| Message ID | 20230216141240.3833272-2-mark.rutland@arm.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | arm_pmu: fix fallout from context handling rewrite | expand |
On 2023-02-16 14:12:38 +0000, Mark Rutland wrote: > Janne reports that perf has been broken on Apple M1 as of commit: > > bd27568117664b8b ("perf: Rewrite core context handling") > > That commit replaced the pmu::filter_match() callback with > pmu::filter(), whose return value has the opposite polarity, with true > implying events should be ignored rather than scheduled. While an > attempt was made to update the logic in armv8pmu_filter() and > armpmu_filter() accordingly, the return value remains inverted in a > couple of cases: > > * If the arm_pmu does not have an arm_pmu::filter() callback, > armpmu_filter() will always return whether the CPU is supported rather > than whether the CPU is not supported. > > As a result, the perf core will not schedule events on supported CPUs, > resulting in a loss of events. Additionally, the perf core will > attempt to schedule events on unsupported CPUs, but this will be > rejected by armpmu_add(), which may result in a loss of events from > other PMUs on those unsupported CPUs. > > * If the arm_pmu does have an arm_pmu::filter() callback, and > armpmu_filter() is called on a CPU which is not supported by the > arm_pmu, armpmu_filter() will return false rather than true. > > As a result, the perf core will attempt to schedule events on > unsupported CPUs, but this will be rejected by armpmu_add(), which may > result in a loss of events from other PMUs on those unsupported CPUs. > > This means a loss of events can be seen with any arm_pmu driver, but > with the ARMv8 PMUv3 driver (which is the only arm_pmu driver with an > arm_pmu::filter() callback) the event loss will be more limited and may > go unnoticed, which is how this issue evaded testing so far. > > Fix the CPU filtering by performing this consistently in > armpmu_filter(), and remove the redundant arm_pmu::filter() callback and > armv8pmu_filter() implementation. > > Commit bd2756811766 also silently removed the CHAIN event filtering from > armv8pmu_filter(), which will be addressed by a separate patch without > using the filter callback. > > Fixes: bd27568117664b8b ("perf: Rewrite core context handling") > Reported-by: Janne Grunau <j@jannau.net> > Link: https://lore.kernel.org/asahi/20230215-arm_pmu_m1_regression-v1-1-f5a266577c8d@jannau.net/ > Signed-off-by: Mark Rutland <mark.rutland@arm.com> > Cc: Will Deacon <will@kernel.org> > Cc: Peter Zijlstra <peterz@infradead.org> > Cc: Ravi Bangoria <ravi.bangoria@amd.com> > Cc: Asahi Lina <lina@asahilina.net> > Cc: Eric Curtin <ecurtin@redhat.com> > --- > arch/arm64/kernel/perf_event.c | 7 ------- > drivers/perf/arm_pmu.c | 8 +------- > include/linux/perf/arm_pmu.h | 1 - > 3 files changed, 1 insertion(+), 15 deletions(-) > > diff --git a/arch/arm64/kernel/perf_event.c b/arch/arm64/kernel/perf_event.c > index a5193f2146a6..3e43538f6b72 100644 > --- a/arch/arm64/kernel/perf_event.c > +++ b/arch/arm64/kernel/perf_event.c > @@ -1023,12 +1023,6 @@ static int armv8pmu_set_event_filter(struct hw_perf_event *event, > return 0; > } > > -static bool armv8pmu_filter(struct pmu *pmu, int cpu) > -{ > - struct arm_pmu *armpmu = to_arm_pmu(pmu); > - return !cpumask_test_cpu(smp_processor_id(), &armpmu->supported_cpus); > -} > - > static void armv8pmu_reset(void *info) > { > struct arm_pmu *cpu_pmu = (struct arm_pmu *)info; > @@ -1258,7 +1252,6 @@ static int armv8_pmu_init(struct arm_pmu *cpu_pmu, char *name, > cpu_pmu->stop = armv8pmu_stop; > cpu_pmu->reset = armv8pmu_reset; > cpu_pmu->set_event_filter = armv8pmu_set_event_filter; > - cpu_pmu->filter = armv8pmu_filter; > > cpu_pmu->pmu.event_idx = armv8pmu_user_event_idx; > > diff --git a/drivers/perf/arm_pmu.c b/drivers/perf/arm_pmu.c > index 9b593f985805..40f70f83daba 