Re: [PATCH][RFC] asm-generic:remove calling flush_write_buffers() in dma_sync_*_for_cpu

[Arnd Bergmann <arnd@arndb.de>]

On Monday 29 June 2009, Alan Cox wrote:
> > Wouldn't it be better to put the flush_write_buffer in the specific
> > operation (swiotlb_sync_*_for_*) rather than the multiplexer?
> > 
> > Maybe in that case, smp_wmb() would be more appropriate because
> > it is defined on all architectures.
> 
> smp_wmb() is stronger and it would slow down x86 if we did that (we'd go
> from no-op on a coherent platform to using mfence/lfence etc)
> 
Really? In my copy of system.h, I read

#ifdef CONFIG_SMP
# ifdef CONFIG_X86_OOSTORE
#  define smp_wmb()      wmb()
# else
#  define smp_wmb()      barrier()
# endif
#else
# define smp_wmb()       barrier()
#endif

That actually looks weaker than flush_write_buffer, as it would turn into
a barrier() in case of !SMP or !X86_OOSTORE, and into an sfence instead of
lock addl on all modern CPUs in case of SMP && X86_OOSTORE.

Of course that raises the question of whether smp_wmb() is too weak in case of
!SMP or X86_PPRO_FENCE, but with the described scenario, I don't think
it does.

	Arnd <><