[PATCH] SPI: bcm2835: move to the transfer_one driver model

[Martin Sperl]

> On 26.03.2015, at 18:35, Mark Brown <broonie at kernel.org> wrote:
> 
> On Thu, Mar 26, 2015 at 11:08:36AM +0100, Martin Sperl wrote:
>> This also allows for GPIO-CS to get used removing the limitation of
>> 2/3 SPI devises on the SPI bus.
> 
> This doesn't apply against current code, please check and resend.  As
> previously mentioned please use subject lines matching the style for the
> subsystem.  A few other things below but it's basically all good.

I have applied it against "for-next" (8befd715f1c1684b3).

The reason I chose for-next is because there all the other patches
I have submitted so far have been added there already...

Against which one do you want to apply it instead?

You want me to merge all those patches into a single patch instead?

>> So the question is if we should depreciate native chip-selects for this
>> driver with one of those future improvements listed below.
> 
> Given that this driver is only going to be used with a single SoC (or
> perhaps a very limited set of SoCs, I don't know if it's the same driver
> in the Pi 2) even if the DT specifies a hardware chip select we should
> be able to look up which pin that's brought out to and put it into GPIO
> mode if that's the most sensible thing for the driver.
The point here is that we need to change the alternate function for that
pin to output make it work.

That is why I was recommending the "simple" approach to "depreciate" 
that function if acceptable - less code to maintain...

Any yes: RPI2 uses the same SPI block and the same driver - the main 
difference is the arms

> 
>> * multiple spi_transfers handled in interrupt alone without waking up the
>>  worker-thread (for some transfers) to reduce context switching
>>  overheads and the corresponding latencies.
> 
> This in particular is something the framework should have: it's
> generally useful and we should be able to do it with either a new
> transfer_irq_safe() (or something) operation or a refactoring of the
> existing one.

Let me see what I can do with the current setup and then we can generalize
from that.

For most parts I see that I would use transfer_one and return 0 immediately
if we can "chain" it in interrupts - only the last transfer would trigger
a complete in the interrupt handler and return. We could even run the 
callback in the irq, if it is permitted...

But as I see there would be a few cases that would also need a complete.
These would be "delays" and maybe a CS-change - there mostly because of
the hard-coded "udelay(10)", but then it might be acceptable to sleep
10us in the irq, because the overhead of the irq handler releasing the CPU
is arround 5us and then you got another 3us inside the scheduler before
the worker-thread wakes up. If you take all that then you may as well sleep
in the interrupt handler - those at least are my measurements for
a RPI.

> 
>> +	/* error in the case of native CS requested with CS-id > 2 */
>> +	dev_err(&spi->dev,
>> +		"setup: only three native chip-selects are supported\n"
>> +		);
> 
> The indentation here is weird - at least the last ); should be with the
> string.

Will fix that when you tell me which branch you want it to get patch to 
apply against.

Martin