1 // SPDX-License-Identifier: GPL-2.0
3 /* Copyright (c) 2019 Facebook */
10 #include <linux/err.h>
11 #include <linux/zalloc.h>
12 #include <api/fs/fs.h>
13 #include <perf/bpf_perf.h>
15 #include "bpf_counter.h"
16 #include "bpf-utils.h"
24 #include "thread_map.h"
26 #include "bpf_skel/bpf_prog_profiler.skel.h"
27 #include "bpf_skel/bperf_u.h"
28 #include "bpf_skel/bperf_leader.skel.h"
29 #include "bpf_skel/bperf_follower.skel.h"
31 #define ATTR_MAP_SIZE 16
33 static inline void *u64_to_ptr(__u64 ptr)
35 return (void *)(unsigned long)ptr;
38 static struct bpf_counter *bpf_counter_alloc(void)
40 struct bpf_counter *counter;
42 counter = zalloc(sizeof(*counter));
44 INIT_LIST_HEAD(&counter->list);
48 static int bpf_program_profiler__destroy(struct evsel *evsel)
50 struct bpf_counter *counter, *tmp;
52 list_for_each_entry_safe(counter, tmp,
53 &evsel->bpf_counter_list, list) {
54 list_del_init(&counter->list);
55 bpf_prog_profiler_bpf__destroy(counter->skel);
58 assert(list_empty(&evsel->bpf_counter_list));
63 static char *bpf_target_prog_name(int tgt_fd)
65 struct bpf_func_info *func_info;
66 struct perf_bpil *info_linear;
67 const struct btf_type *t;
68 struct btf *btf = NULL;
71 info_linear = get_bpf_prog_info_linear(tgt_fd, 1UL << PERF_BPIL_FUNC_INFO);
72 if (IS_ERR_OR_NULL(info_linear)) {
73 pr_debug("failed to get info_linear for prog FD %d\n", tgt_fd);
77 if (info_linear->info.btf_id == 0) {
78 pr_debug("prog FD %d doesn't have valid btf\n", tgt_fd);
82 btf = btf__load_from_kernel_by_id(info_linear->info.btf_id);
83 if (libbpf_get_error(btf)) {
84 pr_debug("failed to load btf for prog FD %d\n", tgt_fd);
88 func_info = u64_to_ptr(info_linear->info.func_info);
89 t = btf__type_by_id(btf, func_info[0].type_id);
91 pr_debug("btf %d doesn't have type %d\n",
92 info_linear->info.btf_id, func_info[0].type_id);
95 name = strdup(btf__name_by_offset(btf, t->name_off));
102 static int bpf_program_profiler_load_one(struct evsel *evsel, u32 prog_id)
104 struct bpf_prog_profiler_bpf *skel;
105 struct bpf_counter *counter;
106 struct bpf_program *prog;
107 char *prog_name = NULL;
111 prog_fd = bpf_prog_get_fd_by_id(prog_id);
113 pr_err("Failed to open fd for bpf prog %u\n", prog_id);
116 counter = bpf_counter_alloc();
122 skel = bpf_prog_profiler_bpf__open();
124 pr_err("Failed to open bpf skeleton\n");
128 skel->rodata->num_cpu = evsel__nr_cpus(evsel);
130 bpf_map__set_max_entries(skel->maps.events, evsel__nr_cpus(evsel));
131 bpf_map__set_max_entries(skel->maps.fentry_readings, 1);
132 bpf_map__set_max_entries(skel->maps.accum_readings, 1);
134 prog_name = bpf_target_prog_name(prog_fd);
136 pr_err("Failed to get program name for bpf prog %u. Does it have BTF?\n", prog_id);
140 bpf_object__for_each_program(prog, skel->obj) {
141 err = bpf_program__set_attach_target(prog, prog_fd, prog_name);
143 pr_err("bpf_program__set_attach_target failed.\n"
144 "Does bpf prog %u have BTF?\n", prog_id);
149 err = bpf_prog_profiler_bpf__load(skel);
151 pr_err("bpf_prog_profiler_bpf__load failed\n");
155 assert(skel != NULL);
156 counter->skel = skel;
157 list_add(&counter->list, &evsel->bpf_counter_list);
162 bpf_prog_profiler_bpf__destroy(skel);
169 static int bpf_program_profiler__load(struct evsel *evsel, struct target *target)
171 char *bpf_str, *bpf_str_, *tok, *saveptr = NULL, *p;
