btrfs: balance dirty metadata pages in btrfs_finish_ordered_io

[Problem description and how we fix it]
We should balance dirty metadata pages at the end of btrfs_finish_ordered_io, since a small, unmergeable random write can potentially produce dirty metadata which is multiple times larger than the data itself. For example, a small, unmergeable 4KiB write may produce:

16KiB dirty leaf (and possibly 16KiB dirty node) in subvolume tree 16KiB dirty leaf (and possibly 16KiB dirty node) in checksum tree 16KiB dirty leaf (and possibly 16KiB dirty node) in extent tree

Although we do call balance dirty pages in write side, but in the buffered write path, most metadata are dirtied only after we reach the dirty background limit (which by far only counts dirty data pages) and wakeup the flusher thread. If there are many small, unmergeable random writes spread in a large btree, we'll find a burst of dirty pages exceeds the dirty_bytes limit after we wakeup the flusher thread - which is not what we expect. In our machine, it caused out-of-memory problem since a page cannot be dropped if it is marked dirty.

Someone may worry about we may sleep in btrfs_btree_balance_dirty_nodelay, but since we do btrfs_finish_ordered_io in a separate worker, it will not stop the flusher consuming dirty pages. Also, we use different worker for metadata writeback endio, sleep in btrfs_finish_ordered_io help us throttle the size of dirty metadata pages.

[Reproduce steps]
To reproduce the problem, we need to do 4KiB write randomly spread in a large btree. In our 2GiB RAM machine:
1) Create 4 subvolumes.
2) Run fio on each subvolume:

   [global]
   direct=0
   rw=randwrite
   ioengine=libaio
   bs=4k
   iodepth=16
   numjobs=1
   group_reporting
   size=128G
   runtime=1800
   norandommap
   time_based
   randrepeat=0

3) Take snapshot on each subvolume and repeat fio on existing files.
4) Repeat step (3) until we get large btrees.
In our case, by observing btrfs_root_item->bytes_used, we have 2GiB of metadata in each subvolume tree and 12GiB of metadata in extent tree.
5) Stop all fio, take snapshot again, and wait until all delayed work is completed.
6) Start all fio. Few seconds later we hit OOM when the flusher starts to work.

It can be reproduced even when using nocow write.

Signed-off-by Ethan Lien <[email protected]>

[Problem description and how we fix it]
We should balance dirty metadata pages at the end of btrfs_finish_ordered_io, since a small, unmergeable random write can potentially produce dirty metadata which is multiple times larger than the data itself. For example, a small, unmergeable 4KiB write may produce:

16KiB dirty leaf (and possibly 16KiB dirty node) in subvolume tree 16KiB dirty leaf (and possibly 16KiB dirty node) in checksum tree 16KiB dirty leaf (and possibly 16KiB dirty node) in extent tree

Although we do call balance dirty pages in write side, but in the buffered write path, most metadata are dirtied only after we reach the dirty background limit (which by far only counts dirty data pages) and wakeup the flusher thread. If there are many small, unmergeable random writes spread in a large btree, we'll find a burst of dirty pages exceeds the dirty_bytes limit after we wakeup the flusher thread - which
is not what we expect. In our machine, it caused out-of-memory problem since a page cannot be dropped if it is marked dirty.

Someone may worry about we may sleep in btrfs_btree_balance_dirty_nodelay,
but since we do btrfs_finish_ordered_io in a separate worker, it will not
stop the flusher consuming dirty pages. Also, we use different worker for
metadata writeback endio, sleep in btrfs_finish_ordered_io help us throttle
the size of dirty metadata pages.

[Reproduce steps]

V2:
Replace btrfs_btree_balance_dirty with
btrfs_btree_balance_dirty_nodelay.
Add reproduce steps.

 fs/btrfs/inode.c | 2 ++
 1 file changed, 2 insertions(+)

fs/btrfs/inode.c +0 -0

index 8e604e7071f1..e54547df24ee 100644

On 28 May 2018, at 1:48, Ethan Lien wrote:

It took me a while to trigger, but this actually deadlocks ;) More below.

Quoted 21 lines

[Problem description and how we fix it]
We should balance dirty metadata pages at the end of btrfs_finish_ordered_io, since a small, unmergeable random write can potentially produce dirty metadata which is multiple times larger than the data itself. For example, a small, unmergeable 4KiB write may produce:

16KiB dirty leaf (and possibly 16KiB dirty node) in subvolume tree 16KiB dirty leaf (and possibly 16KiB dirty node) in checksum tree 16KiB dirty leaf (and possibly 16KiB dirty node) in extent tree

Although we do call balance dirty pages in write side, but in the buffered write path, most metadata are dirtied only after we reach the dirty background limit (which by far only counts dirty data pages) and wakeup the flusher thread. If there are many small, unmergeable random writes spread in a large btree, we'll find a burst of dirty pages exceeds the dirty_bytes limit after we wakeup the flusher thread - which
is not what we expect. In our machine, it caused out-of-memory problem since a page cannot be dropped if it is marked dirty.

