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>

1 files changed, 46 insertions, 0 deletions

diff --git a/tools/perf/util/bpf_skel/bperf_leader.bpf.c b/tools/perf/util/bpf_skel/bperf_leader.bpf.c
new file mode 100644
index 000000000000..4f70d1459e86
--- /dev/null
+++ b/tools/perf/util/bpf_skel/bperf_leader.bpf.c
@@ -0,0 +1,46 @@
+// SPDX-License-Identifier: (GPL-2.0-only OR BSD-2-Clause)
+// Copyright (c) 2021 Facebook
+#include <linux/bpf.h>
+#include <linux/perf_event.h>
+#include <bpf/bpf_helpers.h>
+#include <bpf/bpf_tracing.h>
+#include "bperf.h"
+
+struct {
+	__uint(type, BPF_MAP_TYPE_PERF_EVENT_ARRAY);
+	__uint(key_size, sizeof(__u32));
+	__uint(value_size, sizeof(int));
+	__uint(map_flags, BPF_F_PRESERVE_ELEMS);
+} events SEC(".maps");
+
+reading_map prev_readings SEC(".maps");
+reading_map diff_readings SEC(".maps");
+
+SEC("raw_tp/sched_switch")
+int BPF_PROG(on_switch)
+{
+	struct bpf_perf_event_value val, *prev_val, *diff_val;
+	__u32 key = bpf_get_smp_processor_id();
+	__u32 zero = 0;
+	long err;
+
+	prev_val = bpf_map_lookup_elem(&prev_readings, &zero);
+	if (!prev_val)
+		return 0;
+
+	diff_val = bpf_map_lookup_elem(&diff_readings, &zero);
+	if (!diff_val)
+		return 0;
+
+	err = bpf_perf_event_read_value(&events, key, &val, sizeof(val));
+	if (err)
+		return 0;
+
+	diff_val->counter = val.counter - prev_val->counter;
+	diff_val->enabled = val.enabled - prev_val->enabled;
+	diff_val->running = val.running - prev_val->running;
+	*prev_val = val;
+	return 0;
+}
+
+char LICENSE[] SEC("license") = "Dual BSD/GPL";