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 <errno.h> 8 #include <limits.h> 9 #include <unistd.h> 10 #include <sys/file.h> 11 #include <sys/time.h> 12 #include <sys/resource.h> 13 #include <linux/err.h> 14 #include <linux/perf_event.h> 15 #include <api/fs/fs.h> 16 #include <bpf/bpf.h> 17 #include <perf/bpf_perf.h> 18 19 #include "affinity.h" 20 #include "bpf_counter.h" 21 #include "cgroup.h" 22 #include "counts.h" 23 #include "debug.h" 24 #include "evsel.h" 25 #include "evlist.h" 26 #include "target.h" 27 #include "cpumap.h" 28 #include "thread_map.h" 29 30 #include "bpf_skel/bperf_cgroup.h" 31 #include "bpf_skel/bperf_cgroup.skel.h" 32 33 static struct perf_event_attr cgrp_switch_attr = { 34 .type = PERF_TYPE_SOFTWARE, 35 .config = PERF_COUNT_SW_CGROUP_SWITCHES, 36 .size = sizeof(cgrp_switch_attr), 37 .sample_period = 1, 38 .disabled = 1, 39 }; 40 41 static struct evsel *cgrp_switch; 42 static struct bperf_cgroup_bpf *skel; 43 44 #define FD(evt, cpu) (*(int *)xyarray__entry(evt->core.fd, cpu, 0)) 45 46 static void setup_rodata(struct bperf_cgroup_bpf *sk, int evlist_size) 47 { 48 int map_size, total_cpus = cpu__max_cpu().cpu; 49 50 sk->rodata->num_cpus = total_cpus; 51 sk->rodata->num_events = evlist_size / nr_cgroups; 52 53 if (cgroup_is_v2("perf_event") > 0) 54 sk->rodata->use_cgroup_v2 = 1; 55 56 BUG_ON(evlist_size % nr_cgroups != 0); 57 58 /* we need one copy of events per cpu for reading */ 59 map_size = total_cpus * evlist_size / nr_cgroups; 60 bpf_map__set_max_entries(sk->maps.events, map_size); 61 bpf_map__set_max_entries(sk->maps.cgrp_idx, nr_cgroups); 62 /* previous result is saved in a per-cpu array */ 63 map_size = evlist_size / nr_cgroups; 64 bpf_map__set_max_entries(sk->maps.prev_readings, map_size); 65 /* cgroup result needs all events (per-cpu) */ 66 map_size = evlist_size; 67 bpf_map__set_max_entries(sk->maps.cgrp_readings, map_size); 68 } 69 70 static void test_max_events_program_load(void) 71 { 72 #ifndef NDEBUG 73 /* 74 * Test that the program verifies with the maximum number of events. If 75 * this test fails unfortunately perf needs recompiling with a lower 76 * BPERF_CGROUP__MAX_EVENTS to avoid BPF verifier issues. 77 */ 78 int err, max_events = BPERF_CGROUP__MAX_EVENTS * nr_cgroups; 79 struct bperf_cgroup_bpf *test_skel = bperf_cgroup_bpf__open(); 80 81 if (!test_skel) { 82 pr_err("Failed to open cgroup skeleton\n"); 83 return; 84 } 85 setup_rodata(test_skel, max_events); 86 err = bperf_cgroup_bpf__load(test_skel); 87 if (err) { 88 pr_err("Failed to load cgroup skeleton with max events %d.\n", 89 BPERF_CGROUP__MAX_EVENTS); 90 } 91 bperf_cgroup_bpf__destroy(test_skel); 92 #endif 93 } 94 95 static int bperf_load_program(struct evlist *evlist) 96 { 97 struct bpf_link *link; 98 struct evsel *evsel; 99 struct cgroup *cgrp, *leader_cgrp; 100 unsigned int i; 101 struct perf_cpu cpu; 102 int total_cpus = cpu__max_cpu().cpu; 103 int map_fd, prog_fd, err; 104 105 set_max_rlimit(); 106 107 test_max_events_program_load(); 108 109 skel = bperf_cgroup_bpf__open(); 110 if (!skel) { 111 pr_err("Failed to open cgroup skeleton\n"); 112 return -1; 113 } 114 setup_rodata(skel, evlist->core.nr_entries); 115 116 err = bperf_cgroup_bpf__load(skel); 117 if (err) { 118 pr_err("Failed to load cgroup skeleton\n"); 119 goto out; 120 } 121 122 err = -1; 123 124 cgrp_switch = evsel__new(&cgrp_switch_attr); 125 if (evsel__open_per_cpu(cgrp_switch, evlist->core.all_cpus, -1) < 0) { 126 pr_err("Failed to open cgroup switches event\n"); 127 goto out; 128 } 129 130 perf_cpu_map__for_each_cpu(cpu, i, evlist->core.all_cpus) { 131 link = bpf_program__attach_perf_event(skel->progs.on_cgrp_switch, 132 FD(cgrp_switch, i)); 133 if (IS_ERR(link)) { 134 pr_err("Failed to attach cgroup program\n"); 135 err = PTR_ERR(link); 136 goto out; 137 } 138 } 139 140 /* 141 * Update cgrp_idx map from cgroup-id to event index. 