nvme-fc: on lldd/transport io error, terminate association

Per FC-NVME, when lldd or transport detects an i/o error, the
connection must be terminated, which in turn requires the association
to be termianted.  Currently the transport simply creates a nvme
completion status of transport error and returns the io. The FC-NVME
spec makes the mandate as initiator and host, depending on the error,
can get out of sync on outstanding io counts (sqhd/sqtail).

Implement the association teardown on lldd or transport detected
errors.

Signed-off-by: James Smart <james.smart@broadcom.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>

diff --git a/drivers/nvme/host/fc.c b/drivers/nvme/host/fc.c
index 5b14cbefb7240d5e7d50bb1ade8fd958417282e8..2edae54688e88d88f2b54b7239b20a5ff9c14953 100644 (file)
--- a/drivers/nvme/host/fc.c
+++ b/drivers/nvme/host/fc.c
@@ -1139,6 +1139,7 @@ nvme_fc_xmt_disconnect_assoc(struct nvme_fc_ctrl *ctrl)
 /* *********************** NVME Ctrl Routines **************************** */
 
 static void __nvme_fc_final_op_cleanup(struct request *rq);
+static void nvme_fc_error_recovery(struct nvme_fc_ctrl *ctrl, char *errmsg);
 
 static int
 nvme_fc_reinit_request(void *data, struct request *rq)
@@ -1265,7 +1266,7 @@ nvme_fc_fcpio_done(struct nvmefc_fcp_req *req)
        struct nvme_command *sqe = &op->cmd_iu.sqe;
        __le16 status = cpu_to_le16(NVME_SC_SUCCESS << 1);
        union nvme_result result;
-       bool complete_rq;
+       bool complete_rq, terminate_assoc = true;
 
        /*
         * WARNING:
@@ -1294,6 +1295,14 @@ nvme_fc_fcpio_done(struct nvmefc_fcp_req *req)
         * fabricate a CQE, the following fields will not be set as they
         * are not referenced:
         *      cqe.sqid,  cqe.sqhd,  cqe.command_id
+        *
+        * Failure or error of an individual i/o, in a transport
+        * detected fashion unrelated to the nvme completion status,
+        * potentially cause the initiator and target sides to get out
+        * of sync on SQ head/tail (aka outstanding io count allowed).
+        * Per FC-NVME spec, failure of an individual command requires
+        * the connection to be terminated, which in turn requires the
+        * association to be terminated.
         */
 
        fc_dma_sync_single_for_cpu(ctrl->lport->dev, op->fcp_req.rspdma,
@@ -1359,6 +1368,8 @@ nvme_fc_fcpio_done(struct nvmefc_fcp_req *req)
                goto done;
        }
 
+       terminate_assoc = false;
+
 done:
        if (op->flags & FCOP_FLAGS_AEN) {
                nvme_complete_async_event(&queue->ctrl->ctrl, status, &result);
@@ -1366,7 +1377,7 @@ done:
                atomic_set(&op->state, FCPOP_STATE_IDLE);
                op->flags = FCOP_FLAGS_AEN;     /* clear other flags */
                nvme_fc_ctrl_put(ctrl);
-               return;
+               goto check_error;
        }
 
        complete_rq = __nvme_fc_fcpop_chk_teardowns(ctrl, op);
@@ -1379,6 +1390,10 @@ done:
                nvme_end_request(rq, status, result);
        } else
                __nvme_fc_final_op_cleanup(rq);
+
+check_error:
+       if (terminate_assoc)
+               nvme_fc_error_recovery(ctrl, "transport detected io error");
 }
 
 static int