Adrian Hunter | d0170af | 2015-07-17 19:33:43 +0300 | [diff] [blame] | 1 | /* |
| 2 | * intel-bts.c: Intel Processor Trace support |
| 3 | * Copyright (c) 2013-2015, Intel Corporation. |
| 4 | * |
| 5 | * This program is free software; you can redistribute it and/or modify it |
| 6 | * under the terms and conditions of the GNU General Public License, |
| 7 | * version 2, as published by the Free Software Foundation. |
| 8 | * |
| 9 | * This program is distributed in the hope it will be useful, but WITHOUT |
| 10 | * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or |
| 11 | * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for |
| 12 | * more details. |
| 13 | * |
| 14 | */ |
| 15 | |
| 16 | #include <endian.h> |
| 17 | #include <byteswap.h> |
| 18 | #include <linux/kernel.h> |
| 19 | #include <linux/types.h> |
| 20 | #include <linux/bitops.h> |
| 21 | #include <linux/log2.h> |
| 22 | |
| 23 | #include "cpumap.h" |
| 24 | #include "color.h" |
| 25 | #include "evsel.h" |
| 26 | #include "evlist.h" |
| 27 | #include "machine.h" |
| 28 | #include "session.h" |
| 29 | #include "util.h" |
| 30 | #include "thread.h" |
| 31 | #include "thread-stack.h" |
| 32 | #include "debug.h" |
| 33 | #include "tsc.h" |
| 34 | #include "auxtrace.h" |
| 35 | #include "intel-pt-decoder/intel-pt-insn-decoder.h" |
| 36 | #include "intel-bts.h" |
| 37 | |
| 38 | #define MAX_TIMESTAMP (~0ULL) |
| 39 | |
| 40 | #define INTEL_BTS_ERR_NOINSN 5 |
| 41 | #define INTEL_BTS_ERR_LOST 9 |
| 42 | |
| 43 | #if __BYTE_ORDER == __BIG_ENDIAN |
| 44 | #define le64_to_cpu bswap_64 |
| 45 | #else |
| 46 | #define le64_to_cpu |
| 47 | #endif |
| 48 | |
| 49 | struct intel_bts { |
| 50 | struct auxtrace auxtrace; |
| 51 | struct auxtrace_queues queues; |
| 52 | struct auxtrace_heap heap; |
| 53 | u32 auxtrace_type; |
| 54 | struct perf_session *session; |
| 55 | struct machine *machine; |
| 56 | bool sampling_mode; |
| 57 | bool snapshot_mode; |
| 58 | bool data_queued; |
| 59 | u32 pmu_type; |
| 60 | struct perf_tsc_conversion tc; |
| 61 | bool cap_user_time_zero; |
| 62 | struct itrace_synth_opts synth_opts; |
| 63 | bool sample_branches; |
| 64 | u32 branches_filter; |
| 65 | u64 branches_sample_type; |
| 66 | u64 branches_id; |
| 67 | size_t branches_event_size; |
| 68 | bool synth_needs_swap; |
| 69 | }; |
| 70 | |
| 71 | struct intel_bts_queue { |
| 72 | struct intel_bts *bts; |
| 73 | unsigned int queue_nr; |
| 74 | struct auxtrace_buffer *buffer; |
| 75 | bool on_heap; |
| 76 | bool done; |
| 77 | pid_t pid; |
| 78 | pid_t tid; |
| 79 | int cpu; |
| 80 | u64 time; |
| 81 | struct intel_pt_insn intel_pt_insn; |
| 82 | u32 sample_flags; |
| 83 | }; |
| 84 | |
| 85 | struct branch { |
| 86 | u64 from; |
| 87 | u64 to; |
| 88 | u64 misc; |
| 89 | }; |
| 90 | |
| 91 | static void intel_bts_dump(struct intel_bts *bts __maybe_unused, |
| 92 | unsigned char *buf, size_t len) |
| 93 | { |
| 94 | struct branch *branch; |
| 95 | size_t i, pos = 0, br_sz = sizeof(struct branch), sz; |
| 96 | const char *color = PERF_COLOR_BLUE; |
| 97 | |
| 98 | color_fprintf(stdout, color, |
| 99 | ". ... Intel BTS data: size %zu bytes\n", |
| 100 | len); |
| 101 | |
| 102 | while (len) { |
| 103 | if (len >= br_sz) |
| 104 | sz = br_sz; |
| 105 | else |
| 106 | sz = len; |
| 107 | printf("."); |
| 108 | color_fprintf(stdout, color, " %08x: ", pos); |
| 109 | for (i = 0; i < sz; i++) |
| 110 | color_fprintf(stdout, color, " %02x", buf[i]); |
| 111 | for (; i < br_sz; i++) |
| 112 | color_fprintf(stdout, color, " "); |
| 113 | if (len >= br_sz) { |
| 114 | branch = (struct branch *)buf; |
| 115 | color_fprintf(stdout, color, " %"PRIx64" -> %"PRIx64" %s\n", |
| 116 | le64_to_cpu(branch->from), |
| 117 | le64_to_cpu(branch->to), |
| 118 | le64_to_cpu(branch->misc) & 0x10 ? |
| 119 | "pred" : "miss"); |
