Re: [PATCH] mm: fix list corruptions on shmem shrinklist

[Cong Wang]

On Thu, Aug 3, 2017 at 4:53 PM, Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx> wrote:
> On Thu, 3 Aug 2017 16:25:46 -0700 Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
>
>> On Thu, Aug 3, 2017 at 4:11 PM, Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx> wrote:
>> >
>> > Where is this INIT_LIST_HEAD()?
>>
>> I think it's this one:
>>
>>         list_del_init(&info->shrinklist);
>>
>> in shmem_unused_huge_shrink().
>
> OK.
>
>> > I'm not sure I'm understanding this.  AFAICT all the list operations to
>> > which you refer are synchronized under spin_lock(&sbinfo->shrinklist_lock)?
>>
>> No, notice how shmem_unused_huge_shrink() does the
>>
>>         list_move(&info->shrinklist, &to_remove);
>>
>> and
>>
>>         list_move(&info->shrinklist, &list);
>>
>> to move to (two different) private lists under the shrinklist_lock,
>> but once it is on that private "list/to_remove" list, it is then
>> accessed outside the locked region.
>
> So the code is using sbinfo->shrinklist_lock to protect
> sbinfo->shrinklist AND to protect all the per-inode info->shrinklist's.
> Except it didn't get the coverage complete.


Normally once we move list entries from a global list to a private
one they are no longer visible to others, however in this case
they could still be accessed via setattr() path.

>
> Presumably it's too expensive to extend sbinfo->shrinklist_lock
> coverage in shmem_unused_huge_shrink() (or is it?  - this is huge
> pages).  An alternative would be to add a new
> shmem_inode_info.shrinklist_lock whose mandate is to protect
> shmem_inode_info.shrinklist.

Both find_lock_page() and iput() could sleep, I think this is why
we have to defer these two calls after releasing spinlock.

>
>> Honestly, I don't love this situation, or the patch, but I think the
>> patch is likely the right thing to do.
>
> Well, we could view the premature droppage of sbinfo->shrinklist_lock
> in shmem_unused_huge_shrink() to be a performance optimization and put
> some big fat comments in there explaining what's going on.  But it's
> tricky and it's not known that such an optimization is warranted.

It is not for performance optimization, because we still traverse
the list with the spinlock held. A typical optimization is to use
a list_splice() with spinlock and traverse it without it, this is
used by a few places in networking subsystem.

Thanks.