Quoted 9 lines

Someone may worry about we may sleep in btrfs_btree_balance_dirty_nodelay,
but since we do btrfs_finish_ordered_io in a separate worker, it will not
stop the flusher consuming dirty pages. Also, we use different worker for
metadata writeback endio, sleep in btrfs_finish_ordered_io help us throttle
the size of dirty metadata pages.

In general, slowing down btrfs_finish_ordered_io isn't ideal because it adds latency to places we need to finish quickly. Also, btrfs_finish_ordered_io is used by the free space cache. Even though this happens from its own workqueue, it means completing free space cache writeback may end up waiting on balance_dirty_pages, something like this stack trace:

[..]

Eventually, we have every process in the system waiting on balance_dirty_pages(), and nobody is able to make progress on page writeback.

Quoted 42 lines · On 12.12.2018 15:47 Chris Mason wrote:

On 28 May 2018, at 1:48, Ethan Lien wrote:

It took me a while to trigger, but this actually deadlocks ;) More below.

Quoted 21 lines

[Problem description and how we fix it]
We should balance dirty metadata pages at the end of btrfs_finish_ordered_io, since a small, unmergeable random write can potentially produce dirty metadata which is multiple times larger than the data itself. For example, a small, unmergeable 4KiB write may produce:

16KiB dirty leaf (and possibly 16KiB dirty node) in subvolume tree 16KiB dirty leaf (and possibly 16KiB dirty node) in checksum tree 16KiB dirty leaf (and possibly 16KiB dirty node) in extent tree

Although we do call balance dirty pages in write side, but in the buffered write path, most metadata are dirtied only after we reach the dirty background limit (which by far only counts dirty data pages) and wakeup the flusher thread. If there are many small, unmergeable random writes spread in a large btree, we'll find a burst of dirty pages exceeds the dirty_bytes limit after we wakeup the flusher thread - which
is not what we expect. In our machine, it caused out-of-memory problem since a page cannot be dropped if it is marked dirty.

Quoted 9 lines

Someone may worry about we may sleep in btrfs_btree_balance_dirty_nodelay,
but since we do btrfs_finish_ordered_io in a separate worker, it will not
stop the flusher consuming dirty pages. Also, we use different worker for
metadata writeback endio, sleep in btrfs_finish_ordered_io help us throttle
the size of dirty metadata pages.

In general, slowing down btrfs_finish_ordered_io isn't ideal because it adds latency to places we need to finish quickly. Also, btrfs_finish_ordered_io is used by the free space cache. Even though this happens from its own workqueue, it means completing free space cache writeback may end up waiting on balance_dirty_pages, something like this stack trace:

Quoted 6 lines

[..]

Eventually, we have every process in the system waiting on balance_dirty_pages(), and nobody is able to make progress on page writeback.

I had lockups with this patch as well. If you put e.g. a loop device on top of a btrfs file, loop sets PF_LESS_THROTTLE to avoid a feed back loop causing delays. The task balancing dirty pages in btrfs_finish_ordered_io doesn't have the flag and causes slow-downs. In my case it managed to cause a feedback loop where it queues other btrfs_finish_ordered_io and gets stuck completely.

Quoted 7 lines · Martin Raiber

Quoted 6 lines · On 12.12.2018 15:47 Chris Mason wrote:

On 28 May 2018, at 1:48, Ethan Lien wrote:

Eventually, we have every process in the system waiting on balance_dirty_pages(), and nobody is able to make progress on page writeback.

Quoted 8 lines

I had lockups with this patch as well. If you put e.g. a loop device on
top of a btrfs file, loop sets PF_LESS_THROTTLE to avoid a feed back loop causing delays. The task balancing dirty pages in btrfs_finish_ordered_io doesn't have the flag and causes slow-downs. In
my case it managed to cause a feedback loop where it queues other btrfs_finish_ordered_io and gets stuck completely.

This does not look like the artificial and hard to hit case that's in the original patch. I'm thinking about sending a revert to 4.20-rc6, the
deadlock is IMO worse than OOM.

Quoted 43 lines · Ethan Lien

It took me a while to trigger, but this actually deadlocks ;) More below.

