1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Copyright(C) 2015-2018 Linaro Limited. 4 * 5 * Author: Tor Jeremiassen <tor@ti.com> 6 * Author: Mathieu Poirier <mathieu.poirier@linaro.org> 7 */ 8 9 #include <linux/bitops.h> 10 #include <linux/coresight-pmu.h> 11 #include <linux/err.h> 12 #include <linux/kernel.h> 13 #include <linux/log2.h> 14 #include <linux/types.h> 15 #include <linux/zalloc.h> 16 17 #include <opencsd/ocsd_if_types.h> 18 #include <stdlib.h> 19 20 #include "auxtrace.h" 21 #include "color.h" 22 #include "cs-etm.h" 23 #include "cs-etm-decoder/cs-etm-decoder.h" 24 #include "debug.h" 25 #include "dso.h" 26 #include "evlist.h" 27 #include "intlist.h" 28 #include "machine.h" 29 #include "map.h" 30 #include "perf.h" 31 #include "session.h" 32 #include "map_symbol.h" 33 #include "branch.h" 34 #include "symbol.h" 35 #include "tool.h" 36 #include "thread.h" 37 #include "thread-stack.h" 38 #include <tools/libc_compat.h> 39 #include "util/synthetic-events.h" 40 41 #define MAX_TIMESTAMP (~0ULL) 42 43 struct cs_etm_auxtrace { 44 struct auxtrace auxtrace; 45 struct auxtrace_queues queues; 46 struct auxtrace_heap heap; 47 struct itrace_synth_opts synth_opts; 48 struct perf_session *session; 49 struct machine *machine; 50 struct thread *unknown_thread; 51 52 u8 timeless_decoding; 53 u8 snapshot_mode; 54 u8 data_queued; 55 u8 sample_branches; 56 u8 sample_instructions; 57 58 int num_cpu; 59 u32 auxtrace_type; 60 u64 branches_sample_type; 61 u64 branches_id; 62 u64 instructions_sample_type; 63 u64 instructions_sample_period; 64 u64 instructions_id; 65 u64 **metadata; 66 u64 kernel_start; 67 unsigned int pmu_type; 68 }; 69 70 struct cs_etm_traceid_queue { 71 u8 trace_chan_id; 72 pid_t pid, tid; 73 u64 period_instructions; 74 size_t last_branch_pos; 75 union perf_event *event_buf; 76 struct thread *thread; 77 struct branch_stack *last_branch; 78 struct branch_stack *last_branch_rb; 79 struct cs_etm_packet *prev_packet; 80 struct cs_etm_packet *packet; 81 struct cs_etm_packet_queue packet_queue; 82 }; 83 84 struct cs_etm_queue { 85 struct cs_etm_auxtrace *etm; 86 struct cs_etm_decoder *decoder; 87 struct auxtrace_buffer *buffer; 88 unsigned int queue_nr; 89 u8 pending_timestamp; 90 u64 offset; 91 const unsigned char *buf; 92 size_t buf_len, buf_used; 93 /* Conversion between traceID and index in traceid_queues array */ 94 struct intlist *traceid_queues_list; 95 struct cs_etm_traceid_queue **traceid_queues; 96 }; 97 98 /* RB tree for quick conversion between traceID and metadata pointers */ 99 static struct intlist *traceid_list; 100 101 static int cs_etm__update_queues(struct cs_etm_auxtrace *etm); 102 static int cs_etm__process_queues(struct cs_etm_auxtrace *etm); 103 static int cs_etm__process_timeless_queues(struct cs_etm_auxtrace *etm, 104 pid_t tid); 105 static int cs_etm__get_data_block(struct cs_etm_queue *etmq); 106 static int cs_etm__decode_data_block(struct cs_etm_queue *etmq); 107 108 /* PTMs ETMIDR [11:8] set to b0011 */ 109 #define ETMIDR_PTM_VERSION 0x00000300 110 111 /* 112 * A struct auxtrace_heap_item only has a queue_nr and a timestamp to 113 * work with. One option is to modify to auxtrace_heap_XYZ() API or simply 114 * encode the etm queue number as the upper 16 bit and the channel as 115 * the lower 16 bit. 116 */ 117 #define TO_CS_QUEUE_NR(queue_nr, trace_chan_id) \ 118 (queue_nr << 16 | trace_chan_id) 119 #define TO_QUEUE_NR(cs_queue_nr) (cs_queue_nr >> 16) 120 #define TO_TRACE_CHAN_ID(cs_queue_nr) (cs_queue_nr & 0x0000ffff) 121 122 static u32 cs_etm__get_v7_protocol_version(u32 etmidr) 123 { 124 etmidr &= ETMIDR_PTM_VERSION; 125 126 if (etmidr == ETMIDR_PTM_VERSION) 127 return CS_ETM_PROTO_PTM; 128 129 return CS_ETM_PROTO_ETMV3; 130 } 131 132 static int cs_etm__get_magic(u8 trace_chan_id, u64 *magic) 133 { 134 struct int_node *inode; 135 u64 *metadata; 136 137 inode = intlist__find(traceid_list, trace_chan_id); 138 if (!inode) 139 return -EINVAL; 140 141 metadata = inode->priv; 142 *magic = metadata[CS_ETM_MAGIC]; 143 return 0; 144 } 145 146 int cs_etm__get_cpu(u8 trace_chan_id, int *cpu) 147 { 148 struct int_node *inode; 149 u64 *metadata; 150 151 inode = intlist__find(traceid_list, trace_chan_id); 152 if (!inode) 153 return -EINVAL; 154 155 metadata = inode->priv; 156 *cpu = (int)metadata[CS_ETM_CPU]; 157 return 0; 158 } 159 160 /* 161 * The returned PID format is presented by two bits: 162 * 163 * Bit ETM_OPT_CTXTID: CONTEXTIDR or CONTEXTIDR_EL1 is traced; 164 * Bit ETM_OPT_CTXTID2: CONTEXTIDR_EL2 is traced. 165 * 166 * It's possible that the two bits ETM_OPT_CTXTID and ETM_OPT_CTXTID2 167 * are enabled at the same time when the session runs on an EL2 kernel. 168 * This means the CONTEXTIDR_EL1 and CONTEXTIDR_EL2 both will be 169 * recorded in the trace data, the tool will selectively use 170 * CONTEXTIDR_EL2 as PID. 171 */ 172 int cs_etm__get_pid_fmt(u8 trace_chan_id, u64 *pid_fmt) 173 { 174 struct int_node *inode; 175 u64 *metadata, val; 176 177 inode = intlist__find(traceid_list, trace_chan_id); 178 if (!inode) 179 return -EINVAL; 180 181 metadata = inode->priv; 182 183 if (metadata[CS_ETM_MAGIC] == __perf_cs_etmv3_magic) { 184 val = metadata[CS_ETM_ETMCR]; 185 /* CONTEXTIDR is traced */ 186 if (val & BIT(ETM_OPT_CTXTID)) 187 *pid_fmt = BIT(ETM_OPT_CTXTID); 188 } else { 189 val = metadata[CS_ETMV4_TRCCONFIGR]; 190 /* CONTEXTIDR_EL2 is traced */ 191 if (val & (BIT(ETM4_CFG_BIT_VMID) | BIT(ETM4_CFG_BIT_VMID_OPT))) 192 *pid_fmt = BIT(ETM_OPT_CTXTID2); 193 /* CONTEXTIDR_EL1 is traced */ 194 else if (val & BIT(ETM4_CFG_BIT_CTXTID)) 195 *pid_fmt = BIT(ETM_OPT_CTXTID); 196 } 197 198 return 0; 199 } 200 201 void cs_etm__etmq_set_traceid_queue_timestamp(struct cs_etm_queue *etmq, 202 u8 trace_chan_id) 203 { 204 /* 205 * When a timestamp packet is encountered the backend code 206 * is stopped so that the front end has time to process packets 207 * that were accumulated in the traceID queue. Since there can 208 * be more than one channel per cs_etm_queue, we need to specify 209 * what traceID queue needs servicing. 210 */ 211 etmq->pending_timestamp = trace_chan_id; 212 } 213 214 static u64 cs_etm__etmq_get_timestamp(struct cs_etm_queue *etmq, 215 u8 *trace_chan_id) 216 { 217 struct cs_etm_packet_queue *packet_queue; 218 219 if (!etmq->pending_timestamp) 220 return 0; 221 222 if (trace_chan_id) 223 *trace_chan_id = etmq->pending_timestamp; 224 225 packet_queue = cs_etm__etmq_get_packet_queue(etmq, 226 etmq->pending_timestamp); 227 if (!packet_queue) 228 return 0; 229 230 /* Acknowledge pending status */ 231 etmq->pending_timestamp = 0; 232 233 /* See function cs_etm_decoder__do_{hard|soft}_timestamp() */ 234 return packet_queue->timestamp; 235 } 236 237 static void cs_etm__clear_packet_queue(struct cs_etm_packet_queue *queue) 238 { 239 int i; 240 241 queue->head = 0; 242 queue->tail = 0; 243 queue->packet_count = 0; 244 for (i = 0; i < CS_ETM_PACKET_MAX_BUFFER; i++) { 245 queue->packet_buffer[i].isa = CS_ETM_ISA_UNKNOWN; 246 queue->packet_buffer[i].start_addr = CS_ETM_INVAL_ADDR; 247 queue->packet_buffer[i].end_addr = CS_ETM_INVAL_ADDR; 248 queue->packet_buffer[i].instr_count = 0; 249 queue->packet_buffer[i].last_instr_taken_branch = false; 250 queue->packet_buffer[i].last_instr_size = 0; 251 queue->packet_buffer[i].last_instr_type = 0; 252 queue->packet_buffer[i].last_instr_subtype = 0; 253 queue->packet_buffer[i].last_instr_cond = 0; 254 queue->packet_buffer[i].flags = 0; 255 queue->packet_buffer[i].exception_number = UINT32_MAX; 256 queue->packet_buffer[i].trace_chan_id = UINT8_MAX; 257 queue->packet_buffer[i].cpu = INT_MIN; 258 } 259 } 260 261 static void cs_etm__clear_all_packet_queues(struct cs_etm_queue *etmq) 262 { 263 int idx; 264 struct int_node *inode; 265 struct cs_etm_traceid_queue *tidq; 266 struct intlist *traceid_queues_list = etmq->traceid_queues_list; 267 268 intlist__for_each_entry(inode, traceid_queues_list) { 269 idx = (int)(intptr_t)inode->priv; 270 tidq = etmq->traceid_queues[idx]; 271 cs_etm__clear_packet_queue(&tidq->packet_queue); 272 } 273 } 274 275 static int cs_etm__init_traceid_queue(struct cs_etm_queue *etmq, 276 struct cs_etm_traceid_queue *tidq, 277 u8 trace_chan_id) 278 { 279 int rc = -ENOMEM; 280 struct auxtrace_queue *queue; 281 struct cs_etm_auxtrace *etm = etmq->etm; 282 283 cs_etm__clear_packet_queue(&tidq->packet_queue); 284 285 queue = &etmq->etm->queues.queue_array[etmq->queue_nr]; 286 tidq->tid = queue->tid; 287 tidq->pid = -1; 288 tidq->trace_chan_id = trace_chan_id; 289 290 tidq->packet = zalloc(sizeof(struct cs_etm_packet)); 291 if (!tidq->packet) 292 goto out; 293 294 tidq->prev_packet = zalloc(sizeof(struct cs_etm_packet)); 295 if (!tidq->prev_packet) 296 goto out_free; 297 298 if (etm->synth_opts.last_branch) { 299 size_t sz = sizeof(struct branch_stack); 300 301 sz += etm->synth_opts.last_branch_sz * 302 sizeof(struct branch_entry); 303 tidq->last_branch = zalloc(sz); 304 if (!tidq->last_branch) 305 goto out_free; 306 tidq->last_branch_rb = zalloc(sz); 307 if (!tidq->last_branch_rb) 308 goto out_free; 309 } 310 311 tidq->event_buf = malloc(PERF_SAMPLE_MAX_SIZE); 312 if (!tidq->event_buf) 313 goto out_free; 314 315 return 0; 316 317 out_free: 318 zfree(&tidq->last_branch_rb); 319 zfree(&tidq->last_branch); 320 zfree(&tidq->prev_packet); 321 zfree(&tidq->packet); 322 out: 323 return rc; 324 } 325 326 static struct cs_etm_traceid_queue 327 *cs_etm__etmq_get_traceid_queue(struct cs_etm_queue *etmq, u8 trace_chan_id) 328 { 329 int idx; 330 struct int_node *inode; 331 struct intlist *traceid_queues_list; 332 struct cs_etm_traceid_queue *tidq, **traceid_queues; 333 struct cs_etm_auxtrace *etm = etmq->etm; 334 335 if (etm->timeless_decoding) 336 trace_chan_id = CS_ETM_PER_THREAD_TRACEID; 337 338 traceid_queues_list = etmq->traceid_queues_list; 339 340 /* 341 * Check if the traceid_queue exist for this traceID by looking 342 * in the queue list. 343 */ 344 inode = intlist__find(traceid_queues_list, trace_chan_id); 345 if (inode) { 346 idx = (int)(intptr_t)inode->priv; 347 return etmq->traceid_queues[idx]; 348 } 349 350 /* We couldn't find a traceid_queue for this traceID, allocate one */ 351 tidq = malloc(sizeof(*tidq)); 352 if (!tidq) 353 return NULL; 354 355 memset(tidq, 0, sizeof(*tidq)); 356 357 /* Get a valid index for the new traceid_queue */ 358 idx = intlist__nr_entries(traceid_queues_list); 359 /* Memory for the inode is free'ed in cs_etm_free_traceid_queues () */ 360 inode = intlist__findnew(traceid_queues_list, trace_chan_id); 361 if (!inode) 362 goto out_free; 363 364 /* Associate this traceID with this index */ 365 inode->priv = (void *)(intptr_t)idx; 366 367 if (cs_etm__init_traceid_queue(etmq, tidq, trace_chan_id)) 368 goto out_free; 369 370 /* Grow the traceid_queues array by one unit */ 371 traceid_queues = etmq->traceid_queues; 372 traceid_queues = reallocarray(traceid_queues, 373 idx + 1, 374 sizeof(*traceid_queues)); 375 376 /* 377 * On failure reallocarray() returns NULL and the original block of 378 * memory is left untouched. 