Re: [PATCH] scsi: mac_scsi: Replace NO_IRQ by 0

From: Christophe Leroy
Date: Thu Oct 06 2022 - 07:14:13 EST

Le 06/10/2022 à 11:25, Finn Thain a écrit :
> On Thu, 6 Oct 2022, Christophe Leroy wrote:
>> NO_IRQ is used to check the return of irq_of_parse_and_map().
>> On some architecture NO_IRQ is 0, on other architectures it is -1.
> Yes. The core NCR5380 driver is used on ARM etc. where NO_IRQ is -1 as
> well as on powerpc where it is 0.
>> irq_of_parse_and_map() returns 0 on error, independent of NO_IRQ.
>> So use 0 instead of using NO_IRQ.
> Sorry, I must be missing something.

My mistake.

I started by removing NO_IRQ definition in powerpc and then fixed all
build failures by replacing NO_IRQ by 0. Then I splitted the patch into
one per subsystem, all with the same explaination.

Most places it was just a verification of the value returned by
irq_of_parse_and_map() where it is obviously wrong to use NO_IRQ,
especially on ARM where NO_IRQ doesn't match what irq_of_parse_and_map()
returns in case on error.

But here in the mac_scsi driver it seems a bit different and I have a
look more closely.

> You seem to be saying that this driver could be re-used in the context of
> openfirmware/device trees if it avoided using the NO_IRQ. Do I have that
> right?
> Or are you changing NO_IRQ semantics tree-wide for some reason explained
> somewhere else?

No, I only say that NO_IRQ doesn't match the value returned by
irq_of_parse_and_map(). Ultimately I want to remove the #define NO_IRQ
from arch/powerpc/include/asm/irq.h

That's to be linked to following message from Linus :

> If it is the former, shouldn't you reverse the comment in
> arch/powerpc/include/asm/irq.h, which says the macro is to be used in the
> way this driver (and others) use it?
> If it is the latter, shouldn't you address the use of NO_IRQ in the core
> NCR5380 driver rather than just this wrapper?

Yes I guess so.

> Moreover, wouldn't it make more sense to fix the callers of
> irq_of_parse_and_map(), since they appear to be abusing the NO_IRQ macro?

Indeed. That's what is being done most places.

> For example, drivers/ata/sata_dwc_460ex.c actually does #define NO_IRQ 0
> and then expects irq_of_parse_and_map() will somehow use the same value to
> mean the same thing...

It didn't pop up during the multi-build I did for powerpc, so I guess
that driver is not used for powerpc ? In the ata subsystem I fixed