On Wed, Jan 28, 2015 at 07:22:51PM +0300, Vladimir Davydov wrote:
> +++ b/mm/slub.c
> @@ -2007,6 +2007,7 @@ static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
> int pages;
> int pobjects;
>
> + preempt_disable();
> do {
> pages = 0;
> pobjects = 0;
> @@ -2040,6 +2041,14 @@ static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
>
> } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
> != oldpage);
> + if (unlikely(!s->cpu_partial)) {
> + unsigned long flags;
> +
> + local_irq_save(flags);
> + unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
> + local_irq_restore(flags);
> + }
> + preempt_enable();
> #endif
> }
>
> @@ -3369,7 +3378,7 @@ EXPORT_SYMBOL(kfree);
> * being allocated from last increasing the chance that the last objects
> * are freed in them.
> */
> -int __kmem_cache_shrink(struct kmem_cache *s)
> +int __kmem_cache_shrink(struct kmem_cache *s, bool deactivate)
> {
> int node;
> int i;
> @@ -3381,14 +3390,26 @@ int __kmem_cache_shrink(struct kmem_cache *s)
> unsigned long flags;
> int ret = 0;
>
> + if (deactivate) {
> + /*
> + * Disable empty slabs caching. Used to avoid pinning offline
> + * memory cgroups by kmem pages that can be freed.
> + */
> + s->cpu_partial = 0;
> + s->min_partial = 0;
> +
> + /*
> + * s->cpu_partial is checked locklessly (see put_cpu_partial),
> + * so we have to make sure the change is visible.
> + */
> + kick_all_cpus_sync();
> + }
Argh! what the heck! and without a single mention in the changelog.
Why are you spraying IPIs across the entire machine? Why isn't
synchronize_sched() good enough, that would allow you to get rid of the
local_irq_save/restore as well.
On Fri, Apr 01, 2016 at 11:04:41AM +0200, Peter Zijlstra wrote:
> On Wed, Jan 28, 2015 at 07:22:51PM +0300, Vladimir Davydov wrote:
> > +++ b/mm/slub.c
> > @@ -2007,6 +2007,7 @@ static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
> > int pages;
> > int pobjects;
> >
> > + preempt_disable();
> > do {
> > pages = 0;
> > pobjects = 0;
> > @@ -2040,6 +2041,14 @@ static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
> >
> > } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
> > != oldpage);
> > + if (unlikely(!s->cpu_partial)) {
> > + unsigned long flags;
> > +
> > + local_irq_save(flags);
> > + unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
> > + local_irq_restore(flags);
> > + }
> > + preempt_enable();
> > #endif
> > }
> >
> > @@ -3369,7 +3378,7 @@ EXPORT_SYMBOL(kfree);
> > * being allocated from last increasing the chance that the last objects
> > * are freed in them.
> > */
> > -int __kmem_cache_shrink(struct kmem_cache *s)
> > +int __kmem_cache_shrink(struct kmem_cache *s, bool deactivate)
> > {
> > int node;
> > int i;
> > @@ -3381,14 +3390,26 @@ int __kmem_cache_shrink(struct kmem_cache *s)
> > unsigned long flags;
> > int ret = 0;
> >
> > + if (deactivate) {
> > + /*
> > + * Disable empty slabs caching. Used to avoid pinning offline
> > + * memory cgroups by kmem pages that can be freed.
> > + */
> > + s->cpu_partial = 0;
> > + s->min_partial = 0;
> > +
> > + /*
> > + * s->cpu_partial is checked locklessly (see put_cpu_partial),
> > + * so we have to make sure the change is visible.
> > + */
> > + kick_all_cpus_sync();
> > + }
>
> Argh! what the heck! and without a single mention in the changelog.
This function is only called when a memory cgroup is removed, which is
rather a rare event. I didn't think it would cause any pain. Sorry.
>
> Why are you spraying IPIs across the entire machine? Why isn't
> synchronize_sched() good enough, that would allow you to get rid of the
> local_irq_save/restore as well.
synchronize_sched() is slower. Calling it for every per memcg kmem cache
would slow down cleanup on cgroup removal. The latter is async, so I'm
not sure if it would be a problem though. I think we can try to replace
kick_all_cpus_sync() with synchronize_sched() here.
Regarding local_irq_save/restore - synchronize_sched() wouldn't allow us
to get rid of them, because unfreeze_partials() must be called with irqs
disabled.
Come to think of it, kick_all_cpus_sync() is used as a memory barrier
here, so as to make sure that after it's finished all cpus will use the
new ->cpu_partial value, which makes me wonder if we could replace it
with a simple smp_mb. I mean, this_cpu_cmpxchg(), which is used by
put_cpu_partial to add a page to per-cpu partial list, must issue a full
memory barrier (am I correct?), so we have two possibilities here:
Case 1: smp_mb is called before this_cpu_cmpxchg is called on another
cpu executing put_cpu_partial. In this case, put_cpu_partial will see
cpu_partial == 0 and hence call the second unfreeze_partials, flushing
per cpu partial list.
Case 2: smp_mb is called after this_cpu_cmpxchg. Then
__kmem_cache_shrink ->flush_all -> has_cpu_slab should see (thanks to
the barriers) that there's a slab on a per-cpu list and so flush it
(provided it hasn't already been flushed by put_cpu_partial).
In any case, after __kmem_cache_shrink has finished, we are guaranteed
to not have any slabs on per cpu partial lists.
Does it make sense?
Thanks,
Vladimir
On Fri, Apr 01, 2016 at 01:55:40PM +0300, Vladimir Davydov wrote:
> > > + if (deactivate) {
> > > + /*
> > > + * Disable empty slabs caching. Used to avoid pinning offline
> > > + * memory cgroups by kmem pages that can be freed.
> > > + */
> > > + s->cpu_partial = 0;
> > > + s->min_partial = 0;
> > > +
> > > + /*
> > > + * s->cpu_partial is checked locklessly (see put_cpu_partial),
> > > + * so we have to make sure the change is visible.
> > > + */
> > > + kick_all_cpus_sync();
> > > + }
> >
> > Argh! what the heck! and without a single mention in the changelog.
>
> This function is only called when a memory cgroup is removed, which is
> rather a rare event. I didn't think it would cause any pain. Sorry.
Suppose you have a bunch of CPUs running HPC/RT code and someone causes
the admin CPUs to create/destroy a few cgroups.
> > Why are you spraying IPIs across the entire machine? Why isn't
> > synchronize_sched() good enough, that would allow you to get rid of the
> > local_irq_save/restore as well.
>
> synchronize_sched() is slower. Calling it for every per memcg kmem cache
> would slow down cleanup on cgroup removal.
Right, but who cares? cgroup removal isn't a fast path by any standard.
> Regarding local_irq_save/restore - synchronize_sched() wouldn't allow us
> to get rid of them, because unfreeze_partials() must be called with irqs
> disabled.
OK, I figured it was because it needed to be serialized against this
kick_all_cpus_sync() IPI.
> Come to think of it, kick_all_cpus_sync() is used as a memory barrier
> here, so as to make sure that after it's finished all cpus will use the
> new ->cpu_partial value, which makes me wonder if we could replace it
> with a simple smp_mb. I mean, this_cpu_cmpxchg(), which is used by
> put_cpu_partial to add a page to per-cpu partial list, must issue a full
> memory barrier (am I correct?), so we have two possibilities here:
Nope, this_cpu_cmpxchg() does not imply a memory barrier.