1 // SPDX-License-Identifier: GPL-2.0 2 #include "util/cgroup.h" 3 #include "util/debug.h" 4 #include "util/evlist.h" 5 #include "util/machine.h" 6 #include "util/map.h" 7 #include "util/symbol.h" 8 #include "util/target.h" 9 #include "util/thread.h" 10 #include "util/thread_map.h" 11 #include "util/lock-contention.h" 12 #include <linux/zalloc.h> 13 #include <linux/string.h> 14 #include <bpf/bpf.h> 15 #include <inttypes.h> 16 17 #include "bpf_skel/lock_contention.skel.h" 18 #include "bpf_skel/lock_data.h" 19 20 static struct lock_contention_bpf *skel; 21 22 int lock_contention_prepare(struct lock_contention *con) 23 { 24 int i, fd; 25 int ncpus = 1, ntasks = 1, ntypes = 1, naddrs = 1, ncgrps = 1; 26 struct evlist *evlist = con->evlist; 27 struct target *target = con->target; 28 29 skel = lock_contention_bpf__open(); 30 if (!skel) { 31 pr_err("Failed to open lock-contention BPF skeleton\n"); 32 return -1; 33 } 34 35 bpf_map__set_value_size(skel->maps.stacks, con->max_stack * sizeof(u64)); 36 bpf_map__set_max_entries(skel->maps.lock_stat, con->map_nr_entries); 37 bpf_map__set_max_entries(skel->maps.tstamp, con->map_nr_entries); 38 39 if (con->aggr_mode == LOCK_AGGR_TASK) 40 bpf_map__set_max_entries(skel->maps.task_data, con->map_nr_entries); 41 else 42 bpf_map__set_max_entries(skel->maps.task_data, 1); 43 44 if (con->save_callstack) 45 bpf_map__set_max_entries(skel->maps.stacks, con->map_nr_entries); 46 else 47 bpf_map__set_max_entries(skel->maps.stacks, 1); 48 49 if (target__has_cpu(target)) { 50 skel->rodata->has_cpu = 1; 51 ncpus = perf_cpu_map__nr(evlist->core.user_requested_cpus); 52 } 53 if (target__has_task(target)) { 54 skel->rodata->has_task = 1; 55 ntasks = perf_thread_map__nr(evlist->core.threads); 56 } 57 if (con->filters->nr_types) { 58 skel->rodata->has_type = 1; 59 ntypes = con->filters->nr_types; 60 } 61 if (con->filters->nr_cgrps) { 62 skel->rodata->has_cgroup = 1; 63 ncgrps = con->filters->nr_cgrps; 64 } 65 66 /* resolve lock name filters to addr */ 67 if (con->filters->nr_syms) { 68 struct symbol *sym; 69 struct map *kmap; 70 unsigned long *addrs; 71 72 for (i = 0; i < con->filters->nr_syms; i++) { 73 sym = machine__find_kernel_symbol_by_name(con->machine, 74 con->filters->syms[i], 75 &kmap); 76 if (sym == NULL) { 77 pr_warning("ignore unknown symbol: %s\n", 78 con->filters->syms[i]); 79 continue; 80 } 81 82 addrs = realloc(con->filters->addrs, 83 (con->filters->nr_addrs + 1) * sizeof(*addrs)); 84 if (addrs == NULL) { 85 pr_warning("memory allocation failure\n"); 86 continue; 87 } 88 89 addrs[con->filters->nr_addrs++] = map__unmap_ip(kmap, sym->start); 90 con->filters->addrs = addrs; 91 } 92 naddrs = con->filters->nr_addrs; 93 skel->rodata->has_addr = 1; 94 } 95 96 bpf_map__set_max_entries(skel->maps.cpu_filter, ncpus); 97 bpf_map__set_max_entries(skel->maps.task_filter, ntasks); 98 bpf_map__set_max_entries(skel->maps.type_filter, ntypes); 