tools/perf/util/thread-stack.c

   1 /*
   2  * thread-stack.c: Synthesize a thread's stack using call / return events
   3  * Copyright (c) 2014, Intel Corporation.
   4  *
   5  * This program is free software; you can redistribute it and/or modify it
   6  * under the terms and conditions of the GNU General Public License,
   7  * version 2, as published by the Free Software Foundation.
   8  *
   9  * This program is distributed in the hope it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  12  * more details.
  13  *
  14  */
  15
  16 #include <linux/rbtree.h>
  17 #include <linux/list.h>
  18 #include <errno.h>
  19 #include "thread.h"
  20 #include "event.h"
  21 #include "machine.h"
  22 #include "util.h"
  23 #include "debug.h"
  24 #include "symbol.h"
  25 #include "comm.h"
  26 #include "call-path.h"
  27 #include "thread-stack.h"
  28
  29 #define STACK_GROWTH 2048
  30
  31 /**
  32  * struct thread_stack_entry - thread stack entry.
  33  * @ret_addr: return address
  34  * @timestamp: timestamp (if known)
  35  * @ref: external reference (e.g. db_id of sample)
  36  * @branch_count: the branch count when the entry was created
  37  * @cp: call path
  38  * @no_call: a 'call' was not seen
  39  */
  40 struct thread_stack_entry {
  41         u64 ret_addr;
  42         u64 timestamp;
  43         u64 ref;
  44         u64 branch_count;
  45         struct call_path *cp;
  46         bool no_call;
  47 };
  48
  49 /**
  50  * struct thread_stack - thread stack constructed from 'call' and 'return'
  51  *                       branch samples.
  52  * @stack: array that holds the stack
  53  * @cnt: number of entries in the stack
  54  * @sz: current maximum stack size
  55  * @trace_nr: current trace number
  56  * @branch_count: running branch count
  57  * @kernel_start: kernel start address
  58  * @last_time: last timestamp
  59  * @crp: call/return processor
  60  * @comm: current comm
  61  */
  62 struct thread_stack {
  63         struct thread_stack_entry *stack;
  64         size_t cnt;
  65         size_t sz;
  66         u64 trace_nr;
  67         u64 branch_count;
  68         u64 kernel_start;
  69         u64 last_time;
  70         struct call_return_processor *crp;
  71         struct comm *comm;
  72 };
  73
  74 static int thread_stack__grow(struct thread_stack *ts)
  75 {
  76         struct thread_stack_entry *new_stack;
  77         size_t sz, new_sz;
  78
  79         new_sz = ts->sz + STACK_GROWTH;
  80         sz = new_sz * sizeof(struct thread_stack_entry);
  81
  82         new_stack = realloc(ts->stack, sz);
  83         if (!new_stack)
  84                 return -ENOMEM;
  85
  86         ts->stack = new_stack;
  87         ts->sz = new_sz;
  88
  89         return 0;
  90 }
  91
  92 static struct thread_stack *thread_stack__new(struct thread *thread,
  93                                               struct call_return_processor *crp)
  94 {
  95         struct thread_stack *ts;
  96
  97         ts = zalloc(sizeof(struct thread_stack));
  98         if (!ts)
  99                 return NULL;
 100
 101         if (thread_stack__grow(ts)) {
 102                 free(ts);
 103                 return NULL;
 104         }
 105
 106         if (thread->mg && thread->mg->machine)
 107                 ts->kernel_start = machine__kernel_start(thread->mg->machine);
 108         else
 109                 ts->kernel_start = 1ULL << 63;
 110         ts->crp = crp;
 111
 112         return ts;
 113 }
 114
 115 static int thread_stack__push(struct thread_stack *ts, u64 ret_addr)
 116 {
 117         int err = 0;
 118
 119         if (ts->cnt == ts->sz) {
 120                 err = thread_stack__grow(ts);
 121                 if (err) {
 122                         pr_warning("Out of memory: discarding thread stack\n");
 123                         ts->cnt = 0;
 124                 }
 125         }
 126
 127         ts->stack[ts->cnt++].ret_addr = ret_addr;
 128
 129         return err;
 130 }
 131
 132 static void thread_stack__pop(struct thread_stack *ts, u64 ret_addr)
 133 {
 134         size_t i;
 135
 136         /*
 137          * In some cases there may be functions which are not seen to return.
 138          * For example when setjmp / longjmp has been used.  Or the perf context
 139          * switch in the kernel which doesn't stop and start tracing in exactly
 140          * the same code path.  When that happens the return address will be
 141          * further down the stack.  If the return address is not found at all,
 142          * we assume the opposite (i.e. this is a return for a call that wasn't
 143          * seen for some reason) and leave the stack alone.