| 120 | } else { |
| 121 | color_fprintf(stdout, color, " Bad record!\n"); |
| 122 | } |
| 123 | pos += sz; |
| 124 | buf += sz; |
| 125 | len -= sz; |
| 126 | } |
| 127 | } |
| 128 | |
| 129 | static void intel_bts_dump_event(struct intel_bts *bts, unsigned char *buf, |
| 130 | size_t len) |
| 131 | { |
| 132 | printf(".\n"); |
| 133 | intel_bts_dump(bts, buf, len); |
| 134 | } |
| 135 | |
| 136 | static int intel_bts_lost(struct intel_bts *bts, struct perf_sample *sample) |
| 137 | { |
| 138 | union perf_event event; |
| 139 | int err; |
| 140 | |
| 141 | auxtrace_synth_error(&event.auxtrace_error, PERF_AUXTRACE_ERROR_ITRACE, |
| 142 | INTEL_BTS_ERR_LOST, sample->cpu, sample->pid, |
| 143 | sample->tid, 0, "Lost trace data"); |
| 144 | |
| 145 | err = perf_session__deliver_synth_event(bts->session, &event, NULL); |
| 146 | if (err) |
| 147 | pr_err("Intel BTS: failed to deliver error event, error %d\n", |
| 148 | err); |
| 149 | |
| 150 | return err; |
| 151 | } |
| 152 | |
| 153 | static struct intel_bts_queue *intel_bts_alloc_queue(struct intel_bts *bts, |
| 154 | unsigned int queue_nr) |
| 155 | { |
| 156 | struct intel_bts_queue *btsq; |
| 157 | |
| 158 | btsq = zalloc(sizeof(struct intel_bts_queue)); |
| 159 | if (!btsq) |
| 160 | return NULL; |
| 161 | |
| 162 | btsq->bts = bts; |
| 163 | btsq->queue_nr = queue_nr; |
| 164 | btsq->pid = -1; |
| 165 | btsq->tid = -1; |
| 166 | btsq->cpu = -1; |
| 167 | |
| 168 | return btsq; |
| 169 | } |
| 170 | |
| 171 | static int intel_bts_setup_queue(struct intel_bts *bts, |
| 172 | struct auxtrace_queue *queue, |
| 173 | unsigned int queue_nr) |
| 174 | { |
| 175 | struct intel_bts_queue *btsq = queue->priv; |
| 176 | |
| 177 | if (list_empty(&queue->head)) |
| 178 | return 0; |
| 179 | |
| 180 | if (!btsq) { |
| 181 | btsq = intel_bts_alloc_queue(bts, queue_nr); |
| 182 | if (!btsq) |
| 183 | return -ENOMEM; |
| 184 | queue->priv = btsq; |
| 185 | |
| 186 | if (queue->cpu != -1) |
| 187 | btsq->cpu = queue->cpu; |
| 188 | btsq->tid = queue->tid; |
| 189 | } |
| 190 | |
| 191 | if (bts->sampling_mode) |
| 192 | return 0; |
| 193 | |
| 194 | if (!btsq->on_heap && !btsq->buffer) { |
| 195 | int ret; |
| 196 | |
| 197 | btsq->buffer = auxtrace_buffer__next(queue, NULL); |
| 198 | if (!btsq->buffer) |
| 199 | return 0; |
| 200 | |
| 201 | ret = auxtrace_heap__add(&bts->heap, queue_nr, |
| 202 | btsq->buffer->reference); |
| 203 | if (ret) |
| 204 | return ret; |
| 205 | btsq->on_heap = true; |
| 206 | } |
| 207 | |
| 208 | return 0; |
| 209 | } |
| 210 | |
| 211 | static int intel_bts_setup_queues(struct intel_bts *bts) |
| 212 | { |
| 213 | unsigned int i; |
| 214 | int ret; |
| 215 | |
| 216 | for (i = 0; i < bts->queues.nr_queues; i++) { |
| 217 | ret = intel_bts_setup_queue(bts, &bts->queues.queue_array[i], |
| 218 | i); |
| 219 | if (ret) |
| 220 | return ret; |
| 221 | } |
| 222 | return 0; |
| 223 | } |
| 224 | |
| 225 | static inline int intel_bts_update_queues(struct intel_bts *bts) |
| 226 | { |
| 227 | if (bts->queues.new_data) { |
| 228 | bts->queues.new_data = false; |
| 229 | return intel_bts_setup_queues(bts); |
| 230 | } |
| 231 | return 0; |
| 232 | } |
| 233 | |
| 234 | static unsigned char *intel_bts_find_overlap(unsigned char *buf_a, size_t len_a, |
| 235 | unsigned char *buf_b, size_t len_b) |
| 236 | { |
| 237 | size_t offs, len; |
| 238 | |
| 239 | if (len_a > len_b) |
| 240 | offs = len_a - len_b; |
| 241 | else |
| 242 | offs = 0; |
| 243 | |
| 244 | for (; offs < len_a; offs += sizeof(struct branch)) { |
| 245 | len = len_a - offs; |
| 246 | if (!memcmp(buf_a + offs, buf_b, len)) |
| 247 | return buf_b + len; |
| 248 | } |
| 249 | |
| 250 | return buf_b; |
| 251 | } |
| 252 | |
| 253 | static int intel_bts_do_fix_overlap(struct auxtrace_queue *queue, |
