[PATCH v2 2/2] perf-stat: enable counting events for BPF programs

[Song Liu <songliubraving@fb.com>]

Introduce perf-stat -b option, which counts events for BPF programs, like:

[root@localhost ~]# ~/perf stat -e ref-cycles,cycles -b 254 -I 1000
     1.487903822            115,200      ref-cycles
     1.487903822             86,012      cycles
     2.489147029             80,560      ref-cycles
     2.489147029             73,784      cycles
     3.490341825             60,720      ref-cycles
     3.490341825             37,797      cycles
     4.491540887             37,120      ref-cycles
     4.491540887             31,963      cycles

The example above counts cycles and ref-cycles of BPF program of id 254.
This is similar to bpftool-prog-profile command, but more flexible.

perf-stat -b creates per-cpu perf_event and loads fentry/fexit BPF
programs (monitor-progs) to the target BPF program (target-prog). The
monitor-progs read perf_event before and after the target-prog, and
aggregate the difference in a BPF map. Then the user space reads data
from these maps.

A new struct bpf_counter is introduced to provide common interface that
uses BPF programs/maps to count perf events.

Signed-off-by: Song Liu <songliubraving@fb.com>
---
 tools/perf/Makefile.perf                      |   2 +-
 tools/perf/builtin-stat.c                     |  77 ++++-
 tools/perf/util/Build                         |   1 +
 tools/perf/util/bpf_counter.c                 | 281 ++++++++++++++++++
 tools/perf/util/bpf_counter.h                 |  73 +++++
 .../util/bpf_skel/bpf_prog_profiler.bpf.c     |  96 ++++++
 tools/perf/util/evsel.c                       |  11 +
 tools/perf/util/evsel.h                       |   6 +
 tools/perf/util/stat-display.c                |   4 +-
 tools/perf/util/target.c                      |  34 ++-
 tools/perf/util/target.h                      |  10 +
 11 files changed, 578 insertions(+), 17 deletions(-)
 create mode 100644 tools/perf/util/bpf_counter.c
 create mode 100644 tools/perf/util/bpf_counter.h
 create mode 100644 tools/perf/util/bpf_skel/bpf_prog_profiler.bpf.c

diff --git a/tools/perf/Makefile.perf b/tools/perf/Makefile.perf
index a272bfb60a579..fb7de412152b5 100644
--- a/tools/perf/Makefile.perf
+++ b/tools/perf/Makefile.perf
@@ -1015,7 +1015,7 @@ python-clean:
 
 SKEL_OUT := $(abspath $(OUTPUT)util/bpf_skel)
 SKEL_TMP_OUT := $(abspath $(SKEL_OUT)/.tmp)
-SKELETONS :=
+SKELETONS := $(SKEL_OUT)/bpf_prog_profiler.skel.h
 
 ifdef BUILD_BPF_SKEL
 BPFTOOL := $(SKEL_TMP_OUT)/bpftool-bootstrap
diff --git a/tools/perf/builtin-stat.c b/tools/perf/builtin-stat.c
index f15b2f8aa14d8..a71684446a0e1 100644
--- a/tools/perf/builtin-stat.c
+++ b/tools/perf/builtin-stat.c
@@ -67,6 +67,7 @@
 #include "util/top.h"
 #include "util/affinity.h"
 #include "util/pfm.h"
+#include "util/bpf_counter.h"
 #include "asm/bug.h"
 
 #include <linux/time64.h>
@@ -409,12 +410,31 @@ static int read_affinity_counters(struct timespec *rs)
 	return 0;
 }
 
+static int read_bpf_map_counters(void)
+{
+	struct evsel *counter;
+	int err;
+
+	evlist__for_each_entry(evsel_list, counter) {
+		err = bpf_counter__read(counter);
+		if (err)
+			return err;
+	}
+	return 0;
+}
+
 static void read_counters(struct timespec *rs)
 {
 	struct evsel *counter;
+	int err;
 
-	if (!stat_config.stop_read_counter && (read_affinity_counters(rs) < 0))
-		return;
+	if (!stat_config.stop_read_counter) {
+		err = read_bpf_map_counters();
+		if (err == -EAGAIN)
+			err = read_affinity_counters(rs);
+		if (err < 0)
+			return;
+	}
 
 	evlist__for_each_entry(evsel_list, counter) {
 		if (counter->err)
@@ -496,11 +516,20 @@ static bool handle_interval(unsigned int interval, int *times)
 	return false;
 }
 
