1// SPDX-License-Identifier: GPL-2.0
2
3/* Copyright (c) 2021 Facebook */
4/* Copyright (c) 2021 Google */
5
6#include <assert.h>
7#include <limits.h>
8#include <unistd.h>
9#include <sys/file.h>
10#include <sys/time.h>
11#include <sys/resource.h>
12#include <linux/err.h>
13#include <linux/zalloc.h>
14#include <linux/perf_event.h>
15#include <api/fs/fs.h>
16#include <perf/bpf_perf.h>
17
18#include "affinity.h"
19#include "bpf_counter.h"
20#include "cgroup.h"
21#include "counts.h"
22#include "debug.h"
23#include "evsel.h"
24#include "evlist.h"
25#include "target.h"
26#include "cpumap.h"
27#include "thread_map.h"
28
29#include "bpf_skel/bperf_cgroup.skel.h"
30
31static struct perf_event_attr cgrp_switch_attr = {
32 .type = PERF_TYPE_SOFTWARE,
33 .config = PERF_COUNT_SW_CGROUP_SWITCHES,
34 .size = sizeof(cgrp_switch_attr),
35 .sample_period = 1,
36 .disabled = 1,
37};
38
39static struct evsel *cgrp_switch;
40static struct bperf_cgroup_bpf *skel;
41
42#define FD(evt, cpu) (*(int *)xyarray__entry(evt->core.fd, cpu, 0))
43
44static int bperf_load_program(struct evlist *evlist)
45{
46 struct bpf_link *link;
47 struct evsel *evsel;
48 struct cgroup *cgrp, *leader_cgrp;
49 int i, j;
50 struct perf_cpu cpu;
51 int total_cpus = cpu__max_cpu().cpu;
52 int map_size, map_fd;
53 int prog_fd, err;
54
55 skel = bperf_cgroup_bpf__open();
56 if (!skel) {
57 pr_err("Failed to open cgroup skeleton\n");
58 return -1;
59 }
60
61 skel->rodata->num_cpus = total_cpus;
62 skel->rodata->num_events = evlist->core.nr_entries / nr_cgroups;
63
64 BUG_ON(evlist->core.nr_entries % nr_cgroups != 0);
65
66 /* we need one copy of events per cpu for reading */
67 map_size = total_cpus * evlist->core.nr_entries / nr_cgroups;
68 bpf_map__set_max_entries(skel->maps.events, map_size);
69 bpf_map__set_max_entries(skel->maps.cgrp_idx, nr_cgroups);
70 /* previous result is saved in a per-cpu array */
71 map_size = evlist->core.nr_entries / nr_cgroups;
72 bpf_map__set_max_entries(skel->maps.prev_readings, map_size);
73 /* cgroup result needs all events (per-cpu) */
74 map_size = evlist->core.nr_entries;
75 bpf_map__set_max_entries(skel->maps.cgrp_readings, map_size);
76
77 set_max_rlimit();
78
79 err = bperf_cgroup_bpf__load(skel);
80 if (err) {
81 pr_err("Failed to load cgroup skeleton\n");
82 goto out;
83 }
84
85 if (cgroup_is_v2(subsys: "perf_event") > 0)
86 skel->bss->use_cgroup_v2 = 1;
87
88 err = -1;
89
90 cgrp_switch = evsel__new(attr: &cgrp_switch_attr);
91 if (evsel__open_per_cpu(evsel: cgrp_switch, cpus: evlist->core.all_cpus, cpu_map_idx: -1) < 0) {
92 pr_err("Failed to open cgroup switches event\n");
93 goto out;
94 }
95
96 perf_cpu_map__for_each_cpu(cpu, i, evlist->core.all_cpus) {
97 link = bpf_program__attach_perf_event(skel->progs.on_cgrp_switch,
98 FD(cgrp_switch, i));
99 if (IS_ERR(ptr: link)) {
100 pr_err("Failed to attach cgroup program\n");
101 err = PTR_ERR(ptr: link);
102 goto out;
103 }
104 }
105
106 /*
107 * Update cgrp_idx map from cgroup-id to event index.