| 254 | struct auxtrace_buffer *b) |
| 255 | { |
| 256 | struct auxtrace_buffer *a; |
| 257 | void *start; |
| 258 | |
| 259 | if (b->list.prev == &queue->head) |
| 260 | return 0; |
| 261 | a = list_entry(b->list.prev, struct auxtrace_buffer, list); |
| 262 | start = intel_bts_find_overlap(a->data, a->size, b->data, b->size); |
| 263 | if (!start) |
| 264 | return -EINVAL; |
| 265 | b->use_size = b->data + b->size - start; |
| 266 | b->use_data = start; |
| 267 | return 0; |
| 268 | } |
| 269 | |
| 270 | static int intel_bts_synth_branch_sample(struct intel_bts_queue *btsq, |
| 271 | struct branch *branch) |
| 272 | { |
| 273 | int ret; |
| 274 | struct intel_bts *bts = btsq->bts; |
| 275 | union perf_event event; |
| 276 | struct perf_sample sample = { .ip = 0, }; |
| 277 | |
| 278 | event.sample.header.type = PERF_RECORD_SAMPLE; |
| 279 | event.sample.header.misc = PERF_RECORD_MISC_USER; |
| 280 | event.sample.header.size = sizeof(struct perf_event_header); |
| 281 | |
| 282 | sample.ip = le64_to_cpu(branch->from); |
| 283 | sample.pid = btsq->pid; |
| 284 | sample.tid = btsq->tid; |
| 285 | sample.addr = le64_to_cpu(branch->to); |
| 286 | sample.id = btsq->bts->branches_id; |
| 287 | sample.stream_id = btsq->bts->branches_id; |
| 288 | sample.period = 1; |
| 289 | sample.cpu = btsq->cpu; |
| 290 | sample.flags = btsq->sample_flags; |
| 291 | sample.insn_len = btsq->intel_pt_insn.length; |
| 292 | |
| 293 | if (bts->synth_opts.inject) { |
| 294 | event.sample.header.size = bts->branches_event_size; |
| 295 | ret = perf_event__synthesize_sample(&event, |
| 296 | bts->branches_sample_type, |
| 297 | 0, &sample, |
| 298 | bts->synth_needs_swap); |
| 299 | if (ret) |
| 300 | return ret; |
| 301 | } |
| 302 | |
| 303 | ret = perf_session__deliver_synth_event(bts->session, &event, &sample); |
| 304 | if (ret) |
| 305 | pr_err("Intel BTS: failed to deliver branch event, error %d\n", |
| 306 | ret); |
| 307 | |
| 308 | return ret; |
| 309 | } |
| 310 | |
| 311 | static int intel_bts_get_next_insn(struct intel_bts_queue *btsq, u64 ip) |
| 312 | { |
| 313 | struct machine *machine = btsq->bts->machine; |
| 314 | struct thread *thread; |
| 315 | struct addr_location al; |
| 316 | unsigned char buf[1024]; |
| 317 | size_t bufsz; |
| 318 | ssize_t len; |
| 319 | int x86_64; |
| 320 | uint8_t cpumode; |
| 321 | int err = -1; |
| 322 | |
| 323 | bufsz = intel_pt_insn_max_size(); |
| 324 | |
| 325 | if (machine__kernel_ip(machine, ip)) |
| 326 | cpumode = PERF_RECORD_MISC_KERNEL; |
| 327 | else |
| 328 | cpumode = PERF_RECORD_MISC_USER; |
| 329 | |
| 330 | thread = machine__find_thread(machine, -1, btsq->tid); |
| 331 | if (!thread) |
| 332 | return -1; |
| 333 | |
| 334 | thread__find_addr_map(thread, cpumode, MAP__FUNCTION, ip, &al); |
| 335 | if (!al.map || !al.map->dso) |
| 336 | goto out_put; |
| 337 | |
| 338 | len = dso__data_read_addr(al.map->dso, al.map, machine, ip, buf, bufsz); |
| 339 | if (len <= 0) |
| 340 | goto out_put; |
| 341 | |
| 342 | /* Load maps to ensure dso->is_64_bit has been updated */ |
| 343 | map__load(al.map, machine->symbol_filter); |
| 344 | |
| 345 | x86_64 = al.map->dso->is_64_bit; |
| 346 | |
| 347 | if (intel_pt_get_insn(buf, len, x86_64, &btsq->intel_pt_insn)) |
| 348 | goto out_put; |
| 349 | |
| 350 | err = 0; |
| 351 | out_put: |
| 352 | thread__put(thread); |
| 353 | return err; |
| 354 | } |
| 355 | |
| 356 | static int intel_bts_synth_error(struct intel_bts *bts, int cpu, pid_t pid, |
| 357 | pid_t tid, u64 ip) |
| 358 | { |
| 359 | union perf_event event; |
| 360 | int err; |
| 361 | |
| 362 | auxtrace_synth_error(&event.auxtrace_error, PERF_AUXTRACE_ERROR_ITRACE, |
| 363 | INTEL_BTS_ERR_NOINSN, cpu, pid, tid, ip, |
| 364 | "Failed to get instruction"); |
| 365 | |
