manana <fdmanana@gmail.com>
Signed-off-by: Qu Wenruo <quwenruo@cn.fujitsu.com>
Reviewed-by: Filipe David Manana <fdmanana@suse.com>
Tested-by: Filipe David Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-21 07:55:27 -07:00
Filipe Manana
62e2390e1a Btrfs: clone, don't create invalid hole extent map
When cloning a file that consists of an inline extent, we were creating
an extent map that represents a non-existing trailing hole starting at a
file offset that isn't a multiple of the sector size. This happened because
when processing an inline extent we weren't aligning the extent's length to
the sector size, and therefore incorrectly treating the range
[inline_extent_length; sector_size[ as a hole.

Signed-off-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: Satoru Takeuchi <takeuchi_satoru@jp.fujitsu.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-21 07:55:26 -07:00
Filipe Manana
7064dd5c36 Btrfs: don't monopolize a core when evicting inode
If an inode has a very large number of extent maps, we can spend
a lot of time freeing them, which triggers a soft lockup warning.
Therefore reschedule if we need to when freeing the extent maps
while evicting the inode.

I could trigger this all the time by running xfstests/generic/299 on
a file system with the no-holes feature enabled. That test creates
an inode with 11386677 extent maps.

    $ mkfs.btrfs -f -O no-holes $TEST_DEV
    $ MKFS_OPTIONS="-O no-holes" ./check generic/299
    generic/299 382s ...
    Message from syslogd@debian-vm3 at Aug  7 10:44:29 ...
     kernel:[85304.208017] BUG: soft lockup - CPU#0 stuck for 22s! [umount:25330]
     384s
    Ran: generic/299
    Passed all 1 tests

    $ dmesg
    (...)
    [86304.300017] BUG: soft lockup - CPU#0 stuck for 23s! [umount:25330]
    (...)
    [86304.300036] Call Trace:
    [86304.300036]  [<ffffffff81698ba9>] __slab_free+0x54/0x295
    [86304.300036]  [<ffffffffa02ee9cc>] ? free_extent_map+0x5c/0xb0 [btrfs]
    [86304.300036]  [<ffffffff811a6cd2>] kmem_cache_free+0x282/0x2a0
    [86304.300036]  [<ffffffffa02ee9cc>] free_extent_map+0x5c/0xb0 [btrfs]
    [86304.300036]  [<ffffffffa02e3775>] btrfs_evict_inode+0xd5/0x660 [btrfs]
    [86304.300036]  [<ffffffff811e7c8d>] ? __inode_wait_for_writeback+0x6d/0xc0
    [86304.300036]  [<ffffffff816a389b>] ? _raw_spin_unlock+0x2b/0x40
    [86304.300036]  [<ffffffff811d8cbb>] evict+0xab/0x180
    [86304.300036]  [<ffffffff811d8dce>] dispose_list+0x3e/0x60
    [86304.300036]  [<ffffffff811d9b04>] evict_inodes+0xf4/0x110
    [86304.300036]  [<ffffffff811bd953>] generic_shutdown_super+0x53/0x110
    [86304.300036]  [<ffffffff811bdaa6>] kill_anon_super+0x16/0x30
    [86304.300036]  [<ffffffffa02a78ba>] btrfs_kill_super+0x1a/0xa0 [btrfs]
    [86304.300036]  [<ffffffff811bd3a9>] deactivate_locked_super+0x59/0x80
    [86304.300036]  [<ffffffff811be44e>] deactivate_super+0x4e/0x70
    [86304.300036]  [<ffffffff811dec14>] mntput_no_expire+0x174/0x1f0
    [86304.300036]  [<ffffffff811deab7>] ? mntput_no_expire+0x17/0x1f0
    [86304.300036]  [<ffffffff811e0517>] SyS_umount+0x97/0x100
    (...)

Signed-off-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: Satoru Takeuchi <takeuchi_satoru@jp.fujitsu.com>
Tested-by: Satoru Takeuchi <takeuchi_satoru@jp.fujitsu.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-21 07:55:25 -07:00
Filipe Manana
74121f7cbb Btrfs: fix hole detection during file fsync
The file hole detection logic during a file fsync wasn't correct,
because it didn't look back (in a previous leaf) for the last file
extent item that can be in a leaf to the left of our leaf and that
has a generation lower than the current transaction id. This made it
assume that a hole exists when it really doesn't exist in the file.

Such false positive hole detection happens in the following scenario:

* We have a file that has many file extent items, covering 3 or more
  btree leafs (the first leaf must contain non file extent items too).

* Two ranges of the file are modified, with their extent items being
  located at 2 different leafs and those leafs aren't consecutive.

* When processing the second modified leaf, we weren't checking if
  some file extent item exists that is located in some leaf that is
  between our 2 modified leafs, and therefore assumed the range defined
  between the last file extent item in the first leaf and the first file
  extent item in the second leaf matched a hole.