142 */ 143 cgrp = NULL; 144 i = 0; 145 146 evlist__for_each_entry(evlist, evsel) { 147 if (cgrp == NULL || evsel->cgrp == leader_cgrp) { 148 unsigned int j; 149 150 leader_cgrp = evsel->cgrp; 151 evsel->cgrp = NULL; 152 153 /* open single copy of the events w/o cgroup */ 154 err = evsel__open_per_cpu(evsel, evsel->core.cpus, -1); 155 if (err == 0) 156 evsel->supported = true; 157 158 map_fd = bpf_map__fd(skel->maps.events); 159 perf_cpu_map__for_each_cpu(cpu, j, evsel->core.cpus) { 160 int fd = FD(evsel, j); 161 __u32 idx = evsel->core.idx * total_cpus + cpu.cpu; 162 163 bpf_map_update_elem(map_fd, &idx, &fd, BPF_ANY); 164 } 165 166 evsel->cgrp = leader_cgrp; 167 } 168 169 if (evsel->cgrp == cgrp) 170 continue; 171 172 cgrp = evsel->cgrp; 173 174 if (read_cgroup_id(cgrp) < 0) { 175 pr_debug("Failed to get cgroup id for %s\n", cgrp->name); 176 cgrp->id = 0; 177 } 178 179 map_fd = bpf_map__fd(skel->maps.cgrp_idx); 180 err = bpf_map_update_elem(map_fd, &cgrp->id, &i, BPF_ANY); 181 if (err < 0) { 182 pr_err("Failed to update cgroup index map\n"); 183 goto out; 184 } 185 186 i++; 187 } 188 189 /* 190 * bperf uses BPF_PROG_TEST_RUN to get accurate reading. Check 191 * whether the kernel support it 192 */ 193 prog_fd = bpf_program__fd(skel->progs.trigger_read); 194 err = bperf_trigger_reading(prog_fd, 0); 195 if (err) { 196 pr_warning("The kernel does not support test_run for raw_tp BPF programs.\n" 197 "Therefore, --for-each-cgroup might show inaccurate readings\n"); 198 err = 0; 199 } 200 201 out: 202 return err; 203 } 204 205 static int bperf_cgrp__load(struct evsel *evsel, 206 struct target *target __maybe_unused) 207 { 208 static bool bperf_loaded = false; 209 210 evsel->bperf_leader_prog_fd = -1; 211 evsel->bperf_leader_link_fd = -1; 212 213 if (!bperf_loaded && bperf_load_program(evsel->evlist)) 214 return -1; 215 216 bperf_loaded = true; 217 /* just to bypass bpf_counter_skip() */ 218 evsel->follower_skel = (struct bperf_follower_bpf *)skel; 219 220 return 0; 221 } 222 223 static int bperf_cgrp__install_pe(struct evsel *evsel __maybe_unused, 224 int cpu_map_idx __maybe_unused, 225 int fd __maybe_unused) 226 { 227 /* nothing to do */ 228 return 0; 229 } 230 231 /* 232 * trigger the leader prog on each cpu, so the cgrp_reading map could get 233 * the latest results. 234 */ 235 static int bperf_cgrp__sync_counters(struct evlist *evlist) 236 { 237 struct perf_cpu cpu; 238 unsigned int idx; 239 int prog_fd = bpf_program__fd(skel->progs.trigger_read); 240 241 perf_cpu_map__for_each_cpu(cpu, idx, evlist->core.all_cpus) 242 bperf_trigger_reading(prog_fd, cpu.cpu); 243 244 return 0; 245 } 246 247 static int bperf_cgrp__enable(struct evsel *evsel) 248 { 249 if (evsel->core.idx) 250 return 0; 251 252 bperf_cgrp__sync_counters(evsel->evlist); 253 254 skel->bss->enabled = 1; 255 return 0; 256 } 257 258 static int bperf_cgrp__disable(struct evsel *evsel) 259 { 260 if (evsel->core.idx) 261 return 0; 262 263 bperf_cgrp__sync_counters(evsel->evlist); 264 265 skel->bss->enabled = 0; 266 return 0; 267 } 268 269 static int bperf_cgrp__read(struct evsel *evsel) 270 { 271 struct evlist *evlist = evsel->evlist; 272 int total_cpus = cpu__max_cpu().cpu; 273 struct perf_counts_values *counts; 274 struct bpf_perf_event_value *values; 275 int reading_map_fd, err = 0; 276 277 if (evsel->core.idx) 278 return 0; 279 280 bperf_cgrp__sync_counters(evsel->evlist); 281 282 values = calloc(total_cpus, sizeof(*values)); 283 if (values == NULL) 284 return -ENOMEM; 285 286 reading_map_fd = bpf_map__fd(skel->maps.cgrp_readings); 287 288 evlist__for_each_entry(evlist, evsel) { 289 __u32 idx = evsel->core.idx; 290 unsigned int i; 291 struct perf_cpu cpu; 292 293 err = bpf_map_lookup_elem(reading_map_fd, &idx, values); 294 if (err) { 295 pr_err("bpf map lookup failed: idx=%u, event=%s, cgrp=%s\n", 296 idx, evsel__name(evsel), evsel->cgrp->name); 297 goto out; 298 } 299 300 perf_cpu_map__for_each_cpu(cpu, i, evsel->core.cpus) { 301 counts = perf_counts(evsel->counts, i, 0); 302 counts->val = values[cpu.cpu].counter; 303 counts->ena = values[cpu.cpu].enabled; 304 counts->run = values[cpu.cpu].running; 305 } 306 } 307 308 out: 309 free(values); 310 return err; 311 } 312 313 static int bperf_cgrp__destroy(struct evsel *evsel) 314 { 315 if (evsel->core.idx) 316 return 0; 317 318 bperf_cgroup_bpf__destroy(skel); 319 evsel__delete(cgrp_switch); // it'll destroy on_switch progs too 320 321 return 0; 322 } 323 324 struct bpf_counter_ops bperf_cgrp_ops = { 325 .load = bperf_cgrp__load, 326 .enable = bperf_cgrp__enable, 327 .disable = bperf_cgrp__disable, 328 .read = bperf_cgrp__read, 329 .install_pe = bperf_cgrp__install_pe, 330 .destroy = bperf_cgrp__destroy, 331 }; 332