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: NChris Mason <clm@fb.com>

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: NFilipe Manana <fdmanana@suse.com>
Signed-off-by: NChris Mason <clm@fb.com>

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: NTakashi Iwai <tiwai@suse.de>
Signed-off-by: NChris Mason <clm@fb.com>

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: NLiu Bo <bo.li.liu@oracle.com>
Tested-By: NMartin Steigerwald <martin@lichtvoll.de>
Signed-off-by: NChris Mason <clm@fb.com>

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: NMark Fasheh <mfasheh@suse.de>
Signed-off-by: NChris Mason <clm@fb.com>

During its tree walk, btrfs_drop_snapshot() will skip any shared
subtrees it encounters. This is incorrect when we have qgroups
turned on as those subtrees need to have their contents
accounted. In particular, the case we're concerned with is when
removing our snapshot root leaves the subtree with only one root
reference.

In those cases we need to find the last remaining root and add
each extent in the subtree to the corresponding qgroup exclusive
counts.

This patch implements the shared subtree walk and a new qgroup
operation, BTRFS_QGROUP_OPER_SUB_SUBTREE. When an operation of
this type is encountered during qgroup accounting, we search for
any root references to that extent and in the case that we find
only one reference left, we go ahead and do the math on it's
exclusive counts.
Signed-off-by: NMark Fasheh <mfasheh@suse.de>
Reviewed-by: NJosef Bacik <jbacik@fb.com>
Signed-off-by: NChris Mason <clm@fb.com>

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: NFilipe Manana <fdmanana@suse.com>
Signed-off-by: NChris Mason <clm@fb.com>

Before I extended the no_quota arg to btrfs_dec/inc_ref because I didn't
understand how snapshot delete was using it and assumed that we needed the
quota operations there.  With Mark's work this has turned out to be not the
case, we _always_ need to use no_quota for btrfs_dec/inc_ref, so just drop the
argument and make __btrfs_mod_ref call it's process function with no_quota set
always.  Thanks,
Signed-off-by: NJosef Bacik <jbacik@fb.com>
Signed-off-by: NChris Mason <clm@fb.com>

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: NDavid Sterba <dsterba@suse.cz>
Signed-off-by: NChris Mason <clm@fb.com>

Fix the broken check for calling sys_fallocate() on an active swapfile,
introduced by commit 0790b31b ("fs: disallow all fallocate
operation on active swapfile").
Signed-off-by: NEric Biggers <ebiggers3@gmail.com>
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

The direct-io.c rewrite to use the iov_iter infrastructure stopped updating
the size field in struct dio_submit, and thus rendered the check for
allowing asynchronous completions to always return false.  Fix this by
comparing it to the count of bytes in the iov_iter instead.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reported-by: NTim Chen <tim.c.chen@linux.intel.com>
Tested-by: NTim Chen <tim.c.chen@linux.intel.com>

Correctly assemble the client UUID by OR'ing in the flags rather than
assigning them over the other components.
Reported-by: NHimangi Saraogi <himangi774@gmail.com>
Signed-off-by: NDavid Howells <dhowells@redhat.com>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

Currently umount on symlink blocks following umount:

/vz is separate mount

# ls /vz/ -al | grep test
drwxr-xr-x.  2 root root       4096 Jul 19 01:14 testdir
lrwxrwxrwx.  1 root root         11 Jul 19 01:16 testlink -> /vz/testdir
# umount -l /vz/testlink
umount: /vz/testlink: not mounted (expected)

# lsof /vz
# umount /vz
umount: /vz: device is busy. (unexpected)

In this case mountpoint_last() gets an extra refcount on path->mnt
Signed-off-by: NVasily Averin <vvs@openvz.org>
Acked-by: NIan Kent <raven@themaw.net>
Acked-by: NJeff Layton <jlayton@primarydata.com>
Cc: stable@vger.kernel.org
Signed-off-by: NChristoph Hellwig <hch@lst.de>

The following warnings:

  fs/direct-io.c: In function ‘__blockdev_direct_IO’:
  fs/direct-io.c:1011:12: warning: ‘to’ may be used uninitialized in this function [-Wmaybe-uninitialized]
  fs/direct-io.c:913:16: note: ‘to’ was declared here
  fs/direct-io.c:1011:12: warning: ‘from’ may be used uninitialized in this function [-Wmaybe-uninitialized]
  fs/direct-io.c:913:10: note: ‘from’ was declared here

are false positive because dio_get_page() either fails, or sets both
'from' and 'to'.

