Re: [PATCH] NVMe: Fix compilation on architecturs without readq/writeq

[Ingo Molnar <mingo@elte.hu>]

* Hitoshi Mitake <h.mitake@gmail.com> wrote:

> On Sat, Jan 21, 2012 at 17:28, Ingo Molnar <mingo@elte.hu> wrote:
> >
> > * Linus Torvalds <torvalds@linux-foundation.org> wrote:
> >
> >> On Thu, Jan 19, 2012 at 5:01 PM, Matthew Wilcox
> >> <matthew.r.wilcox@intel.com> wrote:
> >> > The only places that uses readq/writeq are in the initialisation
> >> > path.  Since they're not performance critical, always use readl/writel.
> >>
> >> The arch rules are that i fthe architecture has readq/writeq, they
> >> will be #define'd (they may be inline functions, but there will be a
> >>
> >>   #define readq readq
> >>
> >> to make it visible to the preprocessor as well).
> >>
> >> So if you don't need the atomicity guarantees of a "real" readq, you
> >> can do this instead:
> >>
> >>   #ifndef readq
> >>   static inline u64 readq(void __iomem *addr)
> >>   {
> >>         return readl(addr) | (((u64) readl(addr + 4)) << 32LL);
> >>   }
> >>   #endif
> >>
> >> and then use readq() as if it existed.
> >>
> >> And I do think we should expose this in some generic manner. Because
> >> we currently don't, we already have that pattern copied in quite a few
> >> drivers.
> >>
> >> Maybe <asm-generic/io-nonatomic.h> or something? Making it
> >> clear that its not atomic, but avoiding the silly duplication
> >> we do now..
> >>
> >> This whole mess was introduced in commit dbee8a0affd5 ("x86:
> >> remove 32-bit versions of readq()/writeq()"), and it already
> >> talked about the problems but didn't help with the drivers
> >> that simply don't care.
> >>
> >> All the people in those threads were doing their
> >> self-satisfied "writeq is broken", without much acknowledging
> >> that 99% of users simply don't seem to care.
> >>
> >> "Occupy Writeq - We are the 99%"
> >
> > Agreed, and offering a generic facility for silly duplication
> > was the motivation of the original commit by Hitoshi Mitake.
> >
> > This:
> >
> > | The presense of a writeq() implementation on 32-bit x86 that
> > | splits the 64-bit write into two 32-bit writes turns out to
> > | break the mpt2sas driver (and in general is risky for drivers
> > | as was discussed in
> > |   <http://lkml.kernel.org/r/adaab6c1h7c.fsf@cisco.com>).
> >
> > is actually a mostly bogus statement and creates more problems
> > than it solves.
> >
> > Hitoshi-san, would you be interested in re-adding the generic
> > readq/writeq definitions in a slight variation to 2c5643b1c5, to
> > a separate io-nonatomic.h file, so that drivers that want it can
> > #include that file and be happy?
> 
> It sounds nice. In the previous discussion, I suggested that
> chaning the name of non-atomic readq/writeq to
> readq_nonatomic/writeq_nonatomic. And James Bottomley
> replied that it is fine but not really very useful:
> 
> https://lkml.org/lkml/2011/5/19/13
> 
> The idea of providing non-atomic readq/writeq in the new file
> with the name which express non-atomicity clearly might be
> able to satisfy both of safety and usefulness.
> 
> It will reduce the duplication of the definition. In addition
> readq/writeq users don't have to type the long symbols with
> _nonatomic suffix and can know non-atomicity from the name
> of header file.
> 
> I'd like to hear opinions from James, Roland and folks who
> dislike non-atomic readq/writeq.

Drivers that want the definition add this line to their driver:

#include <asm/io-nonatomic.h>

and then readq()/writeq() does the obvious thing. No need for 
readq_nonatomic()/writeq_nonatomic() - that extra line declares 
things clearly enough and cannot be added accidentally.

Thanks,

	Ingo