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

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Wed 27-10-21 09:08:21, Suren Baghdasaryan wrote:
> On Fri, Oct 22, 2021 at 10:38 AM Suren Baghdasaryan <surenb@xxxxxxxxxx> wrote:
> >
> > On Fri, Oct 22, 2021 at 1:03 AM Michal Hocko <mhocko@xxxxxxxx> wrote:
> > >
> > > On Thu 21-10-21 18:46:58, Suren Baghdasaryan wrote:
> > > > Race between process_mrelease and exit_mmap, where free_pgtables is
> > > > called while __oom_reap_task_mm is in progress, leads to kernel crash
> > > > during pte_offset_map_lock call. oom-reaper avoids this race by setting
> > > > MMF_OOM_VICTIM flag and causing exit_mmap to take and release
> > > > mmap_write_lock, blocking it until oom-reaper releases mmap_read_lock.
> > > > Reusing MMF_OOM_VICTIM for process_mrelease would be the simplest way to
> > > > fix this race, however that would be considered a hack. Fix this race
> > > > by elevating mm->mm_users and preventing exit_mmap from executing until
> > > > process_mrelease is finished. Patch slightly refactors the code to adapt
> > > > for a possible mmget_not_zero failure.
> > > > This fix has considerable negative impact on process_mrelease performance
> > > > and will likely need later optimization.
> > >
> > > I am not sure there is any promise that process_mrelease will run in
> > > parallel with the exiting process. In fact the primary purpose of this
> > > syscall is to provide a reliable way to oom kill from user space. If you
> > > want to optimize process exit resp. its exit_mmap part then you should
> > > be using other means. So I would be careful calling this a regression.
> > >
> > > I do agree that taking the reference count is the right approach here. I
> > > was wrong previously [1] when saying that pinning the mm struct is
> > > sufficient. I have completely forgot about the subtle sync in exit_mmap.
> > > One way we can approach that would be to take exclusive mmap_sem
> > > throughout the exit_mmap unconditionally.
> >
> > I agree, that would probably be the cleanest way.
> >
> > > There was a push back against
> > > that though so arguments would have to be re-evaluated.
> >
> > I'll review that discussion to better understand the reasons for the
> > push back. Thanks for the link.
> 
> Adding Kirill and Andrea.
> 
> I had some time to dig some more. The latency increase is definitely
> coming due to process_mrelease calling the last mmput and exit_aio is
> especially problematic. So, currently process_mrelease not only
> releases memory but does more, including waiting for io to finish.

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.

> 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@xxxxxxxxxxxxxx/ on
> a server with and without a custom patch.

Well, I do not remember all the details of the discussion but I believe
a rather large part of that discussion was a bit misled. The exist
path - and the last mmput in particular - shouldn't trigger mmap_sem
contention. There are only rare cases where somebody can race and take a
lock then (e.g. proc interfaces taking the lock before mmget_notzero).
Certainly not something to optimize for and I believe a correct and
robust code should have a preference. As we can see a lack of proper
synchronization has led to 2 very similar problem nobody revealed during
review because the code is just too tricky.

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...
-- 
Michal Hocko
SUSE Labs




[Index of Archives]     [Linux ARM Kernel]     [Linux ARM]     [Linux Omap]     [Fedora ARM]     [IETF Annouce]     [Bugtraq]     [Linux OMAP]     [Linux MIPS]     [eCos]     [Asterisk Internet PBX]     [Linux API]

  Powered by Linux