On Fri, Oct 22, 2021 at 10:03:29AM +0200, Michal Hocko wrote:

On Thu 21-10-21 18:46:58, Suren Baghdasaryan wrote:

quoted

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. There was a push back against
that though so arguments would have to be re-evaluated.

I have another reason for wanting to take the mmap_sem throughout
exit_mmap.  Liam and I are working on using the Maple tree to replace
the rbtree & vma linked list.  It uses lockdep to check that you haven't
forgotten to take a lock (as of two days ago, that mean the mmap_sem
or the RCU read lock) when walking the tree.

So I'd like to hold it over:

 - unlock_range()
 - unmap_vmas()
 - free_pgtables()
 - while (vma) remove_vma()

Which is basically the whole of exit_mmap().  I'd like to know more
about why there was pushback on holding the mmap_lock across this
-- we're exiting, so nobody else should have a reference to the mm?