1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * thread-stack.c: Synthesize a thread's stack using call / return events 4 * Copyright (c) 2014, Intel Corporation. 5 */ 6 7 #include <linux/rbtree.h> 8 #include <linux/list.h> 9 #include <linux/log2.h> 10 #include <linux/zalloc.h> 11 #include <errno.h> 12 #include <stdlib.h> 13 #include <string.h> 14 #include "thread.h" 15 #include "event.h" 16 #include "machine.h" 17 #include "env.h" 18 #include "debug.h" 19 #include "symbol.h" 20 #include "comm.h" 21 #include "call-path.h" 22 #include "thread-stack.h" 23 24 #define STACK_GROWTH 2048 25 26 /* 27 * State of retpoline detection. 28 * 29 * RETPOLINE_NONE: no retpoline detection 30 * X86_RETPOLINE_POSSIBLE: x86 retpoline possible 31 * X86_RETPOLINE_DETECTED: x86 retpoline detected 32 */ 33 enum retpoline_state_t { 34 RETPOLINE_NONE, 35 X86_RETPOLINE_POSSIBLE, 36 X86_RETPOLINE_DETECTED, 37 }; 38 39 /** 40 * struct thread_stack_entry - thread stack entry. 41 * @ret_addr: return address 42 * @timestamp: timestamp (if known) 43 * @ref: external reference (e.g. db_id of sample) 44 * @branch_count: the branch count when the entry was created 45 * @insn_count: the instruction count when the entry was created 46 * @cyc_count the cycle count when the entry was created 47 * @db_id: id used for db-export 48 * @cp: call path 49 * @no_call: a 'call' was not seen 50 * @trace_end: a 'call' but trace ended 51 * @non_call: a branch but not a 'call' to the start of a different symbol 52 */ 53 struct thread_stack_entry { 54 u64 ret_addr; 55 u64 timestamp; 56 u64 ref; 57 u64 branch_count; 58 u64 insn_count; 59 u64 cyc_count; 60 u64 db_id; 61 struct call_path *cp; 62 bool no_call; 63 bool trace_end; 64 bool non_call; 65 }; 66 67 /** 68 * struct thread_stack - thread stack constructed from 'call' and 'return' 69 * branch samples. 70 * @stack: array that holds the stack 71 * @cnt: number of entries in the stack 72 * @sz: current maximum stack size 73 * @trace_nr: current trace number 74 * @branch_count: running branch count 75 * @insn_count: running instruction count 76 * @cyc_count running cycle count 77 * @kernel_start: kernel start address 78 * @last_time: last timestamp 79 * @crp: call/return processor 80 * @comm: current comm 81 * @arr_sz: size of array if this is the first element of an array 82 * @rstate: used to detect retpolines 83 * @br_stack_rb: branch stack (ring buffer) 84 * @br_stack_sz: maximum branch stack size 85 * @br_stack_pos: current position in @br_stack_rb 86 * @mispred_all: mark all branches as mispredicted 87 */ 88 struct thread_stack { 89 struct thread_stack_entry *stack; 90 size_t cnt; 91 size_t sz; 92 u64 trace_nr; 93 u64 branch_count; 94 u64 insn_count; 95 u64 cyc_count; 96 u64 kernel_start; 97 u64 last_time; 98 struct call_return_processor *crp; 99 struct comm *comm; 100 unsigned int arr_sz; 101 enum retpoline_state_t rstate; 102 struct branch_stack *br_stack_rb; 103 unsigned int br_stack_sz; 104 unsigned int br_stack_pos; 105 bool mispred_all; 106 }; 107 108 /* 109 * Assume pid == tid == 0 identifies the idle task as defined by 110 * perf_session__register_idle_thread(). The idle task is really 1 task per cpu, 111 * and therefore requires a stack for each cpu. 112 */ 113 static inline bool thread_stack__per_cpu(struct thread *thread) 114 { 115 return !(thread__tid(thread) || thread__pid(thread)); 116 } 117 118 static int thread_stack__grow(struct thread_stack *ts) 119 { 120 struct thread_stack_entry *new_stack; 121 size_t sz, new_sz; 122 123 new_sz = ts->sz + STACK_GROWTH; 124 sz = new_sz * sizeof(struct thread_stack_entry); 125 126 new_stack = realloc(ts->stack, sz); 127 if (!new_stack) 128 return -ENOMEM; 129 130 ts->stack = new_stack; 131 ts->sz = new_sz; 132 133 return 0; 134 } 135 136 static int thread_stack__init(struct thread_stack *ts, struct thread *thread, 137 struct call_return_processor *crp, 138 bool callstack, unsigned int br_stack_sz) 139 { 140 int err; 141 142 if (callstack) { 143 err = thread_stack__grow(ts); 144 if (err) 145 return err; 146 } 147 148 if (br_stack_sz) { 149 size_t sz = sizeof(struct branch_stack); 150 151 sz += br_stack_sz * sizeof(struct branch_entry); 