379 */ 380 if (!traceid_queues) 381 goto out_free; 382 383 traceid_queues[idx] = tidq; 384 etmq->traceid_queues = traceid_queues; 385 386 return etmq->traceid_queues[idx]; 387 388 out_free: 389 /* 390 * Function intlist__remove() removes the inode from the list 391 * and delete the memory associated to it. 392 */ 393 intlist__remove(traceid_queues_list, inode); 394 free(tidq); 395 396 return NULL; 397 } 398 399 struct cs_etm_packet_queue 400 *cs_etm__etmq_get_packet_queue(struct cs_etm_queue *etmq, u8 trace_chan_id) 401 { 402 struct cs_etm_traceid_queue *tidq; 403 404 tidq = cs_etm__etmq_get_traceid_queue(etmq, trace_chan_id); 405 if (tidq) 406 return &tidq->packet_queue; 407 408 return NULL; 409 } 410 411 static void cs_etm__packet_swap(struct cs_etm_auxtrace *etm, 412 struct cs_etm_traceid_queue *tidq) 413 { 414 struct cs_etm_packet *tmp; 415 416 if (etm->sample_branches || etm->synth_opts.last_branch || 417 etm->sample_instructions) { 418 /* 419 * Swap PACKET with PREV_PACKET: PACKET becomes PREV_PACKET for 420 * the next incoming packet. 421 */ 422 tmp = tidq->packet; 423 tidq->packet = tidq->prev_packet; 424 tidq->prev_packet = tmp; 425 } 426 } 427 428 static void cs_etm__packet_dump(const char *pkt_string) 429 { 430 const char *color = PERF_COLOR_BLUE; 431 int len = strlen(pkt_string); 432 433 if (len && (pkt_string[len-1] == '\n')) 434 color_fprintf(stdout, color, " %s", pkt_string); 435 else 436 color_fprintf(stdout, color, " %s\n", pkt_string); 437 438 fflush(stdout); 439 } 440 441 static void cs_etm__set_trace_param_etmv3(struct cs_etm_trace_params *t_params, 442 struct cs_etm_auxtrace *etm, int idx, 443 u32 etmidr) 444 { 445 u64 **metadata = etm->metadata; 446 447 t_params[idx].protocol = cs_etm__get_v7_protocol_version(etmidr); 448 t_params[idx].etmv3.reg_ctrl = metadata[idx][CS_ETM_ETMCR]; 449 t_params[idx].etmv3.reg_trc_id = metadata[idx][CS_ETM_ETMTRACEIDR]; 450 } 451 452 static void cs_etm__set_trace_param_etmv4(struct cs_etm_trace_params *t_params, 453 struct cs_etm_auxtrace *etm, int idx) 454 { 455 u64 **metadata = etm->metadata; 456 457 t_params[idx].protocol = CS_ETM_PROTO_ETMV4i; 458 t_params[idx].etmv4.reg_idr0 = metadata[idx][CS_ETMV4_TRCIDR0]; 459 t_params[idx].etmv4.reg_idr1 = metadata[idx][CS_ETMV4_TRCIDR1]; 460 t_params[idx].etmv4.reg_idr2 = metadata[idx][CS_ETMV4_TRCIDR2]; 461 t_params[idx].etmv4.reg_idr8 = metadata[idx][CS_ETMV4_TRCIDR8]; 462 t_params[idx].etmv4.reg_configr = metadata[idx][CS_ETMV4_TRCCONFIGR]; 463 t_params[idx].etmv4.reg_traceidr = metadata[idx][CS_ETMV4_TRCTRACEIDR]; 464 } 465 466 static int cs_etm__init_trace_params(struct cs_etm_trace_params *t_params, 467 struct cs_etm_auxtrace *etm) 468 { 469 int i; 470 u32 etmidr; 471 u64 architecture; 472 473 for (i = 0; i < etm->num_cpu; i++) { 474 architecture = etm->metadata[i][CS_ETM_MAGIC]; 475 476 switch (architecture) { 477 case __perf_cs_etmv3_magic: 478 etmidr = etm->metadata[i][CS_ETM_ETMIDR]; 479 cs_etm__set_trace_param_etmv3(t_params, etm, i, etmidr); 480 break; 481 case __perf_cs_etmv4_magic: 482 cs_etm__set_trace_param_etmv4(t_params, etm, i); 483 break; 484 default: 485 return -EINVAL; 486 } 487 } 488 489 return 0; 490 } 491 492 static int cs_etm__init_decoder_params(struct cs_etm_decoder_params *d_params, 493 struct cs_etm_queue *etmq, 494 enum cs_etm_decoder_operation mode) 495 { 496 int ret = -EINVAL; 497 498 if (!(mode < CS_ETM_OPERATION_MAX)) 499 goto out; 500 501 d_params->packet_printer = cs_etm__packet_dump; 502 d_params->operation = mode; 503 d_params->data = etmq; 504 d_params->formatted = true; 505 d_params->fsyncs = false; 506 d_params->hsyncs = false; 507 d_params->frame_aligned = true; 508 509 ret = 0; 510 out: 511 return ret; 512 } 513 514 static void cs_etm__dump_event(struct cs_etm_auxtrace *etm, 515 struct auxtrace_buffer *buffer) 516 { 517 int ret; 518 const char *color = PERF_COLOR_BLUE; 519 struct cs_etm_decoder_params d_params; 520 struct cs_etm_trace_params *t_params; 521 struct cs_etm_decoder *decoder; 522 size_t buffer_used = 0; 523 524 fprintf(stdout, "\n"); 525 color_fprintf(stdout, color, 526 ". ... CoreSight ETM Trace data: size %zu bytes\n", 527 buffer->size); 528 529 /* Use metadata to fill in trace parameters for trace decoder */ 530 t_params = zalloc(sizeof(*t_params) * etm->num_cpu); 531 532 if (!t_params) 533 return; 534 535 if (cs_etm__init_trace_params(t_params, etm)) 536 goto out_free; 537 538 /* Set decoder parameters to simply print the trace packets */ 539 if (cs_etm__init_decoder_params(&d_params, NULL, 540 CS_ETM_OPERATION_PRINT)) 541 goto out_free; 542 543 decoder = cs_etm_decoder__new(etm->num_cpu, &d_params, t_params); 544 545 if (!decoder) 546 goto out_free; 547 do { 548 size_t consumed; 549 550 ret = cs_etm_decoder__process_data_block( 551 decoder, buffer->offset, 552 &((u8 *)buffer->data)[buffer_used], 553 buffer->size - buffer_used, &consumed); 554 if (ret) 555 break; 556 557 buffer_used += consumed; 558 } while (buffer_used < buffer->size); 559 560 cs_etm_decoder__free(decoder); 561 562 out_free: 563 zfree(&t_params); 564 } 565 566 static int cs_etm__flush_events(struct perf_session *session, 567 struct perf_tool *tool) 568 { 569 int ret; 570 struct cs_etm_auxtrace *etm = container_of(session->auxtrace, 571 struct cs_etm_auxtrace, 572 auxtrace); 573 if (dump_trace) 574 return 0; 575 576 if (!tool->ordered_events) 577 return -EINVAL; 578 579 ret = cs_etm__update_queues(etm); 580 581 if (ret < 0) 582 return ret; 583 584 if (etm->timeless_decoding) 585 return cs_etm__process_timeless_queues(etm, -1); 586 587 return cs_etm__process_queues(etm); 588 } 589 590 static void cs_etm__free_traceid_queues(struct cs_etm_queue *etmq) 591 { 592 int idx; 593 uintptr_t priv; 594 struct int_node *inode, *tmp; 595 struct cs_etm_traceid_queue *tidq; 596 struct intlist *traceid_queues_list = etmq->traceid_queues_list; 597 598 intlist__for_each_entry_safe(inode, tmp, traceid_queues_list) { 599 priv = (uintptr_t)inode->priv; 600 idx = priv; 601 602 /* Free this traceid_queue from the array */ 603 tidq = etmq->traceid_queues[idx]; 604 thread__zput(tidq->thread); 605 zfree(&tidq->event_buf); 606 zfree(&tidq->last_branch); 607 zfree(&tidq->last_branch_rb); 608 zfree(&tidq->prev_packet); 609 zfree(&tidq->packet); 610 zfree(&tidq); 611 612 /* 613 * Function intlist__remove() removes the inode from the list 614 * and delete the memory associated to it. 615 */ 616 intlist__remove(traceid_queues_list, inode); 617 } 618 619 /* Then the RB tree itself */ 620 intlist__delete(traceid_queues_list); 621 etmq->traceid_queues_list = NULL; 622 623 /* finally free the traceid_queues array */ 624 zfree(&etmq->traceid_queues); 625 } 626 627 static void cs_etm__free_queue(void *priv) 628 { 629 struct cs_etm_queue *etmq = priv; 630 631 if (!etmq) 632 return; 633 634 cs_etm_decoder__free(etmq->decoder); 635 cs_etm__free_traceid_queues(etmq); 636 free(etmq); 637 } 638 639 static void cs_etm__free_events(struct perf_session *session) 640 { 641 unsigned int i; 642 struct cs_etm_auxtrace *aux = container_of(session->auxtrace, 643 struct cs_etm_auxtrace, 644 auxtrace); 645 struct auxtrace_queues *queues = &aux->queues; 646 647 for (i = 0; i < queues->nr_queues; i++) { 648 cs_etm__free_queue(queues->queue_array[i].priv); 649 queues->queue_array[i].priv = NULL; 650 } 651 652 auxtrace_queues__free(queues); 653 } 654 655 static void cs_etm__free(struct perf_session *session) 656 { 657 int i; 658 struct int_node *inode, *tmp; 659 struct cs_etm_auxtrace *aux = container_of(session->auxtrace, 660 struct cs_etm_auxtrace, 661 auxtrace); 662 cs_etm__free_events(session); 663 session->auxtrace = NULL; 664 665 /* First remove all traceID/metadata nodes for the RB tree */ 666 intlist__for_each_entry_safe(inode, tmp, traceid_list) 667 intlist__remove(traceid_list, inode); 668 /* Then the RB tree itself */ 669 intlist__delete(traceid_list); 670 671 for (i = 0; i < aux->num_cpu; i++) 672 zfree(&aux->metadata[i]); 673 674 thread__zput(aux->unknown_thread); 675 zfree(&aux->metadata); 676 zfree(&aux); 677 } 678 679 static bool cs_etm__evsel_is_auxtrace(struct perf_session *session, 680 struct evsel *evsel) 681 { 682 struct cs_etm_auxtrace *aux = container_of(session->auxtrace, 683 struct cs_etm_auxtrace, 684 auxtrace); 685 686 return evsel->core.attr.type == aux->pmu_type; 687 } 688 689 static u8 cs_etm__cpu_mode(struct cs_etm_queue *etmq, u64 address) 690 { 691 struct machine *machine; 692 693 machine = etmq->etm->machine; 694 695 if (address >= etmq->etm->kernel_start) { 696 if (machine__is_host(machine)) 697 return PERF_RECORD_MISC_KERNEL; 698 else 699 return PERF_RECORD_MISC_GUEST_KERNEL; 700 } else { 701 if (machine__is_host(machine)) 702 return PERF_RECORD_MISC_USER; 703 else if (perf_guest) 704 return PERF_RECORD_MISC_GUEST_USER; 705 else 706 return PERF_RECORD_MISC_HYPERVISOR; 707 } 708 } 709 710 static u32 cs_etm__mem_access(struct cs_etm_queue *etmq, u8 trace_chan_id, 711 u64 address, size_t size, u8 *buffer) 712 { 713 u8 cpumode; 714 u64 offset; 715 int len; 716 struct thread *thread; 717 struct machine *machine; 718 struct addr_location al; 719 struct cs_etm_traceid_queue *tidq; 720 721 if (!etmq) 722 return 0; 723 724 machine = etmq->etm->machine; 725 cpumode = cs_etm__cpu_mode(etmq, address); 726 tidq = cs_etm__etmq_get_traceid_queue(etmq, trace_chan_id); 727 if (!tidq) 728 return 0; 729 730 thread = tidq->thread; 731 if (!thread) { 732 if (cpumode != PERF_RECORD_MISC_KERNEL) 733 return 0; 734 thread = etmq->etm->unknown_thread; 735 } 736 737 if (!thread__find_map(thread, cpumode, address, &al) || !al.map->dso) 738 return 0; 739 740 if (al.map->dso->data.status == DSO_DATA_STATUS_ERROR && 741 dso__data_status_seen(al.map->dso, DSO_DATA_STATUS_SEEN_ITRACE)) 742 return 0; 743 744 offset = al.map->map_ip(al.map, address); 745 746 map__load(al.map); 747 748 len = dso__data_read_offset(al.map->dso, machine, offset, buffer, size); 749 750 if (len <= 0) 751 return 0; 752 753 return len; 754 } 755 756 static struct cs_etm_queue *cs_etm__alloc_queue(struct cs_etm_auxtrace *etm) 757 { 758 struct cs_etm_decoder_params d_params; 759 struct cs_etm_trace_params *t_params = NULL; 760 struct cs_etm_queue *etmq; 761 762 etmq = zalloc(sizeof(*etmq)); 763 if (!etmq) 764 return NULL; 765 766 etmq->traceid_queues_list = intlist__new(NULL); 767 if (!etmq->traceid_queues_list) 768 goto out_free; 769 770 /* Use metadata to fill in trace parameters for trace decoder */ 771 t_params = zalloc(sizeof(*t_params) * etm->num_cpu); 772 773 if (!t_params) 774 goto out_free; 775 776 if (cs_etm__init_trace_params(t_params, etm)) 777 goto out_free; 778 779 /* Set decoder parameters to decode trace packets */ 780 if (cs_etm__init_decoder_params(&d_params, etmq, 781 CS_ETM_OPERATION_DECODE)) 782 goto out_free; 783 784 etmq->decoder = cs_etm_decoder__new(etm->num_cpu, &d_params, t_params); 785 786 if (!etmq->decoder) 787 goto out_free; 788 789 /* 790 * Register a function to handle all memory accesses required by 791 * the trace decoder library. 