-static void enable_counters(void)
+static int enable_counters(void)
 {
+	struct evsel *evsel;
+	int err;
+
+	evlist__for_each_entry(evsel_list, evsel) {
+		err = bpf_counter__enable(evsel);
+		if (err)
+			return err;
+	}
+
 	if (stat_config.initial_delay < 0) {
 		pr_info(EVLIST_DISABLED_MSG);
-		return;
+		return 0;
 	}
 
 	if (stat_config.initial_delay > 0) {
@@ -518,6 +547,7 @@ static void enable_counters(void)
 		if (stat_config.initial_delay > 0)
 			pr_info(EVLIST_ENABLED_MSG);
 	}
+	return 0;
 }
 
 static void disable_counters(void)
@@ -720,7 +750,7 @@ static int __run_perf_stat(int argc, const char **argv, int run_idx)
 	const bool forks = (argc > 0);
 	bool is_pipe = STAT_RECORD ? perf_stat.data.is_pipe : false;
 	struct affinity affinity;
-	int i, cpu;
+	int i, cpu, err;
 	bool second_pass = false;
 
 	if (forks) {
@@ -738,6 +768,11 @@ static int __run_perf_stat(int argc, const char **argv, int run_idx)
 	if (affinity__setup(&affinity) < 0)
 		return -1;
 
+	evlist__for_each_entry(evsel_list, counter) {
+		if (bpf_counter__load(counter, &target))
+			return -1;
+	}
+
 	evlist__for_each_cpu (evsel_list, i, cpu) {
 		affinity__set(&affinity, cpu);
 
@@ -851,7 +886,7 @@ static int __run_perf_stat(int argc, const char **argv, int run_idx)
 	}
 
 	if (STAT_RECORD) {
-		int err, fd = perf_data__fd(&perf_stat.data);
+		int fd = perf_data__fd(&perf_stat.data);
 
 		if (is_pipe) {
 			err = perf_header__write_pipe(perf_data__fd(&perf_stat.data));
@@ -877,7 +912,9 @@ static int __run_perf_stat(int argc, const char **argv, int run_idx)
 
 	if (forks) {
 		perf_evlist__start_workload(evsel_list);
-		enable_counters();
+		err = enable_counters();
+		if (err)
+			return -1;
 
 		if (interval || timeout || evlist__ctlfd_initialized(evsel_list))
 			status = dispatch_events(forks, timeout, interval, &times);
@@ -896,7 +933,9 @@ static int __run_perf_stat(int argc, const char **argv, int run_idx)
 		if (WIFSIGNALED(status))
 			psignal(WTERMSIG(status), argv[0]);
 	} else {
-		enable_counters();
+		err = enable_counters();
+		if (err)
+			return -1;
 		status = dispatch_events(forks, timeout, interval, &times);
 	}
 
@@ -1087,6 +1126,10 @@ static struct option stat_options[] = {
 		   "stat events on existing process id"),
 	OPT_STRING('t', "tid", &target.tid, "tid",
 		   "stat events on existing thread id"),
+#ifdef BUILD_BPF_SKEL
+	OPT_STRING('b', "bpf-prog", &target.bpf_str, "bpf-prog-id",
+		   "stat events on existing bpf program id"),
+#endif
 	OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
 		    "system-wide collection from all CPUs"),
 	OPT_BOOLEAN('g', "group", &group,
@@ -2058,11 +2101,12 @@ int cmd_stat(int argc, const char **argv)
 		"perf stat [<options>] [<command>]",
 		NULL
 	};
-	int status = -EINVAL, run_idx;
+	int status = -EINVAL, run_idx, err;
 	const char *mode;
 	FILE *output = stderr;
 	unsigned int interval, timeout;
 	const char * const stat_subcommands[] = { "record", "report" };
+	char errbuf[BUFSIZ];
 
 	setlocale(LC_ALL, "");
 
@@ -2173,6 +2217,12 @@ int cmd_stat(int argc, const char **argv)
 	} else if (big_num_opt == 0) /* User passed --no-big-num */
 		stat_config.big_num = false;
 
+	err = target__validate(&target);
+	if (err) {
+		target__strerror(&target, err, errbuf, BUFSIZ);
+		pr_warning("%s\n", errbuf);
+	}
+
 	setup_system_wide(argc);
 
