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 16 #include "bpf_skel/lock_contention.skel.h" 17 #include "bpf_skel/lock_data.h" 18 19 static struct lock_contention_bpf *skel; 20 21 int lock_contention_prepare(struct lock_contention *con) 22 { 23 int i, fd; 24 int ncpus = 1, ntasks = 1, ntypes = 1, naddrs = 1, ncgrps = 1; 25 struct evlist *evlist = con->evlist; 26 struct target *target = con->target; 27 28 skel = lock_contention_bpf__open(); 29 if (!skel) { 30 pr_err("Failed to open lock-contention BPF skeleton\n"); 31 return -1; 32 } 33 34 bpf_map__set_value_size(skel->maps.stacks, con->max_stack * sizeof(u64)); 35 bpf_map__set_max_entries(skel->maps.lock_stat, con->map_nr_entries); 36 bpf_map__set_max_entries(skel->maps.tstamp, con->map_nr_entries); 37 38 if (con->aggr_mode == LOCK_AGGR_TASK) 39 bpf_map__set_max_entries(skel->maps.task_data, con->map_nr_entries); 40 else 41 bpf_map__set_max_entries(skel->maps.task_data, 1); 42 43 if (con->save_callstack) 44 bpf_map__set_max_entries(skel->maps.stacks, con->map_nr_entries); 45 else 46 bpf_map__set_max_entries(skel->maps.stacks, 1); 47 48 if (target__has_cpu(target)) 49 ncpus = perf_cpu_map__nr(evlist->core.user_requested_cpus); 50 if (target__has_task(target)) 51 ntasks = perf_thread_map__nr(evlist->core.threads); 52 if (con->filters->nr_types) 53 ntypes = con->filters->nr_types; 54 if (con->filters->nr_cgrps) 55 ncgrps = con->filters->nr_cgrps; 56 57 /* resolve lock name filters to addr */ 58 if (con->filters->nr_syms) { 59 struct symbol *sym; 60 struct map *kmap; 61 unsigned long *addrs; 62 63 for (i = 0; i < con->filters->nr_syms; i++) { 64 sym = machine__find_kernel_symbol_by_name(con->machine, 65 con->filters->syms[i], 66 &kmap); 67 if (sym == NULL) { 68 pr_warning("ignore unknown symbol: %s\n", 69 con->filters->syms[i]); 70 continue; 71 } 72 73 addrs = realloc(con->filters->addrs, 74 (con->filters->nr_addrs + 1) * sizeof(*addrs)); 75 if (addrs == NULL) { 76 pr_warning("memory allocation failure\n"); 77 continue; 78 } 79 80 addrs[con->filters->nr_addrs++] = map__unmap_ip(kmap, sym->start); 81 con->filters->addrs = addrs; 82 } 83 naddrs = con->filters->nr_addrs; 84 } 85 86 bpf_map__set_max_entries(skel->maps.cpu_filter, ncpus); 87 bpf_map__set_max_entries(skel->maps.task_filter, ntasks); 88 bpf_map__set_max_entries(skel->maps.type_filter, ntypes); 89 bpf_map__set_max_entries(skel->maps.addr_filter, naddrs); 90 bpf_map__set_max_entries(skel->maps.cgroup_filter, ncgrps); 91 92 if (lock_contention_bpf__load(skel) < 0) { 93 pr_err("Failed to load lock-contention BPF skeleton\n"); 94 return -1; 95 } 96 97 if (target__has_cpu(target)) { 98 u32 cpu; 99 u8 val = 1; 100 101 skel->bss->has_cpu = 1; 102 fd = bpf_map__fd(skel->maps.cpu_filter); 103 104 for (i = 0; i < ncpus; i++) { 105 cpu = perf_cpu_map__cpu(evlist->core.user_requested_cpus, i).cpu; 106 bpf_map_update_elem(fd, &cpu, &val, BPF_ANY); 107 } 108 } 109 110 if (target__has_task(target)) { 111 u32 pid; 112 u8 val = 1; 113 114 skel->bss->has_task = 1; 115 fd = bpf_map__fd(skel->maps.task_filter); 116 117 for (i = 0; i < ntasks; i++) { 118 pid = perf_thread_map__pid(evlist->core.threads, i); 119 bpf_map_update_elem(fd, &pid, &val, BPF_ANY); 120 } 121 } 122 123 if (target__none(target) && evlist->workload.pid > 0) { 124 u32 pid = evlist->workload.pid; 125 u8 val = 1; 126 127 skel->bss->has_task = 1; 128 fd = bpf_map__fd(skel->maps.task_filter); 129 bpf_map_update_elem(fd, &pid, &val, BPF_ANY); 130 } 131 132 if (con->filters->nr_types) { 133 u8 val = 1; 134 135 skel->bss->has_type = 1; 136 fd = bpf_map__fd(skel->maps.type_filter); 137 138 for (i = 0; i < con->filters->nr_types; i++) 