Messages in this thread | | | Date | Thu, 12 Feb 2009 12:39:24 -0800 | From | "Paul E. McKenney" <> | Subject | Re: [ltt-dev] [RFC git tree] Userspace RCU (urcu) for Linux (repost) |
| |
On Thu, Feb 12, 2009 at 12:13:29PM -0800, Linus Torvalds wrote: > > > On Thu, 12 Feb 2009, Paul E. McKenney wrote: > > > > Actually the best way to do this would be: > > > > while (ACCESS_ONCE(sig_done) < 1) > > continue; > > > > If ACCESS_ONCE() needs to be made architecture-specific to make this > > really work on Blackfin, we should make that change. > > I really wouldn't want to mix up compiler barriers and cache barriers this > way. > > I think "cpu_relax()" is likely the right thing to piggy-back on for > broken cache-coherency. > > > And, now that you mention it, I have heard rumors that other CPU > > families can violate cache coherence in some circumstances. > > I personally suspect that the BF pseudo-SMP code is just broken, and that > it likely has tons of subtle bugs and races - because we _do_ depend on > cache coherency at least for accessing objects next to each other. I just > never personally felt like I had the energy to care deeply enough. > > But I draw the line at making ACCESS_ONCE() imply anything but a compiler > optimization issue.
In other words, you are arguing for using ACCESS_ONCE() in the loops, but keeping the old ACCESS_ONCE() definition, and declaring BF hardware broken?
I am OK with that, just wanting to make sure I understand what you are asking us to do.
Thanx, Paul
| |