Message-ID: <87jzmd5yq1.fsf@yhuang6-desk2.ccr.corp.intel.com>
Date: Fri, 08 Mar 2024 08:55:18 +0800
From: "Huang, Ying" <ying.huang@...el.com>
To: Ryan Roberts <ryan.roberts@....com>
Cc: Miaohe Lin <linmiaohe@...wei.com>,  Andrew Morton
 <akpm@...ux-foundation.org>,  David Hildenbrand <david@...hat.com>,
  <linux-mm@...ck.org>,  <linux-kernel@...r.kernel.org>,
  <stable@...r.kernel.org>
Subject: Re: [PATCH v1] mm: swap: Fix race between free_swap_and_cache() and
 swapoff()

Ryan Roberts <ryan.roberts@....com> writes:

> On 07/03/2024 08:54, Huang, Ying wrote:
>> Ryan Roberts <ryan.roberts@....com> writes:
>> 
>>> On 07/03/2024 07:34, Huang, Ying wrote:
>>>> Miaohe Lin <linmiaohe@...wei.com> writes:
>>>>
>>>>> On 2024/3/7 13:56, Huang, Ying wrote:
>>>>>> Miaohe Lin <linmiaohe@...wei.com> 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@....com> 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.
>>>
>>> Yes that's the shmem problem I've been trying to convey. Perhaps there are other
>>> (extremely subtle) mechanisms that make this impossible, I don't know.
>>>
>>> Either way, given the length of this discussion, and the subtleties in the
>>> syncrhonization mechanisms that have so far been identified, I think the safest
>>> thing to do is just apply the patch. Then we have explicit syncrhonization that
>>> we can trivially reason about.
>> 
>> Yes.  This is tricky and we can improve it.  So I suggest to,
>> 
>> - Revise the patch description to use shmem race as example except
>>   someone found it's impossible.
>> 
>> - Revise the comments of get_swap_device() about RCU reader side lock
>>   (including IRQ off, spinlock, etc.) can prevent swapoff via
>>   synchronize_rcu() in swapoff().
>> 
>> - Revise the comments of synchronize_rcu() in swapoff(), which can
>>   prevent swapoff in parallel with RCU reader side lock including swap
>>   cache operations, etc.
>
> The only problem with this is that Andrew has already put my v2 into mm-*stable* :-|
>
> So (1) from that list isn't possible. I could do a patch for (2) and (3), but to
> be honest, I think you would do a better job of writing it up than I would - any
> chance you could post the patch?
>

Sure.  I will do that.

--
Best Regards,
Huang, Ying

Powered by blists - more mailing lists