1 /* $NetBSD: tprof.c,v 1.8 2011/02/05 14:04:40 yamt Exp $ */ 2 3 /*- 4 * Copyright (c)2008,2009,2010 YAMAMOTO Takashi, 5 * All rights reserved. 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 1. Redistributions of source code must retain the above copyright 11 * notice, this list of conditions and the following disclaimer. 12 * 2. Redistributions in binary form must reproduce the above copyright 13 * notice, this list of conditions and the following disclaimer in the 14 * documentation and/or other materials provided with the distribution. 15 * 16 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 17 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 18 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 19 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 20 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 21 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 22 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 23 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 24 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 25 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 26 * SUCH DAMAGE. 27 */ 28 29 #include <sys/cdefs.h> 30 __KERNEL_RCSID(0, "$NetBSD: tprof.c,v 1.8 2011/02/05 14:04:40 yamt Exp $"); 31 32 #include <sys/param.h> 33 #include <sys/systm.h> 34 #include <sys/kernel.h> 35 36 #include <sys/cpu.h> 37 #include <sys/conf.h> 38 #include <sys/callout.h> 39 #include <sys/kmem.h> 40 #include <sys/module.h> 41 #include <sys/proc.h> 42 #include <sys/workqueue.h> 43 #include <sys/queue.h> 44 45 #include <dev/tprof/tprof.h> 46 #include <dev/tprof/tprof_ioctl.h> 47 48 /* 49 * locking order: 50 * tprof_reader_lock -> tprof_lock 51 * tprof_startstop_lock -> tprof_lock 52 */ 53 54 /* 55 * protected by: 56 * L: tprof_lock 57 * R: tprof_reader_lock 58 * S: tprof_startstop_lock 59 * s: writer should hold tprof_startstop_lock and tprof_lock 60 * reader should hold tprof_startstop_lock or tprof_lock 61 */ 62 63 typedef struct tprof_buf { 64 u_int b_used; 65 u_int b_size; 66 u_int b_overflow; 67 u_int b_unused; 68 STAILQ_ENTRY(tprof_buf) b_list; 69 tprof_sample_t b_data[]; 70 } tprof_buf_t; 71 #define TPROF_BUF_BYTESIZE(sz) \ 72 (sizeof(tprof_buf_t) + (sz) * sizeof(tprof_sample_t)) 73 #define TPROF_MAX_SAMPLES_PER_BUF 10000 74 75 #define TPROF_MAX_BUF 100 76 77 typedef struct { 78 tprof_buf_t *c_buf; 79 struct work c_work; 80 callout_t c_callout; 81 } __aligned(CACHE_LINE_SIZE) tprof_cpu_t; 82 83 typedef struct tprof_backend { 84 const char *tb_name; 85 const tprof_backend_ops_t *tb_ops; 86 LIST_ENTRY(tprof_backend) tb_list; 87 int tb_usecount; /* S: */ 88 } tprof_backend_t; 89 90 static kmutex_t tprof_lock; 91 static bool tprof_running; /* s: */ 92 static u_int tprof_nworker; /* L: # of running worker LWPs */ 93 static lwp_t *tprof_owner; 94 