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

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite for Android: free password hash cracker in your pocket

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <CAM_iQpU-BQZM2EyxBZ9_maT7WM2LH_Weu4sgBsy=oEZ+aabAbg@mail.gmail.com>
Date:   Fri, 4 Aug 2017 10:58:49 -0700
From:   Cong Wang <xiyou.wangcong@...il.com>
To:     Andrew Morton <akpm@...ux-foundation.org>
Cc:     Linus Torvalds <torvalds@...ux-foundation.org>,
        Linux Kernel Mailing List <linux-kernel@...r.kernel.org>,
        linux-mm <linux-mm@...ck.org>, "# .39.x" <stable@...nel.org>,
        Hugh Dickins <hughd@...gle.com>,
        "Kirill A . Shutemov" <kirill.shutemov@...ux.intel.com>
Subject: Re: [PATCH] mm: fix list corruptions on shmem shrinklist

On Thu, Aug 3, 2017 at 4:53 PM, Andrew Morton <akpm@...ux-foundation.org> wrote:
> On Thu, 3 Aug 2017 16:25:46 -0700 Linus Torvalds <torvalds@...ux-foundation.org> wrote:
>
>> On Thu, Aug 3, 2017 at 4:11 PM, Andrew Morton <akpm@...ux-foundation.org> 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.