On Tue, 2007-07-24 at 10:24 -0700, Linus Torvalds wrote: > > On Tue, 24 Jul 2007, Benjamin Herrenschmidt wrote: > > > > In fact, it's more than that... the bitops that return a value are often > > used to have hand-made spinlock semantics. I'm sure we would get funky > > bugs if loads or stores leaked out of the locked region. I think a full > > "memory" clobber should be kept around for those cases. > > Not helpful. > > The CPU ordering constraints for "test_and_set_bit()" and friends are weak > enough that even if you have a full memory clobber, it still wouldn't work > as a lock. > > That's exactly why we have smp_mb__after_set_bit() and friends. On some > architectures (arm, mips, parsic, powerpc), *that* is where the CPU memory > barrier is, because the "test_and_set_bit()" itself is just a > cache-coherent operation, not an actual barrier.
Well, as I said, our test_and_set_bit() asm (and in general, the asm for all the atomic ops that -return- a value) have at least some level of barriers in them because of that. We do that because people are abusing them as locks. The smp_mb__after_set_bit() I never quite grokked. We do an mb in there but I suspect we don't need if it's only ever used after test_and_set_bit() because of the above. The smb_mb__before_clear_bit() makes more sense as it's supposed to give clear_bit() a spin_unlock semantic. But we do need the "memory" clobber as well. That's one reason why I like Nick's bitop locks patches, providing -explicit- test_and_set_bit_lock() and clear_bit_unlock(), we can fix a whole lot of things and make sure they have the right barriers and not more. (We save a few useless barriers on POWER that way in the page lock path and it's measureable in his benchmark). Cheers, Ben. - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
