nvmet: don't check iosqes,iocqes for discovery controllers
authorSagi Grimberg <sagi@grimberg.me>
Mon, 15 Mar 2021 22:34:51 +0000 (15:34 -0700)
committerChristoph Hellwig <hch@lst.de>
Thu, 18 Mar 2021 04:39:02 +0000 (05:39 +0100)
From the base spec, Figure 78:

  "Controller Configuration, these fields are defined as parameters to
   configure an "I/O Controller (IOC)" and not to configure a "Discovery
   Controller (DC).

   ...
   If the controller does not support I/O queues, then this field shall
   be read-only with a value of 0h

Just perform this check for I/O controllers.

Fixes: a07b4970f464 ("nvmet: add a generic NVMe target")
Reported-by: Belanger, Martin <Martin.Belanger@dell.com>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Chaitanya Kulkarni <chaitanya.kulkarni@wdc.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
drivers/nvme/target/core.c

index be6fcda..a027433 100644 (file)
@@ -1118,9 +1118,20 @@ static void nvmet_start_ctrl(struct nvmet_ctrl *ctrl)
 {
        lockdep_assert_held(&ctrl->lock);
 
-       if (nvmet_cc_iosqes(ctrl->cc) != NVME_NVM_IOSQES ||
-           nvmet_cc_iocqes(ctrl->cc) != NVME_NVM_IOCQES ||
-           nvmet_cc_mps(ctrl->cc) != 0 ||
+       /*
+        * Only I/O controllers should verify iosqes,iocqes.
+        * Strictly speaking, the spec says a discovery controller
+        * should verify iosqes,iocqes are zeroed, however that
+        * would break backwards compatibility, so don't enforce it.
+        */
+       if (ctrl->subsys->type != NVME_NQN_DISC &&
+           (nvmet_cc_iosqes(ctrl->cc) != NVME_NVM_IOSQES ||
+            nvmet_cc_iocqes(ctrl->cc) != NVME_NVM_IOCQES)) {
+               ctrl->csts = NVME_CSTS_CFS;
+               return;
+       }
+
+       if (nvmet_cc_mps(ctrl->cc) != 0 ||
            nvmet_cc_ams(ctrl->cc) != 0 ||
            nvmet_cc_css(ctrl->cc) != 0) {
                ctrl->csts = NVME_CSTS_CFS;