Merge tag 'perf-tools-for-v5.13-2021-04-29' of git://git.kernel.org/pub/scm/linux/kernel/git/acme/linux

Pull perf tool updates from Arnaldo Carvalho de Melo:
 "perf stat:

   - Add support for hybrid PMUs to support systems such as Intel
     Alderlake and its BIG/little core/atom cpus.

   - Introduce 'bperf' to share hardware PMCs with BPF.

   - New --iostat option to collect and present IO stats on Intel
     hardware.

     This functionality is based on recently introduced sysfs attributes
     for Intel® Xeon® Scalable processor family (code name Skylake-SP)
     in commit bb42b3d39781 ("perf/x86/intel/uncore: Expose an Uncore
     unit to IIO PMON mapping")

     It is intended to provide four I/O performance metrics in MB per
     each PCIe root port:

       - Inbound Read: I/O devices below root port read from the host memory
       - Inbound Write: I/O devices below root port write to the host memory
       - Outbound Read: CPU reads from I/O devices below root port
       - Outbound Write: CPU writes to I/O devices below root port

   - Align CSV output for summary.

   - Clarify --null use cases: Assess raw overhead of 'perf stat' or
     measure just wall clock time.

   - Improve readability of shadow stats.

  perf record:

   - Change the COMM when starting tha workload so that --exclude-perf
     doesn't seem to be not honoured.

   - Improve 'Workload failed' message printing events + what was
     exec'ed.

   - Fix cross-arch support for TIME_CONV.

  perf report:

   - Add option to disable raw event ordering.

   - Dump the contents of PERF_RECORD_TIME_CONV in 'perf report -D'.

   - Improvements to --stat output, that shows information about
     PERF_RECORD_ events.

   - Preserve identifier id in OCaml demangler.

  perf annotate:

   - Show full source location with 'l' hotkey in the 'perf annotate'
     TUI.

   - Add line number like in TUI and source location at EOL to the 'perf
     annotate' --stdio mode.

   - Add --demangle and --demangle-kernel to 'perf annotate'.

   - Allow configuring annotate.demangle{,_kernel} in 'perf config'.

   - Fix sample events lost in stdio mode.

  perf data:

   - Allow converting a perf.data file to JSON.

  libperf:

   - Add support for user space counter access.

   - Update topdown documentation to permit rdpmc calls.

  perf test:

   - Add 'perf test' for 'perf stat' CSV output.

   - Add 'perf test' entries to test the hybrid PMU support.

   - Cleanup 'perf test daemon' if its 'perf test' is interrupted.

   - Handle metric reuse in pmu-events parsing 'perf test' entry.

   - Add test for PE executable support.

   - Add timeout for wait for daemon start in its 'perf test' entries.

  Build:

   - Enable libtraceevent dynamic linking.

   - Improve feature detection output.

   - Fix caching of feature checks caching.

   - First round of updates for tools copies of kernel headers.

   - Enable warnings when compiling BPF programs.

  Vendor specific events:

   - Intel:
      - Add missing skylake & icelake model numbers.

   - arm64:
      - Add Hisi hip08 L1, L2 and L3 metrics.
      - Add Fujitsu A64FX PMU events.

   - PowerPC:
      - Initial JSON/events list for power10 platform.
      - Remove unsupported power9 metrics.

   - AMD:
      - Add Zen3 events.
      - Fix broken L2 Cache Hits from L2 HWPF metric.
      - Use lowercases for all the eventcodes and umasks.

  Hardware tracing:

   - arm64:
      - Update CoreSight ETM metadata format.
      - Fix bitmap for CS-ETM option.
      - Support PID tracing in config.
      - Detect pid in VMID for kernel running at EL2.

  Arch specific updates:

   - MIPS:
      - Support MIPS unwinding and dwarf-regs.
      - Generate mips syscalls_n64.c syscall table.