99 bpf_map__set_max_entries(skel->maps.addr_filter, naddrs); 100 bpf_map__set_max_entries(skel->maps.cgroup_filter, ncgrps); 101 102 skel->rodata->stack_skip = con->stack_skip; 103 skel->rodata->aggr_mode = con->aggr_mode; 104 skel->rodata->needs_callstack = con->save_callstack; 105 skel->rodata->lock_owner = con->owner; 106 107 if (con->aggr_mode == LOCK_AGGR_CGROUP || con->filters->nr_cgrps) { 108 if (cgroup_is_v2("perf_event")) 109 skel->rodata->use_cgroup_v2 = 1; 110 } 111 112 if (lock_contention_bpf__load(skel) < 0) { 113 pr_err("Failed to load lock-contention BPF skeleton\n"); 114 return -1; 115 } 116 117 if (target__has_cpu(target)) { 118 u32 cpu; 119 u8 val = 1; 120 121 fd = bpf_map__fd(skel->maps.cpu_filter); 122 123 for (i = 0; i < ncpus; i++) { 124 cpu = perf_cpu_map__cpu(evlist->core.user_requested_cpus, i).cpu; 125 bpf_map_update_elem(fd, &cpu, &val, BPF_ANY); 126 } 127 } 128 129 if (target__has_task(target)) { 130 u32 pid; 131 u8 val = 1; 132 133 fd = bpf_map__fd(skel->maps.task_filter); 134 135 for (i = 0; i < ntasks; i++) { 136 pid = perf_thread_map__pid(evlist->core.threads, i); 137 bpf_map_update_elem(fd, &pid, &val, BPF_ANY); 138 } 139 } 140 141 if (target__none(target) && evlist->workload.pid > 0) { 142 u32 pid = evlist->workload.pid; 143 u8 val = 1; 144 145 fd = bpf_map__fd(skel->maps.task_filter); 146 bpf_map_update_elem(fd, &pid, &val, BPF_ANY); 147 } 148 149 if (con->filters->nr_types) { 150 u8 val = 1; 151 152 fd = bpf_map__fd(skel->maps.type_filter); 153 154 for (i = 0; i < con->filters->nr_types; i++) 155 bpf_map_update_elem(fd, &con->filters->types[i], &val, BPF_ANY); 156 } 157 158 if (con->filters->nr_addrs) { 159 u8 val = 1; 160 161 fd = bpf_map__fd(skel->maps.addr_filter); 162 163 for (i = 0; i < con->filters->nr_addrs; i++) 164 bpf_map_update_elem(fd, &con->filters->addrs[i], &val, BPF_ANY); 165 } 166 167 if (con->filters->nr_cgrps) { 168 u8 val = 1; 169 170 fd = bpf_map__fd(skel->maps.cgroup_filter); 171 172 for (i = 0; i < con->filters->nr_cgrps; i++) 173 bpf_map_update_elem(fd, &con->filters->cgrps[i], &val, BPF_ANY); 174 } 175 176 if (con->aggr_mode == LOCK_AGGR_CGROUP) 177 read_all_cgroups(&con->cgroups); 178 179 bpf_program__set_autoload(skel->progs.collect_lock_syms, false); 180 181 lock_contention_bpf__attach(skel); 182 return 0; 183 } 184 185 /* 186 * Run the BPF program directly using BPF_PROG_TEST_RUN to update the end 187 * timestamp in ktime so that it can calculate delta easily. 188 */ 189 static void mark_end_timestamp(void) 190 { 191 DECLARE_LIBBPF_OPTS(bpf_test_run_opts, opts, 192 .flags = BPF_F_TEST_RUN_ON_CPU, 193 ); 194 int prog_fd = bpf_program__fd(skel->progs.end_timestamp); 195 196 bpf_prog_test_run_opts(prog_fd, &opts); 197 } 198 199 static void update_lock_stat(int map_fd, int pid, u64 end_ts, 200 enum lock_aggr_mode aggr_mode, 201 struct tstamp_data *ts_data) 202 { 203 u64 delta; 204 struct