Re: [PATCH 0/9] mm/madvise: batch tlb flushes for MADV_DONTNEED and MADV_FREE

[Lorenzo Stoakes]

On Mon, Mar 10, 2025 at 11:36:52PM +0000, Roman Gushchin wrote:
> On Mon, Mar 10, 2025 at 04:15:06PM -0700, Shakeel Butt wrote:
> > On Mon, Mar 10, 2025 at 03:39:21PM -0700, Andrew Morton wrote:
> > > On Mon, 10 Mar 2025 10:23:09 -0700 SeongJae Park <sj@xxxxxxxxxx> wrote:
> > >
> > > >  It is unclear if such use case
> > > > is common and the inefficiency is significant.
> > >
> > > Well, we could conduct a survey,
> > >
> > > Can you add some logging to detect when userspace performs such an
> > > madvise() call, then run that kernel on some "typical" machines which
> > > are running "typical" workloads?  That should give us a feeling for how
> > > often userspace does this, and hence will help us understand the usefulness
> > > of this patchset.
> >
> > Just for the clarification, this patchset is very useful for the
> > process_madvise() and the experiment results show that.
>
> +1
>
> Google carried an internal version for a vectorized madvise() which
> was much faster than process_madvise() last time I measured it.
> I hope SJ's patchset will (partially) address this difference,
> which will hopefully allow to drop the internal implementation
> for process_madvise.

Relatedly I also feel, at some point, we ought to remove the UIO_FASTIOV
limit on process_madvise().

But one for a future series...