Commit 8f9167af authored by Tomasz Zawadzki's avatar Tomasz Zawadzki Committed by Tomasz Zawadzki
Browse files

nvme/tcp: check for I/O timeout in poll group completions 



Single qpair completions already checks for timeouts on
a qpair, meanwhile poll group completions were missing it.

To be sure we'd have to check all connected qpairs with
outstanding_reqs on them. Since this is a hot path,
instead create a new list containing of only qpairs
that have outstanding I/O and their controller has timeout_enabled.

This mimics similar behaviour as in nvme_rdma.

Change-Id: Ib851f3501b89808d56b2caea83ca6f03d4c1ce6e
Signed-off-by: default avatarTomasz Zawadzki <tomasz.zawadzki@nutanix.com>
Reviewed-on: https://review.spdk.io/c/spdk/spdk/+/25758


Reviewed-by: default avatarShuhei Matsumoto <smatsumoto@nvidia.com>
Tested-by: default avatarSPDK Automated Test System <spdkbot@gmail.com>
Reviewed-by: default avatarJim Harris <jim.harris@nvidia.com>
Reviewed-by: default avatarKonrad Sztyber <ksztyber@nvidia.com>
parent 161afee9

lib/nvme/nvme_tcp.c

+33 −0
Original line number Diff line number Diff line
@@ -68,6 +68,7 @@ struct nvme_tcp_poll_group { 
	int64_t num_completions;



	TAILQ_HEAD(, nvme_tcp_qpair) needs_poll;

	TAILQ_HEAD(, nvme_tcp_qpair) timeout_enabled;

	struct spdk_nvme_tcp_stat stats;

};
@@ -110,6 +111,8 @@ struct nvme_tcp_qpair { 


	TAILQ_ENTRY(nvme_tcp_qpair)		link_poll;



	TAILQ_ENTRY(nvme_tcp_qpair)		link_timeout;



	uint64_t				icreq_timeout_tsc;



	bool					shared_stats;
@@ -981,6 +984,15 @@ nvme_tcp_qpair_submit_request(struct spdk_nvme_qpair *qpair, 
			  (uint32_t)req->cmd.cid, (uint32_t)req->cmd.opc,

			  req->cmd.cdw10, req->cmd.cdw11, req->cmd.cdw12, tqpair->qpair.queue_depth);

	TAILQ_INSERT_TAIL(&tqpair->outstanding_reqs, tcp_req, link);



	if (TAILQ_ENTRY_NOT_ENQUEUED(tqpair, link_timeout) && qpair->poll_group != NULL &&

	    qpair->ctrlr->timeout_enabled) {

		struct nvme_tcp_poll_group *tgroup;



		tgroup = nvme_tcp_poll_group(qpair->poll_group);

		TAILQ_INSERT_TAIL(&tgroup->timeout_enabled, tqpair, link_timeout);

	}



	return nvme_tcp_qpair_capsule_cmd_send(tqpair, tcp_req);

}
@@ -1030,6 +1042,17 @@ nvme_tcp_req_complete(struct nvme_tcp_req *tcp_req, 
	spdk_trace_record(TRACE_NVME_TCP_COMPLETE, qpair->id, 0, (uintptr_t)tcp_req->pdu, req->cb_arg,

			  (uint32_t)req->cmd.cid, (uint32_t)cpl.status_raw, qpair->queue_depth);

	TAILQ_REMOVE(&tqpair->outstanding_reqs, tcp_req, link);



	if (TAILQ_EMPTY(&tqpair->outstanding_reqs) && qpair->poll_group != NULL &&

	    TAILQ_ENTRY_ENQUEUED(tqpair, link_timeout)) {

		struct nvme_tcp_poll_group *tgroup;



		assert(qpair->ctrlr->timeout_enabled);



		tgroup = nvme_tcp_poll_group(qpair->poll_group);

		TAILQ_REMOVE_CLEAR(&tgroup->timeout_enabled, tqpair, link_timeout);

	}



	nvme_tcp_req_put(tqpair, tcp_req);

	nvme_complete_request(req->cb_fn, req->cb_arg, req->qpair, req, &cpl);

}
@@ -2715,6 +2738,7 @@ nvme_tcp_poll_group_create(void) 
	}



	TAILQ_INIT(&group->needs_poll);

	TAILQ_INIT(&group->timeout_enabled);



	group->sock_group = spdk_sock_group_create(group);

	if (group->sock_group == NULL) {
@@ -2806,6 +2830,9 @@ nvme_tcp_poll_group_remove(struct spdk_nvme_transport_poll_group *tgroup, 
	if (TAILQ_ENTRY_ENQUEUED(tqpair, link_poll)) {

		TAILQ_REMOVE_CLEAR(&group->needs_poll, tqpair, link_poll);

	}

	if (TAILQ_ENTRY_ENQUEUED(tqpair, link_timeout)) {

		TAILQ_REMOVE_CLEAR(&group->timeout_enabled, tqpair, link_timeout);

	}



	return 0;

}
@@ -2846,6 +2873,12 @@ nvme_tcp_poll_group_process_completions(struct spdk_nvme_transport_poll_group *t 
		nvme_tcp_qpair_sock_cb(&tqpair->qpair, group->sock_group, tqpair->sock);

	}



	TAILQ_FOREACH_SAFE(tqpair, &group->timeout_enabled, link_timeout, tmp_tqpair) {

		qpair = &tqpair->qpair;

		assert(qpair->ctrlr->timeout_enabled);

		nvme_tcp_qpair_check_timeout(qpair);

	}



	if (spdk_unlikely(num_events < 0)) {

		return num_events;

	}