152 ts->br_stack_rb = zalloc(sz); 153 if (!ts->br_stack_rb) 154 return -ENOMEM; 155 ts->br_stack_sz = br_stack_sz; 156 } 157 158 if (thread__maps(thread) && maps__machine(thread__maps(thread))) { 159 struct machine *machine = maps__machine(thread__maps(thread)); 160 const char *arch = perf_env__arch(machine->env); 161 162 ts->kernel_start = machine__kernel_start(machine); 163 if (!strcmp(arch, "x86")) 164 ts->rstate = X86_RETPOLINE_POSSIBLE; 165 } else { 166 ts->kernel_start = 1ULL << 63; 167 } 168 ts->crp = crp; 169 170 return 0; 171 } 172 173 static struct thread_stack *thread_stack__new(struct thread *thread, int cpu, 174 struct call_return_processor *crp, 175 bool callstack, 176 unsigned int br_stack_sz) 177 { 178 struct thread_stack *ts = thread__ts(thread), *new_ts; 179 unsigned int old_sz = ts ? ts->arr_sz : 0; 180 unsigned int new_sz = 1; 181 182 if (thread_stack__per_cpu(thread) && cpu > 0) 183 new_sz = roundup_pow_of_two(cpu + 1); 184 185 if (!ts || new_sz > old_sz) { 186 new_ts = calloc(new_sz, sizeof(*ts)); 187 if (!new_ts) 188 return NULL; 189 if (ts) 190 memcpy(new_ts, ts, old_sz * sizeof(*ts)); 191 new_ts->arr_sz = new_sz; 192 free(thread__ts(thread)); 193 thread__set_ts(thread, new_ts); 194 ts = new_ts; 195 } 196 197 if (thread_stack__per_cpu(thread) && cpu > 0 && 198 (unsigned int)cpu < ts->arr_sz) 199 ts += cpu; 200 201 if (!ts->stack && 202 thread_stack__init(ts, thread, crp, callstack, br_stack_sz)) 203 return NULL; 204 205 return ts; 206 } 207 208 static struct thread_stack *thread__cpu_stack(struct thread *thread, int cpu) 209 { 210 struct thread_stack *ts = thread__ts(thread); 211 212 if (cpu < 0) 213 cpu = 0; 214 215 if (!ts || (unsigned int)cpu >= ts->arr_sz) 216 return NULL; 217 218 ts += cpu; 219 220 if (!ts->stack) 221 return NULL; 222 223 return ts; 224 } 225 226 static inline struct thread_stack *thread__stack(struct thread *thread, 227 int cpu) 228 { 229 if (!thread) 230 return NULL; 231 232 if (thread_stack__per_cpu(thread)) 233 return thread__cpu_stack(thread, cpu); 234 235 return thread__ts(thread); 236 } 237 238 static int thread_stack__push(struct thread_stack *ts, u64 ret_addr, 239 bool trace_end) 240 { 241 int err = 0; 242 243 if (ts->cnt == ts->sz) { 244 err = thread_stack__grow(ts); 245 if (err) { 246 pr_warning("Out of memory: discarding thread stack\n"); 247 ts->cnt = 0; 248 } 249 } 250 251 ts->stack[ts->cnt].trace_end = trace_end; 252 ts->stack[ts->cnt++].ret_addr = ret_addr; 253 254 return err; 255 } 256 257 static void thread_stack__pop(struct thread_stack *ts, u64 ret_addr) 258 { 259 size_t i; 260 261 /* 262 * In some cases there may be functions which are not seen to return. 263 * For example when setjmp / longjmp has been used. Or the perf context 264 * switch in the kernel which doesn't stop and start tracing in exactly 265 * the same code path. When that happens the return address will be 266 * further down the stack. If the return address is not found at all, 267 * we assume the opposite (i.e. this is a return for a call that wasn't 268 * seen for some reason) and leave the stack alone. 269 */ 270 for (i = ts->cnt; i; ) { 271 if (ts->stack[--i].ret_addr == ret_addr) { 272 ts->cnt = i; 273 return; 274 } 275 } 276 } 277 278 static void thread_stack__pop_trace_end(struct thread_stack *ts) 279 { 280 size_t i; 281 282 for (i = ts->cnt; i; ) { 283 if (ts->stack[--i].trace_end) 284 ts->cnt = i; 285 else 286 return; 287 } 288 } 289 290 static bool thread_stack__in_kernel(struct thread_stack *ts) 291 { 292 if (!ts->cnt) 293 return false; 294 295 return ts->stack[ts->cnt - 1].cp->in_kernel; 296 } 297 298 static int thread_stack__call_return(struct thread *thread, 299 struct thread_stack *ts, size_t idx, 300 u64 timestamp, u64 ref, bool no_return) 301 { 302 struct call_return_processor *crp = ts->crp; 303 struct thread_stack_entry *tse; 304 struct call_return cr = { 305 .thread = thread, 306 .comm = ts->comm, 307 .db_id = 0, 308 }; 309 u64 *parent_db_id; 310 311 tse = &ts->stack[idx]; 312 cr.cp = tse->cp; 313 cr.call_time = tse->timestamp; 314 cr.return_time = timestamp; 315 cr.branch_count = ts->branch_count - tse->branch_count; 316 cr.insn_count = ts->insn_count - tse->insn_count; 317 cr.cyc_count = ts->cyc_count - tse->cyc_count; 318 cr.db_id = tse->db_id; 319 cr.call_ref = tse->ref; 320 cr.return_ref = ref; 321 if (tse->no_call) 322 cr.flags |= CALL_RETURN_NO_CALL; 323 if (no_return) 324 cr.flags |= CALL_RETURN_NO_RETURN; 325 if (tse->non_call) 326 cr.flags |= CALL_RETURN_NON_CALL; 327 328 /* 329 * The parent db_id must be assigned before exporting the child. Note 330 * it is not possible to export the parent first because its information 331 * is not yet complete because its 'return' has not yet been processed. 