139 bpf_map_update_elem(fd, &con->filters->types[i], &val, BPF_ANY); 140 } 141 142 if (con->filters->nr_addrs) { 143 u8 val = 1; 144 145 skel->bss->has_addr = 1; 146 fd = bpf_map__fd(skel->maps.addr_filter); 147 148 for (i = 0; i < con->filters->nr_addrs; i++) 149 bpf_map_update_elem(fd, &con->filters->addrs[i], &val, BPF_ANY); 150 } 151 152 if (con->filters->nr_cgrps) { 153 u8 val = 1; 154 155 skel->bss->has_cgroup = 1; 156 fd = bpf_map__fd(skel->maps.cgroup_filter); 157 158 for (i = 0; i < con->filters->nr_cgrps; i++) 159 bpf_map_update_elem(fd, &con->filters->cgrps[i], &val, BPF_ANY); 160 } 161 162 /* these don't work well if in the rodata section */ 163 skel->bss->stack_skip = con->stack_skip; 164 skel->bss->aggr_mode = con->aggr_mode; 165 skel->bss->needs_callstack = con->save_callstack; 166 skel->bss->lock_owner = con->owner; 167 168 if (con->aggr_mode == LOCK_AGGR_CGROUP) { 169 if (cgroup_is_v2("perf_event")) 170 skel->bss->use_cgroup_v2 = 1; 171 172 read_all_cgroups(&con->cgroups); 173 } 174 175 bpf_program__set_autoload(skel->progs.collect_lock_syms, false); 176 177 lock_contention_bpf__attach(skel); 178 return 0; 179 } 180 181 int lock_contention_start(void) 182 { 183 skel->bss->enabled = 1; 184 return 0; 185 } 186 187 int lock_contention_stop(void) 188 { 189 skel->bss->enabled = 0; 190 return 0; 191 } 192 193 static const char *lock_contention_get_name(struct lock_contention *con, 194 struct contention_key *key, 195 u64 *stack_trace, u32 flags) 196 { 197 int idx = 0; 198 u64 addr; 199 const char *name = ""; 200 static char name_buf[KSYM_NAME_LEN]; 201 struct symbol *sym; 202 struct map *kmap; 203 struct machine *machine = con->machine; 204 205 if (con->aggr_mode == LOCK_AGGR_TASK) { 206 struct contention_task_data task; 207 int pid = key->pid; 208 int task_fd = bpf_map__fd(skel->maps.task_data); 209 210 /* do not update idle comm which contains CPU number */ 211 if (pid) { 212 struct thread *t = __machine__findnew_thread(machine, /*pid=*/-1, pid); 213 214 if (t == NULL) 215 return name; 216 if (!bpf_map_lookup_elem(task_fd, &pid, &task) && 217 thread__set_comm(t, task.comm, /*timestamp=*/0)) 218 name = task.comm; 219 } 220 return name; 221 } 222 223 if (con->aggr_mode == LOCK_AGGR_ADDR) { 224 int lock_fd = bpf_map__fd(skel->maps.lock_syms); 225 226 /* per-process locks set upper bits of the flags */ 227 if (flags & LCD_F_MMAP_LOCK) 228 return "mmap_lock"; 229 if (flags & LCD_F_SIGHAND_LOCK) 230 return "siglock"; 231 232 /* global locks with symbols */ 233 sym = machine__find_kernel_symbol(machine, key->lock_addr_or_cgroup, &kmap); 234 if (sym) 235 return sym->name; 236 237 /* try semi-global locks collected separately */ 238 if (!bpf_map_lookup_elem(lock_fd, &key->lock_addr_or_cgroup, &flags)) { 239 if (flags == LOCK_CLASS_RQLOCK) 240 return "rq_lock"; 241 } 242 243 return ""; 244 } 245 246 if (con->aggr_mode == LOCK_AGGR_CGROUP) { 247 u64 cgrp_id = key->lock_addr_or_cgroup; 248 struct cgroup *cgrp = __cgroup__find(&con->cgroups, cgrp_id); 249 250 if (cgrp) 251 return cgrp->name; 252 253 snprintf(name_buf, sizeof(name_buf), "cgroup:%lu", cgrp_id); 254 return name_buf; 255 } 256 257 /* LOCK_AGGR_CALLER: skip lock internal functions */ 258 while (machine__is_lock_function(machine, stack_trace[idx]) && 259 idx < con->max_stack - 1) 260 idx++; 261 262 addr = stack_trace[idx]; 263 sym = machine__find_kernel_symbol(machine, addr, &kmap); 264 265 if (sym) { 266 unsigned long offset; 267 268 offset = map__map_ip(kmap, addr) - sym->start; 269 270 if (offset == 0) 271 return sym->name; 272 273 snprintf(name_buf, sizeof(name_buf), "%s+%#lx", sym->name, offset); 274 } else { 275 