On Thu 11-11-21 07:02:42, Suren Baghdasaryan wrote:

On Thu, Nov 11, 2021 at 1:20 AM Michal Hocko [off-list ref] wrote:

quoted

On Wed 10-11-21 17:49:37, Suren Baghdasaryan wrote:

quoted

On Tue, Nov 9, 2021 at 1:10 PM Suren Baghdasaryan [off-list ref] wrote:

quoted

On Tue, Nov 9, 2021 at 12:10 PM Michal Hocko [off-list ref] wrote:

[...]

quoted

quoted

quoted

Yes, those can run concurrently. One thing I completely forgot about is
27ae357fa82b ("mm, oom: fix concurrent munlock and oom reaper unmap, v3")
which is about interaction with the munlock.

Agrh! This interaction with the munlock you mentioned requires us to
take mmap_write_lock before munlock_vma_pages_all and that prevents
__oom_reap_task_mm from running concurrently with unmap_vmas. The
reapers would not be as effective as they are now after such a change
:(

__oom_reap_task_mm will not run concurrently with unmap_vmas even
with the current code. The mmap_sem barrier right before munlock code
prevents that.

You are right, it will run concurrently with another
__oom_reap_task_mm in the exit_mmap. But I thought we wanted to get
rid of that call to __oom_reap_task_mm in exit_mmap or did I
misunderstand?

I do not remember this to be objective or the motivation. IIRC we wanted
to make the locking more robust which would help your process_mrelease
use case. This one currently suffers from a much heavier cost if it
turns out to be the last holder of the reference count on the address
space.
-- 
Michal Hocko
SUSE Labs