1 /*- 2 * Copyright (c) 1982, 1986, 1993 3 * The Regents of the University of California. All rights reserved. 4 * 5 * Redistribution and use in source and binary forms, with or without 6 * modification, are permitted provided that the following conditions 7 * are met: 8 * 1. Redistributions of source code must retain the above copyright 9 * notice, this list of conditions and the following disclaimer. 10 * 2. Redistributions in binary form must reproduce the above copyright 11 * notice, this list of conditions and the following disclaimer in the 12 * documentation and/or other materials provided with the distribution. 13 * 3. All advertising materials mentioning features or use of this software 14 * must display the following acknowledgement: 15 * This product includes software developed by the University of 16 * California, Berkeley and its contributors. 17 * 4. Neither the name of the University nor the names of its contributors 18 * may be used to endorse or promote products derived from this software 19 * without specific prior written permission. 20 * 21 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 22 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 23 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 24 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 25 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 26 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 27 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 28 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 29 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 30 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 31 * SUCH DAMAGE. 32 * 33 * @(#)subr_prof.c 8.3 (Berkeley) 9/23/93 34 */ 35 36 #include <sys/cdefs.h> 37 __FBSDID("$FreeBSD$"); 38 39 #include <sys/param.h> 40 #include <sys/systm.h> 41 #include <sys/sysproto.h> 42 #include <sys/kernel.h> 43 #include <sys/lock.h> 44 #include <sys/mutex.h> 45 #include <sys/proc.h> 46 #include <sys/resourcevar.h> 47 #include <sys/sysctl.h> 48 49 #include <machine/cpu.h> 50 51 #ifdef GPROF 52 #include <sys/malloc.h> 53 #include <sys/gmon.h> 54 #undef MCOUNT 55 56 static MALLOC_DEFINE(M_GPROF, "gprof", "kernel profiling buffer"); 57 58 static void kmstartup(void *); 59 SYSINIT(kmem, SI_SUB_KPROF, SI_ORDER_FIRST, kmstartup, NULL) 60 61 struct gmonparam _gmonparam = { GMON_PROF_OFF }; 62 63 #ifdef GUPROF 64 #include <machine/asmacros.h> 65 66 void 67 nullfunc_loop_profiled() 68 { 69 int i; 70 71 for (i = 0; i < CALIB_SCALE; i++) 72 nullfunc_profiled(); 73 } 74 75 #define nullfunc_loop_profiled_end nullfunc_profiled /* XXX */ 76 77 void 78 nullfunc_profiled() 79 { 80 } 81 #endif /* GUPROF */ 82 83 /* 84 * Update the histograms to support extending the text region arbitrarily. 