On Wed, Aug 28, 2019 at 09:57:08AM +0200, Michal Hocko wrote: > On Tue 27-08-19 10:06:20, Yang Shi wrote: > > > > > > On 8/27/19 5:59 AM, Kirill A. Shutemov wrote: > > > On Tue, Aug 27, 2019 at 03:17:39PM +0300, Kirill A. Shutemov wrote: > > > > On Tue, Aug 27, 2019 at 02:09:23PM +0200, Michal Hocko wrote: > > > > > On Tue 27-08-19 14:01:56, Vlastimil Babka wrote: > > > > > > On 8/27/19 1:02 PM, Kirill A. Shutemov wrote: > > > > > > > On Tue, Aug 27, 2019 at 08:01:39AM +0200, Michal Hocko wrote: > > > > > > > > On Mon 26-08-19 16:15:38, Kirill A. Shutemov wrote: > > > > > > > > > Unmapped completely pages will be freed with current code. Deferred split > > > > > > > > > only applies to partly mapped THPs: at least on 4k of the THP is still > > > > > > > > > mapped somewhere. > > > > > > > > Hmm, I am probably misreading the code but at least current Linus' tree > > > > > > > > reads page_remove_rmap -> [page_remove_anon_compound_rmap ->\ deferred_split_huge_page even > > > > > > > > for fully mapped THP. > > > > > > > Well, you read correctly, but it was not intended. I screwed it up at some > > > > > > > point. > > > > > > > > > > > > > > See the patch below. It should make it work as intened. > > > > > > > > > > > > > > It's not bug as such, but inefficientcy. We add page to the queue where > > > > > > > it's not needed. > > > > > > But that adding to queue doesn't affect whether the page will be freed > > > > > > immediately if there are no more partial mappings, right? I don't see > > > > > > deferred_split_huge_page() pinning the page. > > > > > > So your patch wouldn't make THPs freed immediately in cases where they > > > > > > haven't been freed before immediately, it just fixes a minor > > > > > > inefficiency with queue manipulation? > > > > > Ohh, right. I can see that in free_transhuge_page now. So fully mapped > > > > > THPs really do not matter and what I have considered an odd case is > > > > > really happening more often. > > > > > > > > > > That being said this will not help at all for what Yang Shi is seeing > > > > > and we need a more proactive deferred splitting as I've mentioned > > > > > earlier. > > > > It was not intended to fix the issue. It's fix for current logic. I'm > > > > playing with the work approach now. > > > Below is what I've come up with. It appears to be functional. > > > > > > Any comments? > > > > Thanks, Kirill and Michal. Doing split more proactive is definitely a choice > > to eliminate huge accumulated deferred split THPs, I did think about this > > approach before I came up with memcg aware approach. But, I thought this > > approach has some problems: > > > > First of all, we can't prove if this is a universal win for the most > > workloads or not. For some workloads (as I mentioned about our usecase), we > > do see a lot THPs accumulated for a while, but they are very short-lived for > > other workloads, i.e. kernel build. > > > > Secondly, it may be not fair for some workloads which don't generate too > > many deferred split THPs or those THPs are short-lived. Actually, the cpu > > time is abused by the excessive deferred split THPs generators, isn't it? > > Yes this is indeed true. Do we have any idea on how much time that > actually is? For uncontented case, splitting 1G worth of pages (2MiB x 512) takes a bit more than 50 ms in my setup. But it's best-case scenario: pages not shared across multiple processes, no contention on ptl, page lock, etc. > > With memcg awareness, the deferred split THPs actually are isolated and > > capped by memcg. The long-lived deferred split THPs can't be accumulated too > > many due to the limit of memcg. And, cpu time spent in splitting them would > > just account to the memcgs who generate that many deferred split THPs, who > > generate them who pay for it. This sounds more fair and we could achieve > > much better isolation. > > On the other hand, deferring the split and free up a non trivial amount > of memory is a problem I consider quite serious because it affects not > only the memcg workload which has to do the reclaim but also other > consumers of memory beucase large memory blocks could be used for higher > order allocations. Maybe instead of drive the split from number of pages on queue we can take a hint from compaction that is struggles to get high order pages? We can also try to use schedule_delayed_work() instead of plain schedule_work() to give short-lived page chance to get freed before splitting attempt. > > And, I think the discussion is diverted and mislead by the number of > > excessive deferred split THPs. To be clear, I didn't mean the excessive > > deferred split THPs are problem for us (I agree it may waste memory to have > > that many deferred split THPs not usable), the problem is the oom since they > > couldn't be split by memcg limit reclaim since the shrinker was not memcg > > aware. > > Well, I would like to see how much of a problem the memcg OOM really is > after deferred splitting is more time constrained. Maybe we will find > that there is no special memcg aware solution really needed. > -- > Michal Hocko > SUSE Labs -- Kirill A. Shutemov