Fortunately this didn't result in overriding the log with wrong data,
instead it made the last loop in copy_items() attempt to insert a
duplicated key (for a hole file extent item), which makes the file
fsync code return with -EEXIST to file.c:btrfs_sync_file() which in
turn ends up doing a full transaction commit, which is much more
expensive then writing only to the log tree and wait for it to be
durably persisted (as well as the file's modified extents/pages).
Therefore fix the hole detection logic, so that we don't pay the
cost of doing full transaction commits.

I could trigger this issue with the following test for xfstests (which
never fails, either without or with this patch). The last fsync call
results in a full transaction commit, due to the -EEXIST error mentioned
above. I could also observe this behaviour happening frequently when
running xfstests/generic/075 in a loop.

Test:

    _cleanup()
    {
        _cleanup_flakey
        rm -fr $tmp
    }

    # get standard environment, filters and checks
    . ./common/rc
    . ./common/filter
    . ./common/dmflakey

    # real QA test starts here
    _supported_fs btrfs
    _supported_os Linux
    _require_scratch
    _require_dm_flakey
    _need_to_be_root

    rm -f $seqres.full

    # Create a file with many file extent items, each representing a 4Kb extent.
    # These items span 3 btree leaves, of 16Kb each (default mkfs.btrfs leaf size
    # as of btrfs-progs 3.12).
    _scratch_mkfs -l 16384 >/dev/null 2>&1
    _init_flakey
    SAVE_MOUNT_OPTIONS="$MOUNT_OPTIONS"
    MOUNT_OPTIONS="$MOUNT_OPTIONS -o commit=999"
    _mount_flakey

    # First fsync, inode has BTRFS_INODE_NEEDS_FULL_SYNC flag set.
    $XFS_IO_PROG -f -c "pwrite -S 0x01 -b 4096 0 4096" -c "fsync" \
            $SCRATCH_MNT/foo | _filter_xfs_io

    # For any of the following fsync calls, inode doesn't have the flag
    # BTRFS_INODE_NEEDS_FULL_SYNC set.
    for ((i = 1; i <= 500; i++)); do
        OFFSET=$((4096 * i))
        LEN=4096
        $XFS_IO_PROG -c "pwrite -S 0x01 $OFFSET $LEN" -c "fsync" \
                $SCRATCH_MNT/foo | _filter_xfs_io
    done

    # Commit transaction and bump next transaction's id (to 7).
    sync

    # Truncate will set the BTRFS_INODE_NEEDS_FULL_SYNC flag in the btrfs's
    # inode runtime flags.
    $XFS_IO_PROG -c "truncate 2048000" $SCRATCH_MNT/foo

    # Commit transaction and bump next transaction's id (to 8).
    sync

    # Touch 1 extent item from the first leaf and 1 from the last leaf. The leaf
    # in the middle, containing only file extent items, isn't touched. So the
    # next fsync, when calling btrfs_search_forward(), won't visit that middle
    # leaf. First and 3rd leaf have now a generation with value 8, while the
    # middle leaf remains with a generation with value 6.
    $XFS_IO_PROG \
        -c "pwrite -S 0xee -b 4096 0 4096" \
        -c "pwrite -S 0xff -b 4096 2043904 4096" \
        -c "fsync" \
        $SCRATCH_MNT/foo | _filter_xfs_io

    _load_flakey_table $FLAKEY_DROP_WRITES
    md5sum $SCRATCH_MNT/foo | _filter_scratch
    _unmount_flakey

    _load_flakey_table $FLAKEY_ALLOW_WRITES
    # During mount, we'll replay the log created by the fsync above, and the file's
    # md5 digest should be the same we got before the unmount.
    _mount_flakey
    md5sum $SCRATCH_MNT/foo | _filter_scratch
    _unmount_flakey
    MOUNT_OPTIONS="$SAVE_MOUNT_OPTIONS"

    status=0
    exit

Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-21 07:55:24 -07:00
Filipe Manana
5762b5c958 Btrfs: ensure tmpfile inode is always persisted with link count of 0
If we open a file with O_TMPFILE, don't do any further operation on
it (so that the inode item isn't updated) and then force a transaction
commit, we get a persisted inode item with a link count of 1, and not 0
as it should be.