 144          */
 145         for (i = ts->cnt; i; ) {
 146                 if (ts->stack[--i].ret_addr == ret_addr) {
 147                         ts->cnt = i;
 148                         return;
 149                 }
 150         }
 151 }
 152
 153 static bool thread_stack__in_kernel(struct thread_stack *ts)
 154 {
 155         if (!ts->cnt)
 156                 return false;
 157
 158         return ts->stack[ts->cnt - 1].cp->in_kernel;
 159 }
 160
 161 static int thread_stack__call_return(struct thread *thread,
 162                                      struct thread_stack *ts, size_t idx,
 163                                      u64 timestamp, u64 ref, bool no_return)
 164 {
 165         struct call_return_processor *crp = ts->crp;
 166         struct thread_stack_entry *tse;
 167         struct call_return cr = {
 168                 .thread = thread,
 169                 .comm = ts->comm,
 170                 .db_id = 0,
 171         };
 172
 173         tse = &ts->stack[idx];
 174         cr.cp = tse->cp;
 175         cr.call_time = tse->timestamp;
 176         cr.return_time = timestamp;
 177         cr.branch_count = ts->branch_count - tse->branch_count;
 178         cr.call_ref = tse->ref;
 179         cr.return_ref = ref;
 180         if (tse->no_call)
 181                 cr.flags |= CALL_RETURN_NO_CALL;
 182         if (no_return)
 183                 cr.flags |= CALL_RETURN_NO_RETURN;
 184
 185         return crp->process(&cr, crp->data);
 186 }
 187
 188 static int __thread_stack__flush(struct thread *thread, struct thread_stack *ts)
 189 {
 190         struct call_return_processor *crp = ts->crp;
 191         int err;
 192
 193         if (!crp) {
 194                 ts->cnt = 0;
 195                 return 0;
 196         }
 197
 198         while (ts->cnt) {
 199                 err = thread_stack__call_return(thread, ts, --ts->cnt,
 200                                                 ts->last_time, 0, true);
 201                 if (err) {
 202                         pr_err("Error flushing thread stack!\n");
 203                         ts->cnt = 0;
 204                         return err;
 205                 }
 206         }
 207
 208         return 0;
 209 }
 210
 211 int thread_stack__flush(struct thread *thread)
 212 {
 213         if (thread->ts)
 214                 return __thread_stack__flush(thread, thread->ts);
 215
 216         return 0;
 217 }
 218
 219 int thread_stack__event(struct thread *thread, u32 flags, u64 from_ip,
 220                         u64 to_ip, u16 insn_len, u64 trace_nr)
 221 {
 222         if (!thread)
 223                 return -EINVAL;
 224
 225         if (!thread->ts) {
 226                 thread->ts = thread_stack__new(thread, NULL);
 227                 if (!thread->ts) {
 228                         pr_warning("Out of memory: no thread stack\n");
 229                         return -ENOMEM;
 230                 }
 231                 thread->ts->trace_nr = trace_nr;
 232         }
 233
 234         /*
 235          * When the trace is discontinuous, the trace_nr changes.  In that case
 236          * the stack might be completely invalid.  Better to report nothing than
 237          * to report something misleading, so flush the stack.