792 */ 793 if (cs_etm_decoder__add_mem_access_cb(etmq->decoder, 794 0x0L, ((u64) -1L), 795 cs_etm__mem_access)) 796 goto out_free_decoder; 797 798 zfree(&t_params); 799 return etmq; 800 801 out_free_decoder: 802 cs_etm_decoder__free(etmq->decoder); 803 out_free: 804 intlist__delete(etmq->traceid_queues_list); 805 free(etmq); 806 807 return NULL; 808 } 809 810 static int cs_etm__setup_queue(struct cs_etm_auxtrace *etm, 811 struct auxtrace_queue *queue, 812 unsigned int queue_nr) 813 { 814 int ret = 0; 815 unsigned int cs_queue_nr; 816 u8 trace_chan_id; 817 u64 timestamp; 818 struct cs_etm_queue *etmq = queue->priv; 819 820 if (list_empty(&queue->head) || etmq) 821 goto out; 822 823 etmq = cs_etm__alloc_queue(etm); 824 825 if (!etmq) { 826 ret = -ENOMEM; 827 goto out; 828 } 829 830 queue->priv = etmq; 831 etmq->etm = etm; 832 etmq->queue_nr = queue_nr; 833 etmq->offset = 0; 834 835 if (etm->timeless_decoding) 836 goto out; 837 838 /* 839 * We are under a CPU-wide trace scenario. As such we need to know 840 * when the code that generated the traces started to execute so that 841 * it can be correlated with execution on other CPUs. So we get a 842 * handle on the beginning of traces and decode until we find a 843 * timestamp. The timestamp is then added to the auxtrace min heap 844 * in order to know what nibble (of all the etmqs) to decode first. 845 */ 846 while (1) { 847 /* 848 * Fetch an aux_buffer from this etmq. Bail if no more 849 * blocks or an error has been encountered. 850 */ 851 ret = cs_etm__get_data_block(etmq); 852 if (ret <= 0) 853 goto out; 854 855 /* 856 * Run decoder on the trace block. The decoder will stop when 857 * encountering a timestamp, a full packet queue or the end of 858 * trace for that block. 859 */ 860 ret = cs_etm__decode_data_block(etmq); 861 if (ret) 862 goto out; 863 864 /* 865 * Function cs_etm_decoder__do_{hard|soft}_timestamp() does all 866 * the timestamp calculation for us. 867 */ 868 timestamp = cs_etm__etmq_get_timestamp(etmq, &trace_chan_id); 869 870 /* We found a timestamp, no need to continue. */ 871 if (timestamp) 872 break; 873 874 /* 875 * We didn't find a timestamp so empty all the traceid packet 876 * queues before looking for another timestamp packet, either 877 * in the current data block or a new one. Packets that were 878 * just decoded are useless since no timestamp has been 879 * associated with them. As such simply discard them. 880 */ 881 cs_etm__clear_all_packet_queues(etmq); 882 } 883 884 /* 885 * We have a timestamp. Add it to the min heap to reflect when 886 * instructions conveyed by the range packets of this traceID queue 887 * started to execute. Once the same has been done for all the traceID 888 * queues of each etmq, redenring and decoding can start in 889 * chronological order. 890 * 891 * Note that packets decoded above are still in the traceID's packet 892 * queue and will be processed in cs_etm__process_queues(). 893 */ 894 cs_queue_nr = TO_CS_QUEUE_NR(queue_nr, trace_chan_id); 895 ret = auxtrace_heap__add(&etm->heap, cs_queue_nr, timestamp); 896 out: 897 return ret; 898 } 899 900 static int cs_etm__setup_queues(struct cs_etm_auxtrace *etm) 901 { 902 unsigned int i; 903 int ret; 904 905 if (!etm->kernel_start) 906 etm->kernel_start = machine__kernel_start(etm->machine); 907 908 for (i = 0; i < etm->queues.nr_queues; i++) { 909 ret = cs_etm__setup_queue(etm, &etm->queues.queue_array[i], i); 910 if (ret) 911 return ret; 912 } 913 914 return 0; 915 } 916 917 static int cs_etm__update_queues(struct cs_etm_auxtrace *etm) 918 { 919 if (etm->queues.new_data) { 920 etm->queues.new_data = false; 921 return cs_etm__setup_queues(etm); 922 } 923 924 return 0; 925 } 926 927 static inline 928 void cs_etm__copy_last_branch_rb(struct cs_etm_queue *etmq, 929 struct cs_etm_traceid_queue *tidq) 930 { 931 struct branch_stack *bs_src = tidq->last_branch_rb; 932 struct branch_stack *bs_dst = tidq->last_branch; 933 size_t nr = 0; 934 935 /* 936 * Set the number of records before early exit: ->nr is used to 937 * determine how many branches to copy from ->entries. 938 */ 939 bs_dst->nr = bs_src->nr; 940 941 /* 942 * Early exit when there is nothing to copy. 943 */ 944 if (!bs_src->nr) 945 return; 946 947 /* 948 * As bs_src->entries is a circular buffer, we need to copy from it in 949 * two steps. First, copy the branches from the most recently inserted 950 * branch ->last_branch_pos until the end of bs_src->entries buffer. 951 */ 952 nr = etmq->etm->synth_opts.last_branch_sz - tidq->last_branch_pos; 953 memcpy(&bs_dst->entries[0], 954 &bs_src->entries[tidq->last_branch_pos], 955 sizeof(struct branch_entry) * nr); 956 957 /* 958 * If we wrapped around at least once, the branches from the beginning 959 * of the bs_src->entries buffer and until the ->last_branch_pos element 960 * are older valid branches: copy them over. The total number of 961 * branches copied over will be equal to the number of branches asked by 962 * the user in last_branch_sz. 963 */ 964 if (bs_src->nr >= etmq->etm->synth_opts.last_branch_sz) { 965 memcpy(&bs_dst->entries[nr], 966 &bs_src->entries[0], 967 sizeof(struct branch_entry) * tidq->last_branch_pos); 968 } 969 } 970 971 static inline 972 void cs_etm__reset_last_branch_rb(struct cs_etm_traceid_queue *tidq) 973 { 974 tidq->last_branch_pos = 0; 975 tidq->last_branch_rb->nr = 0; 976 } 977 978 static inline int cs_etm__t32_instr_size(struct cs_etm_queue *etmq, 979 u8 trace_chan_id, u64 addr) 980 { 981 u8 instrBytes[2]; 982 983 cs_etm__mem_access(etmq, trace_chan_id, addr, 984 ARRAY_SIZE(instrBytes), instrBytes); 985 /* 986 * T32 instruction size is indicated by bits[15:11] of the first 987 * 16-bit word of the instruction: 0b11101, 0b11110 and 0b11111 988 * denote a 32-bit instruction. 989 */ 990 return ((instrBytes[1] & 0xF8) >= 0xE8) ? 4 : 2; 991 } 992 993 static inline u64 cs_etm__first_executed_instr(struct cs_etm_packet *packet) 994 { 995 /* Returns 0 for the CS_ETM_DISCONTINUITY packet */ 996 if (packet->sample_type == CS_ETM_DISCONTINUITY) 997 return 0; 998 999 return packet->start_addr; 1000 } 1001 1002 static inline 1003 u64 cs_etm__last_executed_instr(const struct cs_etm_packet *packet) 1004 { 1005 /* Returns 0 for the CS_ETM_DISCONTINUITY packet */ 1006 if (packet->sample_type == CS_ETM_DISCONTINUITY) 1007 return 0; 1008 1009 return packet->end_addr - packet->last_instr_size; 1010 } 1011 1012 static inline u64 cs_etm__instr_addr(struct cs_etm_queue *etmq, 1013 u64 trace_chan_id, 1014 const struct cs_etm_packet *packet, 1015 u64 offset) 1016 { 1017 if (packet->isa == CS_ETM_ISA_T32) { 1018 u64 addr = packet->start_addr; 1019 1020 while (offset) { 1021 addr += cs_etm__t32_instr_size(etmq, 1022 trace_chan_id, addr); 1023 offset--; 1024 } 1025 return addr; 1026 } 1027 1028 /* Assume a 4 byte instruction size (A32/A64) */ 1029 return packet->start_addr + offset * 4; 1030 } 1031 1032 static void cs_etm__update_last_branch_rb(struct cs_etm_queue *etmq, 1033 struct cs_etm_traceid_queue *tidq) 1034 { 1035 struct branch_stack *bs = tidq->last_branch_rb; 1036 struct branch_entry *be; 1037 1038 /* 1039 * The branches are recorded in a circular buffer in reverse 1040 * chronological order: we start recording from the last element of the 1041 * buffer down. After writing the first element of the stack, move the 1042 * insert position back to the end of the buffer. 1043 */ 1044 if (!tidq->last_branch_pos) 1045 tidq->last_branch_pos = etmq->etm->synth_opts.last_branch_sz; 1046 1047 tidq->last_branch_pos -= 1; 1048 1049 be = &bs->entries[tidq->last_branch_pos]; 1050 be->from = cs_etm__last_executed_instr(tidq->prev_packet); 1051 be->to = cs_etm__first_executed_instr(tidq->packet); 1052 /* No support for mispredict */ 1053 be->flags.mispred = 0; 1054 be->flags.predicted = 1; 1055 1056 /* 1057 * Increment bs->nr until reaching the number of last branches asked by 1058 * the user on the command line. 1059 */ 1060 if (bs->nr < etmq->etm->synth_opts.last_branch_sz) 1061 bs->nr += 1; 1062 } 1063 1064 static int cs_etm__inject_event(union perf_event *event, 1065 struct perf_sample *sample, u64 type) 1066 { 1067 event->header.size = perf_event__sample_event_size(sample, type, 0); 1068 return perf_event__synthesize_sample(event, type, 0, sample); 1069 } 1070 1071 1072 static int 1073 cs_etm__get_trace(struct cs_etm_queue *etmq) 1074 { 1075 struct auxtrace_buffer *aux_buffer = etmq->buffer; 1076 struct auxtrace_buffer *old_buffer = aux_buffer; 1077 struct auxtrace_queue *queue; 1078 1079 queue = &etmq->etm->queues.queue_array[etmq->queue_nr]; 1080 1081 aux_buffer = auxtrace_buffer__next(queue, aux_buffer); 1082 1083 /* If no more data, drop the previous auxtrace_buffer and return */ 1084 if (!aux_buffer) { 1085 if (old_buffer) 1086 auxtrace_buffer__drop_data(old_buffer); 1087 etmq->buf_len = 0; 1088 return 0; 1089 } 1090 1091 etmq->buffer = aux_buffer; 1092 1093 /* If the aux_buffer doesn't have data associated, try to load it */ 1094 if (!aux_buffer->data) { 1095 /* get the file desc associated with the perf data file */ 1096 int fd = perf_data__fd(etmq->etm->session->data); 1097 1098 aux_buffer->data = auxtrace_buffer__get_data(aux_buffer, fd); 1099 if (!aux_buffer->data) 1100 return -ENOMEM; 1101 } 1102 1103 /* If valid, drop the previous buffer */ 1104 if (old_buffer) 1105 auxtrace_buffer__drop_data(old_buffer); 1106 1107 etmq->buf_used = 0; 1108 etmq->buf_len = aux_buffer->size; 1109 etmq->buf = aux_buffer->data; 1110 1111 return etmq->buf_len; 1112 } 1113 1114 static void cs_etm__set_pid_tid_cpu(struct cs_etm_auxtrace *etm, 1115 struct cs_etm_traceid_queue *tidq) 1116 { 1117 if ((!tidq->thread) && (tidq->tid != -1)) 1118 tidq->thread = machine__find_thread(etm->machine, -1, 1119 tidq->tid); 1120 1121 if (tidq->thread) 1122 tidq->pid = tidq->thread->pid_; 1123 } 1124 1125 int cs_etm__etmq_set_tid(struct cs_etm_queue *etmq, 1126 pid_t tid, u8 trace_chan_id) 1127 { 1128 int cpu, err = -EINVAL; 1129 struct cs_etm_auxtrace *etm = etmq->etm; 1130 struct cs_etm_traceid_queue *tidq; 1131 1132 tidq = cs_etm__etmq_get_traceid_queue(etmq, trace_chan_id); 1133 if (!tidq) 1134 return err; 1135 1136 if (cs_etm__get_cpu(trace_chan_id, &cpu) < 0) 1137 return err; 1138 1139 err = machine__set_current_tid(etm->machine, cpu, tid, tid); 1140 if (err) 1141 return err; 1142 1143 tidq->tid = tid; 1144 thread__zput(tidq->thread); 1145 1146 cs_etm__set_pid_tid_cpu(etm, tidq); 1147 return 0; 1148 } 1149 1150 bool cs_etm__etmq_is_timeless(struct cs_etm_queue *etmq) 1151 { 1152 return !!etmq->etm->timeless_decoding; 1153 } 1154 1155 static void cs_etm__copy_insn(struct cs_etm_queue *etmq, 1156 u64 trace_chan_id, 1157 const struct cs_etm_packet *packet, 1158 struct perf_sample *sample) 1159 { 1160 /* 1161 * It's pointless to read instructions for the CS_ETM_DISCONTINUITY 1162 * packet, so directly bail out with 'insn_len' = 0. 