On Wed, Nov 21, 2007 at 01:50:16PM +0000, Ralf Baechle wrote: > On Tue, Nov 20, 2007 at 05:02:20PM +0100, Nick Piggin wrote: > > > + TYPE FULL CACHEABLE (SMP CONDITIONAL) IO > > + ========= ======================= =========================== > > ============ > > + GENERAL mb() smp_mb() io_mb() > > + WRITE wmb() smp_wmb() io_wmb() > > + READ rmb() smp_rmb() io_rmb() > > + DATA DEP. read_barrier_depends() smp_read_barrier_depends() > > A while ago I went through the kernel looking at the uses mb(), wmb() and > rmb() and found that every use was either fishy or should have been a > smp_mb(), smp_wmb() or wmp_rmb(). Which leads me to the question if there > is any need for the non-smp_ variants left or can we just bury them?
I don't know driver code very well, however it seems like there is a legitimate need for barriers there, _however_ I think a lot of those went in when writel could be unordered on some architectures. The sane way to do it is to make all drivers default to strongly ordered IO (eg. make readl writel strongly ordered). And then yes indeed I believe most of the mb/wmb/rmb can go away. We still need full memory barrires in order to order cacheable access with IO access (eg. in order to support SMP locking of IO routines). However in that case I believe an easily understood lock-based primitive should be provided, along with a fully ordering implementations of the regular memory barriers, rather than mmiowb (which clearly very few driver writers or anyone else understands). After that, if some architectures want to implement more relaxed variants of the IO accessors, then OK. And we can use io_ memory barriers for that. - To unsubscribe from this list: send the line "unsubscribe linux-arch" in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html
