On Tue, Oct 13, 2015 at 02:52:46PM +0100, Chris Wilson wrote: > On Tue, Oct 13, 2015 at 03:52:08PM +0200, Daniel Vetter wrote: > > Yeah, hence using _rcu list macros. They have the relevant barriers > > already and should work. The only difference is that instead of > > synchronize_rcu on the write side before kfree, we'll use stop_machine on > > the read side. It's still RCU, but with all the cost moved to the read > > side while still keeping the benefit that the read side can be done > > locklessly. > > They imply a mb() on every write not just a barrier(), and we do a fair few > list updates on each buffer. Daniel pointed out that I was reading the version for pentium pro. But the next question will be when does that actually get set? -Chris -- Chris Wilson, Intel Open Source Technology Centre _______________________________________________ Intel-gfx mailing list Intel-gfx@xxxxxxxxxxxxxxxxxxxxx http://lists.freedesktop.org/mailman/listinfo/intel-gfx