175 bpf_str_ = bpf_str = strdup(target->bpf_str);
179 while ((tok = strtok_r(bpf_str, ",", &saveptr)) != NULL) {
180 prog_id = strtoul(tok, &p, 10);
181 if (prog_id == 0 || prog_id == UINT_MAX ||
182 (*p != '\0' && *p != ',')) {
183 pr_err("Failed to parse bpf prog ids %s\n",
189 ret = bpf_program_profiler_load_one(evsel, prog_id);
191 bpf_program_profiler__destroy(evsel);
201 static int bpf_program_profiler__enable(struct evsel *evsel)
203 struct bpf_counter *counter;
206 list_for_each_entry(counter, &evsel->bpf_counter_list, list) {
207 assert(counter->skel != NULL);
208 ret = bpf_prog_profiler_bpf__attach(counter->skel);
210 bpf_program_profiler__destroy(evsel);
217 static int bpf_program_profiler__disable(struct evsel *evsel)
219 struct bpf_counter *counter;
221 list_for_each_entry(counter, &evsel->bpf_counter_list, list) {
222 assert(counter->skel != NULL);
223 bpf_prog_profiler_bpf__detach(counter->skel);
228 static int bpf_program_profiler__read(struct evsel *evsel)
230 // BPF_MAP_TYPE_PERCPU_ARRAY uses /sys/devices/system/cpu/possible
231 // Sometimes possible > online, like on a Ryzen 3900X that has 24
232 // threads but its possible showed 0-31 -acme
233 int num_cpu_bpf = libbpf_num_possible_cpus();
234 struct bpf_perf_event_value values[num_cpu_bpf];
235 struct bpf_counter *counter;
236 struct perf_counts_values *counts;
239 int err, idx, bpf_cpu;
241 if (list_empty(&evsel->bpf_counter_list))
244 perf_cpu_map__for_each_idx(idx, evsel__cpus(evsel)) {
245 counts = perf_counts(evsel->counts, idx, 0);
250 list_for_each_entry(counter, &evsel->bpf_counter_list, list) {
251 struct bpf_prog_profiler_bpf *skel = counter->skel;
253 assert(skel != NULL);
254 reading_map_fd = bpf_map__fd(skel->maps.accum_readings);
256 err = bpf_map_lookup_elem(reading_map_fd, &key, values);
258 pr_err("failed to read value\n");
262 for (bpf_cpu = 0; bpf_cpu < num_cpu_bpf; bpf_cpu++) {
263 idx = perf_cpu_map__idx(evsel__cpus(evsel),
264 (struct perf_cpu){.cpu = bpf_cpu});
267 counts = perf_counts(evsel->counts, idx, 0);
268 counts->val += values[bpf_cpu].counter;
269 counts->ena += values[bpf_cpu].enabled;
270 counts->run += values[bpf_cpu].running;
276 static int bpf_program_profiler__install_pe(struct evsel *evsel, int cpu_map_idx,
279 struct bpf_prog_profiler_bpf *skel;
280 struct bpf_counter *counter;
283 list_for_each_entry(counter, &evsel->bpf_counter_list, list) {
284 skel = counter->skel;
285 assert(skel != NULL);
287 ret = bpf_map_update_elem(bpf_map__fd(skel->maps.events),
288 &cpu_map_idx, &fd, BPF_ANY);
295 struct bpf_counter_ops bpf_program_profiler_ops = {
296 .load = bpf_program_profiler__load,
297 .enable = bpf_program_profiler__enable,
298 .disable = bpf_program_profiler__disable,
299 .read = bpf_program_profiler__read,
300 .destroy = bpf_program_profiler__destroy,
301 .install_pe = bpf_program_profiler__install_pe,
304 static bool bperf_attr_map_compatible(int attr_map_fd)
306 struct bpf_map_info map_info = {0};
307 __u32 map_info_len = sizeof(map_info);
310 err = bpf_obj_get_info_by_fd(attr_map_fd, &map_info, &map_info_len);
314 return (map_info.key_size == sizeof(struct perf_event_attr)) &&
315 (map_info.value_size == sizeof(struct perf_event_attr_map_entry));
318 static int bperf_lock_attr_map(struct target *target)
323 if (target->attr_map) {