| 366 | err = perf_session__deliver_synth_event(bts->session, &event, NULL); |
| 367 | if (err) |
| 368 | pr_err("Intel BTS: failed to deliver error event, error %d\n", |
| 369 | err); |
| 370 | |
| 371 | return err; |
| 372 | } |
| 373 | |
| 374 | static int intel_bts_get_branch_type(struct intel_bts_queue *btsq, |
| 375 | struct branch *branch) |
| 376 | { |
| 377 | int err; |
| 378 | |
| 379 | if (!branch->from) { |
| 380 | if (branch->to) |
| 381 | btsq->sample_flags = PERF_IP_FLAG_BRANCH | |
| 382 | PERF_IP_FLAG_TRACE_BEGIN; |
| 383 | else |
| 384 | btsq->sample_flags = 0; |
| 385 | btsq->intel_pt_insn.length = 0; |
| 386 | } else if (!branch->to) { |
| 387 | btsq->sample_flags = PERF_IP_FLAG_BRANCH | |
| 388 | PERF_IP_FLAG_TRACE_END; |
| 389 | btsq->intel_pt_insn.length = 0; |
| 390 | } else { |
| 391 | err = intel_bts_get_next_insn(btsq, branch->from); |
| 392 | if (err) { |
| 393 | btsq->sample_flags = 0; |
| 394 | btsq->intel_pt_insn.length = 0; |
| 395 | if (!btsq->bts->synth_opts.errors) |
| 396 | return 0; |
| 397 | err = intel_bts_synth_error(btsq->bts, btsq->cpu, |
| 398 | btsq->pid, btsq->tid, |
| 399 | branch->from); |
| 400 | return err; |
| 401 | } |
| 402 | btsq->sample_flags = intel_pt_insn_type(btsq->intel_pt_insn.op); |
| 403 | /* Check for an async branch into the kernel */ |
| 404 | if (!machine__kernel_ip(btsq->bts->machine, branch->from) && |
| 405 | machine__kernel_ip(btsq->bts->machine, branch->to) && |
| 406 | btsq->sample_flags != (PERF_IP_FLAG_BRANCH | |
| 407 | PERF_IP_FLAG_CALL | |
| 408 | PERF_IP_FLAG_SYSCALLRET)) |
| 409 | btsq->sample_flags = PERF_IP_FLAG_BRANCH | |
| 410 | PERF_IP_FLAG_CALL | |
| 411 | PERF_IP_FLAG_ASYNC | |
| 412 | PERF_IP_FLAG_INTERRUPT; |
| 413 | } |
| 414 | |
| 415 | return 0; |
| 416 | } |
| 417 | |
| 418 | static int intel_bts_process_buffer(struct intel_bts_queue *btsq, |
| 419 | struct auxtrace_buffer *buffer) |
| 420 | { |
| 421 | struct branch *branch; |
| 422 | size_t sz, bsz = sizeof(struct branch); |
| 423 | u32 filter = btsq->bts->branches_filter; |
| 424 | int err = 0; |
| 425 | |
| 426 | if (buffer->use_data) { |
| 427 | sz = buffer->use_size; |
| 428 | branch = buffer->use_data; |
| 429 | } else { |
| 430 | sz = buffer->size; |
| 431 | branch = buffer->data; |
| 432 | } |
| 433 | |
| 434 | if (!btsq->bts->sample_branches) |
| 435 | return 0; |
| 436 | |
| 437 | for (; sz > bsz; branch += 1, sz -= bsz) { |
| 438 | if (!branch->from && !branch->to) |
| 439 | continue; |
| 440 | intel_bts_get_branch_type(btsq, branch); |
| 441 | if (filter && !(filter & btsq->sample_flags)) |
| 442 | continue; |
| 443 | err = intel_bts_synth_branch_sample(btsq, branch); |
| 444 | if (err) |
| 445 | break; |
| 446 | } |
| 447 | return err; |
| 448 | } |
| 449 | |
| 450 | static int intel_bts_process_queue(struct intel_bts_queue *btsq, u64 *timestamp) |
| 451 | { |
| 452 | struct auxtrace_buffer *buffer = btsq->buffer, *old_buffer = buffer; |
| 453 | struct auxtrace_queue *queue; |
| 454 | struct thread *thread; |
| 455 | int err; |
| 456 | |
| 457 | if (btsq->done) |
| 458 | return 1; |
| 459 | |
| 460 | if (btsq->pid == -1) { |
| 461 | thread = machine__find_thread(btsq->bts->machine, -1, |
| 462 | btsq->tid); |
| 463 | if (thread) |
| 464 | btsq->pid = thread->pid_; |
| 465 | } else { |
| 466 | thread = machine__findnew_thread(btsq->bts->machine, btsq->pid, |
| 467 | btsq->tid); |
| 468 | } |
| 469 | |
| 470 | queue = &btsq->bts->queues.queue_array[btsq->queue_nr]; |
| 471 | |
| 472 | if (!buffer) |
| 473 | buffer = auxtrace_buffer__next(queue, NULL); |
| 474 | |
| 475 | if (!buffer) { |
| 476 | if (!btsq->bts->sampling_mode) |
| 477 | btsq->done = 1; |
| 478 | err = 1; |
| 479 | goto out_put; |
| 480 | } |
| 481 | |
| 482 | /* Currently there is no support for split buffers */ |
| 483 | if (buffer->consecutive) { |