 	/*
@@ -2246,8 +2296,6 @@ int cmd_stat(int argc, const char **argv)
 		}
 	}
 
-	target__validate(&target);
-
 	if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide))
 		target.per_thread = true;
 
@@ -2378,9 +2426,10 @@ int cmd_stat(int argc, const char **argv)
 		 * tools remain  -acme
 		 */
 		int fd = perf_data__fd(&perf_stat.data);
-		int err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
-							     process_synthesized_event,
-							     &perf_stat.session->machines.host);
+
+		err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
+							 process_synthesized_event,
+							 &perf_stat.session->machines.host);
 		if (err) {
 			pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
 				   "older tools may produce warnings about this file\n.");
diff --git a/tools/perf/util/Build b/tools/perf/util/Build
index e2563d0154eb6..188521f343470 100644
--- a/tools/perf/util/Build
+++ b/tools/perf/util/Build
@@ -135,6 +135,7 @@ perf-y += clockid.o
 
 perf-$(CONFIG_LIBBPF) += bpf-loader.o
 perf-$(CONFIG_LIBBPF) += bpf_map.o
+perf-$(CONFIG_PERF_BPF_SKEL) += bpf_counter.o
 perf-$(CONFIG_BPF_PROLOGUE) += bpf-prologue.o
 perf-$(CONFIG_LIBELF) += symbol-elf.o
 perf-$(CONFIG_LIBELF) += probe-file.o
diff --git a/tools/perf/util/bpf_counter.c b/tools/perf/util/bpf_counter.c
new file mode 100644
index 0000000000000..6cee11cda3665
--- /dev/null
+++ b/tools/perf/util/bpf_counter.c
@@ -0,0 +1,281 @@
+// SPDX-License-Identifier: GPL-2.0
+
+/* Copyright (c) 2019 Facebook */
+
+#include <limits.h>
+#include <sys/time.h>
+#include <sys/resource.h>
+#include <linux/err.h>
+#include <linux/zalloc.h>
+#include <bpf/bpf.h>
+#include <bpf/btf.h>
+#include <bpf/libbpf.h>
+
+#include "bpf_counter.h"
+#include "counts.h"
+#include "debug.h"
+#include "evsel.h"
+#include "target.h"
+
+#include "bpf_skel/bpf_prog_profiler.skel.h"
+
+static inline void *u64_to_ptr(__u64 ptr)
+{
+	return (void *)(unsigned long)ptr;
+}
+
+static void set_max_rlimit(void)
+{
+	struct rlimit rinf = { RLIM_INFINITY, RLIM_INFINITY };
+
+	setrlimit(RLIMIT_MEMLOCK, &rinf);
+}
+
+static inline struct bpf_counter *bpf_counter_alloc(void)
+{
+	struct bpf_counter *counter;
+
+	counter = zalloc(sizeof(*counter));
+	if (counter)
+		INIT_LIST_HEAD(&counter->list);
+	return counter;
+}
+
+static int bpf_program_profiler__destroy(struct evsel *evsel)
+{
+	struct bpf_counter *counter;
+
+	list_for_each_entry(counter, &evsel->bpf_counter_list, list)
+		bpf_prog_profiler_bpf__destroy(counter->skel);
+	INIT_LIST_HEAD(&evsel->bpf_counter_list);
+	return 0;
+}
+
+static char *bpf_target_prog_name(int tgt_fd)
+{
+	struct bpf_prog_info_linear *info_linear;
+	struct bpf_func_info *func_info;
+	const struct btf_type *t;
+	char *name = NULL;
+	struct btf *btf;
+
+	info_linear = bpf_program__get_prog_info_linear(
+		tgt_fd, 1UL << BPF_PROG_INFO_FUNC_INFO);
+	if (IS_ERR_OR_NULL(info_linear)) {
+		pr_debug2("failed to get info_linear for prog FD %d", tgt_fd);
+		return NULL;
+	}
+
+	if (info_linear->info.btf_id == 0 ||
+	    btf__get_from_id(info_linear->info.btf_id, &btf)) {
+		pr_debug2("prog FD %d doesn't have valid btf", tgt_fd);
+		goto out;
+	}
+
+	func_info = u64_to_ptr(info_linear->info.func_info);
+	t = btf__type_by_id(btf, func_info[0].type_id);
+	if (!t) {
+		pr_debug2("btf %d doesn't have type %d",
+		      info_linear->info.btf_id, func_info[0].type_id);
