perf stat: Add/fix bperf cgroup max events workarounds

Commit b830851 bumped the max events to 1024 but this results in
BPF verifier issues if the number of command line events is too
large. Workaround this by:

1) moving the constants to a header file to share between BPF and perf
   C code,
2) testing that the maximum number of events doesn't cause BPF
   verifier issues in debug builds,
3) lower the max events from 1024 to 128,
4) in perf stat, if there are more events than the BPF counters can
   support then disable BPF counter usage.

The rodata setup is factored into its own function to avoid
duplicating it in the testing code.

Signed-off-by: Ian Rogers <irogers@google.com>
Fixes: b830851 ("perf stat bperf cgroup: Increase MAX_EVENTS from 32 to 1024")
Signed-off-by: Namhyung Kim <namhyung@kernel.org>

Author: captain5050

tools/perf/builtin-stat.c

@@ -96,6 +96,10 @@
 #include <perf/evlist.h>
 #include <internal/threadmap.h>
 
+#ifdef HAVE_BPF_SKEL
+#include "util/bpf_skel/bperf_cgroup.h"
+#endif
+
 #define DEFAULT_SEPARATOR	" "
 #define FREEZE_ON_SMI_PATH	"bus/event_source/devices/cpu/freeze_on_smi"
 
@@ -2852,7 +2856,14 @@ int cmd_stat(int argc, const char **argv)
 			goto out;
 		}
 	}
-
+#ifdef HAVE_BPF_SKEL
+	if (target.use_bpf && nr_cgroups &&
+	    (evsel_list->core.nr_entries / nr_cgroups) > BPERF_CGROUP__MAX_EVENTS) {
+		pr_warning("Disabling BPF counters due to more events (%d) than the max (%d)\n",
+			   evsel_list->core.nr_entries / nr_cgroups, BPERF_CGROUP__MAX_EVENTS);
+		target.use_bpf = false;
+	}
+#endif // HAVE_BPF_SKEL
 	evlist__warn_user_requested_cpus(evsel_list, target.cpu_list);
 
 	evlist__for_each_entry(evsel_list, counter) {

tools/perf/util/bpf_counter_cgroup.c

@@ -27,6 +27,7 @@
 #include "cpumap.h"
 #include "thread_map.h"
 
+#include "bpf_skel/bperf_cgroup.h"
 #include "bpf_skel/bperf_cgroup.skel.h"
 
 static struct perf_event_attr cgrp_switch_attr = {
@@ -42,6 +43,55 @@ static struct bperf_cgroup_bpf *skel;
 
 #define FD(evt, cpu) (*(int *)xyarray__entry(evt->core.fd, cpu, 0))
 
+static void setup_rodata(struct bperf_cgroup_bpf *sk, int evlist_size)
+{
+	int map_size, total_cpus = cpu__max_cpu().cpu;
+
+	sk->rodata->num_cpus = total_cpus;
+	sk->rodata->num_events = evlist_size / nr_cgroups;
+
+	if (cgroup_is_v2("perf_event") > 0)
+		sk->rodata->use_cgroup_v2 = 1;
+
+	BUG_ON(evlist_size % nr_cgroups != 0);
+
+	/* we need one copy of events per cpu for reading */
+	map_size = total_cpus * evlist_size / nr_cgroups;
+	bpf_map__set_max_entries(sk->maps.events, map_size);
+	bpf_map__set_max_entries(sk->maps.cgrp_idx, nr_cgroups);
+	/* previous result is saved in a per-cpu array */
+	map_size = evlist_size / nr_cgroups;
+	bpf_map__set_max_entries(sk->maps.prev_readings, map_size);
+	/* cgroup result needs all events (per-cpu) */
+	map_size = evlist_size;
+	bpf_map__set_max_entries(sk->maps.cgrp_readings, map_size);
+}
+
+static void test_max_events_program_load(void)
+{
+#ifndef NDEBUG
+	/*
+	 * Test that the program verifies with the maximum number of events. If
+	 * this test fails unfortunately perf needs recompiling with a lower
+	 * BPERF_CGROUP__MAX_EVENTS to avoid BPF verifier issues.
+	 */
+	int err, max_events = BPERF_CGROUP__MAX_EVENTS * nr_cgroups;
+	struct bperf_cgroup_bpf *test_skel = bperf_cgroup_bpf__open();
+
+	if (!test_skel) {
+		pr_err("Failed to open cgroup skeleton\n");
+		return;
+	}
+	setup_rodata(test_skel, max_events);
+	err = bperf_cgroup_bpf__load(test_skel);
+	if (err) {
+		pr_err("Failed to load cgroup skeleton with max events %d.\n",
+			BPERF_CGROUP__MAX_EVENTS);
+	}
+	bperf_cgroup_bpf__destroy(test_skel);
+#endif
+}
+
 static int bperf_load_program(struct evlist *evlist)
 {
 	struct bpf_link *link;
@@ -50,35 +100,18 @@ static int bperf_load_program(struct evlist *evlist)
 	int i, j;
 	struct perf_cpu cpu;
 	int total_cpus = cpu__max_cpu().cpu;
-	int map_size, map_fd;
-	int prog_fd, err;
+	int map_fd, prog_fd, err;
+
+	set_max_rlimit();
+
+	test_max_events_program_load();
 