| 484 | err = -EINVAL; |
| 485 | goto out_put; |
| 486 | } |
| 487 | |
| 488 | if (!buffer->data) { |
| 489 | int fd = perf_data_file__fd(btsq->bts->session->file); |
| 490 | |
| 491 | buffer->data = auxtrace_buffer__get_data(buffer, fd); |
| 492 | if (!buffer->data) { |
| 493 | err = -ENOMEM; |
| 494 | goto out_put; |
| 495 | } |
| 496 | } |
| 497 | |
| 498 | if (btsq->bts->snapshot_mode && !buffer->consecutive && |
| 499 | intel_bts_do_fix_overlap(queue, buffer)) { |
| 500 | err = -ENOMEM; |
| 501 | goto out_put; |
| 502 | } |
| 503 | |
| 504 | if (!btsq->bts->synth_opts.callchain && thread && |
| 505 | (!old_buffer || btsq->bts->sampling_mode || |
| 506 | (btsq->bts->snapshot_mode && !buffer->consecutive))) |
| 507 | thread_stack__set_trace_nr(thread, buffer->buffer_nr + 1); |
| 508 | |
| 509 | err = intel_bts_process_buffer(btsq, buffer); |
| 510 | |
| 511 | auxtrace_buffer__drop_data(buffer); |
| 512 | |
| 513 | btsq->buffer = auxtrace_buffer__next(queue, buffer); |
| 514 | if (btsq->buffer) { |
| 515 | if (timestamp) |
| 516 | *timestamp = btsq->buffer->reference; |
| 517 | } else { |
| 518 | if (!btsq->bts->sampling_mode) |
| 519 | btsq->done = 1; |
| 520 | } |
| 521 | out_put: |
| 522 | thread__put(thread); |
| 523 | return err; |
| 524 | } |
| 525 | |
| 526 | static int intel_bts_flush_queue(struct intel_bts_queue *btsq) |
| 527 | { |
| 528 | u64 ts = 0; |
| 529 | int ret; |
| 530 | |
| 531 | while (1) { |
| 532 | ret = intel_bts_process_queue(btsq, &ts); |
| 533 | if (ret < 0) |
| 534 | return ret; |
| 535 | if (ret) |
| 536 | break; |
| 537 | } |
| 538 | return 0; |
| 539 | } |
| 540 | |
| 541 | static int intel_bts_process_tid_exit(struct intel_bts *bts, pid_t tid) |
| 542 | { |
| 543 | struct auxtrace_queues *queues = &bts->queues; |
| 544 | unsigned int i; |
| 545 | |
| 546 | for (i = 0; i < queues->nr_queues; i++) { |
| 547 | struct auxtrace_queue *queue = &bts->queues.queue_array[i]; |
| 548 | struct intel_bts_queue *btsq = queue->priv; |
| 549 | |
| 550 | if (btsq && btsq->tid == tid) |
| 551 | return intel_bts_flush_queue(btsq); |
| 552 | } |
| 553 | return 0; |
| 554 | } |
| 555 | |
| 556 | static int intel_bts_process_queues(struct intel_bts *bts, u64 timestamp) |
| 557 | { |
| 558 | while (1) { |
| 559 | unsigned int queue_nr; |
| 560 | struct auxtrace_queue *queue; |
| 561 | struct intel_bts_queue *btsq; |
| 562 | u64 ts = 0; |
| 563 | int ret; |
| 564 | |
| 565 | if (!bts->heap.heap_cnt) |
| 566 | return 0; |
| 567 | |
| 568 | if (bts->heap.heap_array[0].ordinal > timestamp) |
| 569 | return 0; |
| 570 | |
| 571 | queue_nr = bts->heap.heap_array[0].queue_nr; |
| 572 | queue = &bts->queues.queue_array[queue_nr]; |
| 573 | btsq = queue->priv; |
| 574 | |
| 575 | auxtrace_heap__pop(&bts->heap); |
| 576 | |
| 577 | ret = intel_bts_process_queue(btsq, &ts); |
| 578 | if (ret < 0) { |
| 579 | auxtrace_heap__add(&bts->heap, queue_nr, ts); |
| 580 | return ret; |
| 581 | } |
| 582 | |
| 583 | if (!ret) { |
| 584 | ret = auxtrace_heap__add(&bts->heap, queue_nr, ts); |
| 585 | if (ret < 0) |
| 586 | return ret; |
| 587 | } else { |
| 588 | btsq->on_heap = false; |
| 589 | } |
| 590 | } |
| 591 | |
| 592 | return 0; |
| 593 | } |
| 594 | |
| 595 | static int intel_bts_process_event(struct perf_session *session, |
| 596 | union perf_event *event, |
| 597 | struct perf_sample *sample, |
| 598 | struct perf_tool *tool) |
| 599 | { |
| 600 | struct intel_bts *bts = container_of(session->auxtrace, struct intel_bts, |
| 601 | auxtrace); |
| 602 | u64 timestamp; |
| 603 | int err; |
| 604 | |
| 605 | if (dump_trace) |
| 606 | return 0; |
| 607 | |
| 608 | if (!tool->ordered_events) { |
| 609 | pr_err("Intel BTS requires ordered events\n"); |
| 610 | return -EINVAL; |
| 611 | } |
| 612 | |
| 613 | if (sample->time && sample->time != (u64)-1) |