Quoted 22 lines

[Problem description and how we fix it]
We should balance dirty metadata pages at the end of btrfs_finish_ordered_io, since a small, unmergeable random write can potentially produce dirty metadata which is multiple times larger than the data itself. For example, a small, unmergeable 4KiB write may produce:

16KiB dirty leaf (and possibly 16KiB dirty node) in subvolume
tree
16KiB dirty leaf (and possibly 16KiB dirty node) in checksum tree 16KiB dirty leaf (and possibly 16KiB dirty node) in extent tree

Although we do call balance dirty pages in write side, but in the buffered write path, most metadata are dirtied only after we reach the dirty background limit (which by far only counts dirty data pages) and wakeup the flusher thread. If there are many small, unmergeable random writes spread in a large btree, we'll find a burst of dirty pages exceeds the dirty_bytes limit after we wakeup the flusher thread - which
is not what we expect. In our machine, it caused out-of-memory problem since a page cannot be dropped if it is marked dirty.

Quoted 9 lines

Someone may worry about we may sleep in btrfs_btree_balance_dirty_nodelay,
but since we do btrfs_finish_ordered_io in a separate worker, it will not
stop the flusher consuming dirty pages. Also, we use different worker for
metadata writeback endio, sleep in btrfs_finish_ordered_io help us throttle
the size of dirty metadata pages.

In general, slowing down btrfs_finish_ordered_io isn't ideal because it adds latency to places we need to finish quickly. Also, btrfs_finish_ordered_io is used by the free space cache. Even though this happens from its own workqueue, it means completing free space cache writeback may end up waiting on balance_dirty_pages, something like this stack trace:

Quoted 6 lines

[..]

Eventually, we have every process in the system waiting on balance_dirty_pages(), and nobody is able to make progress on paclear page's writebackge writeback.

I had lockups with this patch as well. If you put e.g. a loop device on top of a btrfs file, loop sets PF_LESS_THROTTLE to avoid a feed back loop causing delays. The task balancing dirty pages in btrfs_finish_ordered_io doesn't have the flag and causes slow-downs. In my case it managed to cause a feedback loop where it queues other btrfs_finish_ordered_io and gets stuck completely.

Regards,
Martin Raiber

Martin Raiber 於 2018-12-12 23:22 寫到:

Quoted 50 lines · On 12.12.2018 15:47 Chris Mason wrote:

Quoted 42 lines · Ethan Lien

It took me a while to trigger, but this actually deadlocks ;)  More below.

Quoted 21 lines

[Problem description and how we fix it]
We should balance dirty metadata pages at the end of btrfs_finish_ordered_io, since a small, unmergeable random write can potentially produce dirty metadata which is multiple times larger than the data itself. For example, a small, unmergeable 4KiB write may produce:

    16KiB dirty leaf (and possibly 16KiB dirty node) in subvolume tree 16KiB dirty leaf (and possibly 16KiB dirty node) in checksum tree 16KiB dirty leaf (and possibly 16KiB dirty node) in extent tree

Although we do call balance dirty pages in write side, but in the buffered write path, most metadata are dirtied only after we reach the dirty background limit (which by far only counts dirty data pages) and wakeup the flusher thread. If there are many small, unmergeable random writes spread in a large btree, we'll find a burst of dirty pages exceeds the dirty_bytes limit after we wakeup the flusher thread - which
is not what we expect. In our machine, it caused out-of-memory problem since a page cannot be dropped if it is marked dirty.

Quoted 9 lines

Someone may worry about we may sleep in btrfs_btree_balance_dirty_nodelay,
but since we do btrfs_finish_ordered_io in a separate worker, it will not
stop the flusher consuming dirty pages. Also, we use different worker for
metadata writeback endio, sleep in btrfs_finish_ordered_io help us throttle
the size of dirty metadata pages.

In general, slowing down btrfs_finish_ordered_io isn't ideal because it adds latency to places we need to finish quickly.  Also, btrfs_finish_ordered_io is used by the free space cache.  Even though this happens from its own workqueue, it means completing free space cache writeback may end up waiting on balance_dirty_pages, something like this stack trace:

Quoted 7 lines

[..]

Eventually, we have every process in the system waiting on balance_dirty_pages(), and nobody is able to make progress on paclear page's writebackge
writeback.

Quoted 7 lines

I had lockups with this patch as well. If you put e.g. a loop device on top of a btrfs file, loop sets PF_LESS_THROTTLE to avoid a feed back loop causing delays. The task balancing dirty pages in btrfs_finish_ordered_io doesn't have the flag and causes slow-downs. In my case it managed to cause a feedback loop where it queues other btrfs_finish_ordered_io and gets stuck completely.

The data writepage endio will queue a work for btrfs_finish_ordered_io() in a separate workqueue and clear page's writeback, so throttling in btrfs_finish_ordered_io() should not slow down flusher thread. One suspicious point is while the caller is waiting a range of ordered_extents to complete, they will be blocked until balance_dirty_pages_ratelimited() make some progress, since we finish ordered_extents in btrfs_finish_ordered_io().
Do you have call stack information for stuck processes or using fsync/sync frequently? If this is the case, maybe we should pull this thing out and try balance dirty metadata pages somewhere.