324 scnprintf(path, PATH_MAX, "%s", target->attr_map);
326 scnprintf(path, PATH_MAX, "%s/fs/bpf/%s", sysfs__mountpoint(),
327 BPF_PERF_DEFAULT_ATTR_MAP_PATH);
330 if (access(path, F_OK)) {
331 map_fd = bpf_map_create(BPF_MAP_TYPE_HASH, NULL,
332 sizeof(struct perf_event_attr),
333 sizeof(struct perf_event_attr_map_entry),
334 ATTR_MAP_SIZE, NULL);
338 err = bpf_obj_pin(map_fd, path);
340 /* someone pinned the map in parallel? */
342 map_fd = bpf_obj_get(path);
347 map_fd = bpf_obj_get(path);
352 if (!bperf_attr_map_compatible(map_fd)) {
357 err = flock(map_fd, LOCK_EX);
365 static int bperf_check_target(struct evsel *evsel,
366 struct target *target,
367 enum bperf_filter_type *filter_type,
368 __u32 *filter_entry_cnt)
370 if (evsel->core.leader->nr_members > 1) {
371 pr_err("bpf managed perf events do not yet support groups.\n");
375 /* determine filter type based on target */
376 if (target->system_wide) {
377 *filter_type = BPERF_FILTER_GLOBAL;
378 *filter_entry_cnt = 1;
379 } else if (target->cpu_list) {
380 *filter_type = BPERF_FILTER_CPU;
381 *filter_entry_cnt = perf_cpu_map__nr(evsel__cpus(evsel));
382 } else if (target->tid) {
383 *filter_type = BPERF_FILTER_PID;
384 *filter_entry_cnt = perf_thread_map__nr(evsel->core.threads);
385 } else if (target->pid || evsel->evlist->workload.pid != -1) {
386 *filter_type = BPERF_FILTER_TGID;
387 *filter_entry_cnt = perf_thread_map__nr(evsel->core.threads);
389 pr_err("bpf managed perf events do not yet support these targets.\n");
396 static struct perf_cpu_map *all_cpu_map;
398 static int bperf_reload_leader_program(struct evsel *evsel, int attr_map_fd,
399 struct perf_event_attr_map_entry *entry)
401 struct bperf_leader_bpf *skel = bperf_leader_bpf__open();
402 int link_fd, diff_map_fd, err;
403 struct bpf_link *link = NULL;
406 pr_err("Failed to open leader skeleton\n");
410 bpf_map__set_max_entries(skel->maps.events, libbpf_num_possible_cpus());
411 err = bperf_leader_bpf__load(skel);
413 pr_err("Failed to load leader skeleton\n");
417 link = bpf_program__attach(skel->progs.on_switch);
419 pr_err("Failed to attach leader program\n");
424 link_fd = bpf_link__fd(link);
425 diff_map_fd = bpf_map__fd(skel->maps.diff_readings);
426 entry->link_id = bpf_link_get_id(link_fd);
427 entry->diff_map_id = bpf_map_get_id(diff_map_fd);
428 err = bpf_map_update_elem(attr_map_fd, &evsel->core.attr, entry, BPF_ANY);
431 evsel->bperf_leader_link_fd = bpf_link_get_fd_by_id(entry->link_id);
432 assert(evsel->bperf_leader_link_fd >= 0);
435 * save leader_skel for install_pe, which is called within
436 * following evsel__open_per_cpu call
438 evsel->leader_skel = skel;
439 evsel__open_per_cpu(evsel, all_cpu_map, -1);
442 bperf_leader_bpf__destroy(skel);
443 bpf_link__destroy(link);
447 static int bperf__load(struct evsel *evsel, struct target *target)
449 struct perf_event_attr_map_entry entry = {0xffffffff, 0xffffffff};
450 int attr_map_fd, diff_map_fd = -1, err;
451 enum bperf_filter_type filter_type;
452 __u32 filter_entry_cnt, i;
454 if (bperf_check_target(evsel, target, &filter_type, &filter_entry_cnt))
458 all_cpu_map = perf_cpu_map__new_online_cpus();
463 evsel->bperf_leader_prog_fd = -1;
464 evsel->bperf_leader_link_fd = -1;
467 * Step 1: hold a fd on the leader program and the bpf_link, if
468 * the program is not already gone, reload the program.
