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

[Maarten Lankhorst]

Hey,

The acquire_done() call was always optional. It's meant to indicate that after this point,
ww_acquire_lock may no longer be called and backoff can no longer occur.

It's allowed to call ww_acquire_fini() without ww_acquire_done()

Think of this case:
ww_acquire_init()

ww_acquire_lock_interruptible() -> -ERESTARTSYS

ww_acquire_fini()

Here it wouldn't make sense to call ww_acquire_done().

It's mostly to facilitate this case:

ww_acquire_init()

ww_acquire_lock() a bunch.

/* Got all locks, do the work as no more backoff occurs */
ww_acquire_done()

...

unlock_all()
ww_acquire_fini()

If you call ww_acquire_lock after done, a warning should occur as this should no longer happen.

Kind regards,
~Maarten Lankhorst

Den 2026-01-09 kl. 22:26, skrev Bart Van Assche:
> (+Maarten)
> 
> On 1/9/26 2:06 PM, Marco Elver wrote:
>> 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.
> 
> $ git grep -w ww_acquire_done '**c'|wc -l
> 11
> $ git grep -w ww_acquire_fini '**c'|wc -l
> 33
> 
> The above statistics show that there are more cases where
> ww_acquire_done() is not called rather than cases where
> ww_acquire_done() is called.
> 
> Maarten, since you introduced the ww_mutex code, do you perhaps prefer
> that calling ww_acquire_done() is optional or rather that all users that
> do not call ww_acquire_done() are modified such that they call
> ww_acquire_done()? The full email conversation is available here:
> https://lore.kernel.org/all/20251219154418.3592607-1-elver@google.com/
> 
> Thanks,
> 
> Bart.