Steps to reproduce it (requires a modern xfs_io with -T support):

    $ mkfs.btrfs -f /dev/sdd
    $ mount -o /dev/sdd /mnt
    $ xfs_io -T /mnt &
    $ sync

Then btrfs-debug-tree shows the inode item with a link count of 1:

    $ btrfs-debug-tree /dev/sdd
    (...)
    fs tree key (FS_TREE ROOT_ITEM 0)
    leaf 29556736 items 4 free space 15851 generation 6 owner 5
    fs uuid f164d01b-1b92-481d-a4e4-435fb0f843d0
    chunk uuid 0e3d0e56-bcca-4a1c-aa5f-cec2c6f4f7a6
    	item 0 key (256 INODE_ITEM 0) itemoff 16123 itemsize 160
		inode generation 3 transid 6 size 0 block group 0 mode 40755 links 1
    	item 1 key (256 INODE_REF 256) itemoff 16111 itemsize 12
    		inode ref index 0 namelen 2 name: ..
    	item 2 key (257 INODE_ITEM 0) itemoff 15951 itemsize 160
    		inode generation 6 transid 6 size 0 block group 0 mode 100600 links 1
    	item 3 key (ORPHAN ORPHAN_ITEM 257) itemoff 15951 itemsize 0
		orphan item
    checksum tree key (CSUM_TREE ROOT_ITEM 0)
    (...)

Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-21 07:55:23 -07:00
Filipe Manana
9c3b306e1c Btrfs: race free update of commit root for ro snapshots
This is a better solution for the problem addressed in the following
commit:

    Btrfs: update commit root on snapshot creation after orphan cleanup
    (3821f34888)

The previous solution wasn't the best because of 2 reasons:

    1) It added another full transaction commit, which is more expensive
       than just swapping the commit root with the root;

    2) If a reboot happened after the first transaction commit (the one
       that creates the snapshot) and before the second transaction commit,
       then we would end up with the same problem if a send using that
       snapshot was requested before the first transaction commit after
       the reboot.

This change addresses those 2 issues. The second issue is addressed by
switching the commit root in the dentry lookup VFS callback, which is
also called by the snapshot/subvol creation ioctl and performs orphan
cleanup if needed. Like the vfs, the ioctl locks the parent inode too,
preventing race issues between a dentry lookup and snapshot creation.

Cc: Alex Lyakas <alex.btrfs@zadarastorage.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-21 07:55:21 -07:00
Liu Bo
87fa3bb078 Btrfs: fix regression of btrfs device replace
Commit 49c6f736f34f901117c20960ebd7d5e60f12fcac(
btrfs: dev replace should replace the sysfs entry) added the missing sysfs entry
in the process of device replace, but didn't take missing devices into account,
so now we have

BUG: unable to handle kernel NULL pointer dereference at 0000000000000088
IP: [<ffffffffa0268551>] btrfs_kobj_rm_device+0x21/0x40 [btrfs]
...

To reproduce it,
1. mkfs.btrfs -f disk1 disk2
2. mkfs.ext4 disk1
3. mount disk2 /mnt -odegraded
4. btrfs replace start -B 1 disk3 /mnt
--------------------------

This fixes the problem.

Reported-by: Chris Murphy <lists@colorremedies.com>
Signed-off-by: Liu Bo <bo.li.liu@oracle.com>
Reviewed-by: Satoru Takeuchi <takeuchi_satoru@jp.fujitsu.com>
Tested-by: Satoru Takeuchi <takeuchi_satoru@jp.fujitsu.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-21 07:55:20 -07:00
Miao Xie
95669976bd Btrfs: don't consider the missing device when allocating new chunks
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>
2014-08-19 08:52:19 -07:00
Miao Xie
7df69d3e94 Btrfs: Fix wrong device size when we are resizing the device
total_bytes of device is just a in-memory variant which is used to record
the size of the device, and it might be changed before we resize a device,
if the resize operation fails, it will be fallbacked. But some code used it
to update on-disk metadata of the device, it would cause the problem that
on-disk metadata of the devices was not consistent. We should use the other
variant named disk_total_bytes to update the on-disk metadata of device,
because that variant is updated only when the resize operation is successful.
Fix it.

Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-19 08:52:18 -07:00
Miao Xie
5d68da3b8e Btrfs: don't write any data into a readonly device when scrub
We should not write data into a readonly device especially seed device when
doing scrub, skip those devices.

Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Reviewed-by: David Sterba <dsterba@suse.cz>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-19 08:52:17 -07:00
Miao Xie
ff61d17c63 Btrfs: Fix the problem that the replace destroys the seed filesystem
The seed filesystem was destroyed by the device replace, the reproduce
method is:
 # mkfs.btrfs -f <dev0>
 # btrfstune -S 1 <dev0>
 # mount <dev0> <mnt>
 # btrfs device add <dev1> <mnt>
 # umount <mnt>
 # mount <dev1> <mnt>
 # btrfs replace start -f <dev0> <dev2> <mnt>
 # umount <mnt>
 # mount <dev0> <mnt>

It is because we erase the super block on the seed device. It is wrong,
we should not change anything on the seed device.

Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Reviewed-by: David Sterba <dsterba@suse.cz>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-19 08:52:16 -07:00
Qu Wenruo
2c91943b50 btrfs: Return right extent when fiemap gives unaligned offset and len.
When page aligned start and len passed to extent_fiemap(), the result is
good, but when start and len is not aligned, e.g. start = 1 and len =
4095 is passed to extent_fiemap(), it returns no extent.

