Re: identify the race condition in this code and win the respect of linux-rdma developers!

On Fri, Sep 16, 2011 at 07:36:17PM +0000, Hefty, Sean wrote:

I was suggesting that there were only 2 states, not 3, with
ibv_get_cq_event() simply retrieving a queued event from the kernel
without touching the CQ state.

But that isn't true, as you point out in your next message
ibv_get_cq_event calls mlx4_cq_event which does alter the CQ state,
particularly in how it behaves WRT to mlx4_arm_cq.

Hard to say what the arm_sn does, but my guess is that it is creating
a one-shot that is sensitive to ibv_get_cq_event, eg you will never
get more than one event queued in the kernel queue, which creates the
3 state system I described.. (at least that matches my experience)

Certainly, considering that arm_sn is a 3 bit counter, calling
mlx4_cq_event many times then calling arm_cq will alias the counter
from the point of view of the chip and that can't possibly be good.

I changed the code to do this:

	do {
		poll
		rearm
		poll
		get_event
	} while (1);

which I believe fixes the issue.

In the past I looked at all this and concluded this was the only
possible correct way to write the loop. You have to guarentee rearm is
called before every get_event and you have to do a poll after every
rearm.

Thanks for the help.  You are now the winner of my respect.  :)

NP :)

Jason
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html