Quoted 9 lines

Quoted 8 lines

Quoted 1 line

I had lockups with this patch as well. If you put e.g. a loop device on top of a btrfs file, loop sets PF_LESS_THROTTLE to avoid a feed back loop causing delays. The task balancing dirty pages in btrfs_finish_ordered_io doesn't have the flag and causes slow-downs. In my case it managed to cause a feedback loop where it queues other btrfs_finish_ordered_io and gets stuck completely.

Quoted 11 lines

The data writepage endio will queue a work for btrfs_finish_ordered_io() in a separate workqueue and clear page's writeback, so throttling in btrfs_finish_ordered_io() should not slow down flusher thread. One suspicious point is while the caller is waiting a range of ordered_extents to complete, they will be blocked until balance_dirty_pages_ratelimited() make some progress, since we finish ordered_extents in btrfs_finish_ordered_io().
Do you have call stack information for stuck processes or using fsync/sync frequently? If this is the case, maybe we should pull this thing out and try balance dirty metadata pages somewhere.

Yeah like,

[875317.071433] Call Trace:
[875317.071438]  ? __schedule+0x306/0x7f0
[875317.071442]  schedule+0x32/0x80
[875317.071447]  btrfs_start_ordered_extent+0xed/0x120
[875317.071450]  ? remove_wait_queue+0x60/0x60
[875317.071454]  btrfs_wait_ordered_range+0xa0/0x100
[875317.071457]  btrfs_sync_file+0x1d6/0x400
[875317.071461]  ? do_fsync+0x38/0x60
[875317.071463]  ? btrfs_fdatawrite_range+0x50/0x50
[875317.071465]  do_fsync+0x38/0x60
[875317.071468]  __x64_sys_fsync+0x10/0x20
[875317.071470]  do_syscall_64+0x55/0x100
[875317.071473]  entry_SYSCALL_64_after_hwframe+0x44/0xa9

so I guess the problem is that the calling balance_dirty_pages causes fsyncs to the same btrfs (via my unusual setup of loop+fuse)? Those fsyncs are deadlocked because they are called indirectly from btrfs_finish_ordered_io... It is a unusal setup, which is why I did not post it to the mailing list initially.

Martin Raiber 於 2018-12-17 22:00 寫到:

Quoted 41 lines

Quoted 12 lines

Quoted 11 lines

Quoted 1 line

I had lockups with this patch as well. If you put e.g. a loop device on
top of a btrfs file, loop sets PF_LESS_THROTTLE to avoid a feed back
loop causing delays. The task balancing dirty pages in btrfs_finish_ordered_io doesn't have the flag and causes slow-downs. In
my case it managed to cause a feedback loop where it queues other btrfs_finish_ordered_io and gets stuck completely.

Quoted 12 lines

The data writepage endio will queue a work for btrfs_finish_ordered_io() in a separate workqueue and clear page's writeback, so throttling in btrfs_finish_ordered_io() should not slow
down flusher thread. One suspicious point is while the caller is waiting a range of ordered_extents to complete, they will be blocked until balance_dirty_pages_ratelimited() make some progress, since we finish ordered_extents in btrfs_finish_ordered_io().
Do you have call stack information for stuck processes or using fsync/sync frequently? If this is the case, maybe we should pull this thing out and try balance dirty metadata pages somewhere.

Yeah like,

[875317.071433] Call Trace:
[875317.071438]  ? __schedule+0x306/0x7f0
[875317.071442]  schedule+0x32/0x80
[875317.071447]  btrfs_start_ordered_extent+0xed/0x120
[875317.071450]  ? remove_wait_queue+0x60/0x60
[875317.071454]  btrfs_wait_ordered_range+0xa0/0x100
[875317.071457]  btrfs_sync_file+0x1d6/0x400
[875317.071461]  ? do_fsync+0x38/0x60
[875317.071463]  ? btrfs_fdatawrite_range+0x50/0x50
[875317.071465]  do_fsync+0x38/0x60
[875317.071468]  __x64_sys_fsync+0x10/0x20
[875317.071470]  do_syscall_64+0x55/0x100
[875317.071473]  entry_SYSCALL_64_after_hwframe+0x44/0xa9

Quoted 6 lines

so I guess the problem is that the calling balance_dirty_pages causes fsyncs to the same btrfs (via my unusual setup of loop+fuse)? Those fsyncs are deadlocked because they are called indirectly from btrfs_finish_ordered_io... It is a unusal setup, which is why I did not
post it to the mailing list initially.

