Re: net: BUG in unix_notinflight

On Tue, Mar 7, 2017 at 12:37 AM, Dmitry Vyukov [off-list ref] wrote:

On Mon, Mar 6, 2017 at 11:34 PM, Cong Wang [off-list ref] wrote:

quoted

The problem here is there is no lock protecting concurrent unix_detach_fds()
even though unix_notinflight() is already serialized, if we call
unix_notinflight()
twice on the same file pointer, we trigger this bug...

I don't know what is the right lock here to serialize it.

What exactly here needs to be protected?

1484 static void unix_detach_fds(struct scm_cookie *scm, struct sk_buff *skb)
1485 {
1486         int i;
1487
1488         scm->fp = UNIXCB(skb).fp;
1489         UNIXCB(skb).fp = NULL;
1490
1491         for (i = scm->fp->count-1; i >= 0; i--)
1492                 unix_notinflight(scm->fp->user, scm->fp->fp[i]);
1493 }

Whole unix_notinflight happens under global unix_gc_lock.

Is it that 2 threads call unix_detach_fds for the same skb, and then
call unix_notinflight for the same fd twice?

Not the same skb, but their UNIXCB(skb).fp points to the same place,
therefore we call unix_notinflight() twice on the same fp->user and
fp->fp[i], although we have refcounting but still able to trigger this
warning.