85 * This is done slightly naively (no sparse regions), so will waste slight 86 * amounts of memory, but will overall work nicely enough to allow profiling 87 * of KLDs. 88 */ 89 void 90 kmupetext(uintfptr_t nhighpc) 91 { 92 struct gmonparam np; /* slightly large */ 93 struct gmonparam *p = &_gmonparam; 94 char *cp; 95 96 GIANT_REQUIRED; 97 bcopy(p, &np, sizeof(*p)); 98 np.highpc = ROUNDUP(nhighpc, HISTFRACTION * sizeof(HISTCOUNTER)); 99 if (np.highpc <= p->highpc) 100 return; 101 np.textsize = np.highpc - p->lowpc; 102 np.kcountsize = np.textsize / HISTFRACTION; 103 np.hashfraction = HASHFRACTION; 104 np.fromssize = np.textsize / HASHFRACTION; 105 np.tolimit = np.textsize * ARCDENSITY / 100; 106 if (np.tolimit < MINARCS) 107 np.tolimit = MINARCS; 108 else if (np.tolimit > MAXARCS) 109 np.tolimit = MAXARCS; 110 np.tossize = np.tolimit * sizeof(struct tostruct); 111 cp = malloc(np.kcountsize + np.fromssize + np.tossize, 112 M_GPROF, M_WAITOK); 113 /* 114 * Check for something else extending highpc while we slept. 115 */ 116 if (np.highpc <= p->highpc) { 117 free(cp, M_GPROF); 118 return; 119 } 120 np.tos = (struct tostruct *)cp; 121 cp += np.tossize; 122 np.kcount = (HISTCOUNTER *)cp; 123 cp += np.kcountsize; 124 np.froms = (u_short *)cp; 125 #ifdef GUPROF 126 /* Reinitialize pointers to overhead counters. */ 127 np.cputime_count = &KCOUNT(&np, PC_TO_I(&np, cputime)); 128 np.mcount_count = &KCOUNT(&np, PC_TO_I(&np, mcount)); 129 np.mexitcount_count = &KCOUNT(&np, PC_TO_I(&np, mexitcount)); 130 #endif 131 critical_enter(); 132 bcopy(p->tos, np.tos, p->tossize); 133 bzero((char *)np.tos + p->tossize, np.tossize - p->tossize); 134 bcopy(p->kcount, np.kcount, p->kcountsize); 135 bzero((char *)np.kcount + p->kcountsize, np.kcountsize - 136 p->kcountsize); 137 bcopy(p->froms, np.froms, p->fromssize); 138 bzero((char *)np.froms + p->fromssize, np.fromssize - p->fromssize); 139 cp = (char *)p->tos; 140 bcopy(&np, p, sizeof(*p)); 141 critical_exit(); 142 free(cp, M_GPROF); 143 } 144 145 static void 146 kmstartup(dummy) 147 void *dummy; 148 { 149 char *cp; 150 struct gmonparam *p = &_gmonparam; 151 #ifdef GUPROF 152 int cputime_overhead; 153 int empty_loop_time; 154 int i; 155 int mcount_overhead; 156 int mexitcount_overhead; 157 int nullfunc_loop_overhead; 158 int nullfunc_loop_profiled_time; 159 uintfptr_t tmp_addr; 160 #endif 161 162 /* 163 * Round lowpc and highpc to multiples of the density we're using 164 * so the rest of the scaling (here and in gprof) stays in ints. 165 */ 166 p->lowpc = ROUNDDOWN((u_long)btext, HISTFRACTION * sizeof(HISTCOUNTER)); 167 p->highpc = ROUNDUP((u_long)etext, HISTFRACTION * sizeof(HISTCOUNTER)); 168 p->textsize = p->highpc - p->lowpc; 169 printf("Profiling kernel, textsize=%lu [%x..%x]\n", 170 p->textsize, p->lowpc, p->highpc); 171 p->kcountsize = p->textsize / HISTFRACTION; 172 p->hashfraction = HASHFRACTION; 173 p->fromssize = p->textsize / HASHFRACTION; 174 p->tolimit = p->textsize * ARCDENSITY / 100; 175 if (p->tolimit < MINARCS) 176 p->tolimit = MINARCS; 177 else if (p->tolimit > MAXARCS) 178 p->tolimit = MAXARCS; 179 p->tossize = p->tolimit * sizeof(struct tostruct); 180 cp = (char *)malloc(p->kcountsize + p->fromssize + p->tossize, 181 M_GPROF, M_WAITOK | M_ZERO); 182 p->tos = (struct tostruct *)cp; 183 cp += p->tossize; 184 p->kcount = (HISTCOUNTER *)cp; 185 cp += p->kcountsize; 186 p->froms = (u_short *)cp; 187 188 #ifdef GUPROF 189 /* Initialize pointers to overhead counters. */ 190 p->cputime_count = &KCOUNT(p, PC_TO_I(p, cputime)); 191 p->mcount_count = &KCOUNT(p, PC_TO_I(p, mcount)); 192 p->mexitcount_count = &KCOUNT(p, PC_TO_I(p, mexitcount)); 193 194 /* 195 * Disable interrupts to avoid interference while we calibrate 196 * things. 