On Tue 13-03-18 10:55:18, Shakeel Butt wrote:
> On Tue, Mar 13, 2018 at 6:49 AM, Michal Hocko <mhocko@xxxxxxxxxx> wrote:
> > On Wed 21-02-18 14:37:56, Shakeel Butt wrote:
> > [...]
> >> +#ifdef CONFIG_MEMCG
> >> +static inline struct mem_cgroup *memalloc_memcg_save(struct mem_cgroup *memcg)
> >> +{
> >> + struct mem_cgroup *old_memcg = current->target_memcg;
> >> + current->target_memcg = memcg;
> >> + return old_memcg;
> >> +}
> >
> > So you are relying that the caller will handle the reference counting
> > properly? I do not think this is a good idea.
>
> For the fsnotify use-case, this assumption makes sense as fsnotify has
> an abstraction of fsnotify_group which is created by the
> person/process interested in the events and thus can be used to hold
> the reference to the person/process's memcg.
OK, but there is not any direct connection between fsnotify_group and
task_struct lifetimes, is it? This makes the API suspectible to
use-after-free bugs.
> Another use-case I have
> in mind is the filesystem mount. Basically attaching a mount with a
> memcg and thus all user pages and kmem allocations (inodes, dentries)
> for that mount will be charged to the attached memcg.
So you charge page cache to the origin task but metadata to a different
memcg?
> In this use-case
> the super_block is the perfect structure to hold the reference to the
> memcg.
>
> If in future we find a use-case where this assumption does not make
> sense we can evolve the API and since this is kernel internal API, it
> should not be hard to evolve.
>
> > Also do we need some kind
> > of debugging facility to detect unbalanced save/restore scopes?
> >
>
> I am not sure, I didn't find other similar patterns (like PF_MEMALLOC)
> having debugging facility.
Maybe we need something more generic here.
> Maybe we can add such debugging facility
> when we find more users other than kmalloc & kmem_cache_alloc. Vmalloc
> may be one but I could not think of a use-case for vmalloc for remote
> charging, so, no need to add more code at this time.
>
> > [...]
> >> @@ -2260,7 +2269,10 @@ struct kmem_cache *memcg_kmem_get_cache(struct kmem_cache *cachep)
> >> if (current->memcg_kmem_skip_account)
> >> return cachep;
> >>
> >> - memcg = get_mem_cgroup_from_mm(current->mm);
> >> + if (current->target_memcg)
> >> + memcg = get_mem_cgroup(current->target_memcg);
> >> + if (!memcg)
> >> + memcg = get_mem_cgroup_from_mm(current->mm);
> >> kmemcg_id = READ_ONCE(memcg->kmemcg_id);
> >> if (kmemcg_id < 0)
> >> goto out;
> >
> > You are also adding one branch for _each_ charge path even though the
> > usecase is rather limited.
> >
>
> I understand the concern but the charging path, IMO, is much complex
> than just one or couple of additional branches. I can run a simple
> microbenchmark to see if there is anything noticeable here.
Charging path is still a _hot path_. Especially when the kmem accounting
is enabled by default. You cannot simply downplay the overhead. We have
_one_ user but all users should pay the price. This is simply hard to
justify. Maybe we can thing of something that would put the burden on
the charging context?
--
Michal Hocko
SUSE Labs
