[PATCH v5 20/36] locking/ww_mutex: Support Clang's context analysis

[Marco Elver]

On Fri, Jan 09, 2026 at 12:16PM -0800, Bart Van Assche wrote:
> On 12/19/25 8:40 AM, Marco Elver wrote:
> > Add support for Clang's context analysis for ww_mutex.
> > 
> > The programming model for ww_mutex is subtly more complex than other
> > locking primitives when using ww_acquire_ctx. Encoding the respective
> > pre-conditions for ww_mutex lock/unlock based on ww_acquire_ctx state
> > using Clang's context analysis makes incorrect use of the API harder.
> 
> That's a very short description. It should have been explained in the
> patch description how the ww_acquire_ctx changes affect callers of the
> ww_acquire_{init,done,fini}() functions.

How so? The API is the same (now statically enforced), and there's no
functional change at runtime. Or did I miss something?

> >   static inline void ww_acquire_init(struct ww_acquire_ctx *ctx,
> >   				   struct ww_class *ww_class)
> > +	__acquires(ctx) __no_context_analysis
> > [ ... ]
> >   static inline void ww_acquire_done(struct ww_acquire_ctx *ctx)
> > +	__releases(ctx) __acquires_shared(ctx) __no_context_analysis
> >   {
> > [ ... ]
> >   static inline void ww_acquire_fini(struct ww_acquire_ctx *ctx)
> > +	__releases_shared(ctx) __no_context_analysis
> 
> The above changes make it mandatory to call ww_acquire_done() before
> calling ww_acquire_fini(). In Documentation/locking/ww-mutex-design.rst
> there is an example where there is no ww_acquire_done() call between
> ww_acquire_init() and ww_acquire_fini() (see also line 202).

It might be worth updating the example with what the kernel-doc
documentation recommends (below).

> The
> function dma_resv_lockdep() in drivers/dma-buf/dma-resv.c doesn't call
> ww_acquire_done() at all. Does this mean that the above annotations are
> wrong?

If there's 1 out of N ww_mutex users that missed ww_acquire_done()
there's a good chance that 1 case is wrong.

But generally, depends if we want to enforce ww_acquire_done() or not
which itself is no-op in non-lockdep builds, however, with
DEBUG_WW_MUTEXES it's no longer no-op so it might be a good idea to
enforce it to get proper lockdep checking.

> Is there a better solution than removing the __acquire() and
> __release() annotations from the above three functions?

The kernel-doc comment for ww_acquire_done() says:

	/**
	 * ww_acquire_done - marks the end of the acquire phase
	 * @ctx: the acquire context
	 *
>>	 * Marks the end of the acquire phase, any further w/w mutex lock calls using
>>	 * this context are forbidden.
>>	 *
>>	 * Calling this function is optional, it is just useful to document w/w mutex
>>	 * code and clearly designated the acquire phase from actually using the locked
>>	 * data structures.
	 */
	static inline void ww_acquire_done(struct ww_acquire_ctx *ctx)
		__releases(ctx) __acquires_shared(ctx) __no_context_analysis
	{
	#ifdef DEBUG_WW_MUTEXES
		lockdep_assert_held(ctx);

		DEBUG_LOCKS_WARN_ON(ctx->done_acquire);
		ctx->done_acquire = 1;
	#endif
	}

It states it's optional, but it's unclear if that's true with
DEBUG_WW_MUTEXES builds. I'd vote for enforcing use of
ww_acquire_done(). If there's old code that's not using it, it should be
added there to get proper lockdep checking.