Paul Bolle said ...
Maybe it's better to move initializing "to" and "from" out of
dio_get_page(). That _might_ make it easier for both the the reader and
the compiler to understand what's going on. Something like this:

Christoph Hellwig said ...
The fix of moving the code definitively looks nicer, while I think
uninitialized_var is horrible wart that won't get anywhere near my code.

Boaz Harrosh: I agree with Christoph and Paul
Signed-off-by: NBoaz Harrosh <boaz@plexistor.com>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

If a filesystem uses simple_xattr to support user extended attributes,
LTP setxattr01 and xfstests generic/062 fail with "Cannot allocate
memory": simple_xattr_alloc()'s wrap-around test mistakenly excludes
values of zero size.  Fix that off-by-one (but apparently no filesystem
needs them yet).
Signed-off-by: NHugh Dickins <hughd@google.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Jeff Layton <jlayton@poochiereds.net>
Cc: Aristeu Rozanski <aris@redhat.com>
Signed-off-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

Commit 079148b9 ("coredump: factor out the setting of PF_DUMPCORE")
cleaned up the setting of PF_DUMPCORE by removing it from all the
linux_binfmt->core_dump() and moving it to zap_threads().But this ended
up clearing all the previously set flags.  This causes issues during
core generation when tsk->flags is checked again (eg.  for PF_USED_MATH
to dump floating point registers).  Fix this.
Signed-off-by: NSilesh C V <svellattu@mvista.com>
Acked-by: NOleg Nesterov <oleg@redhat.com>
Cc: Mandeep Singh Baines <msb@chromium.org>
Cc: <stable@vger.kernel.org>	[3.10+]
Signed-off-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

Commit 8c7424cf "nfsd4: don't try to encode conflicting owner if low
on space" forgot to free conf->data in nfsd4_encode_lockt and before
sign conf->data to NULL in nfsd4_encode_lock_denied, causing a leak.

Worse, kfree() can be called on an uninitialized pointer in the case of
a succesful lock (or one that fails for a reason other than a conflict).

(Note that lock->lk_denied.ld_owner.data appears it should be zero here,
until you notice that it's one arm of a union the other arm of which is
written to in the succesful case by the

	memcpy(&lock->lk_resp_stateid, &lock_stp->st_stid.sc_stateid,
	                                sizeof(stateid_t));

in nfsd4_lock().  In the 32-bit case this overwrites ld_owner.data.)
Signed-off-by: NKinglong Mee <kinglongmee@gmail.com>
Fixes: 8c7424cf ""nfsd4: don't try to encode conflicting owner if low on space"
Signed-off-by: NJ. Bruce Fields <bfields@redhat.com>

Here some additional changes to set a capability flag so that clients can
detect when it's appropriate to return -ENOSYS from open.

This amends the following commit introduced in 3.14:

  7678ac50  fuse: support clients that don't implement 'open'

However we can only add the flag to 3.15 and later since there was no
protocol version update in 3.14.
Signed-off-by: NMiklos Szeredi <mszeredi@suse.cz>
Cc: <stable@vger.kernel.org> # v3.15+

Default s_time_gran is 1, don't overwrite that if userspace didn't
explicitly specify one.
Signed-off-by: NMiklos Szeredi <mszeredi@suse.cz>
Cc: <stable@vger.kernel.org> # v3.15+

commit 99994cde btrfs: dev delete should remove sysfs entry
added a btrfs_kobj_rm_device, which dereferences device->bdev...
right after we check whether device->bdev might be NULL.

I don't honestly know if it's possible to have a NULL device->bdev
here, but assuming that it is (given the test), we need to move
the kobject removal to be under that test.

