Re: [PATCH v4 0/7] mm: reparent slab memory on cgroup removal

[Roman Gushchin <guro@fb.com>]

On Wed, Jun 05, 2019 at 12:39:24AM -0700, Greg Thelen wrote:
> Roman Gushchin <guro@fb.com> wrote:
> 
> > # Why do we need this?
> >
> > We've noticed that the number of dying cgroups is steadily growing on most
> > of our hosts in production. The following investigation revealed an issue
> > in userspace memory reclaim code [1], accounting of kernel stacks [2],
> > and also the mainreason: slab objects.
> >
> > The underlying problem is quite simple: any page charged
> > to a cgroup holds a reference to it, so the cgroup can't be reclaimed unless
> > all charged pages are gone. If a slab object is actively used by other cgroups,
> > it won't be reclaimed, and will prevent the origin cgroup from being reclaimed.
> >
> > Slab objects, and first of all vfs cache, is shared between cgroups, which are
> > using the same underlying fs, and what's even more important, it's shared
> > between multiple generations of the same workload. So if something is running
> > periodically every time in a new cgroup (like how systemd works), we do
> > accumulate multiple dying cgroups.
> >
> > Strictly speaking pagecache isn't different here, but there is a key difference:
> > we disable protection and apply some extra pressure on LRUs of dying cgroups,
> > and these LRUs contain all charged pages.
> > My experiments show that with the disabled kernel memory accounting the number
> > of dying cgroups stabilizes at a relatively small number (~100, depends on
> > memory pressure and cgroup creation rate), and with kernel memory accounting
> > it grows pretty steadily up to several thousands.
> >
> > Memory cgroups are quite complex and big objects (mostly due to percpu stats),
> > so it leads to noticeable memory losses. Memory occupied by dying cgroups
> > is measured in hundreds of megabytes. I've even seen a host with more than 100Gb
> > of memory wasted for dying cgroups. It leads to a degradation of performance
> > with the uptime, and generally limits the usage of cgroups.
> >
> > My previous attempt [3] to fix the problem by applying extra pressure on slab
> > shrinker lists caused a regressions with xfs and ext4, and has been reverted [4].
> > The following attempts to find the right balance [5, 6] were not successful.
> >
> > So instead of trying to find a maybe non-existing balance, let's do reparent
> > the accounted slabs to the parent cgroup on cgroup removal.
> >
> >
> > # Implementation approach
> >
> > There is however a significant problem with reparenting of slab memory:
> > there is no list of charged pages. Some of them are in shrinker lists,
> > but not all. Introducing of a new list is really not an option.
> >
> > But fortunately there is a way forward: every slab page has a stable pointer
> > to the corresponding kmem_cache. So the idea is to reparent kmem_caches
> > instead of slab pages.
> >
> > It's actually simpler and cheaper, but requires some underlying changes:
> > 1) Make kmem_caches to hold a single reference to the memory cgroup,
> >    instead of a separate reference per every slab page.
> > 2) Stop setting page->mem_cgroup pointer for memcg slab pages and use
> >    page->kmem_cache->memcg indirection instead. It's used only on
> >    slab page release, so it shouldn't be a big issue.
> > 3) Introduce a refcounter for non-root slab caches. It's required to
> >    be able to destroy kmem_caches when they become empty and release
> >    the associated memory cgroup.
> >
> > There is a bonus: currently we do release empty kmem_caches on cgroup
> > removal, however all other are waiting for the releasing of the memory cgroup.
> > These refactorings allow kmem_caches to be released as soon as they
> > become inactive and free.
> >
> > Some additional implementation details are provided in corresponding
> > commit messages.
> >
> > # Results
> >
> > Below is the average number of dying cgroups on two groups of our production
> > hosts. They do run some sort of web frontend workload, the memory pressure
> > is moderate. As we can see, with the kernel memory reparenting the number
> > stabilizes in 60s range; however with the original version it grows almost
> > linearly and doesn't show any signs of plateauing. The difference in slab
> > and percpu usage between patched and unpatched versions also grows linearly.
> > In 7 days it exceeded 200Mb.
> >
> > day           0    1    2    3    4    5    6    7
> > original     56  362  628  752 1070 1250 1490 1560
> > patched      23   46   51   55   60   57   67   69
> > mem diff(Mb) 22   74  123  152  164  182  214  241
> 
> No objection to the idea, but a question...

Hi Greg!

> In patched kernel, does slabinfo (or similar) show the list reparented
> slab caches?  A pile of zombie kmem_caches is certainly better than a
> pile of zombie mem_cgroup.  But it still seems like it'll might cause
> degradation - does cache_reap() walk an ever growing set of zombie
> caches?

It's not a pile of zombie kmem_caches vs a pile of zombie mem_cgroups.
It's a smaller pile of zombie kmem_caches vs a larger pile of zombie kmem_caches
*and* a pile of zombie mem_cgroups. The patchset makes the number of zombie
kmem_caches lower, not bigger.

Re slabinfo and other debug interfaces: I do not change anything here.

> 
> We've found it useful to add a slabinfo_full file which includes zombie
> kmem_cache with their memcg_name.  This can help hunt down zombies.

I'm not sure we need to add a permanent debug interface, because something like
drgn ( https://github.com/osandov/drgn ) can be used instead.

If you think that we lack some necessary debug interfaces, I'm totally open
here, but it's not a part of this patchset. Let's talk about them separately.

Thank you for looking into it!

Roman