Re: [PATCH] gpiolib: fix the speed of descriptor label setting with SRCU

[Bartosz Golaszewski <brgl@xxxxxxxx>]

On Tue, May 7, 2024 at 4:24 PM Paul E. McKenney <paulmck@xxxxxxxxxx> wrote:
>
> On Tue, May 07, 2024 at 02:13:46PM +0200, Bartosz Golaszewski wrote:
> > From: Bartosz Golaszewski <bartosz.golaszewski@xxxxxxxxxx>
> >
> > Commit 1f2bcb8c8ccd ("gpio: protect the descriptor label with SRCU")
> > caused a massive drop in performance of requesting GPIO lines due to the
> > call to synchronize_srcu() on each label change. Rework the code to not
> > wait until all read-only users are done with reading the label but
> > instead atomically replace the label pointer and schedule its release
> > after all read-only critical sections are done.
> >
> > To that end wrap the descriptor label in a struct that also contains the
> > rcu_head struct required for deferring tasks using call_srcu() and stop
> > using kstrdup_const() as we're required to allocate memory anyway. Just
> > allocate enough for the label string and rcu_head in one go.
> >
> > Reported-by: Neil Armstrong <neil.armstrong@xxxxxxxxxx>
> > Closes: https://lore.kernel.org/linux-gpio/CAMRc=Mfig2oooDQYTqo23W3PXSdzhVO4p=G4+P8y1ppBOrkrJQ@xxxxxxxxxxxxxx/
> > Fixes: 1f2bcb8c8ccd ("gpio: protect the descriptor label with SRCU")
> > Suggested-by: Paul E. McKenney <paulmck@xxxxxxxxxx>
> > Signed-off-by: Bartosz Golaszewski <bartosz.golaszewski@xxxxxxxxxx>
>
> Looks good to me!
>
> Acked-by: Paul E. McKenney <paulmck@xxxxxxxxxx>
>
> One semi-related question...  Why the per-descriptor srcu_struct?
>
> If the srcu_struct was shared among all of these, you could just do one
> synchronize_srcu() and one cleanup_srcu_struct() instead of needing to
> do one per gdev->desc[] entry.
>
> You might be able to go further and have one srcu_struct for all the
> gpio devices.
>
> Or did you guys run tests and find some performance problem with sharing
> srcu_struct structures?   (I wouldn't expect one, but sometimes the
> hardware has a better imagination than I do.)
>

I guess my goal was not to make synchronize_srcu() for descriptor X
wait for read-only operations on descriptor Y. But with that gone, I
suppose you're right, we can improve this patch further by switching
to a single SRCU descriptor.

I'll send a v2.

Bart