332 */ 333 parent_db_id = idx ? &(tse - 1)->db_id : NULL; 334 335 return crp->process(&cr, parent_db_id, crp->data); 336 } 337 338 static int __thread_stack__flush(struct thread *thread, struct thread_stack *ts) 339 { 340 struct call_return_processor *crp = ts->crp; 341 int err; 342 343 if (!crp) { 344 ts->cnt = 0; 345 ts->br_stack_pos = 0; 346 if (ts->br_stack_rb) 347 ts->br_stack_rb->nr = 0; 348 return 0; 349 } 350 351 while (ts->cnt) { 352 err = thread_stack__call_return(thread, ts, --ts->cnt, 353 ts->last_time, 0, true); 354 if (err) { 355 pr_err("Error flushing thread stack!\n"); 356 ts->cnt = 0; 357 return err; 358 } 359 } 360 361 return 0; 362 } 363 364 int thread_stack__flush(struct thread *thread) 365 { 366 struct thread_stack *ts = thread__ts(thread); 367 unsigned int pos; 368 int err = 0; 369 370 if (ts) { 371 for (pos = 0; pos < ts->arr_sz; pos++) { 372 int ret = __thread_stack__flush(thread, ts + pos); 373 374 if (ret) 375 err = ret; 376 } 377 } 378 379 return err; 380 } 381 382 static void thread_stack__update_br_stack(struct thread_stack *ts, u32 flags, 383 u64 from_ip, u64 to_ip) 384 { 385 struct branch_stack *bs = ts->br_stack_rb; 386 struct branch_entry *be; 387 388 if (!ts->br_stack_pos) 389 ts->br_stack_pos = ts->br_stack_sz; 390 391 ts->br_stack_pos -= 1; 392 393 be = &bs->entries[ts->br_stack_pos]; 394 be->from = from_ip; 395 be->to = to_ip; 396 be->flags.value = 0; 397 be->flags.abort = !!(flags & PERF_IP_FLAG_TX_ABORT); 398 be->flags.in_tx = !!(flags & PERF_IP_FLAG_IN_TX); 399 /* No support for mispredict */ 400 be->flags.mispred = ts->mispred_all; 401 402 if (bs->nr < ts->br_stack_sz) 403 bs->nr += 1; 404 } 405 406 int thread_stack__event(struct thread *thread, int cpu, u32 flags, u64 from_ip, 407 u64 to_ip, u16 insn_len, u64 trace_nr, bool callstack, 408 unsigned int br_stack_sz, bool mispred_all) 409 { 410 struct thread_stack *ts = thread__stack(thread, cpu); 411 412 if (!thread) 413 return -EINVAL; 414 415 if (!ts) { 416 ts = thread_stack__new(thread, cpu, NULL, callstack, br_stack_sz); 417 if (!ts) { 418 pr_warning("Out of memory: no thread stack\n"); 419 return -ENOMEM; 420 } 421 ts->trace_nr = trace_nr; 422 ts->mispred_all = mispred_all; 423 } 424 425 /* 426 * When the trace is discontinuous, the trace_nr changes. In that case 427 * the stack might be completely invalid. Better to report nothing than 428 * to report something misleading, so flush the stack. 429 */ 430 if (trace_nr != ts->trace_nr) { 431 if (ts->trace_nr) 432 __thread_stack__flush(thread, ts); 433 ts->trace_nr = trace_nr; 434 } 435 436 if (br_stack_sz) 437 thread_stack__update_br_stack(ts, flags, from_ip, to_ip); 438 439 /* 440 * Stop here if thread_stack__process() is in use, or not recording call 441 * stack. 442 */ 443 if (ts->crp || !callstack) 444 return 0; 445 446 if (flags & PERF_IP_FLAG_CALL) { 447 u64 ret_addr; 448 449 if (!to_ip) 450 return 0; 451 ret_addr = from_ip + insn_len; 452 if (ret_addr == to_ip) 453 return 0; /* Zero-length calls are excluded */ 454 return thread_stack__push(ts, ret_addr, 455 flags & PERF_IP_FLAG_TRACE_END); 456 } else if (flags & PERF_IP_FLAG_TRACE_BEGIN) { 457 /* 458 * If the caller did not change the trace number (which would 459 * have flushed the stack) then try to make sense of the stack. 