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] 2436236e
  # 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.comSigned-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>

tools/perf/Documentation/perf-stat.txt

@@ -93,6 +93,17 @@ report::
 
         1.102235068 seconds time elapsed
 
+--bpf-counters::
+	Use BPF programs to aggregate readings from perf_events.  This
+	allows multiple perf-stat sessions that are counting the same metric (cycles,
+	instructions, etc.) to share hardware counters.
+
+--bpf-attr-map::
+	With option "--bpf-counters", different perf-stat sessions share
+	information about shared BPF programs and maps via a pinned hashmap.
+	Use "--bpf-attr-map" to specify the path of this pinned hashmap.
+	The default path is /sys/fs/bpf/perf_attr_map.
+
 ifdef::HAVE_LIBPFM[]
 --pfm-events events::
 Select a PMU event using libpfm4 syntax (see http://perfmon2.sf.net)

tools/perf/Makefile.perf

@@ -1007,6 +1007,7 @@ python-clean:
 SKEL_OUT := $(abspath $(OUTPUT)util/bpf_skel)
 SKEL_TMP_OUT := $(abspath $(SKEL_OUT)/.tmp)
 SKELETONS := $(SKEL_OUT)/bpf_prog_profiler.skel.h
+SKELETONS += $(SKEL_OUT)/bperf_leader.skel.h $(SKEL_OUT)/bperf_follower.skel.h
 
 ifdef BUILD_BPF_SKEL
 BPFTOOL := $(SKEL_TMP_OUT)/bootstrap/bpftool

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"),

tools/perf/util/bpf_skel/bperf.h

+// SPDX-License-Identifier: (GPL-2.0-only OR BSD-2-Clause)
+// Copyright (c) 2021 Facebook
+
+#ifndef __BPERF_STAT_H
+#define __BPERF_STAT_H
+
+typedef struct {
+	__uint(type, BPF_MAP_TYPE_PERCPU_ARRAY);
+	__uint(key_size, sizeof(__u32));
+	__uint(value_size, sizeof(struct bpf_perf_event_value));
+	__uint(max_entries, 1);
+} reading_map;
+
+#endif /* __BPERF_STAT_H */

tools/perf/util/bpf_skel/bperf_follower.bpf.c

+// 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"
+#include "bperf_u.h"
+
+reading_map diff_readings SEC(".maps");
+reading_map accum_readings SEC(".maps");
+
+struct {
+	__uint(type, BPF_MAP_TYPE_HASH);
+	__uint(key_size, sizeof(__u32));
+	__uint(value_size, sizeof(__u32));
+} filter SEC(".maps");
+
+enum bperf_filter_type type = 0;
+int enabled = 0;
+
+SEC("fexit/XXX")
+int BPF_PROG(fexit_XXX)
+{
+	struct bpf_perf_event_value *diff_val, *accum_val;
+	__u32 filter_key, zero = 0;
+	__u32 *accum_key;
+
+	if (!enabled)
+		return 0;
+
+	switch (type) {
+	case BPERF_FILTER_GLOBAL:
+		accum_key = &zero;
+		goto do_add;
+	case BPERF_FILTER_CPU:
+		filter_key = bpf_get_smp_processor_id();
+		break;
+	case BPERF_FILTER_PID:
+		filter_key = bpf_get_current_pid_tgid() & 0xffffffff;
+		break;
+	case BPERF_FILTER_TGID:
+		filter_key = bpf_get_current_pid_tgid() >> 32;
+		break;
+	default:
+		return 0;
+	}
+
+	accum_key = bpf_map_lookup_elem(&filter, &filter_key);
+	if (!accum_key)
+		return 0;
+
+do_add:
+	diff_val = bpf_map_lookup_elem(&diff_readings, &zero);
+	if (!diff_val)
+		return 0;
+
+	accum_val = bpf_map_lookup_elem(&accum_readings, accum_key);
+	if (!accum_val)
+		return 0;
+
+	accum_val->counter += diff_val->counter;
+	accum_val->enabled += diff_val->enabled;
+	accum_val->running += diff_val->running;
+
+	return 0;
+}
+
+char LICENSE[] SEC("license") = "Dual BSD/GPL";

tools/perf/util/bpf_skel/bperf_leader.bpf.c

+// 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";

tools/perf/util/bpf_skel/bperf_u.h

+// SPDX-License-Identifier: (GPL-2.0-only OR BSD-2-Clause)
+// Copyright (c) 2021 Facebook
+
+#ifndef __BPERF_STAT_U_H
+#define __BPERF_STAT_U_H
+
+enum bperf_filter_type {
+	BPERF_FILTER_GLOBAL = 1,
+	BPERF_FILTER_CPU,
+	BPERF_FILTER_PID,
+	BPERF_FILTER_TGID,
+};
+
+#endif /* __BPERF_STAT_U_H */

tools/perf/util/evsel.h

@@ -20,6 +20,8 @@ union perf_event;
 struct bpf_counter_ops;
 struct target;
 struct hashmap;
+struct bperf_leader_bpf;
+struct bperf_follower_bpf;
 
 typedef int (evsel__sb_cb_t)(union perf_event *event, void *data);
 
@@ -130,8 +132,24 @@ struct evsel {
 	 * See also evsel__has_callchain().
 	 */
 	__u64			synth_sample_type;
-	struct list_head	bpf_counter_list;
+
+	/*
+	 * bpf_counter_ops serves two use cases:
+	 *   1. perf-stat -b          counting events used byBPF programs
+	 *   2. perf-stat --use-bpf   use BPF programs to aggregate counts
+	 */
 	struct bpf_counter_ops	*bpf_counter_ops;
+
+	/* for perf-stat -b */
+	struct list_head	bpf_counter_list;
+
+	/* for perf-stat --use-bpf */
+	int			bperf_leader_prog_fd;
+	int			bperf_leader_link_fd;
+	union {
+		struct bperf_leader_bpf *leader_skel;
+		struct bperf_follower_bpf *follower_skel;
+	};
 };
 
 struct perf_missing_features {

tools/perf/util/target.h

@@ -16,6 +16,8 @@ struct target {
 	bool	     uses_mmap;
 	bool	     default_per_cpu;
 	bool	     per_thread;
+	bool	     use_bpf;
+	const char   *attr_map;
 };
 
 enum target_errno {
@@ -66,7 +68,7 @@ static inline bool target__has_cpu(struct target *target)
 
 static inline bool target__has_bpf(struct target *target)
 {
-	return target->bpf_str;
+	return target->bpf_str || target->use_bpf;
 }
 
 static inline bool target__none(struct target *target)