469 * Use flock() to ensure exclusive access to the perf_event_attr
472 attr_map_fd = bperf_lock_attr_map(target);
473 if (attr_map_fd < 0) {
474 pr_err("Failed to lock perf_event_attr map\n");
478 err = bpf_map_lookup_elem(attr_map_fd, &evsel->core.attr, &entry);
480 err = bpf_map_update_elem(attr_map_fd, &evsel->core.attr, &entry, BPF_ANY);
485 evsel->bperf_leader_link_fd = bpf_link_get_fd_by_id(entry.link_id);
486 if (evsel->bperf_leader_link_fd < 0 &&
487 bperf_reload_leader_program(evsel, attr_map_fd, &entry)) {
492 * The bpf_link holds reference to the leader program, and the
493 * leader program holds reference to the maps. Therefore, if
494 * link_id is valid, diff_map_id should also be valid.
496 evsel->bperf_leader_prog_fd = bpf_prog_get_fd_by_id(
497 bpf_link_get_prog_id(evsel->bperf_leader_link_fd));
498 assert(evsel->bperf_leader_prog_fd >= 0);
500 diff_map_fd = bpf_map_get_fd_by_id(entry.diff_map_id);
501 assert(diff_map_fd >= 0);
504 * bperf uses BPF_PROG_TEST_RUN to get accurate reading. Check
505 * whether the kernel support it
507 err = bperf_trigger_reading(evsel->bperf_leader_prog_fd, 0);
509 pr_err("The kernel does not support test_run for raw_tp BPF programs.\n"
510 "Therefore, --use-bpf might show inaccurate readings\n");
514 /* Step 2: load the follower skeleton */
515 evsel->follower_skel = bperf_follower_bpf__open();
516 if (!evsel->follower_skel) {
518 pr_err("Failed to open follower skeleton\n");
522 /* attach fexit program to the leader program */
523 bpf_program__set_attach_target(evsel->follower_skel->progs.fexit_XXX,
524 evsel->bperf_leader_prog_fd, "on_switch");
526 /* connect to leader diff_reading map */
527 bpf_map__reuse_fd(evsel->follower_skel->maps.diff_readings, diff_map_fd);
529 /* set up reading map */
530 bpf_map__set_max_entries(evsel->follower_skel->maps.accum_readings,
532 /* set up follower filter based on target */
533 bpf_map__set_max_entries(evsel->follower_skel->maps.filter,
535 err = bperf_follower_bpf__load(evsel->follower_skel);
537 pr_err("Failed to load follower skeleton\n");
538 bperf_follower_bpf__destroy(evsel->follower_skel);
539 evsel->follower_skel = NULL;
543 for (i = 0; i < filter_entry_cnt; i++) {
547 if (filter_type == BPERF_FILTER_PID ||
548 filter_type == BPERF_FILTER_TGID)
549 key = perf_thread_map__pid(evsel->core.threads, i);
550 else if (filter_type == BPERF_FILTER_CPU)
551 key = perf_cpu_map__cpu(evsel->core.cpus, i).cpu;
555 filter_map_fd = bpf_map__fd(evsel->follower_skel->maps.filter);
556 bpf_map_update_elem(filter_map_fd, &key, &i, BPF_ANY);
559 evsel->follower_skel->bss->type = filter_type;
561 err = bperf_follower_bpf__attach(evsel->follower_skel);
564 if (err && evsel->bperf_leader_link_fd >= 0)
565 close(evsel->bperf_leader_link_fd);
566 if (err && evsel->bperf_leader_prog_fd >= 0)
567 close(evsel->bperf_leader_prog_fd);
568 if (diff_map_fd >= 0)
571 flock(attr_map_fd, LOCK_UN);
577 static int bperf__install_pe(struct evsel *evsel, int cpu_map_idx, int fd)
579 struct bperf_leader_bpf *skel = evsel->leader_skel;
581 return bpf_map_update_elem(bpf_map__fd(skel->maps.events),
582 &cpu_map_idx, &fd, BPF_ANY);
586 * trigger the leader prog on each cpu, so the accum_reading map could get
587 * the latest readings.
