Re: [PATCH v3] userfaultfd: opportunistic TLB-flush batching for present pages in MOVE

[Peter Xu <peterx@redhat.com>]

On Fri, Aug 08, 2025 at 09:29:58AM -0700, Lokesh Gidra wrote:
> On Thu, Aug 7, 2025 at 12:17 PM Peter Xu <peterx@redhat.com> wrote:
> >
> > Hi, Lokesh,
> >
> > On Thu, Aug 07, 2025 at 03:39:02AM -0700, Lokesh Gidra wrote:
> > > MOVE ioctl's runtime is dominated by TLB-flush cost, which is required
> > > for moving present pages. Mitigate this cost by opportunistically
> > > batching present contiguous pages for TLB flushing.
> > >
> > > Without batching, in our testing on an arm64 Android device with UFFD GC,
> > > which uses MOVE ioctl for compaction, we observed that out of the total
> > > time spent in move_pages_pte(), over 40% is in ptep_clear_flush(), and
> > > ~20% in vm_normal_folio().
> > >
> > > With batching, the proportion of vm_normal_folio() increases to over
> > > 70% of move_pages_pte() without any changes to vm_normal_folio().
> >
> > Do you know why vm_normal_folio() could be expensive? I still see quite
> > some other things this path needs to do.
> >
> Let's discuss this in Andrew's reply thread.

Sorry to get back to this late.  Thanks for the link, Andrew!

> 
> > > Furthermore, time spent within move_pages_pte() is only ~20%, which
> > > includes TLB-flush overhead.
> >
> > Indeed this should already prove the optimization, I'm just curious whether
> > you've run some benchmark on the GC app to show the real world benefit.
> >
> I did! The same benchmark through which I gathered these numbers, when
> run on cuttlefish (qemu android instance on x86_64), the completion
> time of the benchmark went down from ~45mins to ~20mins. The benchmark
> is very GC intensive and the overhead of IPI on vCPUs seems to be
> enormous leading to this drastic improvement.
> 
> In another instance, system_server, one of the most critical system
> processes on android, saw over 50% reduction in GC compaction time on
> an arm64 android device.

Would you mind add some of these numbers into the commit message when you
repost?

Thanks,

-- 
Peter Xu