108 */
109 cgrp = NULL;
110 i = 0;
111
112 evlist__for_each_entry(evlist, evsel) {
113 if (cgrp == NULL || evsel->cgrp == leader_cgrp) {
114 leader_cgrp = evsel->cgrp;
115 evsel->cgrp = NULL;
116
117 /* open single copy of the events w/o cgroup */
118 err = evsel__open_per_cpu(evsel, cpus: evsel->core.cpus, cpu_map_idx: -1);
119 if (err == 0)
120 evsel->supported = true;
121
122 map_fd = bpf_map__fd(skel->maps.events);
123 perf_cpu_map__for_each_cpu(cpu, j, evsel->core.cpus) {
124 int fd = FD(evsel, j);
125 __u32 idx = evsel->core.idx * total_cpus + cpu.cpu;
126
127 bpf_map_update_elem(map_fd, &idx, &fd, BPF_ANY);
128 }
129
130 evsel->cgrp = leader_cgrp;
131 }
132
133 if (evsel->cgrp == cgrp)
134 continue;
135
136 cgrp = evsel->cgrp;
137
138 if (read_cgroup_id(cgrp) < 0) {
139 pr_err("Failed to get cgroup id\n");
140 err = -1;
141 goto out;
142 }
143
144 map_fd = bpf_map__fd(skel->maps.cgrp_idx);
145 err = bpf_map_update_elem(map_fd, &cgrp->id, &i, BPF_ANY);
146 if (err < 0) {
147 pr_err("Failed to update cgroup index map\n");
148 goto out;
149 }
150
151 i++;
152 }
153
154 /*
155 * bperf uses BPF_PROG_TEST_RUN to get accurate reading. Check
156 * whether the kernel support it
157 */
158 prog_fd = bpf_program__fd(skel->progs.trigger_read);
159 err = bperf_trigger_reading(prog_fd, 0);
160 if (err) {
161 pr_warning("The kernel does not support test_run for raw_tp BPF programs.\n"
162 "Therefore, --for-each-cgroup might show inaccurate readings\n");
163 err = 0;
164 }
165
166out:
167 return err;
168}
169
170static int bperf_cgrp__load(struct evsel *evsel,
171 struct target *target __maybe_unused)
172{
173 static bool bperf_loaded = false;
174
175 evsel->bperf_leader_prog_fd = -1;
176 evsel->bperf_leader_link_fd = -1;
177
178 if (!bperf_loaded && bperf_load_program(evlist: evsel->evlist))
179 return -1;
180
181 bperf_loaded = true;
182 /* just to bypass bpf_counter_skip() */
183 evsel->follower_skel = (struct bperf_follower_bpf *)skel;
184
185 return 0;
186}
187
188static int bperf_cgrp__install_pe(struct evsel *evsel __maybe_unused,
189 int cpu __maybe_unused, int fd __maybe_unused)
190{
191 /* nothing to do */
192 return 0;
193}
194
195/*
196 * trigger the leader prog on each cpu, so the cgrp_reading map could get
197 * the latest results.
198 */
199static int bperf_cgrp__sync_counters(struct evlist *evlist)
200{
201 struct perf_cpu cpu;
202 int idx;
203 int prog_fd = bpf_program__fd(skel->progs.trigger_read);
204
205 perf_cpu_map__for_each_cpu(cpu, idx, evlist->core.all_cpus)
206 bperf_trigger_reading(prog_fd, cpu.cpu);
207
208 return 0;
209}
210
211static int bperf_cgrp__enable(struct evsel *evsel)
212{
213 if (evsel->core.idx)
214 return 0;
215
216 bperf_cgrp__sync_counters(evlist: evsel->evlist);
217
218 skel->bss->enabled = 1;
219 return 0;
220}
221
222static int bperf_cgrp__disable(struct evsel *evsel)
223{
224 if (evsel->core.idx)
225 return 0;
226
227 bperf_cgrp__sync_counters(evlist: evsel->evlist);
228
229 skel->bss->enabled = 0;
230 return 0;
231}
232
233static int bperf_cgrp__read(struct evsel *evsel)
234{
235 struct evlist *evlist = evsel->evlist;
236 int total_cpus = cpu__max_cpu().cpu;
237 struct perf_counts_values *counts;
238 struct bpf_perf_event_value *values;
239 int reading_map_fd, err = 0;
240
241 if (evsel->core.idx)
242 return 0;
243
244 bperf_cgrp__sync_counters(evlist: evsel->evlist);
245
246 values = calloc(total_cpus, sizeof(*values));
247 if (values == NULL)
248 return -ENOMEM;
249
250 reading_map_fd = bpf_map__fd(skel->maps.cgrp_readings);
251
252 evlist__for_each_entry(evlist, evsel) {
253 __u32 idx = evsel->core.idx;
254 int i;
255 struct perf_cpu cpu;
256
257 err = bpf_map_lookup_elem(reading_map_fd, &idx, values);
258 if (err) {
259 pr_err("bpf map lookup failed: idx=%u, event=%s, cgrp=%s\n",
260 idx, evsel__name(evsel), evsel->cgrp->name);
261 goto out;
262 }
263
264 perf_cpu_map__for_each_cpu(cpu, i, evsel->core.cpus) {
265 counts = perf_counts(counts: evsel->counts, cpu_map_idx: i, thread: 0);
266 counts->val = values[cpu.cpu].counter;
267 counts->ena = values[cpu.cpu].enabled;
268 counts->run = values[cpu.cpu].running;
269 }
270 }
271
272out:
273 free(values);
274 return err;
275}
276
277static int bperf_cgrp__destroy(struct evsel *evsel)
278{
279 if (evsel->core.idx)
280 return 0;
281
282 bperf_cgroup_bpf__destroy(skel);
283 evsel__delete(evsel: cgrp_switch); // it'll destroy on_switch progs too
284
285 return 0;
286}
287
288struct bpf_counter_ops bperf_cgrp_ops = {
289 .load = bperf_cgrp__load,
290 .enable = bperf_cgrp__enable,
291 .disable = bperf_cgrp__disable,
292 .read = bperf_cgrp__read,
293 .install_pe = bperf_cgrp__install_pe,
294 .destroy = bperf_cgrp__destroy,
295};
296

source code of linux/tools/perf/util/bpf_counter_cgroup.c