static STAILQ_HEAD(, tprof_buf) tprof_list; /* L: global buffer list */ 95 static u_int tprof_nbuf_on_list; /* L: # of buffers on tprof_list */ 96 static struct workqueue *tprof_wq; 97 static tprof_cpu_t tprof_cpus[MAXCPUS] __aligned(CACHE_LINE_SIZE); 98 static u_int tprof_samples_per_buf; 99 100 static tprof_backend_t *tprof_backend; /* S: */ 101 static LIST_HEAD(, tprof_backend) tprof_backends = 102 LIST_HEAD_INITIALIZER(tprof_backend); /* S: */ 103 104 static kmutex_t tprof_reader_lock; 105 static kcondvar_t tprof_reader_cv; /* L: */ 106 static off_t tprof_reader_offset; /* R: */ 107 108 static kmutex_t tprof_startstop_lock; 109 static kcondvar_t tprof_cv; /* L: */ 110 111 static struct tprof_stat tprof_stat; /* L: */ 112 113 static tprof_cpu_t * 114 tprof_cpu(struct cpu_info *ci) 115 { 116 117 return &tprof_cpus[cpu_index(ci)]; 118 } 119 120 static tprof_cpu_t * 121 tprof_curcpu(void) 122 { 123 124 return tprof_cpu(curcpu()); 125 } 126 127 static tprof_buf_t * 128 tprof_buf_alloc(void) 129 { 130 tprof_buf_t *new; 131 u_int size = tprof_samples_per_buf; 132 133 new = kmem_alloc(TPROF_BUF_BYTESIZE(size), KM_SLEEP); 134 new->b_used = 0; 135 new->b_size = size; 136 new->b_overflow = 0; 137 return new; 138 } 139 140 static void 141 tprof_buf_free(tprof_buf_t *buf) 142 { 143 144 kmem_free(buf, TPROF_BUF_BYTESIZE(buf->b_size)); 145 } 146 147 static tprof_buf_t * 148 tprof_buf_switch(tprof_cpu_t *c, tprof_buf_t *new) 149 { 150 tprof_buf_t *old; 151 152 old = c->c_buf; 153 c->c_buf = new; 154 return old; 155 } 156 157 static tprof_buf_t * 158 tprof_buf_refresh(void) 159 { 160 tprof_cpu_t * const c = tprof_curcpu(); 161 tprof_buf_t *new; 162 163 new = tprof_buf_alloc(); 164 return tprof_buf_switch(c, new); 165 } 166 167 static void 168 tprof_worker(struct work *wk, void *dummy) 169 { 170 tprof_cpu_t * const c = tprof_curcpu(); 171 tprof_buf_t *buf; 172 bool shouldstop; 173 174 KASSERT(wk == &c->c_work); 175 KASSERT(dummy == NULL); 176 177 /* 178 * get a per cpu buffer. 179 */ 180 buf = tprof_buf_refresh(); 181 182 /* 183 * and put it on the global list for read(2). 184 */ 185 mutex_enter(&tprof_lock); 186 shouldstop = !tprof_running; 187 if (shouldstop) { 188 KASSERT(tprof_nworker > 0); 189 tprof_nworker--; 190 cv_broadcast(&tprof_cv); 191 cv_broadcast(&tprof_reader_cv); 192 } 193 if (buf->b_used == 0) { 194 tprof_stat.ts_emptybuf++; 195 } else if (tprof_nbuf_on_list < TPROF_MAX_BUF) { 196 tprof_stat.ts_sample += buf->b_used; 197 tprof_stat.ts_overflow += buf->b_overflow; 198 tprof_stat.ts_buf++; 199 STAILQ_INSERT_TAIL(&tprof_list, buf, b_list); 200 tprof_nbuf_on_list++; 201 buf = NULL; 202 cv_broadcast(&tprof_reader_cv); 203 } else { 204 tprof_stat.ts_dropbuf_sample += buf->b_used; 205 tprof_stat.ts_dropbuf++; 206 } 207 mutex_exit(&tprof_lock); 208 if (buf) { 209 tprof_buf_free(buf); 