There were apparently good reasons for moving away from OOM notifier callback: https://lkml.org/lkml/2018/7/12/314 https://lkml.org/lkml/2018/8/2/322 In particular the OOM notifier is worse than the shrinker because: 1. It is last-resort, which means the system has already gone through heroics to prevent OOM. Those heroic reclaim efforts are expensive and impact application performance. 2. It lacks understanding of NUMA or other OOM constraints. 3. It has a higher potential for bugs due to the subtlety of the callback context. Given the above, I think the shrinker API certainly makes the most sense _if_ the balloon size is static. In that case memory should be reclaimed from the balloon early and proportionally to balloon size, which the shrinker API achieves. However, if the balloon is inflating and intentionally causing memory pressure then this results in the inefficiency pointed out earlier. If the balloon is inflating but not causing memory pressure then there is no problem with either API. This suggests another route: rather than cause memory pressure to shrink the page cache, the balloon could issue the equivalent of "echo 3 > /proc/sys/vm/drop_caches". Of course ideally, we want to be more fine grained than "drop everything". We really want an API that says "drop everything that hasn't been accessed in the last 5 minutes". This would eliminate the need for the balloon to cause memory pressure at all which avoids the inefficiency in question. Furthermore, this pairs nicely with the FREE_PAGE_HINT feature. On Mon, Feb 3, 2020 at 9:04 AM Michael S. Tsirkin <mst at redhat.com> wrote:> On Mon, Feb 03, 2020 at 05:34:20PM +0100, David Hildenbrand wrote: > > On 03.02.20 17:18, Alexander Duyck wrote: > > > On Mon, 2020-02-03 at 08:11 -0500, Michael S. Tsirkin wrote: > > >> On Thu, Jan 30, 2020 at 11:59:46AM -0800, Tyler Sanderson wrote: > > >>> > > >>> On Thu, Jan 30, 2020 at 7:31 AM Wang, Wei W <wei.w.wang at intel.com> > wrote: > > >>> > > >>> On Thursday, January 30, 2020 11:03 PM, David Hildenbrand wrote: > > >>> > On 29.01.20 20:11, Tyler Sanderson wrote: > > >>> > > > > >>> > > > > >>> > > On Wed, Jan 29, 2020 at 2:31 AM David Hildenbrand < > david at redhat.com > > >>> > > <mailto:david at redhat.com>> wrote: > > >>> > > > > >>> > > On 29.01.20 01:22, Tyler Sanderson via Virtualization > wrote: > > >>> > > > A primary advantage of virtio balloon over other > memory reclaim > > >>> > > > mechanisms is that it can pressure the guest's page > cache into > > >>> > > shrinking. > > >>> > > > > > >>> > > > However, since the balloon driver changed to using the > shrinker > > >>> API > > >>> > > > > > >>> > > > > >>> > < > https://github.com/torvalds/linux/commit/71994620bb25a8b109388fefa9 > > >>> > e99a28e355255a#diff-fd202acf694d9eba19c8c64da3e480c9> this > > >>> > > > use case has become a bit more tricky. I'm wondering > what the > > >>> > intended > > >>> > > > device implementation is. > > >>> > > > > > >>> > > > When inflating the balloon against page cache (i.e. no > free > > >>> memory > > >>> > > > remains) vmscan.c will both shrink page cache, but > also invoke > > >>> the > > >>> > > > shrinkers -- including the balloon's shrinker. So the > balloon > > >>> driver > > >>> > > > allocates memory which requires reclaim, vmscan gets > this memory > > >>> > by > > >>> > > > shrinking the balloon, and then the driver adds the > memory back > > >>> to > > >>> > the > > >>> > > > balloon. Basically a busy no-op. > > >>> > > >>> Per my understanding, the balloon allocation won?t invoke > shrinker as > > >>> __GFP_DIRECT_RECLAIM isn't set, no? > > >>> > > >>> I could be wrong about the mechanism, but the device sees lots of > activity on > > >>> the deflate