   - PowerPC:
      - Add support for PERF_SAMPLE_WEIGH_STRUCT on PowerPC.
      - Support pipeline stage cycles for powerpc.

  libbeauty:

   - Fix fsconfig generator"

* tag 'perf-tools-for-v5.13-2021-04-29' of git://git.kernel.org/pub/scm/linux/kernel/git/acme/linux: (132 commits)
  perf build: Defer printing detected features to the end of all feature checks
  tools build: Allow deferring printing the results of feature detection
  perf build: Regenerate the FEATURE_DUMP file after extra feature checks
  perf session: Dump PERF_RECORD_TIME_CONV event
  perf session: Add swap operation for event TIME_CONV
  perf jit: Let convert_timestamp() to be backwards-compatible
  perf tools: Change fields type in perf_record_time_conv
  perf tools: Enable libtraceevent dynamic linking
  perf Documentation: Document intel-hybrid support
  perf tests: Skip 'perf stat metrics (shadow stat) test' for hybrid
  perf tests: Support 'Convert perf time to TSC' test for hybrid
  perf tests: Support 'Session topology' test for hybrid
  perf tests: Support 'Parse and process metrics' test for hybrid
  perf tests: Support 'Track with sched_switch' test for hybrid
  perf tests: Skip 'Setup struct perf_event_attr' test for hybrid
  perf tests: Add hybrid cases for 'Roundtrip evsel->name' test
  perf tests: Add hybrid cases for 'Parse event definition strings' test
  perf record: Uniquify hybrid event name
  perf stat: Warn group events from different hybrid PMU
  perf stat: Filter out unmatched aggregation for hybrid event
  ...

1 files changed, 539 insertions, 5 deletions

diff --git a/tools/perf/util/bpf_counter.c b/tools/perf/util/bpf_counter.c
index 04f89120b323..ddb52f748c8e 100644
--- a/tools/perf/util/bpf_counter.c
+++ b/tools/perf/util/bpf_counter.c
@@ -5,6 +5,7 @@
 #include <assert.h>
 #include <limits.h>
 #include <unistd.h>
+#include <sys/file.h>
 #include <sys/time.h>
 #include <sys/resource.h>
 #include <linux/err.h>
@@ -12,14 +13,24 @@
 #include <bpf/bpf.h>
 #include <bpf/btf.h>
 #include <bpf/libbpf.h>
+#include <api/fs/fs.h>
+#include <perf/bpf_perf.h>
 
 #include "bpf_counter.h"
 #include "counts.h"
 #include "debug.h"
 #include "evsel.h"
+#include "evlist.h"
 #include "target.h"
+#include "cpumap.h"
+#include "thread_map.h"
 
 #include "bpf_skel/bpf_prog_profiler.skel.h"
+#include "bpf_skel/bperf_u.h"
+#include "bpf_skel/bperf_leader.skel.h"
+#include "bpf_skel/bperf_follower.skel.h"
+
+#define ATTR_MAP_SIZE 16
 
 static inline void *u64_to_ptr(__u64 ptr)
 {
@@ -204,6 +215,17 @@ static int bpf_program_profiler__enable(struct evsel *evsel)
 	return 0;
 }
 
+static int bpf_program_profiler__disable(struct evsel *evsel)
+{
+	struct bpf_counter *counter;
+
+	list_for_each_entry(counter, &evsel->bpf_counter_list, list) {
+		assert(counter->skel != NULL);
+		bpf_prog_profiler_bpf__detach(counter->skel);
+	}
+	return 0;
+}
+
 static int bpf_program_profiler__read(struct evsel *evsel)
 {
 	// perf_cpu_map uses /sys/devices/system/cpu/online
@@ -269,22 +291,527 @@ static int bpf_program_profiler__install_pe(struct evsel *evsel, int cpu,
 struct bpf_counter_ops bpf_program_profiler_ops = {
 	.load       = bpf_program_profiler__load,
 	.enable	    = bpf_program_profiler__enable,
+	.disable    = bpf_program_profiler__disable,
 	.read       = bpf_program_profiler__read,
 	.destroy    = bpf_program_profiler__destroy,
 	.install_pe = bpf_program_profiler__install_pe,
 };
 
