Re: stable request: mm, page_alloc: actually ignore mempolicies for high priority allocations

[Vlastimil Babka <vbabka@suse.cz>]

On 11/8/18 9:30 AM, Mike Manning wrote:
> On 08/11/2018 07:54, Vlastimil Babka wrote:
>> +CC linux-mm
>>
>> On 11/7/18 6:33 PM, Mike Manning wrote:
>>> Hello, Please consider backporting to 4.14.y the following commit from
>>> kernel-net-next by Vlastimil Babka [CC'ed]:
>>>
>>> d6a24df00638 ("mm, page_alloc: actually ignore mempolicies for high
>>> priority allocations") It cherry-picks cleanly and builds fine.
>>>
>>> The reason for the request is that the commit 1d26c112959f ("mm,
>>> page_alloc:do not break __GFP_THISNODE by zonelist reset") that was
>>> previously backported to 4.14.y broke some of our functionality after we
>>> upgraded from an earlier 4.14 kernel without the fix.
>> Well, that's very surprising! Could you be more specific about what
>> exactly got broken?
> 
> Thank you for your reply. I agree, we were also very surprised when
> bisecting our updated 4.14 kernel, as this change is apparently
> completely unrelated to our application running in userspace. But the
> problem was 100% reproducible on a baremetal setup running automated
> performance multi-stream testing, so only seen under load.

So what was the workload doing, and what were the symptoms, at least
from a high level perspective? And was it a vanilla 4.14.y kernel, or
with some additional patches, out of tree modules etc?

> With the fix
> reverted from the 4.14 kernel, the problem went away, and this is with
> many repeated runs (the load test is part of a suite that is
> automatically run quite a few times every day, and this test was failing
> since the upgrade).
> 
>>
>>> The reason this is
>>> happening is not clear, with this commit only found by bisect.
>>> Fortunately the requested commit resolves the issue.
>> I would like to understand the problem first, because I currently can't
>> imagine how the first commit could break something and the second fix it.
> 
> I agree, but from an empirical point of view, 2 options present:
> 
> 1) The original commit was not suitable for backport to 4.14 and should
> be reverted.
> 
> 2) For the same reason that the original commit was suitable for
> backport to 4.14, the requested commit should also be backported.

I don't think that covers all possibilities.

You didn't say what the observed problem was, so I can imagine it was
either allocation failures, OOM's, or worse performance (probably
related to network).

The original commit should be a non-functional change for allocations
that don't use __GFP_THISNODE. The zonelist reassignment that was
removed by the patch might have changed order of zones to allocate from,
but the set of available zones for the allocation should be unchanged,
unless the zonelist generation code is broken (and then we should better
find out). Otherwise I can only imagine some minor performance impact.

The patch you are requesting is a functional change, with positive
effects expected on all 3 potential problems I listed above. It should
however make a difference only in the context of processes restricted by
bind mempolicies... or potentially some out-of-tree modules. And again,
it shouldn't be related to the original commit.

>>> Best Regards,
>>>
>>> Mike Manning
>>>
>