queue. The balloon is being shrunk. And this only starts > once all > > >>> free memory is depleted and we're inflating into page cache. > > >> > > >> So given this looks like a regression, maybe we should revert the > > >> patch in question 71994620bb25 ("virtio_balloon: replace oom notifier > with shrinker") > > >> Besides, with VIRTIO_BALLOON_F_FREE_PAGE_HINT > > >> shrinker also ignores VIRTIO_BALLOON_F_MUST_TELL_HOST which isn't nice > > >> at all. > > >> > > >> So it looks like all this rework introduced more issues than it > > >> addressed ... > > >> > > >> I also CC Alex Duyck for an opinion on this. > > >> Alex, what do you use to put pressure on page cache? > > > > > > I would say reverting probably makes sense. I'm not sure there is much > > > value to having a shrinker running deflation when you are actively > trying > > > to increase the balloon. It would make more sense to wait until you are > > > actually about to start hitting oom. > > > > I think the shrinker makes sense for free page hinting feature > > (everything on free_page_list). > > > > So instead of only reverting, I think we should split it up and always > > register the shrinker for VIRTIO_BALLOON_F_FREE_PAGE_HINT and the OOM > > notifier (as before) for VIRTIO_BALLOON_F_MUST_TELL_HOST. > > OK ... I guess that means we need to fix shrinker to take > VIRTIO_BALLOON_F_MUST_TELL_HOST into account correctly. > Hosts ignore it at the moment but it's a fragile thing > to do what it does and ignore used buffers. > > > (Of course, adapting what is being done in the shrinker and in the OOM > > notifier) > > > > -- > > Thanks, > > > > David / dhildenb > >-------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.linuxfoundation.org/pipermail/virtualization/attachments/20200203/6c3740dd/attachment-0001.html>
On Mon, Feb 03, 2020 at 12:32:05PM -0800, Tyler Sanderson wrote:> There were apparently good reasons for moving away from OOM notifier callback: > https://lkml.org/lkml/2018/7/12/314 > https://lkml.org/lkml/2018/8/2/322 > > In particular the OOM notifier is worse than the shrinker because: > > 1. It is last-resort, which means the system has already gone through heroics > to prevent OOM. Those heroic reclaim efforts are expensive and impact > application performance. > 2. It lacks understanding of NUMA or other OOM constraints. > 3. It has a higher potential for bugs due to the subtlety?of the callback > context. > > Given the above, I think the shrinker?API certainly makes the most sense _if_ > the balloon size is static. In that case memory should be reclaimed from the > balloon early and proportionally to balloon size, which the shrinker API > achieves.OK that sounds like VIRTIO_BALLOON_F_FREE_PAGE_HINT then.> However, if the balloon is inflating and intentionally causing memory pressure > then this results in the inefficiency pointed out earlier.And that sounds like VIRTIO_BALLOON_F_DEFLATE_ON_OOM.> If the balloon is inflating but not causing memory pressure then there is no > problem with either API. > > This suggests another route: rather than cause memory pressure to shrink the > page cache, the balloon could issue the equivalent?of "echo 3 > /proc/sys/vm/ > drop_caches". > Of course ideally, we want to be more fine grained than "drop everything". We > really want an API that says "drop everything that hasn't been accessed in the > last 5 minutes". > > This would eliminate the need for the balloon to cause memory pressure at all > which?avoids the inefficiency in question. Furthermore, this pairs nicely with > the FREE_PAGE_HINT feature.Well we still do have a regression. So we probably should revert for now, and separately look for better solutions.