nvme-fc: fix io timeout to abort I/O
authorJames Smart <james.smart@broadcom.com>
Fri, 16 Oct 2020 21:06:27 +0000 (14:06 -0700)
committerChristoph Hellwig <hch@lst.de>
Fri, 23 Oct 2020 10:52:16 +0000 (12:52 +0200)
Currently, an I/O timeout unconditionally invokes
nvme_fc_error_recovery() which checks for LIVE or CONNECTING state.  If
live, the routine resets the controller which initiates a reconnect -
which is valid.  If CONNECTING, err_work is scheduled.  Err_work then
calls the terminate_io routine, which also checks for CONNECTING and
noops any further action on outstanding I/O.  The result is nothing
happened to the timed out io.  As such, if the command was dropped on
the wire, it will never timeout / complete, and the connect process
will hang.

Change the behavior of the io timeout routine to unconditionally abort
the I/O.  I/O completion handling will note that an io failed due to an
abort and will terminate the connection / association as needed.  If the
abort was unable to happen, continue with a call to
nvme_fc_error_recovery(). To ensure something different happens in
nvme_fc_error_recovery() rework it so at it will abort all I/Os on the
association to force a failure.

As I/O aborts now may occur outside of delete_association, counting for
completion must be wary and only count those aborted during
delete_association when TERMIO is set on the controller.

Signed-off-by: James Smart <james.smart@broadcom.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
drivers/nvme/host/fc.c

index e2e09e2..3e72b7d 100644 (file)
@@ -1837,8 +1837,10 @@ __nvme_fc_abort_op(struct nvme_fc_ctrl *ctrl, struct nvme_fc_fcp_op *op)
        opstate = atomic_xchg(&op->state, FCPOP_STATE_ABORTED);
        if (opstate != FCPOP_STATE_ACTIVE)
                atomic_set(&op->state, opstate);
-       else if (test_bit(FCCTRL_TERMIO, &ctrl->flags))
+       else if (test_bit(FCCTRL_TERMIO, &ctrl->flags)) {
+               op->flags |= FCOP_FLAGS_TERMIO;
                ctrl->iocnt++;
+       }
        spin_unlock_irqrestore(&ctrl->lock, flags);
 
        if (opstate != FCPOP_STATE_ACTIVE)
@@ -1874,7 +1876,8 @@ __nvme_fc_fcpop_chk_teardowns(struct nvme_fc_ctrl *ctrl,
 
        if (opstate == FCPOP_STATE_ABORTED) {
                spin_lock_irqsave(&ctrl->lock, flags);
-               if (test_bit(FCCTRL_TERMIO, &ctrl->flags)) {
+               if (test_bit(FCCTRL_TERMIO, &ctrl->flags) &&
+                   op->flags & FCOP_FLAGS_TERMIO) {
                        if (!--ctrl->iocnt)
                                wake_up(&ctrl->ioabort_wait);
                }
@@ -2446,15 +2449,20 @@ nvme_fc_timeout(struct request *rq, bool reserved)
 {
        struct nvme_fc_fcp_op *op = blk_mq_rq_to_pdu(rq);
        struct nvme_fc_ctrl *ctrl = op->ctrl;
+       struct nvme_fc_cmd_iu *cmdiu = &op->cmd_iu;
+       struct nvme_command *sqe = &cmdiu->sqe;
 
        /*
-        * we can't individually ABTS an io without affecting the queue,
-        * thus killing the queue, and thus the association.
-        * So resolve by performing a controller reset, which will stop
-        * the host/io stack, terminate the association on the link,
-        * and recreate an association on the link.
+        * Attempt to abort the offending command. Command completion
+        * will detect the aborted io and will fail the connection.
         */
-       nvme_fc_error_recovery(ctrl, "io timeout error");
+       dev_info(ctrl->ctrl.device,
+               "NVME-FC{%d.%d}: io timeout: opcode %d fctype %d w10/11: "
+               "x%08x/x%08x\n",
+               ctrl->cnum, op->queue->qnum, sqe->common.opcode,
+               sqe->connect.fctype, sqe->common.cdw10, sqe->common.cdw11);
+       if (__nvme_fc_abort_op(ctrl, op))
+               nvme_fc_error_recovery(ctrl, "io timeout abort failed");
 
        /*
         * the io abort has been initiated. Have the reset timer
@@ -2726,6 +2734,7 @@ nvme_fc_complete_rq(struct request *rq)
        struct nvme_fc_ctrl *ctrl = op->ctrl;
 
        atomic_set(&op->state, FCPOP_STATE_IDLE);
+       op->flags &= ~FCOP_FLAGS_TERMIO;
 
        nvme_fc_unmap_data(ctrl, rq, op);
        nvme_complete_rq(rq);
@@ -3090,26 +3099,19 @@ out_free_queue:
        return ret;
 }
 