To me this is not like a real deadlock. The fsync call invokes two steps:
(1) flushing dirty data pages, (2) update corresponding metadata to point to those flushed data. Since step1 consume dirty pages and step2 produce more dirty pages, in this patch we leave step1 unchanged and block step2 in btrfs_finish_ordered_io(), which seems reasonable to a OOM fix. The problem is, if there are other processes continually writing new data, the fsync call will need to wait the metadata update for a long time, even its dirty data has been flushed long time ago.

Back to the deadlock problem, what Chris found is really a deadlock, and it can be fixed by adding a check of free space inode.

It took me a while to trigger, but this actually deadlocks ;) More below.

Quoted 21 lines

[Problem description and how we fix it]
We should balance dirty metadata pages at the end of btrfs_finish_ordered_io, since a small, unmergeable random write can potentially produce dirty metadata which is multiple times larger than the data itself. For example, a small, unmergeable 4KiB write may produce:

16KiB dirty leaf (and possibly 16KiB dirty node) in subvolume tree 16KiB dirty leaf (and possibly 16KiB dirty node) in checksum tree 16KiB dirty leaf (and possibly 16KiB dirty node) in extent tree

Although we do call balance dirty pages in write side, but in the buffered write path, most metadata are dirtied only after we reach the dirty background limit (which by far only counts dirty data pages) and wakeup the flusher thread. If there are many small, unmergeable random writes spread in a large btree, we'll find a burst of dirty pages exceeds the dirty_bytes limit after we wakeup the flusher thread - which
is not what we expect. In our machine, it caused out-of-memory problem since a page cannot be dropped if it is marked dirty.

Quoted 9 lines

Someone may worry about we may sleep in btrfs_btree_balance_dirty_nodelay,
but since we do btrfs_finish_ordered_io in a separate worker, it will not
stop the flusher consuming dirty pages. Also, we use different worker for
metadata writeback endio, sleep in btrfs_finish_ordered_io help us throttle
the size of dirty metadata pages.

In general, slowing down btrfs_finish_ordered_io isn't ideal because it adds latency to places we need to finish quickly. Also, btrfs_finish_ordered_io is used by the free space cache. Even though this happens from its own workqueue, it means completing free space cache writeback may end up waiting on balance_dirty_pages, something like this stack trace:

12260 kworker/u96:16+btrfs-freespace-write D

[<0>] balance_dirty_pages+0x6e6/0x7ad
[<0>] balance_dirty_pages_ratelimited+0x6bb/0xa90
[<0>] btrfs_finish_ordered_io+0x3da/0x770
[<0>] normal_work_helper+0x1c5/0x5a0
[<0>] process_one_work+0x1ee/0x5a0
[<0>] worker_thread+0x46/0x3d0
[<0>] kthread+0xf5/0x130
[<0>] ret_from_fork+0x24/0x30

[<0>] 0xffffffffffffffff

Transaction commit will wait on the freespace cache:

838 btrfs-transacti D

[<0>] btrfs_start_ordered_extent+0x154/0x1e0
[<0>] btrfs_wait_ordered_range+0xbd/0x110
[<0>] __btrfs_wait_cache_io+0x49/0x1a0
[<0>] btrfs_write_dirty_block_groups+0x10b/0x3b0
[<0>] commit_cowonly_roots+0x215/0x2b0
[<0>] btrfs_commit_transaction+0x37e/0x910
[<0>] transaction_kthread+0x14d/0x180
[<0>] kthread+0xf5/0x130
[<0>] ret_from_fork+0x24/0x30

[<0>] 0xffffffffffffffff

And then writepages ends up waiting on transaction commit:

9520 kworker/u96:13+flush-btrfs-1 D

[<0>] wait_current_trans+0xac/0xe0
[<0>] start_transaction+0x21b/0x4b0
[<0>] cow_file_range_inline+0x10b/0x6b0
[<0>] cow_file_range.isra.69+0x329/0x4a0
[<0>] run_delalloc_range+0x105/0x3c0
[<0>] writepage_delalloc+0x119/0x180
[<0>] __extent_writepage+0x10c/0x390
[<0>] extent_write_cache_pages+0x26f/0x3d0
[<0>] extent_writepages+0x4f/0x80
[<0>] do_writepages+0x17/0x60
[<0>] __writeback_single_inode+0x59/0x690
[<0>] writeback_sb_inodes+0x291/0x4e0
[<0>] __writeback_inodes_wb+0x87/0xb0
[<0>] wb_writeback+0x3bb/0x500
[<0>] wb_workfn+0x40d/0x610
[<0>] process_one_work+0x1ee/0x5a0
[<0>] worker_thread+0x1e0/0x3d0
[<0>] kthread+0xf5/0x130
[<0>] ret_from_fork+0x24/0x30

[<0>] 0xffffffffffffffff

Eventually, we have every process in the system waiting on balance_dirty_pages(), and nobody is able to make progress on page writeback.

