MMIO and gcc re-ordering issue

[Trent Piepho]

On Tue, 3 Jun 2008, Matthew Wilcox wrote:
> On Tue, Jun 03, 2008 at 12:43:21PM -0700, Trent Piepho wrote:
>> IOW, there are four ways one can defined endianness/swapping:
>> 1) Little-endian
>> 2) Big-endian
>> 3) Native-endian aka non-byte-swapping
>> 4) Foreign-endian aka byte-swapping
>>
>> 1 and 2 are by far the most used.  Some code wants 3.  No one wants 4.  Yet
>> our API is providing 3 & 4, the two which are the least useful.
>
> You've fundamentally misunderstood.
>
> readX/writeX and __readX/__writeX provide little-endian access.
> __raw_readX provide native-endian.
>
> If you want 2 or 4, define your own accessors.  Some architectures define
> other accessors (eg gsc_readX on parisc is native (big) endian, and

How about providing 1 and 2, and if you want 3 or 4 define your own accessors?

>> Is it enough to provide only "all or none" for ordering strictness?  For
>> instance on powerpc, one can get a speedup by dropping strict ordering for
>> IO
>> vs cacheable memory, but still keeping ordering for IO vs IO and IO vs
>> locks. This is much easier to program for than no ordering at all.  In
>> fact, if one
>> doesn't use coherent DMA, it's basically the same as fully strict ordering.
>
> I don't understand why you keep talking about DMA.  Are you talking
> about ordering between readX() and DMA?  PCI proides those guarantees.

I guess you haven't been reading the whole thread.  The reason it started was
because gcc can re-order powerpc (and everyone else's too) IO accesses vs
accesses to cachable memory (but not spin-locks), which ends up only being a
problem with coherent DMA.