LoongArch: Fix PMU counter allocation for mixed-type event groups
Subsystems (2)
PATCH v3 1/1 LoongArch: Fix PMU counter allocation for mixed-type event groups
When validating a perf event group, validate_group() unconditionally attempts to allocate hardware PMU counters for the leader, sibling events and the new event being added.
This is incorrect for mixed-type groups. If a PERF_TYPE_SOFTWARE event is part of the group, the current code still tries to allocate a hardware PMU counter for it, which can wrongly consume hardware PMU resources and cause spurious allocation failures.
Fix this by only allocating PMU counters for hardware events during group validation, and skipping software events.
A trimmed down reproducer is as simple as this:
#include <stdio.h>
#include <assert.h>
#include <unistd.h>
#include <string.h>
#include <sys/syscall.h>
#include <linux/perf_event.h>
int
main (int argc, char *argv[])
{
struct perf_event_attr attr = { 0 };
int fds[5];
attr.disabled = 1;
attr.exclude_kernel = 1;
attr.exclude_hv = 1;
attr.read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
PERF_FORMAT_TOTAL_TIME_RUNNING | PERF_FORMAT_ID | PERF_FORMAT_GROUP;
attr.size = sizeof (attr);
attr.type = PERF_TYPE_SOFTWARE;
attr.config = PERF_COUNT_SW_DUMMY;
fds[0] = syscall (SYS_perf_event_open, &attr, 0, -1, -1, 0);
assert (fds[0] >= 0);
attr.type = PERF_TYPE_HARDWARE;
attr.config = PERF_COUNT_HW_CPU_CYCLES;
fds[1] = syscall (SYS_perf_event_open, &attr, 0, -1, fds[0], 0);
assert (fds[1] >= 0);
attr.type = PERF_TYPE_HARDWARE;
attr.config = PERF_COUNT_HW_INSTRUCTIONS;
fds[2] = syscall (SYS_perf_event_open, &attr, 0, -1, fds[0], 0);
assert (fds[2] >= 0);
attr.type = PERF_TYPE_HARDWARE;
attr.config = PERF_COUNT_HW_BRANCH_MISSES;
fds[3] = syscall (SYS_perf_event_open, &attr, 0, -1, fds[0], 0);
assert (fds[3] >= 0);
attr.type = PERF_TYPE_HARDWARE;
attr.config = PERF_COUNT_HW_CACHE_REFERENCES;
fds[4] = syscall (SYS_perf_event_open, &attr, 0, -1, fds[0], 0);
assert (fds[4] >= 0);
printf ("PASSED\n");
return 0;
}---
Changes in v3:
- Fix typo.
Changes in v2:
- Factor out duplicated perf event type checks into an inline helper.
--- arch/loongarch/kernel/perf_event.c | 21 ++++++++++++++++++--- 1 file changed, 18 insertions(+), 3 deletions(-)
arch/loongarch/kernel/perf_event.c +18 -3
@@ -626,6 +626,18 @@ static const struct loongarch_perf_event *loongarch_pmu_map_cache_event(u64 confreturn pev; } +static inline bool loongarch_pmu_event_requires_counter (const struct perf_event *event) +{ +switch (event->attr.type) { +case PERF_TYPE_HARDWARE : +case PERF_TYPE_HW_CACHE : +case PERF_TYPE_RAW : +return true; +default : +return false; + } +} +static int validate_group (struct perf_event *event) {struct cpu_hw_events fake_cpuc;
@@ -633,15 +645,18 @@ static int validate_group(struct perf_event *event)memset (&fake_cpuc,0 ,sizeof (fake_cpuc)); -if (loongarch_pmu_alloc_counter (&fake_cpuc, &leader->hw) <0 ) +if (loongarch_pmu_event_requires_counter (leader) && +loongarch_pmu_alloc_counter (&fake_cpuc, &leader->hw) <0 )return -EINVAL ;for_each_sibling_event (sibling, leader) { -if (loongarch_pmu_alloc_counter (&fake_cpuc, &sibling->hw) <0 ) +if (loongarch_pmu_event_requires_counter (sibling) && +loongarch_pmu_alloc_counter (&fake_cpuc, &sibling->hw) <0 )return -EINVAL ; } -if (loongarch_pmu_alloc_counter (&fake_cpuc, &event->hw) <0 ) +if (loongarch_pmu_event_requires_counter (event) && +loongarch_pmu_alloc_counter (&fake_cpuc, &event->hw) <0 )return -EINVAL ;return 0 ;
Signature · 1 line
2.52.0
Applied, thanks.
