On Wed, Feb 17, 2010 at 08:43:14PM +1100, Anton Blanchard wrote: > > Hi Nick, > > > Ah, good to see this one come back. I also tested tbench over localhost > > btw which actually did show some speedup on the G5. > > > > BTW. this was the last thing left: > > http://www.mail-archive.com/linuxppc-dev@lists.ozlabs.org/msg29738.html > > > > Don't know if you took a look at that again, but maybe it's worth > > looking at. Hmm, we do actually seem to be growing number of smp_mb* > > calls in core kernel too. > > Interesting idea! I do worry we will get a late night visit from the > architecture police. From memory they want the complete larx, stcwx, bne, > isync sequence to guarantee an acquire barrier.
Yes I suppose the branch can be executed "non speculatively" before the lwsync is completed. Wheras the larx/stcwx will have to complete before the branch outcome can be known. I suppose probably not worthwhile avoiding the full IO sync by adding yet more crap to make this work. Thanks for putting my mind to rest though :) I'd still be interested to know how expensive the full sync is when you have a lot of IOs in flight. Question, are you going to do the hint and isync->lwsync thing for userspace as well? Too bad the kernel doesn't export synchronisation primitives to userspace... _______________________________________________ Linuxppc-dev mailing list Linuxppc-dev@lists.ozlabs.org https://lists.ozlabs.org/listinfo/linuxppc-dev