| 614 | timestamp = perf_time_to_tsc(sample->time, &bts->tc); |
| 615 | else |
| 616 | timestamp = 0; |
| 617 | |
| 618 | err = intel_bts_update_queues(bts); |
| 619 | if (err) |
| 620 | return err; |
| 621 | |
| 622 | err = intel_bts_process_queues(bts, timestamp); |
| 623 | if (err) |
| 624 | return err; |
| 625 | if (event->header.type == PERF_RECORD_EXIT) { |
Adrian Hunter | 53ff6bc | 2015-08-18 12:07:05 +0300 | [diff] [blame] | 626 | err = intel_bts_process_tid_exit(bts, event->fork.tid); |
Adrian Hunter | d0170af | 2015-07-17 19:33:43 +0300 | [diff] [blame] | 627 | if (err) |
| 628 | return err; |
| 629 | } |
| 630 | |
| 631 | if (event->header.type == PERF_RECORD_AUX && |
| 632 | (event->aux.flags & PERF_AUX_FLAG_TRUNCATED) && |
| 633 | bts->synth_opts.errors) |
| 634 | err = intel_bts_lost(bts, sample); |
| 635 | |
| 636 | return err; |
| 637 | } |
| 638 | |
| 639 | static int intel_bts_process_auxtrace_event(struct perf_session *session, |
| 640 | union perf_event *event, |
| 641 | struct perf_tool *tool __maybe_unused) |
| 642 | { |
| 643 | struct intel_bts *bts = container_of(session->auxtrace, struct intel_bts, |
| 644 | auxtrace); |
| 645 | |
| 646 | if (bts->sampling_mode) |
| 647 | return 0; |
| 648 | |
| 649 | if (!bts->data_queued) { |
| 650 | struct auxtrace_buffer *buffer; |
| 651 | off_t data_offset; |
| 652 | int fd = perf_data_file__fd(session->file); |
| 653 | int err; |
| 654 | |
| 655 | if (perf_data_file__is_pipe(session->file)) { |
| 656 | data_offset = 0; |
| 657 | } else { |
| 658 | data_offset = lseek(fd, 0, SEEK_CUR); |
| 659 | if (data_offset == -1) |
| 660 | return -errno; |
| 661 | } |
| 662 | |
| 663 | err = auxtrace_queues__add_event(&bts->queues, session, event, |
| 664 | data_offset, &buffer); |
| 665 | if (err) |
| 666 | return err; |
| 667 | |
| 668 | /* Dump here now we have copied a piped trace out of the pipe */ |
| 669 | if (dump_trace) { |
| 670 | if (auxtrace_buffer__get_data(buffer, fd)) { |
| 671 | intel_bts_dump_event(bts, buffer->data, |
| 672 | buffer->size); |
| 673 | auxtrace_buffer__put_data(buffer); |
| 674 | } |
| 675 | } |
| 676 | } |
| 677 | |
| 678 | return 0; |
| 679 | } |
| 680 | |
| 681 | static int intel_bts_flush(struct perf_session *session __maybe_unused, |
| 682 | struct perf_tool *tool __maybe_unused) |
| 683 | { |
| 684 | struct intel_bts *bts = container_of(session->auxtrace, struct intel_bts, |
| 685 | auxtrace); |
| 686 | int ret; |
| 687 | |
| 688 | if (dump_trace || bts->sampling_mode) |
| 689 | return 0; |
| 690 | |
| 691 | if (!tool->ordered_events) |
| 692 | return -EINVAL; |
| 693 | |
| 694 | ret = intel_bts_update_queues(bts); |
| 695 | if (ret < 0) |
| 696 | return ret; |
| 697 | |
| 698 | return intel_bts_process_queues(bts, MAX_TIMESTAMP); |
| 699 | } |
| 700 | |
| 701 | static void intel_bts_free_queue(void *priv) |
| 702 | { |
| 703 | struct intel_bts_queue *btsq = priv; |
| 704 | |
| 705 | if (!btsq) |
| 706 | return; |
| 707 | free(btsq); |
| 708 | } |
| 709 | |
| 710 | static void intel_bts_free_events(struct perf_session *session) |
| 711 | { |
| 712 | struct intel_bts *bts = container_of(session->auxtrace, struct intel_bts, |
| 713 | auxtrace); |
| 714 | struct auxtrace_queues *queues = &bts->queues; |
| 715 | unsigned int i; |
| 716 | |
| 717 | for (i = 0; i < queues->nr_queues; i++) { |
| 718 | intel_bts_free_queue(queues->queue_array[i].priv); |
| 719 | queues->queue_array[i].priv = NULL; |
| 720 | } |
| 721 | auxtrace_queues__free(queues); |
| 722 | } |
| 723 | |
| 724 | static void intel_bts_free(struct perf_session *session) |
| 725 | { |
| 726 | struct intel_bts *bts = container_of(session->auxtrace, struct intel_bts, |
| 727 | auxtrace); |
| 728 | |
| 729 | auxtrace_heap__free(&bts->heap); |