+
 /*
- * This routine stops operation of the controller on the host side.
- * On the host os stack side: Admin and IO queues are stopped,
- *   outstanding ios on them terminated via FC ABTS.
- * On the link side: the association is terminated.
+ * This routine runs through all outstanding commands on the association
+ * and aborts them.  This routine is typically be called by the
+ * delete_association routine. It is also called due to an error during
+ * reconnect. In that scenario, it is most likely a command that initializes
+ * the controller, including fabric Connect commands on io queues, that
+ * may have timed out or failed thus the io must be killed for the connect
+ * thread to see the error.
  */
 static void
-nvme_fc_delete_association(struct nvme_fc_ctrl *ctrl)
+__nvme_fc_abort_outstanding_ios(struct nvme_fc_ctrl *ctrl, bool start_queues)
 {
-       struct nvmefc_ls_rcv_op *disls = NULL;
-       unsigned long flags;
-
-       if (!test_and_clear_bit(ASSOC_ACTIVE, &ctrl->flags))
-               return;
-
-       spin_lock_irqsave(&ctrl->lock, flags);
-       set_bit(FCCTRL_TERMIO, &ctrl->flags);
-       ctrl->iocnt = 0;
-       spin_unlock_irqrestore(&ctrl->lock, flags);
-
        /*
         * If io queues are present, stop them and terminate all outstanding
         * ios on them. As FC allocates FC exchange for each io, the
@@ -3127,6 +3129,8 @@ nvme_fc_delete_association(struct nvme_fc_ctrl *ctrl)
                blk_mq_tagset_busy_iter(&ctrl->tag_set,
                                nvme_fc_terminate_exchange, &ctrl->ctrl);
                blk_mq_tagset_wait_completed_request(&ctrl->tag_set);
+               if (start_queues)
+                       nvme_start_queues(&ctrl->ctrl);
        }
 
        /*
@@ -3143,13 +3147,34 @@ nvme_fc_delete_association(struct nvme_fc_ctrl *ctrl)
 
        /*
         * clean up the admin queue. Same thing as above.
-        * use blk_mq_tagset_busy_itr() and the transport routine to
-        * terminate the exchanges.
         */
        blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
        blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
                                nvme_fc_terminate_exchange, &ctrl->ctrl);
        blk_mq_tagset_wait_completed_request(&ctrl->admin_tag_set);
+}
+
+/*
+ * This routine stops operation of the controller on the host side.
+ * On the host os stack side: Admin and IO queues are stopped,
+ *   outstanding ios on them terminated via FC ABTS.
+ * On the link side: the association is terminated.
+ */
+static void
+nvme_fc_delete_association(struct nvme_fc_ctrl *ctrl)
+{
+       struct nvmefc_ls_rcv_op *disls = NULL;
+       unsigned long flags;
+
+       if (!test_and_clear_bit(ASSOC_ACTIVE, &ctrl->flags))
+               return;
+
+       spin_lock_irqsave(&ctrl->lock, flags);
+       set_bit(FCCTRL_TERMIO, &ctrl->flags);
+       ctrl->iocnt = 0;
+       spin_unlock_irqrestore(&ctrl->lock, flags);
+
+       __nvme_fc_abort_outstanding_ios(ctrl, false);
 
        /* kill the aens as they are a separate path */
        nvme_fc_abort_aen_ops(ctrl);
@@ -3263,22 +3288,27 @@ static void
 __nvme_fc_terminate_io(struct nvme_fc_ctrl *ctrl)
 {
        /*
-        * if state is connecting - the error occurred as part of a
-        * reconnect attempt. The create_association error paths will
-        * clean up any outstanding io.
-        *
-        * if it's a different state - ensure all pending io is
-        * terminated. Given this can delay while waiting for the
-        * aborted io to return, we recheck adapter state below
-        * before changing state.
+        * if state is CONNECTING - the error occurred as part of a
+        * reconnect attempt. Abort any ios on the association and
+        * let the create_association error paths resolve things.
         */
-       if (ctrl->ctrl.state != NVME_CTRL_CONNECTING) {
-               nvme_stop_keep_alive(&ctrl->ctrl);
-
-               /* will block will waiting for io to terminate */
-               nvme_fc_delete_association(ctrl);
+       if (ctrl->ctrl.state == NVME_CTRL_CONNECTING) {
+               __nvme_fc_abort_outstanding_ios(ctrl, true);
+               return;
        }
 
+       /*
+        * For any other state, kill the association. As this routine
+        * is a common io abort routine for resetting and such, after
+        * the association is terminated, ensure that the state is set
+        * to CONNECTING.
+        */
+
+       nvme_stop_keep_alive(&ctrl->ctrl);
+
+       /* will block will waiting for io to terminate */
+       nvme_fc_delete_association(ctrl);
+
        if (ctrl->ctrl.state != NVME_CTRL_CONNECTING &&
            !nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING))
                dev_err(ctrl->ctrl.device,