From 7fac83aaf2eecc9e7e7b72da694c49bb4ce7fdfc Mon Sep 17 00:00:00 2001
From: Song Liu <songliubraving@fb.com>
Date: Tue, 16 Mar 2021 14:18:35 -0700
Subject: perf stat: Introduce 'bperf' to share hardware PMCs with BPF
The perf tool uses performance monitoring counters (PMCs) to monitor
system performance. The PMCs are limited hardware resources. For
example, Intel CPUs have 3x fixed PMCs and 4x programmable PMCs per cpu.
Modern data center systems use these PMCs in many different ways: system
level monitoring, (maybe nested) container level monitoring, per process
monitoring, profiling (in sample mode), etc. In some cases, there are
more active perf_events than available hardware PMCs. To allow all
perf_events to have a chance to run, it is necessary to do expensive
time multiplexing of events.
On the other hand, many monitoring tools count the common metrics
(cycles, instructions). It is a waste to have multiple tools create
multiple perf_events of "cycles" and occupy multiple PMCs.
bperf tries to reduce such wastes by allowing multiple perf_events of
"cycles" or "instructions" (at different scopes) to share PMUs. Instead
of having each perf-stat session to read its own perf_events, bperf uses
BPF programs to read the perf_events and aggregate readings to BPF maps.
Then, the perf-stat session(s) reads the values from these BPF maps.
Please refer to the comment before the definition of bperf_ops for the
description of bperf architecture.
bperf is off by default. To enable it, pass --bpf-counters option to
perf-stat. bperf uses a BPF hashmap to share information about BPF
programs and maps used by bperf. This map is pinned to bpffs. The
default path is /sys/fs/bpf/perf_attr_map. The user could change the
path with option --bpf-attr-map.
Committer testing:
# dmesg|grep "Performance Events" -A5
[ 0.225277] Performance Events: Fam17h+ core perfctr, AMD PMU driver.
[ 0.225280] ... version: 0
[ 0.225280] ... bit width: 48
[ 0.225281] ... generic registers: 6
[ 0.225281] ... value mask: 0000ffffffffffff
[ 0.225281] ... max period: 00007fffffffffff
#
# for a in $(seq 6) ; do perf stat -a -e cycles,instructions sleep 100000 & done
[1] 2436231
[2] 2436232
[3] 2436233
[4] 2436234
[5] 2436235
[6] 2436236
# perf stat -a -e cycles,instructions sleep 0.1
Performance counter stats for 'system wide':
310,326,987 cycles (41.87%)
236,143,290 instructions # 0.76 insn per cycle (41.87%)
0.100800885 seconds time elapsed
#
We can see that the counters were enabled for this workload 41.87% of
the time.
Now with --bpf-counters:
# for a in $(seq 32) ; do perf stat --bpf-counters -a -e cycles,instructions sleep 100000 & done
[1] 2436514
[2] 2436515
[3] 2436516
[4] 2436517
[5] 2436518
[6] 2436519
[7] 2436520
[8] 2436521
[9] 2436522
[10] 2436523
[11] 2436524
[12] 2436525
[13] 2436526
[14] 2436527
[15] 2436528
[16] 2436529
[17] 2436530
[18] 2436531
[19] 2436532
[20] 2436533
[21] 2436534
[22] 2436535
[23] 2436536
[24] 2436537
[25] 2436538
[26] 2436539
[27] 2436540
[28] 2436541
[29] 2436542
[30] 2436543
[31] 2436544
[32] 2436545
#
# ls -la /sys/fs/bpf/perf_attr_map
-rw-------. 1 root root 0 Mar 23 14:53 /sys/fs/bpf/perf_attr_map
# bpftool map | grep bperf | wc -l
64
#
# bpftool map | tail
1265: percpu_array name accum_readings flags 0x0
key 4B value 24B max_entries 1 memlock 4096B
1266: hash name filter flags 0x0
