On Thu, Apr 14, 2022 at 01:36:59PM +0100, Mark Rutland wrote: > As I suspected, you're hitting a known performance oddity with QEMU TCG > mode where pointer authentication is *incredibly* slow when using the > architected QARMA5 algorithm (enabled by default with `-cpu max`). > This overhead has nothing to do with the *nature* of the unwinder, and > is an artifact of the *platform* and the *structure* of the code. > There's plenty that can be done to avoid that overhead FWIW, from a quick look, disabling KASAN instrumentation for the stacktrace object alone (with no other changes) has a significant impact (compounded by the TCG QARMA5 slowdown), and I note that x86 doesn't both instrumenting its stacktrace code anyway, so we could consider doing likewise. Atop that, replacing set_bit() with __set_bit() brings the regular unwinder *really* close to the earlier SCS unwinder figures. I know that the on_accessible_stack() calculations and checks could be ammortized with some refactoring (which I'd planned to do anyway), so I think it's plausible that with some changes to the existing unwinder we can bring the difference into the noise. > generic kasan w/ `-cpu max` > --------------------------- > > master-no-stack-traces: 12.66 > master: 18.39 (+45.2%) > master-no-stack-depot: 17.85 (+40.1%) > up-scs-stacks-v3: 13.54 (+7.0%) master-noasan: 15.67 (+23.8%) master-noasan-__set_bit: 14.61 (+15.5%) > Generic KASAN w/ `-cpu max,pauth-impdef=true` > --------------------------------------------- > > master-no-stack-traces: 2.69 > master: 3.35 (+24.5%) > master-no-stack-depot: 3.54 (+31.5%) > up-scs-stacks-v3: 2.80 (+4.1%) master-noasan: 3.05 (+13.0%) master-noasan-__set_bit: 2.96 (+10.0%) > Generic KASAN w/ `-cpu max,pauth=false` > --------------------------------------- > > master-no-stack-traces: 1.92 > master: 2.27 (+18.2%) > master-no-stack-depot: 2.22 (+15.6%) > up-scs-stacks-v3: 2.06 (+7.3%) master-noasan: 2.14 (+11.4%) master-noasan-__set_bit: 2.10 (+9.4%) Thanks, Mark.
