Re: [PATCH] staging: nvec: make i2c controller register writes robust

["Thierry Reding" <thierry.reding@gmail.com>]

[-- Attachment #1: Type: text/plain, Size: 1991 bytes --]

On Wed May 1, 2024 at 9:03 PM CEST, Marc Dietrich wrote:
> Hi Ben,
>
> On Mon, 22 Apr 2024, Ben Dooks wrote:
>
> > On 21/04/2024 11:46, Marc Dietrich wrote:
> >> The i2c controller needs to read back the data written to its registers.
> >> This way we can avoid the long delay in the interrupt handler.
> >>
> >> Signed-off-by: Marc Dietrich <marvin24@gmx.de>
> >> ---
> >>   drivers/staging/nvec/nvec.c | 41 ++++++++++++++++++++++---------------
> >>   1 file changed, 24 insertions(+), 17 deletions(-)
> >>
> >> diff --git a/drivers/staging/nvec/nvec.c b/drivers/staging/nvec/nvec.c
> >> index 45df190c2f94..214839f51048 100644
> >> --- a/drivers/staging/nvec/nvec.c
> >> +++ b/drivers/staging/nvec/nvec.c
> >> @@ -570,6 +570,22 @@ static void nvec_tx_set(struct nvec_chip *nvec)
> >>   		(uint)nvec->tx->size, nvec->tx->data[1]);
> >>   }
> >>
> >> +/**
> >> + * i2c_writel - safely write to an I2C client controller register
> >> + * @val: value to be written
> >> + * @reg: register to write to
> >> + *
> >> + * A write to an I2C controller register needs to be read back to make
> >> sure
> >> + * that the value has arrived.
> >> + */
> >> +static void i2c_writel(u32 val, void *reg)
> >> +{
> >> +	writel_relaxed(val, reg);
> >> +
> >> +	/* read back register to make sure that register writes completed */
> >> +	readl_relaxed(reg);
> >> +}
> >
> > I thought the default behaviour of writel() should be to force writes
> > out of any CPU buffers. Are there any bus isuses here causing the code
> > to be necessary (and if so, why is there another buffer breaking the
> > writel behaviour?)
>
> if fear that's a question only NVIDIA can answer.

Ben,

in case you didn't see the discussion on v2 of this patch, it's here:

	https://lore.kernel.org/all/D1N2OIXAF6QQ.3TCYLBU42CJ3U@gmail.com/

in a nutshell: there's indeed another buffer here that causes these
writes to be queued and the read-back flushes that queue.

Thierry

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 833 bytes --]