1163 */ 1164 if (packet->sample_type == CS_ETM_DISCONTINUITY) { 1165 sample->insn_len = 0; 1166 return; 1167 } 1168 1169 /* 1170 * T32 instruction size might be 32-bit or 16-bit, decide by calling 1171 * cs_etm__t32_instr_size(). 1172 */ 1173 if (packet->isa == CS_ETM_ISA_T32) 1174 sample->insn_len = cs_etm__t32_instr_size(etmq, trace_chan_id, 1175 sample->ip); 1176 /* Otherwise, A64 and A32 instruction size are always 32-bit. */ 1177 else 1178 sample->insn_len = 4; 1179 1180 cs_etm__mem_access(etmq, trace_chan_id, sample->ip, 1181 sample->insn_len, (void *)sample->insn); 1182 } 1183 1184 static int cs_etm__synth_instruction_sample(struct cs_etm_queue *etmq, 1185 struct cs_etm_traceid_queue *tidq, 1186 u64 addr, u64 period) 1187 { 1188 int ret = 0; 1189 struct cs_etm_auxtrace *etm = etmq->etm; 1190 union perf_event *event = tidq->event_buf; 1191 struct perf_sample sample = {.ip = 0,}; 1192 1193 event->sample.header.type = PERF_RECORD_SAMPLE; 1194 event->sample.header.misc = cs_etm__cpu_mode(etmq, addr); 1195 event->sample.header.size = sizeof(struct perf_event_header); 1196 1197 sample.ip = addr; 1198 sample.pid = tidq->pid; 1199 sample.tid = tidq->tid; 1200 sample.id = etmq->etm->instructions_id; 1201 sample.stream_id = etmq->etm->instructions_id; 1202 sample.period = period; 1203 sample.cpu = tidq->packet->cpu; 1204 sample.flags = tidq->prev_packet->flags; 1205 sample.cpumode = event->sample.header.misc; 1206 1207 cs_etm__copy_insn(etmq, tidq->trace_chan_id, tidq->packet, &sample); 1208 1209 if (etm->synth_opts.last_branch) 1210 sample.branch_stack = tidq->last_branch; 1211 1212 if (etm->synth_opts.inject) { 1213 ret = cs_etm__inject_event(event, &sample, 1214 etm->instructions_sample_type); 1215 if (ret) 1216 return ret; 1217 } 1218 1219 ret = perf_session__deliver_synth_event(etm->session, event, &sample); 1220 1221 if (ret) 1222 pr_err( 1223 "CS ETM Trace: failed to deliver instruction event, error %d\n", 1224 ret); 1225 1226 return ret; 1227 } 1228 1229 /* 1230 * The cs etm packet encodes an instruction range between a branch target 1231 * and the next taken branch. Generate sample accordingly. 1232 */ 1233 static int cs_etm__synth_branch_sample(struct cs_etm_queue *etmq, 1234 struct cs_etm_traceid_queue *tidq) 1235 { 1236 int ret = 0; 1237 struct cs_etm_auxtrace *etm = etmq->etm; 1238 struct perf_sample sample = {.ip = 0,}; 1239 union perf_event *event = tidq->event_buf; 1240 struct dummy_branch_stack { 1241 u64 nr; 1242 u64 hw_idx; 1243 struct branch_entry entries; 1244 } dummy_bs; 1245 u64 ip; 1246 1247 ip = cs_etm__last_executed_instr(tidq->prev_packet); 1248 1249 event->sample.header.type = PERF_RECORD_SAMPLE; 1250 event->sample.header.misc = cs_etm__cpu_mode(etmq, ip); 1251 event->sample.header.size = sizeof(struct perf_event_header); 1252 1253 sample.ip = ip; 1254 sample.pid = tidq->pid; 1255 sample.tid = tidq->tid; 1256 sample.addr = cs_etm__first_executed_instr(tidq->packet); 1257 sample.id = etmq->etm->branches_id; 1258 sample.stream_id = etmq->etm->branches_id; 1259 sample.period = 1; 1260 sample.cpu = tidq->packet->cpu; 1261 sample.flags = tidq->prev_packet->flags; 1262 sample.cpumode = event->sample.header.misc; 1263 1264 cs_etm__copy_insn(etmq, tidq->trace_chan_id, tidq->prev_packet, 1265 &sample); 1266 1267 /* 1268 * perf report cannot handle events without a branch stack 1269 */ 1270 if (etm->synth_opts.last_branch) { 1271 dummy_bs = (struct dummy_branch_stack){ 1272 .nr = 1, 1273 .hw_idx = -1ULL, 1274 .entries = { 1275 .from = sample.ip, 1276 .to = sample.addr, 1277 }, 1278 }; 1279 sample.branch_stack = (struct branch_stack *)&dummy_bs; 1280 } 1281 1282 if (etm->synth_opts.inject) { 1283 ret = cs_etm__inject_event(event, &sample, 1284 etm->branches_sample_type); 1285 if (ret) 1286 return ret; 1287 } 1288 1289 ret = perf_session__deliver_synth_event(etm->session, event, &sample); 1290 1291 if (ret) 1292 pr_err( 1293 "CS ETM Trace: failed to deliver instruction event, error %d\n", 1294 ret); 1295 1296 return ret; 1297 } 1298 1299 struct cs_etm_synth { 1300 struct perf_tool dummy_tool; 1301 struct perf_session *session; 1302 }; 1303 1304 static int cs_etm__event_synth(struct perf_tool *tool, 1305 union perf_event *event, 1306 struct perf_sample *sample __maybe_unused, 1307 struct machine *machine __maybe_unused) 1308 { 1309 struct cs_etm_synth *cs_etm_synth = 1310 container_of(tool, struct cs_etm_synth, dummy_tool); 1311 1312 return perf_session__deliver_synth_event(cs_etm_synth->session, 1313 event, NULL); 1314 } 1315 1316 static int cs_etm__synth_event(struct perf_session *session, 1317 struct perf_event_attr *attr, u64 id) 1318 { 1319 struct cs_etm_synth cs_etm_synth; 1320 1321 memset(&cs_etm_synth, 0, sizeof(struct cs_etm_synth)); 1322 cs_etm_synth.session = session; 1323 1324 return perf_event__synthesize_attr(&cs_etm_synth.dummy_tool, attr, 1, 1325 &id, cs_etm__event_synth); 1326 } 1327 1328 static int cs_etm__synth_events(struct cs_etm_auxtrace *etm, 1329 struct perf_session *session) 1330 { 1331 struct evlist *evlist = session->evlist; 1332 struct evsel *evsel; 1333 struct perf_event_attr attr; 1334 bool found = false; 1335 u64 id; 1336 int err; 1337 1338 evlist__for_each_entry(evlist, evsel) { 1339 if (evsel->core.attr.type == etm->pmu_type) { 1340 found = true; 1341 break; 1342 } 1343 } 1344 1345 if (!found) { 1346 pr_debug("No selected events with CoreSight Trace data\n"); 1347 return 0; 1348 } 1349 1350 memset(&attr, 0, sizeof(struct perf_event_attr)); 1351 attr.size = sizeof(struct perf_event_attr); 1352 attr.type = PERF_TYPE_HARDWARE; 1353 attr.sample_type = evsel->core.attr.sample_type & PERF_SAMPLE_MASK; 1354 attr.sample_type |= PERF_SAMPLE_IP | PERF_SAMPLE_TID | 1355 PERF_SAMPLE_PERIOD; 1356 if (etm->timeless_decoding) 1357 attr.sample_type &= ~(u64)PERF_SAMPLE_TIME; 1358 else 1359 attr.sample_type |= PERF_SAMPLE_TIME; 1360 1361 attr.exclude_user = evsel->core.attr.exclude_user; 1362 attr.exclude_kernel = evsel->core.attr.exclude_kernel; 1363 attr.exclude_hv = evsel->core.attr.exclude_hv; 1364 attr.exclude_host = evsel->core.attr.exclude_host; 1365 attr.exclude_guest = evsel->core.attr.exclude_guest; 1366 attr.sample_id_all = evsel->core.attr.sample_id_all; 1367 attr.read_format = evsel->core.attr.read_format; 1368 1369 /* create new id val to be a fixed offset from evsel id */ 1370 id = evsel->core.id[0] + 1000000000; 1371 1372 if (!id) 1373 id = 1; 1374 1375 if (etm->synth_opts.branches) { 1376 attr.config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS; 1377 attr.sample_period = 1; 1378 attr.sample_type |= PERF_SAMPLE_ADDR; 1379 err = cs_etm__synth_event(session, &attr, id); 1380 if (err) 1381 return err; 1382 etm->sample_branches = true; 1383 etm->branches_sample_type = attr.sample_type; 1384 etm->branches_id = id; 1385 id += 1; 1386 attr.sample_type &= ~(u64)PERF_SAMPLE_ADDR; 1387 } 1388 1389 if (etm->synth_opts.last_branch) { 1390 attr.sample_type |= PERF_SAMPLE_BRANCH_STACK; 1391 /* 1392 * We don't use the hardware index, but the sample generation 1393 * code uses the new format branch_stack with this field, 1394 * so the event attributes must indicate that it's present. 1395 */ 1396 attr.branch_sample_type |= PERF_SAMPLE_BRANCH_HW_INDEX; 1397 } 1398 1399 if (etm->synth_opts.instructions) { 1400 attr.config = PERF_COUNT_HW_INSTRUCTIONS; 1401 attr.sample_period = etm->synth_opts.period; 1402 etm->instructions_sample_period = attr.sample_period; 1403 err = cs_etm__synth_event(session, &attr, id); 1404 if (err) 1405 return err; 1406 etm->sample_instructions = true; 1407 etm->instructions_sample_type = attr.sample_type; 1408 etm->instructions_id = id; 1409 id += 1; 1410 } 1411 1412 return 0; 1413 } 1414 1415 static int cs_etm__sample(struct cs_etm_queue *etmq, 1416 struct cs_etm_traceid_queue *tidq) 1417 { 1418 struct cs_etm_auxtrace *etm = etmq->etm; 1419 int ret; 1420 u8 trace_chan_id = tidq->trace_chan_id; 1421 u64 instrs_prev; 1422 1423 /* Get instructions remainder from previous packet */ 1424 instrs_prev = tidq->period_instructions; 1425 1426 tidq->period_instructions += tidq->packet->instr_count; 1427 1428 /* 1429 * Record a branch when the last instruction in 1430 * PREV_PACKET is a branch. 1431 */ 1432 if (etm->synth_opts.last_branch && 1433 tidq->prev_packet->sample_type == CS_ETM_RANGE && 1434 tidq->prev_packet->last_instr_taken_branch) 1435 cs_etm__update_last_branch_rb(etmq, tidq); 1436 1437 if (etm->sample_instructions && 1438 tidq->period_instructions >= etm->instructions_sample_period) { 1439 /* 1440 * Emit instruction sample periodically 1441 * TODO: allow period to be defined in cycles and clock time 1442 */ 1443 1444 /* 1445 * Below diagram demonstrates the instruction samples 1446 * generation flows: 1447 * 1448 * Instrs Instrs Instrs Instrs 1449 * Sample(n) Sample(n+1) Sample(n+2) Sample(n+3) 1450 * | | | | 1451 * V V V V 1452 * -------------------------------------------------- 1453 * ^ ^ 1454 * | | 1455 * Period Period 1456 * instructions(Pi) instructions(Pi') 1457 * 1458 * | | 1459 * \---------------- -----------------/ 1460 * V 1461 * tidq->packet->instr_count 1462 * 1463 * Instrs Sample(n...) are the synthesised samples occurring 1464 * every etm->instructions_sample_period instructions - as 1465 * defined on the perf command line. Sample(n) is being the 1466 * last sample before the current etm packet, n+1 to n+3 1467 * samples are generated from the current etm packet. 1468 * 1469 * tidq->packet->instr_count represents the number of 1470 * instructions in the current etm packet. 1471 * 1472 * Period instructions (Pi) contains the the number of 1473 * instructions executed after the sample point(n) from the 1474 * previous etm packet. This will always be less than 1475 * etm->instructions_sample_period. 1476 * 1477 * When generate new samples, it combines with two parts 1478 * instructions, one is the tail of the old packet and another 1479 * is the head of the new coming packet, to generate 1480 * sample(n+1); sample(n+2) and sample(n+3) consume the 1481 * instructions with sample period. After sample(n+3), the rest 1482 * instructions will be used by later packet and it is assigned 1483 * to tidq->period_instructions for next round calculation. 1484 */ 1485 1486 /* 1487 * Get the initial offset into the current packet instructions; 1488 * entry conditions ensure that instrs_prev is less than 1489 * etm->instructions_sample_period. 