[PATCH] nvme-core: remove head->effects to fix use-after-free

[Keith Busch]

On Fri, Nov 17, 2023 at 02:28:46PM +0100, Christoph Hellwig wrote:
> On Wed, Nov 15, 2023 at 10:52:01PM -0500, Keith Busch wrote:
> > 
> > Yes, in section 5.16.1.6, "Commands Supported and Effects":
> > 
> >   This log page is used to describe the commands that the controller
> >   supports and the effects of those commands on the state of the NVM
> >   subsystem.
> > 
> > Oddly enough, Figure 202 says the scope of the log page is "Controller"
> > rather than "Subsystem". Sounds like ECN potential. You can memcmp the
> > effects log from each controller for a sanity check if you think some
> > subsystem controllers messed that up.
> 
> If we really want to be 111% sure we could read the effects for all
> controllers and do a logical OR of them, but I think the reason for the
> per-controller scope is that for odd subsystems where different
> controllers don't actually access the same namespaces these flags
> could be different, i.e. one that only does KV, one that does ZNS,
> one that does NVM and one that is just an administrative controller.

The effects log is per-CSI so different command sets won't create
conflicts.

Namespaces that are not shared don't really matter here because this
problem is unique to mulitpath.

It doesn't make sense for effects logs to be different per-controller
for the same shared namespace. The spec doesn't seem to explicitly
prevent that, but hints that all hosts should be seeing the same thing
no matter which controller they're connected to:

"
  If the namespace is attached to multiple controllers, the host(s)
  associated with those controllers should coordinate their commands to
  meet the Command Submission and Execution requirements
"

That couldn't be a reliable suggestion if the hosts observe diverging
effects. For the controllers that a host does see, though, I agree we
should use the most cautious effects reported with logical OR of them.