Re: [PATCH] nvme: Remove namespace when nvme_identify_ns_descs() failed

[Keith Busch <kbusch@kernel.org>]

On Wed, Dec 04, 2024 at 08:14:08AM +0100, Hannes Reinecke wrote:
> On 12/3/24 20:15, Keith Busch wrote:
> > On Fri, Nov 29, 2024 at 03:06:08PM +0100, Hannes Reinecke wrote:
> > > When a namespace gets unmapped on the target during scanning
> > > nvme_identify_ns_descs() returns with a non-retryable error.
> > > With the currrent code we will ignore that error on the grounds
> > > that we failed to get information, and hence cannot make any
> > > decisions whether to keep or remove that namespace.
> > > But a non-retryable error implies that the namespace is _not_
> > > present as we cannot retry that command and will never get
> > > information about that namespace.
> > > And we need to remove the namespace during scanning, as otherwise
> > > the AEN informing us about a namespace change will find the NSID
> > > present, but nvme_validate_ns() will fail, and the namespace
> > > will never be updated with the correct information.
> > 
> > The scanning only checks namespaces returned in the "active" namespace
> > list. Every namespace not in the active list gets removed already. Why
> > is this unmapped namespace appearing on the active list?
> 
> Timing. Imagine a system used as a backing store for kubernetes, where
> namespaces come and go at a _really_ fast pace.
> 1) AEN triggers a rescan
> 2) List of active namespace is retrieved
> -> NSID A gets unmapped (or moved to another node in the cluster)
> 3) Scan of NSID A returns an error with DNR set.
> Without this patch we keep the namespace around, so eventually we'll
> trip over the 'non-matching UUID' check once the NSID is reused.

I'm still not sure that makes sense. The target shouldn't attach the new
namespace until the host acknowledges the removal of the older NSID via
the Namespace Change List log. Until the log is read, the inventory for
removed namespaces should be latched. Otherwise, timing might remove+add
a specific NSID before the host requests the NS Descriptor for the
racing removal, then it would just get the "non-matching UUID" issue
anyway.