snprintf(name_buf, sizeof(name_buf), "%#lx", (unsigned long)addr); 276 } 277 278 return name_buf; 279 } 280 281 int lock_contention_read(struct lock_contention *con) 282 { 283 int fd, stack, err = 0; 284 struct contention_key *prev_key, key = {}; 285 struct contention_data data = {}; 286 struct lock_stat *st = NULL; 287 struct machine *machine = con->machine; 288 u64 *stack_trace; 289 size_t stack_size = con->max_stack * sizeof(*stack_trace); 290 291 fd = bpf_map__fd(skel->maps.lock_stat); 292 stack = bpf_map__fd(skel->maps.stacks); 293 294 con->fails.task = skel->bss->task_fail; 295 con->fails.stack = skel->bss->stack_fail; 296 con->fails.time = skel->bss->time_fail; 297 con->fails.data = skel->bss->data_fail; 298 299 stack_trace = zalloc(stack_size); 300 if (stack_trace == NULL) 301 return -1; 302 303 if (con->aggr_mode == LOCK_AGGR_TASK) { 304 struct thread *idle = __machine__findnew_thread(machine, 305 /*pid=*/0, 306 /*tid=*/0); 307 thread__set_comm(idle, "swapper", /*timestamp=*/0); 308 } 309 310 if (con->aggr_mode == LOCK_AGGR_ADDR) { 311 DECLARE_LIBBPF_OPTS(bpf_test_run_opts, opts, 312 .flags = BPF_F_TEST_RUN_ON_CPU, 313 ); 314 int prog_fd = bpf_program__fd(skel->progs.collect_lock_syms); 315 316 bpf_prog_test_run_opts(prog_fd, &opts); 317 } 318 319 /* make sure it loads the kernel map */ 320 map__load(maps__first(machine->kmaps)->map); 321 322 prev_key = NULL; 323 while (!bpf_map_get_next_key(fd, prev_key, &key)) { 324 s64 ls_key; 325 const char *name; 326 327 /* to handle errors in the loop body */ 328 err = -1; 329 330 bpf_map_lookup_elem(fd, &key, &data); 331 if (con->save_callstack) { 332 bpf_map_lookup_elem(stack, &key.stack_id, stack_trace); 333 334 if (!match_callstack_filter(machine, stack_trace)) { 335 con->nr_filtered += data.count; 336 goto next; 337 } 338 } 339 340 switch (con->aggr_mode) { 341 case LOCK_AGGR_CALLER: 342 ls_key = key.stack_id; 343 break; 344 case LOCK_AGGR_TASK: 345 ls_key = key.pid; 346 break; 347 case LOCK_AGGR_ADDR: 348 case LOCK_AGGR_CGROUP: 349 ls_key = key.lock_addr_or_cgroup; 350 break; 351 default: 352 goto next; 353 } 354 355 st = lock_stat_find(ls_key); 356 if (st != NULL) { 357 st->wait_time_total += data.total_time; 358 if (st->wait_time_max < data.max_time) 359 st->wait_time_max = data.max_time; 360 if (st->wait_time_min > data.min_time) 361 st->wait_time_min = data.min_time; 362 363 st->nr_contended += data.count; 364 if (st->nr_contended) 365 st->avg_wait_time = st->wait_time_total / st->nr_contended; 366 goto next; 367 } 368 369 name = lock_contention_get_name(con, &key, stack_trace, data.flags); 370 st = lock_stat_findnew(ls_key, name, data.flags); 371 if (st == NULL) 372 break; 373 374 st->nr_contended = data.count; 375 st->wait_time_total = data.total_time; 376 st->wait_time_max = data.max_time; 377 st->wait_time_min = data.min_time; 378 379 if (data.count) 380 st->avg_wait_time = data.total_time / data.count; 381 382 if (con->aggr_mode == LOCK_AGGR_CALLER && verbose > 0) { 383 st->callstack = memdup(stack_trace, stack_size); 384 if (st->callstack == NULL) 385 break; 386 } 387 388 next: 389 prev_key = &key; 390 391 /* we're fine now, reset the error */ 392 err = 0; 393 } 394 395 free(stack_trace); 396 397 return err; 398 } 399 400 int lock_contention_finish(struct lock_contention *con) 401 { 402 if (skel) { 403 skel->bss->enabled = 0; 404 lock_contention_bpf__destroy(skel); 405 } 406 407 while (!RB_EMPTY_ROOT(&con->cgroups)) { 408 struct rb_node *node = rb_first(&con->cgroups); 409 struct cgroup *cgrp = rb_entry(node, struct cgroup, node); 410 411 rb_erase(node, &con->cgroups); 412 cgroup__put(cgrp); 413 } 414 415 return 0; 416 } 417