On Thu, Jul 14, 2016 at 9:49 AM, Peter Zijlstra <pet...@infradead.org> wrote: > On Thu, Jul 14, 2016 at 09:43:40AM -0700, John Stultz wrote: >> On Thu, Jul 14, 2016 at 6:18 AM, Peter Zijlstra <pet...@infradead.org> wrote: >> > On Wed, Jul 13, 2016 at 10:51:02PM +0200, Peter Zijlstra wrote: >> >> So, IIRC, the trade-off is a full memory barrier in read_lock and >> >> read_unlock() vs sync_sched() in write. >> >> >> >> Full memory barriers are expensive and while the combined cost might >> >> well exceed the cost of the sync_sched() it doesn't suffer the latency >> >> issues. >> >> >> >> Not sure if we can frob the two in a single codebase, but I can have a >> >> poke if Oleg or Paul doesn't beat me to it. >> > >> > OK, not too horrible if I say so myself :-) >> > >> > The below is a compile tested only first draft so far. I'll go give it >> > some runtime next. >> >> Unfortunately it didn't apply cleanly to the 4.4 based tree I'm >> working with, so I had to manually apply the entirety of the >> percpu-rwsem.c changes myself. Hopefully I didn't screw it up. >> >> So running with this, I'm still seeing some pretty large delays. 80ms >> peak, with lots of >20ms values as well. >> So it doesn't seem to have the positive effect that Paul's change provided. > > Well that is weird, did you put a tracepoint/printk in > synchronize_sched() to ensure we don't end up calling that?
So I am seeing synchronize_sched called, and its taking the !rcu_gp_is_expedited path when I see the particularly bad latencies. I wonder if I just mucked up applying the patch? thanks -john
