Re: fs, net: deadlock between bind/splice on af_unix

From: Cong Wang
Date: Fri Jan 27 2017 - 00:12:42 EST

On Thu, Jan 26, 2017 at 3:29 PM, Mateusz Guzik <mguzik@xxxxxxxxxx> wrote:
> On Tue, Jan 17, 2017 at 01:21:48PM -0800, Cong Wang wrote:
>> On Mon, Jan 16, 2017 at 1:32 AM, Dmitry Vyukov <dvyukov@xxxxxxxxxx> wrote:
>> > On Fri, Dec 9, 2016 at 7:41 AM, Al Viro <viro@xxxxxxxxxxxxxxxxxx> wrote:
>> >> On Thu, Dec 08, 2016 at 10:32:00PM -0800, Cong Wang wrote:
>> >>
>> >>> > Why do we do autobind there, anyway, and why is it conditional on
>> >>> > SOCK_PASSCRED? Note that e.g. for SOCK_STREAM we can bloody well get
>> >>> > to sending stuff without autobind ever done - just use socketpair()
>> >>> > to create that sucker and we won't be going through the connect()
>> >>> > at all.
>> >>>
>> >>> In the case Dmitry reported, unix_dgram_sendmsg() calls unix_autobind(),
>> >>> not SOCK_STREAM.
>> >>
>> >> Yes, I've noticed. What I'm asking is what in there needs autobind triggered
>> >> on sendmsg and why doesn't the same need affect the SOCK_STREAM case?
>> >>
>> >>> I guess some lock, perhaps the u->bindlock could be dropped before
>> >>> acquiring the next one (sb_writer), but I need to double check.
>> >>
>> >> Bad idea, IMO - do you *want* autobind being able to come through while
>> >> bind(2) is busy with mknod?
>> >
>> >
>> > Ping. This is still happening on HEAD.
>> >
>> Thanks for your reminder. Mind to give the attached patch (compile only)
>> a try? I take another approach to fix this deadlock, which moves the
>> unix_mknod() out of unix->bindlock. Not sure if there is any unexpected
>> impact with this way.
> I don't think this is the right approach.
> Currently the file creation is potponed until unix_bind can no longer
> fail otherwise. With it reordered, it may be someone races you with a
> different path and now you are left with a file to clean up. Except it
> is quite unclear for me if you can unlink it.

What races do you mean here? If you mean someone could get a
refcount of that file, it could happen no matter we have bindlock or not
since it is visible once created. The filesystem layer should take care of
the file refcount so all we need to do here is calling path_put() as in my
patch. Or if you mean two threads calling unix_bind() could race without
binlock, only one of them should succeed the other one just fails out.