460 * Possibly, tracing began after returning to the current 461 * address, so try to pop that. Also, do not expect a call made 462 * when the trace ended, to return, so pop that. 463 */ 464 thread_stack__pop(ts, to_ip); 465 thread_stack__pop_trace_end(ts); 466 } else if ((flags & PERF_IP_FLAG_RETURN) && from_ip) { 467 thread_stack__pop(ts, to_ip); 468 } 469 470 return 0; 471 } 472 473 void thread_stack__set_trace_nr(struct thread *thread, int cpu, u64 trace_nr) 474 { 475 struct thread_stack *ts = thread__stack(thread, cpu); 476 477 if (!ts) 478 return; 479 480 if (trace_nr != ts->trace_nr) { 481 if (ts->trace_nr) 482 __thread_stack__flush(thread, ts); 483 ts->trace_nr = trace_nr; 484 } 485 } 486 487 static void __thread_stack__free(struct thread *thread, struct thread_stack *ts) 488 { 489 __thread_stack__flush(thread, ts); 490 zfree(&ts->stack); 491 zfree(&ts->br_stack_rb); 492 } 493 494 static void thread_stack__reset(struct thread *thread, struct thread_stack *ts) 495 { 496 unsigned int arr_sz = ts->arr_sz; 497 498 __thread_stack__free(thread, ts); 499 memset(ts, 0, sizeof(*ts)); 500 ts->arr_sz = arr_sz; 501 } 502 503 void thread_stack__free(struct thread *thread) 504 { 505 struct thread_stack *ts = thread__ts(thread); 506 unsigned int pos; 507 508 if (ts) { 509 for (pos = 0; pos < ts->arr_sz; pos++) 510 __thread_stack__free(thread, ts + pos); 511 free(thread__ts(thread)); 512 thread__set_ts(thread, NULL); 513 } 514 } 515 516 static inline u64 callchain_context(u64 ip, u64 kernel_start) 517 { 518 return ip < kernel_start ? PERF_CONTEXT_USER : PERF_CONTEXT_KERNEL; 519 } 520 521 void thread_stack__sample(struct thread *thread, int cpu, 522 struct ip_callchain *chain, 523 size_t sz, u64 ip, u64 kernel_start) 524 { 525 struct thread_stack *ts = thread__stack(thread, cpu); 526 u64 context = callchain_context(ip, kernel_start); 527 u64 last_context; 528 size_t i, j; 529 530 if (sz < 2) { 531 chain->nr = 0; 532 return; 533 } 534 535 chain->ips[0] = context; 536 chain->ips[1] = ip; 537 538 if (!ts) { 539 chain->nr = 2; 540 return; 541 } 542 543 last_context = context; 544 545 for (i = 2, j = 1; i < sz && j <= ts->cnt; i++, j++) { 546 ip = ts->stack[ts->cnt - j].ret_addr; 547 context = callchain_context(ip, kernel_start); 548 if (context != last_context) { 549 if (i >= sz - 1) 550 break; 551 chain->ips[i++] = context; 552 last_context = context; 553 } 554 chain->ips[i] = ip; 555 } 556 557 chain->nr = i; 558 } 559 560 /* 561 * Hardware sample records, created some time after the event occurred, need to 562 * have subsequent addresses removed from the call chain. 563 */ 564 void thread_stack__sample_late(struct thread *thread, int cpu, 565 struct ip_callchain *chain, size_t sz, 566 u64 sample_ip, u64 kernel_start) 567 { 568 struct thread_stack *ts = thread__stack(thread, cpu); 569 u64 sample_context = callchain_context(sample_ip, kernel_start); 570 u64 last_context, context, ip; 571 size_t nr = 0, j; 572 573 if (sz < 2) { 574 chain->nr = 0; 575 return; 576 } 577 578 if (!ts) 579 goto out; 580 581 /* 582 * When tracing kernel space, kernel addresses occur at the top of the 583 * call chain after the event occurred but before tracing stopped. 584 * Skip them. 585 */ 586 for (j = 1; j <= ts->cnt; j++) { 587 ip = ts->stack[ts->cnt - j].ret_addr; 588 context = callchain_context(ip, kernel_start); 589 if (context == PERF_CONTEXT_USER || 590 (context == sample_context && ip == sample_ip)) 591 break; 592 } 593 594 last_context = sample_ip; /* Use sample_ip as an invalid context */ 595 596 for (; nr < sz && j <= ts->cnt; nr++, j++) { 597 ip = ts->stack[ts->cnt - j].ret_addr; 598 context = callchain_context(ip, kernel_start); 599 if (context != last_context) { 600 if (nr >= sz - 1) 601 break; 602 chain->ips[nr++] = context; 603 last_context = context; 604 } 605 chain->ips[nr] = ip; 606 } 607 out: 608 if (nr) { 609 chain->nr = nr; 610 } else { 611 chain->ips[0] = sample_context; 612 chain->ips[1] = sample_ip; 613 chain->nr = 2; 614 } 615 } 616 617 void thread_stack__br_sample(struct thread *thread, int cpu, 618 struct branch_stack *dst, unsigned int sz) 619 { 620 struct thread_stack *ts = thread__stack(thread, cpu); 621 const size_t bsz = sizeof(struct branch_entry); 622 struct branch_stack *src; 623 struct branch_entry *be; 624 unsigned int nr; 625 626 dst->nr = 0; 627 628 if (!ts) 629 return; 630 631 src = ts->br_stack_rb; 632 if (!src->nr) 633 return; 634 635 dst->nr = min((unsigned int)src->nr, sz); 636 637 be = &dst->entries[0]; 638 nr = min(ts->br_stack_sz - ts->br_stack_pos, (unsigned int)dst->nr); 639 memcpy(be, &src->entries[ts->br_stack_pos], bsz * nr); 640 641 if (src->nr >= ts->br_stack_sz) { 642 sz -= nr; 643 be = &dst->entries[nr]; 644 nr = min(ts->br_stack_pos, sz); 645 memcpy(be, &src->entries[0], bsz * ts->br_stack_pos); 646 } 647 } 648 649 /* Start of user space branch entries */ 650 static bool us_start(struct branch_entry *be, u64 kernel_start, bool *start) 651 { 652 if (!