1490 */ 1491 u64 offset = etm->instructions_sample_period - instrs_prev; 1492 u64 addr; 1493 1494 /* Prepare last branches for instruction sample */ 1495 if (etm->synth_opts.last_branch) 1496 cs_etm__copy_last_branch_rb(etmq, tidq); 1497 1498 while (tidq->period_instructions >= 1499 etm->instructions_sample_period) { 1500 /* 1501 * Calculate the address of the sampled instruction (-1 1502 * as sample is reported as though instruction has just 1503 * been executed, but PC has not advanced to next 1504 * instruction) 1505 */ 1506 addr = cs_etm__instr_addr(etmq, trace_chan_id, 1507 tidq->packet, offset - 1); 1508 ret = cs_etm__synth_instruction_sample( 1509 etmq, tidq, addr, 1510 etm->instructions_sample_period); 1511 if (ret) 1512 return ret; 1513 1514 offset += etm->instructions_sample_period; 1515 tidq->period_instructions -= 1516 etm->instructions_sample_period; 1517 } 1518 } 1519 1520 if (etm->sample_branches) { 1521 bool generate_sample = false; 1522 1523 /* Generate sample for tracing on packet */ 1524 if (tidq->prev_packet->sample_type == CS_ETM_DISCONTINUITY) 1525 generate_sample = true; 1526 1527 /* Generate sample for branch taken packet */ 1528 if (tidq->prev_packet->sample_type == CS_ETM_RANGE && 1529 tidq->prev_packet->last_instr_taken_branch) 1530 generate_sample = true; 1531 1532 if (generate_sample) { 1533 ret = cs_etm__synth_branch_sample(etmq, tidq); 1534 if (ret) 1535 return ret; 1536 } 1537 } 1538 1539 cs_etm__packet_swap(etm, tidq); 1540 1541 return 0; 1542 } 1543 1544 static int cs_etm__exception(struct cs_etm_traceid_queue *tidq) 1545 { 1546 /* 1547 * When the exception packet is inserted, whether the last instruction 1548 * in previous range packet is taken branch or not, we need to force 1549 * to set 'prev_packet->last_instr_taken_branch' to true. This ensures 1550 * to generate branch sample for the instruction range before the 1551 * exception is trapped to kernel or before the exception returning. 1552 * 1553 * The exception packet includes the dummy address values, so don't 1554 * swap PACKET with PREV_PACKET. This keeps PREV_PACKET to be useful 1555 * for generating instruction and branch samples. 1556 */ 1557 if (tidq->prev_packet->sample_type == CS_ETM_RANGE) 1558 tidq->prev_packet->last_instr_taken_branch = true; 1559 1560 return 0; 1561 } 1562 1563 static int cs_etm__flush(struct cs_etm_queue *etmq, 1564 struct cs_etm_traceid_queue *tidq) 1565 { 1566 int err = 0; 1567 struct cs_etm_auxtrace *etm = etmq->etm; 1568 1569 /* Handle start tracing packet */ 1570 if (tidq->prev_packet->sample_type == CS_ETM_EMPTY) 1571 goto swap_packet; 1572 1573 if (etmq->etm->synth_opts.last_branch && 1574 tidq->prev_packet->sample_type == CS_ETM_RANGE) { 1575 u64 addr; 1576 1577 /* Prepare last branches for instruction sample */ 1578 cs_etm__copy_last_branch_rb(etmq, tidq); 1579 1580 /* 1581 * Generate a last branch event for the branches left in the 1582 * circular buffer at the end of the trace. 1583 * 1584 * Use the address of the end of the last reported execution 1585 * range 1586 */ 1587 addr = cs_etm__last_executed_instr(tidq->prev_packet); 1588 1589 err = cs_etm__synth_instruction_sample( 1590 etmq, tidq, addr, 1591 tidq->period_instructions); 1592 if (err) 1593 return err; 1594 1595 tidq->period_instructions = 0; 1596 1597 } 1598 1599 if (etm->sample_branches && 1600 tidq->prev_packet->sample_type == CS_ETM_RANGE) { 1601 err = cs_etm__synth_branch_sample(etmq, tidq); 1602 if (err) 1603 return err; 1604 } 1605 1606 swap_packet: 1607 cs_etm__packet_swap(etm, tidq); 1608 1609 /* Reset last branches after flush the trace */ 1610 if (etm->synth_opts.last_branch) 1611 cs_etm__reset_last_branch_rb(tidq); 1612 1613 return err; 1614 } 1615 1616 static int cs_etm__end_block(struct cs_etm_queue *etmq, 1617 struct cs_etm_traceid_queue *tidq) 1618 { 1619 int err; 1620 1621 /* 1622 * It has no new packet coming and 'etmq->packet' contains the stale 1623 * packet which was set at the previous time with packets swapping; 1624 * so skip to generate branch sample to avoid stale packet. 1625 * 1626 * For this case only flush branch stack and generate a last branch 1627 * event for the branches left in the circular buffer at the end of 1628 * the trace. 1629 */ 1630 if (etmq->etm->synth_opts.last_branch && 1631 tidq->prev_packet->sample_type == CS_ETM_RANGE) { 1632 u64 addr; 1633 1634 /* Prepare last branches for instruction sample */ 1635 cs_etm__copy_last_branch_rb(etmq, tidq); 1636 1637 /* 1638 * Use the address of the end of the last reported execution 1639 * range. 1640 */ 1641 addr = cs_etm__last_executed_instr(tidq->prev_packet); 1642 1643 err = cs_etm__synth_instruction_sample( 1644 etmq, tidq, addr, 1645 tidq->period_instructions); 1646 if (err) 1647 return err; 1648 1649 tidq->period_instructions = 0; 1650 } 1651 1652 return 0; 1653 } 1654 /* 1655 * cs_etm__get_data_block: Fetch a block from the auxtrace_buffer queue 1656 * if need be. 1657 * Returns: < 0 if error 1658 * = 0 if no more auxtrace_buffer to read 1659 * > 0 if the current buffer isn't empty yet 1660 */ 1661 static int cs_etm__get_data_block(struct cs_etm_queue *etmq) 1662 { 1663 int ret; 1664 1665 if (!etmq->buf_len) { 1666 ret = cs_etm__get_trace(etmq); 1667 if (ret <= 0) 1668 return ret; 1669 /* 1670 * We cannot assume consecutive blocks in the data file 1671 * are contiguous, reset the decoder to force re-sync. 1672 */ 1673 ret = cs_etm_decoder__reset(etmq->decoder); 1674 if (ret) 1675 return ret; 1676 } 1677 1678 return etmq->buf_len; 1679 } 1680 1681 static bool cs_etm__is_svc_instr(struct cs_etm_queue *etmq, u8 trace_chan_id, 1682 struct cs_etm_packet *packet, 1683 u64 end_addr) 1684 { 1685 /* Initialise to keep compiler happy */ 1686 u16 instr16 = 0; 1687 u32 instr32 = 0; 1688 u64 addr; 1689 1690 switch (packet->isa) { 1691 case CS_ETM_ISA_T32: 1692 /* 1693 * The SVC of T32 is defined in ARM DDI 0487D.a, F5.1.247: 1694 * 1695 * b'15 b'8 1696 * +-----------------+--------+ 1697 * | 1 1 0 1 1 1 1 1 | imm8 | 1698 * +-----------------+--------+ 1699 * 1700 * According to the specification, it only defines SVC for T32 1701 * with 16 bits instruction and has no definition for 32bits; 1702 * so below only read 2 bytes as instruction size for T32. 1703 */ 1704 addr = end_addr - 2; 1705 cs_etm__mem_access(etmq, trace_chan_id, addr, 1706 sizeof(instr16), (u8 *)&instr16); 1707 if ((instr16 & 0xFF00) == 0xDF00) 1708 return true; 1709 1710 break; 1711 case CS_ETM_ISA_A32: 1712 /* 1713 * The SVC of A32 is defined in ARM DDI 0487D.a, F5.1.247: 1714 * 1715 * b'31 b'28 b'27 b'24 1716 * +---------+---------+-------------------------+ 1717 * | !1111 | 1 1 1 1 | imm24 | 1718 * +---------+---------+-------------------------+ 1719 */ 1720 addr = end_addr - 4; 1721 cs_etm__mem_access(etmq, trace_chan_id, addr, 1722 sizeof(instr32), (u8 *)&instr32); 1723 if ((instr32 & 0x0F000000) == 0x0F000000 && 1724 (instr32 & 0xF0000000) != 0xF0000000) 1725 return true; 1726 1727 break; 1728 case CS_ETM_ISA_A64: 1729 /* 1730 * The SVC of A64 is defined in ARM DDI 0487D.a, C6.2.294: 1731 * 1732 * b'31 b'21 b'4 b'0 1733 * +-----------------------+---------+-----------+ 1734 * | 1 1 0 1 0 1 0 0 0 0 0 | imm16 | 0 0 0 0 1 | 1735 * +-----------------------+---------+-----------+ 1736 */ 1737 addr = end_addr - 4; 1738 cs_etm__mem_access(etmq, trace_chan_id, addr, 1739 sizeof(instr32), (u8 *)&instr32); 1740 if ((instr32 & 0xFFE0001F) == 0xd4000001) 1741 return true; 1742 1743 break; 1744 case CS_ETM_ISA_UNKNOWN: 1745 default: 1746 break; 1747 } 1748 1749 return false; 1750 } 1751 1752 static bool cs_etm__is_syscall(struct cs_etm_queue *etmq, 1753 struct cs_etm_traceid_queue *tidq, u64 magic) 1754 { 1755 u8 trace_chan_id = tidq->trace_chan_id; 1756 struct cs_etm_packet *packet = tidq->packet; 1757 struct cs_etm_packet *prev_packet = tidq->prev_packet; 1758 1759 if (magic == __perf_cs_etmv3_magic) 1760 if (packet->exception_number == CS_ETMV3_EXC_SVC) 1761 return true; 1762 1763 /* 1764 * ETMv4 exception type CS_ETMV4_EXC_CALL covers SVC, SMC and 1765 * HVC cases; need to check if it's SVC instruction based on 1766 * packet address. 1767 */ 1768 if (magic == __perf_cs_etmv4_magic) { 1769 if (packet->exception_number == CS_ETMV4_EXC_CALL && 1770 cs_etm__is_svc_instr(etmq, trace_chan_id, prev_packet, 1771 prev_packet->end_addr)) 1772 return true; 1773 } 1774 1775 return false; 1776 } 1777 1778 static bool cs_etm__is_async_exception(struct cs_etm_traceid_queue *tidq, 1779 u64 magic) 1780 { 1781 struct cs_etm_packet *packet = tidq->packet; 1782 1783 if (magic == __perf_cs_etmv3_magic) 1784 if (packet->exception_number == CS_ETMV3_EXC_DEBUG_HALT || 1785 packet->exception_number == CS_ETMV3_EXC_ASYNC_DATA_ABORT || 1786 packet->exception_number == CS_ETMV3_EXC_PE_RESET || 1787 packet->exception_number == CS_ETMV3_EXC_IRQ || 1788 packet->exception_number == CS_ETMV3_EXC_FIQ) 1789 return true; 1790 1791 if (magic == __perf_cs_etmv4_magic) 1792 if (packet->exception_number == CS_ETMV4_EXC_RESET || 1793 packet->exception_number == CS_ETMV4_EXC_DEBUG_HALT || 1794 packet->exception_number == CS_ETMV4_EXC_SYSTEM_ERROR || 1795 packet->exception_number == CS_ETMV4_EXC_INST_DEBUG || 1796 packet->exception_number == CS_ETMV4_EXC_DATA_DEBUG || 1797 packet->exception_number == CS_ETMV4_EXC_IRQ || 1798 packet->exception_number == CS_ETMV4_EXC_FIQ) 1799 return true; 1800 1801 return false; 1802 } 1803 1804 static bool cs_etm__is_sync_exception(struct cs_etm_queue *etmq, 1805 struct cs_etm_traceid_queue *tidq, 1806 u64 magic) 1807 { 1808 u8 trace_chan_id = tidq->trace_chan_id; 1809 struct cs_etm_packet *packet = tidq->packet; 1810 struct cs_etm_packet *prev_packet = tidq->prev_packet; 1811 1812 if (magic == __perf_cs_etmv3_magic) 1813 if (packet->exception_number == CS_ETMV3_EXC_SMC || 1814 packet->exception_number == CS_ETMV3_EXC_HYP || 1815 packet->exception_number == CS_ETMV3_EXC_JAZELLE_THUMBEE || 1816 packet->exception_number == CS_ETMV3_EXC_UNDEFINED_INSTR || 1817 packet->exception_number == CS_ETMV3_EXC_PREFETCH_ABORT || 1818 packet->exception_number == CS_ETMV3_EXC_DATA_FAULT || 1819 packet->exception_number == CS_ETMV3_EXC_GENERIC) 1820 return true; 1821 1822 if (magic == __perf_cs_etmv4_magic) { 1823 if (packet->exception_number == CS_ETMV4_EXC_TRAP || 1824 packet->exception_number == CS_ETMV4_EXC_ALIGNMENT || 1825 packet->exception_number == CS_ETMV4_EXC_INST_FAULT || 1826 packet->exception_number == CS_ETMV4_EXC_DATA_FAULT) 1827 return true; 1828 1829 /* 1830 * For CS_ETMV4_EXC_CALL, except SVC other instructions 1831 * (SMC, HVC) are taken as sync exceptions. 1832 */ 1833 if (packet->exception_number == CS_ETMV4_EXC_CALL && 1834 !cs_etm__is_svc_instr(etmq, trace_chan_id, prev_packet, 1835 prev_packet->end_addr)) 1836 return true; 1837 1838 /* 1839 * ETMv4 has 5 bits for exception number; if the numbers 1840 * are in the range ( CS_ETMV4_EXC_FIQ, CS_ETMV4_EXC_END ] 1841 * they are implementation defined exceptions. 