[PATCH -stable 4.1 1/4] perf/core: Fix dynamic interrupt throttle

From: Tan Xiaojun
Date: Mon Mar 06 2017 - 01:59:27 EST


From: Peter Zijlstra <peterz@xxxxxxxxxxxxx>

There were two problems with the dynamic interrupt throttle mechanism,
both triggered by the same action.

When you (or perf_fuzzer) write a huge value into
/proc/sys/kernel/perf_event_max_sample_rate the computed
perf_sample_allowed_ns becomes 0. This effectively disables the whole
dynamic throttle.

This is fixed by ensuring update_perf_cpu_limits() never sets the
value to 0. However, we allow disabling of the dynamic throttle by
writing 100 to /proc/sys/kernel/perf_cpu_time_max_percent. This will
generate a warning in dmesg.

The second problem is that by setting the max_sample_rate to a huge
number, the adaptive process can take a few tries, since it halfs the
limit each time. Change that to directly compute a new value based on
the observed duration.

Signed-off-by: Peter Zijlstra (Intel) <peterz@xxxxxxxxxxxxx>
Cc: Alexander Shishkin <alexander.shishkin@xxxxxxxxxxxxxxx>
Cc: Andy Lutomirski <luto@xxxxxxxxxxxxxx>
Cc: Arnaldo Carvalho de Melo <acme@xxxxxxxxxx>
Cc: Borislav Petkov <bp@xxxxxxxxx>
Cc: Brian Gerst <brgerst@xxxxxxxxx>
Cc: David Ahern <dsahern@xxxxxxxxx>
Cc: Denys Vlasenko <dvlasenk@xxxxxxxxxx>
Cc: H. Peter Anvin <hpa@xxxxxxxxx>
Cc: Jiri Olsa <jolsa@xxxxxxxxxx>
Cc: Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx>
Cc: Namhyung Kim <namhyung@xxxxxxxxxx>
Cc: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
Cc: Stephane Eranian <eranian@xxxxxxxxxx>
Cc: Thomas Gleixner <tglx@xxxxxxxxxxxxx>
Cc: Vince Weaver <vincent.weaver@xxxxxxxxx>
Signed-off-by: Ingo Molnar <mingo@xxxxxxxxxx>
Signed-off-by: Tan Xiaojun <tanxiaojun@xxxxxxxxxx>
---
kernel/events/core.c | 87 ++++++++++++++++++++++++++++++----------------------
1 file changed, 51 insertions(+), 36 deletions(-)

diff --git a/kernel/events/core.c b/kernel/events/core.c
index 6da64f0..3089a004 100644
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -198,8 +198,11 @@ void update_perf_cpu_limits(void)
u64 tmp = perf_sample_period_ns;

tmp *= sysctl_perf_cpu_time_max_percent;
- do_div(tmp, 100);
- ACCESS_ONCE(perf_sample_allowed_ns) = tmp;
+ tmp = div_u64(tmp, 100);
+ if (!tmp)
+ tmp = 1;
+
+ WRITE_ONCE(perf_sample_allowed_ns, tmp);
}

static int perf_rotate_context(struct perf_cpu_context *cpuctx);
@@ -231,7 +234,13 @@ int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
if (ret || !write)
return ret;

- update_perf_cpu_limits();
+ if (sysctl_perf_cpu_time_max_percent == 100) {
+ printk(KERN_WARNING
+ "perf: Dynamic interrupt throttling disabled, can hang your system!\n");
+ WRITE_ONCE(perf_sample_allowed_ns, 0);
+ } else {
+ update_perf_cpu_limits();
+ }

return 0;
}
@@ -245,62 +254,68 @@ int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
#define NR_ACCUMULATED_SAMPLES 128
static DEFINE_PER_CPU(u64, running_sample_length);

