On Fri, Jul 01, 2022 at 05:43:26PM +0800, David Gow wrote: > On Fri, Jul 1, 2022 at 5:16 PM Vincent Whitchurch > <vincent.whitchurch@xxxxxxxx> wrote: > > On Fri, Jul 01, 2022 at 11:08:27AM +0200, David Gow wrote: > > > On Thu, Jun 30, 2022 at 9:29 PM Andrey Konovalov <andreyknvl@xxxxxxxxx> wrote: > > > > Stack trace collection code might trigger KASAN splats when walking > > > > stack frames, but this can be resolved by using unchecked accesses. > > > > The main reason to disable instrumentation here is for performance > > > > reasons, see the upcoming patch for arm64 [1] for some details. > > > > > > > > [1] https://git.kernel.org/pub/scm/linux/kernel/git/arm64/linux.git/commit/?id=802b91118d11 > > > > > > Ah -- that does it! Using READ_ONCE_NOCHECK() in dump_trace() gets rid > > > of the nasty recursive KASAN failures we were getting in the tests. > > > > > > I'll send out v5 with those files instrumented again. > > > > Hmm, do we really want that? In the patch Andrey linked to above he > > removed the READ_ONCE_NOCHECK() and added the KASAN_SANITIZE on the > > corresponding files for arm64, just like it's already the case in this > > patch for UML. > > Personally, I'm okay with the performance overhead so far: in my tests > with a collection of ~350 KUnit tests, the total difference in runtime > was about ~.2 seconds, and was within the margin of error caused by > fluctuations in the compilation time. > > As an example, without the stacktrace code instrumented: > [17:36:50] Testing complete. Passed: 364, Failed: 0, Crashed: 0, > Skipped: 47, Errors: 0 > [17:36:50] Elapsed time: 15.114s total, 0.003s configuring, 8.518s > building, 6.433s running > > versus with it instrumented: > [17:35:40] Testing complete. Passed: 364, Failed: 0, Crashed: 0, > Skipped: 47, Errors: 0 > [17:35:40] Elapsed time: 15.497s total, 0.003s configuring, 8.691s > building, 6.640s running OK, good to know. > That being said, I'm okay with disabling it again and adding a comment > if it's slow enough in some other usecase to cause problems (or even > just be annoying). That could either be done in a v6 of this patchset, > or a follow-up patch, depending on what people would prefer. But I'd > not have a problem with leaving it instrumented for now. I don't have any strong opinion either way either, so you don't have to change it back on my account. Thanks.
