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 
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 	__u32 i, cpu;
50 	__u32 nr_cpus = evlist->core.all_cpus->nr;
51 	int total_cpus = cpu__max_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("perf_event") > 0)
86 		skel->bss->use_cgroup_v2 = 1;
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 	for (i = 0; i < nr_cpus; i++) {
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, evlist->core.all_cpus, -1);
119 			if (err) {
120 				pr_err("Failed to open first cgroup events\n");
121 				goto out;
122 			}
123 
124 			map_fd = bpf_map__fd(skel->maps.events);
125 			for (cpu = 0; cpu < nr_cpus; cpu++) {
126 				int fd = FD(evsel, cpu);
127 				__u32 idx = evsel->core.idx * total_cpus +
128 					evlist->core.all_cpus->map[cpu];
129 
130 				err = bpf_map_update_elem(map_fd, &idx, &fd,
131 							  BPF_ANY);
132 				if (err < 0) {
133 					pr_err("Failed to update perf_event fd\n");
134 					goto out;
135 				}
136 			}
137 
138 			evsel->cgrp = leader_cgrp;
139 		}
140 		evsel->supported = true;
141 
142 		if (evsel->cgrp == cgrp)
143 			continue;
144 
145 		cgrp = evsel->cgrp;
146 
147 		if (read_cgroup_id(cgrp) < 0) {
148 			pr_err("Failed to get cgroup id\n");
149 			err = -1;
150 			goto out;
151 		}
152 
153 		map_fd = bpf_map__fd(skel->maps.cgrp_idx);
154 		err = bpf_map_update_elem(map_fd, &cgrp->id, &i, BPF_ANY);
155 		if (err < 0) {
156 			pr_err("Failed to update cgroup index map\n");
157 			goto out;
158 		}
159 
160 		i++;
161 	}
162 
163 	/*
164 	 * bperf uses BPF_PROG_TEST_RUN to get accurate reading. Check
165 	 * whether the kernel support it
166 	 */
167 	prog_fd = bpf_program__fd(skel->progs.trigger_read);
168 	err = bperf_trigger_reading(prog_fd, 0);
169 	if (err) {
170 		pr_warning("The kernel does not support test_run for raw_tp BPF programs.\n"
171 			   "Therefore, --for-each-cgroup might show inaccurate readings\n");
172 		err = 0;
173 	}
174 
175 out:
176 	return err;
177 }
178 
179 static int bperf_cgrp__load(struct evsel *evsel,
180 			    struct target *target __maybe_unused)
181 {
182 	static bool bperf_loaded = false;
183 
184 	evsel->bperf_leader_prog_fd = -1;
185 	evsel->bperf_leader_link_fd = -1;
186 
187 	if (!bperf_loaded && bperf_load_program(evsel->evlist))
188 		return -1;
189 
190 	bperf_loaded = true;
191 	/* just to bypass bpf_counter_skip() */
192 	evsel->follower_skel = (struct bperf_follower_bpf *)skel;
193 
194 	return 0;
195 }
196 
197 static int bperf_cgrp__install_pe(struct evsel *evsel __maybe_unused,
198 				  int cpu __maybe_unused, int fd __maybe_unused)
199 {
200 	/* nothing to do */
201 	return 0;
202 }
203 
204 /*
205  * trigger the leader prog on each cpu, so the cgrp_reading map could get
206  * the latest results.
207  */
208 static int bperf_cgrp__sync_counters(struct evlist *evlist)
209 {
210 	int i, cpu;
211 	int nr_cpus = evlist->core.all_cpus->nr;
212 	int prog_fd = bpf_program__fd(skel->progs.trigger_read);
213 
214 	for (i = 0; i < nr_cpus; i++) {
215 		cpu = evlist->core.all_cpus->map[i];
216 		bperf_trigger_reading(prog_fd, cpu);
217 	}
218 
219 	return 0;
220 }
221 
222 static int bperf_cgrp__enable(struct evsel *evsel)
223 {
224 	if (evsel->core.idx)
225 		return 0;
226 
227 	bperf_cgrp__sync_counters(evsel->evlist);
228 
229 	skel->bss->enabled = 1;
230 	return 0;
231 }
232 
233 static int bperf_cgrp__disable(struct evsel *evsel)
234 {
235 	if (evsel->core.idx)
236 		return 0;
237 
238 	bperf_cgrp__sync_counters(evsel->evlist);
239 
240 	skel->bss->enabled = 0;
241 	return 0;
242 }
243 
244 static int bperf_cgrp__read(struct evsel *evsel)
245 {
246 	struct evlist *evlist = evsel->evlist;
247 	int i, cpu, nr_cpus = evlist->core.all_cpus->nr;
248 	int total_cpus = cpu__max_cpu();
249 	struct perf_counts_values *counts;
250 	struct bpf_perf_event_value *values;
251 	int reading_map_fd, err = 0;
252 	__u32 idx;
253 
254 	if (evsel->core.idx)
255 		return 0;
256 
257 	bperf_cgrp__sync_counters(evsel->evlist);
258 
259 	values = calloc(total_cpus, sizeof(*values));
260 	if (values == NULL)
261 		return -ENOMEM;
262 
263 	reading_map_fd = bpf_map__fd(skel->maps.cgrp_readings);
264 
265 	evlist__for_each_entry(evlist, evsel) {
266 		idx = evsel->core.idx;
267 		err = bpf_map_lookup_elem(reading_map_fd, &idx, values);
268 		if (err) {
269 			pr_err("bpf map lookup falied: idx=%u, event=%s, cgrp=%s\n",
270 			       idx, evsel__name(evsel), evsel->cgrp->name);
271 			goto out;
272 		}
273 
274 		for (i = 0; i < nr_cpus; i++) {
275 			cpu = evlist->core.all_cpus->map[i];
276 
277 			counts = perf_counts(evsel->counts, i, 0);
278 			counts->val = values[cpu].counter;
279 			counts->ena = values[cpu].enabled;
280 			counts->run = values[cpu].running;
281 		}
282 	}
283 
284 out:
285 	free(values);
286 	return err;
287 }
288 
289 static int bperf_cgrp__destroy(struct evsel *evsel)
290 {
291 	if (evsel->core.idx)
292 		return 0;
293 
294 	bperf_cgroup_bpf__destroy(skel);
295 	evsel__delete(cgrp_switch);  // it'll destroy on_switch progs too
296 
297 	return 0;
298 }
299 
300 struct bpf_counter_ops bperf_cgrp_ops = {
301 	.load       = bperf_cgrp__load,
302 	.enable     = bperf_cgrp__enable,
303 	.disable    = bperf_cgrp__disable,
304 	.read       = bperf_cgrp__read,
305 	.install_pe = bperf_cgrp__install_pe,
306 	.destroy    = bperf_cgrp__destroy,
307 };
308