 238          */
 239         if (trace_nr != thread->ts->trace_nr) {
 240                 if (thread->ts->trace_nr)
 241                         __thread_stack__flush(thread, thread->ts);
 242                 thread->ts->trace_nr = trace_nr;
 243         }
 244
 245         /* Stop here if thread_stack__process() is in use */
 246         if (thread->ts->crp)
 247                 return 0;
 248
 249         if (flags & PERF_IP_FLAG_CALL) {
 250                 u64 ret_addr;
 251
 252                 if (!to_ip)
 253                         return 0;
 254                 ret_addr = from_ip + insn_len;
 255                 if (ret_addr == to_ip)
 256                         return 0; /* Zero-length calls are excluded */
 257                 return thread_stack__push(thread->ts, ret_addr);
 258         } else if (flags & PERF_IP_FLAG_RETURN) {
 259                 if (!from_ip)
 260                         return 0;
 261                 thread_stack__pop(thread->ts, to_ip);
 262         }
 263
 264         return 0;
 265 }
 266
 267 void thread_stack__set_trace_nr(struct thread *thread, u64 trace_nr)
 268 {
 269         if (!thread || !thread->ts)
 270                 return;
 271
 272         if (trace_nr != thread->ts->trace_nr) {
 273                 if (thread->ts->trace_nr)
 274                         __thread_stack__flush(thread, thread->ts);
 275                 thread->ts->trace_nr = trace_nr;
 276         }
 277 }
 278
 279 void thread_stack__free(struct thread *thread)
 280 {
 281         if (thread->ts) {
 282                 __thread_stack__flush(thread, thread->ts);
 283                 zfree(&thread->ts->stack);
 284                 zfree(&thread->ts);
 285         }
 286 }
 287
 288 static inline u64 callchain_context(u64 ip, u64 kernel_start)
 289 {
 290         return ip < kernel_start ? PERF_CONTEXT_USER : PERF_CONTEXT_KERNEL;
 291 }
 292
 293 void thread_stack__sample(struct thread *thread, struct ip_callchain *chain,
 294                           size_t sz, u64 ip, u64 kernel_start)
 295 {
 296         u64 context = callchain_context(ip, kernel_start);
 297         u64 last_context;
 298         size_t i, j;
 299
 300         if (sz < 2) {
 301                 chain->nr = 0;
 302                 return;
 303         }
 304
 305         chain->ips[0] = context;
 306         chain->ips[1] = ip;
 307
 308         if (!thread || !thread->ts) {
 309                 chain->nr = 2;
 310                 return;
 311         }
 312
 313         last_context = context;
 314
 315         for (i = 2, j = 1; i < sz && j <= thread->ts->cnt; i++, j++) {
 316                 ip = thread->ts->stack[thread->ts->cnt - j].ret_addr;
 317                 context = callchain_context(ip, kernel_start);
 318                 if (context != last_context) {
 319                         if (i >= sz - 1)
 320                                 break;
 321                         chain->ips[i++] = context;
 322                         last_context = context;
 323                 }
 324                 chain->ips[i] = ip;
 325         }
 326
 327         chain->nr = i;
 328 }
 329
 330 struct call_return_processor *
 331 call_return_processor__new(int (*process)(struct call_return *cr, void *data),
 332                            void *data)
 333 {
 334         struct call_return_processor *crp;
 335
 336         crp = zalloc(sizeof(struct call_return_processor));
 337         if (!crp)
 338                 return NULL;
 339         crp->cpr = call_path_root__new();
 340         if (!crp->cpr)
 341                 goto out_free;
 342         crp->process = process;
 343         crp->data = data;
 344         return crp;
 345
 346 out_free:
 347         free(crp);
 348         return NULL;
 349 }
 350
 351 void call_return_processor__free(struct call_return_processor *crp)
 352 {
 353         if (crp) {
 354                 call_path_root__free(crp->cpr);
 355                 free(crp);
 356         }
 357 }
 358
 359 static int thread_stack__push_cp(struct thread_stack *ts, u64 ret_addr,
 360                                  u64 timestamp, u64 ref, struct call_path *cp,
 361                                  bool no_call)
 362 {
 363         struct thread_stack_entry *tse;
 364         int err;
 365
 366         if (ts->cnt == ts->sz) {
 367                 err = thread_stack__grow(ts);
 368                 if (err)
 369                         return err;
 370         }
 371
 372         tse = &ts->stack[ts->cnt++];
 373         tse->ret_addr = ret_addr;
 374         tse->timestamp = timestamp;
 375         tse->ref = ref;
 376         tse->branch_count = ts->branch_count;
 377         tse->cp = cp;
 378         tse->no_call = no_call;
 379
 380         return 0;
 381 }
 382
 383 static int thread_stack__pop_cp(struct thread *thread, struct thread_stack *ts,
 384                                 u64 ret_addr, u64 timestamp, u64 ref,
 385                                 struct symbol *sym)
 386 {
 387         int err;
 388
 389         if (!ts->cnt)
 390                 return 1;
 391
 392         if (ts->cnt == 1) {
 393                 struct thread_stack_entry *tse = &ts->stack[0];
 394
 395                 if (tse->cp->sym == sym)
 396                         return thread_stack__call_return(thread, ts, --ts->cnt,
 397                                                          timestamp, ref, false);
 398         }
 399
