[PATCH 46/49] perf record: Add --thread-stats option support

From: Jiri Olsa
Date: Tue Jan 09 2018 - 10:37:53 EST


Add per-thread stats to have an idea what's happening
in the main reading loop.

$ perf --debug threads=2 record ...
SNIP
pid write poll skip
1s 8914 136B 1 0
2s 8914 512K 43 79
3s 8914 3M 214 385
4s 8914 3M 121 291

$ perf --debug threads=2 record --threads ...
SNIP
pid write poll skip
1s 9770 144B 1 0
9772 0B 1 0
9773 0B 1 0
9774 0B 1 0
2s 9770 290K 35 37
9772 272K 36 34
9773 274K 35 35
9774 304K 39 39
3s 9770 1120K 140 140
9772 1088K 138 138
9773 1120K 140 140
9774 1123K 140 140
4s 9770 1161K 146 146
9772 1121K 142 142
9773 1135K 142 142
9774 1159K 145 145

Link: http://lkml.kernel.org/n/tip-z9un5mjzsh47u9m12ijn7pfq@xxxxxxxxxxxxxx
Signed-off-by: Jiri Olsa <jolsa@xxxxxxxxxx>
---
tools/perf/builtin-record.c | 79 +++++++++++++++++++++++++++++++++++++++++++--
1 file changed, 76 insertions(+), 3 deletions(-)

diff --git a/tools/perf/builtin-record.c b/tools/perf/builtin-record.c
index ba72aed6a411..3941bfa3d8aa 100644
--- a/tools/perf/builtin-record.c
+++ b/tools/perf/builtin-record.c
@@ -69,7 +69,14 @@ enum {
RECORD_THREAD__STOP = 1,
};

+struct thread_stat {
+ u64 bytes_written;
+ u64 poll;
+ u64 poll_skip;
+};
+
struct record_thread {
+ int pid;
struct perf_mmap **mmap;
int mmap_nr;
struct perf_mmap **ovw_mmap;
@@ -80,6 +87,7 @@ struct record_thread {
u64 bytes_written;
pthread_t pt;
int state;
+ struct thread_stat stats;
};

struct record {
@@ -149,7 +157,8 @@ static int record__write(struct record *rec, struct perf_mmap *map,
return -1;
}

- thread->bytes_written += size;
+ thread->bytes_written += size;
+ thread->stats.bytes_written += size;

if (switch_output_size(rec))
trigger_hit(&switch_output_trigger);
@@ -1155,6 +1164,11 @@ record__threads_config(struct record *rec)
return ret;
}

+static inline pid_t gettid(void)
+{
+ return (pid_t) syscall(__NR_gettid);
+}
+
static void*
record_thread__process(struct record *rec)
{
@@ -1166,6 +1180,8 @@ record_thread__process(struct record *rec)
break;

if (hits == thread->samples) {
+ thread->stats.poll++;
+
err = fdarray__poll(&thread->pollfd, 500);
/*
* Propagate error, only if there's any. Ignore positive
@@ -1178,6 +1194,8 @@ record_thread__process(struct record *rec)
if (fdarray__filter(&thread->pollfd, POLLERR|POLLHUP,
perf_mmap__put_filtered, NULL) == 0)
break;
+ } else {
+ thread->stats.poll_skip++;
}
}

@@ -1210,6 +1228,7 @@ static void *worker(void *arg)
struct record *rec = th->rec;

thread = th;
+ thread->pid = gettid();
thread->state = RECORD_THREAD__RUNNING;

signal_main(rec);
@@ -1251,6 +1270,50 @@ static int record__threads_stop(struct record *rec)
return err;
}

+static void record_thread__display(struct record_thread *t, unsigned long s)
+{
+ char buf_size[20];
+ char buf_time[20];
+
+ unit_number__scnprintf(buf_size, sizeof(buf_size), t->stats.bytes_written);
+
+ if (s)
+ scnprintf(buf_time, sizeof(buf_time), "%5lus", s);
+ else
+ buf_time[0] = 0;
+
+ fprintf(stderr, "%6s %6d %10s %10" PRIu64" %10" PRIu64"\n",
+ buf_time, t->pid, buf_size, t->stats.poll, t->stats.poll_skip);
+}
+
+static void record__threads_stats(struct record *rec)
+{
+ struct record_thread *threads = rec->threads;
+ static time_t last, last_header, start;
+ time_t current = time(NULL);
+ int i;
+
+ if (last == current)
+ return;
+
+ if (!start)
+ start = current - 1;
+
+ last = current;
+
+ if (!last_header || (last_header + 10 < current)) {
+ fprintf(stderr, "%6s %6s %10s %10s %10s\n", " ", "pid", "write", "poll", "skip");
+ last_header = current;
+ }
+
+ for (i = 0; i < rec->threads_cnt; i++) {
+ struct record_thread *t = threads + i;
+
+ record_thread__display(t, !i ? current - start : 0);
+ memset(&t->stats, 0, sizeof(t->stats));
+ }
+}
+
static int __cmd_record(struct record *rec, int argc, const char **argv)
{
int err;
@@ -1329,6 +1392,7 @@ static int __cmd_record(struct record *rec, int argc, const char **argv)
}

thread = &rec->threads[0];
+ thread->pid = gettid();

err = bpf__apply_obj_config();
if (err) {
@@ -1528,7 +1592,10 @@ static int __cmd_record(struct record *rec, int argc, const char **argv)
if (hits == thread->samples) {
if (done || draining)
break;
- err = fdarray__poll(&thread->pollfd, -1);
+
+ err = fdarray__poll(&thread->pollfd, 1000);
+ thread->stats.poll++;
+
/*
* Propagate error, only if there's any. Ignore positive
* number of returned events and interrupt error.
@@ -1537,10 +1604,16 @@ static int __cmd_record(struct record *rec, int argc, const char **argv)
err = 0;
rec->waking++;

- if (perf_evlist__filter_pollfd(rec->evlist, POLLERR | POLLHUP) == 0)
+ if (fdarray__filter(&thread->pollfd, POLLERR|POLLHUP,
+ perf_mmap__put_filtered, NULL) == 0)
draining = true;
+ } else {
+ thread->stats.poll_skip++;
}

+ if (debug_threads)
+ record__threads_stats(rec);
+
/*
* When perf is starting the traced process, at the end events
* die with the process and we wait for that. Thus no need to
--
2.13.6