Re: Fwd: Need NVME QUIRK BOGUS for SAMSUNG MZ1WV480HCGL-000MV (Samsung SM-953 Datacenter SSD)

[John Meneghini <jmeneghi@redhat.com>]

Yes, this is what I thought.  This is all the result of the duplicate NID check added to deal with TP4034 Dispersed Namespaces.

One suggestion I have would be to limit this check to nvme-of subsystems only. These are the only devices I am aware of out 
there which support TP4034.  Moreover, all nvme-of devices report a valid NID.  It's required with NVMe Over Fabrics. The PCIe 
devices, I expect, don't care. You don't really need a valid NID with a private namespace - which is what most PCIe devices are.

I'll wager that if you change nvme_global_check_duplicate_ids() to check only nvme-of subsystems, and simply continue with PCIe 
subsystems, 90% of these nvme quirks can be removed.

John Meneghini
Senior Principal Platform Storage Engineer
RHEL SST - Platform Storage Group
jmeneghi@redhat.com

On 7/11/23 13:21, Keith Busch wrote:
> On Tue, Jul 11, 2023 at 09:47:00AM -0700, Linus Torvalds wrote:
>> On Tue, 11 Jul 2023 at 05:06, Christoph Hellwig <hch@lst.de> wrote:
>> For example, we have this completely unacceptable garbage:
>>
>>          ret = nvme_global_check_duplicate_ids(ctrl->subsys, &info->ids);
>>          if (ret) {
>>                  dev_err(ctrl->device,
>>                          "globally duplicate IDs for nsid %d\n", info->nsid);
>>                  nvme_print_device_info(ctrl);
>>                  return ret;
>>          }
>>
>> iow, the code even checks for and *notices* that there are duplicate
>> IDs, and what does it do? It then errors out.
> 
> This check came from a recent half-baked spec feature called "Dispersed
> Namespaces" that caused breakage and data corruption when used in Linux.
> Rather than attempt to support that mostly vendor specific feature, the
> driver attempted to fence that off as unmaintainable. This check wasn't
> aimed at enforcing "correctness", but it certainly found a lot of that
> as collatoral damage. Let's see if we can find a better way to detect
> the difference with a sane fallback as you suggest.
>