Re: [PATCH] i2c: imx: increase retries on arbitration loss

[Primoz Fiser]

Hi all,

On 16. 12. 22 13:51, Francesco Dolcini wrote:
> On Fri, Dec 16, 2022 at 01:23:29PM +0100, Primoz Fiser wrote:
> > Hi all,
> >
> > On 16. 12. 22 12:13, Uwe Kleine-König wrote:
> > > On Fri, Dec 16, 2022 at 12:02:27PM +0100, Oleksij Rempel wrote:
> > > > On Fri, Dec 16, 2022 at 11:41:08AM +0100, Primoz Fiser wrote:
> > > > > Hi Marco,
> > > > >
> > > > > On 16. 12. 22 10:45, Marco Felsch wrote:
> > > > > > Hi Primoz,
> > > > > >
> > > > > > On 22-12-16, Primoz Fiser wrote:
> > > > > > > By default, retries value is set to 0 (no retries). Set retries to more
> > > > > > > sensible value of 3 to allow i2c core to re-attempt transfer in case of
> > > > > > > i2c arbitration loss (i2c-imx returns -EAGAIN errno is such case).
> > > > > >
> > > > > > apart the fact that the number of retries vary a lot and so the client
> > > > > > driver behaviour can vary a lot which is not good IMHO, why do you think
> > > > > > that 3 is a sufficient number?
> > > > >
> > > > > IMHO it is better than leaving it at 0 (no retries)?
> > > > >
> > > > > Setting it to sensible value like 3 will at least attempt to make transfer
> > > > > in case arbitration-loss occurs.
> > > > >
> > > > > > If an arbitration loss happen, why do you think that retrying it 3 times
> > > > > > changes that?
> > > > >
> > > > > I our case, setting retries to non-zero value solves issues with PMIC
> > > > > shutdown on phyboard-mira which in some rare cases fails with "Failed to
> > > > > shutdown (err =  -11)" (-EAGAIN).
> > > > >
> > > > > To me it makes common sense retries is set to non-zero value especially for
> > > > > such rare conditions/situations.
> > > >
> > > > https://lore.kernel.org/all/Ys1bw9zuIwWS+bqw@shikoro/
> >
> > Ohh I see.
> >
> > Reading through the thread I guess we aren't getting this mainlined at all
> > :)
> >
> > The only solid point in the thread seems to be that in that case we are not
> > covering up the potential i2c hardware issues?
>
> I believe that in this case we should just have a warning in the kernel.
> The retry potentially work-around a transient issue and we do not hide any hardware
> issue at the same time. It seems an easy win-win solution.

I would agree about throwing a warning message in retry case.

Not sure how would it affect other i2c bus drivers using retries > 0. 
Retries might be pretty rare with i2c-imx but some other drivers set 
this to 5 for example. At least using _ratelimited printk is a must 
using this approach.

> > Yeah fair point but on the other hand, goal of this patch would be to
> > improve robustness in case of otherwise good performing hardware. From user
> > perspective I just want it to work despite it retrying under the hood from
> > time to time. I think Francesco had the same idea.
>
> Unfortunately I was missing the exact background that made us do this
> change, we just had it sitting in our fork for too long :-/
> This is one of the reason I gave up on it.
>
> Quoting Uwe [1]
> > sometimes there is no practical way around such work arounds. I happens
> > from time to time that the reason for problem is known, but fixing the
> > hardware is no option, then you need such workrounds. (This applies to
> > both, retrying the transfers and resetting the bus.)

I wouldn't say this is exactly a workaround if "retries mechanism" is 
standard part of the i2c core. Other drivers use it as well. it is just 
a setting to improve bus robustness.

But OK, I guess we can live with this one-liner in the downstream kernel.

> Francesco
>
> [1] https://lore.kernel.org/all/20220715083400.q226rrwxsgt4eomp@xxxxxxxxxxxxxx/