> > On Mon, Feb 3, 2020 at 9:04 AM Michael S. Tsirkin <mst at redhat.com> wrote: > > On Mon, Feb 03, 2020 at 05:34:20PM +0100, David Hildenbrand wrote: > > On 03.02.20 17:18, Alexander Duyck wrote: > > > On Mon, 2020-02-03 at 08:11 -0500, Michael S. Tsirkin wrote: > > >> On Thu, Jan 30, 2020 at 11:59:46AM -0800, Tyler Sanderson wrote: > > >>> > > >>> On Thu, Jan 30, 2020 at 7:31 AM Wang, Wei W <wei.w.wang at intel.com> > wrote: > > >>> > > >>>? ? ?On Thursday, January 30, 2020 11:03 PM, David Hildenbrand wrote: > > >>>? ? ?> On 29.01.20 20:11, Tyler Sanderson wrote: > > >>>? ? ?> > > > >>>? ? ?> > > > >>>? ? ?> > On Wed, Jan 29, 2020 at 2:31 AM David Hildenbrand < > david at redhat.com > > >>>? ? ?> > <mailto:david at redhat.com>> wrote: > > >>>? ? ?> > > > >>>? ? ?> >? ? ?On 29.01.20 01:22, Tyler Sanderson via Virtualization > wrote: > > >>>? ? ?> >? ? ?> A primary advantage of virtio balloon over other memory > reclaim > > >>>? ? ?> >? ? ?> mechanisms is that it can pressure the guest's page > cache into > > >>>? ? ?> >? ? ?shrinking. > > >>>? ? ?> >? ? ?> > > >>>? ? ?> >? ? ?> However, since the balloon driver changed to using the > shrinker > > >>>? ? ?API > > >>>? ? ?> >? ? ?> > > >>>? ? ?> > > > >>>? ? ?> <https://github.com/torvalds/linux/commit/ > 71994620bb25a8b109388fefa9 > > >>>? ? ?> e99a28e355255a#diff-fd202acf694d9eba19c8c64da3e480c9> this > > >>>? ? ?> >? ? ?> use case has become a bit more tricky. I'm wondering > what the > > >>>? ? ?> intended > > >>>? ? ?> >? ? ?> device implementation is. > > >>>? ? ?> >? ? ?> > > >>>? ? ?> >? ? ?> When inflating the balloon against page cache (i.e. no > free > > >>>? ? ?memory > > >>>? ? ?> >? ? ?> remains) vmscan.c will both shrink page cache, but also > invoke > > >>>? ? ?the > > >>>? ? ?> >? ? ?> shrinkers -- including the balloon's shrinker. So the > balloon > > >>>? ? ?driver > > >>>? ? ?> >? ? ?> allocates memory which requires reclaim, vmscan gets > this memory > > >>>? ? ?> by > > >>>? ? ?> >? ? ?> shrinking the balloon, and then the driver adds the > memory back > > >>>? ? ?to > > >>>? ? ?> the > > >>>? ? ?> >? ? ?> balloon. Basically a busy no-op. > > >>> > > >>>? ? ?Per my understanding, the balloon allocation won?t invoke > shrinker as > > >>>? ? ?__GFP_DIRECT_RECLAIM isn't set, no? > > >>> > > >>> I could be wrong about the mechanism, but the device sees lots of > activity on > > >>> the deflate queue. The balloon is being shrunk. And this only starts > once all > > >>> free memory is depleted and we're inflating into page cache. > > >> > > >> So given this looks like a regression, maybe we should revert the > > >> patch in question 71994620bb25 ("virtio_balloon: replace oom notifier > with shrinker") > > >> Besides, with VIRTIO_BALLOON_F_FREE_PAGE_HINT > > >> shrinker also ignores VIRTIO_BALLOON_F_MUST_TELL_HOST which isn't nice > > >> at all. > > >> > > >> So it looks like all this rework introduced more issues than it > > >> addressed ... > > >> > > >> I also CC Alex Duyck for an opinion on this. > > >> Alex, what do you use to put pressure on page cache? > > > > > > I would say reverting probably makes sense. I'm not sure there is much > > > value to having a shrinker running deflation when you are actively > trying > > > to increase the balloon. It would make more sense to wait until you are > > > actually about to start hitting oom. > > > > I think the shrinker makes sense for free page hinting feature > > (everything on free_page_list). > > > > So instead of only reverting, I think we should split it up and always > > register the shrinker for VIRTIO_BALLOON_F_FREE_PAGE_HINT and the OOM > > notifier (as before) for VIRTIO_BALLOON_F_MUST_TELL_HOST. > > OK ... I guess that means we need to fix shrinker to take > VIRTIO_BALLOON_F_MUST_TELL_HOST into account correctly. > Hosts ignore it at the moment but it's a fragile thing > to do what it does and ignore used buffers. > > > (Of course, adapting what is being done in the shrinker and in the OOM > > notifier) > > > > -- > > Thanks, > > > > David / dhildenb > >