+static __u32 bpf_link_get_id(int fd)
+{
+	struct bpf_link_info link_info = {0};
+	__u32 link_info_len = sizeof(link_info);
+
+	bpf_obj_get_info_by_fd(fd, &link_info, &link_info_len);
+	return link_info.id;
+}
+
+static __u32 bpf_link_get_prog_id(int fd)
+{
+	struct bpf_link_info link_info = {0};
+	__u32 link_info_len = sizeof(link_info);
+
+	bpf_obj_get_info_by_fd(fd, &link_info, &link_info_len);
+	return link_info.prog_id;
+}
+
+static __u32 bpf_map_get_id(int fd)
+{
+	struct bpf_map_info map_info = {0};
+	__u32 map_info_len = sizeof(map_info);
+
+	bpf_obj_get_info_by_fd(fd, &map_info, &map_info_len);
+	return map_info.id;
+}
+
+static bool bperf_attr_map_compatible(int attr_map_fd)
+{
+	struct bpf_map_info map_info = {0};
+	__u32 map_info_len = sizeof(map_info);
+	int err;
+
+	err = bpf_obj_get_info_by_fd(attr_map_fd, &map_info, &map_info_len);
+
+	if (err)
+		return false;
+	return (map_info.key_size == sizeof(struct perf_event_attr)) &&
+		(map_info.value_size == sizeof(struct perf_event_attr_map_entry));
+}
+
+static int bperf_lock_attr_map(struct target *target)
+{
+	char path[PATH_MAX];
+	int map_fd, err;
+
+	if (target->attr_map) {
+		scnprintf(path, PATH_MAX, "%s", target->attr_map);
+	} else {
+		scnprintf(path, PATH_MAX, "%s/fs/bpf/%s", sysfs__mountpoint(),
+			  BPF_PERF_DEFAULT_ATTR_MAP_PATH);
+	}
+
+	if (access(path, F_OK)) {
+		map_fd = bpf_create_map(BPF_MAP_TYPE_HASH,
+					sizeof(struct perf_event_attr),
+					sizeof(struct perf_event_attr_map_entry),
+					ATTR_MAP_SIZE, 0);
+		if (map_fd < 0)
+			return -1;
+
+		err = bpf_obj_pin(map_fd, path);
+		if (err) {
+			/* someone pinned the map in parallel? */
+			close(map_fd);
+			map_fd = bpf_obj_get(path);
+			if (map_fd < 0)
+				return -1;
+		}
+	} else {
+		map_fd = bpf_obj_get(path);
+		if (map_fd < 0)
+			return -1;
+	}
+
+	if (!bperf_attr_map_compatible(map_fd)) {
+		close(map_fd);
+		return -1;
+
+	}
+	err = flock(map_fd, LOCK_EX);
+	if (err) {
+		close(map_fd);
+		return -1;
+	}
+	return map_fd;
+}
+
+/* trigger the leader program on a cpu */
+static int bperf_trigger_reading(int prog_fd, int cpu)
+{
+	DECLARE_LIBBPF_OPTS(bpf_test_run_opts, opts,
+			    .ctx_in = NULL,
+			    .ctx_size_in = 0,
+			    .flags = BPF_F_TEST_RUN_ON_CPU,
+			    .cpu = cpu,
+			    .retval = 0,
+		);
+
+	return bpf_prog_test_run_opts(prog_fd, &opts);
+}
+
+static int bperf_check_target(struct evsel *evsel,
+			      struct target *target,
+			      enum bperf_filter_type *filter_type,
+			      __u32 *filter_entry_cnt)
+{
+	if (evsel->leader->core.nr_members > 1) {
+		pr_err("bpf managed perf events do not yet support groups.\n");
+		return -1;
+	}
+
+	/* determine filter type based on target */
+	if (target->system_wide) {
+		*filter_type = BPERF_FILTER_GLOBAL;
+		*filter_entry_cnt = 1;
+	} else if (target->cpu_list) {
+		*filter_type = BPERF_FILTER_CPU;
+		*filter_entry_cnt = perf_cpu_map__nr(evsel__cpus(evsel));
+	} else if (target->tid) {
+		*filter_type = BPERF_FILTER_PID;
+		*filter_entry_cnt = perf_thread_map__nr(evsel->core.threads);
+	} else if (target->pid || evsel->evlist->workload.pid != -1) {
+		*filter_type = BPERF_FILTER_TGID;