+		goto out;
+	}
+	name = strdup(btf__name_by_offset(btf, t->name_off));
+out:
+	free(info_linear);
+	return name;
+}
+
+static int bpf_program_profiler_load_one(struct evsel *evsel, u32 prog_id)
+{
+	struct bpf_prog_profiler_bpf *skel;
+	struct bpf_counter *counter;
+	struct bpf_program *prog;
+	char *prog_name;
+	int prog_fd;
+	int err;
+
+	prog_fd = bpf_prog_get_fd_by_id(prog_id);
+	if (prog_fd < 0) {
+		pr_debug("Failed to open fd for bpf prog %u\n", prog_id);
+		return -1;
+	}
+	counter = bpf_counter_alloc();
+	if (!counter)
+		return -1;
+
+	skel = bpf_prog_profiler_bpf__open();
+	if (!skel) {
+		pr_debug("Failed to load bpf skeleton\n");
+		free(counter);
+		return -1;
+	}
+	skel->rodata->num_cpu = evsel__nr_cpus(evsel);
+
+	bpf_map__resize(skel->maps.events, evsel__nr_cpus(evsel));
+	bpf_map__resize(skel->maps.fentry_readings, 1);
+	bpf_map__resize(skel->maps.accum_readings, 1);
+
+	prog_name = bpf_target_prog_name(prog_fd);
+
+	bpf_object__for_each_program(prog, skel->obj) {
+		err = bpf_program__set_attach_target(prog, prog_fd, prog_name);
+		if (err)
+			pr_debug("bpf_program__set_attach_target failed\n");
+	}
+	set_max_rlimit();
+	err = bpf_prog_profiler_bpf__load(skel);
+	if (err)
+		pr_debug("bpf_prog_profiler_bpf__load failed\n");
+
+	counter->skel = skel;
+	list_add(&counter->list, &evsel->bpf_counter_list);
+	return 0;
+}
+
+static int bpf_program_profiler__load(struct evsel *evsel, struct target *target)
+{
+	char *bpf_str, *bpf_str_, *tok, *saveptr = NULL, *p;
+	u32 prog_id;
+	int ret;
+
+	bpf_str_ = bpf_str = strdup(target->bpf_str);
+	if (!bpf_str)
+		return -1;
+
+	while ((tok = strtok_r(bpf_str, ",", &saveptr)) != NULL) {
+		prog_id = strtoul(tok, &p, 10);
+		if (prog_id == 0 || prog_id == UINT_MAX ||
+		    (*p != '\0' && *p != ',')) {
+			pr_debug("Failed to parse bpf prog ids %s\n",
+				 target->bpf_str);
+			return -1;
+		}
+
+		ret = bpf_program_profiler_load_one(evsel, prog_id);
+		if (ret) {
+			bpf_program_profiler__destroy(evsel);
+			free(bpf_str_);
+			return -1;
+		}
+		bpf_str = NULL;
+	}
+	free(bpf_str_);
+	return 0;
+}
+
+static int bpf_program_profiler__enable(struct evsel *evsel)
+{
+	struct bpf_counter *counter;
+	int ret;
+
+	list_for_each_entry(counter, &evsel->bpf_counter_list, list) {
+		ret = bpf_prog_profiler_bpf__attach(counter->skel);
+		if (ret) {
+			bpf_program_profiler__destroy(evsel);
+			return ret;
+		}
+	}
+	return 0;
+}
+
+static int bpf_program_profiler__read(struct evsel *evsel)
+{
+	int num_cpu = evsel__nr_cpus(evsel);
+	struct bpf_perf_event_value values[num_cpu];
+	struct bpf_counter *counter;
+	int reading_map_fd;
+	__u32 key = 0;
+	int err, cpu;
+
+	if (list_empty(&evsel->bpf_counter_list))
+		return -EAGAIN;
+
+	for (cpu = 0; cpu < num_cpu; cpu++) {
+		perf_counts(evsel->counts, cpu, 0)->val = 0;
+		perf_counts(evsel->counts, cpu, 0)->ena = 0;
+		perf_counts(evsel->counts, cpu, 0)->run = 0;
+	}
+	list_for_each_entry(counter, &evsel->bpf_counter_list, list) {
+		struct bpf_prog_profiler_bpf *skel = counter->skel;
+
+		reading_map_fd = bpf_map__fd(skel->maps.accum_readings);
+
+		err = bpf_map_lookup_elem(reading_map_fd, &key, values);
+		if (err) {
+			fprintf(stderr, "failed to read value\n");
+			return err;
+		}
+
+		for (cpu = 0; cpu < num_cpu; cpu++) {
+			perf_counts(evsel->counts, cpu, 0)->val += values[cpu].counter;