key 4B value 4B max_entries 1 memlock 4096B
1267: array name bperf_fo.bss flags 0x400
key 4B value 8B max_entries 1 memlock 4096B
btf_id 996
pids perf(2436545)
1268: percpu_array name accum_readings flags 0x0
key 4B value 24B max_entries 1 memlock 4096B
1269: hash name filter flags 0x0
key 4B value 4B max_entries 1 memlock 4096B
1270: array name bperf_fo.bss flags 0x400
key 4B value 8B max_entries 1 memlock 4096B
btf_id 997
pids perf(2436541)
1285: array name pid_iter.rodata flags 0x480
key 4B value 4B max_entries 1 memlock 4096B
btf_id 1017 frozen
pids bpftool(2437504)
1286: array flags 0x0
key 4B value 32B max_entries 1 memlock 4096B
#
# bpftool map dump id 1268 | tail
value (CPU 21):
8f f3 bc ca 00 00 00 00 80 fd 2a d1 4d 00 00 00
80 fd 2a d1 4d 00 00 00
value (CPU 22):
7e d5 64 4d 00 00 00 00 a4 8a 2e ee 4d 00 00 00
a4 8a 2e ee 4d 00 00 00
value (CPU 23):
a7 78 3e 06 01 00 00 00 b2 34 94 f6 4d 00 00 00
b2 34 94 f6 4d 00 00 00
Found 1 element
# bpftool map dump id 1268 | tail
value (CPU 21):
c6 8b d9 ca 00 00 00 00 20 c6 fc 83 4e 00 00 00
20 c6 fc 83 4e 00 00 00
value (CPU 22):
9c b4 d2 4d 00 00 00 00 3e 0c df 89 4e 00 00 00
3e 0c df 89 4e 00 00 00
value (CPU 23):
18 43 66 06 01 00 00 00 5b 69 ed 83 4e 00 00 00
5b 69 ed 83 4e 00 00 00
Found 1 element
# bpftool map dump id 1268 | tail
value (CPU 21):
f2 6e db ca 00 00 00 00 92 67 4c ba 4e 00 00 00
92 67 4c ba 4e 00 00 00
value (CPU 22):
dc 8e e1 4d 00 00 00 00 d9 32 7a c5 4e 00 00 00
d9 32 7a c5 4e 00 00 00
value (CPU 23):
bd 2b 73 06 01 00 00 00 7c 73 87 bf 4e 00 00 00
7c 73 87 bf 4e 00 00 00
Found 1 element
#
# perf stat --bpf-counters -a -e cycles,instructions sleep 0.1
Performance counter stats for 'system wide':
119,410,122 cycles
152,105,479 instructions # 1.27 insn per cycle
0.101395093 seconds time elapsed
#
See? We had the counters enabled all the time.
Signed-off-by: Song Liu <songliubraving@fb.com>
Reviewed-by: Jiri Olsa <jolsa@kernel.org>
Acked-by: Namhyung Kim <namhyung@kernel.org>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: kernel-team@fb.com
Link: http://lore.kernel.org/lkml/20210316211837.910506-2-songliubraving@fb.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
---
tools/perf/builtin-stat.c | 10 ++++++++++
1 file changed, 10 insertions(+)
(limited to 'tools/perf/builtin-stat.c')
diff --git a/tools/perf/builtin-stat.c b/tools/perf/builtin-stat.c
index d140ffbb5b34..d34e22dce2c8 100644
--- a/tools/perf/builtin-stat.c
+++ b/tools/perf/builtin-stat.c
@@ -792,6 +792,12 @@ static int __run_perf_stat(int argc, const char **argv, int run_idx)
}
evlist__for_each_cpu (evsel_list, i, cpu) {
+ /*
+ * bperf calls evsel__open_per_cpu() in bperf__load(), so
+ * no need to call it again here.
+ */
+ if (target.use_bpf)
+ break;
affinity__set(&affinity, cpu);
evlist__for_each_entry(evsel_list, counter) {
@@ -1146,6 +1152,10 @@ static struct option stat_options[] = {
#ifdef HAVE_BPF_SKEL
OPT_STRING('b', "bpf-prog", &target.bpf_str, "bpf-prog-id",
"stat events on existing bpf program id"),
+ OPT_BOOLEAN(0, "bpf-counters", &target.use_bpf,
+ "use bpf program to count events"),
+ OPT_STRING(0, "bpf-attr-map", &target.attr_map, "attr-map-path",
+ "path to perf_event_attr map"),
#endif
OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
"system-wide collection from all CPUs"),
--
cgit v1.2.3