+		*filter_entry_cnt = perf_thread_map__nr(evsel->core.threads);
+	} else {
+		pr_err("bpf managed perf events do not yet support these targets.\n");
+		return -1;
+	}
+
+	return 0;
+}
+
+static	struct perf_cpu_map *all_cpu_map;
+
+static int bperf_reload_leader_program(struct evsel *evsel, int attr_map_fd,
+				       struct perf_event_attr_map_entry *entry)
+{
+	struct bperf_leader_bpf *skel = bperf_leader_bpf__open();
+	int link_fd, diff_map_fd, err;
+	struct bpf_link *link = NULL;
+
+	if (!skel) {
+		pr_err("Failed to open leader skeleton\n");
+		return -1;
+	}
+
+	bpf_map__resize(skel->maps.events, libbpf_num_possible_cpus());
+	err = bperf_leader_bpf__load(skel);
+	if (err) {
+		pr_err("Failed to load leader skeleton\n");
+		goto out;
+	}
+
+	err = -1;
+	link = bpf_program__attach(skel->progs.on_switch);
+	if (!link) {
+		pr_err("Failed to attach leader program\n");
+		goto out;
+	}
+
+	link_fd = bpf_link__fd(link);
+	diff_map_fd = bpf_map__fd(skel->maps.diff_readings);
+	entry->link_id = bpf_link_get_id(link_fd);
+	entry->diff_map_id = bpf_map_get_id(diff_map_fd);
+	err = bpf_map_update_elem(attr_map_fd, &evsel->core.attr, entry, BPF_ANY);
+	assert(err == 0);
+
+	evsel->bperf_leader_link_fd = bpf_link_get_fd_by_id(entry->link_id);
+	assert(evsel->bperf_leader_link_fd >= 0);
+
+	/*
+	 * save leader_skel for install_pe, which is called within
+	 * following evsel__open_per_cpu call
+	 */
+	evsel->leader_skel = skel;
+	evsel__open_per_cpu(evsel, all_cpu_map, -1);
+
+out:
+	bperf_leader_bpf__destroy(skel);
+	bpf_link__destroy(link);
+	return err;
+}
+
+static int bperf__load(struct evsel *evsel, struct target *target)
+{
+	struct perf_event_attr_map_entry entry = {0xffffffff, 0xffffffff};
+	int attr_map_fd, diff_map_fd = -1, err;
+	enum bperf_filter_type filter_type;
+	__u32 filter_entry_cnt, i;
+
+	if (bperf_check_target(evsel, target, &filter_type, &filter_entry_cnt))
+		return -1;
+
+	if (!all_cpu_map) {
+		all_cpu_map = perf_cpu_map__new(NULL);
+		if (!all_cpu_map)
+			return -1;
+	}
+
+	evsel->bperf_leader_prog_fd = -1;
+	evsel->bperf_leader_link_fd = -1;
+
+	/*
+	 * Step 1: hold a fd on the leader program and the bpf_link, if
+	 * the program is not already gone, reload the program.
+	 * Use flock() to ensure exclusive access to the perf_event_attr
+	 * map.
+	 */
+	attr_map_fd = bperf_lock_attr_map(target);
+	if (attr_map_fd < 0) {
+		pr_err("Failed to lock perf_event_attr map\n");
+		return -1;
+	}
+
+	err = bpf_map_lookup_elem(attr_map_fd, &evsel->core.attr, &entry);
+	if (err) {
+		err = bpf_map_update_elem(attr_map_fd, &evsel->core.attr, &entry, BPF_ANY);
+		if (err)
+			goto out;
+	}
+
+	evsel->bperf_leader_link_fd = bpf_link_get_fd_by_id(entry.link_id);
+	if (evsel->bperf_leader_link_fd < 0 &&
+	    bperf_reload_leader_program(evsel, attr_map_fd, &entry))
+		goto out;
+
+	/*
+	 * The bpf_link holds reference to the leader program, and the
+	 * leader program holds reference to the maps. Therefore, if
+	 * link_id is valid, diff_map_id should also be valid.
+	 */
+	evsel->bperf_leader_prog_fd = bpf_prog_get_fd_by_id(
+		bpf_link_get_prog_id(evsel->bperf_leader_link_fd));
+	assert(evsel->bperf_leader_prog_fd >= 0);
