On Thu, Jun 06, 2019 at 10:28:12AM +0000, Jan Glauber wrote: > On Thu, Jun 06, 2019 at 10:41:54AM +0100, Will Deacon wrote: > > On Thu, Jun 06, 2019 at 08:03:27AM +0000, Jan Glauber wrote: > > > On Wed, Jun 05, 2019 at 01:16:46PM -0700, Linus Torvalds wrote: > > > > On Wed, Jun 5, 2019 at 6:49 AM Jan Glauber <jglau...@cavium.com> wrote: > > > > > > > > > > Add an upper bound to the loop to force the fallback to spinlocks > > > > > after some time. A retry value of 100 should not impact any hardware > > > > > that does not have this issue. > > > > > > > > > > With the retry limit the performance of an open-close testcase > > > > > improved between 60-70% on ThunderX2. > > > > > > > > Btw, did you do any kind of performance analysis across different > > > > retry limit values? > > > > > > I tried 15/50/100/200/500, results were largely identical up to 100. > > > For SMT=4 a higher retry value might be better, but unless we can add a > > > sysctl value 100 looked like a good compromise to me. > > > > Perhaps I'm just getting confused pre-morning-coffee, but I thought the > > original complaint (and the reason for this patch even existing) was that > > when many CPUs were hammering the lockref then performance tanked? In which > > case, increasing the threshold as the number of CPUs increases seems > > counter-intuitive to me because it suggests that the larger the system, > > the harder we should try to make the cmpxchg work. > > For SMT=4 the top hit I see is queued_spin_lock_slowpath(). Maybe this is more > costly with more threads, so trying harder to use lockref-cmpxchg makes > the microbenchmark faster in that case?
To clarify, with 224 threads & CPUs queued_spin_lock_slowpath is the top hit even without a retry limit in lockref. This could be unrelated to the lockref fallback, it looks like it's coming from the spinlock in: do_sys_open -> get_unused_fd_flags -> __alloc_fd --Jan