David Brazdil | 0f672f6 | 2019-12-10 10:32:29 +0000 | [diff] [blame] | 1 | // SPDX-License-Identifier: GPL-2.0 |
| 2 | #include <errno.h> |
| 3 | #include <stdlib.h> |
| 4 | #include <bpf/bpf.h> |
| 5 | #include <bpf/btf.h> |
| 6 | #include <bpf/libbpf.h> |
| 7 | #include <linux/btf.h> |
| 8 | #include <linux/err.h> |
| 9 | #include "bpf-event.h" |
| 10 | #include "debug.h" |
| 11 | #include "dso.h" |
| 12 | #include "symbol.h" |
| 13 | #include "machine.h" |
| 14 | #include "env.h" |
| 15 | #include "session.h" |
| 16 | #include "map.h" |
| 17 | #include "evlist.h" |
| 18 | #include "record.h" |
| 19 | #include "util/synthetic-events.h" |
| 20 | |
| 21 | #define ptr_to_u64(ptr) ((__u64)(unsigned long)(ptr)) |
| 22 | |
| 23 | static int snprintf_hex(char *buf, size_t size, unsigned char *data, size_t len) |
| 24 | { |
| 25 | int ret = 0; |
| 26 | size_t i; |
| 27 | |
| 28 | for (i = 0; i < len; i++) |
| 29 | ret += snprintf(buf + ret, size - ret, "%02x", data[i]); |
| 30 | return ret; |
| 31 | } |
| 32 | |
| 33 | static int machine__process_bpf_event_load(struct machine *machine, |
| 34 | union perf_event *event, |
| 35 | struct perf_sample *sample __maybe_unused) |
| 36 | { |
| 37 | struct bpf_prog_info_linear *info_linear; |
| 38 | struct bpf_prog_info_node *info_node; |
| 39 | struct perf_env *env = machine->env; |
| 40 | int id = event->bpf.id; |
| 41 | unsigned int i; |
| 42 | |
| 43 | /* perf-record, no need to handle bpf-event */ |
| 44 | if (env == NULL) |
| 45 | return 0; |
| 46 | |
| 47 | info_node = perf_env__find_bpf_prog_info(env, id); |
| 48 | if (!info_node) |
| 49 | return 0; |
| 50 | info_linear = info_node->info_linear; |
| 51 | |
| 52 | for (i = 0; i < info_linear->info.nr_jited_ksyms; i++) { |
| 53 | u64 *addrs = (u64 *)(uintptr_t)(info_linear->info.jited_ksyms); |
| 54 | u64 addr = addrs[i]; |
| 55 | struct map *map; |
| 56 | |
| 57 | map = map_groups__find(&machine->kmaps, addr); |
| 58 | |
| 59 | if (map) { |
| 60 | map->dso->binary_type = DSO_BINARY_TYPE__BPF_PROG_INFO; |
| 61 | map->dso->bpf_prog.id = id; |
| 62 | map->dso->bpf_prog.sub_id = i; |
| 63 | map->dso->bpf_prog.env = env; |
| 64 | } |
| 65 | } |
| 66 | return 0; |
| 67 | } |
| 68 | |
| 69 | int machine__process_bpf(struct machine *machine, union perf_event *event, |
| 70 | struct perf_sample *sample) |
| 71 | { |
| 72 | if (dump_trace) |
| 73 | perf_event__fprintf_bpf(event, stdout); |
| 74 | |
| 75 | switch (event->bpf.type) { |
| 76 | case PERF_BPF_EVENT_PROG_LOAD: |
| 77 | return machine__process_bpf_event_load(machine, event, sample); |
| 78 | |
| 79 | case PERF_BPF_EVENT_PROG_UNLOAD: |
| 80 | /* |
| 81 | * Do not free bpf_prog_info and btf of the program here, |
| 82 | * as annotation still need them. They will be freed at |
| 83 | * the end of the session. |
| 84 | */ |
| 85 | break; |
| 86 | default: |
| 87 | pr_debug("unexpected bpf event type of %d\n", event->bpf.type); |
| 88 | break; |
| 89 | } |
| 90 | return 0; |
| 91 | } |
| 92 | |
| 93 | static int perf_env__fetch_btf(struct perf_env *env, |
| 94 | u32 btf_id, |
| 95 | struct btf *btf) |
| 96 | { |
| 97 | struct btf_node *node; |
| 98 | u32 data_size; |
| 99 | const void *data; |
| 100 | |
| 101 | data = btf__get_raw_data(btf, &data_size); |
| 102 | |
| 103 | node = malloc(data_size + sizeof(struct btf_node)); |