+			perf_counts(evsel->counts, cpu, 0)->ena += values[cpu].enabled;
+			perf_counts(evsel->counts, cpu, 0)->run += values[cpu].running;
+		}
+	}
+	return 0;
+}
+
+static int bpf_program_profiler__install_pe(struct evsel *evsel, int cpu,
+					    int fd)
+{
+	struct bpf_prog_profiler_bpf *skel;
+	struct bpf_counter *counter;
+	int ret;
+
+	list_for_each_entry(counter, &evsel->bpf_counter_list, list) {
+		skel = counter->skel;
+		ret = bpf_map_update_elem(bpf_map__fd(skel->maps.events),
+					  &cpu, &fd, BPF_ANY);
+		if (ret)
+			return ret;
+	}
+	return 0;
+}
+
+struct bpf_counter_ops bpf_program_profiler_ops = {
+	.load       = bpf_program_profiler__load,
+	.enable	    = bpf_program_profiler__enable,
+	.read       = bpf_program_profiler__read,
+	.destroy    = bpf_program_profiler__destroy,
+	.install_pe = bpf_program_profiler__install_pe,
+};
+
+int bpf_counter__install_pe(struct evsel *evsel, int cpu, int fd)
+{
+	if (list_empty(&evsel->bpf_counter_list))
+		return 0;
+	return evsel->bpf_counter_ops->install_pe(evsel, cpu, fd);
+}
+
+int bpf_counter__load(struct evsel *evsel, struct target *target)
+{
+	if (target__has_bpf(target))
+		evsel->bpf_counter_ops = &bpf_program_profiler_ops;
+
+	if (evsel->bpf_counter_ops)
+		return evsel->bpf_counter_ops->load(evsel, target);
+	return 0;
+}
+
+int bpf_counter__enable(struct evsel *evsel)
+{
+	if (list_empty(&evsel->bpf_counter_list))
+		return 0;
+	return evsel->bpf_counter_ops->enable(evsel);
+}
+
+int bpf_counter__read(struct evsel *evsel)
+{
+	if (list_empty(&evsel->bpf_counter_list))
+		return -EAGAIN;
+	return evsel->bpf_counter_ops->read(evsel);
+}
+
+int bpf_counter__destroy(struct evsel *evsel)
+{
+	if (list_empty(&evsel->bpf_counter_list))
+		return 0;
+	evsel->bpf_counter_ops->destroy(evsel);
+	evsel->bpf_counter_ops = NULL;
+	return 0;
+}
diff --git a/tools/perf/util/bpf_counter.h b/tools/perf/util/bpf_counter.h
new file mode 100644
index 0000000000000..d6f10077478ee
--- /dev/null
+++ b/tools/perf/util/bpf_counter.h
@@ -0,0 +1,73 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef __PERF_BPF_COUNTER_H
+#define __PERF_BPF_COUNTER_H 1
+
+#include <linux/list.h>
+
+struct evsel;
+struct target;
+struct bpf_counter;
+
+typedef int (*bpf_counter_evsel_op)(struct evsel *evsel);
+typedef int (*bpf_counter_evsel_target_op)(struct evsel *evsel,
+					   struct target *target);
+typedef int (*bpf_counter_evsel_install_pe_op)(struct evsel *evsel,
+					       int cpu,
+					       int fd);
+
+struct bpf_counter_ops {
+	bpf_counter_evsel_target_op load;
+	bpf_counter_evsel_op enable;
+	bpf_counter_evsel_op read;
+	bpf_counter_evsel_op destroy;
+	bpf_counter_evsel_install_pe_op install_pe;
+};
+
+struct bpf_counter {
+	void *skel;
+	struct list_head list;
+};
+
+#ifdef BUILD_BPF_SKEL
+
+int bpf_counter__load(struct evsel *evsel, struct target *target);
+int bpf_counter__enable(struct evsel *evsel);
+int bpf_counter__read(struct evsel *evsel);
+int bpf_counter__destroy(struct evsel *evsel);
+int bpf_counter__install_pe(struct evsel *evsel, int cpu, int fd);
+
+#else
+
+#include<linux/err.h>
+
+static inline int bpf_counter__load(struct evsel *evsel __maybe_unused,
+				    struct target *target __maybe_unused)
+{
+	return 0;
+}
+
+static inline int bpf_counter__enable(struct evsel *evsel __maybe_unused)
+{