The problem is that start and len is all rounded down which causes the
problem. This patch will round down start and round up (start + len) to
return right extent.

Reported-by: Chandan Rajendra <chandan@linux.vnet.ibm.com>
Signed-off-by: Qu Wenruo <quwenruo@cn.fujitsu.com>
Reviewed-by: David Sterba <dsterba@suse.cz>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-19 08:52:14 -07:00
Wang Shilong
e2eca69dc6 Btrfs: fix wrong extent mapping for DirectIO
btrfs_next_leaf() will use current leaf's last key to search
and then return a bigger one. So it may still return a file extent
item that is smaller than expected value and we will
get an overflow here for @em->len.

This is easy to reproduce for Btrfs Direct writting, it did not
cause any problem, because writting will re-insert right mapping later.

However, by hacking code to make DIO support compression, wrong extent
mapping is kept and it encounter merging failure(EEXIST) quickly.

Fix this problem by looping to find next file extent item that is bigger
than @start or we could not find anything more.

Signed-off-by: Wang Shilong <wangsl.fnst@cn.fujitsu.com>
Reviewed-by: David Sterba <dsterba@suse.cz>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-19 08:52:13 -07:00
Wang Shilong
9a025a0860 Btrfs: fix wrong write range for filemap_fdatawrite_range()
filemap_fdatawrite_range() expect the third arg to be @end
not @len, fix it.

Signed-off-by: Wang Shilong <wangsl.fnst@cn.fujitsu.com>
Reviewed-by: David Sterba <dsterba@suse.cz>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-19 08:52:12 -07:00
Miao Xie
3a7d55c84c Btrfs: fix wrong missing device counter decrease
The missing devices are accounted by its own fs device, for example
the missing devices in seed filesystem will be accounted by the fs device
of the seed filesystem, not by the new filesystem which is based on
the seed filesystem, so when we remove the missing device in the
seed filesystem, we should decrease the counter of its own fs device.
Fix it.

Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-19 08:52:10 -07:00
Miao Xie
69611ac810 Btrfs: fix unzeroed members in fs_devices when creating a fs from seed fs
We forgot to zero some members in fs_devices when we create new fs_devices
from the one of the seed fs. It would cause the problem that we got wrong
chunk profile when allocating chunks. Fix it.

Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-19 08:36:32 -07:00
Anand Jain
77bdae4d13 btrfs: check generation as replace duplicates devid+uuid
When FS in unmounted we need to check generation number as well
since devid+uuid combination could match with the missing replaced
disk when it reappears, and without this patch it might pair with
the replaced disk again.

 device_list_add() function is called in the following threads,
	mount device option
	mount argument
	ioctl BTRFS_IOC_SCAN_DEV (btrfs dev scan)
	ioctl BTRFS_IOC_DEVICES_READY (btrfs dev ready <dev>)
 they have been unit tested to work fine with this patch.

 If the user knows what he is doing and really want to pair with
 replaced disk (which is not a standard operation), then he should
 first clear the kernel btrfs device list in the memory by doing
 the module unload/load and followed with the mount -o device option.

Signed-off-by: Anand Jain <anand.jain@oracle.com>
Signed-off-by: Wang Shilong <wangsl.fnst@cn.fujitsu.com>
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-19 08:36:30 -07:00
Anand Jain
b96de000bc Btrfs: device_list_add() should not update list when mounted
device_list_add() is called when user runs btrfs dev scan, which would add
any btrfs device into the btrfs_fs_devices list.

Now think of a mounted btrfs. And a new device which contains the a SB
from the mounted btrfs devices.

In this situation when user runs btrfs dev scan, the current code would
just replace existing device with the new device.

Which is to note that old device is neither closed nor gracefully
removed from the btrfs.

The FS is still operational with the old bdev however the device name
is the btrfs_device is new which is provided by the btrfs dev scan.

reproducer:

devmgt[1] detach /dev/sdc

replace the missing disk /dev/sdc

btrfs rep start -f 1 /dev/sde /btrfs
Label: none  uuid: 5dc0aaf4-4683-4050-b2d6-5ebe5f5cd120
        Total devices 2 FS bytes used 32.00KiB
        devid    1 size 958.94MiB used 115.88MiB path /dev/sde
        devid    2 size 958.94MiB used 103.88MiB path /dev/sdd

make /dev/sdc to reappear

devmgt attach host2

btrfs dev scan

btrfs fi show -m
Label: none  uuid: 5dc0aaf4-4683-4050-b2d6-5ebe5f5cd120^M
        Total devices 2 FS bytes used 32.00KiB^M
        devid    1 size 958.94MiB used 115.88MiB path /dev/sdc <- Wrong.
        devid    2 size 958.94MiB used 103.88MiB path /dev/sdd

since /dev/sdc has been replaced with /dev/sde, the /dev/sdc shouldn't be
part of the btrfs-fsid when it reappears. If user want it to be part of it
then sys admin should be using btrfs device add instead.