| 104 | if (!node) |
| 105 | return -1; |
| 106 | |
| 107 | node->id = btf_id; |
| 108 | node->data_size = data_size; |
| 109 | memcpy(node->data, data, data_size); |
| 110 | |
| 111 | perf_env__insert_btf(env, node); |
| 112 | return 0; |
| 113 | } |
| 114 | |
| 115 | static int synthesize_bpf_prog_name(char *buf, int size, |
| 116 | struct bpf_prog_info *info, |
| 117 | struct btf *btf, |
| 118 | u32 sub_id) |
| 119 | { |
| 120 | u8 (*prog_tags)[BPF_TAG_SIZE] = (void *)(uintptr_t)(info->prog_tags); |
| 121 | void *func_infos = (void *)(uintptr_t)(info->func_info); |
| 122 | u32 sub_prog_cnt = info->nr_jited_ksyms; |
| 123 | const struct bpf_func_info *finfo; |
| 124 | const char *short_name = NULL; |
| 125 | const struct btf_type *t; |
| 126 | int name_len; |
| 127 | |
| 128 | name_len = snprintf(buf, size, "bpf_prog_"); |
| 129 | name_len += snprintf_hex(buf + name_len, size - name_len, |
| 130 | prog_tags[sub_id], BPF_TAG_SIZE); |
| 131 | if (btf) { |
| 132 | finfo = func_infos + sub_id * info->func_info_rec_size; |
| 133 | t = btf__type_by_id(btf, finfo->type_id); |
| 134 | short_name = btf__name_by_offset(btf, t->name_off); |
| 135 | } else if (sub_id == 0 && sub_prog_cnt == 1) { |
| 136 | /* no subprog */ |
| 137 | if (info->name[0]) |
| 138 | short_name = info->name; |
| 139 | } else |
| 140 | short_name = "F"; |
| 141 | if (short_name) |
| 142 | name_len += snprintf(buf + name_len, size - name_len, |
| 143 | "_%s", short_name); |
| 144 | return name_len; |
| 145 | } |
| 146 | |
| 147 | /* |
| 148 | * Synthesize PERF_RECORD_KSYMBOL and PERF_RECORD_BPF_EVENT for one bpf |
| 149 | * program. One PERF_RECORD_BPF_EVENT is generated for the program. And |
| 150 | * one PERF_RECORD_KSYMBOL is generated for each sub program. |
| 151 | * |
| 152 | * Returns: |
| 153 | * 0 for success; |
| 154 | * -1 for failures; |
| 155 | * -2 for lack of kernel support. |
| 156 | */ |
| 157 | static int perf_event__synthesize_one_bpf_prog(struct perf_session *session, |
| 158 | perf_event__handler_t process, |
| 159 | struct machine *machine, |
| 160 | int fd, |
| 161 | union perf_event *event, |
| 162 | struct record_opts *opts) |
| 163 | { |
| 164 | struct perf_record_ksymbol *ksymbol_event = &event->ksymbol; |
| 165 | struct perf_record_bpf_event *bpf_event = &event->bpf; |
| 166 | struct bpf_prog_info_linear *info_linear; |
| 167 | struct perf_tool *tool = session->tool; |
| 168 | struct bpf_prog_info_node *info_node; |
| 169 | struct bpf_prog_info *info; |
| 170 | struct btf *btf = NULL; |
| 171 | struct perf_env *env; |
| 172 | u32 sub_prog_cnt, i; |
| 173 | int err = 0; |
| 174 | u64 arrays; |
| 175 | |
| 176 | /* |
| 177 | * for perf-record and perf-report use header.env; |
| 178 | * otherwise, use global perf_env. |
| 179 | */ |
| 180 | env = session->data ? &session->header.env : &perf_env; |
| 181 | |
| 182 | arrays = 1UL << BPF_PROG_INFO_JITED_KSYMS; |
| 183 | arrays |= 1UL << BPF_PROG_INFO_JITED_FUNC_LENS; |
| 184 | arrays |= 1UL << BPF_PROG_INFO_FUNC_INFO; |
| 185 | arrays |= 1UL << BPF_PROG_INFO_PROG_TAGS; |
| 186 | arrays |= 1UL << BPF_PROG_INFO_JITED_INSNS; |
| 187 | arrays |= 1UL << BPF_PROG_INFO_LINE_INFO; |
| 188 | arrays |= 1UL << BPF_PROG_INFO_JITED_LINE_INFO; |
| 189 | |
