Nouveau - Jan 2020 - [RFT 00/13] iomap: Constify ioreadX() iomem argument

Le 08/01/2020 ? 09:18, Krzysztof Kozlowski a ?crit?:
> On Wed, 8 Jan 2020 at 09:13, Geert Uytterhoeven <geert at
linux-m68k.org> wrote:
>>
>> Hi Krzysztof,
>>
>> On Wed, Jan 8, 2020 at 9:07 AM Geert Uytterhoeven <geert at
linux-m68k.org> wrote:
>>> On Tue, Jan 7, 2020 at 5:53 PM Krzysztof Kozlowski <krzk at
kernel.org> wrote:
>>>> The ioread8/16/32() and others have inconsistent interface
among the
>>>> architectures: some taking address as const, some not.
>>>>
>>>> It seems there is nothing really stopping all of them to take
>>>> pointer to const.
>>>
>>> Shouldn't all of them take const volatile __iomem pointers?
>>> It seems the "volatile" is missing from all but the
implementations in
>>> include/asm-generic/io.h.
>>
>> As my "volatile" comment applies to iowrite*(), too, probably
that should be
>> done in a separate patch.
>>
>> Hence with patches 1-5 squashed, and for patches 11-13:
>> Reviewed-by: Geert Uytterhoeven <geert+renesas at glider.be>
> 
> I'll add to this one also changes to ioreadX_rep() and add another
> patch for volatile for reads and writes. I guess your review will be
> appreciated once more because of ioreadX_rep()
> 
volatile should really only be used where deemed necessary:

https://www.kernel.org/doc/html/latest/process/volatile-considered-harmful.html

It is said: " ...  accessor functions might use volatile on 
architectures where direct I/O memory access does work. Essentially, 
each accessor call becomes a little critical section on its own and 
ensures that the access happens as expected by the programmer."

Christophe

Hi Christophe,

On Wed, Jan 8, 2020 at 9:35 AM Christophe Leroy <christophe.leroy at
c-s.fr> wrote:
> Le 08/01/2020 ? 09:18, Krzysztof Kozlowski a ?crit :
> > On Wed, 8 Jan 2020 at 09:13, Geert Uytterhoeven <geert at
linux-m68k.org> wrote:
> >> On Wed, Jan 8, 2020 at 9:07 AM Geert Uytterhoeven <geert at
linux-m68k.org> wrote:
> >>> On Tue, Jan 7, 2020 at 5:53 PM Krzysztof Kozlowski <krzk at
kernel.org> wrote:
> >>>> The ioread8/16/32() and others have inconsistent interface
among the
> >>>> architectures: some taking address as const, some not.
> >>>>
> >>>> It seems there is nothing really stopping all of them to
take
> >>>> pointer to const.
> >>>
> >>> Shouldn't all of them take const volatile __iomem
pointers?
> >>> It seems the "volatile" is missing from all but the
implementations in
> >>> include/asm-generic/io.h.
> >>
> >> As my "volatile" comment applies to iowrite*(), too,
probably that should be
> >> done in a separate patch.
> >>
> >> Hence with patches 1-5 squashed, and for patches 11-13:
> >> Reviewed-by: Geert Uytterhoeven <geert+renesas at glider.be>
> >
> > I'll add to this one also changes to ioreadX_rep() and add another
> > patch for volatile for reads and writes. I guess your review will be
> > appreciated once more because of ioreadX_rep()
>
> volatile should really only be used where deemed necessary:
>
>
https://www.kernel.org/doc/html/latest/process/volatile-considered-harmful.html
>
> It is said: " ...  accessor functions might use volatile on
> architectures where direct I/O memory access does work. Essentially,
> each accessor call becomes a little critical section on its own and
> ensures that the access happens as expected by the programmer."
That is exactly the use case here: all above are accessor functions.

Why would ioreadX() not need volatile, while readY() does?

Gr{oetje,eeting}s,

                        Geert

-- 
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert at
linux-m68k.org

In personal conversations with technical people, I call myself a hacker. But
when I'm talking to journalists I just say "programmer" or
something like that.
                                -- Linus Torvalds

On Wed, Jan 8, 2020 at 9:36 AM Christophe Leroy <christophe.leroy at
c-s.fr> wrote:
> Le 08/01/2020 ? 09:18, Krzysztof Kozlowski a ?crit :
> > On Wed, 8 Jan 2020 at 09:13, Geert Uytterhoeven <geert at
linux-m68k.org> wrote:
> > I'll add to this one also changes to ioreadX_rep() and add another
> > patch for volatile for reads and writes. I guess your review will be
> > appreciated once more because of ioreadX_rep()
> >
>
> volatile should really only be used where deemed necessary:
>
>
https://www.kernel.org/doc/html/latest/process/volatile-considered-harmful.html
>
> It is said: " ...  accessor functions might use volatile on
> architectures where direct I/O memory access does work. Essentially,
> each accessor call becomes a little critical section on its own and
> ensures that the access happens as expected by the programmer."
The I/O accessors are one of the few places in which 'volatile'
generally
makes sense, at least for the implementations that do a plain pointer
dereference (probably none of the ones in question here).

In case of readl/writel, this is what we do in asm-generic:

static inline u32 __raw_readl(const volatile void __iomem *addr)
{
        return *(const volatile u32 __force *)addr;
}

The __force-cast that removes the __iomem here also means that
the 'volatile' keyword could be dropped from the argument list,
as it has no real effect any more, but then there are a few drivers
that mark their iomem pointers as either 'volatile void __iomem*' or
(worse) 'volatile void *', so we keep it in the argument list to not
add warnings for those drivers.

