Fix a problem where the blockback driver could run out of requests,
despite the fact that we allocate enough request and reqlist
structures to satisfy the maximum possible number of requests.
The problem was that we were sending responses back to the other
end (blockfront) before freeing resources. The Citrix Windows
driver is pretty agressive about queueing, and would queue more I/O
to us immediately after we sent responses to it. We would run into
a resource shortage and stall out I/O until we freed resources.
It isn't clear whether the request shortage condition was an
indirect cause of the I/O hangs we've been seeing between Windows
with the Citrix PV drivers and FreeBSD's blockback, but the above
problem is certainly a bug.
Sponsored by: Spectra Logic
Submitted by: ken
Reviewed by: royger
dev/xen/blkback/blkback.c:
- Break xbb_send_response() into two sub-functions,
xbb_queue_response() and xbb_push_responses().
Remove xbb_send_response(), because it is no longer
used.
- Adjust xbb_complete_reqlist() so that it calls the
two new functions, and holds the mutex around both
calls. The mutex insures that another context
can't come along and push responses before we've
freed our resources.
- Change xbb_release_reqlist() so that it requires
the mutex to be held instead of acquiring the mutex
itself. Both callers could easily hold the mutex
while calling it, and one really needs to hold the
mutex during the call.
- Add two new counters, accessible via sysctl
variables. The first one counts the number of
I/Os that are queued and waiting to be pushed
(reqs_queued_for_completion). The second one
(reqs_completed_with_error) counts the number of
requests we've completed with an error status.
1ca25732e0