| 730 | intel_bts_free_events(session); |
| 731 | session->auxtrace = NULL; |
| 732 | free(bts); |
| 733 | } |
| 734 | |
| 735 | struct intel_bts_synth { |
| 736 | struct perf_tool dummy_tool; |
| 737 | struct perf_session *session; |
| 738 | }; |
| 739 | |
| 740 | static int intel_bts_event_synth(struct perf_tool *tool, |
| 741 | union perf_event *event, |
| 742 | struct perf_sample *sample __maybe_unused, |
| 743 | struct machine *machine __maybe_unused) |
| 744 | { |
| 745 | struct intel_bts_synth *intel_bts_synth = |
| 746 | container_of(tool, struct intel_bts_synth, dummy_tool); |
| 747 | |
| 748 | return perf_session__deliver_synth_event(intel_bts_synth->session, |
| 749 | event, NULL); |
| 750 | } |
| 751 | |
| 752 | static int intel_bts_synth_event(struct perf_session *session, |
| 753 | struct perf_event_attr *attr, u64 id) |
| 754 | { |
| 755 | struct intel_bts_synth intel_bts_synth; |
| 756 | |
| 757 | memset(&intel_bts_synth, 0, sizeof(struct intel_bts_synth)); |
| 758 | intel_bts_synth.session = session; |
| 759 | |
| 760 | return perf_event__synthesize_attr(&intel_bts_synth.dummy_tool, attr, 1, |
| 761 | &id, intel_bts_event_synth); |
| 762 | } |
| 763 | |
| 764 | static int intel_bts_synth_events(struct intel_bts *bts, |
| 765 | struct perf_session *session) |
| 766 | { |
| 767 | struct perf_evlist *evlist = session->evlist; |
| 768 | struct perf_evsel *evsel; |
| 769 | struct perf_event_attr attr; |
| 770 | bool found = false; |
| 771 | u64 id; |
| 772 | int err; |
| 773 | |
| 774 | evlist__for_each(evlist, evsel) { |
| 775 | if (evsel->attr.type == bts->pmu_type && evsel->ids) { |
| 776 | found = true; |
| 777 | break; |
| 778 | } |
| 779 | } |
| 780 | |
| 781 | if (!found) { |
| 782 | pr_debug("There are no selected events with Intel BTS data\n"); |
| 783 | return 0; |
| 784 | } |
| 785 | |
| 786 | memset(&attr, 0, sizeof(struct perf_event_attr)); |
| 787 | attr.size = sizeof(struct perf_event_attr); |
| 788 | attr.type = PERF_TYPE_HARDWARE; |
| 789 | attr.sample_type = evsel->attr.sample_type & PERF_SAMPLE_MASK; |
| 790 | attr.sample_type |= PERF_SAMPLE_IP | PERF_SAMPLE_TID | |
| 791 | PERF_SAMPLE_PERIOD; |
| 792 | attr.sample_type &= ~(u64)PERF_SAMPLE_TIME; |
| 793 | attr.sample_type &= ~(u64)PERF_SAMPLE_CPU; |
| 794 | attr.exclude_user = evsel->attr.exclude_user; |
| 795 | attr.exclude_kernel = evsel->attr.exclude_kernel; |
| 796 | attr.exclude_hv = evsel->attr.exclude_hv; |
| 797 | attr.exclude_host = evsel->attr.exclude_host; |
| 798 | attr.exclude_guest = evsel->attr.exclude_guest; |
| 799 | attr.sample_id_all = evsel->attr.sample_id_all; |
| 800 | attr.read_format = evsel->attr.read_format; |
| 801 | |
| 802 | id = evsel->id[0] + 1000000000; |
| 803 | if (!id) |
| 804 | id = 1; |
| 805 | |
| 806 | if (bts->synth_opts.branches) { |
| 807 | attr.config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS; |
| 808 | attr.sample_period = 1; |
| 809 | attr.sample_type |= PERF_SAMPLE_ADDR; |
| 810 | pr_debug("Synthesizing 'branches' event with id %" PRIu64 " sample type %#" PRIx64 "\n", |
| 811 | id, (u64)attr.sample_type); |
| 812 | err = intel_bts_synth_event(session, &attr, id); |
| 813 | if (err) { |
| 814 | pr_err("%s: failed to synthesize 'branches' event type\n", |
| 815 | __func__); |
| 816 | return err; |
| 817 | } |
| 818 | bts->sample_branches = true; |
| 819 | bts->branches_sample_type = attr.sample_type; |
| 820 | bts->branches_id = id; |
| 821 | /* |
| 822 | * We only use sample types from PERF_SAMPLE_MASK so we can use |
| 823 | * __perf_evsel__sample_size() here. |
| 824 | */ |
| 825 | bts->branches_event_size = sizeof(struct sample_event) + |
| 826 | __perf_evsel__sample_size(attr.sample_type); |
| 827 | } |
| 828 | |
| 829 | bts->synth_needs_swap = evsel->needs_swap; |
| 830 | |
| 831 | return 0; |
| 832 | } |
