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

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On 08/11/2018 09:32, Vlastimil Babka wrote:
> 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?
We carry patches, but they are not at all in this area. The tests, which
are in relation to traffic forwarding, are not portable to a stock linux
image.
>> 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).
No errors are reported in the journal such as OOM.
> 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 failure is with ND being tested with Spirent under load, so it
causes a complete test failure, as it stops any subsequent performance
testing for IPv6 traffic forwarding.
> 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.

We do not have any processes explicitly restricted by bind mempolicies.

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




[Index of Archives]     [Linux Kernel]     [Kernel Development Newbies]     [Linux USB Devel]     [Video for Linux]     [Linux Audio Users]     [Yosemite Hiking]     [Linux Kernel]     [Linux SCSI]

  Powered by Linux