On Thu 12-04-18 07:40:41, Naoya Horiguchi wrote: > On Thu, Apr 12, 2018 at 08:18:59AM +0200, Michal Hocko wrote: > > On Wed 11-04-18 17:09:25, Naoya Horiguchi wrote: > > > Hi everyone, > > > > > > I wrote patches introducing separate vm event counters for hugepage migration > > > (both for hugetlb and thp.) > > > Hugepage migration is different from normal page migration in event frequency > > > and/or how likely it succeeds, so maintaining statistics for them in mixed > > > counters might not be helpful both for develors and users. > > > > This is quite a lot of code to be added se we should better document > > what it is intended for. Sure I understand your reasonaning about huge > > pages are more likely to fail but is this really worth a separate > > counter? Do you have an example of how this would be useful? > > Our customers periodically collect some log info to understand what > happened after system failures happen. Then if we have separate counters > for hugepage migration and the values show some anomaly, that might > help admins and developers understand the issue more quickly. > We have other ways to get this info like checking /proc/pid/pagemap and > /proc/kpageflags, but they are costly and most users decide not to > collect them in periodical logging. Wouldn't tracepoints be more suitable for that purpose? They can collect more valuable information. > > If we are there then what about different huge page sizes (for hugetlb)? > > Do we need per-hstate stats? > > Yes, per-hstate counters are better. And existing hugetlb counters > htlb_buddy_alloc_* are also affected by this point. The thing is that this would bloat the code and the vmstat output even more. I am not really convinced this is a great idea for something that tracepoints would handle as well. -- Michal Hocko SUSE Labs