| 833 | |
| 834 | static const char * const intel_bts_info_fmts[] = { |
| 835 | [INTEL_BTS_PMU_TYPE] = " PMU Type %"PRId64"\n", |
| 836 | [INTEL_BTS_TIME_SHIFT] = " Time Shift %"PRIu64"\n", |
| 837 | [INTEL_BTS_TIME_MULT] = " Time Muliplier %"PRIu64"\n", |
| 838 | [INTEL_BTS_TIME_ZERO] = " Time Zero %"PRIu64"\n", |
| 839 | [INTEL_BTS_CAP_USER_TIME_ZERO] = " Cap Time Zero %"PRId64"\n", |
| 840 | [INTEL_BTS_SNAPSHOT_MODE] = " Snapshot mode %"PRId64"\n", |
| 841 | }; |
| 842 | |
| 843 | static void intel_bts_print_info(u64 *arr, int start, int finish) |
| 844 | { |
| 845 | int i; |
| 846 | |
| 847 | if (!dump_trace) |
| 848 | return; |
| 849 | |
| 850 | for (i = start; i <= finish; i++) |
| 851 | fprintf(stdout, intel_bts_info_fmts[i], arr[i]); |
| 852 | } |
| 853 | |
| 854 | u64 intel_bts_auxtrace_info_priv[INTEL_BTS_AUXTRACE_PRIV_SIZE]; |
| 855 | |
| 856 | int intel_bts_process_auxtrace_info(union perf_event *event, |
| 857 | struct perf_session *session) |
| 858 | { |
| 859 | struct auxtrace_info_event *auxtrace_info = &event->auxtrace_info; |
| 860 | size_t min_sz = sizeof(u64) * INTEL_BTS_SNAPSHOT_MODE; |
| 861 | struct intel_bts *bts; |
| 862 | int err; |
| 863 | |
| 864 | if (auxtrace_info->header.size < sizeof(struct auxtrace_info_event) + |
| 865 | min_sz) |
| 866 | return -EINVAL; |
| 867 | |
| 868 | bts = zalloc(sizeof(struct intel_bts)); |
| 869 | if (!bts) |
| 870 | return -ENOMEM; |
| 871 | |
| 872 | err = auxtrace_queues__init(&bts->queues); |
| 873 | if (err) |
| 874 | goto err_free; |
| 875 | |
| 876 | bts->session = session; |
| 877 | bts->machine = &session->machines.host; /* No kvm support */ |
| 878 | bts->auxtrace_type = auxtrace_info->type; |
| 879 | bts->pmu_type = auxtrace_info->priv[INTEL_BTS_PMU_TYPE]; |
| 880 | bts->tc.time_shift = auxtrace_info->priv[INTEL_BTS_TIME_SHIFT]; |
| 881 | bts->tc.time_mult = auxtrace_info->priv[INTEL_BTS_TIME_MULT]; |
| 882 | bts->tc.time_zero = auxtrace_info->priv[INTEL_BTS_TIME_ZERO]; |
| 883 | bts->cap_user_time_zero = |
| 884 | auxtrace_info->priv[INTEL_BTS_CAP_USER_TIME_ZERO]; |
| 885 | bts->snapshot_mode = auxtrace_info->priv[INTEL_BTS_SNAPSHOT_MODE]; |
| 886 | |
| 887 | bts->sampling_mode = false; |
| 888 | |
| 889 | bts->auxtrace.process_event = intel_bts_process_event; |
| 890 | bts->auxtrace.process_auxtrace_event = intel_bts_process_auxtrace_event; |
| 891 | bts->auxtrace.flush_events = intel_bts_flush; |
| 892 | bts->auxtrace.free_events = intel_bts_free_events; |
| 893 | bts->auxtrace.free = intel_bts_free; |
| 894 | session->auxtrace = &bts->auxtrace; |
| 895 | |
| 896 | intel_bts_print_info(&auxtrace_info->priv[0], INTEL_BTS_PMU_TYPE, |
| 897 | INTEL_BTS_SNAPSHOT_MODE); |
| 898 | |
| 899 | if (dump_trace) |
| 900 | return 0; |
| 901 | |
| 902 | if (session->itrace_synth_opts && session->itrace_synth_opts->set) |
| 903 | bts->synth_opts = *session->itrace_synth_opts; |
| 904 | else |
| 905 | itrace_synth_opts__set_default(&bts->synth_opts); |
| 906 | |
| 907 | if (bts->synth_opts.calls) |
| 908 | bts->branches_filter |= PERF_IP_FLAG_CALL | PERF_IP_FLAG_ASYNC | |
| 909 | PERF_IP_FLAG_TRACE_END; |
| 910 | if (bts->synth_opts.returns) |
| 911 | bts->branches_filter |= PERF_IP_FLAG_RETURN | |
| 912 | PERF_IP_FLAG_TRACE_BEGIN; |
| 913 | |
| 914 | err = intel_bts_synth_events(bts, session); |
| 915 | if (err) |
| 916 | goto err_free_queues; |
| 917 | |
| 918 | err = auxtrace_queues__process_index(&bts->queues, session); |
| 919 | if (err) |
| 920 | goto err_free_queues; |
| 921 | |
| 922 | if (bts->queues.populated) |
| 923 | bts->data_queued = true; |
| 924 | |
| 925 | return 0; |
| 926 | |
| 927 | err_free_queues: |
| 928 | auxtrace_queues__free(&bts->queues); |
| 929 | session->auxtrace = NULL; |
| 930 | err_free: |
| 931 | free(bts); |
| 932 | return err; |
| 933 | } |