Re: [PATCH v1] mm: swap: Fix race between free_swap_and_cache() and swapoff()

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

 



Miaohe Lin <linmiaohe@xxxxxxxxxx> writes:

> On 2024/3/7 13:56, Huang, Ying wrote:
>> Miaohe Lin <linmiaohe@xxxxxxxxxx> writes:
>> 
>>> On 2024/3/6 17:31, Ryan Roberts wrote:
>>>> On 06/03/2024 08:51, Miaohe Lin wrote:
>>>>> On 2024/3/6 10:52, Huang, Ying wrote:
>>>>>> Ryan Roberts <ryan.roberts@xxxxxxx> writes:
>>>>>>
>>>>>>> There was previously a theoretical window where swapoff() could run and
>>>>>>> teardown a swap_info_struct while a call to free_swap_and_cache() was
>>>>>>> running in another thread. This could cause, amongst other bad
>>>>>>> possibilities, swap_page_trans_huge_swapped() (called by
>>>>>>> free_swap_and_cache()) to access the freed memory for swap_map.
>>>>>>>
>>>>>>> This is a theoretical problem and I haven't been able to provoke it from
>>>>>>> a test case. But there has been agreement based on code review that this
>>>>>>> is possible (see link below).
>>>>>>>
>>>>>>> Fix it by using get_swap_device()/put_swap_device(), which will stall
>>>>>>> swapoff(). There was an extra check in _swap_info_get() to confirm that
>>>>>>> the swap entry was valid. This wasn't present in get_swap_device() so
>>>>>>> I've added it. I couldn't find any existing get_swap_device() call sites
>>>>>>> where this extra check would cause any false alarms.
>>>>>>>
>>>>>>> Details of how to provoke one possible issue (thanks to David Hilenbrand
>>>>>>> for deriving this):
>>>>>>>
>>>>>>> --8<-----
>>>>>>>
>>>>>>> __swap_entry_free() might be the last user and result in
>>>>>>> "count == SWAP_HAS_CACHE".
>>>>>>>
>>>>>>> swapoff->try_to_unuse() will stop as soon as soon as si->inuse_pages==0.
>>>>>>>
>>>>>>> So the question is: could someone reclaim the folio and turn
>>>>>>> si->inuse_pages==0, before we completed swap_page_trans_huge_swapped().
>>>>>>>
>>>>>>> Imagine the following: 2 MiB folio in the swapcache. Only 2 subpages are
>>>>>>> still references by swap entries.
>>>>>>>
>>>>>>> Process 1 still references subpage 0 via swap entry.
>>>>>>> Process 2 still references subpage 1 via swap entry.
>>>>>>>
>>>>>>> Process 1 quits. Calls free_swap_and_cache().
>>>>>>> -> count == SWAP_HAS_CACHE
>>>>>>> [then, preempted in the hypervisor etc.]
>>>>>>>
>>>>>>> Process 2 quits. Calls free_swap_and_cache().
>>>>>>> -> count == SWAP_HAS_CACHE
>>>>>>>
>>>>>>> Process 2 goes ahead, passes swap_page_trans_huge_swapped(), and calls
>>>>>>> __try_to_reclaim_swap().
>>>>>>>
>>>>>>> __try_to_reclaim_swap()->folio_free_swap()->delete_from_swap_cache()->
>>>>>>> put_swap_folio()->free_swap_slot()->swapcache_free_entries()->
>>>>>>> swap_entry_free()->swap_range_free()->
>>>>>>> ...
>>>>>>> WRITE_ONCE(si->inuse_pages, si->inuse_pages - nr_entries);
>>>>>>>
>>>>>>> What stops swapoff to succeed after process 2 reclaimed the swap cache
>>>>>>> but before process1 finished its call to swap_page_trans_huge_swapped()?
>>>>>>>
>>>>>>> --8<-----
>>>>>>
>>>>>> I think that this can be simplified.  Even for a 4K folio, this could
>>>>>> happen.
>>>>>>
>>>>>> CPU0                                     CPU1
>>>>>> ----                                     ----
>>>>>>
>>>>>> zap_pte_range
>>>>>>   free_swap_and_cache
>>>>>>   __swap_entry_free
>>>>>>   /* swap count become 0 */
>>>>>>                                          swapoff
>>>>>>                                            try_to_unuse
>>>>>>                                              filemap_get_folio
>>>>>>                                              folio_free_swap
>>>>>>                                              /* remove swap cache */
>>>>>>                                            /* free si->swap_map[] */
>>>>>>
>>>>>>   swap_page_trans_huge_swapped <-- access freed si->swap_map !!!
>>>>>
>>>>> Sorry for jumping the discussion here. IMHO, free_swap_and_cache is called with pte lock held.
>>>>
>>>> I don't beleive it has the PTL when called by shmem.
>>>
>>> In the case of shmem, folio_lock is used to guard against the race.
>> 
>> I don't find folio is lock for shmem.  find_lock_entries() will only
>> lock the folio if (!xa_is_value()), that is, not swap entry.  Can you
>> point out where the folio is locked for shmem?
>
> You're right, folio is locked if not swap entry. That's my mistake. But it seems above race is still nonexistent.
> shmem_unuse() will first be called to read all the shared memory data that resides in the swap device back into
> memory when doing swapoff. In that case, all the swapped pages are moved to page cache thus there won't be any
> xa_is_value(folio) cases when calling shmem_undo_range(). free_swap_and_cache() even won't be called from
> shmem_undo_range() after shmem_unuse(). Or am I miss something?

I think the following situation is possible.  Right?

CPU0                               CPU1
----                               ----
shmem_undo_range
  shmem_free_swap
    xa_cmpxchg_irq
    free_swap_and_cache
      __swap_entry_free
      /* swap count become 0 */
                                   swapoff
                                     try_to_unuse
                                       shmem_unuse /* cannot find swap entry */
                                       find_next_to_unuse
                                       filemap_get_folio
                                       folio_free_swap
                                       /* remove swap cache */
                                       /* free si->swap_map[] */
      swap_page_trans_huge_swapped <-- access freed si->swap_map !!!

shmem_undo_range can run earlier.

--
Best Regards,
Huang, Ying




[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