Re: [PATCH] task_work: simplify the task_work_add() interface

[Christoph Hellwig]

On Thu, Mar 10, 2022 at 05:43:25AM -0700, Jens Axboe wrote:
> On 2/23/22 12:27 AM, Christoph Hellwig wrote:
> > Provide a low-level task_work_add_nonotify interface that just adds
> > the work to the list and open code the TWA_SIGNAL and TWA_NONE callers
> > using it.  task_work_add() itself now only handles the common TWA_RESUME
> > case and can drop the notify argument.
> 
> Not sure this is much of a cleanup, and a potential fast case of
> TWA_NONE will now still still set TIF_NOTIFY_RESUME. Also:

No, the old TWA_NONE case is switched to task_work_add_nonotify and
does not set TIF_NOTIFY_RESUME.

> 
> > diff --git a/fs/io_uring.c b/fs/io_uring.c
> > index 77b9c7e4793bf..94116a102dc61 100644
> > --- a/fs/io_uring.c
> > +++ b/fs/io_uring.c
> > @@ -9606,7 +9606,7 @@ static __cold void io_ring_exit_work(struct work_struct *work)
> >  					ctx_node);
> >  		/* don't spin on a single task if cancellation failed */
> >  		list_rotate_left(&ctx->tctx_list);
> > -		ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
> > +		ret = task_work_add_nonotify(node->task, &exit.task_work);
> >  		if (WARN_ON_ONCE(ret))
> >  			continue;
> 
> This one is now no longer setting TIF_NOTIFY_SIGNAL.

Yes, this was a rebase bug.

> If you want to get rid of the argument, why not just have separate
> helpers? task_work_add_signal(), task_work_add_resume(),
> task_work_add(). Setting TWA_RESUME unconditionally because it's the
> common use case doesn't seem ideal.

In this series, task_work_add_nonotify is what you seems to call
task_work_add, task_work_add is this series is what you call
task_work_add_resume and task_work_add_signal is open coded because
there aren't a whole lot of users.  But if you want I can add
task_work_add_signal and rename task_work_add to task_work_add_resume,
but I think keeping the task_work_add_nonotify name for the low-level
helper is a lot more descriptive.