[1] github.com/anajain/devmgt.git

Signed-off-by: Anand Jain <anand.jain@oracle.com>
Signed-off-by: Wang Shilong <wangsl.fnst@cn.fujitsu.com>
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Reviewed-by: Satoru Takeuchi <takeuchi_satoru@jp.fujitsu.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-19 08:36:28 -07:00
chandan
1707e26d6a Btrfs: fill_holes: Fix slot number passed to hole_mergeable() call.
For a non-existent key, btrfs_search_slot() sets path->slots[0] to the slot
where the key could have been present, which in this case would be the slot
containing the extent item which would be the next neighbor of the file range
being punched. The current code passes an incremented path->slots[0] and we
skip to the wrong file extent item. This would mean that we would fail to
merge the "yet to be created" hole with the next neighboring hole (if one
exists). Fix this.

Signed-off-by: Chandan Rajendra <chandan@linux.vnet.ibm.com>
Reviewed-by: Wang Shilong <wangsl.fnst@cn.fujitsu.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-19 08:36:26 -07:00
Miao Xie
7a5c3c9be1 Btrfs: fix put dio bio twice when we submit dio bio fail
The caller of btrfs_submit_direct_hook() will put the original dio bio
when btrfs_submit_direct_hook() return a error number, so we needn't
put the original bio in btrfs_submit_direct_hook().

Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-19 08:36:24 -07:00
Chris Mason
8d875f95da btrfs: disable strict file flushes for renames and truncates
Truncates and renames are often used to replace old versions of a file
with new versions.  Applications often expect this to be an atomic
replacement, even if they haven't done anything to make sure the new
version is fully on disk.

Btrfs has strict flushing in place to make sure that renaming over an
old file with a new file will fully flush out the new file before
allowing the transaction commit with the rename to complete.

This ordering means the commit code needs to be able to lock file pages,
and there are a few paths in the filesystem where we will try to end a
transaction with the page lock held.  It's rare, but these things can
deadlock.

This patch removes the ordered flushes and switches to a best effort
filemap_flush like ext4 uses. It's not perfect, but it should fix the
deadlocks.

Signed-off-by: Chris Mason <clm@fb.com>
2014-08-15 07:43:42 -07:00
Filipe Manana
27b9a8122f Btrfs: fix csum tree corruption, duplicate and outdated checksums
Under rare circumstances we can end up leaving 2 versions of a checksum
for the same file extent range.

The reason for this is that after calling btrfs_next_leaf we process
slot 0 of the leaf it returns, instead of processing the slot set in
path->slots[0]. Most of the time (by far) path->slots[0] is 0, but after
btrfs_next_leaf() releases the path and before it searches for the next
leaf, another task might cause a split of the next leaf, which migrates
some of its keys to the leaf we were processing before calling
btrfs_next_leaf(). In this case btrfs_next_leaf() returns again the
same leaf but with path->slots[0] having a slot number corresponding
to the first new key it got, that is, a slot number that didn't exist
before calling btrfs_next_leaf(), as the leaf now has more keys than
it had before. So we must really process the returned leaf starting at
path->slots[0] always, as it isn't always 0, and the key at slot 0 can
have an offset much lower than our search offset/bytenr.

For example, consider the following scenario, where we have:

sums->bytenr: 40157184, sums->len: 16384, sums end: 40173568
four 4kb file data blocks with offsets 40157184, 40161280, 40165376, 40169472

  Leaf N:

    slot = 0                           slot = btrfs_header_nritems() - 1
  |-------------------------------------------------------------------|
  | [(CSUM CSUM 39239680), size 8] ... [(CSUM CSUM 40116224), size 4] |
  |-------------------------------------------------------------------|

  Leaf N + 1:

      slot = 0                          slot = btrfs_header_nritems() - 1
  |--------------------------------------------------------------------|
  | [(CSUM CSUM 40161280), size 32] ... [((CSUM CSUM 40615936), size 8 |
  |--------------------------------------------------------------------|

Because we are at the last slot of leaf N, we call btrfs_next_leaf() to
find the next highest key, which releases the current path and then searches
for that next key. However after releasing the path and before finding that
next key, the item at slot 0 of leaf N + 1 gets moved to leaf N, due to a call
to ctree.c:push_leaf_left() (via ctree.c:split_leaf()), and therefore
btrfs_next_leaf() will returns us a path again with leaf N but with the slot
pointing to its new last key (CSUM CSUM 40161280). This new version of leaf N
is then:

    slot = 0                        slot = btrfs_header_nritems() - 2  slot = btrfs_header_nritems() - 1
  |----------------------------------------------------------------------------------------------------|
  | [(CSUM CSUM 39239680), size 8] ... [(CSUM CSUM 40116224), size 4]  [(CSUM CSUM 40161280), size 32] |
  |----------------------------------------------------------------------------------------------------|

And incorrecly using slot 0, makes us set next_offset to 39239680 and we jump
into the "insert:" label, which will set tmp to:

    tmp = min((sums->len - total_bytes) >> blocksize_bits,
        (next_offset - file_key.offset) >> blocksize_bits) =
    min((16384 - 0) >> 12, (39239680 - 40157184) >> 12) =
    min(4, (u64)-917504 = 18446744073708634112 >> 12) = 4

and

   ins_size = csum_size * tmp = 4 * 4 = 16 bytes.