*start) 653 *start = be->to && be->to < kernel_start; 654 655 return *start; 656 } 657 658 /* 659 * Start of branch entries after the ip fell in between 2 branches, or user 660 * space branch entries. 661 */ 662 static bool ks_start(struct branch_entry *be, u64 sample_ip, u64 kernel_start, 663 bool *start, struct branch_entry *nb) 664 { 665 if (!*start) { 666 *start = (nb && sample_ip >= be->to && sample_ip <= nb->from) || 667 be->from < kernel_start || 668 (be->to && be->to < kernel_start); 669 } 670 671 return *start; 672 } 673 674 /* 675 * Hardware sample records, created some time after the event occurred, need to 676 * have subsequent addresses removed from the branch stack. 677 */ 678 void thread_stack__br_sample_late(struct thread *thread, int cpu, 679 struct branch_stack *dst, unsigned int sz, 680 u64 ip, u64 kernel_start) 681 { 682 struct thread_stack *ts = thread__stack(thread, cpu); 683 struct branch_entry *d, *s, *spos, *ssz; 684 struct branch_stack *src; 685 unsigned int nr = 0; 686 bool start = false; 687 688 dst->nr = 0; 689 690 if (!ts) 691 return; 692 693 src = ts->br_stack_rb; 694 if (!src->nr) 695 return; 696 697 spos = &src->entries[ts->br_stack_pos]; 698 ssz = &src->entries[ts->br_stack_sz]; 699 700 d = &dst->entries[0]; 701 s = spos; 702 703 if (ip < kernel_start) { 704 /* 705 * User space sample: start copying branch entries when the 706 * branch is in user space. 707 */ 708 for (s = spos; s < ssz && nr < sz; s++) { 709 if (us_start(s, kernel_start, &start)) { 710 *d++ = *s; 711 nr += 1; 712 } 713 } 714 715 if (src->nr >= ts->br_stack_sz) { 716 for (s = &src->entries[0]; s < spos && nr < sz; s++) { 717 if (us_start(s, kernel_start, &start)) { 718 *d++ = *s; 719 nr += 1; 720 } 721 } 722 } 723 } else { 724 struct branch_entry *nb = NULL; 725 726 /* 727 * Kernel space sample: start copying branch entries when the ip 728 * falls in between 2 branches (or the branch is in user space 729 * because then the start must have been missed). 730 */ 731 for (s = spos; s < ssz && nr < sz; s++) { 732 if (ks_start(s, ip, kernel_start, &start, nb)) { 733 *d++ = *s; 734 nr += 1; 735 } 736 nb = s; 737 } 738 739 if (src->nr >= ts->br_stack_sz) { 740 for (s = &src->entries[0]; s < spos && nr < sz; s++) { 741 if (ks_start(s, ip, kernel_start, &start, nb)) { 742 *d++ = *s; 743 nr += 1; 744 } 745 nb = s; 746 } 747 } 748 } 749 750 dst->nr = nr; 751 } 752 753 struct call_return_processor * 754 call_return_processor__new(int (*process)(struct call_return *cr, u64 *parent_db_id, void *data), 755 void *data) 756 { 757 struct call_return_processor *crp; 758 759 crp = zalloc(sizeof(struct call_return_processor)); 760 if (!crp) 761 return NULL; 762 crp->cpr = call_path_root__new(); 763 if (!crp->cpr) 764 goto out_free; 765 crp->process = process; 766 crp->data = data; 767 return crp; 768 769 out_free: 770 free(crp); 771 return NULL; 772 } 773 774 void call_return_processor__free(struct call_return_processor *crp) 775 { 776 if (crp) { 777 call_path_root__free(crp->cpr); 778 free(crp); 779 } 780 } 781 782 static int thread_stack__push_cp(struct thread_stack *ts, u64 ret_addr, 783 u64 timestamp, u64 ref, struct call_path *cp, 784 bool no_call, bool trace_end) 785 { 786 struct thread_stack_entry *tse; 787 int err; 788 789 if (!cp) 790 return -ENOMEM; 791 792 if (ts->cnt == ts->sz) { 793 err = thread_stack__grow(ts); 794 if (err) 795 return err; 796 } 797 798 tse = &ts->stack[ts->cnt++]; 799 tse->ret_addr = ret_addr; 800 tse->timestamp = timestamp; 801 tse->ref = ref; 802 tse->branch_count = ts->branch_count; 803 tse->insn_count = ts->insn_count; 804 tse->cyc_count = ts->cyc_count; 805 tse->cp = cp; 806 