1842 * 1843 * For this case, simply take it as sync exception. 1844 */ 1845 if (packet->exception_number > CS_ETMV4_EXC_FIQ && 1846 packet->exception_number <= CS_ETMV4_EXC_END) 1847 return true; 1848 } 1849 1850 return false; 1851 } 1852 1853 static int cs_etm__set_sample_flags(struct cs_etm_queue *etmq, 1854 struct cs_etm_traceid_queue *tidq) 1855 { 1856 struct cs_etm_packet *packet = tidq->packet; 1857 struct cs_etm_packet *prev_packet = tidq->prev_packet; 1858 u8 trace_chan_id = tidq->trace_chan_id; 1859 u64 magic; 1860 int ret; 1861 1862 switch (packet->sample_type) { 1863 case CS_ETM_RANGE: 1864 /* 1865 * Immediate branch instruction without neither link nor 1866 * return flag, it's normal branch instruction within 1867 * the function. 1868 */ 1869 if (packet->last_instr_type == OCSD_INSTR_BR && 1870 packet->last_instr_subtype == OCSD_S_INSTR_NONE) { 1871 packet->flags = PERF_IP_FLAG_BRANCH; 1872 1873 if (packet->last_instr_cond) 1874 packet->flags |= PERF_IP_FLAG_CONDITIONAL; 1875 } 1876 1877 /* 1878 * Immediate branch instruction with link (e.g. BL), this is 1879 * branch instruction for function call. 1880 */ 1881 if (packet->last_instr_type == OCSD_INSTR_BR && 1882 packet->last_instr_subtype == OCSD_S_INSTR_BR_LINK) 1883 packet->flags = PERF_IP_FLAG_BRANCH | 1884 PERF_IP_FLAG_CALL; 1885 1886 /* 1887 * Indirect branch instruction with link (e.g. BLR), this is 1888 * branch instruction for function call. 1889 */ 1890 if (packet->last_instr_type == OCSD_INSTR_BR_INDIRECT && 1891 packet->last_instr_subtype == OCSD_S_INSTR_BR_LINK) 1892 packet->flags = PERF_IP_FLAG_BRANCH | 1893 PERF_IP_FLAG_CALL; 1894 1895 /* 1896 * Indirect branch instruction with subtype of 1897 * OCSD_S_INSTR_V7_IMPLIED_RET, this is explicit hint for 1898 * function return for A32/T32. 1899 */ 1900 if (packet->last_instr_type == OCSD_INSTR_BR_INDIRECT && 1901 packet->last_instr_subtype == OCSD_S_INSTR_V7_IMPLIED_RET) 1902 packet->flags = PERF_IP_FLAG_BRANCH | 1903 PERF_IP_FLAG_RETURN; 1904 1905 /* 1906 * Indirect branch instruction without link (e.g. BR), usually 1907 * this is used for function return, especially for functions 1908 * within dynamic link lib. 1909 */ 1910 if (packet->last_instr_type == OCSD_INSTR_BR_INDIRECT && 1911 packet->last_instr_subtype == OCSD_S_INSTR_NONE) 1912 packet->flags = PERF_IP_FLAG_BRANCH | 1913 PERF_IP_FLAG_RETURN; 1914 1915 /* Return instruction for function return. */ 1916 if (packet->last_instr_type == OCSD_INSTR_BR_INDIRECT && 1917 packet->last_instr_subtype == OCSD_S_INSTR_V8_RET) 1918 packet->flags = PERF_IP_FLAG_BRANCH | 1919 PERF_IP_FLAG_RETURN; 1920 1921 /* 1922 * Decoder might insert a discontinuity in the middle of 1923 * instruction packets, fixup prev_packet with flag 1924 * PERF_IP_FLAG_TRACE_BEGIN to indicate restarting trace. 1925 */ 1926 if (prev_packet->sample_type == CS_ETM_DISCONTINUITY) 1927 prev_packet->flags |= PERF_IP_FLAG_BRANCH | 1928 PERF_IP_FLAG_TRACE_BEGIN; 1929 1930 /* 1931 * If the previous packet is an exception return packet 1932 * and the return address just follows SVC instruction, 1933 * it needs to calibrate the previous packet sample flags 1934 * as PERF_IP_FLAG_SYSCALLRET. 1935 */ 1936 if (prev_packet->flags == (PERF_IP_FLAG_BRANCH | 1937 PERF_IP_FLAG_RETURN | 1938 PERF_IP_FLAG_INTERRUPT) && 1939 cs_etm__is_svc_instr(etmq, trace_chan_id, 1940 packet, packet->start_addr)) 1941 prev_packet->flags = PERF_IP_FLAG_BRANCH | 1942 PERF_IP_FLAG_RETURN | 1943 PERF_IP_FLAG_SYSCALLRET; 1944 break; 1945 case CS_ETM_DISCONTINUITY: 1946 /* 1947 * The trace is discontinuous, if the previous packet is 1948 * instruction packet, set flag PERF_IP_FLAG_TRACE_END 1949 * for previous packet. 1950 */ 1951 if (prev_packet->sample_type == CS_ETM_RANGE) 1952 prev_packet->flags |= PERF_IP_FLAG_BRANCH | 1953 PERF_IP_FLAG_TRACE_END; 1954 break; 1955 case CS_ETM_EXCEPTION: 1956 ret = cs_etm__get_magic(packet->trace_chan_id, &magic); 1957 if (ret) 1958 return ret; 1959 1960 /* The exception is for system call. */ 1961 if (cs_etm__is_syscall(etmq, tidq, magic)) 1962 packet->flags = PERF_IP_FLAG_BRANCH | 1963 PERF_IP_FLAG_CALL | 1964 PERF_IP_FLAG_SYSCALLRET; 1965 /* 1966 * The exceptions are triggered by external signals from bus, 1967 * interrupt controller, debug module, PE reset or halt. 1968 */ 1969 else if (cs_etm__is_async_exception(tidq, magic)) 1970 packet->flags = PERF_IP_FLAG_BRANCH | 1971 PERF_IP_FLAG_CALL | 1972 PERF_IP_FLAG_ASYNC | 1973 PERF_IP_FLAG_INTERRUPT; 1974 /* 1975 * Otherwise, exception is caused by trap, instruction & 1976 * data fault, or alignment errors. 1977 */ 1978 else if (cs_etm__is_sync_exception(etmq, tidq, magic)) 1979 packet->flags = PERF_IP_FLAG_BRANCH | 1980 PERF_IP_FLAG_CALL | 1981 PERF_IP_FLAG_INTERRUPT; 1982 1983 /* 1984 * When the exception packet is inserted, since exception 1985 * packet is not used standalone for generating samples 1986 * and it's affiliation to the previous instruction range 1987 * packet; so set previous range packet flags to tell perf 1988 * it is an exception taken branch. 1989 */ 1990 if (prev_packet->sample_type == CS_ETM_RANGE) 1991 prev_packet->flags = packet->flags; 1992 break; 1993 case CS_ETM_EXCEPTION_RET: 1994 /* 1995 * When the exception return packet is inserted, since 1996 * exception return packet is not used standalone for 1997 * generating samples and it's affiliation to the previous 1998 * instruction range packet; so set previous range packet 1999 * flags to tell perf it is an exception return branch. 2000 * 2001 * The exception return can be for either system call or 2002 * other exception types; unfortunately the packet doesn't 2003 * contain exception type related info so we cannot decide 2004 * the exception type purely based on exception return packet. 2005 * If we record the exception number from exception packet and 2006 * reuse it for exception return packet, this is not reliable 2007 * due the trace can be discontinuity or the interrupt can 2008 * be nested, thus the recorded exception number cannot be 2009 * used for exception return packet for these two cases. 2010 * 2011 * For exception return packet, we only need to distinguish the 2012 * packet is for system call or for other types. Thus the 2013 * decision can be deferred when receive the next packet which 2014 * contains the return address, based on the return address we 2015 * can read out the previous instruction and check if it's a 2016 * system call instruction and then calibrate the sample flag 2017 * as needed. 2018 */ 2019 if (prev_packet->sample_type == CS_ETM_RANGE) 2020 prev_packet->flags = PERF_IP_FLAG_BRANCH | 2021 PERF_IP_FLAG_RETURN | 2022 PERF_IP_FLAG_INTERRUPT; 2023 break; 2024 case CS_ETM_EMPTY: 2025 default: 2026 break; 2027 } 2028 2029 return 0; 2030 } 2031 2032 static int cs_etm__decode_data_block(struct cs_etm_queue *etmq) 2033 { 2034 int ret = 0; 2035 size_t processed = 0; 2036 2037 /* 2038 * Packets are decoded and added to the decoder's packet queue 2039 * until the decoder packet processing callback has requested that 2040 * processing stops or there is nothing left in the buffer. Normal 2041 * operations that stop processing are a timestamp packet or a full 2042 * decoder buffer queue. 2043 */ 2044 ret = cs_etm_decoder__process_data_block(etmq->decoder, 2045 etmq->offset, 2046 &etmq->buf[etmq->buf_used], 2047 etmq->buf_len, 2048 &processed); 2049 if (ret) 2050 goto out; 2051 2052 etmq->offset += processed; 2053 etmq->buf_used += processed; 2054 etmq->buf_len -= processed; 2055 2056 out: 2057 return ret; 2058 } 2059 2060 static int cs_etm__process_traceid_queue(struct cs_etm_queue *etmq, 2061 struct cs_etm_traceid_queue *tidq) 2062 { 2063 int ret; 2064 struct cs_etm_packet_queue *packet_queue; 2065 2066 packet_queue = &tidq->packet_queue; 2067 2068 /* Process each packet in this chunk */ 2069 while (1) { 2070 ret = cs_etm_decoder__get_packet(packet_queue, 2071 tidq->packet); 2072 if (ret <= 0) 2073 /* 2074 * Stop processing this chunk on 2075 * end of data or error 2076 */ 2077 break; 2078 2079 /* 2080 * Since packet addresses are swapped in packet 2081 * handling within below switch() statements, 2082 * thus setting sample flags must be called 2083 * prior to switch() statement to use address 2084 * information before packets swapping. 2085 */ 2086 ret = cs_etm__set_sample_flags(etmq, tidq); 2087 if (ret < 0) 2088 break; 2089 2090 switch (tidq->packet->sample_type) { 2091 case CS_ETM_RANGE: 2092 /* 2093 * If the packet contains an instruction 2094 * range, generate instruction sequence 2095 * events. 2096 */ 2097 cs_etm__sample(etmq, tidq); 2098 break; 2099 case CS_ETM_EXCEPTION: 2100 case CS_ETM_EXCEPTION_RET: 2101 /* 2102 * If the exception packet is coming, 2103 * make sure the previous instruction 2104 * range packet to be handled properly. 2105 */ 2106 cs_etm__exception(tidq); 2107 break; 2108 case CS_ETM_DISCONTINUITY: 2109 /* 2110 * Discontinuity in trace, flush 2111 * previous branch stack 2112 */ 2113 cs_etm__flush(etmq, tidq); 2114 break; 2115 case CS_ETM_EMPTY: 2116 /* 2117 * Should not receive empty packet, 2118 * report error. 2119 */ 2120 pr_err("CS ETM Trace: empty packet\n"); 2121 return -EINVAL; 2122 default: 2123 break; 2124 } 2125 } 2126 2127 return ret; 2128 } 2129 2130 static void cs_etm__clear_all_traceid_queues(struct cs_etm_queue *etmq) 2131 { 2132 int idx; 2133 struct int_node *inode; 2134 struct cs_etm_traceid_queue *tidq; 2135 struct intlist *traceid_queues_list = etmq->traceid_queues_list; 2136 2137 intlist__for_each_entry(inode, traceid_queues_list) { 2138 idx = (int)(intptr_t)inode->priv; 2139 tidq = etmq->traceid_queues[idx]; 2140 2141 /* Ignore return value */ 2142 cs_etm__process_traceid_queue(etmq, tidq); 2143 2144 /* 2145 * Generate an instruction sample with the remaining 2146 * branchstack entries. 2147 */ 2148 cs_etm__flush(etmq, tidq); 2149 } 2150 } 2151 2152 static int cs_etm__run_decoder(struct cs_etm_queue *etmq) 2153 { 2154 int err = 0; 2155 struct cs_etm_traceid_queue *tidq; 2156 2157 tidq = cs_etm__etmq_get_traceid_queue(etmq, CS_ETM_PER_THREAD_TRACEID); 2158 if (!tidq) 2159 return -EINVAL; 2160 2161 /* Go through each buffer in the queue and decode them one by one */ 2162 while (1) { 2163 err = cs_etm__get_data_block(etmq); 2164 if (err <= 0) 2165 return err; 2166 2167 /* Run trace decoder until buffer consumed or end of trace */ 2168 do { 2169 err = cs_etm__decode_data_block(etmq); 2170 if (err) 2171 return err; 2172 2173 /* 2174 * Process each packet in this chunk, nothing to do if 2175 * an error occurs other than hoping the next one will 2176 * be better. 