+
+	diff_map_fd = bpf_map_get_fd_by_id(entry.diff_map_id);
+	assert(diff_map_fd >= 0);
+
+	/*
+	 * bperf uses BPF_PROG_TEST_RUN to get accurate reading. Check
+	 * whether the kernel support it
+	 */
+	err = bperf_trigger_reading(evsel->bperf_leader_prog_fd, 0);
+	if (err) {
+		pr_err("The kernel does not support test_run for raw_tp BPF programs.\n"
+		       "Therefore, --use-bpf might show inaccurate readings\n");
+		goto out;
+	}
+
+	/* Step 2: load the follower skeleton */
+	evsel->follower_skel = bperf_follower_bpf__open();
+	if (!evsel->follower_skel) {
+		pr_err("Failed to open follower skeleton\n");
+		goto out;
+	}
+
+	/* attach fexit program to the leader program */
+	bpf_program__set_attach_target(evsel->follower_skel->progs.fexit_XXX,
+				       evsel->bperf_leader_prog_fd, "on_switch");
+
+	/* connect to leader diff_reading map */
+	bpf_map__reuse_fd(evsel->follower_skel->maps.diff_readings, diff_map_fd);
+
+	/* set up reading map */
+	bpf_map__set_max_entries(evsel->follower_skel->maps.accum_readings,
+				 filter_entry_cnt);
+	/* set up follower filter based on target */
+	bpf_map__set_max_entries(evsel->follower_skel->maps.filter,
+				 filter_entry_cnt);
+	err = bperf_follower_bpf__load(evsel->follower_skel);
+	if (err) {
+		pr_err("Failed to load follower skeleton\n");
+		bperf_follower_bpf__destroy(evsel->follower_skel);
+		evsel->follower_skel = NULL;
+		goto out;
+	}
+
+	for (i = 0; i < filter_entry_cnt; i++) {
+		int filter_map_fd;
+		__u32 key;
+
+		if (filter_type == BPERF_FILTER_PID ||
+		    filter_type == BPERF_FILTER_TGID)
+			key = evsel->core.threads->map[i].pid;
+		else if (filter_type == BPERF_FILTER_CPU)
+			key = evsel->core.cpus->map[i];
+		else
+			break;
+
+		filter_map_fd = bpf_map__fd(evsel->follower_skel->maps.filter);
+		bpf_map_update_elem(filter_map_fd, &key, &i, BPF_ANY);
+	}
+
+	evsel->follower_skel->bss->type = filter_type;
+
+	err = bperf_follower_bpf__attach(evsel->follower_skel);
+
+out:
+	if (err && evsel->bperf_leader_link_fd >= 0)
+		close(evsel->bperf_leader_link_fd);
+	if (err && evsel->bperf_leader_prog_fd >= 0)
+		close(evsel->bperf_leader_prog_fd);
+	if (diff_map_fd >= 0)
+		close(diff_map_fd);
+
+	flock(attr_map_fd, LOCK_UN);
+	close(attr_map_fd);
+
+	return err;
+}
+
+static int bperf__install_pe(struct evsel *evsel, int cpu, int fd)
+{
+	struct bperf_leader_bpf *skel = evsel->leader_skel;
+
+	return bpf_map_update_elem(bpf_map__fd(skel->maps.events),
+				   &cpu, &fd, BPF_ANY);
+}
+
+/*
+ * trigger the leader prog on each cpu, so the accum_reading map could get
+ * the latest readings.
+ */
+static int bperf_sync_counters(struct evsel *evsel)
+{
+	int num_cpu, i, cpu;
+
+	num_cpu = all_cpu_map->nr;
+	for (i = 0; i < num_cpu; i++) {
+		cpu = all_cpu_map->map[i];
+		bperf_trigger_reading(evsel->bperf_leader_prog_fd, cpu);
+	}
+	return 0;
+}
+
+static int bperf__enable(struct evsel *evsel)
+{
+	evsel->follower_skel->bss->enabled = 1;
+	return 0;
+}
+
+static int bperf__disable(struct evsel *evsel)
+{
+	evsel->follower_skel->bss->enabled = 0;
+	return 0;
+}
+
+static int bperf__read(struct evsel *evsel)
+{
+	struct bperf_follower_bpf *skel = evsel->follower_skel;