contention_key stat_key = {}; 205 struct contention_data stat_data; 206 207 if (ts_data->timestamp >= end_ts) 208 return; 209 210 delta = end_ts - ts_data->timestamp; 211 212 switch (aggr_mode) { 213 case LOCK_AGGR_CALLER: 214 stat_key.stack_id = ts_data->stack_id; 215 break; 216 case LOCK_AGGR_TASK: 217 stat_key.pid = pid; 218 break; 219 case LOCK_AGGR_ADDR: 220 stat_key.lock_addr_or_cgroup = ts_data->lock; 221 break; 222 case LOCK_AGGR_CGROUP: 223 /* TODO */ 224 return; 225 default: 226 return; 227 } 228 229 if (bpf_map_lookup_elem(map_fd, &stat_key, &stat_data) < 0) 230 return; 231 232 stat_data.total_time += delta; 233 stat_data.count++; 234 235 if (delta > stat_data.max_time) 236 stat_data.max_time = delta; 237 if (delta < stat_data.min_time) 238 stat_data.min_time = delta; 239 240 bpf_map_update_elem(map_fd, &stat_key, &stat_data, BPF_EXIST); 241 } 242 243 /* 244 * Account entries in the tstamp map (which didn't see the corresponding 245 * lock:contention_end tracepoint) using end_ts. 246 */ 247 static void account_end_timestamp(struct lock_contention *con) 248 { 249 int ts_fd, stat_fd; 250 int *prev_key, key; 251 u64 end_ts = skel->bss->end_ts; 252 int total_cpus; 253 enum lock_aggr_mode aggr_mode = con->aggr_mode; 254 struct tstamp_data ts_data, *cpu_data; 255 256 /* Iterate per-task tstamp map (key = TID) */ 257 ts_fd = bpf_map__fd(skel->maps.tstamp); 258 stat_fd = bpf_map__fd(skel->maps.lock_stat); 259 260 prev_key = NULL; 261 while (!bpf_map_get_next_key(ts_fd, prev_key, &key)) { 262 if (bpf_map_lookup_elem(ts_fd, &key, &ts_data) == 0) { 263 int pid = key; 264 265 if (aggr_mode == LOCK_AGGR_TASK && con->owner) 266 pid = ts_data.flags; 267 268 update_lock_stat(stat_fd, pid, end_ts, aggr_mode, 269 &ts_data); 270 } 271 272 prev_key = &key; 273 } 274 275 /* Now it'll check per-cpu tstamp map which doesn't have TID. */ 276 if (aggr_mode == LOCK_AGGR_TASK || aggr_mode == LOCK_AGGR_CGROUP) 277 return; 278 279 total_cpus = cpu__max_cpu().cpu; 280 ts_fd = bpf_map__fd(skel->maps.tstamp_cpu); 281 282 cpu_data = calloc(total_cpus, sizeof(*cpu_data)); 283 if (cpu_data == NULL) 284 return; 285 286 prev_key = NULL; 287 while (!bpf_map_get_next_key(ts_fd, prev_key, &key)) { 288 if (bpf_map_lookup_elem(ts_fd, &key, cpu_data) < 0) 289 goto next; 290 291 for (int i = 0; i < total_cpus; i++) { 292 if (cpu_data[i].lock == 0) 293 continue; 294 295 update_lock_stat(stat_fd, -1, end_ts, aggr_mode, 296 &cpu_data[i]); 297 } 298 299 next: 300 prev_key = &key; 301 } 302 free(cpu_data); 303 } 304 305 int lock_contention_start(void) 306 { 307 skel->bss->enabled = 1; 308 return 0; 309 } 310 311 int lock_contention_stop(void) 312 { 313 skel->bss->enabled = 0; 314 mark_end_timestamp(); 315 return 0; 316 } 317 318 static const char *lock_contention_get_name(struct lock_contention *con, 319 struct contention_key *key, 320 u64 *stack_trace, u32 flags) 321 { 322 int idx = 0; 323 u64 addr; 