Quoted 2 lines

[Reproduce steps]

[ ... ]

Quoted 18 lines

V2:
Replace btrfs_btree_balance_dirty with
btrfs_btree_balance_dirty_nodelay.
Add reproduce steps.

 fs/btrfs/inode.c | 2 ++
 1 file changed, 2 insertions(+)

fs/btrfs/inode.c +0 -0

index 8e604e7071f1..e54547df24ee 100644
@@ -3158,6 +3158,8 @@ static int btrfs_finish_ordered_io(struct
 btrfs_ordered_extent *ordered_extent)
 	/* once for the tree */
 	btrfs_put_ordered_extent(ordered_extent);
 

fs/btrfs/inode.c static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent) · line 3164

Quoted 4 lines

fs/btrfs/inode.c +2 -0

+	btrfs_btree_balance_dirty_nodelay(fs_info);
+
 	return ret;
 }

The original OOM you describe feels like an MM bug to me, but I'm going to try the repro steps here. Since the freespace cache has its own workqueue, we could fix the deadlock just by wrapping the balance_dirty_pages call in a check for the freespace inode. But, I think we'll get better performance by nudging the fix outside of btrfs_finish_ordered_io. I'll see if I can reproduce.

I haven't been able to trigger the OOM this morning. Ethan, is this something you can still hit on upstream kernels with the balance_dirty_pages() removed?

-chris

Chris Mason 於 2018-12-12 22:47 寫到:

Quoted 119 lines · Ethan Lien

It took me a while to trigger, but this actually deadlocks ;) More below.

Quoted 21 lines

[Problem description and how we fix it]
We should balance dirty metadata pages at the end of btrfs_finish_ordered_io, since a small, unmergeable random write can potentially produce dirty metadata which is multiple times larger than the data itself. For example, a small, unmergeable 4KiB write may produce:

16KiB dirty leaf (and possibly 16KiB dirty node) in subvolume tree 16KiB dirty leaf (and possibly 16KiB dirty node) in checksum tree 16KiB dirty leaf (and possibly 16KiB dirty node) in extent tree

Although we do call balance dirty pages in write side, but in the buffered write path, most metadata are dirtied only after we reach the dirty background limit (which by far only counts dirty data pages) and wakeup the flusher thread. If there are many small, unmergeable random writes spread in a large btree, we'll find a burst of dirty pages exceeds the dirty_bytes limit after we wakeup the flusher thread - which
is not what we expect. In our machine, it caused out-of-memory problem since a page cannot be dropped if it is marked dirty.

Quoted 9 lines

Someone may worry about we may sleep in btrfs_btree_balance_dirty_nodelay,
but since we do btrfs_finish_ordered_io in a separate worker, it will not
stop the flusher consuming dirty pages. Also, we use different worker for
metadata writeback endio, sleep in btrfs_finish_ordered_io help us throttle
the size of dirty metadata pages.

In general, slowing down btrfs_finish_ordered_io isn't ideal because it adds latency to places we need to finish quickly. Also, btrfs_finish_ordered_io is used by the free space cache. Even though this happens from its own workqueue, it means completing free space cache writeback may end up waiting on balance_dirty_pages, something like this stack trace:

12260 kworker/u96:16+btrfs-freespace-write D

[<0>] balance_dirty_pages+0x6e6/0x7ad
[<0>] balance_dirty_pages_ratelimited+0x6bb/0xa90
[<0>] btrfs_finish_ordered_io+0x3da/0x770
[<0>] normal_work_helper+0x1c5/0x5a0
[<0>] process_one_work+0x1ee/0x5a0
[<0>] worker_thread+0x46/0x3d0
[<0>] kthread+0xf5/0x130
[<0>] ret_from_fork+0x24/0x30

[<0>] 0xffffffffffffffff

Transaction commit will wait on the freespace cache:

838 btrfs-transacti D

[<0>] btrfs_start_ordered_extent+0x154/0x1e0
[<0>] btrfs_wait_ordered_range+0xbd/0x110
[<0>] __btrfs_wait_cache_io+0x49/0x1a0
[<0>] btrfs_write_dirty_block_groups+0x10b/0x3b0
[<0>] commit_cowonly_roots+0x215/0x2b0
[<0>] btrfs_commit_transaction+0x37e/0x910
[<0>] transaction_kthread+0x14d/0x180
[<0>] kthread+0xf5/0x130
[<0>] ret_from_fork+0x24/0x30

[<0>] 0xffffffffffffffff

And then writepages ends up waiting on transaction commit:

