Re: [PATCH] generic/551: prevent OOM on NFS on systems with no swap memory

Re: [PATCH] generic/551: prevent OOM on NFS on systems with no swap memory

[Christoph Hellwig]

On Wed, Mar 25, 2026 at 11:58:22AM +0800, Yongcheng Yang wrote:
> From: Scott Mayhew <smayhew@redhat.com>
> 
> We have frequently observed the oom-killer killing aio-dio-write-verify
> when generic/551 is run on NFS filesystems on virtual machines in AWS.
> 
> Virtual machines in AWS typically don't have a swap partition, so check
> for that condition when testing NFS and only use 90% of available memory
> when generating the list of write operations passed to
> aio-dio-write-verify.

I don't think this is a good idea.  The proper fix is reduce whatever
crazy large memory allocations this workloads causes in NFS.  I suspect
it might be page lists or similar, and just breaking them into somewhat
smaller chunks and/or using potentially failing allocations to
dynamically adjust would help.

Re: [PATCH] generic/551: prevent OOM on NFS on systems with no swap memory

[Darrick J. Wong]

On Tue, Mar 24, 2026 at 10:52:05PM -0700, Christoph Hellwig wrote:
> On Wed, Mar 25, 2026 at 11:58:22AM +0800, Yongcheng Yang wrote:
> > From: Scott Mayhew <smayhew@redhat.com>
> > 
> > We have frequently observed the oom-killer killing aio-dio-write-verify
> > when generic/551 is run on NFS filesystems on virtual machines in AWS.
> > 
> > Virtual machines in AWS typically don't have a swap partition, so check
> > for that condition when testing NFS and only use 90% of available memory
> > when generating the list of write operations passed to
> > aio-dio-write-verify.
> 
> I don't think this is a good idea.  The proper fix is reduce whatever
> crazy large memory allocations this workloads causes in NFS.  I suspect
> it might be page lists or similar, and just breaking them into somewhat
> smaller chunks and/or using potentially failing allocations to
> dynamically adjust would help.

I run fstests on XFS every night on a fleets of VMs with no swap and
never hit OOM.

Or at least I didn't until IT mandated CrowdStrike last week and now
it's anyone's guess if the test results are valid. <grumble>

--D