Commit 4d5f288c authored by Ben Walker's avatar Ben Walker
Browse files

nvmf/rdma: Fix double complete when RNIC goes offline 



A request could be completed twice, once for an error
on an IBV_SEND operation and again on an outstanding
IBV_RDMA_WRITE operation, if the RNIC goes offline
while a complete + data transfer are occurring.

This fixes GitHub issue #414

Change-Id: I2338b4d4582c5ee2512cfbd1e89048a10d3ecf1c
Signed-off-by: default avatarBen Walker <benjamin.walker@intel.com>
Reviewed-on: https://review.gerrithub.io/425646


Chandler-Test-Pool: SPDK Automated Test System <sys_sgsw@intel.com>
Tested-by: default avatarSPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: default avatarJim Harris <james.r.harris@intel.com>
Reviewed-by: default avatarShuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com>
parent e7988759

lib/nvmf/rdma.c

+8 −2
Original line number Diff line number Diff line
@@ -2670,8 +2670,6 @@ spdk_nvmf_rdma_poller_poll(struct spdk_nvmf_rdma_transport *rtransport, 


			switch (wc[i].opcode) {

			case IBV_WC_SEND:

			case IBV_WC_RDMA_WRITE:

			case IBV_WC_RDMA_READ:

				rdma_req = get_rdma_req_from_wc(&wc[i]);

				rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair);
@@ -2688,6 +2686,14 @@ spdk_nvmf_rdma_poller_poll(struct spdk_nvmf_rdma_transport *rtransport, 
				 * the queue pair disconnects or recovers. */

				TAILQ_INSERT_TAIL(&rqpair->incoming_queue, rdma_recv, link);

				break;

			case IBV_WC_RDMA_WRITE:

			case IBV_WC_RDMA_READ:

				/* If the data transfer fails still force the queue into the error state,

				 * but the rdma_req objects should only be manipulated in response to

				 * SEND and RECV operations. */

				rdma_req = get_rdma_req_from_wc(&wc[i]);

				rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair);

				break;

			default:

				SPDK_ERRLOG("Received an unknown opcode on the CQ: %d\n", wc[i].opcode);

				continue;