9520 kworker/u96:13+flush-btrfs-1 D

[<0>] wait_current_trans+0xac/0xe0
[<0>] start_transaction+0x21b/0x4b0
[<0>] cow_file_range_inline+0x10b/0x6b0
[<0>] cow_file_range.isra.69+0x329/0x4a0
[<0>] run_delalloc_range+0x105/0x3c0
[<0>] writepage_delalloc+0x119/0x180
[<0>] __extent_writepage+0x10c/0x390
[<0>] extent_write_cache_pages+0x26f/0x3d0
[<0>] extent_writepages+0x4f/0x80
[<0>] do_writepages+0x17/0x60
[<0>] __writeback_single_inode+0x59/0x690
[<0>] writeback_sb_inodes+0x291/0x4e0
[<0>] __writeback_inodes_wb+0x87/0xb0
[<0>] wb_writeback+0x3bb/0x500
[<0>] wb_workfn+0x40d/0x610
[<0>] process_one_work+0x1ee/0x5a0
[<0>] worker_thread+0x1e0/0x3d0
[<0>] kthread+0xf5/0x130
[<0>] ret_from_fork+0x24/0x30

[<0>] 0xffffffffffffffff

Eventually, we have every process in the system waiting on balance_dirty_pages(), and nobody is able to make progress on page writeback.

Quoted 2 lines

[Reproduce steps]

[ ... ]

Quoted 18 lines

V2:
Replace btrfs_btree_balance_dirty with
btrfs_btree_balance_dirty_nodelay.
Add reproduce steps.

 fs/btrfs/inode.c | 2 ++
 1 file changed, 2 insertions(+)

fs/btrfs/inode.c +0 -0

index 8e604e7071f1..e54547df24ee 100644
@@ -3158,6 +3158,8 @@ static int btrfs_finish_ordered_io(struct
 btrfs_ordered_extent *ordered_extent)
 	/* once for the tree */
 	btrfs_put_ordered_extent(ordered_extent);
 

Quoted 12 lines

fs/btrfs/inode.c static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent) · line 3164

Quoted 4 lines

fs/btrfs/inode.c +2 -0

+	btrfs_btree_balance_dirty_nodelay(fs_info);
+
 	return ret;
 }

The original OOM you describe feels like an MM bug to me, but I'm going to try the repro steps here. Since the freespace cache has its own workqueue, we could fix the deadlock just by wrapping the balance_dirty_pages call in a check for the freespace inode. But, I think we'll get better performance by nudging the fix outside of btrfs_finish_ordered_io. I'll see if I can reproduce.

Before this patch, I tried adding a new workqueue for metadata writeback,
and kick off async flush work on fs_info->btree_inode in btrfs_finish_ordered_io(). It works, but it can't guarantee we control dirty
pages under MM's dirty_bytes limit if btrfs_finish_ordered_io() still running
at full speed.

Quoted 3 lines

I haven't been able to trigger the OOM this morning. Ethan, is this something you can still hit on upstream kernels with the balance_dirty_pages() removed?

I hit the OOM problem in 4.4 kernel. I'll try if I can trigger it in uptodate kernel.

Quoted 146 lines

Chris Mason 於 2018-12-12 22:47 寫到:

Quoted 119 lines · Ethan Lien

It took me a while to trigger, but this actually deadlocks ;) More below.

Quoted 21 lines

[Problem description and how we fix it]
We should balance dirty metadata pages at the end of btrfs_finish_ordered_io, since a small, unmergeable random write can potentially produce dirty metadata which is multiple times larger than the data itself. For example, a small, unmergeable 4KiB write may produce:

16KiB dirty leaf (and possibly 16KiB dirty node) in subvolume tree 16KiB dirty leaf (and possibly 16KiB dirty node) in checksum tree 16KiB dirty leaf (and possibly 16KiB dirty node) in extent tree

Although we do call balance dirty pages in write side, but in the buffered write path, most metadata are dirtied only after we reach the dirty background limit (which by far only counts dirty data pages) and wakeup the flusher thread. If there are many small, unmergeable random writes spread in a large btree, we'll find a burst of dirty pages exceeds the dirty_bytes limit after we wakeup the flusher thread - which
is not what we expect. In our machine, it caused out-of-memory problem since a page cannot be dropped if it is marked dirty.

Quoted 9 lines

Someone may worry about we may sleep in btrfs_btree_balance_dirty_nodelay,
but since we do btrfs_finish_ordered_io in a separate worker, it will not
stop the flusher consuming dirty pages. Also, we use different worker for
metadata writeback endio, sleep in btrfs_finish_ordered_io help us throttle
the size of dirty metadata pages.