100644 > --- a/drivers/perf/arm_pmu.c > +++ b/drivers/perf/arm_pmu.c > @@ -550,13 +550,7 @@ static void armpmu_disable(struct pmu *pmu) > static bool armpmu_filter(struct pmu *pmu, int cpu) > { > struct arm_pmu *armpmu = to_arm_pmu(pmu); > - bool ret; > - > - ret = cpumask_test_cpu(cpu, &armpmu->supported_cpus); > - if (ret && armpmu->filter) > - return armpmu->filter(pmu, cpu); > - > - return ret; > + return !cpumask_test_cpu(cpu, &armpmu->supported_cpus); > } > > static ssize_t cpus_show(struct device *dev, > diff --git a/include/linux/perf/arm_pmu.h b/include/linux/perf/arm_pmu.h > index ef914a600087..525b5d64e394 100644 > --- a/include/linux/perf/arm_pmu.h > +++ b/include/linux/perf/arm_pmu.h > @@ -100,7 +100,6 @@ struct arm_pmu { > void (*stop)(struct arm_pmu *); > void (*reset)(void *); > int (*map_event)(struct perf_event *event); > - bool (*filter)(struct pmu *pmu, int cpu); > int num_events; > bool secure_access; /* 32-bit ARM only */ > #define ARMV8_PMUV3_MAX_COMMON_EVENTS 0x40 This works as well. I limited the patch to the minimal fix this this late in the cycle. Tested-by: Janne Grunau <j@jannau.net> thanks, Janne
On Thu, Feb 16, 2023 at 03:35:19PM +0100, Janne Grunau wrote: > On 2023-02-16 14:12:38 +0000, Mark Rutland wrote: > > Fix the CPU filtering by performing this consistently in > > armpmu_filter(), and remove the redundant arm_pmu::filter() callback and > > armv8pmu_filter() implementation. > > > > Commit bd2756811766 also silently removed the CHAIN event filtering from > > armv8pmu_filter(), which will be addressed by a separate patch without > > using the filter callback. [...] > This works as well. I limited the patch to the minimal fix this > this late in the cycle. I did appreciate that you'd made the effort for the minimal fix; had the issue with CHAIN events not existed I would have acked that as-is and done the simplification later. Given the CHAIN issue and given the simplification make the code "obviously correct" I think it's preferable to do both bits now. > Tested-by: Janne Grunau <j@jannau.net> Thanks! Hopefully Will or Peter can pick this up shortly; I'm assuming that Will can take this via the arm64 tree. Mark.
On Thu, Feb 16, 2023 at 03:13:11PM +0000, Mark Rutland wrote: > On Thu, Feb 16, 2023 at 03:35:19PM +0100, Janne Grunau wrote: > > On 2023-02-16 14:12:38 +0000, Mark Rutland wrote: > > > Fix the CPU filtering by performing this consistently in > > > armpmu_filter(), and remove the redundant arm_pmu::filter() callback and > > > armv8pmu_filter() implementation. > > > > > > Commit bd2756811766 also silently removed the CHAIN event filtering from > > > armv8pmu_filter(), which will be addressed by a separate patch without > > > using the filter callback. > > [...] > > > This works as well. I limited the patch to the minimal fix this > > this late in the cycle. > > I did appreciate that you'd made the effort for the minimal fix; had the issue > with CHAIN events not existed I would have acked that as-is and done the > simplification later. Given the CHAIN issue and given the simplification make > the code "obviously correct" I think it's preferable to do both bits now. > > > Tested-by: Janne Grunau <j@jannau.net> > > Thanks! > > Hopefully Will or Peter can pick this up shortly; I'm assuming that Will can > take this via the arm64 tree. I'll grab 'em. Will
