Re: [PATCHv2 05/17] nvme: add Clang context annotations for nvme_ns_head::current_path

[Nilay Shroff]

On 6/29/26 11:30 AM, Marco Elver wrote:
> On Sun, 28 Jun 2026 at 18:07, Paul E. McKenney <paulmck@xxxxxxxxxx> wrote:
> > On Sun, Jun 28, 2026 at 08:00:00AM +0200, Marco Elver wrote:
> > > On Sat, 27 Jun 2026 at 19:14, Paul E. McKenney <paulmck@xxxxxxxxxx> wrote:
> > > > On Sat, Jun 27, 2026 at 05:38:44PM +0200, Marco Elver wrote:
> > > > > On Fri, 26 Jun 2026 at 20:36, Paul E. McKenney <paulmck@xxxxxxxxxx> wrote:
> > > > > > On Fri, Jun 26, 2026 at 08:40:50AM +0200, Christoph Hellwig wrote:
> > > > > > > On Sun, Jun 14, 2026 at 06:45:20PM +0530, Nilay Shroff wrote:
> > > > > > > > +++ b/drivers/nvme/host/multipath.c
> > > > > > > > @@ -231,8 +231,16 @@ bool nvme_mpath_clear_current_path(struct nvme_ns *ns)
> > > > > > > >      bool changed = false;
> > > > > > > >      int node;
> > > > > > > >
> > > > > > > > +   /*
> > > > > > > > +    * This helper is used by namespace failover/teardown and I/O policy
> > > > > > > > +    * update paths. We only compare the head->current_path[] pointer value
> > > > > > > > +    * and do not dereference the referenced namespace, so suppress the
> > > > > > > > +    * context analysis warning for this lockless inspection of the
> > > > > > > > +    * __rcu_guarded pointer.
> > > > > > > > +    */
> > > > > > > >      for_each_node(node) {
> > > > > > > > -           if (ns == rcu_access_pointer(head->current_path[node])) {
> > > > > > > > +           if (context_unsafe(ns ==
> > > > > > > > +                           rcu_access_pointer(head->current_path[node]))) {
> > > > > > >
> > > > > > > I think we need a helper for this, as for a simple pointer value
> > > > > > > comparison without a dereference we don't really need either
> > > > > > > rcu_access_pointer nor locking.
> > > > > > >
> > > > > > > Maybe somthing like a
> > > > > > >
> > > > > > >        rcu_compare_pointer(rcu_pointer, nonrcu_pointer)
> > > > > > >
> > > > > > > ?
> > > > > >
> > > > > > We could provide something like this:
> > > > > >
> > > > > > #define rcu_compare_pointer(rcu_pointer, nonrcu_pointer) \
> > > > > >          context_unsafe(rcu_access_pointer(rcu_pointer) == (nonrcu_pointer))
> > > > > >
> > > > > > Or maybe rcu_pointer_equals()?
> > > > > >
> > > > > > Marco, thoughts?
> > > > >
> > > > > I see 2 options:
> > > > >
> > > > > 1. rcu_compare_pointer / rcu_pointer_equals as proposed. It's more
> > > > > explicit but does add some friction during Context Analysis
> > > > > enablement. But this only makes sense if there are cases where
> > > > > rcu_access_pointer() should be used under the RCU reader lock, which
> > > > > led me to the next suggestion...
> > > > >
> > > > > 2. Redefine rcu_access_pointer() to just not require the RCU read-side
> > > > > lock to be held as:
> > > > >
> > > > > > #define rcu_access_pointer(p) context_unsafe(__rcu_access_pointer((p), __UNIQUE_ID(rcu), __rcu))
> > > > >
> > > > > [ Or alternatively wrap __rcu_access_pointer() in context_unsafe()
> > > > >    (similar to rcu_assign_pointer and friends). ]
> > > > >
> > > > > I think rcu_access_pointer() is in the same category as the other RCU
> > > > > pointer helpers, although currently only the pointer update helpers
> > > > > imply context_unsafe(); I think it might have been an oversight
> > > > > initially, and we should change rcu_access_pointer() to match. There
> > > > > should be no reason why an rcu_access_pointer() should be protected by
> > > > > the RCU read-side critical section, since it's not meant for
> > > > > dereferencing the pointer (that'd be a bug; its documentation also
> > > > > says "Return the value of the specified RCU-protected pointer, but
> > > > > omit the lockdep checks for being in an RCU read-side critical section
> > > > > [...]").
> > > > >
> > > > > If you agree there should be no cases where an rcu_access_pointer()
> > > > > should be guarded by an RCU read-side critical section, then I think
> > > > > this is the simplest and correct design, and avoids expanding the RCU
> > > > > API.
> > > >
> > > > I don't know of any uses of rcu_access_pointer() within an RCU read-side
> > > > critical section, but code in a function that might be called both within
> > > > and outside of a critical section might well use rcu_access_pointer().
> > > > In other words, it should be OK to use rcu_access_pointer() within an
> > > > RCU read-side critical section as well as outside of one.
> > >
> > > Thanks, yes, that's what I wanted to confirm; i.e. it's ok to use
> > > rcu_access_pointer() within and outside an RCU read-side critical
> > > section.
> > >
> > > In which case, my proposal (2) to make rcu_access_pointer() not warn
> > > on accessing __rcu_guarded pointers outside an RCU read-side critical
> > > section should be the simpler and more generic change (vs. adding a
> > > new rcu_pointer_equals() helper).
> >
> > As shown below?
> >
> > My guess is that this change makes it unnecessary to have a separate
> > RCU-protected-pointer comparison macro, but please let me know if I am
> > missing something.
>
> Looks good to me. Yes, this makes a new comparison macro unnecessary.
>
> Nilay, does this work and simplify the patch for you?

Yeah, this work for me. In fact, I have already tested this patch from
Paul and sent my Tested-by tag[1]. You may have missed it.

[1] https://lore.kernel.org/all/2a3c6a56-a9a6-461e-80ca-c0f2b4203bff@xxxxxxxxxxxxx/

Thanks,
--Nilay