| 190 | info_linear = bpf_program__get_prog_info_linear(fd, arrays); |
| 191 | if (IS_ERR_OR_NULL(info_linear)) { |
| 192 | info_linear = NULL; |
| 193 | pr_debug("%s: failed to get BPF program info. aborting\n", __func__); |
| 194 | return -1; |
| 195 | } |
| 196 | |
| 197 | if (info_linear->info_len < offsetof(struct bpf_prog_info, prog_tags)) { |
| 198 | pr_debug("%s: the kernel is too old, aborting\n", __func__); |
| 199 | return -2; |
| 200 | } |
| 201 | |
| 202 | info = &info_linear->info; |
| 203 | |
| 204 | /* number of ksyms, func_lengths, and tags should match */ |
| 205 | sub_prog_cnt = info->nr_jited_ksyms; |
| 206 | if (sub_prog_cnt != info->nr_prog_tags || |
| 207 | sub_prog_cnt != info->nr_jited_func_lens) |
| 208 | return -1; |
| 209 | |
| 210 | /* check BTF func info support */ |
| 211 | if (info->btf_id && info->nr_func_info && info->func_info_rec_size) { |
| 212 | /* btf func info number should be same as sub_prog_cnt */ |
| 213 | if (sub_prog_cnt != info->nr_func_info) { |
| 214 | pr_debug("%s: mismatch in BPF sub program count and BTF function info count, aborting\n", __func__); |
| 215 | err = -1; |
| 216 | goto out; |
| 217 | } |
| 218 | if (btf__get_from_id(info->btf_id, &btf)) { |
| 219 | pr_debug("%s: failed to get BTF of id %u, aborting\n", __func__, info->btf_id); |
| 220 | err = -1; |
| 221 | btf = NULL; |
| 222 | goto out; |
| 223 | } |
| 224 | perf_env__fetch_btf(env, info->btf_id, btf); |
| 225 | } |
| 226 | |
| 227 | /* Synthesize PERF_RECORD_KSYMBOL */ |
| 228 | for (i = 0; i < sub_prog_cnt; i++) { |
| 229 | __u32 *prog_lens = (__u32 *)(uintptr_t)(info->jited_func_lens); |
| 230 | __u64 *prog_addrs = (__u64 *)(uintptr_t)(info->jited_ksyms); |
| 231 | int name_len; |
| 232 | |
| 233 | *ksymbol_event = (struct perf_record_ksymbol) { |
| 234 | .header = { |
| 235 | .type = PERF_RECORD_KSYMBOL, |
| 236 | .size = offsetof(struct perf_record_ksymbol, name), |
| 237 | }, |
| 238 | .addr = prog_addrs[i], |
| 239 | .len = prog_lens[i], |
| 240 | .ksym_type = PERF_RECORD_KSYMBOL_TYPE_BPF, |
| 241 | .flags = 0, |
| 242 | }; |
| 243 | |
| 244 | name_len = synthesize_bpf_prog_name(ksymbol_event->name, |
| 245 | KSYM_NAME_LEN, info, btf, i); |
| 246 | ksymbol_event->header.size += PERF_ALIGN(name_len + 1, |
| 247 | sizeof(u64)); |
| 248 | |
| 249 | memset((void *)event + event->header.size, 0, machine->id_hdr_size); |
| 250 | event->header.size += machine->id_hdr_size; |
| 251 | err = perf_tool__process_synth_event(tool, event, |
| 252 | machine, process); |
| 253 | } |
| 254 | |
| 255 | if (!opts->no_bpf_event) { |
| 256 | /* Synthesize PERF_RECORD_BPF_EVENT */ |
| 257 | *bpf_event = (struct perf_record_bpf_event) { |
| 258 | .header = { |
| 259 | .type = PERF_RECORD_BPF_EVENT, |
| 260 | .size = sizeof(struct perf_record_bpf_event), |
| 261 | }, |
| 262 | .type = PERF_BPF_EVENT_PROG_LOAD, |
| 263 | .flags = 0, |
| 264 | .id = info->id, |
| 265 | }; |
| 266 | memcpy(bpf_event->tag, info->tag, BPF_TAG_SIZE); |
| 267 | memset((void *)event + event->header.size, 0, machine->id_hdr_size); |
| 268 | event->header.size += machine->id_hdr_size; |
| 269 | |
| 270 | /* save bpf_prog_info to env */ |
| 271 | info_node = malloc(sizeof(struct bpf_prog_info_node)); |
| 272 | if (!info_node) { |
| 273 | err = -1; |
| 274 | goto out; |
| 275 | } |
| 276 | |