210 } 211 if (!shouldstop) { 212 callout_schedule(&c->c_callout, hz); 213 } 214 } 215 216 static void 217 tprof_kick(void *vp) 218 { 219 struct cpu_info * const ci = vp; 220 tprof_cpu_t * const c = tprof_cpu(ci); 221 222 workqueue_enqueue(tprof_wq, &c->c_work, ci); 223 } 224 225 static void 226 tprof_stop1(void) 227 { 228 CPU_INFO_ITERATOR cii; 229 struct cpu_info *ci; 230 231 KASSERT(mutex_owned(&tprof_startstop_lock)); 232 KASSERT(tprof_nworker == 0); 233 234 for (CPU_INFO_FOREACH(cii, ci)) { 235 tprof_cpu_t * const c = tprof_cpu(ci); 236 tprof_buf_t *old; 237 238 old = tprof_buf_switch(c, NULL); 239 if (old != NULL) { 240 tprof_buf_free(old); 241 } 242 callout_destroy(&c->c_callout); 243 } 244 workqueue_destroy(tprof_wq); 245 } 246 247 static int 248 tprof_start(const struct tprof_param *param) 249 { 250 CPU_INFO_ITERATOR cii; 251 struct cpu_info *ci; 252 int error; 253 uint64_t freq; 254 tprof_backend_t *tb; 255 256 KASSERT(mutex_owned(&tprof_startstop_lock)); 257 if (tprof_running) { 258 error = EBUSY; 259 goto done; 260 } 261 262 tb = tprof_backend; 263 if (tb == NULL) { 264 error = ENOENT; 265 goto done; 266 } 267 if (tb->tb_usecount > 0) { 268 error = EBUSY; 269 goto done; 270 } 271 272 tb->tb_usecount++; 273 freq = tb->tb_ops->tbo_estimate_freq(); 274 tprof_samples_per_buf = MIN(freq * 2, TPROF_MAX_SAMPLES_PER_BUF); 275 276 error = workqueue_create(&tprof_wq, "tprofmv", tprof_worker, NULL, 277 PRI_NONE, IPL_SOFTCLOCK, WQ_MPSAFE | WQ_PERCPU); 278 if (error != 0) { 279 goto done; 280 } 281 282 for (CPU_INFO_FOREACH(cii, ci)) { 283 tprof_cpu_t * const c = tprof_cpu(ci); 284 tprof_buf_t *new; 285 tprof_buf_t *old; 286 287 new = tprof_buf_alloc(); 288 old = tprof_buf_switch(c, new); 289 if (old != NULL) { 290 tprof_buf_free(old); 291 } 292 callout_init(&c->c_callout, CALLOUT_MPSAFE); 293 callout_setfunc(&c->c_callout, tprof_kick, ci); 294 } 295 296 error = tb->tb_ops->tbo_start(NULL); 297 if (error != 0) { 298 tprof_stop1(); 299 goto done; 300 } 301 302 mutex_enter(&tprof_lock); 303 tprof_running = true; 304 mutex_exit(&tprof_lock); 305 for (CPU_INFO_FOREACH(cii, ci)) { 306 tprof_cpu_t * const c = tprof_cpu(ci); 307 308 mutex_enter(&tprof_lock); 309 tprof_nworker++; 310 mutex_exit(&tprof_lock); 311 workqueue_enqueue(tprof_wq, &c->c_work, ci); 312 } 313 done: 314 return error; 315 } 316 317 static void 318 tprof_stop(void) 319 { 320 tprof_backend_t *tb; 321 322 KASSERT(mutex_owned(&tprof_startstop_lock)); 323 if (!tprof_running) { 324 goto done; 325 } 326 327 tb = tprof_backend; 328 KASSERT(tb->tb_usecount > 0); 329 tb->tb_ops->tbo_stop(NULL); 330 tb->tb_usecount--; 331 332 mutex_enter(&tprof_lock); 333 tprof_running = false; 334 cv_broadcast(&tprof_reader_cv); 335 while (tprof_nworker > 0) { 336 cv_wait(&tprof_cv, &tprof_lock); 337 } 338 mutex_exit(&tprof_lock); 339 340 tprof_stop1(); 341 done: 342 ; 343 } 344 345 /* 346 * tprof_clear: drain unread samples. 