> >>> + writel(n, adap->bar2 + q->udb + 8);
> >>> +#if defined(CONFIG_X86_32) || defined(CONFIG_X86_64)
> >>> + asm volatile("sfence" : : : "memory");
> >>> +#endif
> >> There is absolutely no way I'm letting anyone put crap like this
> >> into a driver.
> >>
> >> Use a portable inteface, and if one does not exist create one.
> >
> > I guess you'll have to add a wc_wmb() function for all the hw platforms
> > supported by the kernel. I see libibverbs defines this for the user
> > side in include/infiniband/arch.h, and that could be used as the meat of
> > the hw platform-specific implementations.
> >
> I see that normal wmb() code for x86_64 architecture is same as what
> above #ifdef condition is doing. To be more clear I looked at the
> assembly code for wmb and find that it is converted into 'sfence'
> instruction. So, I think above code should be replaced with wmb call
> which will also take care of portability on different architecture. I
> will submit the series again soon.
>From my recollection of the x86 architecture, the memory barriers
are hardly ever needed, certainly not in the places where, for example
a ppc needs them. I'd actually suspect that the normal wmb() for
x86 should be a nop.
About the only place where any on the fence instructions are needed
are in relation to write combining accesses.
In particular I don't believe they are ever needed to synchronise
uncached accesses with each other, or with cached accesses (which
are snooped).
David
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