Huacai
Quoted 124 lines · Lisa Robinson
When validating a perf event group,
validate_group()unconditionally attempts to allocate hardware PMU counters for the leader, sibling events and the new event being added.This is incorrect for mixed-type groups. If a
PERF_TYPE_SOFTWAREevent is part of the group, the current code still tries to allocate a hardware PMU counter for it, which can wrongly consume hardware PMU resources and cause spurious allocation failures.Fix this by only allocating PMU counters for hardware events during group validation, and skipping software events.
A trimmed down reproducer is as simple as this:
#include <stdio.h> #include <assert.h> #include <unistd.h> #include <string.h> #include <sys/syscall.h> #include <linux/perf_event.h> int main (int argc, char *argv[]) { struct perf_event_attr attr = { 0 }; int fds[5];attr.disabled = 1; attr.exclude_kernel = 1; attr.exclude_hv = 1; attr.read_format = PERF_FORMAT_TOTAL_TIME_ENABLED | PERF_FORMAT_TOTAL_TIME_RUNNING | PERF_FORMAT_ID | PERF_FORMAT_GROUP; attr.size = sizeof (attr); attr.type = PERF_TYPE_SOFTWARE; attr.config = PERF_COUNT_SW_DUMMY; fds[0] = syscall (SYS_perf_event_open, &attr, 0, -1, -1, 0); assert (fds[0] >= 0); attr.type = PERF_TYPE_HARDWARE; attr.config = PERF_COUNT_HW_CPU_CYCLES; fds[1] = syscall (SYS_perf_event_open, &attr, 0, -1, fds[0], 0); assert (fds[1] >= 0); attr.type = PERF_TYPE_HARDWARE; attr.config = PERF_COUNT_HW_INSTRUCTIONS; fds[2] = syscall (SYS_perf_event_open, &attr, 0, -1, fds[0], 0); assert (fds[2] >= 0); attr.type = PERF_TYPE_HARDWARE; attr.config = PERF_COUNT_HW_BRANCH_MISSES; fds[3] = syscall (SYS_perf_event_open, &attr, 0, -1, fds[0], 0); assert (fds[3] >= 0); attr.type = PERF_TYPE_HARDWARE; attr.config = PERF_COUNT_HW_CACHE_REFERENCES; fds[4] = syscall (SYS_perf_event_open, &attr, 0, -1, fds[0], 0); assert (fds[4] >= 0); printf ("PASSED\n"); return 0;}
Fixes b37042b2bb7c ("LoongArch: Add perf events support")Signed-off-by Lisa Robinson <[email protected]>---
Changes in v3:
- Fix typo.
Changes in v2:
- Factor out duplicated perf event type checks into an inline helper.--- arch/loongarch/kernel/perf_event.c | 21 ++++++++++++++++++--- 1 file changed, 18 insertions(+), 3 deletions(-)arch/loongarch/kernel/perf_event.c +18
-4@@ -626,6 +626,18 @@ static const struct loongarch_perf_event *loongarch_pmu_map_cache_event(u64 confreturn pev; } +static inline bool loongarch_pmu_event_requires_counter (const struct perf_event *event) +{ +switch (event->attr.type) { +case PERF_TYPE_HARDWARE : +case PERF_TYPE_HW_CACHE : +case PERF_TYPE_RAW : +return true; +default : +return false; + } +} +static int validate_group (struct perf_event *event) {struct cpu_hw_events fake_cpuc;@@ -633,15 +645,18 @@ static int validate_group(struct perf_event *event)memset (&fake_cpuc,0 ,sizeof (fake_cpuc)); -if (loongarch_pmu_alloc_counter (&fake_cpuc, &leader->hw) <0 ) +if (loongarch_pmu_event_requires_counter (leader) && +loongarch_pmu_alloc_counter (&fake_cpuc, &leader->hw) <0 )return -EINVAL ;for_each_sibling_event (sibling, leader) { -if (loongarch_pmu_alloc_counter (&fake_cpuc, &sibling->hw) <0 ) +if (loongarch_pmu_event_requires_counter (sibling) && +loongarch_pmu_alloc_counter (&fake_cpuc, &sibling->hw) <0 )return -EINVAL ; } -if (loongarch_pmu_alloc_counter (&fake_cpuc, &event->hw) <0 ) +if (loongarch_pmu_event_requires_counter (event) && +loongarch_pmu_alloc_counter (&fake_cpuc, &event->hw) <0 )return -EINVAL ;return 0 ; --2.52.0