On Wed, Oct 27, 2021 at 10:52 AM Matthew Wilcox [off-list ref] wrote:

On Wed, Oct 27, 2021 at 10:42:29AM -0700, Suren Baghdasaryan wrote:

quoted

On Wed, Oct 27, 2021 at 10:35 AM Matthew Wilcox [off-list ref] wrote:

quoted

On Wed, Oct 27, 2021 at 09:08:21AM -0700, Suren Baghdasaryan wrote:

quoted

Unconditional mmap_write_lock around free_pgtables in exit_mmap seems
to me the most semantically correct way forward and the pushback is on
the basis of regressing performance of the exit path. I would like to
measure that regression to confirm this. I don't have access to a big
machine but will ask someone in another Google team to try the test
Michal wrote here
https://lore.kernel.org/all/20170725142626.GJ26723@dhcp22.suse.cz/ (local) on
a server with and without a custom patch.

Sorry to hijack this, but could you ask that team to also test this
patch?  I think there's probably a good-sized win here, but I have no
profiles to share at this point.  I've only done light testing, and
it may have bugs.

NB: I only did the exit() path here.  fork() conversion is left as an
exercise for the reader^W^W Liam.

To clarify, this patch does not change the mmap_write_lock portion of
exit_mmap. Do you want to test it in isolation or with the locking
changes in exit_mmap I mentioned?

Correct, it does not.  I think it's interesting to test it in isolation,
but if you want to test it in in combination, that could also be
interesting (see if we regain some of the expected performance loss).
I just don't have a NUMA box of my own to test on, so I'm hoping to
exploit your test infrastructure ;-)

By the way, my vmavec patch should also be helpful on small systems
like phones ... ;-)

Sounds good. I'll try to queue up the patches so that it's easy to
test them both in isolation and together.