+	return 0;
+}
+
+static inline int bpf_counter__read(struct evsel *evsel __maybe_unused)
+{
+	return -EAGAIN;
+}
+
+static inline int bpf_counter__destroy(struct evsel *evsel __maybe_unused)
+{
+	return 0;
+}
+
+static inline int bpf_counter__install_pe(struct evsel *evsel __maybe_unused,
+					  int cpu __maybe_unused,
+					  int fd __maybe_unused)
+{
+	return 0;
+}
+
+#endif /* BUILD_BPF_SKEL */
+
+#endif /* __PERF_BPF_COUNTER_H */
diff --git a/tools/perf/util/bpf_skel/bpf_prog_profiler.bpf.c b/tools/perf/util/bpf_skel/bpf_prog_profiler.bpf.c
new file mode 100644
index 0000000000000..cdde2218af86a
--- /dev/null
+++ b/tools/perf/util/bpf_skel/bpf_prog_profiler.bpf.c
@@ -0,0 +1,96 @@
+// SPDX-License-Identifier: (GPL-2.0-only OR BSD-2-Clause)
+// Copyright (c) 2020 Facebook
+#include <linux/bpf.h>
+#include <bpf/bpf_helpers.h>
+#include <bpf/bpf_tracing.h>
+
+/* map of perf event fds, num_cpu * num_metric entries */
+struct {
+	__uint(type, BPF_MAP_TYPE_PERF_EVENT_ARRAY);
+	__uint(key_size, sizeof(__u32));
+	__uint(value_size, sizeof(int));
+} events SEC(".maps");
+
+/* readings at fentry */
+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);
+} fentry_readings SEC(".maps");
+
+/* accumulated readings */
+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);
+} accum_readings SEC(".maps");
+
+const volatile __u32 num_cpu = 1;
+
+SEC("fentry/XXX")
+int BPF_PROG(fentry_XXX)
+{
+	__u32 key = bpf_get_smp_processor_id();
+	struct bpf_perf_event_value reading;
+	struct bpf_perf_event_value *ptr;
+	__u32 zero = 0;
+	long err;
+
+	/* look up before reading, to reduce error */
+	ptr = bpf_map_lookup_elem(&fentry_readings, &zero);
+	if (!ptr)
+		return 0;
+
+	err = bpf_perf_event_read_value(&events, key, &reading,
+					sizeof(reading));
+	if (err)
+		return 0;
+
+	*ptr = reading;
+	return 0;
+}
+
+static inline void
+fexit_update_maps(struct bpf_perf_event_value *after)
+{
+	struct bpf_perf_event_value *before, diff, *accum;
+	__u32 zero = 0;
+
+	before = bpf_map_lookup_elem(&fentry_readings, &zero);
+	/* only account samples with a valid fentry_reading */
+	if (before && before->counter) {
+		struct bpf_perf_event_value *accum;
+
+		diff.counter = after->counter - before->counter;
+		diff.enabled = after->enabled - before->enabled;
+		diff.running = after->running - before->running;
+
+		accum = bpf_map_lookup_elem(&accum_readings, &zero);
+		if (accum) {
+			accum->counter += diff.counter;
+			accum->enabled += diff.enabled;
+			accum->running += diff.running;
+		}
+	}
+}
+
+SEC("fexit/XXX")
+int BPF_PROG(fexit_XXX)
+{
+	struct bpf_perf_event_value reading;
+	__u32 cpu = bpf_get_smp_processor_id();
+	__u32 one = 1, zero = 0;
+	int err;
+
+	/* read all events before updating the maps, to reduce error */
+	err = bpf_perf_event_read_value(&events, cpu, &reading, sizeof(reading));
+	if (err)
+		return 0;
+
+	fexit_update_maps(&reading);
+	return 0;
+}
+
+char LICENSE[] SEC("license") = "Dual BSD/GPL";
diff --git a/tools/perf/util/evsel.c b/tools/perf/util/evsel.c
index 1cad6051d8b08..c6a50467076f8 100644
--- a/tools/perf/util/evsel.c
+++ b/tools/perf/util/evsel.c
@@ -25,6 +25,7 @@
 #include <stdlib.h>
 #include <perf/evsel.h>
 #include "asm/bug.h"
+#include "bpf_counter.h"
 #include "callchain.h"
 #include "cgroup.h"
 #include "counts.h"
@@ -51,6 +52,10 @@
 #include <internal/lib.h>
 