589 static int bperf_sync_counters(struct evsel *evsel)
593 num_cpu = perf_cpu_map__nr(all_cpu_map);
594 for (i = 0; i < num_cpu; i++) {
595 cpu = perf_cpu_map__cpu(all_cpu_map, i).cpu;
596 bperf_trigger_reading(evsel->bperf_leader_prog_fd, cpu);
601 static int bperf__enable(struct evsel *evsel)
603 evsel->follower_skel->bss->enabled = 1;
607 static int bperf__disable(struct evsel *evsel)
609 evsel->follower_skel->bss->enabled = 0;
613 static int bperf__read(struct evsel *evsel)
615 struct bperf_follower_bpf *skel = evsel->follower_skel;
616 __u32 num_cpu_bpf = cpu__max_cpu().cpu;
617 struct bpf_perf_event_value values[num_cpu_bpf];
618 struct perf_counts_values *counts;
619 int reading_map_fd, err = 0;
623 bperf_sync_counters(evsel);
624 reading_map_fd = bpf_map__fd(skel->maps.accum_readings);
626 for (i = 0; i < bpf_map__max_entries(skel->maps.accum_readings); i++) {
627 struct perf_cpu entry;
630 err = bpf_map_lookup_elem(reading_map_fd, &i, values);
633 switch (evsel->follower_skel->bss->type) {
634 case BPERF_FILTER_GLOBAL:
637 perf_cpu_map__for_each_cpu(entry, j, evsel__cpus(evsel)) {
638 counts = perf_counts(evsel->counts, j, 0);
639 counts->val = values[entry.cpu].counter;
640 counts->ena = values[entry.cpu].enabled;
641 counts->run = values[entry.cpu].running;
644 case BPERF_FILTER_CPU:
645 cpu = perf_cpu_map__cpu(evsel__cpus(evsel), i).cpu;
647 counts = perf_counts(evsel->counts, i, 0);
648 counts->val = values[cpu].counter;
649 counts->ena = values[cpu].enabled;
650 counts->run = values[cpu].running;
652 case BPERF_FILTER_PID:
653 case BPERF_FILTER_TGID:
654 counts = perf_counts(evsel->counts, 0, i);
659 for (cpu = 0; cpu < num_cpu_bpf; cpu++) {
660 counts->val += values[cpu].counter;
661 counts->ena += values[cpu].enabled;
662 counts->run += values[cpu].running;
673 static int bperf__destroy(struct evsel *evsel)
675 bperf_follower_bpf__destroy(evsel->follower_skel);
676 close(evsel->bperf_leader_prog_fd);
677 close(evsel->bperf_leader_link_fd);
682 * bperf: share hardware PMCs with BPF
684 * perf uses performance monitoring counters (PMC) to monitor system
685 * performance. The PMCs are limited hardware resources. For example,
686 * Intel CPUs have 3x fixed PMCs and 4x programmable PMCs per cpu.
688 * Modern data center systems use these PMCs in many different ways:
689 * system level monitoring, (maybe nested) container level monitoring, per
690 * process monitoring, profiling (in sample mode), etc. In some cases,
691 * there are more active perf_events than available hardware PMCs. To allow
692 * all perf_events to have a chance to run, it is necessary to do expensive
693 * time multiplexing of events.
695 * On the other hand, many monitoring tools count the common metrics
696 * (cycles, instructions). It is a waste to have multiple tools create
697 * multiple perf_events of "cycles" and occupy multiple PMCs.
699 * bperf tries to reduce such wastes by allowing multiple perf_events of
700 * "cycles" or "instructions" (at different scopes) to share PMUs. Instead
701 * of having each perf-stat session to read its own perf_events, bperf uses
702 * BPF programs to read the perf_events and aggregate readings to BPF maps.