 400         if (ts->stack[ts->cnt - 1].ret_addr == ret_addr) {
 401                 return thread_stack__call_return(thread, ts, --ts->cnt,
 402                                                  timestamp, ref, false);
 403         } else {
 404                 size_t i = ts->cnt - 1;
 405
 406                 while (i--) {
 407                         if (ts->stack[i].ret_addr != ret_addr)
 408                                 continue;
 409                         i += 1;
 410                         while (ts->cnt > i) {
 411                                 err = thread_stack__call_return(thread, ts,
 412                                                                 --ts->cnt,
 413                                                                 timestamp, ref,
 414                                                                 true);
 415                                 if (err)
 416                                         return err;
 417                         }
 418                         return thread_stack__call_return(thread, ts, --ts->cnt,
 419                                                          timestamp, ref, false);
 420                 }
 421         }
 422
 423         return 1;
 424 }
 425
 426 static int thread_stack__bottom(struct thread *thread, struct thread_stack *ts,
 427                                 struct perf_sample *sample,
 428                                 struct addr_location *from_al,
 429                                 struct addr_location *to_al, u64 ref)
 430 {
 431         struct call_path_root *cpr = ts->crp->cpr;
 432         struct call_path *cp;
 433         struct symbol *sym;
 434         u64 ip;
 435
 436         if (sample->ip) {
 437                 ip = sample->ip;
 438                 sym = from_al->sym;
 439         } else if (sample->addr) {
 440                 ip = sample->addr;
 441                 sym = to_al->sym;
 442         } else {
 443                 return 0;
 444         }
 445
 446         cp = call_path__findnew(cpr, &cpr->call_path, sym, ip,
 447                                 ts->kernel_start);
 448         if (!cp)
 449                 return -ENOMEM;
 450
 451         return thread_stack__push_cp(thread->ts, ip, sample->time, ref, cp,
 452                                      true);
 453 }
 454
 455 static int thread_stack__no_call_return(struct thread *thread,
 456                                         struct thread_stack *ts,
 457                                         struct perf_sample *sample,
 458                                         struct addr_location *from_al,
 459                                         struct addr_location *to_al, u64 ref)
 460 {
 461         struct call_path_root *cpr = ts->crp->cpr;
 462         struct call_path *cp, *parent;
 463         u64 ks = ts->kernel_start;
 464         int err;
 465
 466         if (sample->ip >= ks && sample->addr < ks) {
 467                 /* Return to userspace, so pop all kernel addresses */
 468                 while (thread_stack__in_kernel(ts)) {
 469                         err = thread_stack__call_return(thread, ts, --ts->cnt,
 470                                                         sample->time, ref,
 471                                                         true);
 472                         if (err)
 473                                 return err;
 474                 }
 475
 476                 /* If the stack is empty, push the userspace address */
 477                 if (!ts->cnt) {
 478                         cp = call_path__findnew(cpr, &cpr->call_path,
 479                                                 to_al->sym, sample->addr,
 480                                                 ts->kernel_start);
 481                         if (!cp)
 482                                 return -ENOMEM;
 483                         return thread_stack__push_cp(ts, 0, sample->time, ref,
 484                                                      cp, true);
 485                 }
 486         } else if (thread_stack__in_kernel(ts) && sample->ip < ks) {
 487                 /* Return to userspace, so pop all kernel addresses */
 488                 while (thread_stack__in_kernel(ts)) {
 489                         err = thread_stack__call_return(thread, ts, --ts->cnt,
 490                                                         sample->time, ref,
 491                                                         true);
 492                         if (err)
 493                                 return err;
 494                 }
 495         }
 496
 497         if (ts->cnt)
 498                 parent = ts->stack[ts->cnt - 1].cp;
 499         else
 500                 parent = &cpr->call_path;
 501
 502         /* This 'return' had no 'call', so push and pop top of stack */
 503         cp = call_path__findnew(cpr, parent, from_al->sym, sample->ip,
 504                                 ts->kernel_start);
 505         if (!cp)
 506                 return -ENOMEM;
 507
 508         err = thread_stack__push_cp(ts, sample->addr, sample->time, ref, cp,
 509                                     true);
 510         if (err)
 511                 return err;
 512
 513         return thread_stack__pop_cp(thread, ts, sample->addr, sample->time, ref,
 514                                     to_al->sym);
 515 }
 516
 517 static int thread_stack__trace_begin(struct thread *thread,
 518                                      struct thread_stack *ts, u64 timestamp,