+static u64 __report_avg;
+static u64 __report_allowed;
+
static void perf_duration_warn(struct irq_work *w)
{
- u64 allowed_ns = ACCESS_ONCE(perf_sample_allowed_ns);
- u64 avg_local_sample_len;
- u64 local_samples_len;
-
- local_samples_len = __this_cpu_read(running_sample_length);
- avg_local_sample_len = local_samples_len/NR_ACCUMULATED_SAMPLES;
-
printk_ratelimited(KERN_WARNING
- "perf interrupt took too long (%lld > %lld), lowering "
- "kernel.perf_event_max_sample_rate to %d\n",
- avg_local_sample_len, allowed_ns >> 1,
- sysctl_perf_event_sample_rate);
+ "perf: interrupt took too long (%lld > %lld), lowering "
+ "kernel.perf_event_max_sample_rate to %d\n",
+ __report_avg, __report_allowed,
+ sysctl_perf_event_sample_rate);
}

static DEFINE_IRQ_WORK(perf_duration_work, perf_duration_warn);

void perf_sample_event_took(u64 sample_len_ns)
{
- u64 allowed_ns = ACCESS_ONCE(perf_sample_allowed_ns);
- u64 avg_local_sample_len;
- u64 local_samples_len;
+ u64 max_len = READ_ONCE(perf_sample_allowed_ns);
+ u64 running_len;
+ u64 avg_len;
+ u32 max;

- if (allowed_ns == 0)
+ if (max_len == 0)
return;

- /* decay the counter by 1 average sample */
- local_samples_len = __this_cpu_read(running_sample_length);
- local_samples_len -= local_samples_len/NR_ACCUMULATED_SAMPLES;
- local_samples_len += sample_len_ns;
- __this_cpu_write(running_sample_length, local_samples_len);
+ /* Decay the counter by 1 average sample. */
+ running_len = __this_cpu_read(running_sample_length);
+ running_len -= running_len/NR_ACCUMULATED_SAMPLES;
+ running_len += sample_len_ns;
+ __this_cpu_write(running_sample_length, running_len);

/*
- * note: this will be biased artifically low until we have
- * seen NR_ACCUMULATED_SAMPLES. Doing it this way keeps us
+ * Note: this will be biased artifically low until we have
+ * seen NR_ACCUMULATED_SAMPLES. Doing it this way keeps us
* from having to maintain a count.
*/
- avg_local_sample_len = local_samples_len/NR_ACCUMULATED_SAMPLES;
-
- if (avg_local_sample_len <= allowed_ns)
+ avg_len = running_len/NR_ACCUMULATED_SAMPLES;
+ if (avg_len <= max_len)
return;

- if (max_samples_per_tick <= 1)
- return;
+ __report_avg = avg_len;
+ __report_allowed = max_len;

- max_samples_per_tick = DIV_ROUND_UP(max_samples_per_tick, 2);
- sysctl_perf_event_sample_rate = max_samples_per_tick * HZ;
- perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate;
+ /*
+ * Compute a throttle threshold 25% below the current duration.
+ */
+ avg_len += avg_len / 4;
+ max = (TICK_NSEC / 100) * sysctl_perf_cpu_time_max_percent;
+ if (avg_len < max)
+ max /= (u32)avg_len;
+ else
+ max = 1;

- update_perf_cpu_limits();
+ WRITE_ONCE(perf_sample_allowed_ns, avg_len);
+ WRITE_ONCE(max_samples_per_tick, max);
+
+ sysctl_perf_event_sample_rate = max * HZ;
+ perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate;

if (!irq_work_queue(&perf_duration_work)) {
- early_printk("perf interrupt took too long (%lld > %lld), lowering "
+ early_printk("perf: interrupt took too long (%lld > %lld), lowering "
"kernel.perf_event_max_sample_rate to %d\n",
- avg_local_sample_len, allowed_ns >> 1,
+ __report_avg, __report_allowed,
sysctl_perf_event_sample_rate);
}
}
--
1.9.1