In other words, we insert a new csum item in the tree with key
(CSUM_OBJECTID CSUM_KEY 40157184 = sums->bytenr) that contains the checksums
for all the data (4 blocks of 4096 bytes each = sums->len). Which is wrong,
because the item with key (CSUM CSUM 40161280) (the one that was moved from
leaf N + 1 to the end of leaf N) contains the old checksums of the last 12288
bytes of our data and won't get those old checksums removed.

So this leaves us 2 different checksums for 3 4kb blocks of data in the tree,
and breaks the logical rule:

   Key_N+1.offset >= Key_N.offset + length_of_data_its_checksums_cover

An obvious bad effect of this is that a subsequent csum tree lookup to get
the checksum of any of the blocks with logical offset of 40161280, 40165376
or 40169472 (the last 3 4kb blocks of file data), will get the old checksums.

Cc: stable@vger.kernel.org
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-15 07:43:40 -07:00
Takashi Iwai
4eb1f66dce Btrfs: Fix memory corruption by ulist_add_merge() on 32bit arch
We've got bug reports that btrfs crashes when quota is enabled on
32bit kernel, typically with the Oops like below:
 BUG: unable to handle kernel NULL pointer dereference at 00000004
 IP: [<f9234590>] find_parent_nodes+0x360/0x1380 [btrfs]
 *pde = 00000000
 Oops: 0000 [#1] SMP
 CPU: 0 PID: 151 Comm: kworker/u8:2 Tainted: G S      W 3.15.2-1.gd43d97e-default #1
 Workqueue: btrfs-qgroup-rescan normal_work_helper [btrfs]
 task: f1478130 ti: f147c000 task.ti: f147c000
 EIP: 0060:[<f9234590>] EFLAGS: 00010213 CPU: 0
 EIP is at find_parent_nodes+0x360/0x1380 [btrfs]
 EAX: f147dda8 EBX: f147ddb0 ECX: 00000011 EDX: 00000000
 ESI: 00000000 EDI: f147dda4 EBP: f147ddf8 ESP: f147dd38
  DS: 007b ES: 007b FS: 00d8 GS: 00e0 SS: 0068
 CR0: 8005003b CR2: 00000004 CR3: 00bf3000 CR4: 00000690
 Stack:
  00000000 00000000 f147dda4 00000050 00000001 00000000 00000001 00000050
  00000001 00000000 d3059000 00000001 00000022 000000a8 00000000 00000000
  00000000 000000a1 00000000 00000000 00000001 00000000 00000000 11800000
 Call Trace:
  [<f923564d>] __btrfs_find_all_roots+0x9d/0xf0 [btrfs]
  [<f9237bb1>] btrfs_qgroup_rescan_worker+0x401/0x760 [btrfs]
  [<f9206148>] normal_work_helper+0xc8/0x270 [btrfs]
  [<c025e38b>] process_one_work+0x11b/0x390
  [<c025eea1>] worker_thread+0x101/0x340
  [<c026432b>] kthread+0x9b/0xb0
  [<c0712a71>] ret_from_kernel_thread+0x21/0x30
  [<c0264290>] kthread_create_on_node+0x110/0x110

This indicates a NULL corruption in prefs_delayed list.  The further
investigation and bisection pointed that the call of ulist_add_merge()
results in the corruption.

ulist_add_merge() takes u64 as aux and writes a 64bit value into
old_aux.  The callers of this function in backref.c, however, pass a
pointer of a pointer to old_aux.  That is, the function overwrites
64bit value on 32bit pointer.  This caused a NULL in the adjacent
variable, in this case, prefs_delayed.

Here is a quick attempt to band-aid over this: a new function,
ulist_add_merge_ptr() is introduced to pass/store properly a pointer
value instead of u64.  There are still ugly void ** cast remaining
in the callers because void ** cannot be taken implicitly.  But, it's
safer than explicit cast to u64, anyway.

Bugzilla: https://bugzilla.novell.com/show_bug.cgi?id=887046
Cc: <stable@vger.kernel.org> [v3.11+]
Signed-off-by: Takashi Iwai <tiwai@suse.de>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-15 07:43:19 -07:00
Liu Bo
ce62003f69 Btrfs: fix compressed write corruption on enospc
When failing to allocate space for the whole compressed extent, we'll
fallback to uncompressed IO, but we've forgotten to redirty the pages
which belong to this compressed extent, and these 'clean' pages will
simply skip 'submit' part and go to endio directly, at last we got data
corruption as we write nothing.