(Coverity spotted this)
Signed-off-by: NEric Sandeen <sandeen@redhat.com>
Signed-off-by: NChris Mason <clm@fb.com>

xfstests generic/127 detected this problem.

With commit 7fc34a62, now fsync will only flush
data within the passed range.  This is the cause of the above problem,
-- btrfs's fsync has a stage called 'sync log' which will wait for all the
ordered extents it've recorded to finish.

In xfstests/generic/127, with mixed operations such as truncate, fallocate,
punch hole, and mapwrite, we get some pre-allocated extents, and mapwrite will
mmap, and then msync.  And I find that msync will wait for quite a long time
(about 20s in my case), thanks to ftrace, it turns out that the previous
fallocate calls 'btrfs_wait_ordered_range()' to flush dirty pages, but as the
range of dirty pages may be larger than 'btrfs_wait_ordered_range()' wants,
there can be some ordered extents created but not getting corresponding pages
flushed, then they're left in memory until we fsync which runs into the
stage 'sync log', and fsync will just wait for the system writeback thread
to flush those pages and get ordered extents finished, so the latency is
inevitable.

This adds a flush similar to btrfs_start_ordered_extent() in
btrfs_wait_logged_extents() to fix that.
Reviewed-by: NMiao Xie <miaox@cn.fujitsu.com>
Signed-off-by: NLiu Bo <bo.li.liu@oracle.com>
Signed-off-by: NChris Mason <clm@fb.com>

Cc: cluster-devel@redhat.com
Signed-off-by: NFabian Frederick <fabf@skynet.be>
Signed-off-by: NSteven Whitehouse <swhiteho@redhat.com>

Signed-off-by: NGeert Uytterhoeven <geert@linux-m68k.org>
Cc: cluster-devel@redhat.com
Signed-off-by: NSteven Whitehouse <swhiteho@redhat.com>

This patch removes the GLF_NOCACHE flag from the glocks associated with
flocks. There should be no good reason not to cache glocks for flocks:
they only force the glock to be demoted before they can be reacquired,
which can slow down performance and even cause glock hangs, especially
in cases where the flocks are held in Shared (SH) mode.
Signed-off-by: NBob Peterson <rpeterso@redhat.com>
Signed-off-by: NSteven Whitehouse <swhiteho@redhat.com>

This patch allows flock glocks to use a non-blocking dequeue rather
than dq_wait. It also reverts the previous patch I had posted regarding
dq_wait. The reverted patch isn't necessarily a bad idea, but I decided
this might avoid unforeseen side effects, and was therefore safer.
Signed-off-by: NBob Peterson <rpeterso@redhat.com>
Signed-off-by: NSteven Whitehouse <swhiteho@redhat.com>

kcalloc manages count*sizeof overflow.

Cc: cluster-devel@redhat.com
Signed-off-by: NFabian Frederick <fabf@skynet.be>
Signed-off-by: NSteven Whitehouse <swhiteho@redhat.com>

Normally GFP_KERNEL is ok here, but there is now a rarely used code path
relating to deallocation of unlinked inodes (in certain corner cases)
which if hit at times of memory shortage can cause recursion while
trying to free memory.

One solution would be to try and move the gfs2_glock_get() call so
that it is no longer called while another glock is held, but that
doesn't look at all easy, so GFP_NOFS is the best solution for the
time being.
Signed-off-by: NSteven Whitehouse <swhiteho@redhat.com>

We must not leave items on the LRU list with GLF_LOCK set, since
they can be removed if the glock is brought back into use, which
may then potentially result in a hang, waiting for GLF_LOCK to
clear.

It doesn't happen very often, since it requires a glock that has
not been used for a long time to be brought back into use at the
same moment that the shrinker is part way through disposing of
glocks.