| 277 | info_node->info_linear = info_linear; |
| 278 | perf_env__insert_bpf_prog_info(env, info_node); |
| 279 | info_linear = NULL; |
| 280 | |
| 281 | /* |
| 282 | * process after saving bpf_prog_info to env, so that |
| 283 | * required information is ready for look up |
| 284 | */ |
| 285 | err = perf_tool__process_synth_event(tool, event, |
| 286 | machine, process); |
| 287 | } |
| 288 | |
| 289 | out: |
| 290 | free(info_linear); |
| 291 | free(btf); |
| 292 | return err ? -1 : 0; |
| 293 | } |
| 294 | |
| 295 | int perf_event__synthesize_bpf_events(struct perf_session *session, |
| 296 | perf_event__handler_t process, |
| 297 | struct machine *machine, |
| 298 | struct record_opts *opts) |
| 299 | { |
| 300 | union perf_event *event; |
| 301 | __u32 id = 0; |
| 302 | int err; |
| 303 | int fd; |
| 304 | |
| 305 | event = malloc(sizeof(event->bpf) + KSYM_NAME_LEN + machine->id_hdr_size); |
| 306 | if (!event) |
| 307 | return -1; |
| 308 | while (true) { |
| 309 | err = bpf_prog_get_next_id(id, &id); |
| 310 | if (err) { |
| 311 | if (errno == ENOENT) { |
| 312 | err = 0; |
| 313 | break; |
| 314 | } |
| 315 | pr_debug("%s: can't get next program: %s%s\n", |
| 316 | __func__, strerror(errno), |
| 317 | errno == EINVAL ? " -- kernel too old?" : ""); |
| 318 | /* don't report error on old kernel or EPERM */ |
| 319 | err = (errno == EINVAL || errno == EPERM) ? 0 : -1; |
| 320 | break; |
| 321 | } |
| 322 | fd = bpf_prog_get_fd_by_id(id); |
| 323 | if (fd < 0) { |
| 324 | pr_debug("%s: failed to get fd for prog_id %u\n", |
| 325 | __func__, id); |
| 326 | continue; |
| 327 | } |
| 328 | |
| 329 | err = perf_event__synthesize_one_bpf_prog(session, process, |
| 330 | machine, fd, |
| 331 | event, opts); |
| 332 | close(fd); |
| 333 | if (err) { |
| 334 | /* do not return error for old kernel */ |
| 335 | if (err == -2) |
| 336 | err = 0; |
| 337 | break; |
| 338 | } |
| 339 | } |
| 340 | free(event); |
| 341 | return err; |
| 342 | } |
| 343 | |
| 344 | static void perf_env__add_bpf_info(struct perf_env *env, u32 id) |
| 345 | { |
| 346 | struct bpf_prog_info_linear *info_linear; |
| 347 | struct bpf_prog_info_node *info_node; |
| 348 | struct btf *btf = NULL; |
| 349 | u64 arrays; |
| 350 | u32 btf_id; |
| 351 | int fd; |
| 352 | |
| 353 | fd = bpf_prog_get_fd_by_id(id); |
| 354 | if (fd < 0) |
| 355 | return; |
| 356 | |
| 357 | arrays = 1UL << BPF_PROG_INFO_JITED_KSYMS; |
| 358 | arrays |= 1UL << BPF_PROG_INFO_JITED_FUNC_LENS; |
| 359 | arrays |= 1UL << BPF_PROG_INFO_FUNC_INFO; |
| 360 | arrays |= 1UL << BPF_PROG_INFO_PROG_TAGS; |
| 361 | arrays |= 1UL << BPF_PROG_INFO_JITED_INSNS; |
| 362 | arrays |= 1UL << BPF_PROG_INFO_LINE_INFO; |
| 363 | arrays |= 1UL << BPF_PROG_INFO_JITED_LINE_INFO; |
| 364 | |
| 365 | info_linear = bpf_program__get_prog_info_linear(fd, arrays); |
| 366 | if (IS_ERR_OR_NULL(info_linear)) { |
| 367 | pr_debug("%s: failed to get BPF program info. aborting\n", __func__); |
| 368 | goto out; |
| 369 | } |
| 370 | |
| 371 | btf_id = info_linear->info.btf_id; |
| 372 | |
| 373 | info_node = malloc(sizeof(struct bpf_prog_info_node)); |
| 374 | if (info_node) { |
| 375 | info_node->info_linear = info_linear; |
| 376 | perf_env__insert_bpf_prog_info(env, info_node); |
| 377 | } else |
| 378 | free(info_linear); |
| 379 | |