Signed-off-by: Liu Bo <bo.li.liu@oracle.com>
Tested-By: Martin Steigerwald <martin@lichtvoll.de>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-15 07:43:18 -07:00
Mark Fasheh
f90e579c2b btrfs: correctly handle return from ulist_add
ulist_add() can return '1' on sucess, which qgroup_subtree_accounting()
doesn't take into account. As a result, that value can be bubbled up to
callers, causing an error to be printed. Fix this by only returning the
value of ulist_add() when it indicates an error.

Signed-off-by: Mark Fasheh <mfasheh@suse.de>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-15 07:43:16 -07:00
Mark Fasheh
1152651a08 btrfs: qgroup: account shared subtrees during snapshot delete
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>
2014-08-15 07:43:14 -07:00
Filipe Manana
6f7ff6d783 Btrfs: read lock extent buffer while walking backrefs
Before processing the extent buffer, acquire a read lock on it, so
that we're safe against concurrent updates on the extent buffer.

Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-15 07:43:13 -07:00
Josef Bacik
e339a6b097 Btrfs: __btrfs_mod_ref should always use no_quota
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>
2014-08-15 07:43:11 -07:00
David Sterba
ba7b6e62f4 btrfs: adjust statfs calculations according to raid profiles
This has been discussed in thread:
http://thread.gmane.org/gmane.comp.file-systems.btrfs/32528

and this patch implements this proposal:
http://thread.gmane.org/gmane.comp.file-systems.btrfs/32536

Works fine for "clean" raid profiles where the raid factor correction
does the right job. Otherwise it's pessimistic and may show low space
although there's still some left.

The df nubmers are lightly wrong in case of mixed block groups, but this
is not a major usecase and can be addressed later.

The RAID56 numbers are wrong almost the same way as before and will be
addressed separately.

CC: Hugo Mills <hugo@carfax.org.uk>
CC: cwillu <cwillu@cwillu.com>
CC: Josef Bacik <jbacik@fb.com>
Signed-off-by: David Sterba <dsterba@suse.cz>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-15 07:43:10 -07:00
Linus Torvalds
19583ca584 Linux 3.16 2014-08-03 15:25:02 -07:00
Linus Torvalds
8d71844b51 Merge branch 'timers-urgent-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip
Pull timer fixes from Thomas Gleixner:
 "Two fixes in the timer area:
   - a long-standing lock inversion due to a printk
   - suspend-related hrtimer corruption in sched_clock"

* 'timers-urgent-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
  timer: Fix lock inversion between hrtimer_bases.lock and scheduler locks
  sched_clock: Avoid corrupting hrtimer tree during suspend
2014-08-03 09:58:20 -07:00
Linus Torvalds
3f9c08f7ce Merge branch 'fixes' of git://ftp.arm.linux.org.uk/~rmk/linux-arm
Pull ARM fixes from Russell King:
 "A few fixes for ARM.  Some of these are correctness issues:
   - TLBs must be flushed after the old mappings are removed by the DMA
     mapping code, but before the new mappings are established.
   - An off-by-one entry error in the Keystone LPAE setup code.

  Fixes include:
   - ensuring that the identity mapping for LPAE does not remove the
     kernel image from the identity map.
   - preventing userspace from trapping into kgdb.
   - fixing a preemption issue in the Intel iwmmxt code.
   - fixing a build error with nommu.

  Other changes include:
   - Adding a note about which areas of memory are expected to be
     accessible while the identity mapping tables are in place"

* 'fixes' of git://ftp.arm.linux.org.uk/~rmk/linux-arm:
  ARM: 8124/1: don't enter kgdb when userspace executes a kgdb break instruction
  ARM: idmap: add identity mapping usage note
  ARM: 8115/1: LPAE: reduce damage caused by idmap to virtual memory layout
  ARM: fix alignment of keystone page table fixup
  ARM: 8112/1: only select ARM_PATCH_PHYS_VIRT if MMU is enabled
  ARM: 8100/1: Fix preemption disable in iwmmxt_task_enable()
  ARM: DMA: ensure that old section mappings are flushed from the TLB
2014-08-02 10:57:39 -07:00
Omar Sandoval
6bf755db4d ARM: 8124/1: don't enter kgdb when userspace executes a kgdb break instruction
The kgdb breakpoint hooks (kgdb_brk_fn and kgdb_compiled_brk_fn)
should only be entered when a kgdb break instruction is executed
from the kernel. Otherwise, if kgdb is enabled, a userspace program
can cause the kernel to drop into the debugger by executing either
KGDB_BREAKINST or KGDB_COMPILED_BREAK.

