RE: [dm-devel] patch to dm-emc.c

[Edward Goggin <egoggin emc com>]

On Thursday, August 10, 2006 5:28 PM, Mike Christie wrote
 
> But I thought, a request should eventually complete and so we 
> should be
> woken up from get_request_wait, and we will try allocate a 
> request again
> and then the write request count should have gone down and we 
> should be
> able to proceed.

Mike,

You are right.  My logic was bad ... I was confusing mapped and
target device queues.  Turns out the delays that I am seeing
(multipathd checker thread is stuck waiting for up to 15 minutes
waiting for a free target device queue) are due to ios being
held and timed out by the SCSI LLD (Emulex).  I never waited
more than 10 minutes in my testing earlier in the month before
looking for some type of deadlock.

I've changed the patch to call blk_get_request() with __GFP_WAIT
always from a worker thread context since using pre-allocated
request structures does not help the problem.

Thanks,

Ed