The fix is to set GLF_LOCK at a later time, when we already know
that the other locks can be obtained. Also, we now only release
the lru_lock in case a resched is needed, rather than on every
iteration.
Signed-off-by: NSteven Whitehouse <swhiteho@redhat.com>

Function gfs2_glock_dq_wait is supposed to dequeue a glock and then
wait for the lock to be demoted. The problem is, if this is a shared
lock, its demote will depend on the other holders, which means you
might end up waiting forever because the other process is blocked.
This problem is especially apparent when dealing with nested flocks.
Signed-off-by: NBob Peterson <rpeterso@redhat.com>
Signed-off-by: NSteven Whitehouse <swhiteho@redhat.com>

Commit 1ab6c499 (fs: convert fs shrinkers to new scan/count API)
accidentally removed locking from quota shrinker. Fix it -
dqcache_shrink_scan() should use dq_list_lock to protect the
scan on free_dquots list.

CC: stable@vger.kernel.org
Fixes: 1ab6c499Signed-off-by: NNiu Yawei <yawei.niu@intel.com>
Signed-off-by: NJan Kara <jack@suse.cz>

When quota is on, it is expected that unused quota inodes have a
value of NULLFSINO. The changes to support a separate project quota
in 3.12 broken this rule for non-project quota inode enabled
filesystem, as the code now refuses to write the group quota inode
if neither group or project quotas are enabled. This regression was
introduced by commit d892d586 ("xfs: Start using pquotaino from the
superblock").

In this case, we should be writing NULLFSINO rather than nothing to
ensure that we leave the group quota inode in a valid state while
quotas are enabled.

Failure to do so doesn't cause a current kernel to break - the
separate project quota inodes introduced translation code to always
treat a zero inode as NULLFSINO. This was introduced by commit
01026297 ("xfs: Initialize all quota inodes to be NULLFSINO") with is
also in 3.12 but older kernels do not do this and hence taking a
filesystem back to an older kernel can result in quotas failing
initialisation at mount time. When that happens, we see this in
dmesg:

[ 1649.215390] XFS (sdb): Mounting Filesystem
[ 1649.316894] XFS (sdb): Failed to initialize disk quotas.
[ 1649.316902] XFS (sdb): Ending clean mount

By ensuring that we write NULLFSINO to quota inodes that aren't
active, we avoid this problem. We have to be really careful when
determining if the quota inodes are active or not, because we don't
want to write a NULLFSINO if the quota inodes are active and we
simply aren't updating them.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

The allocation stack switch at xfs_bmapi_allocate() has served it's
purpose, but is no longer a sufficient solution to the stack usage
problem we have in the XFS allocation path.

Whilst the kernel stack size is now 16k, that is not a valid reason
for undoing all our "keep stack usage down" modifications. What it
does allow us to do is have the freedom to refine and perfect the
modifications knowing that if we get it wrong it won't blow up in
our faces - we have a safety net now.

This is important because we still have the issue of older kernels
having smaller stacks and that they are still supported and are
demonstrating a wide range of different stack overflows.  Red Hat
has several open bugs for allocation based stack overflows from
directory modifications and direct IO block allocation and these
problems still need to be solved. If we can solve them upstream,
then distro's won't need to bake their own unique solutions.

To that end, I've observed that every allocation based stack
overflow report has had a specific characteristic - it has happened
during or directly after a bmap btree block split. That event
requires a new block to be allocated to the tree, and so we
effectively stack one allocation stack on top of another, and that's
when we get into trouble.

A further observation is that bmap btree block splits are much rarer
than writeback allocation - over a range of different workloads I've
observed the ratio of bmap btree inserts to splits ranges from 100:1
(xfstests run) to 10000:1 (local VM image server with sparse files
that range in the hundreds of thousands to millions of extents).
Either way, bmap btree split events are much, much rarer than
allocation events.

Finally, we have to move the kswapd state to the allocation workqueue
work when allocation is done on behalf of kswapd. This is proving to
cause significant perturbation in performance under memory pressure
and appears to be generating allocation deadlock warnings under some
workloads, so avoiding the use of a workqueue for the majority of
kswapd writeback allocation will minimise the impact of such
behaviour.

Hence it makes sense to move the stack switch to xfs_btree_split()
and only do it for bmap btree splits. Stack switches during
allocation will be much rarer, so there won't be significant
performacne overhead caused by switching stacks. The worse case
stack from all allocation paths will be split, not just writeback.
And the majority of memory allocations will be done in the correct
context (e.g. kswapd) without causing additional latency, and so we
simplify the memory reclaim interactions between processes,
workqueues and kswapd.

