Re: [RFC] tcp: race in receive part

Jiri Olsa a écrit :

Hi,

in RHEL4 we can see a race in the tcp layer. We were not able to reproduce 
this on the upstream kernel, but since the issue occurs very rarelly
(once per 8 days), we just might not be lucky.

I'm affraid this might be a long email, I'll try to structure it nicely.. :)

Thanks for your mail and detailed analysis

RACE DESCRIPTION
================

There's a nice pdf describing the issue (and sollution using locks) on
https://bugzilla.redhat.com/attachment.cgi?id=345014

I could not reach this url unfortunatly 

--> "You are not authorized to access bug #494404. "

The race fires, when following code paths meet, and the tp->rcv_nxt and
__add_wait_queue updates stay in CPU caches.

CPU1                         CPU2


sys_select                   receive packet
  ...                        ...
  __add_wait_queue           update tp->rcv_nxt
  ...                        ...
  tp->rcv_nxt check          sock_def_readable
  ...                        {
  schedule                      ...
                                if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
                                        wake_up_interruptible(sk->sk_sleep)
                                ...
                             }

If there were no cache the code would work ok, since the wait_queue and
rcv_nxt are opposit to each other.

Meaning that once tp->rcv_nxt is updated by CPU2, the CPU1 either already
passed the tp->rcv_nxt check and sleeps, or will get the new value for
tp->rcv_nxt and will return with new data mask.  
In both cases the process (CPU1) is being added to the wait queue, so the
waitqueue_active (CPU2) call cannot miss and will wake up CPU1.

The bad case is when the __add_wait_queue changes done by CPU1 stay in its
cache , and so does the tp->rcv_nxt update on CPU2 side.  The CPU1 will then
endup calling schedule and sleep forever if there are no more data on the
socket.

Adding smp_mb() calls before sock_def_readable call and after __add_wait_queue
should prevent the above bad scenario.

The upstream patch is attached. It seems to prevent the issue.



CPU BUGS
========

The customer has been able to reproduce this problem only on one CPU model:
Xeon E5345*2. They didn't reproduce on XEON MV, for example.

Is there an easy way to reproduce the problem ?

That CPU model happens to have 2 possible issues, that might cause the issue:
(see errata http://www.intel.com/Assets/PDF/specupdate/315338.pdf)

AJ39 and AJ18. The first one can be workarounded by BIOS upgrade,
the other one has following notes:

AJ18 only matters on unaligned accesses, tcp code doesnt do this.

poll_wait() calls add_wait_queue() which contains a 
spin_lock_irqsave()/spin_unlock_irqrestore() pair

Documentation/memory-barriers.txt states in line 1123 :

	Memory operations issued after the LOCK will be completed after the LOCK
	operation has completed.

and line 1131 states :

	Memory operations issued before the UNLOCK will be completed before the
	UNLOCK operation has completed.

So yes, there is no full smp_mb() in poll_wait()

+
+	/* Get in sync with tcp_data_queue, tcp_urg
+	   and tcp_rcv_established function. */
+	smp_mb();

If this barrier is really necessary, I guess it should be done in poll_wait() itself.

Documentation/memory-barriers.txt misses some information about poll_wait()

Oh well... if smp_mb() is needed, I believe it should be done
right before "if (waitqueue_active(sk->sk_sleep) ... "

 	read_lock(&sk->sk_callback_lock);
+	smp_mb();
 	if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
 		wake_up_interruptible(sk->sk_sleep)

It would match other parts in kernel (see fs/splice.c, fs/aio.c, ...)

Strange thing is that read_lock() on x86 is a full memory barrier, as it uses
"lock subl $0x1,(%eax)"

Maybe we could define a smp_mb_after_read_lock()  (a compiler barrier() on x86)