tse->no_call = no_call; 807 tse->trace_end = trace_end; 808 tse->non_call = false; 809 tse->db_id = 0; 810 811 return 0; 812 } 813 814 static int thread_stack__pop_cp(struct thread *thread, struct thread_stack *ts, 815 u64 ret_addr, u64 timestamp, u64 ref, 816 struct symbol *sym) 817 { 818 int err; 819 820 if (!ts->cnt) 821 return 1; 822 823 if (ts->cnt == 1) { 824 struct thread_stack_entry *tse = &ts->stack[0]; 825 826 if (tse->cp->sym == sym) 827 return thread_stack__call_return(thread, ts, --ts->cnt, 828 timestamp, ref, false); 829 } 830 831 if (ts->stack[ts->cnt - 1].ret_addr == ret_addr && 832 !ts->stack[ts->cnt - 1].non_call) { 833 return thread_stack__call_return(thread, ts, --ts->cnt, 834 timestamp, ref, false); 835 } else { 836 size_t i = ts->cnt - 1; 837 838 while (i--) { 839 if (ts->stack[i].ret_addr != ret_addr || 840 ts->stack[i].non_call) 841 continue; 842 i += 1; 843 while (ts->cnt > i) { 844 err = thread_stack__call_return(thread, ts, 845 --ts->cnt, 846 timestamp, ref, 847 true); 848 if (err) 849 return err; 850 } 851 return thread_stack__call_return(thread, ts, --ts->cnt, 852 timestamp, ref, false); 853 } 854 } 855 856 return 1; 857 } 858 859 static int thread_stack__bottom(struct thread_stack *ts, 860 struct perf_sample *sample, 861 struct addr_location *from_al, 862 struct addr_location *to_al, u64 ref) 863 { 864 struct call_path_root *cpr = ts->crp->cpr; 865 struct call_path *cp; 866 struct symbol *sym; 867 u64 ip; 868 869 if (sample->ip) { 870 ip = sample->ip; 871 sym = from_al->sym; 872 } else if (sample->addr) { 873 ip = sample->addr; 874 sym = to_al->sym; 875 } else { 876 return 0; 877 } 878 879 cp = call_path__findnew(cpr, &cpr->call_path, sym, ip, 880 ts->kernel_start); 881 882 return thread_stack__push_cp(ts, ip, sample->time, ref, cp, 883 true, false); 884 } 885 886 static int thread_stack__pop_ks(struct thread *thread, struct thread_stack *ts, 887 struct perf_sample *sample, u64 ref) 888 { 889 u64 tm = sample->time; 890 int err; 891 892 /* Return to userspace, so pop all kernel addresses */ 893 while (thread_stack__in_kernel(ts)) { 894 err = thread_stack__call_return(thread, ts, --ts->cnt, 895 tm, ref, true); 896 if (err) 897 return err; 898 } 899 900 return 0; 901 } 902 903 static int thread_stack__no_call_return(struct thread *thread, 904 struct thread_stack *ts, 905 struct perf_sample *sample, 906 struct addr_location *from_al, 907 struct addr_location *to_al, u64 ref) 908 { 909 struct call_path_root *cpr = ts->crp->cpr; 910 struct call_path *root = &cpr->call_path; 911 struct symbol *fsym = from_al->sym; 912 struct symbol *tsym = to_al->sym; 913 struct call_path *cp, *parent; 914 u64 ks = ts->kernel_start; 915 u64 addr = sample->addr; 916 u64 tm = sample->time; 917 u64 ip = sample->ip; 918 int err; 919 920 if (ip >= ks && addr < ks) { 921 /* Return to userspace, so pop all kernel addresses */ 922 err = thread_stack__pop_ks(thread, ts, sample, ref); 923 if (err) 924 return err; 925 926 /* If the stack is empty, push the userspace address */ 927 if (!ts->cnt) { 928 cp = call_path__findnew(cpr, root, tsym, addr, ks); 929 return thread_stack__push_cp(ts, 0, tm, ref, cp, true, 930 false); 931 } 932 } else if (thread_stack__in_kernel(ts) && ip < ks) { 933 /* Return to userspace, so pop all kernel addresses */ 934 err = thread_stack__pop_ks(thread, ts, sample, ref); 935 if (err) 936 return err; 937 } 938 939 if (ts->cnt) 940 parent = ts->stack[ts->cnt - 1].cp; 941 else 942 parent = root; 943 944 if (parent->sym == from_al->sym) { 945 /* 946 * At the bottom of the stack, assume the missing 'call' was 947 * before the trace started. So, pop the current symbol and push 948 * the 'to' symbol. 949 */ 950 if (ts->cnt == 1) { 951 err = thread_stack__call_return(thread, ts, --ts->cnt, 952 tm, ref, false); 953 if (err) 954 return err; 955 } 956 957 if (!ts->cnt) { 958 cp = call_path__findnew(cpr, root, tsym, addr, ks); 959 960 return thread_stack__push_cp(ts, addr, tm, ref, cp, 961 true, false); 962 } 963 964 /* 965 * Otherwise assume the 'return' is being used as a jump (e.g. 966 * retpoline) and just push the 'to' symbol. 