Re: [PATCH v4 1/1] memcg/hugetlb: Add hugeTLB counters to memcg

[David Rientjes <rientjes@google.com>]

On Thu, 14 Nov 2024, Johannes Weiner wrote:

> > > > The reason that I opted not to include a breakdown of each hugetlb
> > > > size in memory.stat is only because I wanted to keep the addition that
> > > > this patch makes as minimal as possible, while still addressing
> > > > the goal of bridging the gap between memory.stat and memory.current.
> > > > Users who are curious about this breakdown can see how much memory
> > > > is used by each hugetlb size by enabling the hugetlb controller as well.
> > > > 
> > > 
> > > While the patch may be minimal, this is solidifying a kernel API that 
> > > users will start to count on.  Users who may be interested in their 
> > > hugetlb usage may not have control over the configuration of their kernel?
> > > 
> > > Does it make sense to provide a breakdown in memory.stat so that users can 
> > > differentiate between mapping one 1GB hugetlb page and 512 2MB hugetlb 
> > > pages, which are different global resources?
> > > 
> > > > It's true that this is the case as well for total hugeltb usage, but
> > > > I felt that not including hugetlb memory usage in memory.stat when it
> > > > is accounted by memory.current would cause confusion for the users
> > > > not being able to see that memory.current = sum of memory.stat. On the
> > > > other hand, seeing the breakdown of how much each hugetlb size felt more
> > > > like an optimization, and not a solution that bridges a confusion.
> > > > 
> > > 
> > > If broken down into hugetlb_2048kB and hugetlb_1048576kB on x86, for 
> > > example, users could still do sum of memory.stat, no?>
> > > 
> > 
> > Friendly ping on this, would there be any objections to splitting the 
> > memory.stat metrics out to be per hugepage size?
> 
> I don't think it has to be either/or. We can add the total here, and a
> per-size breakdown in a separate patch (with its own changelog)?
> 
> That said, a per-size breakdown might make more sense in the hugetlb
> cgroup controller. You're mentioning separate global resources, which
> suggests this is about more explicitly controlled hugetlb use.
> 
> From a memcg POV, all hugetlb is the same. It's just (non-swappable)
> memory consumed by the cgroup.
> 

Ok, that's fair.  We have a local patch that tracks hugetlb usage, 
admittedly for all hugetlb sizes, in struct mem_cgroup_per_node so that we 
can provide a breakdown in memory.numa_stat because we can't get the 
per-node breakdown from hugetlb_cgroup.  If there is interest in that 
breakdown, we could easily propose the patch.