On Mon, 2026-06-01 at 00:32 +0200, Danilo Krummrich wrote:
> On Mon Jun 1, 2026 at 12:00 AM CEST, Markus Probst wrote:
> > On Sun, 2026-05-31 at 23:49 +0200, Danilo Krummrich wrote:
> > > On Sun May 31, 2026 at 9:42 PM CEST, Markus Probst wrote:
> > > > I just noticed, is it even possible to use SRCU here? Currently the mutex not
> > > > only ensures that no drvdata access happens after drvdata drop, but also that
> > > > the receive_buf waits for the probe to complete, as the drvdata hasn't been
> > > > initialized yet.
> > > 
> > > Yeah, if you drop the completion, you need the mutex.
> > Is the performance impact on an mutex or on srcu + completion higher?
> 
> Weighing in the completion, the mutex probably wins as it will always be
> uncontested under normal operation.
> 
> > > 
> > > (In case it wasn't discussed in previous versions already, there is also the
> > > option to just attach separate private data to the receive callback, which would
> > > avoid this synchonization problem in the first place.
> > > 
> > > You could have serdev::Device<Core>::open(), which takes its own private data
> > > and a corresponding close(), this way you'd allow drivers to control whether
> > > they want the serial line "open" or not. You just need to make sure it is closed
> > > eventually.)
> > This would add complexity with types, as we need to ensure that write,
> > set_baudrate, set_parity etc. does not run when closed.
> 
> Right, but looking at a few serdev drivers, there seem to be a few cases where
> drivers need to close and re-open.
That is something we could consider once there is a user.

I am certain it would make it more complex, not less.
> 
> > > That said, I don't know what turns out to be the better approach. And maybe it
> > > simply isn't something this initial series has to tackle? I think your driver
> > > does not implement the receive callback?
> > The initial driver with only leds: no.
> > 
> > That changes once the driver also takes care of hwmon (ADC sensor and
> > fan failure) and input (power button and possibly other buttons).
> 
> Maybe drop it from this initial series then and revisit with the first user?
I would prefer it with this series.

Except for probe and remove there should be no overhead if receive_buf
is not implemented.

As it will be needed eventually and it is basic serdev functionality, I
don't see a reason to not include it now.

Thanks
- Markus Probst