Acked-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: Omar Sandoval <osandov@osandov.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
2014-08-02 15:20:30 +01:00
Russell King
c5cc87fa8d ARM: idmap: add identity mapping usage note
Add a note about the usage of the identity mapping; we do not support
accesses outside of the identity map region and kernel image while a
CPU is using the identity map.  This is because the identity mapping
may overwrite vmalloc space, IO mappings, the vectors pages, etc.

Acked-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
2014-08-02 15:20:26 +01:00
Linus Torvalds
1f2609460c Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs
Pull vfs fixes from Al Viro:
 "This contains a couple of fixes - one is the aio fix from Christoph,
  the other a fallocate() one from Eric"

* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs:
  vfs: fix check for fallocate on active swapfile
  direct-io: fix AIO regression
2014-08-01 18:01:41 -07:00
Linus Torvalds
f88cf230a4 Merge branch 'x86-urgent-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip
Pull x86 fix from Peter Anvin:
 "A single fix to not invoke the espfix code on Xen PV, as it turns out
  to oops the guest when invoked after all.  This patch leaves some
  amount of dead code, in particular unnecessary initialization of the
  espfix stacks when they won't be used, but in the interest of keeping
  the patch minimal that cleanup can wait for the next cycle"

* 'x86-urgent-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
  x86_64/entry/xen: Do not invoke espfix64 on Xen
2014-08-01 17:37:01 -07:00
Linus Torvalds
ecb679fc2c Staging driver bugfixes for 3.16-rc8
Here are some tiny staging driver bugfixes that I've had in my tree for
 the past week that resolve some reported issues.  Nothing major at all,
 but it would be good to get them merged for 3.16-rc8 or -final.
 
 Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
 -----BEGIN PGP SIGNATURE-----
 Version: GnuPG v2
 
 iEYEABECAAYFAlPcBekACgkQMUfUDdst+yljIACggnJMs6Hj7yn4Ml5YFws77lTD
 R6AAnR7zzqrw+5pqOpH+ILZRZohmpwwE
 =SM27
 -----END PGP SIGNATURE-----

Merge tag 'staging-3.16-rc8' of git://git.kernel.org/pub/scm/linux/kernel/git/gregkh/staging

Pull staging driver bugfixes from Greg KH:
 "Here are some tiny staging driver bugfixes that I've had in my tree
  for the past week that resolve some reported issues.  Nothing major at
  all, but it would be good to get them merged for 3.16-rc8 or -final"

* tag 'staging-3.16-rc8' of git://git.kernel.org/pub/scm/linux/kernel/git/gregkh/staging:
  staging: vt6655: Fix disassociated messages every 10 seconds
  staging: vt6655: Fix Warning on boot handle_irq_event_percpu.
  staging: rtl8723au: rtw_resume(): release semaphore before exit on error
  iio:bma180: Missing check for frequency fractional part
  iio:bma180: Fix scale factors to report correct acceleration units
  iio: buffer: Fix demux table creation
2014-08-01 17:16:05 -07:00
Linus Torvalds
818be5894e Fix dm bufio shrinker to properly zero-fill all fields.
Fix race in dm cache that caused improper reporting of the number of
 dirty blocks in the cache.
 -----BEGIN PGP SIGNATURE-----
 Version: GnuPG v1
 
 iQEcBAABAgAGBQJT28K2AAoJEMUj8QotnQNau30IAOAiW683+TnSpq4MXRywUNKl
 gaLtC65Q04KesZ2FtZYaYUOdfF8nsmuyBiSlHl8RLzGWCYJcQQAMdXVEgCZ+qXKX
 OhD85zhySXUamvKfq2wX452kAK6O2eR//Azc3d57uWhGboXTZrqTDc4QLRDJQoAF
 8b4g4r/NCV6fAtHtEfB9JtLyrk1kOUGpHdF2rSFy20IUKs1RPZRZzNYEh5KB9ZWI
 DeoZY6GrqR1bLlOAL3Usd43fYGdgv+Mn1HaR+Xgn7LVl2HpRIts4M8Y54F0xko4F
 T26rJMZlMppolgZBXElrDm8ly6bDPglfU7ymJMleiPEdCLhpLX/jqoPEtgsqGI8=
 =pQe5
 -----END PGP SIGNATURE-----

Merge tag 'dm-3.16-fixes-3' of git://git.kernel.org/pub/scm/linux/kernel/git/device-mapper/linux-dm

Pull device mapper fixes from Mike Snitzer:
 "Fix dm bufio shrinker to properly zero-fill all fields.

  Fix race in dm cache that caused improper reporting of the number of
  dirty blocks in the cache"

* tag 'dm-3.16-fixes-3' of git://git.kernel.org/pub/scm/linux/kernel/git/device-mapper/linux-dm:
  dm cache: fix race affecting dirty block count
  dm bufio: fully initialize shrinker
2014-08-01 12:50:05 -07:00