| 380 | if (btf_id == 0) |
| 381 | goto out; |
| 382 | |
| 383 | if (btf__get_from_id(btf_id, &btf)) { |
| 384 | pr_debug("%s: failed to get BTF of id %u, aborting\n", |
| 385 | __func__, btf_id); |
| 386 | goto out; |
| 387 | } |
| 388 | perf_env__fetch_btf(env, btf_id, btf); |
| 389 | |
| 390 | out: |
| 391 | free(btf); |
| 392 | close(fd); |
| 393 | } |
| 394 | |
| 395 | static int bpf_event__sb_cb(union perf_event *event, void *data) |
| 396 | { |
| 397 | struct perf_env *env = data; |
| 398 | |
| 399 | if (event->header.type != PERF_RECORD_BPF_EVENT) |
| 400 | return -1; |
| 401 | |
| 402 | switch (event->bpf.type) { |
| 403 | case PERF_BPF_EVENT_PROG_LOAD: |
| 404 | perf_env__add_bpf_info(env, event->bpf.id); |
| 405 | |
| 406 | case PERF_BPF_EVENT_PROG_UNLOAD: |
| 407 | /* |
| 408 | * Do not free bpf_prog_info and btf of the program here, |
| 409 | * as annotation still need them. They will be freed at |
| 410 | * the end of the session. |
| 411 | */ |
| 412 | break; |
| 413 | default: |
| 414 | pr_debug("unexpected bpf event type of %d\n", event->bpf.type); |
| 415 | break; |
| 416 | } |
| 417 | |
| 418 | return 0; |
| 419 | } |
| 420 | |
| 421 | int bpf_event__add_sb_event(struct evlist **evlist, |
| 422 | struct perf_env *env) |
| 423 | { |
| 424 | struct perf_event_attr attr = { |
| 425 | .type = PERF_TYPE_SOFTWARE, |
| 426 | .config = PERF_COUNT_SW_DUMMY, |
| 427 | .sample_id_all = 1, |
| 428 | .watermark = 1, |
| 429 | .bpf_event = 1, |
| 430 | .size = sizeof(attr), /* to capture ABI version */ |
| 431 | }; |
| 432 | |
| 433 | /* |
| 434 | * Older gcc versions don't support designated initializers, like above, |
| 435 | * for unnamed union members, such as the following: |
| 436 | */ |
| 437 | attr.wakeup_watermark = 1; |
| 438 | |
| 439 | return perf_evlist__add_sb_event(evlist, &attr, bpf_event__sb_cb, env); |
| 440 | } |
| 441 | |
| 442 | void bpf_event__print_bpf_prog_info(struct bpf_prog_info *info, |
| 443 | struct perf_env *env, |
| 444 | FILE *fp) |
| 445 | { |
| 446 | __u32 *prog_lens = (__u32 *)(uintptr_t)(info->jited_func_lens); |
| 447 | __u64 *prog_addrs = (__u64 *)(uintptr_t)(info->jited_ksyms); |
| 448 | char name[KSYM_NAME_LEN]; |
| 449 | struct btf *btf = NULL; |
| 450 | u32 sub_prog_cnt, i; |
| 451 | |
| 452 | sub_prog_cnt = info->nr_jited_ksyms; |
| 453 | if (sub_prog_cnt != info->nr_prog_tags || |
| 454 | sub_prog_cnt != info->nr_jited_func_lens) |
| 455 | return; |
| 456 | |
| 457 | if (info->btf_id) { |
| 458 | struct btf_node *node; |
| 459 | |
| 460 | node = perf_env__find_btf(env, info->btf_id); |
| 461 | if (node) |
| 462 | btf = btf__new((__u8 *)(node->data), |
| 463 | node->data_size); |
| 464 | } |
| 465 | |
| 466 | if (sub_prog_cnt == 1) { |
| 467 | synthesize_bpf_prog_name(name, KSYM_NAME_LEN, info, btf, 0); |
| 468 | fprintf(fp, "# bpf_prog_info %u: %s addr 0x%llx size %u\n", |
| 469 | info->id, name, prog_addrs[0], prog_lens[0]); |
| 470 | return; |
| 471 | } |
| 472 | |
| 473 | fprintf(fp, "# bpf_prog_info %u:\n", info->id); |
| 474 | for (i = 0; i < sub_prog_cnt; i++) { |
| 475 | synthesize_bpf_prog_name(name, KSYM_NAME_LEN, info, btf, i); |
| 476 | |
| 477 | fprintf(fp, "# \tsub_prog %u: %s addr 0x%llx size %u\n", |
| 478 | i, name, prog_addrs[i], prog_lens[i]); |
| 479 | } |
| 480 | } |