[PATCH V4 0/5] bugs fix for auto-reload mmap read and rdpmc read

From: kan . liang
Date: Mon Feb 12 2018 - 17:21:57 EST


From: Kan Liang <kan.liang@xxxxxxxxxxxxxxx>

------

Changes since V3:
- Apply Peter's patch to fix event update for auto-reload event
- Based on Peter's patch, specially handle case A.
- Introduce specific read function for auto-reload event, not just
large PEBS.

Changes since V2:
- Refined the changelog
- Introduced specific read function for large PEBS.
The previous generic PEBS read function is confusing.
Disabled PMU in pmu::read() path for large PEBS.
Handled the corner case when reload_times == 0.
- Modified the parameter of intel_pmu_save_and_restart_reload()
Discarded local64_cmpxchg
- Added fixes tag
- Added WARN to handle reload_times == 0 || reload_val == 0

Changes since V1:
- Check PERF_X86_EVENT_AUTO_RELOAD before call
intel_pmu_save_and_restore()
- Introduce a special purpose intel_pmu_save_and_restart()
just for AUTO_RELOAD.
- New patch to disable userspace RDPMC usage if large PEBS is enabled.

------

There is a bug when mmap read event->count with large PEBS enabled.
Here is an example.
#./read_count
0x71f0
0x122c0
0x1000000001c54
0x100000001257d
0x200000000bdc5

The bug is caused by two issues.
- In x86_perf_event_update, the calculation of event->count does not
take the auto-reload values into account.
- In x86_pmu_read, it doesn't count the undrained values in large PEBS
buffers.

The first issue was introduced with the auto-reload mechanism enabled
since commit 851559e35fd5 ("perf/x86/intel: Use the PEBS auto reload
mechanism when possible")

Patch 1 fixed the issue in x86_perf_event_update.

The second issue was introduced since commit b8241d20699e
("perf/x86/intel: Implement batched PEBS interrupt handling (large PEBS
interrupt threshold)")

Patch 2-4 fixed the issue in x86_pmu_read.

Besides the two issues, the userspace RDPMC usage is broken for large
PEBS as well.
The RDPMC issue was also introduced since commit b8241d20699e
("perf/x86/intel: Implement batched PEBS interrupt handling (large PEBS
interrupt threshold)")

Patch 5 fixed the RDPMC issue.

The source code of read_count is as below.

struct cpu {
int fd;
struct perf_event_mmap_page *buf;
};

int perf_open(struct cpu *ctx, int cpu)
{
struct perf_event_attr attr = {
.type = PERF_TYPE_HARDWARE,
.size = sizeof(struct perf_event_attr),
.sample_period = 100000,
.config = 0,
.sample_type = PERF_SAMPLE_IP | PERF_SAMPLE_TID |
PERF_SAMPLE_TIME | PERF_SAMPLE_CPU,
.precise_ip = 3,
.mmap = 1,
.comm = 1,
.task = 1,
.mmap2 = 1,
.sample_id_all = 1,
.comm_exec = 1,
};
ctx->buf = NULL;
ctx->fd = syscall(__NR_perf_event_open, &attr, -1, cpu, -1, 0);
if (ctx->fd < 0) {
perror("perf_event_open");
return -1;
}
return 0;
}

void perf_close(struct cpu *ctx)
{
close(ctx->fd);
if (ctx->buf)
munmap(ctx->buf, pagesize);
}

int main(int ac, char **av)
{
struct cpu ctx;
u64 count;

perf_open(&ctx, 0);

while (1) {
sleep(5);

if (read(ctx.fd, &count, 8) != 8) {
perror("counter read");
break;
}
printf("0x%llx\n", count);

}
perf_close(&ctx);
}


Kan Liang (5):
perf/x86/intel: Fix event update for auto-reload
perf/x86: Introduce read function for x86_pmu
perf/x86/intel/ds: Introduce read function for auto-reload event
perf/x86/intel: Fix pmu read for auto-reload
perf/x86: Fix: disable userspace RDPMC usage for large PEBS

arch/x86/events/core.c | 20 ++++-----
arch/x86/events/intel/core.c | 9 +++++
arch/x86/events/intel/ds.c | 96 ++++++++++++++++++++++++++++++++++++++++++--
arch/x86/events/perf_event.h | 3 ++
4 files changed, 115 insertions(+), 13 deletions(-)

--
2.7.4