703 * Then, the perf-stat session(s) reads the values from these BPF maps.
706 * shared progs and maps <- || -> per session progs and maps
710 * --------------- fexit || -----------------
711 * | --------||----> | follower prog |
712 * --------------- / || --- -----------------
713 * cs -> | leader prog |/ ||/ | |
714 * --> --------------- /|| -------------- ------------------
715 * / | | / || | filter map | | accum_readings |
716 * / ------------ ------------ || -------------- ------------------
717 * | | prev map | | diff map | || |
718 * | ------------ ------------ || |
720 * = \ ==================================================== | ============
724 * BPF_PROG_TEST_RUN BPF_MAP_LOOKUP_ELEM
727 * \------ perf-stat ----------------------/
729 * The figure above shows the architecture of bperf. Note that the figure
730 * is divided into 3 regions: shared progs and maps (top left), per session
731 * progs and maps (top right), and user space (bottom).
733 * The leader prog is triggered on each context switch (cs). The leader
734 * prog reads perf_events and stores the difference (current_reading -
735 * previous_reading) to the diff map. For the same metric, e.g. "cycles",
736 * multiple perf-stat sessions share the same leader prog.
738 * Each perf-stat session creates a follower prog as fexit program to the
739 * leader prog. It is possible to attach up to BPF_MAX_TRAMP_PROGS (38)
740 * follower progs to the same leader prog. The follower prog checks current
741 * task and processor ID to decide whether to add the value from the diff
742 * map to its accumulated reading map (accum_readings).
744 * Finally, perf-stat user space reads the value from accum_reading map.
746 * Besides context switch, it is also necessary to trigger the leader prog
747 * before perf-stat reads the value. Otherwise, the accum_reading map may
748 * not have the latest reading from the perf_events. This is achieved by
749 * triggering the event via sys_bpf(BPF_PROG_TEST_RUN) to each CPU.
751 * Comment before the definition of struct perf_event_attr_map_entry
752 * describes how different sessions of perf-stat share information about
756 struct bpf_counter_ops bperf_ops = {
758 .enable = bperf__enable,
759 .disable = bperf__disable,
761 .install_pe = bperf__install_pe,
762 .destroy = bperf__destroy,
765 extern struct bpf_counter_ops bperf_cgrp_ops;
767 static inline bool bpf_counter_skip(struct evsel *evsel)
769 return evsel->bpf_counter_ops == NULL;
772 int bpf_counter__install_pe(struct evsel *evsel, int cpu_map_idx, int fd)
774 if (bpf_counter_skip(evsel))
776 return evsel->bpf_counter_ops->install_pe(evsel, cpu_map_idx, fd);
779 int bpf_counter__load(struct evsel *evsel, struct target *target)
782 evsel->bpf_counter_ops = &bpf_program_profiler_ops;
783 else if (cgrp_event_expanded && target->use_bpf)
784 evsel->bpf_counter_ops = &bperf_cgrp_ops;
785 else if (target->use_bpf || evsel->bpf_counter ||
786 evsel__match_bpf_counter_events(evsel->name))
787 evsel->bpf_counter_ops = &bperf_ops;
789 if (evsel->bpf_counter_ops)
790 return evsel->bpf_counter_ops->load(evsel, target);
794 int bpf_counter__enable(struct evsel *evsel)
796 if (bpf_counter_skip(evsel))
798 return evsel->bpf_counter_ops->enable(evsel);
801 int bpf_counter__disable(struct evsel *evsel)
803 if (bpf_counter_skip(evsel))
805 return evsel->bpf_counter_ops->disable(evsel);
808 int bpf_counter__read(struct evsel *evsel)
810 if (bpf_counter_skip(evsel))
812 return evsel->bpf_counter_ops->read(evsel);
815 void bpf_counter__destroy(struct evsel *evsel)
817 if (bpf_counter_skip(evsel))
819 evsel->bpf_counter_ops->destroy(evsel);
820 evsel->bpf_counter_ops = NULL;
821 evsel->bpf_skel = NULL;