967 */ 968 cp = call_path__findnew(cpr, parent, tsym, addr, ks); 969 970 err = thread_stack__push_cp(ts, 0, tm, ref, cp, true, false); 971 if (!err) 972 ts->stack[ts->cnt - 1].non_call = true; 973 974 return err; 975 } 976 977 /* 978 * Assume 'parent' has not yet returned, so push 'to', and then push and 979 * pop 'from'. 980 */ 981 982 cp = call_path__findnew(cpr, parent, tsym, addr, ks); 983 984 err = thread_stack__push_cp(ts, addr, tm, ref, cp, true, false); 985 if (err) 986 return err; 987 988 cp = call_path__findnew(cpr, cp, fsym, ip, ks); 989 990 err = thread_stack__push_cp(ts, ip, tm, ref, cp, true, false); 991 if (err) 992 return err; 993 994 return thread_stack__call_return(thread, ts, --ts->cnt, tm, ref, false); 995 } 996 997 static int thread_stack__trace_begin(struct thread *thread, 998 struct thread_stack *ts, u64 timestamp, 999 u64 ref) 1000 { 1001 struct thread_stack_entry *tse; 1002 int err; 1003 1004 if (!ts->cnt) 1005 return 0; 1006 1007 /* Pop trace end */ 1008 tse = &ts->stack[ts->cnt - 1]; 1009 if (tse->trace_end) { 1010 err = thread_stack__call_return(thread, ts, --ts->cnt, 1011 timestamp, ref, false); 1012 if (err) 1013 return err; 1014 } 1015 1016 return 0; 1017 } 1018 1019 static int thread_stack__trace_end(struct thread_stack *ts, 1020 struct perf_sample *sample, u64 ref) 1021 { 1022 struct call_path_root *cpr = ts->crp->cpr; 1023 struct call_path *cp; 1024 u64 ret_addr; 1025 1026 /* No point having 'trace end' on the bottom of the stack */ 1027 if (!ts->cnt || (ts->cnt == 1 && ts->stack[0].ref == ref)) 1028 return 0; 1029 1030 cp = call_path__findnew(cpr, ts->stack[ts->cnt - 1].cp, NULL, 0, 1031 ts->kernel_start); 1032 1033 ret_addr = sample->ip + sample->insn_len; 1034 1035 return thread_stack__push_cp(ts, ret_addr, sample->time, ref, cp, 1036 false, true); 1037 } 1038 1039 static bool is_x86_retpoline(const char *name) 1040 { 1041 const char *p = strstr(name, "__x86_indirect_thunk_"); 1042 1043 return p == name || !strcmp(name, "__indirect_thunk_start"); 1044 } 1045 1046 /* 1047 * x86 retpoline functions pollute the call graph. This function removes them. 1048 * This does not handle function return thunks, nor is there any improvement 1049 * for the handling of inline thunks or extern thunks. 1050 */ 1051 static int thread_stack__x86_retpoline(struct thread_stack *ts, 1052 struct perf_sample *sample, 1053 struct addr_location *to_al) 1054 { 1055 struct thread_stack_entry *tse = &ts->stack[ts->cnt - 1]; 1056 struct call_path_root *cpr = ts->crp->cpr; 1057 struct symbol *sym = tse->cp->sym; 1058 struct symbol *tsym = to_al->sym; 1059 struct call_path *cp; 1060 1061 if (sym && is_x86_retpoline(sym->name)) { 1062 /* 1063 * This is a x86 retpoline fn. It pollutes the call graph by 1064 * showing up everywhere there is an indirect branch, but does 1065 * not itself mean anything. Here the top-of-stack is removed, 1066 * by decrementing the stack count, and then further down, the 1067 * resulting top-of-stack is replaced with the actual target. 1068 * The result is that the retpoline functions will no longer 1069 * appear in the call graph. Note this only affects the call 1070 * graph, since all the original branches are left unchanged. 1071 */ 1072 ts->cnt -= 1; 1073 sym = ts->stack[ts->cnt - 2].cp->sym; 1074 if (sym && sym == tsym && to_al->addr != tsym->start) { 1075 /* 1076 * Target is back to the middle of the symbol we came 1077 * from so assume it is an indirect jmp and forget it 1078 * altogether. 1079 */ 1080 ts->cnt -= 1; 1081 return 0; 1082 } 1083 } else if (sym && sym == tsym) { 1084 /* 1085 * Target is back to the symbol we came from so assume it is an 1086 * indirect jmp and forget it altogether. 