On 2019/6/12 12:10, Jayachandran Chandrasekharan Nair wrote: > On Wed, May 22, 2019 at 05:04:17PM +0100, Will Deacon wrote: >> On Sat, May 18, 2019 at 12:00:34PM +0200, Ard Biesheuvel wrote: >>> On Sat, 18 May 2019 at 06:25, Jayachandran Chandrasekharan Nair >>> <jn...@marvell.com> wrote: >>>> >>>> On Mon, May 06, 2019 at 07:10:40PM +0100, Will Deacon wrote: >>>>> On Mon, May 06, 2019 at 06:13:12AM +0000, Jayachandran Chandrasekharan >>>>> Nair wrote: >>>>>> Perhaps someone from ARM can chime in here how the cas/yield combo >>>>>> is expected to work when there is contention. ThunderX2 does not >>>>>> do much with the yield, but I don't expect any ARM implementation >>>>>> to treat YIELD as a hint not to yield, but to get/keep exclusive >>>>>> access to the last failed CAS location. >>>>> >>>>> Just picking up on this as "someone from ARM". >>>>> >>>>> The yield instruction in our implementation of cpu_relax() is *only* there >>>>> as a scheduling hint to QEMU so that it can treat it as an internal >>>>> scheduling hint and run some other thread; see 1baa82f48030 ("arm64: >>>>> Implement cpu_relax as yield"). We can't use WFE or WFI blindly here, as >>>>> it >>>>> could be a long time before we see a wake-up event such as an interrupt. >>>>> Our >>>>> implementation of smp_cond_load_acquire() is much better for that kind of >>>>> thing, but doesn't help at all for a contended CAS loop where the variable >>>>> is actually changing constantly. >>>> >>>> Looking thru the perf output of this case (open/close of a file from >>>> multiple CPUs), I see that refcount is a significant factor in most >>>> kernel configurations - and that too uses cmpxchg (without yield). >>>> x86 has an optimized inline version of refcount that helps >>>> significantly. Do you think this is worth looking at for arm64? >>>> >>> >>> I looked into this a while ago [0], but at the time, we decided to >>> stick with the generic implementation until we encountered a use case >>> that benefits from it. Worth a try, I suppose ... >>> >>> [0] >>> https://lore.kernel.org/linux-arm-kernel/20170903101622.12093-1-ard.biesheu...@linaro.org/ >> >> If JC can show that we benefit from this, it would be interesting to see if >> we can implement the refcount-full saturating arithmetic using the >> LDMIN/LDMAX instructions instead of the current cmpxchg() loops. > > Now that the lockref change is mainline, I think we need to take another > look at this patch. > > Using a fixed up version of Ard's patch above along with Jan's lockref > change upstream, I get significant improvement in scaling for my file > open/read/close testcase[1]. Like I wrote earlier, if I take a > standard Ubuntu arm64 kernel configuration, most of the time for my > test[1] is spent in refcount operations. > > With Ard's changes applied[2], I see that the lockref CAS code becomes > the top function and then the retry limit will kick in as expected. In > my testcase, I see that the queued spinlock case is about 2.5 times > faster than the unbound CAS loop when 224 CPUs are enabled (SMT 4, > 28core, 2socket). > > JC > > [1] https://github.com/jchandra-cavm/refcount-test > [2] https://github.com/jchandra-cavm/linux/commits/refcount-fixes
FWIW, with the patch (Ard's patch plus fixes) above, running the same testcase on ARM64 Kunpeng920 96 CPU core system, I can see about 50% performance boost. I also tested Jan's lockref change without Ard's patch, performance is almost the same. Thanks Hanjun