197 */ 198 critical_enter(); 199 200 /* 201 * Determine overheads. 202 * XXX this needs to be repeated for each useful timer/counter. 203 */ 204 cputime_overhead = 0; 205 startguprof(p); 206 for (i = 0; i < CALIB_SCALE; i++) 207 cputime_overhead += cputime(); 208 209 empty_loop(); 210 startguprof(p); 211 empty_loop(); 212 empty_loop_time = cputime(); 213 214 nullfunc_loop_profiled(); 215 216 /* 217 * Start profiling. There won't be any normal function calls since 218 * interrupts are disabled, but we will call the profiling routines 219 * directly to determine their overheads. 220 */ 221 p->state = GMON_PROF_HIRES; 222 223 startguprof(p); 224 nullfunc_loop_profiled(); 225 226 startguprof(p); 227 for (i = 0; i < CALIB_SCALE; i++) 228 #if defined(__i386__) && __GNUC__ >= 2 229 __asm("pushl %0; call __mcount; popl %%ecx" 230 : 231 : "i" (profil) 232 : "ax", "bx", "cx", "dx", "memory"); 233 #elif defined(lint) 234 #else 235 #error 236 #endif 237 mcount_overhead = KCOUNT(p, PC_TO_I(p, profil)); 238 239 startguprof(p); 240 for (i = 0; i < CALIB_SCALE; i++) 241 #if defined(__i386__) && __GNUC__ >= 2 242 __asm("call " __XSTRING(HIDENAME(mexitcount)) "; 1:" 243 : : : "ax", "bx", "cx", "dx", "memory"); 244 __asm("movl $1b,%0" : "=rm" (tmp_addr)); 245 #elif defined(lint) 246 #else 247 #error 248 #endif 249 mexitcount_overhead = KCOUNT(p, PC_TO_I(p, tmp_addr)); 250 251 p->state = GMON_PROF_OFF; 252 stopguprof(p); 253 254 critical_exit(); 255 256 nullfunc_loop_profiled_time = 0; 257 for (tmp_addr = (uintfptr_t)nullfunc_loop_profiled; 258 tmp_addr < (uintfptr_t)nullfunc_loop_profiled_end; 259 tmp_addr += HISTFRACTION * sizeof(HISTCOUNTER)) 260 nullfunc_loop_profiled_time += KCOUNT(p, PC_TO_I(p, tmp_addr)); 261 #define CALIB_DOSCALE(count) (((count) + CALIB_SCALE / 3) / CALIB_SCALE) 262 #define c2n(count, freq) ((int)((count) * 1000000000LL / freq)) 263 printf("cputime %d, empty_loop %d, nullfunc_loop_profiled %d, mcount %d, mexitcount %d\n", 264 CALIB_DOSCALE(c2n(cputime_overhead, p->profrate)), 265 CALIB_DOSCALE(c2n(empty_loop_time, p->profrate)), 266 CALIB_DOSCALE(c2n(nullfunc_loop_profiled_time, p->profrate)), 267 CALIB_DOSCALE(c2n(mcount_overhead, p->profrate)), 268 CALIB_DOSCALE(c2n(mexitcount_overhead, p->profrate))); 269 cputime_overhead -= empty_loop_time; 270 mcount_overhead -= empty_loop_time; 271 mexitcount_overhead -= empty_loop_time; 272 273 /*- 274 * Profiling overheads are determined by the times between the 275 * following events: 276 * MC1: mcount() is called 277 * MC2: cputime() (called from mcount()) latches the timer 278 * MC3: mcount() completes 279 * ME1: mexitcount() is called 280 * ME2: cputime() (called from mexitcount()) latches the timer 281 * ME3: mexitcount() completes. 