+	__u32 num_cpu_bpf = cpu__max_cpu();
+	struct bpf_perf_event_value values[num_cpu_bpf];
+	int reading_map_fd, err = 0;
+	__u32 i, j, num_cpu;
+
+	bperf_sync_counters(evsel);
+	reading_map_fd = bpf_map__fd(skel->maps.accum_readings);
+
+	for (i = 0; i < bpf_map__max_entries(skel->maps.accum_readings); i++) {
+		__u32 cpu;
+
+		err = bpf_map_lookup_elem(reading_map_fd, &i, values);
+		if (err)
+			goto out;
+		switch (evsel->follower_skel->bss->type) {
+		case BPERF_FILTER_GLOBAL:
+			assert(i == 0);
+
+			num_cpu = all_cpu_map->nr;
+			for (j = 0; j < num_cpu; j++) {
+				cpu = all_cpu_map->map[j];
+				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;
+			}
+			break;
+		case BPERF_FILTER_CPU:
+			cpu = evsel->core.cpus->map[i];
+			perf_counts(evsel->counts, i, 0)->val = values[cpu].counter;
+			perf_counts(evsel->counts, i, 0)->ena = values[cpu].enabled;
+			perf_counts(evsel->counts, i, 0)->run = values[cpu].running;
+			break;
+		case BPERF_FILTER_PID:
+		case BPERF_FILTER_TGID:
+			perf_counts(evsel->counts, 0, i)->val = 0;
+			perf_counts(evsel->counts, 0, i)->ena = 0;
+			perf_counts(evsel->counts, 0, i)->run = 0;
+
+			for (cpu = 0; cpu < num_cpu_bpf; cpu++) {
+				perf_counts(evsel->counts, 0, i)->val += values[cpu].counter;
+				perf_counts(evsel->counts, 0, i)->ena += values[cpu].enabled;
+				perf_counts(evsel->counts, 0, i)->run += values[cpu].running;
+			}
+			break;
+		default:
+			break;
+		}
+	}
+out:
+	return err;
+}
+
+static int bperf__destroy(struct evsel *evsel)
+{
+	bperf_follower_bpf__destroy(evsel->follower_skel);
+	close(evsel->bperf_leader_prog_fd);
+	close(evsel->bperf_leader_link_fd);
+	return 0;
+}
+
+/*
+ * bperf: share hardware PMCs with BPF
+ *
+ * perf uses performance monitoring counters (PMC) 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.
+ *
+ *                                ||
+ *       shared progs and maps <- || -> per session progs and maps
+ *                                ||
+ *   ---------------              ||
+ *   | perf_events |              ||
+ *   ---------------       fexit  ||      -----------------
+ *          |             --------||----> | follower prog |
+ *       --------------- /        || ---  -----------------
+ * cs -> | leader prog |/         ||/        |         |
+ *   --> ---------------         /||  --------------  ------------------
+ *  /       |         |         / ||  | filter map |  | accum_readings |
+ * /  ------------  ------------  ||  --------------  ------------------
+ * |  | prev map |  | diff map |  ||                        |
+ * |  ------------  ------------  ||                        |
+ *  \                             ||                        |
+ * = \ ==================================================== | ============
+ *    \                                                    /   user space
+ *     \                                                  /
+ *      \                                                /
+ *    BPF_PROG_TEST_RUN                    BPF_MAP_LOOKUP_ELEM
+ *        \                                            /
+ *         \                                          /
