Grygorii,

thanks a lot for your input. Much appreciated!

> That would be great, but note:
> 1) only i2c_transfer() operations are locked, so if driver is doing 
> i2c_transfer(1)
> i2c_transfer(2) <- suspend in the middle
> <- suspend in between
> i2c_transfer(3)
> It will not help.

Will it not improve the situation by ensuring that at least the transfer
with its (potenitally) multiple messages got completed? That we are at
least in a bus-free state (assuming single-master here) before
suspending?

> Everything depends on timings :( - in my practice 10000 suspend iteration tests
> where required to run many times to catch 3 buggy I2C client drivers.

Matches my experiences that creating a reliable test case for that is
not that easy as I thought. Or I am missing something obvious.

> 2) It's normal to abort suspend if system is busy, so if I2C core will be able
> to catch active I2C operation - it should abort, but again I do not see how it
>  can be detected 100% with current I2C core design or without reworking huge number of drivers.

I agree. After second thought, waiting for i2c_transfer to finish maybe
won't be enough, I am afraid. We don't know if STOP has been put on the
wires yet. My best bet now is that we implement such a
'is-transfer-ongoing'-check in the suspend function of the master
driver? That check should be optional, but recommended.

> 3) So, only one thing I2C core potentially can do - catch invalid access and
> report it. "wait for transfer to finish" wouldn't work as for me.

And we do this in suspend_noirq function of the i2c core.

> > I at least know of some Renesas boards which needed the I2C connected
> > PMIC to do a system reset (not sure about suspend, need to recheck
> > that). That still today causes problems because interrupts are disabled
> > then.
> 
> this was triggered few times already (sry, don't have links), as of now,
>  and as I know, the only ways to W/A this is:
> - to create barametal platform driver (some time in ASM)
> - or delegate final suspend operation to another system controller (co-processor),
>   as example TI am335x SoCs,
> - or implement I2C driver in hw - TI AVS/SmartReflex. 

Yes. Please note that this is only needed for reset, not suspend. So, it
is a bit easier. Still, it might make more sense to use a platform based
solution. I'll think about that.

> Sry, but 99% percent of I2C client drivers *should not* access I2C bus after
> .suspend_noirq() stage it's BUG-BUG!! Any W/A will just hide real problems.

I do believe you, still is there documentation about such things? I like
to understand more but didn't dig up something up to now. E.g. I grepped
for "noirq" in Documentation/power.

> "master_xfer_irqless" might be a not bad idea, but, in my opinion, it 
> should be used explicitly by platform code only, and each usage should
> be proved to exist. 

Yes, we can think about it once it is really needed.

> Some additional info:

Thanks a lot for that!

> I'm attaching some very old patch (don't ask me why it was not sent :()
> I did for Android system - which likes suspend very much. Some
> part of below diff are obsolete now (like omap_i2c_suspend()),
> but .noirq() callback are still valid and can show over all idea.
> Really helped to catch min 3 buggy client drivers with timers, delayed
> or periodic works.

Ok, so what do you think about my plan to:

1) encourage drivers to check if there is still an ongoing transfer in
their .suspend function (or the core can do that, too, if we agree that
checking for a taken adapter lock is sufficient)

-> to ensure transfers don't get interrupted in the middle

2) use a .suspend_noirq callback in i2c_bus_type.pm to reject and WARN
about transfers still going on in that phase

-> this ensures that buggy drivers are caught

3) write some documentation about our findings / assumptions /
recommendations to a file in Documentation/i2c/

-> this ensures we won't forget why we did things like they are ;)

?

Kind regards,

   Wolfram