347 */ 348 349 static void 350 tprof_clear(void) 351 { 352 tprof_buf_t *buf; 353 354 mutex_enter(&tprof_reader_lock); 355 mutex_enter(&tprof_lock); 356 while ((buf = STAILQ_FIRST(&tprof_list)) != NULL) { 357 if (buf != NULL) { 358 STAILQ_REMOVE_HEAD(&tprof_list, b_list); 359 KASSERT(tprof_nbuf_on_list > 0); 360 tprof_nbuf_on_list--; 361 mutex_exit(&tprof_lock); 362 tprof_buf_free(buf); 363 mutex_enter(&tprof_lock); 364 } 365 } 366 KASSERT(tprof_nbuf_on_list == 0); 367 mutex_exit(&tprof_lock); 368 tprof_reader_offset = 0; 369 mutex_exit(&tprof_reader_lock); 370 371 memset(&tprof_stat, 0, sizeof(tprof_stat)); 372 } 373 374 static tprof_backend_t * 375 tprof_backend_lookup(const char *name) 376 { 377 tprof_backend_t *tb; 378 379 KASSERT(mutex_owned(&tprof_startstop_lock)); 380 381 LIST_FOREACH(tb, &tprof_backends, tb_list) { 382 if (!strcmp(tb->tb_name, name)) { 383 return tb; 384 } 385 } 386 return NULL; 387 } 388 389 /* -------------------- backend interfaces */ 390 391 /* 392 * tprof_sample: record a sample on the per-cpu buffer. 393 * 394 * be careful; can be called in NMI context. 395 * we are bluntly assuming that curcpu() and curlwp->l_proc->p_pid are safe. 396 */ 397 398 void 399 tprof_sample(tprof_backend_cookie_t *cookie, const tprof_frame_info_t *tfi) 400 { 401 tprof_cpu_t * const c = tprof_curcpu(); 402 tprof_buf_t * const buf = c->c_buf; 403 tprof_sample_t *sp; 404 const uintptr_t pc = tfi->tfi_pc; 405 u_int idx; 406 407 idx = buf->b_used; 408 if (__predict_false(idx >= buf->b_size)) { 409 buf->b_overflow++; 410 return; 411 } 412 sp = &buf->b_data[idx]; 413 sp->s_pid = curlwp->l_proc->p_pid; 414 sp->s_flags = (tfi->tfi_inkernel) ? TPROF_SAMPLE_INKERNEL : 0; 415 sp->s_pc = pc; 416 buf->b_used = idx + 1; 417 } 418 419 /* 420 * tprof_backend_register: 421 */ 422 423 int 424 tprof_backend_register(const char *name, const tprof_backend_ops_t *ops, 425 int vers) 426 { 427 tprof_backend_t *tb; 428 429 if (vers != TPROF_BACKEND_VERSION) { 430 return EINVAL; 431 } 432 433 mutex_enter(&tprof_startstop_lock); 434 tb = tprof_backend_lookup(name); 435 if (tb != NULL) { 436 mutex_exit(&tprof_startstop_lock); 437 return EEXIST; 438 } 439 #if 1 /* XXX for now */ 440 if (!LIST_EMPTY(&tprof_backends)) { 441 mutex_exit(&tprof_startstop_lock); 442 return ENOTSUP; 443 } 444 #endif 445 tb = kmem_alloc(sizeof(*tb), KM_SLEEP); 446 tb->tb_name = name; 447 tb->tb_ops = ops; 448 tb->tb_usecount = 0; 449 LIST_INSERT_HEAD(&tprof_backends, tb, tb_list); 450 #if 1 /* XXX for now */ 451 if (tprof_backend == NULL) { 452 tprof_backend = tb; 453 } 454 #endif 455 mutex_exit(&tprof_startstop_lock); 456 457 return 0; 458 } 459 460 /* 461 * tprof_backend_unregister: 462 */ 463 464 int 465 tprof_backend_unregister(const