1087 */ 1088 ts->cnt -= 1; 1089 return 0; 1090 } 1091 1092 cp = call_path__findnew(cpr, ts->stack[ts->cnt - 2].cp, tsym, 1093 sample->addr, ts->kernel_start); 1094 if (!cp) 1095 return -ENOMEM; 1096 1097 /* Replace the top-of-stack with the actual target */ 1098 ts->stack[ts->cnt - 1].cp = cp; 1099 1100 return 0; 1101 } 1102 1103 int thread_stack__process(struct thread *thread, struct comm *comm, 1104 struct perf_sample *sample, 1105 struct addr_location *from_al, 1106 struct addr_location *to_al, u64 ref, 1107 struct call_return_processor *crp) 1108 { 1109 struct thread_stack *ts = thread__stack(thread, sample->cpu); 1110 enum retpoline_state_t rstate; 1111 int err = 0; 1112 1113 if (ts && !ts->crp) { 1114 /* Supersede thread_stack__event() */ 1115 thread_stack__reset(thread, ts); 1116 ts = NULL; 1117 } 1118 1119 if (!ts) { 1120 ts = thread_stack__new(thread, sample->cpu, crp, true, 0); 1121 if (!ts) 1122 return -ENOMEM; 1123 ts->comm = comm; 1124 } 1125 1126 rstate = ts->rstate; 1127 if (rstate == X86_RETPOLINE_DETECTED) 1128 ts->rstate = X86_RETPOLINE_POSSIBLE; 1129 1130 /* Flush stack on exec */ 1131 if (ts->comm != comm && thread__pid(thread) == thread__tid(thread)) { 1132 err = __thread_stack__flush(thread, ts); 1133 if (err) 1134 return err; 1135 ts->comm = comm; 1136 } 1137 1138 /* If the stack is empty, put the current symbol on the stack */ 1139 if (!ts->cnt) { 1140 err = thread_stack__bottom(ts, sample, from_al, to_al, ref); 1141 if (err) 1142 return err; 1143 } 1144 1145 ts->branch_count += 1; 1146 ts->insn_count += sample->insn_cnt; 1147 ts->cyc_count += sample->cyc_cnt; 1148 ts->last_time = sample->time; 1149 1150 if (sample->flags & PERF_IP_FLAG_CALL) { 1151 bool trace_end = sample->flags & PERF_IP_FLAG_TRACE_END; 1152 struct call_path_root *cpr = ts->crp->cpr; 1153 struct call_path *cp; 1154 u64 ret_addr; 1155 1156 if (!sample->ip || !sample->addr) 1157 return 0; 1158 1159 ret_addr = sample->ip + sample->insn_len; 1160 if (ret_addr == sample->addr) 1161 return 0; /* Zero-length calls are excluded */ 1162 1163 cp = call_path__findnew(cpr, ts->stack[ts->cnt - 1].cp, 1164 to_al->sym, sample->addr, 1165 ts->kernel_start); 1166 err = thread_stack__push_cp(ts, ret_addr, sample->time, ref, 1167 cp, false, trace_end); 1168 1169 /* 1170 * A call to the same symbol but not the start of the symbol, 1171 * may be the start of a x86 retpoline. 1172 */ 1173 if (!err && rstate == X86_RETPOLINE_POSSIBLE && to_al->sym && 1174 from_al->sym == to_al->sym && 1175 to_al->addr != to_al->sym->start) 1176 ts->rstate = X86_RETPOLINE_DETECTED; 1177 1178 } else if (sample->flags & PERF_IP_FLAG_RETURN) { 1179 if (!sample->addr) { 1180 u32 return_from_kernel = PERF_IP_FLAG_SYSCALLRET | 1181 PERF_IP_FLAG_INTERRUPT; 1182 1183 if (!(sample->flags & return_from_kernel)) 1184 return 0; 1185 1186 /* Pop kernel stack */ 1187 return thread_stack__pop_ks(thread, ts, sample, ref); 1188 } 1189 1190 if (!sample->ip) 1191 return 0; 1192 1193 /* x86 retpoline 'return' doesn't match the stack */ 1194 if (rstate == X86_RETPOLINE_DETECTED && ts->cnt > 2 && 1195 ts->stack[ts->cnt - 1].ret_addr != sample->addr) 1196 return thread_stack__x86_retpoline(ts, sample, to_al); 1197 1198 err = thread_stack__pop_cp(thread, ts, sample->addr, 1199 sample->time, ref, from_al->sym); 1200 if (err) { 1201 if (err < 0) 1202 return err; 1203 err = thread_stack__no_call_return(thread, ts, sample, 1204 from_al, to_al, ref); 1205 } 1206 } else if (sample->flags & PERF_IP_FLAG_TRACE_BEGIN) { 1207 err = thread_stack__trace_begin(thread, ts, sample->time, ref); 1208 } else if (sample->flags & PERF_IP_FLAG_TRACE_END) { 1209 err = thread_stack__trace_end(ts, sample, ref); 1210 } else if (sample->flags & PERF_IP_FLAG_BRANCH && 1211 from_al->sym != to_al->sym && to_al->sym && 1212 to_al->addr == to_al->sym->start) { 1213 struct call_path_root *cpr = ts->crp->cpr; 1214 struct call_path *cp; 1215 1216 /* 1217 * The compiler might optimize a call/ret combination by making 1218 * it a jmp. Make that visible by recording on the stack a 1219 * branch to the start of a different symbol. Note, that means 1220 * when a ret pops the stack, all jmps must be popped off first. 1221 */ 1222 cp = call_path__findnew(cpr, ts->stack[ts->cnt - 1].cp, 1223 to_al->sym, sample->addr, 1224 ts->kernel_start); 1225 err = thread_stack__push_cp(ts, 0, sample->time, ref, cp, false, 1226 false); 1227 if (!err) 1228 ts->stack[ts->cnt - 1].non_call = true; 1229 } 1230 1231 return err; 1232 } 1233 1234 size_t thread_stack__depth(struct thread *thread, int cpu) 1235 { 1236 struct thread_stack *ts = thread__stack(thread, cpu); 1237 1238 if (!ts) 1239 return 0; 1240 return ts->cnt; 1241 } 1242