> -----Original Message----- > From: Jiri Olsa [mailto:[email protected]] > Sent: Friday, October 2, 2020 10:00 PM > To: Namhyung Kim <[email protected]>; liwei (GF) > <[email protected]> > Cc: Mark Rutland <[email protected]>; Andi Kleen <[email protected]>; > Alexander Shishkin <[email protected]>; Alexey Budankov > <[email protected]>; Adrian Hunter > <[email protected]>; Arnaldo Carvalho de Melo <[email protected]>; > linux-kernel <[email protected]>; Peter Zijlstra > <[email protected]>; Andi Kleen <[email protected]>; Libin (Huawei) > <[email protected]>; Ingo Molnar <[email protected]>; > [email protected] > Subject: Re: [PATCH 1/2] perf stat: Fix segfault when counting armv8_pmu > events > > On Thu, Sep 24, 2020 at 11:36:23PM +0900, Namhyung Kim wrote: > > On Wed, Sep 23, 2020 at 10:19:00PM +0200, Jiri Olsa wrote: > > > On Wed, Sep 23, 2020 at 11:15:06PM +0900, Namhyung Kim wrote: > > > > I think the problem is that armv8_pmu has a cpumask, > > > > and the user requested per-task events. > > > > > > > > The code tried to open the event with a dummy cpu map > > > > since it's not a cpu event, but the pmu has cpu map and > > > > it's passed to evsel. So there's confusion somewhere > > > > whether it should use evsel->cpus or a dummy map. > > > > > > you're right, I have following cpus file in pmu: > > > > > > # cat /sys/devices/armv8_pmuv3_0/cpus > > > 0-3 > > > > > > covering all the cpus.. and once you have cpumask/cpus file, > > > you're system wide by default in current code, but we should > > > not crash ;-) > > > > > > I tried to cover this case in patch below and I probably broke > > > some other use cases, but perhaps we could allow to open counters > > > per cpus for given workload > > > > > > I'll try to look at this more tomorrow > > > > I'm thinking about a different approach, we can ignore cpu map > > for the ARM cpu PMU and use the dummy, not tested ;-) > > > > Thanks > > Namhyung > > > > > > diff --git a/tools/lib/perf/evlist.c b/tools/lib/perf/evlist.c > > index 2208444ecb44..cfcdbd7be066 100644 > > --- a/tools/lib/perf/evlist.c > > +++ b/tools/lib/perf/evlist.c > > @@ -45,6 +45,9 @@ static void __perf_evlist__propagate_maps(struct > perf_evlist *evlist, > > if (!evsel->own_cpus || evlist->has_user_cpus) { > > perf_cpu_map__put(evsel->cpus); > > evsel->cpus = perf_cpu_map__get(evlist->cpus); > > + } else if (!evsel->system_wide && > perf_cpu_map__empty(evlist->cpus)) { > > + perf_cpu_map__put(evsel->cpus); > > + evsel->cpus = perf_cpu_map__get(evlist->cpus); > > } else if (evsel->cpus != evsel->own_cpus) { > > perf_cpu_map__put(evsel->cpus); > > evsel->cpus = perf_cpu_map__get(evsel->own_cpus); > > > > Wei Li, > is this fixing your problem? As I have seen the same crash and suggested another patch: [PATCH] perf evlist: fix memory corruption for Kernel PMU event https://lore.kernel.org/lkml/[email protected]/ Also, I have tested Namhyung Kim's patch on ARM64. It does fix the crash for me. So: Tested-by: Barry Song <[email protected]> Please put the below fixes-tag in commit log: Fixes: 7736627b865d ("perf stat: Use affinity for closing file descriptors") Thanks Barry