 #include <linux/ctype.h>
+#include <bpf/bpf.h>
+#include <bpf/libbpf.h>
+#include <bpf/btf.h>
+#include "rlimit.h"
 
 struct perf_missing_features perf_missing_features;
 
@@ -247,6 +252,7 @@ void evsel__init(struct evsel *evsel,
 	evsel->bpf_obj	   = NULL;
 	evsel->bpf_fd	   = -1;
 	INIT_LIST_HEAD(&evsel->config_terms);
+	INIT_LIST_HEAD(&evsel->bpf_counter_list);
 	perf_evsel__object.init(evsel);
 	evsel->sample_size = __evsel__sample_size(attr->sample_type);
 	evsel__calc_id_pos(evsel);
@@ -1365,6 +1371,7 @@ void evsel__exit(struct evsel *evsel)
 {
 	assert(list_empty(&evsel->core.node));
 	assert(evsel->evlist == NULL);
+	bpf_counter__destroy(evsel);
 	evsel__free_counts(evsel);
 	perf_evsel__free_fd(&evsel->core);
 	perf_evsel__free_id(&evsel->core);
@@ -1770,6 +1777,8 @@ static int evsel__open_cpu(struct evsel *evsel, struct perf_cpu_map *cpus,
 		evsel->core.attr.sample_id_all = 0;
 
 	display_attr(&evsel->core.attr);
+	if (!list_empty(&evsel->bpf_counter_list))
+		evsel->core.attr.inherit = 0;
 
 	for (cpu = start_cpu; cpu < end_cpu; cpu++) {
 
@@ -1788,6 +1797,8 @@ static int evsel__open_cpu(struct evsel *evsel, struct perf_cpu_map *cpus,
 
 			FD(evsel, cpu, thread) = fd;
 
+			bpf_counter__install_pe(evsel, cpu, fd);
+
 			if (unlikely(test_attr__enabled)) {
 				test_attr__open(&evsel->core.attr, pid, cpus->map[cpu],
 						fd, group_fd, flags);
diff --git a/tools/perf/util/evsel.h b/tools/perf/util/evsel.h
index 79a860d8e3eef..1731fba702bf4 100644
--- a/tools/perf/util/evsel.h
+++ b/tools/perf/util/evsel.h
@@ -10,6 +10,7 @@
 #include <internal/evsel.h>
 #include <perf/evsel.h>
 #include "symbol_conf.h"
+#include "bpf_counter.h"
 #include <internal/cpumap.h>
 
 struct bpf_object;
@@ -17,6 +18,8 @@ struct cgroup;
 struct perf_counts;
 struct perf_stat_evsel;
 union perf_event;
+struct bpf_counter_ops;
+struct target;
 
 typedef int (evsel__sb_cb_t)(union perf_event *event, void *data);
 
@@ -127,6 +130,8 @@ struct evsel {
 	 * See also evsel__has_callchain().
 	 */
 	__u64			synth_sample_type;
+	struct list_head	bpf_counter_list;
+	struct bpf_counter_ops	*bpf_counter_ops;
 };
 
 struct perf_missing_features {
@@ -423,4 +428,5 @@ static inline bool evsel__is_dummy_event(struct evsel *evsel)
 struct perf_env *evsel__env(struct evsel *evsel);
 
 int evsel__store_ids(struct evsel *evsel, struct evlist *evlist);
+
 #endif /* __PERF_EVSEL_H */
diff --git a/tools/perf/util/stat-display.c b/tools/perf/util/stat-display.c
index 4b57c0c076323..1536cfe0e50b4 100644
--- a/tools/perf/util/stat-display.c
+++ b/tools/perf/util/stat-display.c
@@ -1029,7 +1029,9 @@ static void print_header(struct perf_stat_config *config,
 	if (!config->csv_output) {
 		fprintf(output, "\n");
 		fprintf(output, " Performance counter stats for ");
-		if (_target->system_wide)
+		if (_target->bpf_str)
+			fprintf(output, "\'BPF program(s) %s", _target->bpf_str);
+		else if (_target->system_wide)
 			fprintf(output, "\'system wide");
 		else if (_target->cpu_list)
 			fprintf(output, "\'CPU(s) %s", _target->cpu_list);
diff --git a/tools/perf/util/target.c b/tools/perf/util/target.c
index a3db13dea937c..0f383418e3df5 100644
--- a/tools/perf/util/target.c
+++ b/tools/perf/util/target.c
@@ -56,6 +56,34 @@ enum target_errno target__validate(struct target *target)
 			ret = TARGET_ERRNO__UID_OVERRIDE_SYSTEM;
 	}
 