char *name) 466 { 467 tprof_backend_t *tb; 468 469 mutex_enter(&tprof_startstop_lock); 470 tb = tprof_backend_lookup(name); 471 #if defined(DIAGNOSTIC) 472 if (tb == NULL) { 473 mutex_exit(&tprof_startstop_lock); 474 panic("%s: not found '%s'", __func__, name); 475 } 476 #endif /* defined(DIAGNOSTIC) */ 477 if (tb->tb_usecount > 0) { 478 mutex_exit(&tprof_startstop_lock); 479 return EBUSY; 480 } 481 #if 1 /* XXX for now */ 482 if (tprof_backend == tb) { 483 tprof_backend = NULL; 484 } 485 #endif 486 LIST_REMOVE(tb, tb_list); 487 mutex_exit(&tprof_startstop_lock); 488 489 kmem_free(tb, sizeof(*tb)); 490 491 return 0; 492 } 493 494 /* -------------------- cdevsw interfaces */ 495 496 void tprofattach(int); 497 498 static int 499 tprof_open(dev_t dev, int flags, int type, struct lwp *l) 500 { 501 502 if (minor(dev) != 0) { 503 return EXDEV; 504 } 505 mutex_enter(&tprof_lock); 506 if (tprof_owner != NULL) { 507 mutex_exit(&tprof_lock); 508 return EBUSY; 509 } 510 tprof_owner = curlwp; 511 mutex_exit(&tprof_lock); 512 513 return 0; 514 } 515 516 static int 517 tprof_close(dev_t dev, int flags, int type, struct lwp *l) 518 { 519 520 KASSERT(minor(dev) == 0); 521 522 mutex_enter(&tprof_startstop_lock); 523 mutex_enter(&tprof_lock); 524 tprof_owner = NULL; 525 mutex_exit(&tprof_lock); 526 tprof_stop(); 527 tprof_clear(); 528 mutex_exit(&tprof_startstop_lock); 529 530 return 0; 531 } 532 533 static int 534 tprof_read(dev_t dev, struct uio *uio, int flags) 535 { 536 tprof_buf_t *buf; 537 size_t bytes; 538 size_t resid; 539 size_t done; 540 int error = 0; 541 542 KASSERT(minor(dev) == 0); 543 mutex_enter(&tprof_reader_lock); 544 while (uio->uio_resid > 0 && error == 0) { 545 /* 546 * take the first buffer from the list. 547 */ 548 mutex_enter(&tprof_lock); 549 buf = STAILQ_FIRST(&tprof_list); 550 if (buf == NULL) { 551 if (tprof_nworker == 0) { 552 mutex_exit(&tprof_lock); 553 error = 0; 554 break; 555 } 556 mutex_exit(&tprof_reader_lock); 557 error = cv_wait_sig(&tprof_reader_cv, &tprof_lock); 558 mutex_exit(&tprof_lock); 559 mutex_enter(&tprof_reader_lock); 560 continue; 561 } 562 STAILQ_REMOVE_HEAD(&tprof_list, b_list); 563 KASSERT(tprof_nbuf_on_list > 0); 564 tprof_nbuf_on_list--; 565 mutex_exit(&tprof_lock); 566 567 /* 568 * copy it out. 569 */ 570 bytes = MIN(buf->b_used * sizeof(tprof_sample_t) - 571 tprof_reader_offset, uio->uio_resid); 572 resid = uio->uio_resid; 573 error = uiomove((char *)buf->b_data + tprof_reader_offset, 574 bytes, uio); 575 done = resid - uio->uio_resid; 576 tprof_reader_offset += done; 577 578 /* 579 * if we didn't consume the whole buffer, 580 * put it back to the list. 581 */ 582 if (tprof_reader_offset < 583 buf->b_used * sizeof(tprof_sample_t)) { 584 mutex_enter(&tprof_lock); 585 STAILQ_INSERT_HEAD(&tprof_list, buf, b_list); 586 tprof_nbuf_on_list++; 