2177 */ 2178 err = cs_etm__process_traceid_queue(etmq, tidq); 2179 2180 } while (etmq->buf_len); 2181 2182 if (err == 0) 2183 /* Flush any remaining branch stack entries */ 2184 err = cs_etm__end_block(etmq, tidq); 2185 } 2186 2187 return err; 2188 } 2189 2190 static int cs_etm__process_timeless_queues(struct cs_etm_auxtrace *etm, 2191 pid_t tid) 2192 { 2193 unsigned int i; 2194 struct auxtrace_queues *queues = &etm->queues; 2195 2196 for (i = 0; i < queues->nr_queues; i++) { 2197 struct auxtrace_queue *queue = &etm->queues.queue_array[i]; 2198 struct cs_etm_queue *etmq = queue->priv; 2199 struct cs_etm_traceid_queue *tidq; 2200 2201 if (!etmq) 2202 continue; 2203 2204 tidq = cs_etm__etmq_get_traceid_queue(etmq, 2205 CS_ETM_PER_THREAD_TRACEID); 2206 2207 if (!tidq) 2208 continue; 2209 2210 if ((tid == -1) || (tidq->tid == tid)) { 2211 cs_etm__set_pid_tid_cpu(etm, tidq); 2212 cs_etm__run_decoder(etmq); 2213 } 2214 } 2215 2216 return 0; 2217 } 2218 2219 static int cs_etm__process_queues(struct cs_etm_auxtrace *etm) 2220 { 2221 int ret = 0; 2222 unsigned int cs_queue_nr, queue_nr; 2223 u8 trace_chan_id; 2224 u64 timestamp; 2225 struct auxtrace_queue *queue; 2226 struct cs_etm_queue *etmq; 2227 struct cs_etm_traceid_queue *tidq; 2228 2229 while (1) { 2230 if (!etm->heap.heap_cnt) 2231 goto out; 2232 2233 /* Take the entry at the top of the min heap */ 2234 cs_queue_nr = etm->heap.heap_array[0].queue_nr; 2235 queue_nr = TO_QUEUE_NR(cs_queue_nr); 2236 trace_chan_id = TO_TRACE_CHAN_ID(cs_queue_nr); 2237 queue = &etm->queues.queue_array[queue_nr]; 2238 etmq = queue->priv; 2239 2240 /* 2241 * Remove the top entry from the heap since we are about 2242 * to process it. 2243 */ 2244 auxtrace_heap__pop(&etm->heap); 2245 2246 tidq = cs_etm__etmq_get_traceid_queue(etmq, trace_chan_id); 2247 if (!tidq) { 2248 /* 2249 * No traceID queue has been allocated for this traceID, 2250 * which means something somewhere went very wrong. No 2251 * other choice than simply exit. 2252 */ 2253 ret = -EINVAL; 2254 goto out; 2255 } 2256 2257 /* 2258 * Packets associated with this timestamp are already in 2259 * the etmq's traceID queue, so process them. 2260 */ 2261 ret = cs_etm__process_traceid_queue(etmq, tidq); 2262 if (ret < 0) 2263 goto out; 2264 2265 /* 2266 * Packets for this timestamp have been processed, time to 2267 * move on to the next timestamp, fetching a new auxtrace_buffer 2268 * if need be. 2269 */ 2270 refetch: 2271 ret = cs_etm__get_data_block(etmq); 2272 if (ret < 0) 2273 goto out; 2274 2275 /* 2276 * No more auxtrace_buffers to process in this etmq, simply 2277 * move on to another entry in the auxtrace_heap. 2278 */ 2279 if (!ret) 2280 continue; 2281 2282 ret = cs_etm__decode_data_block(etmq); 2283 if (ret) 2284 goto out; 2285 2286 timestamp = cs_etm__etmq_get_timestamp(etmq, &trace_chan_id); 2287 2288 if (!timestamp) { 2289 /* 2290 * Function cs_etm__decode_data_block() returns when 2291 * there is no more traces to decode in the current 2292 * auxtrace_buffer OR when a timestamp has been 2293 * encountered on any of the traceID queues. Since we 2294 * did not get a timestamp, there is no more traces to 2295 * process in this auxtrace_buffer. As such empty and 2296 * flush all traceID queues. 2297 */ 2298 cs_etm__clear_all_traceid_queues(etmq); 2299 2300 /* Fetch another auxtrace_buffer for this etmq */ 2301 goto refetch; 2302 } 2303 2304 /* 2305 * Add to the min heap the timestamp for packets that have 2306 * just been decoded. They will be processed and synthesized 2307 * during the next call to cs_etm__process_traceid_queue() for 2308 * this queue/traceID. 2309 */ 2310 cs_queue_nr = TO_CS_QUEUE_NR(queue_nr, trace_chan_id); 2311 ret = auxtrace_heap__add(&etm->heap, cs_queue_nr, timestamp); 2312 } 2313 2314 out: 2315 return ret; 2316 } 2317 2318 static int cs_etm__process_itrace_start(struct cs_etm_auxtrace *etm, 2319 union perf_event *event) 2320 { 2321 struct thread *th; 2322 2323 if (etm->timeless_decoding) 2324 return 0; 2325 2326 /* 2327 * Add the tid/pid to the log so that we can get a match when 2328 * we get a contextID from the decoder. 2329 */ 2330 th = machine__findnew_thread(etm->machine, 2331 event->itrace_start.pid, 2332 event->itrace_start.tid); 2333 if (!th) 2334 return -ENOMEM; 2335 2336 thread__put(th); 2337 2338 return 0; 2339 } 2340 2341 static int cs_etm__process_switch_cpu_wide(struct cs_etm_auxtrace *etm, 2342 union perf_event *event) 2343 { 2344 struct thread *th; 2345 bool out = event->header.misc & PERF_RECORD_MISC_SWITCH_OUT; 2346 2347 /* 2348 * Context switch in per-thread mode are irrelevant since perf 2349 * will start/stop tracing as the process is scheduled. 2350 */ 2351 if (etm->timeless_decoding) 2352 return 0; 2353 2354 /* 2355 * SWITCH_IN events carry the next process to be switched out while 2356 * SWITCH_OUT events carry the process to be switched in. As such 2357 * we don't care about IN events. 2358 */ 2359 if (!out) 2360 return 0; 2361 2362 /* 2363 * Add the tid/pid to the log so that we can get a match when 2364 * we get a contextID from the decoder. 2365 */ 2366 th = machine__findnew_thread(etm->machine, 2367 event->context_switch.next_prev_pid, 2368 event->context_switch.next_prev_tid); 2369 if (!th) 2370 return -ENOMEM; 2371 2372 thread__put(th); 2373 2374 return 0; 2375 } 2376 2377 static int cs_etm__process_event(struct perf_session *session, 2378 union perf_event *event, 2379 struct perf_sample *sample, 2380 struct perf_tool *tool) 2381 { 2382 int err = 0; 2383 u64 timestamp; 2384 struct cs_etm_auxtrace *etm = container_of(session->auxtrace, 2385 struct cs_etm_auxtrace, 2386 auxtrace); 2387 2388 if (dump_trace) 2389 return 0; 2390 2391 if (!tool->ordered_events) { 2392 pr_err("CoreSight ETM Trace requires ordered events\n"); 2393 return -EINVAL; 2394 } 2395 2396 if (sample->time && (sample->time != (u64) -1)) 2397 timestamp = sample->time; 2398 else 2399 timestamp = 0; 2400 2401 if (timestamp || etm->timeless_decoding) { 2402 err = cs_etm__update_queues(etm); 2403 if (err) 2404 return err; 2405 } 2406 2407 if (etm->timeless_decoding && 2408 event->header.type == PERF_RECORD_EXIT) 2409 return cs_etm__process_timeless_queues(etm, 2410 event->fork.tid); 2411 2412 if (event->header.type == PERF_RECORD_ITRACE_START) 2413 return cs_etm__process_itrace_start(etm, event); 2414 else if (event->header.type == PERF_RECORD_SWITCH_CPU_WIDE) 2415 return cs_etm__process_switch_cpu_wide(etm, event); 2416 2417 if (!etm->timeless_decoding && 2418 event->header.type == PERF_RECORD_AUX) 2419 return cs_etm__process_queues(etm); 2420 2421 return 0; 2422 } 2423 2424 static int cs_etm__process_auxtrace_event(struct perf_session *session, 2425 union perf_event *event, 2426 struct perf_tool *tool __maybe_unused) 2427 { 2428 struct cs_etm_auxtrace *etm = container_of(session->auxtrace, 2429 struct cs_etm_auxtrace, 2430 auxtrace); 2431 if (!etm->data_queued) { 2432 struct auxtrace_buffer *buffer; 2433 off_t data_offset; 2434 int fd = perf_data__fd(session->data); 2435 bool is_pipe = perf_data__is_pipe(session->data); 2436 int err; 2437 2438 if (is_pipe) 2439 data_offset = 0; 2440 else { 2441 data_offset = lseek(fd, 0, SEEK_CUR); 2442 if (data_offset == -1) 2443 return -errno; 2444 } 2445 2446 err = auxtrace_queues__add_event(&etm->queues, session, 2447 event, data_offset, &buffer); 2448 if (err) 2449 return err; 2450 2451 if (dump_trace) 2452 if (auxtrace_buffer__get_data(buffer, fd)) { 2453 cs_etm__dump_event(etm, buffer); 2454 auxtrace_buffer__put_data(buffer); 2455 } 2456 } 2457 2458 return 0; 2459 } 2460 2461 static bool cs_etm__is_timeless_decoding(struct cs_etm_auxtrace *etm) 2462 { 2463 struct evsel *evsel; 2464 struct evlist *evlist = etm->session->evlist; 2465 bool timeless_decoding = true; 2466 2467 /* 2468 * Circle through the list of event and complain if we find one 2469 * with the time bit set. 2470 */ 2471 evlist__for_each_entry(evlist, evsel) { 2472 if ((evsel->core.attr.sample_type & PERF_SAMPLE_TIME)) 2473 timeless_decoding = false; 2474 } 2475 2476 return timeless_decoding; 2477 } 2478 2479 static const char * const cs_etm_global_header_fmts[] = { 2480 [CS_HEADER_VERSION] = " Header version %llx\n", 2481 [CS_PMU_TYPE_CPUS] = " PMU type/num cpus %llx\n", 2482 [CS_ETM_SNAPSHOT] = " Snapshot %llx\n", 2483 }; 2484 2485 static const char * const cs_etm_priv_fmts[] = { 2486 [CS_ETM_MAGIC] = " Magic number %llx\n", 2487 [CS_ETM_CPU] = " CPU %lld\n", 2488 [CS_ETM_NR_TRC_PARAMS] = " NR_TRC_PARAMS %llx\n", 2489 [CS_ETM_ETMCR] = " ETMCR %llx\n", 2490 [CS_ETM_ETMTRACEIDR] = " ETMTRACEIDR %llx\n", 2491 [CS_ETM_ETMCCER] = " ETMCCER %llx\n", 2492 [CS_ETM_ETMIDR] = " ETMIDR %llx\n", 2493 }; 2494 2495 static const char * const cs_etmv4_priv_fmts[] = { 2496 [CS_ETM_MAGIC] = " Magic number %llx\n", 2497 [CS_ETM_CPU] = " CPU %lld\n", 2498 [CS_ETM_NR_TRC_PARAMS] = " NR_TRC_PARAMS %llx\n", 2499 [CS_ETMV4_TRCCONFIGR] = " TRCCONFIGR %llx\n", 2500 [CS_ETMV4_TRCTRACEIDR] = " TRCTRACEIDR %llx\n", 2501 [CS_ETMV4_TRCIDR0] = " TRCIDR0 %llx\n", 2502 [CS_ETMV4_TRCIDR1] = " TRCIDR1 %llx\n", 2503 [CS_ETMV4_TRCIDR2] = " TRCIDR2 %llx\n", 2504 [CS_ETMV4_TRCIDR8] = " TRCIDR8 %llx\n", 2505 [CS_ETMV4_TRCAUTHSTATUS] = " TRCAUTHSTATUS %llx\n", 2506 }; 2507 2508 static const char * const param_unk_fmt = 2509 " Unknown parameter [%d] %llx\n"; 2510 static const char * const magic_unk_fmt = 2511 " Magic number Unknown %llx\n"; 2512 2513 static int cs_etm__print_cpu_metadata_v0(__u64 *val, int *offset) 2514 { 2515 int i = *offset, j, nr_params = 0, fmt_offset; 2516 __u64 magic; 2517 2518 /* check magic value */ 2519 magic = val[i + CS_ETM_MAGIC]; 2520 if ((magic != __perf_cs_etmv3_magic) && 2521 (magic != __perf_cs_etmv4_magic)) { 2522 /* failure - note bad magic value */ 2523 fprintf(stdout, magic_unk_fmt, magic); 2524 return -EINVAL; 2525 } 2526 2527 /* print common header block */ 2528 fprintf(stdout, cs_etm_priv_fmts[CS_ETM_MAGIC], val[i++]); 2529 fprintf(stdout, cs_etm_priv_fmts[CS_ETM_CPU], val[i++]); 2530 2531 if (magic == __perf_cs_etmv3_magic) { 2532 nr_params = CS_ETM_NR_TRC_PARAMS_V0; 2533 fmt_offset = CS_ETM_ETMCR; 2534 /* after common block, offset format index past NR_PARAMS */ 2535 for (j = fmt_offset; j < nr_params + fmt_offset; j++, i++) 2536 fprintf(stdout, cs_etm_priv_fmts[j], val[i]); 2537 } else if (magic == __perf_cs_etmv4_magic) { 2538 nr_params = CS_ETMV4_NR_TRC_PARAMS_V0; 2539 fmt_offset = CS_ETMV4_TRCCONFIGR; 2540 /* after common block, offset format index past NR_PARAMS */ 2541 for (j = fmt_offset; j < nr_params + fmt_offset; j++, i++) 2542 fprintf(stdout, cs_etmv4_priv_fmts[j], val[i]); 2543 } 2544 *offset = i; 2545 return 0; 2546 } 2547 2548 static int cs_etm__print_cpu_metadata_v1(__u64 *val, int *offset) 2549 { 2550 int i = *offset, j, total_params = 0; 2551 __u64 magic; 2552 2553 magic = val[i + CS_ETM_MAGIC]; 2554 /* total params to print is NR_PARAMS + common block size