Re: [PATCH 1/2] mm: hugetlb: proc: check for hugetlb shared PMD in /proc/PID/smaps

[Mike Kravetz]

On 02/01/23 08:47, Michal Hocko wrote:
> On Mon 30-01-23 14:08:47, Mike Kravetz wrote:
> > On 01/30/23 13:36, Michal Hocko wrote:
> > > On Fri 27-01-23 17:12:05, Mike Kravetz wrote:
> > > > On 01/27/23 15:04, Andrew Morton wrote:
> > > > > On Fri, 27 Jan 2023 17:23:39 +0100 David Hildenbrand <david@xxxxxxxxxx> wrote:
> > > > > > On 26.01.23 23:27, Mike Kravetz wrote:
> > > 
> > > Yes, this looks simple enough. My only concern would be that this
> > > special casing might be required on other places which is hard to
> > > evaluate. I thought PSS reported by smaps would be broken as well but it
> > > seems pss is not really accounted for hugetlb mappings at all.
> > > 
> > > Have you tried to look into {in,de}creasing the map count of the page when
> > > the the pte is {un}shared for it?
> > 
> > A quick thought is that it would not be too difficult.  It would need
> > to include the following:
> > - At PMD share time in huge_pmd_share(),
> >   Go through all entries in the PMD, and increment map and ref count for
> >   all referenced pages.  huge_pmd_share is just adding another sharing
> >   process.
> > - At PMD unshare time in huge_pmd_unshare(),
> >   Go through all entries in the PMD, and decrement map and ref count for
> >   all referenced pages.  huge_pmd_unshare is just removing one sharing
> >   process.
> > - At page fault time, check if we are adding a new entry to a shared PMD.
> >   If yes, add 'num_of_sharing__processes - 1' to the ref and map count.
> > 
> > In each of the above operations, we are holding the PTL lock (which is
> > really the split/PMD lock) so synchronization should not be an issue.
> > 
> > Although I mention processes sharing the PMD above, it is really mappings/vmas
> > sharing the PMD.  You could have two mappings of the same object in the same
> > process sharing PMDs.
> > 
> > I'll code this up and see how it looks.
> 
> Thanks!
>  
> > However, unless you have an objection I would prefer the simple patches
> > move forward, especially for stable backports.
> 
> Yes, the current patch is much simpler and more suitable for stable
> backports. If the explicit map count modifications are not all that
> terrible then this would sound like a more appropriate long term plan
> though.

The approach mentioned above seems to be simple enough.  Patch is below.

I 'tested' with the same method and tests used to measure fault scalabilty
when developing vma based locking [1].  I figured this would be a good stress
of the share, unshare and fault paths.  With the patch, we are doing more
with the page table lock held, so I expected to see a little difference
in scalability, but not as much as actually measured:

				next-20230131
test		instances	unmodified	patched
--------------------------------------------------------------------------
Combined faults 24		61888.4		58314.8
Combined forks  24		  157.3		  130.1

These tests could seem a bit like a micro-benchmark targeting these code
paths.  However, I put them together based on the description of a
customer workload that prompted the vma based locking work.  And, performance
of these tests seems to reflect performance of their workloads.

This extra overhead is the cost needed to make shared PMD map counts be
accurate and in line with what is normal and expected.  I think it is
worth the cost.  Other opinions?  Of course, the patch below may have
issues so please take a look.

[1] https://lore.kernel.org/linux-mm/20220914221810.95771-1-mike.kravetz@xxxxxxxxxx/