587 cv_broadcast(&tprof_reader_cv); 588 mutex_exit(&tprof_lock); 589 } else { 590 tprof_buf_free(buf); 591 tprof_reader_offset = 0; 592 } 593 } 594 mutex_exit(&tprof_reader_lock); 595 596 return error; 597 } 598 599 static int 600 tprof_ioctl(dev_t dev, u_long cmd, void *data, int flags, struct lwp *l) 601 { 602 const struct tprof_param *param; 603 int error = 0; 604 605 KASSERT(minor(dev) == 0); 606 607 switch (cmd) { 608 case TPROF_IOC_GETVERSION: 609 *(int *)data = TPROF_VERSION; 610 break; 611 case TPROF_IOC_START: 612 param = data; 613 mutex_enter(&tprof_startstop_lock); 614 error = tprof_start(param); 615 mutex_exit(&tprof_startstop_lock); 616 break; 617 case TPROF_IOC_STOP: 618 mutex_enter(&tprof_startstop_lock); 619 tprof_stop(); 620 mutex_exit(&tprof_startstop_lock); 621 break; 622 case TPROF_IOC_GETSTAT: 623 mutex_enter(&tprof_lock); 624 memcpy(data, &tprof_stat, sizeof(tprof_stat)); 625 mutex_exit(&tprof_lock); 626 break; 627 default: 628 error = EINVAL; 629 break; 630 } 631 632 return error; 633 } 634 635 const struct cdevsw tprof_cdevsw = { 636 .d_open = tprof_open, 637 .d_close = tprof_close, 638 .d_read = tprof_read, 639 .d_write = nowrite, 640 .d_ioctl = tprof_ioctl, 641 .d_stop = nostop, 642 .d_tty = notty, 643 .d_poll = nopoll, 644 .d_mmap = nommap, 645 .d_kqfilter = nokqfilter, 646 .d_flag = D_OTHER | D_MPSAFE, 647 }; 648 649 void 650 tprofattach(int nunits) 651 { 652 653 /* nothing */ 654 } 655 656 MODULE(MODULE_CLASS_DRIVER, tprof, NULL); 657 658 static void 659 tprof_driver_init(void) 660 { 661 662 mutex_init(&tprof_lock, MUTEX_DEFAULT, IPL_NONE); 663 mutex_init(&tprof_reader_lock, MUTEX_DEFAULT, IPL_NONE); 664 mutex_init(&tprof_startstop_lock, MUTEX_DEFAULT, IPL_NONE); 665 cv_init(&tprof_cv, "tprof"); 666 cv_init(&tprof_reader_cv, "tprof_rd"); 667 STAILQ_INIT(&tprof_list); 668 } 669 670 static void 671 tprof_driver_fini(void) 672 { 673 674 mutex_destroy(&tprof_lock); 675 mutex_destroy(&tprof_reader_lock); 676 mutex_destroy(&tprof_startstop_lock); 677 cv_destroy(&tprof_cv); 678 cv_destroy(&tprof_reader_cv); 679 } 680 681 static int 682 tprof_modcmd(modcmd_t cmd, void *arg) 683 { 684 685 switch (cmd) { 686 case MODULE_CMD_INIT: 687 tprof_driver_init(); 688 #if defined(_MODULE) 689 { 690 devmajor_t bmajor = NODEVMAJOR; 691 devmajor_t cmajor = NODEVMAJOR; 692 int error; 693 694 error = devsw_attach("tprof", NULL, &bmajor, 695 &tprof_cdevsw, &cmajor); 696 if (error) { 697 tprof_driver_fini(); 698 return error; 699 } 700 } 701 #endif /* defined(_MODULE) */ 702 return 0; 703 704 case MODULE_CMD_FINI: 705 #if defined(_MODULE) 706 { 707 int error; 708 error = devsw_detach(NULL, &tprof_cdevsw); 709 if (error) { 710 return error; 711 } 712 } 713 #endif /* defined(_MODULE) */ 714 tprof_driver_fini(); 715 return 0; 716 717 default: 718 return ENOTTY; 719 } 720 } 721