diff --git a/arch/arm64/kernel/perf_event.c b/arch/arm64/kernel/perf_event.c index a5193f2146a6..3e43538f6b72 100644 --- a/arch/arm64/kernel/perf_event.c +++ b/arch/arm64/kernel/perf_event.c @@ -1023,12 +1023,6 @@ static int armv8pmu_set_event_filter(struct hw_perf_event *event, return 0; } -static bool armv8pmu_filter(struct pmu *pmu, int cpu) -{ - struct arm_pmu *armpmu = to_arm_pmu(pmu); - return !cpumask_test_cpu(smp_processor_id(), &armpmu->supported_cpus); -} - static void armv8pmu_reset(void *info) { struct arm_pmu *cpu_pmu = (struct arm_pmu *)info; @@ -1258,7 +1252,6 @@ static int armv8_pmu_init(struct arm_pmu *cpu_pmu, char *name, cpu_pmu->stop = armv8pmu_stop; cpu_pmu->reset = armv8pmu_reset; cpu_pmu->set_event_filter = armv8pmu_set_event_filter; - cpu_pmu->filter = armv8pmu_filter; cpu_pmu->pmu.event_idx = armv8pmu_user_event_idx; diff --git a/drivers/perf/arm_pmu.c b/drivers/perf/arm_pmu.c index 9b593f985805..40f70f83daba 100644 --- a/drivers/perf/arm_pmu.c +++ b/drivers/perf/arm_pmu.c @@ -550,13 +550,7 @@ static void armpmu_disable(struct pmu *pmu) static bool armpmu_filter(struct pmu *pmu, int cpu) { struct arm_pmu *armpmu = to_arm_pmu(pmu); - bool ret; - - ret = cpumask_test_cpu(cpu, &armpmu->supported_cpus); - if (ret && armpmu->filter) - return armpmu->filter(pmu, cpu); - - return ret; + return !cpumask_test_cpu(cpu, &armpmu->supported_cpus); } static ssize_t cpus_show(struct device *dev, diff --git a/include/linux/perf/arm_pmu.h b/include/linux/perf/arm_pmu.h index ef914a600087..525b5d64e394 100644 --- a/include/linux/perf/arm_pmu.h +++ b/include/linux/perf/arm_pmu.h @@ -100,7 +100,6 @@ struct arm_pmu { void (*stop)(struct arm_pmu *); void (*reset)(void *); int (*map_event)(struct perf_event *event); - bool (*filter)(struct pmu *pmu, int cpu); int num_events; bool secure_access; /* 32-bit ARM only */ #define ARMV8_PMUV3_MAX_COMMON_EVENTS 0x40
Janne reports that perf has been broken on Apple M1 as of commit: bd27568117664b8b ("perf: Rewrite core context handling") That commit replaced the pmu::filter_match() callback with pmu::filter(), whose return value has the opposite polarity, with true implying events should be ignored rather than scheduled. While an attempt was made to update the logic in armv8pmu_filter() and armpmu_filter() accordingly, the return value remains inverted in a couple of cases: * If the arm_pmu does not have an arm_pmu::filter() callback, armpmu_filter() will always return whether the CPU is supported rather than whether the CPU is not supported. As a result, the perf core will not schedule events on supported CPUs, resulting in a loss of events. Additionally, the perf core will attempt to schedule events on unsupported CPUs, but this will be rejected by armpmu_add(), which may result in a loss of events from other PMUs on those unsupported CPUs. * If the arm_pmu does have an arm_pmu::filter() callback, and armpmu_filter() is called on a CPU which is not supported by the arm_pmu, armpmu_filter() will return false rather than true. As a result, the perf core will attempt to schedule events on unsupported CPUs, but this will be rejected by armpmu_add(), which may result in a loss of events from other PMUs on those unsupported CPUs. This means a loss of events can be seen with any arm_pmu driver, but with the ARMv8 PMUv3 driver (which is the only arm_pmu driver with an arm_pmu::filter() callback) the event loss will be more limited and may go unnoticed, which is how this issue evaded testing so far. Fix the CPU filtering by performing this consistently in armpmu_filter(), and remove the redundant arm_pmu::filter() callback and armv8pmu_filter() implementation. Commit bd2756811766 also silently removed the CHAIN event filtering from armv8pmu_filter(), which will be addressed by a separate patch without using the filter callback. Fixes: bd27568117664b8b ("perf: Rewrite core context handling") Reported-by: Janne Grunau <j@jannau.net> Link: https://lore.kernel.org/asahi/20230215-arm_pmu_m1_regression-v1-1-f5a266577c8d@jannau.net/ Signed-off-by: Mark Rutland <mark.rutland@arm.com> Cc: Will Deacon <will@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Ravi Bangoria <ravi.bangoria@amd.com> Cc: Asahi Lina <lina@asahilina.net> Cc: Eric Curtin <ecurtin@redhat.com> --- arch/arm64/kernel/perf_event.c | 7 ------- drivers/perf/arm_pmu.c | 8 +------- include/linux/perf/arm_pmu.h | 1 - 3 files changed, 1 insertion(+), 15 deletions(-)