Re: [PATCH 4/5] mm: rework non-root kmem_cache lifecycle management

[Roman Gushchin <guro@fb.com>]

On Thu, Apr 18, 2019 at 07:05:24AM -0700, Shakeel Butt wrote:
> On Wed, Apr 17, 2019 at 8:07 PM Roman Gushchin <guro@fb.com> wrote:
> >
> > On Wed, Apr 17, 2019 at 06:55:12PM -0700, Shakeel Butt wrote:
> > > On Wed, Apr 17, 2019 at 5:39 PM Roman Gushchin <guro@fb.com> wrote:
> > > >
> > > > On Wed, Apr 17, 2019 at 04:41:01PM -0700, Shakeel Butt wrote:
> > > > > On Wed, Apr 17, 2019 at 2:55 PM Roman Gushchin <guroan@gmail.com> wrote:
> > > > > >
> > > > > > This commit makes several important changes in the lifecycle
> > > > > > of a non-root kmem_cache, which also affect the lifecycle
> > > > > > of a memory cgroup.
> > > > > >
> > > > > > Currently each charged slab page has a page->mem_cgroup pointer
> > > > > > to the memory cgroup and holds a reference to it.
> > > > > > Kmem_caches are held by the cgroup. On offlining empty kmem_caches
> > > > > > are freed, all other are freed on cgroup release.
> > > > >
> > > > > No, they are not freed (i.e. destroyed) on offlining, only
> > > > > deactivated. All memcg kmem_caches are freed/destroyed on memcg's
> > > > > css_free.
> > > >
> > > > You're right, my bad. I was thinking about the corresponding sysfs entry
> > > > when was writing it. We try to free it from the deactivation path too.
> > > >
> > > > >
> > > > > >
> > > > > > So the current scheme can be illustrated as:
> > > > > > page->mem_cgroup->kmem_cache.
> > > > > >
> > > > > > To implement the slab memory reparenting we need to invert the scheme
> > > > > > into: page->kmem_cache->mem_cgroup.
> > > > > >
> > > > > > Let's make every page to hold a reference to the kmem_cache (we
> > > > > > already have a stable pointer), and make kmem_caches to hold a single
> > > > > > reference to the memory cgroup.
> > > > >
> > > > > What about memcg_kmem_get_cache()? That function assumes that by
> > > > > taking reference on memcg, it's kmem_caches will stay. I think you
> > > > > need to get reference on the kmem_cache in memcg_kmem_get_cache()
> > > > > within the rcu lock where you get the memcg through css_tryget_online.
> > > >
> > > > Yeah, a very good question.
> > > >
> > > > I believe it's safe because css_tryget_online() guarantees that
> > > > the cgroup is online and won't go offline before css_free() in
> > > > slab_post_alloc_hook(). I do initialize kmem_cache's refcount to 1
> > > > and drop it on offlining, so it protects the online kmem_cache.
> > > >
> > >
> > > Let's suppose a thread doing a remote charging calls
> > > memcg_kmem_get_cache() and gets an empty kmem_cache of the remote
> > > memcg having refcnt equal to 1. That thread got a reference on the
> > > remote memcg but no reference on the kmem_cache. Let's suppose that
> > > thread got stuck in the reclaim and scheduled away. In the meantime
> > > that remote memcg got offlined and decremented the refcnt of all of
> > > its kmem_caches. The empty kmem_cache which the thread stuck in
> > > reclaim have pointer to can get deleted and may be using an already
> > > destroyed kmem_cache after coming back from reclaim.
> > >
> > > I think the above situation is possible unless the thread gets the
> > > reference on the kmem_cache in memcg_kmem_get_cache().
> >
> > Yes, you're right and I'm writing a nonsense: css_tryget_online()
> > can't prevent the cgroup from being offlined.
> >
> 
> The reason I knew about that race is because I tried something similar
> but for different use-case:
> 
> https://lkml.org/lkml/2018/3/26/472
> 
> > So, the problem with getting a reference in memcg_kmem_get_cache()
> > is that it's an atomic operation on the hot path, something I'd like
> > to avoid.
> >
> > I can make the refcounter percpu, but it'll add some complexity and size
> > to the kmem_cache object. Still an option, of course.
> >
> 
> I kind of prefer this option.
> 
> > I wonder if we can use rcu_read_lock() instead, and bump the refcounter
> > only if we're going into reclaim.
> >
> > What do you think?
> 
> Should it be just reclaim or anything that can reschedule the current thread?
> 
> I can tell how we resolve the similar issue for our
> eager-kmem_cache-deletion use-case. Our solution (hack) works only for
> CONFIG_SLAB (we only use SLAB) and non-preemptible kernel. The
> underlying motivation was to reduce the overhead of slab reaper of
> traversing thousands of empty offlined kmem caches. CONFIG_SLAB
> disables interrupts before accessing the per-cpu caches and reenables
> the interrupts if it has to fallback to the page allocation. We use
> this window to call memcg_kmem_get_cache() and only increment the
> refcnt of kmem_cache if going to the fallback. Thus no need to do
> atomic operation on the hot path.
> 
> Anyways, I think having percpu refcounter for each memcg kmem_cache is
> not that costy for CONFIG_MEMCG_KMEM users and to me that seems like
> the most simple solution.
> 
> Shakeel

Ok, sounds like a percpu refcounter is the best option.
I'll try this approach in v2.

Thanks!