The worst stack I've been able to generate with this patch in place
is 5600 bytes deep. It's very revealing because we exit XFS at:

37)     1768      64   kmem_cache_alloc+0x13b/0x170

about 1800 bytes of stack consumed, and the remaining 3800 bytes
(and 36 functions) is memory reclaim, swap and the IO stack. And
this occurs in the inode allocation from an open(O_CREAT) syscall,
not writeback.

The amount of stack being used is much less than I've previously be
able to generate - fs_mark testing has been able to generate stack
usage of around 7k without too much trouble; with this patch it's
only just getting to 5.5k. This is primarily because the metadata
allocation paths (e.g. directory blocks) are no longer causing
double splits on the same stack, and hence now stack tracing is
showing swapping being the worst stack consumer rather than XFS.

Performance of fs_mark inode create workloads is unchanged.
Performance of fs_mark async fsync workloads is consistently good
with context switches reduced by around 150,000/s (30%).
Performance of dbench, streaming IO and postmark is unchanged.
Allocation deadlock warnings have not been seen on the workloads
that generated them since adding this patch.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

This reverts commit 1f6d6482.

This commit resulted in regressions in performance in low
memory situations where kswapd was doing writeback of delayed
allocation blocks. It resulted in significant parallelism of the
kswapd work and with the special kswapd flags meant that hundreds of
active allocation could dip into kswapd specific memory reserves and
avoid being throttled. This cause a large amount of performance
variation, as well as random OOM-killer invocations that didn't
previously exist.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

As of commit f8567a38 it is now possible to
have put_reqs_available() called from irq context.  While put_reqs_available()
is per cpu, it did not protect itself from interrupts on the same CPU.  This
lead to aio_complete() corrupting the available io requests count when run
under a heavy O_DIRECT workloads as reported by Robert Elliott.  Fix this by
disabling irq updates around the per cpu batch updates of reqs_available.

Many thanks to Robert and folks for testing and tracking this down.
Reported-by: NRobert Elliot <Elliott@hp.com>
Tested-by: NRobert Elliot <Elliott@hp.com>
Signed-off-by: NBenjamin LaHaise <bcrl@kvack.org>
Cc: Jens Axboe <axboe@kernel.dk>, Christoph Hellwig <hch@infradead.org>
Cc: stable@vger.kenel.org

kcalloc manages count*sizeof overflow.
Signed-off-by: NFabian Frederick <fabf@skynet.be>
Signed-off-by: NMiklos Szeredi <mszeredi@suse.cz>

tmp_page to be freed if fuse_write_file_get() returns NULL.
Signed-off-by: NMaxim Patlasov <mpatlasov@parallels.com>
Signed-off-by: NMiklos Szeredi <mszeredi@suse.cz>

Once we've started sending unstable NFS writes, we do not want to
clear pg_moreio, or we may end up sending the very last request as
a stable write if the commit lists are still empty.

Do, however, reset pg_moreio in the case where we end up having to
recoalesce the write if an attempt to use pNFS failed.
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

Cc: Weston Andros Adamson <dros@primarydata.com>
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

Use nfs_lock_and_join_requests to merge all subrequests into the head request -
this cancels and dereferences all subrequests.
Signed-off-by: NWeston Andros Adamson <dros@primarydata.com>
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>

Change nfs_find_and_lock_request so nfs_page_async_flush can handle multiple
requests in a page. There is only one request for a page the first time
nfs_page_async_flush is called, but if a write or commit fails, async_flush
is called again and there may be multiple requests associated with the page.
The solution is to merge all the requests in a page group into a single
request before calling nfs_pageio_add_request.

Rename nfs_find_and_lock_request to nfs_lock_and_join_requests and
change it to first lock all requests for the page, then cancel and merge
all subrequests into the head request.
Signed-off-by: NWeston Andros Adamson <dros@primarydata.com>
Signed-off-by: NTrond Myklebust <trond.myklebust@primarydata.com>