On Thu, Sep 29, 2016 at 04:45:50PM +0300, Vladimir Davydov wrote: > On Thu, Sep 29, 2016 at 11:00:50AM +0900, Joonsoo Kim wrote: > > On Wed, Sep 28, 2016 at 11:09:53AM +0300, Vladimir Davydov wrote: > > > On Tue, Sep 27, 2016 at 10:03:47PM -0400, Johannes Weiner wrote: > > > > [CC Vladimir] > > > > > > > > These are the delayed memcg cache allocations, where in a fresh memcg > > > > that doesn't have per-memcg caches yet, every accounted allocation > > > > schedules a kmalloc work item in __memcg_schedule_kmem_cache_create() > > > > until the cache is finally available. It looks like those can be many > > > > more than the number of slab caches in existence, if there is a storm > > > > of slab allocations before the workers get a chance to run. > > > > > > > > Vladimir, what do you think of embedding the work item into the > > > > memcg_cache_array? That way we make sure we have exactly one work per > > > > cache and not an unbounded number of them. The downside of course is > > > > that we'd have to keep these things around as long as the memcg is in > > > > existence, but that's the only place I can think of that allows us to > > > > serialize this. > > > > > > We could set the entry of the root_cache->memcg_params.memcg_caches > > > array corresponding to the cache being created to a special value, say > > > (void*)1, and skip scheduling cache creation work on kmalloc if the > > > caller sees it. I'm not sure it's really worth it though, because > > > work_struct isn't that big (at least, in comparison with the cache > > > itself) to avoid embedding it at all costs. > > > > Hello, Johannes and Vladimir. > > > > I'm not familiar with memcg so have a question about this solution. > > This solution will solve the current issue but if burst memcg creation > > happens, similar issue would happen again. My understanding is correct? > > Yes, I think you're right - embedding the work_struct responsible for > cache creation in kmem_cache struct won't help if a thousand of > different cgroups call kmem_cache_alloc() simultaneously for a cache > they haven't used yet. > > Come to think of it, we could fix the issue by simply introducing a > special single-threaded workqueue used exclusively for cache creation > works - cache creation is done mostly under the slab_mutex, anyway. This > way, we wouldn't have to keep those used-once work_structs for the whole > kmem_cache life time. > > > > > I think that the other cause of the problem is that we call > > synchronize_sched() which is rather slow with holding a slab_mutex and > > it blocks further kmem_cache creation. Should we fix that, too? > > Well, the patch you posted looks pretty obvious and it helps the > reporter, so personally I don't see any reason for not applying it. Oops... I forgot to mention why I asked that. There is another report that similar problem also happens in SLUB. In there, synchronize_sched() is called in cache shrinking path with holding the slab_mutex. I guess that it blocks further kmem_cache creation. If we uses special single-threaded workqueue, number of kworker would be limited but kmem_cache creation will be delayed for a long time in burst memcg creation/destroy scenario. https://bugzilla.kernel.org/show_bug.cgi?id=172991 Do we need to remove synchronize_sched() in SLUB and find other solution? Thanks. -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>