+	/* BPF and CPU are mutually exclusive */
+	if (target->bpf_str && target->cpu_list) {
+		target->cpu_list = NULL;
+		if (ret == TARGET_ERRNO__SUCCESS)
+			ret = TARGET_ERRNO__BPF_OVERRIDE_CPU;
+	}
+
+	/* BPF and PID/TID are mutually exclusive */
+	if (target->bpf_str && target->tid) {
+		target->tid = NULL;
+		if (ret == TARGET_ERRNO__SUCCESS)
+			ret = TARGET_ERRNO__BPF_OVERRIDE_PID;
+	}
+
+	/* BPF and UID are mutually exclusive */
+	if (target->bpf_str && target->uid_str) {
+		target->uid_str = NULL;
+		if (ret == TARGET_ERRNO__SUCCESS)
+			ret = TARGET_ERRNO__BPF_OVERRIDE_UID;
+	}
+
+	/* BPF and THREADS are mutually exclusive */
+	if (target->bpf_str && target->per_thread) {
+		target->per_thread = false;
+		if (ret == TARGET_ERRNO__SUCCESS)
+			ret = TARGET_ERRNO__BPF_OVERRIDE_THREAD;
+	}
+
 	/* THREAD and SYSTEM/CPU are mutually exclusive */
 	if (target->per_thread && (target->system_wide || target->cpu_list)) {
 		target->per_thread = false;
@@ -109,6 +137,10 @@ static const char *target__error_str[] = {
 	"PID/TID switch overriding SYSTEM",
 	"UID switch overriding SYSTEM",
 	"SYSTEM/CPU switch overriding PER-THREAD",
+	"BPF switch overriding CPU",
+	"BPF switch overriding PID/TID",
+	"BPF switch overriding UID",
+	"BPF switch overriding THREAD",
 	"Invalid User: %s",
 	"Problems obtaining information for user %s",
 };
@@ -134,7 +166,7 @@ int target__strerror(struct target *target, int errnum,
 
 	switch (errnum) {
 	case TARGET_ERRNO__PID_OVERRIDE_CPU ...
-	     TARGET_ERRNO__SYSTEM_OVERRIDE_THREAD:
+	     TARGET_ERRNO__BPF_OVERRIDE_THREAD:
 		snprintf(buf, buflen, "%s", msg);
 		break;
 
diff --git a/tools/perf/util/target.h b/tools/perf/util/target.h
index 6ef01a83b24e9..f132c6c2eef81 100644
--- a/tools/perf/util/target.h
+++ b/tools/perf/util/target.h
@@ -10,6 +10,7 @@ struct target {
 	const char   *tid;
 	const char   *cpu_list;
 	const char   *uid_str;
+	const char   *bpf_str;
 	uid_t	     uid;
 	bool	     system_wide;
 	bool	     uses_mmap;
@@ -36,6 +37,10 @@ enum target_errno {
 	TARGET_ERRNO__PID_OVERRIDE_SYSTEM,
 	TARGET_ERRNO__UID_OVERRIDE_SYSTEM,
 	TARGET_ERRNO__SYSTEM_OVERRIDE_THREAD,
+	TARGET_ERRNO__BPF_OVERRIDE_CPU,
+	TARGET_ERRNO__BPF_OVERRIDE_PID,
+	TARGET_ERRNO__BPF_OVERRIDE_UID,
+	TARGET_ERRNO__BPF_OVERRIDE_THREAD,
 
 	/* for target__parse_uid() */
 	TARGET_ERRNO__INVALID_UID,
@@ -59,6 +64,11 @@ static inline bool target__has_cpu(struct target *target)
 	return target->system_wide || target->cpu_list;
 }
 
+static inline bool target__has_bpf(struct target *target)
+{
+	return target->bpf_str;
+}
+
 static inline bool target__none(struct target *target)
 {
 	return !target__has_task(target) && !target__has_cpu(target);
-- 
2.24.1