Re: [PATCH 1/1] RFC: add pidfd_send_signal flag to reclaim mm while killing a process

[Suren Baghdasaryan]

On Fri, Nov 13, 2020 at 5:18 PM Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx> wrote:
>
> On Fri, 13 Nov 2020 17:09:37 -0800 Suren Baghdasaryan <surenb@xxxxxxxxxx> wrote:
>
> > > > > Seems to me that the ability to reap another process's memory is a
> > > > > generally useful one, and that it should not be tied to delivering a
> > > > > signal in this fashion.
> > > > >
> > > > > And we do have the new process_madvise(MADV_PAGEOUT).  It may need a
> > > > > few changes and tweaks, but can't that be used to solve this problem?
> > > >
> > > > Thank you for the feedback, Andrew. process_madvise(MADV_DONTNEED) was
> > > > one of the options recently discussed in
> > > > https://lore.kernel.org/linux-api/CAJuCfpGz1kPM3G1gZH+09Z7aoWKg05QSAMMisJ7H5MdmRrRhNQ@xxxxxxxxxxxxxx
> > > > . The thread describes some of the issues with that approach but if we
> > > > limit it to processes with pending SIGKILL only then I think that
> > > > would be doable.
> > >
> > > Why would it be necessary to read /proc/pid/maps?  I'd have thought
> > > that a starting effort would be
> > >
> > >         madvise((void *)0, (void *)-1, MADV_PAGEOUT)
> > >
> > > (after translation into process_madvise() speak).  Which is equivalent
> > > to the proposed process_madvise(MADV_DONTNEED_MM)?
> >
> > Yep, this is very similar to option #3 in
> > https://lore.kernel.org/linux-api/CAJuCfpGz1kPM3G1gZH+09Z7aoWKg05QSAMMisJ7H5MdmRrRhNQ@xxxxxxxxxxxxxx
> > and I actually have a tested prototype for that.
>
> Why is the `vector=NULL' needed?  Can't `vector' point at a single iovec
> which spans the whole address range?

That would be the option #4 from the same discussion and the issues
noted there are "process_madvise return value can't handle such a
large number of bytes and there is MAX_RW_COUNT limit on max number of
bytes one process_madvise call can handle". In my prototype I have a
special handling for such "bulk operation" to work around the
MAX_RW_COUNT limitation.

>
> > If that's the
> > preferred method then I can post it quite quickly.
>
> I assume you've tested that prototype.  How did its usefulness compare
> with this SIGKILL-based approach?

Just to make sure I understand correctly your question, you are asking
about performance comparison of:

// approach in this RFC
pidfd_send_signal(SIGKILL, SYNC_REAP_MM)

vs

// option #4 in the previous RFC
kill(SIGKILL); process_madvise(vector=NULL, MADV_DONTNEED);

If so, I have results for the current RFC approach but the previous
approach was testing on an older device, so don't have
apples-to-apples comparison results at the moment. I can collect the
data for fair comparison if desired, however I don't expect a
noticeable performance difference since they both do pretty much the
same thing (even on different devices my results are quite close). I
think it's more a question of which API would be more appropriate.

>