Re: [PATCH v2 1/2] mm: protect free_pgtables with mmap_lock write lock in exit_mmap

[Michal Hocko <mhocko@suse.com>]

On Wed 24-11-21 15:59:05, Suren Baghdasaryan wrote:
> oom-reaper and process_mrelease system call should protect against
> races with exit_mmap which can destroy page tables while they
> walk the VMA tree. oom-reaper protects from that race by setting
> MMF_OOM_VICTIM and by relying on exit_mmap to set MMF_OOM_SKIP
> before taking and releasing mmap_write_lock. process_mrelease has
> to elevate mm->mm_users to prevent such race. Both oom-reaper and
> process_mrelease hold mmap_read_lock when walking the VMA tree.
> The locking rules and mechanisms could be simpler if exit_mmap takes
> mmap_write_lock while executing destructive operations such as
> free_pgtables.
> Change exit_mmap to hold the mmap_write_lock when calling
> free_pgtables. Operations like unmap_vmas() and unlock_range() are not
> destructive and could run under mmap_read_lock but for simplicity we
> take one mmap_write_lock during almost the entire operation. Note
> also that because oom-reaper checks VM_LOCKED flag, unlock_range()
> should not be allowed to race with it.
> In most cases this lock should be uncontended. Previously, Kirill
> reported ~4% regression caused by a similar change [1]. We reran the
> same test and although the individual results are quite noisy, the
> percentiles show lower regression with 1.6% being the worst case [2].
> The change allows oom-reaper and process_mrelease to execute safely
> under mmap_read_lock without worries that exit_mmap might destroy page
> tables from under them.
> 
> [1] https://lore.kernel.org/all/20170725141723.ivukwhddk2voyhuc@node.shutemov.name/
> [2] https://lore.kernel.org/all/CAJuCfpGC9-c9P40x7oy=jy5SphMcd0o0G_6U1-+JAziGKG6dGA@mail.gmail.com/
> 
> Signed-off-by: Suren Baghdasaryan <surenb@google.com>
> ---
> changes in v2
> - Moved mmap_write_unlock to cover remove_vma loop as well, per Matthew Wilcox
> 
>  mm/mmap.c | 16 ++++++++--------
>  1 file changed, 8 insertions(+), 8 deletions(-)
> 
> diff --git a/mm/mmap.c b/mm/mmap.c
> index bfb0ea164a90..f4e09d390a07 100644
> --- a/mm/mmap.c
> +++ b/mm/mmap.c
> @@ -3142,25 +3142,27 @@ void exit_mmap(struct mm_struct *mm)
>  		 * to mmu_notifier_release(mm) ensures mmu notifier callbacks in
>  		 * __oom_reap_task_mm() will not block.
>  		 *
> -		 * This needs to be done before calling munlock_vma_pages_all(),
> +		 * This needs to be done before calling unlock_range(),
>  		 * which clears VM_LOCKED, otherwise the oom reaper cannot
>  		 * reliably test it.
>  		 */
>  		(void)__oom_reap_task_mm(mm);
>  
>  		set_bit(MMF_OOM_SKIP, &mm->flags);

Why do you keep this in place?

Other than that looks OK to me. Maybe we want to add an explicit note
that vm_ops::close cannot take mmap_sem in any form. The changelog
should also mention that you have considered remove_vma and its previous
no MM locking assumption. You can argue that fput is async and close
callback shouldn't really need mmap_sem.
-- 
Michal Hocko
SUSE Labs