]> www.infradead.org Git - linux-platform-drivers-x86.git/commitdiff
nvme-fc: clear q_live at beginning of association teardown
authorJames Smart <jsmart2021@gmail.com>
Tue, 11 May 2021 04:56:35 +0000 (21:56 -0700)
committerChristoph Hellwig <hch@lst.de>
Wed, 19 May 2021 06:40:24 +0000 (08:40 +0200)
The __nvmf_check_ready() routine used to bounce all filesystem io if the
controller state isn't LIVE.  However, a later patch changed the logic so
that it rejection ends up being based on the Q live check.  The FC
transport has a slightly different sequence from rdma and tcp for
shutting down queues/marking them non-live.  FC marks its queue non-live
after aborting all ios and waiting for their termination, leaving a
rather large window for filesystem io to continue to hit the transport.
Unfortunately this resulted in filesystem I/O or applications seeing I/O
errors.

Change the FC transport to mark the queues non-live at the first sign of
teardown for the association (when I/O is initially terminated).

Fixes: 73a5379937ec ("nvme-fabrics: allow to queue requests for live queues")
Signed-off-by: James Smart <jsmart2021@gmail.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Himanshu Madhani <himanshu.madhani@oracle.com>
Reviewed-by: Hannes Reinecke <hare@suse.de>
Signed-off-by: Christoph Hellwig <hch@lst.de>
drivers/nvme/host/fc.c

index d9ab9e7871d0f7dc4b52e18e4c96091cfeac8733..256e87721a01f25a13aa439cb383adc3e4fa0a0a 100644 (file)
@@ -2461,6 +2461,18 @@ nvme_fc_terminate_exchange(struct request *req, void *data, bool reserved)
 static void
 __nvme_fc_abort_outstanding_ios(struct nvme_fc_ctrl *ctrl, bool start_queues)
 {
+       int q;
+
+       /*
+        * if aborting io, the queues are no longer good, mark them
+        * all as not live.
+        */
+       if (ctrl->ctrl.queue_count > 1) {
+               for (q = 1; q < ctrl->ctrl.queue_count; q++)
+                       clear_bit(NVME_FC_Q_LIVE, &ctrl->queues[q].flags);
+       }
+       clear_bit(NVME_FC_Q_LIVE, &ctrl->queues[0].flags);
+
        /*
         * If io queues are present, stop them and terminate all outstanding
         * ios on them. As FC allocates FC exchange for each io, the