In general, slowing down btrfs_finish_ordered_io isn't ideal because it adds latency to places we need to finish quickly. Also, btrfs_finish_ordered_io is used by the free space cache. Even though this happens from its own workqueue, it means completing free space cache writeback may end up waiting on balance_dirty_pages, something like this stack trace:

12260 kworker/u96:16+btrfs-freespace-write D

[<0>] balance_dirty_pages+0x6e6/0x7ad
[<0>] balance_dirty_pages_ratelimited+0x6bb/0xa90
[<0>] btrfs_finish_ordered_io+0x3da/0x770
[<0>] normal_work_helper+0x1c5/0x5a0
[<0>] process_one_work+0x1ee/0x5a0
[<0>] worker_thread+0x46/0x3d0
[<0>] kthread+0xf5/0x130
[<0>] ret_from_fork+0x24/0x30

[<0>] 0xffffffffffffffff

Transaction commit will wait on the freespace cache:

838 btrfs-transacti D

[<0>] btrfs_start_ordered_extent+0x154/0x1e0
[<0>] btrfs_wait_ordered_range+0xbd/0x110
[<0>] __btrfs_wait_cache_io+0x49/0x1a0
[<0>] btrfs_write_dirty_block_groups+0x10b/0x3b0
[<0>] commit_cowonly_roots+0x215/0x2b0
[<0>] btrfs_commit_transaction+0x37e/0x910
[<0>] transaction_kthread+0x14d/0x180
[<0>] kthread+0xf5/0x130
[<0>] ret_from_fork+0x24/0x30

[<0>] 0xffffffffffffffff

And then writepages ends up waiting on transaction commit:

9520 kworker/u96:13+flush-btrfs-1 D

[<0>] wait_current_trans+0xac/0xe0
[<0>] start_transaction+0x21b/0x4b0
[<0>] cow_file_range_inline+0x10b/0x6b0
[<0>] cow_file_range.isra.69+0x329/0x4a0
[<0>] run_delalloc_range+0x105/0x3c0
[<0>] writepage_delalloc+0x119/0x180
[<0>] __extent_writepage+0x10c/0x390
[<0>] extent_write_cache_pages+0x26f/0x3d0
[<0>] extent_writepages+0x4f/0x80
[<0>] do_writepages+0x17/0x60
[<0>] __writeback_single_inode+0x59/0x690
[<0>] writeback_sb_inodes+0x291/0x4e0
[<0>] __writeback_inodes_wb+0x87/0xb0
[<0>] wb_writeback+0x3bb/0x500
[<0>] wb_workfn+0x40d/0x610
[<0>] process_one_work+0x1ee/0x5a0
[<0>] worker_thread+0x1e0/0x3d0
[<0>] kthread+0xf5/0x130
[<0>] ret_from_fork+0x24/0x30

[<0>] 0xffffffffffffffff

Eventually, we have every process in the system waiting on balance_dirty_pages(), and nobody is able to make progress on page writeback.

Quoted 2 lines

[Reproduce steps]

[ ... ]

Quoted 18 lines

V2:
Replace btrfs_btree_balance_dirty with
btrfs_btree_balance_dirty_nodelay.
Add reproduce steps.

 fs/btrfs/inode.c | 2 ++
 1 file changed, 2 insertions(+)

fs/btrfs/inode.c +0 -0

index 8e604e7071f1..e54547df24ee 100644
@@ -3158,6 +3158,8 @@ static int btrfs_finish_ordered_io(struct
 btrfs_ordered_extent *ordered_extent)
 	/* once for the tree */
 	btrfs_put_ordered_extent(ordered_extent);
 

Quoted 12 lines

fs/btrfs/inode.c static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent) · line 3164

Quoted 4 lines

fs/btrfs/inode.c +2 -0

+	btrfs_btree_balance_dirty_nodelay(fs_info);
+
 	return ret;
 }

The original OOM you describe feels like an MM bug to me, but I'm going to try the repro steps here. Since the freespace cache has its own workqueue, we could fix the deadlock just by wrapping the balance_dirty_pages call in a check for the freespace inode. But, I think we'll get better performance by nudging the fix outside of btrfs_finish_ordered_io. I'll see if I can reproduce.

Before this patch, I tried adding a new workqueue for metadata writeback,
and kick off async flush work on fs_info->btree_inode in btrfs_finish_ordered_io(). It works, but it can't guarantee we control dirty
pages under MM's dirty_bytes limit if btrfs_finish_ordered_io() still running
at full speed.

Quoted 3 lines

I haven't been able to trigger the OOM this morning. Ethan, is this something you can still hit on upstream kernels with the balance_dirty_pages() removed?

Quoted 2 lines

I hit the OOM problem in 4.4 kernel. I'll try if I can trigger it in uptodate kernel.

Any followup to your testing? Otherwise I'm going to add revert of the patch.