 519                                      u64 ref)
 520 {
 521         struct thread_stack_entry *tse;
 522         int err;
 523
 524         if (!ts->cnt)
 525                 return 0;
 526
 527         /* Pop trace end */
 528         tse = &ts->stack[ts->cnt - 1];
 529         if (tse->cp->sym == NULL && tse->cp->ip == 0) {
 530                 err = thread_stack__call_return(thread, ts, --ts->cnt,
 531                                                 timestamp, ref, false);
 532                 if (err)
 533                         return err;
 534         }
 535
 536         return 0;
 537 }
 538
 539 static int thread_stack__trace_end(struct thread_stack *ts,
 540                                    struct perf_sample *sample, u64 ref)
 541 {
 542         struct call_path_root *cpr = ts->crp->cpr;
 543         struct call_path *cp;
 544         u64 ret_addr;
 545
 546         /* No point having 'trace end' on the bottom of the stack */
 547         if (!ts->cnt || (ts->cnt == 1 && ts->stack[0].ref == ref))
 548                 return 0;
 549
 550         cp = call_path__findnew(cpr, ts->stack[ts->cnt - 1].cp, NULL, 0,
 551                                 ts->kernel_start);
 552         if (!cp)
 553                 return -ENOMEM;
 554
 555         ret_addr = sample->ip + sample->insn_len;
 556
 557         return thread_stack__push_cp(ts, ret_addr, sample->time, ref, cp,
 558                                      false);
 559 }
 560
 561 int thread_stack__process(struct thread *thread, struct comm *comm,
 562                           struct perf_sample *sample,
 563                           struct addr_location *from_al,
 564                           struct addr_location *to_al, u64 ref,
 565                           struct call_return_processor *crp)
 566 {
 567         struct thread_stack *ts = thread->ts;
 568         int err = 0;
 569
 570         if (ts) {
 571                 if (!ts->crp) {
 572                         /* Supersede thread_stack__event() */
 573                         thread_stack__free(thread);
 574                         thread->ts = thread_stack__new(thread, crp);
 575                         if (!thread->ts)
 576                                 return -ENOMEM;
 577                         ts = thread->ts;
 578                         ts->comm = comm;
 579                 }
 580         } else {
 581                 thread->ts = thread_stack__new(thread, crp);
 582                 if (!thread->ts)
 583                         return -ENOMEM;
 584                 ts = thread->ts;
 585                 ts->comm = comm;
 586         }
 587
 588         /* Flush stack on exec */
 589         if (ts->comm != comm && thread->pid_ == thread->tid) {
 590                 err = __thread_stack__flush(thread, ts);
 591                 if (err)
 592                         return err;
 593                 ts->comm = comm;
 594         }
 595
 596         /* If the stack is empty, put the current symbol on the stack */
 597         if (!ts->cnt) {
 598                 err = thread_stack__bottom(thread, ts, sample, from_al, to_al,
 599                                            ref);
 600                 if (err)
 601                         return err;
 602         }
 603
 604         ts->branch_count += 1;
 605         ts->last_time = sample->time;
 606
 607         if (sample->flags & PERF_IP_FLAG_CALL) {
 608                 struct call_path_root *cpr = ts->crp->cpr;
 609                 struct call_path *cp;
 610                 u64 ret_addr;
 611
 612                 if (!sample->ip || !sample->addr)
 613                         return 0;
 614
 615                 ret_addr = sample->ip + sample->insn_len;
 616                 if (ret_addr == sample->addr)
 617                         return 0; /* Zero-length calls are excluded */
 618
 619                 cp = call_path__findnew(cpr, ts->stack[ts->cnt - 1].cp,
 620                                         to_al->sym, sample->addr,
 621                                         ts->kernel_start);
 622                 if (!cp)
 623                         return -ENOMEM;
 624                 err = thread_stack__push_cp(ts, ret_addr, sample->time, ref,
 625                                             cp, false);
 626         } else if (sample->flags & PERF_IP_FLAG_RETURN) {
 627                 if (!sample->ip || !sample->addr)
 628                         return 0;
 629
 630                 err = thread_stack__pop_cp(thread, ts, sample->addr,
 631                                            sample->time, ref, from_al->sym);
 632                 if (err) {
 633                         if (err < 0)
 634                                 return err;
 635                         err = thread_stack__no_call_return(thread, ts, sample,
 636                                                            from_al, to_al, ref);
 637                 }
 638         } else if (sample->flags & PERF_IP_FLAG_TRACE_BEGIN) {
 639                 err = thread_stack__trace_begin(thread, ts, sample->time, ref);
 640         } else if (sample->flags & PERF_IP_FLAG_TRACE_END) {
 641                 err = thread_stack__trace_end(ts, sample, ref);
 642         }
 643
 644         return err;
 645 }
 646
 647 size_t thread_stack__depth(struct thread *thread)
 648 {
 649         if (!thread->ts)
 650                 return 0;
 651         return thread->ts->cnt;
 652 }