Re: [PATCHv3 3/5] nvme: implement I/O Command Sets Command Set support

[Sagi Grimberg <sagi@grimberg.me>]

On 6/24/20 2:49 PM, Keith Busch wrote:
> On Wed, Jun 24, 2020 at 12:03:41PM -0700, Sagi Grimberg wrote:
>>
>>>>>> If the controller does not support the CNS value, it may return Invalid
>>>>>> Field with DNR set. That error currently gets propogated back to
>>>>>> nvme_init_ns_head(), which then abandons the namespace. Just as the code
>>>>>> coments say, we had been historically been clearing such errors because
>>>>>> we have other ways to identify the namespace, but now we're not clearing
>>>>>> that error.
>>>>>
>>>>> I don't understand what you are saying Keith.
>>>>>
>>>>> My comment was for this block:
>>>>> -- 
>>>>>        if (!status && nvme_multi_css(ctrl) && !csi_seen) {
>>>>>            dev_warn(ctrl->device, "Command set not reported for nsid:%d\n",
>>>>>                 nsid);
>>>>>            status = -EINVAL;
>>>>>        }
>>>>> -- 
>>>>>
>>>>> I was saying that !status doesn't necessarily mean success, but it can
>>>>> also mean that its an retry-able error status (due to transport or
>>>>> controller). If we see a retry-able error we should still clear the
>>>>> status so we don't abandon the namespace.
>>>>>
>>>>> This for example would achieve that:
>>>
>>> We're not talking about the same thing. I am only talking about what
>>> introduced the DNR check, from commit 59c7c3caaaf87.
>>>
>>> I know you added it because you want to abort comparing identifiers on a
>>> rescan when retrieving the identifiers failed. That's fine, but I am
>>> saying this fails namespace creation in the first place for some types
>>> of devices that used to succeed.
>>
>> OK, now I think I understand (thanks for clarifying that the discussion
>> is not on patch 3/5, but rather on 59c7c3caaaf87).
>>
>> So the original proposal was to check NVME_SC_HOST_PATH_ERROR (and now
>> we have NVME_SC_HOST_ABORTED_CMD) but with the review Christoph gave
>> it seemed to make more sense that we generalize and check the DNR bit.
> 
> Okay, I didn't question this approach when it first went through, so
> sorry about this digression, but I really don't get how this DNR check
> helps at all.
> 
> The code currently returns an error if DNR is set.

Right.

> Based on the commit
> messages, it sounds like you need that error to skip comparing
> identifiers through nvme's scan_work calling revalidate_disk():
> suppressing the error has revalidate_disk() fail with -ENODEV when
> comparing identifiers fails.

Your understanding is correct.

> Since we do return the error when DNR is set, we skip comparing
> identifiers and return blk_status_to_errno(nvme_error_status(ret))
> instead. How is that errno an improvement?

See nvme_revalidate_disk:
out:
         /*
          * Only fail the function if we got a fatal error back from the
          * device, otherwise ignore the error and just move on.
          */
         if (ret == -ENOMEM || (ret > 0 && !(ret & NVME_SC_DNR)))
                 ret = 0;
         else if (ret > 0)
                 ret = blk_status_to_errno(nvme_error_status(ret));
         return ret;

So we don't actually propagate the error back if its a non-DNR (hence
retry-able error). The errno was there before in order to not leak NVMe
errors to the block layer.

> And then if DNR is not set, we suppress the error and proceed with
> comparing identifiers??

Wait, I think that I re-read it it's backwards. The intent was to indeed
clear the error for the DNR case and propagate the error for the non-DNR
case!

The code needs to be:
--
diff --git a/drivers/nvme/host/core.c b/drivers/nvme/host/core.c
index 2afed32d3892..3e84ab6c2bd3 100644
--- a/drivers/nvme/host/core.c
+++ b/drivers/nvme/host/core.c
@@ -1130,7 +1130,7 @@ static int nvme_identify_ns_descs(struct nvme_ctrl 
*ctrl, unsigned nsid,
                   * Don't treat an error as fatal, as we potentially 
already
                   * have a NGUID or EUI-64.
                   */
-               if (status > 0 && !(status & NVME_SC_DNR))
+               if (status > 0 && (status & NVME_SC_DNR))
                         status = 0;
                 goto free_data;
         }
--

This way, if the controller failed the identify it will be with DNR
status and we will silently ignore, and if the transport failed its
a non-DNR status, and we propagate the status back, skip the id compare,
and then silently ignore the error in nvme_revalidate_disk (as above).

Looking into the original fix we had internally, this was the case, and
got fat-fingered in I can only assume...

Will send a fix right away, good catch keith!