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
31 static 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
39 static struct evsel *cgrp_switch;
40 static struct bperf_cgroup_bpf *skel;
41
42 #define FD(evt, cpu) (*(int *)xyarray__entry(evt->core.fd, cpu, 0))
43
bperf_load_program(struct evlist * evlist)44 static 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 if (cgroup_is_v2("perf_event") > 0)
65 skel->rodata->use_cgroup_v2 = 1;
66
67 BUG_ON(evlist->core.nr_entries % nr_cgroups != 0);
68
69 /* we need one copy of events per cpu for reading */
70 map_size = total_cpus * evlist->core.nr_entries / nr_cgroups;
71 bpf_map__set_max_entries(skel->maps.events, map_size);
72 bpf_map__set_max_entries(skel->maps.cgrp_idx, nr_cgroups);
73 /* previous result is saved in a per-cpu array */
74 map_size = evlist->core.nr_entries / nr_cgroups;
75 bpf_map__set_max_entries(skel->maps.prev_readings, map_size);
76 /* cgroup result needs all events (per-cpu) */
77 map_size = evlist->core.nr_entries;
78 bpf_map__set_max_entries(skel->maps.cgrp_readings, map_size);
79
80 set_max_rlimit();
81
82 err = bperf_cgroup_bpf__load(skel);
83 if (err) {
84 pr_err("Failed to load cgroup skeleton\n");
85 goto out;
86 }
87
88 err = -1;
89
90 cgrp_switch = evsel__new(&cgrp_switch_attr);
91 if (evsel__open_per_cpu(cgrp_switch, evlist->core.all_cpus, -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(link)) {
100 pr_err("Failed to attach cgroup program\n");
101 err = PTR_ERR(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, evsel->core.cpus, -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_debug("Failed to get cgroup id for %s\n", cgrp->name);
140 cgrp->id = 0;
141 }
142
143 map_fd = bpf_map__fd(skel->maps.cgrp_idx);
144 err = bpf_map_update_elem(map_fd, &cgrp->id, &i, BPF_ANY);
145 if (err < 0) {
146 pr_err("Failed to update cgroup index map\n");
147 goto out;
148 }
149
150 i++;
151 }
152
153 /*
154 * bperf uses BPF_PROG_TEST_RUN to get accurate reading. Check
155 * whether the kernel support it
156 */
157 prog_fd = bpf_program__fd(skel->progs.trigger_read);
158 err = bperf_trigger_reading(prog_fd, 0);
159 if (err) {
160 pr_warning("The kernel does not support test_run for raw_tp BPF programs.\n"
161 "Therefore, --for-each-cgroup might show inaccurate readings\n");
162 err = 0;
163 }
164
165 out:
166 return err;
167 }
168
bperf_cgrp__load(struct evsel * evsel,struct target * target __maybe_unused)169 static int bperf_cgrp__load(struct evsel *evsel,
170 struct target *target __maybe_unused)
171 {
172 static bool bperf_loaded = false;
173
174 evsel->bperf_leader_prog_fd = -1;
175 evsel->bperf_leader_link_fd = -1;
176
177 if (!bperf_loaded && bperf_load_program(evsel->evlist))
178 return -1;
179
180 bperf_loaded = true;
181 /* just to bypass bpf_counter_skip() */
182 evsel->follower_skel = (struct bperf_follower_bpf *)skel;
183
184 return 0;
185 }
186
bperf_cgrp__install_pe(struct evsel * evsel __maybe_unused,int cpu __maybe_unused,int fd __maybe_unused)187 static int bperf_cgrp__install_pe(struct evsel *evsel __maybe_unused,
188 int cpu __maybe_unused, int fd __maybe_unused)
189 {
190 /* nothing to do */
191 return 0;
192 }
193
194 /*
195 * trigger the leader prog on each cpu, so the cgrp_reading map could get
196 * the latest results.
197 */
bperf_cgrp__sync_counters(struct evlist * evlist)198 static int bperf_cgrp__sync_counters(struct evlist *evlist)
199 {
200 struct perf_cpu cpu;
201 int idx;
202 int prog_fd = bpf_program__fd(skel->progs.trigger_read);
203
204 perf_cpu_map__for_each_cpu(cpu, idx, evlist->core.all_cpus)
205 bperf_trigger_reading(prog_fd, cpu.cpu);
206
207 return 0;
208 }
209
bperf_cgrp__enable(struct evsel * evsel)210 static int bperf_cgrp__enable(struct evsel *evsel)
211 {
212 if (evsel->core.idx)
213 return 0;
214
215 bperf_cgrp__sync_counters(evsel->evlist);
216
217 skel->bss->enabled = 1;
218 return 0;
219 }
220
bperf_cgrp__disable(struct evsel * evsel)221 static int bperf_cgrp__disable(struct evsel *evsel)
222 {
223 if (evsel->core.idx)
224 return 0;
225
226 bperf_cgrp__sync_counters(evsel->evlist);
227
228 skel->bss->enabled = 0;
229 return 0;
230 }
231
bperf_cgrp__read(struct evsel * evsel)232 static int bperf_cgrp__read(struct evsel *evsel)
233 {
234 struct evlist *evlist = evsel->evlist;
235 int total_cpus = cpu__max_cpu().cpu;
236 struct perf_counts_values *counts;
237 struct bpf_perf_event_value *values;
238 int reading_map_fd, err = 0;
239
240 if (evsel->core.idx)
241 return 0;
242
243 bperf_cgrp__sync_counters(evsel->evlist);
244
245 values = calloc(total_cpus, sizeof(*values));
246 if (values == NULL)
247 return -ENOMEM;
248
249 reading_map_fd = bpf_map__fd(skel->maps.cgrp_readings);
250
251 evlist__for_each_entry(evlist, evsel) {
252 __u32 idx = evsel->core.idx;
253 int i;
254 struct perf_cpu cpu;
255
256 err = bpf_map_lookup_elem(reading_map_fd, &idx, values);
257 if (err) {
258 pr_err("bpf map lookup failed: idx=%u, event=%s, cgrp=%s\n",
259 idx, evsel__name(evsel), evsel->cgrp->name);
260 goto out;
261 }
262
263 perf_cpu_map__for_each_cpu(cpu, i, evsel->core.cpus) {
264 counts = perf_counts(evsel->counts, i, 0);
265 counts->val = values[cpu.cpu].counter;
266 counts->ena = values[cpu.cpu].enabled;
267 counts->run = values[cpu.cpu].running;
268 }
269 }
270
271 out:
272 free(values);
273 return err;
274 }
275
bperf_cgrp__destroy(struct evsel * evsel)276 static int bperf_cgrp__destroy(struct evsel *evsel)
277 {
278 if (evsel->core.idx)
279 return 0;
280
281 bperf_cgroup_bpf__destroy(skel);
282 evsel__delete(cgrp_switch); // it'll destroy on_switch progs too
283
284 return 0;
285 }
286
287 struct bpf_counter_ops bperf_cgrp_ops = {
288 .load = bperf_cgrp__load,
289 .enable = bperf_cgrp__enable,
290 .disable = bperf_cgrp__disable,
291 .read = bperf_cgrp__read,
292 .install_pe = bperf_cgrp__install_pe,
293 .destroy = bperf_cgrp__destroy,
294 };
295