It may be time to change these drivers to not use volatile for __iomem
pointers, but that seems out of scope for what Krzysztof is trying
to do. Ideally we would be consistent here though, either using volatile
all the time or never.

        Arnd

Hi Geert,

Le 08/01/2020 ? 09:43, Geert Uytterhoeven a ?crit?:
> Hi Christophe,
> 
> On Wed, Jan 8, 2020 at 9:35 AM Christophe Leroy <christophe.leroy at
c-s.fr> wrote:
>> Le 08/01/2020 ? 09:18, Krzysztof Kozlowski a ?crit :
>>> On Wed, 8 Jan 2020 at 09:13, Geert Uytterhoeven <geert at
linux-m68k.org> wrote:
>>>> On Wed, Jan 8, 2020 at 9:07 AM Geert Uytterhoeven <geert at
linux-m68k.org> wrote:
>>>>> On Tue, Jan 7, 2020 at 5:53 PM Krzysztof Kozlowski <krzk
at kernel.org> wrote:
>>>>>> The ioread8/16/32() and others have inconsistent
interface among the
>>>>>> architectures: some taking address as const, some not.
>>>>>>
>>>>>> It seems there is nothing really stopping all of them
to take
>>>>>> pointer to const.
>>>>>
>>>>> Shouldn't all of them take const volatile __iomem
pointers?
>>>>> It seems the "volatile" is missing from all but
the implementations in
>>>>> include/asm-generic/io.h.
>>>>
>>>> As my "volatile" comment applies to iowrite*(), too,
probably that should be
>>>> done in a separate patch.
>>>>
>>>> Hence with patches 1-5 squashed, and for patches 11-13:
>>>> Reviewed-by: Geert Uytterhoeven <geert+renesas at
glider.be>
>>>
>>> I'll add to this one also changes to ioreadX_rep() and add
another
>>> patch for volatile for reads and writes. I guess your review will
be
>>> appreciated once more because of ioreadX_rep()
>>
>> volatile should really only be used where deemed necessary:
>>
>>
https://www.kernel.org/doc/html/latest/process/volatile-considered-harmful.html
>>
>> It is said: " ...  accessor functions might use volatile on
>> architectures where direct I/O memory access does work. Essentially,
>> each accessor call becomes a little critical section on its own and
>> ensures that the access happens as expected by the programmer."
> 
> That is exactly the use case here: all above are accessor functions.
> 
> Why would ioreadX() not need volatile, while readY() does?
> 
My point was: it might be necessary for some arches and not for others.

And as pointed by Arnd, the volatile is really only necessary for the 
dereference itself, should the arch use dereferencing.

So I guess the best would be to go in the other direction: remove 
volatile keyword wherever possible instead of adding it where it is not 
needed.

Christophe

On Wed, Jan 08, 2020 at 09:44:36AM +0100, Arnd Bergmann
wrote:
> On Wed, Jan 8, 2020 at 9:36 AM Christophe Leroy <christophe.leroy at
c-s.fr> wrote:
> > Le 08/01/2020 ? 09:18, Krzysztof Kozlowski a ?crit :
> > > On Wed, 8 Jan 2020 at 09:13, Geert Uytterhoeven <geert at
linux-m68k.org> wrote:
> > > I'll add to this one also changes to ioreadX_rep() and add
another
> > > patch for volatile for reads and writes. I guess your review will
be
> > > appreciated once more because of ioreadX_rep()
> > >
> >
> > volatile should really only be used where deemed necessary:
> >
> >
https://www.kernel.org/doc/html/latest/process/volatile-considered-harmful.html
> >
> > It is said: " ...  accessor functions might use volatile on
> > architectures where direct I/O memory access does work. Essentially,
> > each accessor call becomes a little critical section on its own and
> > ensures that the access happens as expected by the programmer."
> 
> The I/O accessors are one of the few places in which 'volatile'
generally
> makes sense, at least for the implementations that do a plain pointer
> dereference (probably none of the ones in question here).
> 
> In case of readl/writel, this is what we do in asm-generic:
> 
> static inline u32 __raw_readl(const volatile void __iomem *addr)
> {
>         return *(const volatile u32 __force *)addr;
> }
SuperH is another example:
1. ioread8_rep(void __iomem *addr, void *dst, unsigned long count)
   calls mmio_insb()

2. static inline void mmio_insb(void __iomem *addr, u8 *dst, int count)
   calls __raw_readb()

3. #define __raw_readb(a)          (__chk_io_ptr(a), *(volatile u8  __force
*)(a))

Even if interface was not marked as volatile, in fact its implementation
was casting to volatile.
> The __force-cast that removes the __iomem here also means that
> the 'volatile' keyword could be dropped from the argument list,
> as it has no real effect any more, but then there are a few drivers
> that mark their iomem pointers as either 'volatile void __iomem*'
or
> (worse) 'volatile void *', so we keep it in the argument list to
not
> add warnings for those drivers.
> 
> It may be time to change these drivers to not use volatile for __iomem
> pointers, but that seems out of scope for what Krzysztof is trying
> to do. Ideally we would be consistent here though, either using volatile
> all the time or never.
Indeed. I guess there are no objections around const so let me send v2
for const only.

Best regards,
Krzysztof