 	skel = bperf_cgroup_bpf__open();
 	if (!skel) {
 		pr_err("Failed to open cgroup skeleton\n");
 		return -1;
 	}
-
-	skel->rodata->num_cpus = total_cpus;
-	skel->rodata->num_events = evlist->core.nr_entries / nr_cgroups;
-
-	if (cgroup_is_v2("perf_event") > 0)
-		skel->rodata->use_cgroup_v2 = 1;
-
-	BUG_ON(evlist->core.nr_entries % nr_cgroups != 0);
-
-	/* we need one copy of events per cpu for reading */
-	map_size = total_cpus * evlist->core.nr_entries / nr_cgroups;
-	bpf_map__set_max_entries(skel->maps.events, map_size);
-	bpf_map__set_max_entries(skel->maps.cgrp_idx, nr_cgroups);
-	/* previous result is saved in a per-cpu array */
-	map_size = evlist->core.nr_entries / nr_cgroups;
-	bpf_map__set_max_entries(skel->maps.prev_readings, map_size);
-	/* cgroup result needs all events (per-cpu) */
-	map_size = evlist->core.nr_entries;
-	bpf_map__set_max_entries(skel->maps.cgrp_readings, map_size);
-
-	set_max_rlimit();
+	setup_rodata(skel, evlist->core.nr_entries);
 
 	err = bperf_cgroup_bpf__load(skel);
 	if (err) {

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

@@ -1,14 +1,12 @@
 // SPDX-License-Identifier: (GPL-2.0-only OR BSD-2-Clause)
 // Copyright (c) 2021 Facebook
 // Copyright (c) 2021 Google
+#include "bperf_cgroup.h"
 #include "vmlinux.h"
 #include <bpf/bpf_helpers.h>
 #include <bpf/bpf_tracing.h>
 #include <bpf/bpf_core_read.h>
 
-#define MAX_LEVELS  10  // max cgroup hierarchy level: arbitrary
-#define MAX_EVENTS  1024  // max events per cgroup: arbitrary
-
 // NOTE: many of map and global data will be modified before loading
 //       from the userspace (perf tool) using the skeleton helpers.
 
@@ -97,7 +95,7 @@ static inline int get_cgroup_v1_idx(__u32 *cgrps, int size)
 	cgrp = BPF_CORE_READ(p, cgroups, subsys[perf_subsys_id], cgroup);
 	level = BPF_CORE_READ(cgrp, level);
 
-	for (cnt = 0; i < MAX_LEVELS; i++) {
+	for (cnt = 0; i < BPERF_CGROUP__MAX_LEVELS; i++) {
 		__u64 cgrp_id;
 
 		if (i > level)
@@ -123,7 +121,7 @@ static inline int get_cgroup_v2_idx(__u32 *cgrps, int size)
 	__u32 *elem;
 	int cnt;
 
-	for (cnt = 0; i < MAX_LEVELS; i++) {
+	for (cnt = 0; i < BPERF_CGROUP__MAX_LEVELS; i++) {
 		__u64 cgrp_id = bpf_get_current_ancestor_cgroup_id(i);
 
 		if (cgrp_id == 0)
@@ -148,17 +146,17 @@ static int bperf_cgroup_count(void)
 	register int c = 0;
 	struct bpf_perf_event_value val, delta, *prev_val, *cgrp_val;
 	__u32 cpu = bpf_get_smp_processor_id();
-	__u32 cgrp_idx[MAX_LEVELS];
+	__u32 cgrp_idx[BPERF_CGROUP__MAX_LEVELS];
 	int cgrp_cnt;
 	__u32 key, cgrp;
 	long err;
 
 	if (use_cgroup_v2)
-		cgrp_cnt = get_cgroup_v2_idx(cgrp_idx, MAX_LEVELS);
+		cgrp_cnt = get_cgroup_v2_idx(cgrp_idx, BPERF_CGROUP__MAX_LEVELS);
 	else
-		cgrp_cnt = get_cgroup_v1_idx(cgrp_idx, MAX_LEVELS);
+		cgrp_cnt = get_cgroup_v1_idx(cgrp_idx, BPERF_CGROUP__MAX_LEVELS);
 
-	for ( ; idx < MAX_EVENTS; idx++) {
+	for ( ; idx < BPERF_CGROUP__MAX_EVENTS; idx++) {
 		if (idx == num_events)
 			break;
 
@@ -186,7 +184,7 @@ static int bperf_cgroup_count(void)
 			delta.enabled = val.enabled - prev_val->enabled;
 			delta.running = val.running - prev_val->running;
 
-			for (c = 0; c < MAX_LEVELS; c++) {
+			for (c = 0; c < BPERF_CGROUP__MAX_LEVELS; c++) {
 				if (c == cgrp_cnt)
 					break;
 

tools/perf/util/bpf_skel/bperf_cgroup.h

@@ -0,0 +1,15 @@
+/* SPDX-License-Identifier: (GPL-2.0-only OR BSD-2-Clause) */
+/* Data structures shared between BPF and tools. */
+#ifndef __BPERF_CGROUP_H
+#define __BPERF_CGROUP_H
+
+// These constants impact code size of bperf_cgroup.bpf.c that may result in BPF
+// verifier issues. They are exposed to control the size and also to disable BPF
+// counters when the number of user events is too large.
+
+// max cgroup hierarchy level: arbitrary
+#define BPERF_CGROUP__MAX_LEVELS  10
+// max events per cgroup: arbitrary
+#define BPERF_CGROUP__MAX_EVENTS  128
+
+#endif /* __BPERF_CGROUP_H */