324 const char *name = ""; 325 static char name_buf[KSYM_NAME_LEN]; 326 struct symbol *sym; 327 struct map *kmap; 328 struct machine *machine = con->machine; 329 330 if (con->aggr_mode == LOCK_AGGR_TASK) { 331 struct contention_task_data task; 332 int pid = key->pid; 333 int task_fd = bpf_map__fd(skel->maps.task_data); 334 335 /* do not update idle comm which contains CPU number */ 336 if (pid) { 337 struct thread *t = machine__findnew_thread(machine, /*pid=*/-1, pid); 338 339 if (t == NULL) 340 return name; 341 if (!bpf_map_lookup_elem(task_fd, &pid, &task) && 342 thread__set_comm(t, task.comm, /*timestamp=*/0)) 343 name = task.comm; 344 } 345 return name; 346 } 347 348 if (con->aggr_mode == LOCK_AGGR_ADDR) { 349 int lock_fd = bpf_map__fd(skel->maps.lock_syms); 350 351 /* per-process locks set upper bits of the flags */ 352 if (flags & LCD_F_MMAP_LOCK) 353 return "mmap_lock"; 354 if (flags & LCD_F_SIGHAND_LOCK) 355 return "siglock"; 356 357 /* global locks with symbols */ 358 sym = machine__find_kernel_symbol(machine, key->lock_addr_or_cgroup, &kmap); 359 if (sym) 360 return sym->name; 361 362 /* try semi-global locks collected separately */ 363 if (!bpf_map_lookup_elem(lock_fd, &key->lock_addr_or_cgroup, &flags)) { 364 if (flags == LOCK_CLASS_RQLOCK) 365 return "rq_lock"; 366 } 367 368 return ""; 369 } 370 371 if (con->aggr_mode == LOCK_AGGR_CGROUP) { 372 u64 cgrp_id = key->lock_addr_or_cgroup; 373 struct cgroup *cgrp = __cgroup__find(&con->cgroups, cgrp_id); 374 375 if (cgrp) 376 return cgrp->name; 377 378 snprintf(name_buf, sizeof(name_buf), "cgroup:%" PRIu64 "", cgrp_id); 379 return name_buf; 380 } 381 382 /* LOCK_AGGR_CALLER: skip lock internal functions */ 383 while (machine__is_lock_function(machine, stack_trace[idx]) && 384 idx < con->max_stack - 1) 385 idx++; 386 387 addr = stack_trace[idx]; 388 sym = machine__find_kernel_symbol(machine, addr, &kmap); 389 390 if (sym) { 391 unsigned long offset; 392 393 offset = map__map_ip(kmap, addr) - sym->start; 394 395 if (offset == 0) 396 return sym->name; 397 398 snprintf(name_buf, sizeof(name_buf), "%s+%#lx", sym->name, offset); 399 } else { 400 snprintf(name_buf, sizeof(name_buf), "%#lx", (unsigned long)addr); 401 } 402 403 return name_buf; 404 } 405 406 int lock_contention_read(struct lock_contention *con) 407 { 408 int fd, stack, err = 0; 409 struct contention_key *prev_key, key = {}; 410 struct contention_data data = {}; 411 struct lock_stat *st = NULL; 412 struct machine *machine = con->machine; 413 u64 *stack_trace; 414 size_t stack_size = con->max_stack * sizeof(*stack_trace); 415 416 fd = bpf_map__fd(skel->maps.lock_stat); 417 stack = bpf_map__fd(skel->maps.stacks); 418 419 con->fails.task = skel->bss->task_fail; 420 con->fails.stack = skel->bss->stack_fail; 421 con->fails.time = skel->bss->time_fail; 422 con->fails.data = skel->bss->data_fail; 423 424 stack_trace = zalloc(stack_size); 425 if (stack_trace == NULL) 426 return -1; 427 428 account_end_timestamp(con); 429 430 if (con->aggr_mode == LOCK_AGGR_TASK) { 431 struct thread *idle = machine__findnew_thread(machine, 432 /*pid=*/0, 433 /*tid=*/0); 434 thread__set_comm(idle, "swapper", /*timestamp=*/0); 435 } 436 437 if (con->aggr_mode == LOCK_AGGR_ADDR) { 438 DECLARE_LIBBPF_OPTS(bpf_test_run_opts, opts, 439 .flags = BPF_F_TEST_RUN_ON_CPU, 440 ); 441 int prog_fd = bpf_program__fd(skel->progs.collect_lock_syms); 442 443 bpf_prog_test_run_opts(prog_fd, &opts); 444 } 445 446 /* make sure it loads the kernel map */ 447 maps__load_first(machine->kmaps); 448 449 prev_key = NULL; 450 while (!bpf_map_get_next_key(fd, prev_key, &key)) { 451 s64 ls_key; 452 const char *name; 453 454 /* to handle errors in the loop body */ 455 err = -1; 456 457 bpf_map_lookup_elem(fd, &key, &data); 458 if (con->save_callstack) { 459 bpf_map_lookup_elem(stack, &key.stack_id, stack_trace); 460 461 if (!match_callstack_filter(machine, stack_trace)) { 462 con->nr_filtered += data.count; 463 goto next; 464 } 465 } 466 467 switch (con->aggr_mode) { 468 case LOCK_AGGR_CALLER: 469 ls_key = key.stack_id; 470 break; 471 case LOCK_AGGR_TASK: 472 ls_key = key.pid; 473 break; 474 case LOCK_AGGR_ADDR: 475 case LOCK_AGGR_CGROUP: 476 ls_key = key.lock_addr_or_cgroup; 477 break; 478 default: 479 goto next; 480 } 481 482 st = lock_stat_find(ls_key); 483 if (st != NULL) { 484 st->wait_time_total += data.total_time; 485 if (st->wait_time_max < data.max_time) 486 st->wait_time_max = data.max_time; 487 if (st->wait_time_min > data.min_time) 488 st->wait_time_min = data.min_time; 489 490 st->nr_contended += data.count; 491 if (st->nr_contended) 492 st->avg_wait_time = st->wait_time_total / st->nr_contended; 493 goto next; 494 } 495 496 name = lock_contention_get_name(con, &key, stack_trace, data.flags); 497 st = lock_stat_findnew(ls_key, name, data.flags); 498 if (st == NULL) 499 break; 500 501 st->nr_contended = data.count; 502 st->wait_time_total = data.total_time; 503 st->wait_time_max = data.max_time; 504 st->wait_time_min = data.min_time; 505 506 if (data.count) 507 st->avg_wait_time = data.total_time / data.count; 508 509 if (con->aggr_mode == LOCK_AGGR_CALLER && verbose > 0) { 510 st->callstack = memdup(stack_trace, stack_size); 511 if (st->callstack == NULL) 512 break; 513 } 514 515 next: 516 prev_key = &key; 517 518 /* we're fine now, reset the error */ 519 err = 0; 520 } 521 522 free(stack_trace); 523 524 return err; 525 } 526 527 int lock_contention_finish(struct lock_contention *con) 528 { 529 if (skel) { 530 skel->bss->enabled = 0; 531 lock_contention_bpf__destroy(skel); 532 } 533 534 while (!RB_EMPTY_ROOT(&con->cgroups)) { 535 struct rb_node *node = rb_first(&con->cgroups); 536 struct cgroup *cgrp = rb_entry(node, struct cgroup, node); 537 538 rb_erase(node, &con->cgroups); 539 cgroup__put(cgrp); 540 } 541 542 return 0; 543 } 544