+ *          \------  perf-stat ----------------------/
+ *
+ * The figure above shows the architecture of bperf. Note that the figure
+ * is divided into 3 regions: shared progs and maps (top left), per session
+ * progs and maps (top right), and user space (bottom).
+ *
+ * The leader prog is triggered on each context switch (cs). The leader
+ * prog reads perf_events and stores the difference (current_reading -
+ * previous_reading) to the diff map. For the same metric, e.g. "cycles",
+ * multiple perf-stat sessions share the same leader prog.
+ *
+ * Each perf-stat session creates a follower prog as fexit program to the
+ * leader prog. It is possible to attach up to BPF_MAX_TRAMP_PROGS (38)
+ * follower progs to the same leader prog. The follower prog checks current
+ * task and processor ID to decide whether to add the value from the diff
+ * map to its accumulated reading map (accum_readings).
+ *
+ * Finally, perf-stat user space reads the value from accum_reading map.
+ *
+ * Besides context switch, it is also necessary to trigger the leader prog
+ * before perf-stat reads the value. Otherwise, the accum_reading map may
+ * not have the latest reading from the perf_events. This is achieved by
+ * triggering the event via sys_bpf(BPF_PROG_TEST_RUN) to each CPU.
+ *
+ * Comment before the definition of struct perf_event_attr_map_entry
+ * describes how different sessions of perf-stat share information about
+ * the leader prog.
+ */
+
+struct bpf_counter_ops bperf_ops = {
+	.load       = bperf__load,
+	.enable     = bperf__enable,
+	.disable    = bperf__disable,
+	.read       = bperf__read,
+	.install_pe = bperf__install_pe,
+	.destroy    = bperf__destroy,
+};
+
+static inline bool bpf_counter_skip(struct evsel *evsel)
+{
+	return list_empty(&evsel->bpf_counter_list) &&
+		evsel->follower_skel == NULL;
+}
+
 int bpf_counter__install_pe(struct evsel *evsel, int cpu, int fd)
 {
-	if (list_empty(&evsel->bpf_counter_list))
+	if (bpf_counter_skip(evsel))
 		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))
+	if (target->bpf_str)
 		evsel->bpf_counter_ops = &bpf_program_profiler_ops;
+	else if (target->use_bpf || evsel->bpf_counter ||
+		 evsel__match_bpf_counter_events(evsel->name))
+		evsel->bpf_counter_ops = &bperf_ops;
 
 	if (evsel->bpf_counter_ops)
 		return evsel->bpf_counter_ops->load(evsel, target);
@@ -293,21 +820,28 @@ int bpf_counter__load(struct evsel *evsel, struct target *target)
 
 int bpf_counter__enable(struct evsel *evsel)
 {
-	if (list_empty(&evsel->bpf_counter_list))
+	if (bpf_counter_skip(evsel))
 		return 0;
 	return evsel->bpf_counter_ops->enable(evsel);
 }
 
+int bpf_counter__disable(struct evsel *evsel)
+{
+	if (bpf_counter_skip(evsel))
+		return 0;
+	return evsel->bpf_counter_ops->disable(evsel);
+}
+
 int bpf_counter__read(struct evsel *evsel)
 {
-	if (list_empty(&evsel->bpf_counter_list))
+	if (bpf_counter_skip(evsel))
 		return -EAGAIN;
 	return evsel->bpf_counter_ops->read(evsel);
 }
 
 void bpf_counter__destroy(struct evsel *evsel)
 {
-	if (list_empty(&evsel->bpf_counter_list))
+	if (bpf_counter_skip(evsel))
 		return;
 	evsel->bpf_counter_ops->destroy(evsel);
 	evsel->bpf_counter_ops = NULL;