Re: arch/x86/Kconfig selects invalid HAVE_READQ, HAVE_WRITEQ vars

[Hitoshi Mitake <h.mitake@gmail.com>]

On Tue, Apr 21, 2009 at 01:03, Ingo Molnar <mingo@elte.hu> wrote:
>
> * Hitoshi Mitake <h.mitake@gmail.com> wrote:
>
>> On Mon, Apr 20, 2009 at 19:53, Ingo Molnar <mingo@elte.hu> wrote:
>> >
>> > * H. Peter Anvin <hpa@zytor.com> wrote:
>> >
>> >> Roland Dreier wrote:
>> >> >
>> >> > Notice that it reads from addr+4 *before* it reads from addr, rather
>> >> > than after as in your example (and in fact your example depends on
>> >> > undefined compiler semantics, since there is no sequence point between
>> >> > the two operands of the | operator).  Now, I don't know that hardware,
>> >> > so I don't know if it makes a difference, but the niu example I gave in
>> >> > my original email shows that given hardware with clear-on-read
>> >> > registers, the order does very much matter.
>> >> >
>> >>
>> >> At least for x86, the order should be low-high, because that is the
>> >> order that those two transactions would be seen on a 32-bit bus
>> >> downstream from the CPU if the CPU issued a 64-bit transaction.
>> >>
>> >> The only sane way to handle this as something other than per-driver
>> >> hacks would be something like:
>> >>
>> >> #include <linux/io64.h>               /* Any 64-bit I/O OK */
>> >>
>> >> #include <linux/io64lh.h>     /* Low-high splitting OK */
>> >>
>> >> #include <linux/io64hl.h>     /* High-low splitting OK */
>> >>
>> >> #include <linux/io64atomic.h> /* 64-bit I/O must be atomic */
>> >>
>> >> ... i.e. letting the driver choose what fallback method it will accept.
>> >
>> > Yeah - with the default being the natural low-high order.
>> >
>> > The other argument is that if a driver really wants some rare, oddly
>> > different order it should better define its own method that is not
>> > named in the same (or in a similar) way as an existing generic API.
>> > Otherwise, confusion will ensue.
>> I think this is a good way.
>> readq/writeq are already in Linus's tree, removing these is not a good idea.
>>
>> And I've sent the patch to fix a little problem of Kconfig about
>> readq/writeq to you.
>> http://marc.info/?l=linux-kernel&m=123521109218008&w=2
>> Did you notice?
>>
>> Adding cautions about accessing order or non-atomic to Kconfig's help
>> part may be benefit.
>
> It's better to add add such non-interactive help text as Makefile
> comments:
>
> #
> # This option ...
> #
>
> and they should be invisible in make menuconfig. This is a facility
> provided by architectures.
I'll move the help text from Kconfig to Makefile.
(My original patch also doesn't make help text visible in make menuconfig.)

>
> Note, the whole patchset is still incomplete - readq/writeq wrappers
> should be provided on all 32-bit architectures. Are those in the
> works?

I'm not working on porting readq/writeq on all 32-bit architectures.
If I port these, HAVE_READQ will be needless. Because there's no reason
to judge that architecture provides readq/writeq.

Porting readq/writeq on all architectures is radical way to solve.
But the problem related to order of accessing and non-atomic still exists.

I think there are 3 ways to choose:

1) Removing readq/writeq from x86_32
This is the way Roland mentioned.
This way removes the bugs related to order of accessing and non-atomic forever.
But driver programmers must implement their own version of readq/writeq,
and Andrew Morton said such case is sucks.
http://marc.info/?l=linux-kernel&m=122625885124798&w=2

2) Adding HAVE_READQ and HAVE_WRITEQ to Kconfigs of architectures
which provides readq/writeq
   despite of 32/64bit
This is the nearest with current state of Linux.
But some day non-atomic or order of accessing which driver programmers
didn't expect may cause
subtle bugs.

3) Porting readq/writeq on all architectures despite of 32/64bit
This is a very radical way.
This frees us from the problem of "#ifdef readq <implement driver own
version> #endif"
or HAVE_READQ forever.
But the possibility of subtle bugs caused by non-atomic or order of
accessing still exists.

Which one should we choose?

I suggest 2) (or 3)).
Because there's no problem since ported readq/writeq on x86_32.
And as H. Peter Anvin mentioned non-atomic is generally fine.