On 26.03.19 г. 17:09 ч., Zygo Blaxell wrote:
> On Tue, Mar 26, 2019 at 10:42:31AM +0200, Nikolay Borisov wrote:
>>
>>
>> On 26.03.19 г. 6:30 ч., Zygo Blaxell wrote:
>>> On Mon, Mar 25, 2019 at 10:50:28PM -0400, Zygo Blaxell wrote:
>>>> Running balance, rsync, and dedupe, I get kernel warnings every few
>>>> minutes on 5.0.4. No warnings on 5.0.3 under similar conditions.
>>>>
>>>> Mount options are: flushoncommit,space_cache=v2,compress=zstd.
>>>>
>>>> There are two different stacks on the warnings. This one comes from
>>>> btrfs balance:
>>>
>>> [snip]
>>>
>>> Possibly unrelated, but I'm also repeatably getting this in 5.0.4 and
>>> not 5.0.3, after about 5 hours of uptime. Different processes, same
>>> kernel stack:
>>>
>>> [Mon Mar 25 23:35:17 2019] kworker/u8:4: page allocation failure:
>>> order:0, mode:0x404000(GFP_NOWAIT|__GFP_COMP),
>>> nodemask=(null),cpuset=/,mems_allowed=0
>>> [Mon Mar 25 23:35:17 2019] CPU: 2 PID: 29518 Comm: kworker/u8:4
>>> Tainted: G W 5.0.4-zb64-303ce93b05c9+ #1
>>
>> What commits does this kernel include because it doesn't seem to be a
>> pristine upstream 5.0.4 ? Also what you are seeing below is definitely a
>> bug in MM. The question is whether it's due to your doing faulty
>> backports in the kernel or it's due to something that got automatically
>> backported to 5.0.4
>
> That was the first thing I thought of, so I reverted to vanilla 5.0.4,
> repeated the test, and obtained the same result.
>
> You may have a point about non-btrfs patches in 5.0.4, though.
> I previously tested 5.0.3 with most of the 5.0.4 fs/btrfs commits
> already included by cherry-pick:
>
> 1098803b8cb7 Btrfs: fix deadlock between clone/dedupe and rename
> 3486142a68e3 Btrfs: fix corruption reading shared and compressed
> extents after hole punching
> fb9c36acfab1 btrfs: scrub: fix circular locking dependency warning
> 9d7b327affb8 Btrfs: setup a nofs context for memory allocation at
> __btrfs_set_acl
> 80dcd07c27df Btrfs: setup a nofs context for memory allocation at
> btrfs_create_tree()
>
> The commits that are in 5.0.4 but not in my last 5.0.3 test run are:
>
> ebbb48419e8a btrfs: init csum_list before possible free
> 88e610ae4c3a btrfs: ensure that a DUP or RAID1 block group has exactly
> two stripes
> 9c58f2ada4fa btrfs: drop the lock on error in btrfs_dev_replace_cancel
>
> and I don't see how those commits could lead to the observed changes
> in behavior. I didn't include them for 5.0.3 because my test scenario
> doesn't execute the code they touch. So the problem might be outside
> of btrfs completely.
I think it might very well be outside of btrfs because you are seeing an
order 0 failure when you have plenty of order 0 free pages. That's
definitely something you might want to report to mm.
>