for v1 */ 2555 total_params = val[i + CS_ETM_NR_TRC_PARAMS] + CS_ETM_COMMON_BLK_MAX_V1; 2556 2557 if (magic == __perf_cs_etmv3_magic) { 2558 for (j = 0; j < total_params; j++, i++) { 2559 /* if newer record - could be excess params */ 2560 if (j >= CS_ETM_PRIV_MAX) 2561 fprintf(stdout, param_unk_fmt, j, val[i]); 2562 else 2563 fprintf(stdout, cs_etm_priv_fmts[j], val[i]); 2564 } 2565 } else if (magic == __perf_cs_etmv4_magic) { 2566 for (j = 0; j < total_params; j++, i++) { 2567 /* if newer record - could be excess params */ 2568 if (j >= CS_ETMV4_PRIV_MAX) 2569 fprintf(stdout, param_unk_fmt, j, val[i]); 2570 else 2571 fprintf(stdout, cs_etmv4_priv_fmts[j], val[i]); 2572 } 2573 } else { 2574 /* failure - note bad magic value and error out */ 2575 fprintf(stdout, magic_unk_fmt, magic); 2576 return -EINVAL; 2577 } 2578 *offset = i; 2579 return 0; 2580 } 2581 2582 static void cs_etm__print_auxtrace_info(__u64 *val, int num) 2583 { 2584 int i, cpu = 0, version, err; 2585 2586 /* bail out early on bad header version */ 2587 version = val[0]; 2588 if (version > CS_HEADER_CURRENT_VERSION) { 2589 /* failure.. return */ 2590 fprintf(stdout, " Unknown Header Version = %x, ", version); 2591 fprintf(stdout, "Version supported <= %x\n", CS_HEADER_CURRENT_VERSION); 2592 return; 2593 } 2594 2595 for (i = 0; i < CS_HEADER_VERSION_MAX; i++) 2596 fprintf(stdout, cs_etm_global_header_fmts[i], val[i]); 2597 2598 for (i = CS_HEADER_VERSION_MAX; cpu < num; cpu++) { 2599 if (version == 0) 2600 err = cs_etm__print_cpu_metadata_v0(val, &i); 2601 else if (version == 1) 2602 err = cs_etm__print_cpu_metadata_v1(val, &i); 2603 if (err) 2604 return; 2605 } 2606 } 2607 2608 /* 2609 * Read a single cpu parameter block from the auxtrace_info priv block. 2610 * 2611 * For version 1 there is a per cpu nr_params entry. If we are handling 2612 * version 1 file, then there may be less, the same, or more params 2613 * indicated by this value than the compile time number we understand. 2614 * 2615 * For a version 0 info block, there are a fixed number, and we need to 2616 * fill out the nr_param value in the metadata we create. 2617 */ 2618 static u64 *cs_etm__create_meta_blk(u64 *buff_in, int *buff_in_offset, 2619 int out_blk_size, int nr_params_v0) 2620 { 2621 u64 *metadata = NULL; 2622 int hdr_version; 2623 int nr_in_params, nr_out_params, nr_cmn_params; 2624 int i, k; 2625 2626 metadata = zalloc(sizeof(*metadata) * out_blk_size); 2627 if (!metadata) 2628 return NULL; 2629 2630 /* read block current index & version */ 2631 i = *buff_in_offset; 2632 hdr_version = buff_in[CS_HEADER_VERSION]; 2633 2634 if (!hdr_version) { 2635 /* read version 0 info block into a version 1 metadata block */ 2636 nr_in_params = nr_params_v0; 2637 metadata[CS_ETM_MAGIC] = buff_in[i + CS_ETM_MAGIC]; 2638 metadata[CS_ETM_CPU] = buff_in[i + CS_ETM_CPU]; 2639 metadata[CS_ETM_NR_TRC_PARAMS] = nr_in_params; 2640 /* remaining block params at offset +1 from source */ 2641 for (k = CS_ETM_COMMON_BLK_MAX_V1 - 1; k < nr_in_params; k++) 2642 metadata[k + 1] = buff_in[i + k]; 2643 /* version 0 has 2 common params */ 2644 nr_cmn_params = 2; 2645 } else { 2646 /* read version 1 info block - input and output nr_params may differ */ 2647 /* version 1 has 3 common params */ 2648 nr_cmn_params = 3; 2649 nr_in_params = buff_in[i + CS_ETM_NR_TRC_PARAMS]; 2650 2651 /* if input has more params than output - skip excess */ 2652 nr_out_params = nr_in_params + nr_cmn_params; 2653 if (nr_out_params > out_blk_size) 2654 nr_out_params = out_blk_size; 2655 2656 for (k = CS_ETM_MAGIC; k < nr_out_params; k++) 2657 metadata[k] = buff_in[i + k]; 2658 2659 /* record the actual nr params we copied */ 2660 metadata[CS_ETM_NR_TRC_PARAMS] = nr_out_params - nr_cmn_params; 2661 } 2662 2663 /* adjust in offset by number of in params used */ 2664 i += nr_in_params + nr_cmn_params; 2665 *buff_in_offset = i; 2666 return metadata; 2667 } 2668 2669 int cs_etm__process_auxtrace_info(union perf_event *event, 2670 struct perf_session *session) 2671 { 2672 struct perf_record_auxtrace_info *auxtrace_info = &event->auxtrace_info; 2673 struct cs_etm_auxtrace *etm = NULL; 2674 struct int_node *inode; 2675 unsigned int pmu_type; 2676 int event_header_size = sizeof(struct perf_event_header); 2677 int info_header_size; 2678 int total_size = auxtrace_info->header.size; 2679 int priv_size = 0; 2680 int num_cpu, trcidr_idx; 2681 int err = 0; 2682 int i, j; 2683 u64 *ptr, *hdr = NULL; 2684 u64 **metadata = NULL; 2685 u64 hdr_version; 2686 2687 /* 2688 * sizeof(auxtrace_info_event::type) + 2689 * sizeof(auxtrace_info_event::reserved) == 8 2690 */ 2691 info_header_size = 8; 2692 2693 if (total_size < (event_header_size + info_header_size)) 2694 return -EINVAL; 2695 2696 priv_size = total_size - event_header_size - info_header_size; 2697 2698 /* First the global part */ 2699 ptr = (u64 *) auxtrace_info->priv; 2700 2701 /* Look for version of the header */ 2702 hdr_version = ptr[0]; 2703 if (hdr_version > CS_HEADER_CURRENT_VERSION) { 2704 /* print routine will print an error on bad version */ 2705 if (dump_trace) 2706 cs_etm__print_auxtrace_info(auxtrace_info->priv, 0); 2707 return -EINVAL; 2708 } 2709 2710 hdr = zalloc(sizeof(*hdr) * CS_HEADER_VERSION_MAX); 2711 if (!hdr) 2712 return -ENOMEM; 2713 2714 /* Extract header information - see cs-etm.h for format */ 2715 for (i = 0; i < CS_HEADER_VERSION_MAX; i++) 2716 hdr[i] = ptr[i]; 2717 num_cpu = hdr[CS_PMU_TYPE_CPUS] & 0xffffffff; 2718 pmu_type = (unsigned int) ((hdr[CS_PMU_TYPE_CPUS] >> 32) & 2719 0xffffffff); 2720 2721 /* 2722 * Create an RB tree for traceID-metadata tuple. Since the conversion 2723 * has to be made for each packet that gets decoded, optimizing access 2724 * in anything other than a sequential array is worth doing. 2725 */ 2726 traceid_list = intlist__new(NULL); 2727 if (!traceid_list) { 2728 err = -ENOMEM; 2729 goto err_free_hdr; 2730 } 2731 2732 metadata = zalloc(sizeof(*metadata) * num_cpu); 2733 if (!metadata) { 2734 err = -ENOMEM; 2735 goto err_free_traceid_list; 2736 } 2737 2738 /* 2739 * The metadata is stored in the auxtrace_info section and encodes 2740 * the configuration of the ARM embedded trace macrocell which is 2741 * required by the trace decoder to properly decode the trace due 2742 * to its highly compressed nature. 2743 */ 2744 for (j = 0; j < num_cpu; j++) { 2745 if (ptr[i] == __perf_cs_etmv3_magic) { 2746 metadata[j] = 2747 cs_etm__create_meta_blk(ptr, &i, 2748 CS_ETM_PRIV_MAX, 2749 CS_ETM_NR_TRC_PARAMS_V0); 2750 2751 /* The traceID is our handle */ 2752 trcidr_idx = CS_ETM_ETMTRACEIDR; 2753 2754 } else if (ptr[i] == __perf_cs_etmv4_magic) { 2755 metadata[j] = 2756 cs_etm__create_meta_blk(ptr, &i, 2757 CS_ETMV4_PRIV_MAX, 2758 CS_ETMV4_NR_TRC_PARAMS_V0); 2759 2760 /* The traceID is our handle */ 2761 trcidr_idx = CS_ETMV4_TRCTRACEIDR; 2762 } 2763 2764 if (!metadata[j]) { 2765 err = -ENOMEM; 2766 goto err_free_metadata; 2767 } 2768 2769 /* Get an RB node for this CPU */ 2770 inode = intlist__findnew(traceid_list, metadata[j][trcidr_idx]); 2771 2772 /* Something went wrong, no need to continue */ 2773 if (!inode) { 2774 err = -ENOMEM; 2775 goto err_free_metadata; 2776 } 2777 2778 /* 2779 * The node for that CPU should not be taken. 2780 * Back out if that's the case. 2781 */ 2782 if (inode->priv) { 2783 err = -EINVAL; 2784 goto err_free_metadata; 2785 } 2786 /* All good, associate the traceID with the metadata pointer */ 2787 inode->priv = metadata[j]; 2788 } 2789 2790 /* 2791 * Each of CS_HEADER_VERSION_MAX, CS_ETM_PRIV_MAX and 2792 * CS_ETMV4_PRIV_MAX mark how many double words are in the 2793 * global metadata, and each cpu's metadata respectively. 2794 * The following tests if the correct number of double words was 2795 * present in the auxtrace info section. 2796 */ 2797 if (i * 8 != priv_size) { 2798 err = -EINVAL; 2799 goto err_free_metadata; 2800 } 2801 2802 etm = zalloc(sizeof(*etm)); 2803 2804 if (!etm) { 2805 err = -ENOMEM; 2806 goto err_free_metadata; 2807 } 2808 2809 err = auxtrace_queues__init(&etm->queues); 2810 if (err) 2811 goto err_free_etm; 2812 2813 etm->session = session; 2814 etm->machine = &session->machines.host; 2815 2816 etm->num_cpu = num_cpu; 2817 etm->pmu_type = pmu_type; 2818 etm->snapshot_mode = (hdr[CS_ETM_SNAPSHOT] != 0); 2819 etm->metadata = metadata; 2820 etm->auxtrace_type = auxtrace_info->type; 2821 etm->timeless_decoding = cs_etm__is_timeless_decoding(etm); 2822 2823 etm->auxtrace.process_event = cs_etm__process_event; 2824 etm->auxtrace.process_auxtrace_event = cs_etm__process_auxtrace_event; 2825 etm->auxtrace.flush_events = cs_etm__flush_events; 2826 etm->auxtrace.free_events = cs_etm__free_events; 2827 etm->auxtrace.free = cs_etm__free; 2828 etm->auxtrace.evsel_is_auxtrace = cs_etm__evsel_is_auxtrace; 2829 session->auxtrace = &etm->auxtrace; 2830 2831 etm->unknown_thread = thread__new(999999999, 999999999); 2832 if (!etm->unknown_thread) { 2833 err = -ENOMEM; 2834 goto err_free_queues; 2835 } 2836 2837 /* 2838 * Initialize list node so that at thread__zput() we can avoid 2839 * segmentation fault at list_del_init(). 2840 */ 2841 INIT_LIST_HEAD(&etm->unknown_thread->node); 2842 2843 err = thread__set_comm(etm->unknown_thread, "unknown", 0); 2844 if (err) 2845 goto err_delete_thread; 2846 2847 if (thread__init_maps(etm->unknown_thread, etm->machine)) { 2848 err = -ENOMEM; 2849 goto err_delete_thread; 2850 } 2851 2852 if (dump_trace) { 2853 cs_etm__print_auxtrace_info(auxtrace_info->priv, num_cpu); 2854 return 0; 2855 } 2856 2857 if (session->itrace_synth_opts->set) { 2858 etm->synth_opts = *session->itrace_synth_opts; 2859 } else { 2860 itrace_synth_opts__set_default(&etm->synth_opts, 2861 session->itrace_synth_opts->default_no_sample); 2862 etm->synth_opts.callchain = false; 2863 } 2864 2865 err = cs_etm__synth_events(etm, session); 2866 if (err) 2867 goto err_delete_thread; 2868 2869 err = auxtrace_queues__process_index(&etm->queues, session); 2870 if (err) 2871 goto err_delete_thread; 2872 2873 etm->data_queued = etm->queues.populated; 2874 2875 return 0; 2876 2877 err_delete_thread: 2878 thread__zput(etm->unknown_thread); 2879 err_free_queues: 2880 auxtrace_queues__free(&etm->queues); 2881 session->auxtrace = NULL; 2882 err_free_etm: 2883 zfree(&etm); 2884 err_free_metadata: 2885 /* No need to check @metadata[j], free(NULL) is supported */ 2886 for (j = 0; j < num_cpu; j++) 2887 zfree(&metadata[j]); 2888 zfree(&metadata); 2889 err_free_traceid_list: 2890 intlist__delete(traceid_list); 2891 err_free_hdr: 2892 zfree(&hdr); 2893 /* 2894 * At this point, as a minimum we have valid header. Dump the rest of 2895 * the info section - the print routines will error out on structural 2896 * issues. 2897 */ 2898 if (dump_trace) 2899 cs_etm__print_auxtrace_info(auxtrace_info->priv, num_cpu); 2900 return err; 2901 } 2902