On 03.02.20 21:32, Tyler Sanderson wrote:> There were apparently good reasons for moving away from OOM notifier > callback: > https://lkml.org/lkml/2018/7/12/314 > https://lkml.org/lkml/2018/8/2/322 > > In particular the OOM notifier is worse than the shrinker because:The issue is that DEFLATE_ON_OOM is under-specified.> > 1. It is last-resort, which means the system has already gone through > heroics to prevent OOM. Those heroic reclaim efforts are expensive > and impact application performance.That's *exactly* what "deflate on OOM" suggests. Assume you are using virtio-balloon for some weird way of memory hotunplug (which is what some people do) and you want to minimize the footprint of your guest. Then you really only want to give the guest more memory (or rather, let it take back memory automatically in this case) in case it really needs more memory. It should try to reclaim first. Under-specified.> 2. It lacks understanding of NUMA or other OOM constraints.Ballooning in general lacks the understanding of NUMA.> 3. It has a higher potential for bugs due to the subtlety?of the > callback context.While that is a valid point, it doesn't explain why existing functionality is changed. Personally, I think DEFLATE_ON_OOM should never have been introduced (at least not in this form). -- Thanks, David / dhildenb
On Tue, Feb 4, 2020 at 12:29 AM David Hildenbrand <david at redhat.com> wrote:> On 03.02.20 21:32, Tyler Sanderson wrote: > > There were apparently good reasons for moving away from OOM notifier > > callback: > > https://lkml.org/lkml/2018/7/12/314 > > https://lkml.org/lkml/2018/8/2/322 > > > > In particular the OOM notifier is worse than the shrinker because: > > The issue is that DEFLATE_ON_OOM is under-specified. > > > > > 1. It is last-resort, which means the system has already gone through > > heroics to prevent OOM. Those heroic reclaim efforts are expensive > > and impact application performance. > > That's *exactly* what "deflate on OOM" suggests. >It seems there are some use cases where "deflate on OOM" is desired and others where "deflate on pressure" is desired. This suggests adding a new feature bit "DEFLATE_ON_PRESSURE" that registers the shrinker, and reverting DEFLATE_ON_OOM to use the OOM notifier callback. This lets users configure the balloon for their use case.> > Assume you are using virtio-balloon for some weird way of memory > hotunplug (which is what some people do) and you want to minimize the > footprint of your guest. Then you really only want to give the guest > more memory (or rather, let it take back memory automatically in this > case) in case it really needs more memory. It should try to reclaim first. > > Under-specified. > > > > 2. It lacks understanding of NUMA or other OOM constraints. > > Ballooning in general lacks the understanding of NUMA. > > > 3. It has a higher potential for bugs due to the subtlety of the > > callback context. > > While that is a valid point, it doesn't explain why existing > functionality is changed. > > Personally, I think DEFLATE_ON_OOM should never have been introduced (at > least not in this form). >I'm actually not sure how you would safely do memory overcommit without DEFLATE_ON_OOM. So I think it unlocks a huge use case.> > > -- > Thanks, > > David / dhildenb > >-------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.linuxfoundation.org/pipermail/virtualization/attachments/20200204/9c4eafa7/attachment.html>