Re: [PATCH 1/1] mm: prevent a race between process_mrelease and exit_mmap

[Suren Baghdasaryan <surenb@google.com>]

On Mon, Nov 1, 2021 at 1:37 AM Michal Hocko <mhocko@suse.com> wrote:
>
> On Fri 29-10-21 09:07:39, Suren Baghdasaryan wrote:
> > On Fri, Oct 29, 2021 at 6:03 AM Michal Hocko <mhocko@suse.com> wrote:
> [...]
> > > Well, I still do not see why that is a problem. This syscall is meant to
> > > release the address space not to do it fast.
> >
> > It's the same problem for a userspace memory reaper as for the
> > oom-reaper. The goal is to release the memory of the victim and to
> > quickly move on to the next one if needed.
>
> The purpose of the oom_reaper is to _guarantee_ a forward progress. It
> doesn't have to be quick or optimized for speed.

Fair enough. Then the same guarantees should apply to userspace memory
reapers. I think you clarified that well in your replies in
https://lore.kernel.org/all/20170725154514.GN26723@dhcp22.suse.cz:

Because there is no _guarantee_ that the final __mmput will release
the memory in finite time. And we cannot guarantee that longterm.
...
__mmput calls into exit_aio and that can wait for completion and there
is no way to guarantee this will finish in finite time.

>
> [...]
>
> > > Btw. the above code will not really tell you much on a larger machine
> > > unless you manage to trigger mmap_sem contection. Otherwise you are
> > > measuring the mmap_sem writelock fast path and that should be really
> > > within a noise comparing to the whole address space destruction time. If
> > > that is not the case then we have a real problem with the locking...
> >
> > My understanding of that discussion is that the concern was that even
> > taking uncontended mmap_sem writelock would regress the exit path.
> > That was what I wanted to confirm. Am I misreading it?
>
> No, your reading match my recollection. I just think that code
> robustness in exchange of a rw semaphore write lock fast path is a
> reasonable price to pay even if that has some effect on micro
> benchmarks.

I'm with you on this one, that's why I wanted to measure the price we
would pay. Below are the test results:

Test: https://lore.kernel.org/all/20170725142626.GJ26723@dhcp22.suse.cz/
Compiled: gcc -O2 -static test.c -o test
Test machine: 128 core / 256 thread 2x AMD EPYC 7B12 64-Core Processor
(family 17h)

baseline (Linus master, f31531e55495ca3746fb895ffdf73586be8259fa)
p50 (median)   87412
p95                  168210
p99                  190058
average           97843.8
stdev               29.85%

unconditional mmap_write_lock in exit_mmap (last column is the change
from the baseline)
p50 (median)   88312     +1.03%
p95                  170797   +1.54%
p99                  191813   +0.92%
average           97659.5  -0.19%
stdev               32.41%

unconditional mmap_write_lock in exit_mmap + Matthew's patch (last
column is the change from the baseline)
p50 (median)   88807      +1.60%
p95                  167783     -0.25%
p99                  187853     -1.16%
average           97491.4    -0.36%
stdev               30.61%

stdev is quite high in all cases, so the test is very noisy.
The impact seems quite low IMHO. WDYT?

> --
> Michal Hocko
> SUSE Labs