282 * The times between the events vary slightly depending on instruction 283 * combination and cache misses, etc. Attempt to determine the 284 * minimum times. These can be subtracted from the profiling times 285 * without much risk of reducing the profiling times below what they 286 * would be when profiling is not configured. Abbreviate: 287 * ab = minimum time between MC1 and MC3 288 * a = minumum time between MC1 and MC2 289 * b = minimum time between MC2 and MC3 290 * cd = minimum time between ME1 and ME3 291 * c = minimum time between ME1 and ME2 292 * d = minimum time between ME2 and ME3. 293 * These satisfy the relations: 294 * ab <= mcount_overhead (just measured) 295 * a + b <= ab 296 * cd <= mexitcount_overhead (just measured) 297 * c + d <= cd 298 * a + d <= nullfunc_loop_profiled_time (just measured) 299 * a >= 0, b >= 0, c >= 0, d >= 0. 300 * Assume that ab and cd are equal to the minimums. 301 */ 302 p->cputime_overhead = CALIB_DOSCALE(cputime_overhead); 303 p->mcount_overhead = CALIB_DOSCALE(mcount_overhead - cputime_overhead); 304 p->mexitcount_overhead = CALIB_DOSCALE(mexitcount_overhead 305 - cputime_overhead); 306 nullfunc_loop_overhead = nullfunc_loop_profiled_time - empty_loop_time; 307 p->mexitcount_post_overhead = CALIB_DOSCALE((mcount_overhead 308 - nullfunc_loop_overhead) 309 / 4); 310 p->mexitcount_pre_overhead = p->mexitcount_overhead 311 + p->cputime_overhead 312 - p->mexitcount_post_overhead; 313 p->mcount_pre_overhead = CALIB_DOSCALE(nullfunc_loop_overhead) 314 - p->mexitcount_post_overhead; 315 p->mcount_post_overhead = p->mcount_overhead 316 + p->cputime_overhead 317 - p->mcount_pre_overhead; 318 printf( 319 "Profiling overheads: mcount: %d+%d, %d+%d; mexitcount: %d+%d, %d+%d nsec\n", 320 c2n(p->cputime_overhead, p->profrate), 321 c2n(p->mcount_overhead, p->profrate), 322 c2n(p->mcount_pre_overhead, p->profrate), 323 c2n(p->mcount_post_overhead, p->profrate), 324 c2n(p->cputime_overhead, p->profrate), 325 c2n(p->mexitcount_overhead, p->profrate), 326 c2n(p->mexitcount_pre_overhead, p->profrate), 327 c2n(p->mexitcount_post_overhead, p->profrate)); 328 printf( 329 "Profiling overheads: mcount: %d+%d, %d+%d; mexitcount: %d+%d, %d+%d cycles\n", 330 p->cputime_overhead, p->mcount_overhead, 331 p->mcount_pre_overhead, p->mcount_post_overhead, 332 p->cputime_overhead, p->mexitcount_overhead, 333 p->mexitcount_pre_overhead, p->mexitcount_post_overhead); 334 #endif /* GUPROF */ 335 } 336 337 /* 338 * Return kernel profiling information. 339 */ 340 static int 341 sysctl_kern_prof(SYSCTL_HANDLER_ARGS) 342 { 343 int *name = (int *) arg1; 344 u_int namelen = arg2; 345 struct gmonparam *gp = &_gmonparam; 346 int error; 347 int state; 348 349 /* all sysctl names at this level are terminal */ 350 if (namelen != 1) 351 return (ENOTDIR); /* overloaded */ 352 353 switch (name[0]) { 354 case GPROF_STATE: 355 state = gp->state; 356 error = sysctl_handle_int(oidp, &state, 0, req); 357 if (error) 358 return (error); 359 if (!req->newptr) 360 return (0); 361 if (state == GMON_PROF_OFF) { 362 gp->state = state; 363 PROC_LOCK(&proc0); 364 stopprofclock(&proc0); 365 PROC_UNLOCK(&proc0); 366 stopguprof(gp); 367 } else if (state == GMON_PROF_ON) { 368 gp->state = GMON_PROF_OFF; 369 stopguprof(gp); 370 gp->profrate = profhz; 371 PROC_LOCK(&proc0); 372 startprofclock(&proc0); 373 PROC_UNLOCK(&proc0); 374 gp->state = state; 375 #ifdef GUPROF 376 } else if (state == GMON_PROF_HIRES) { 377 gp->state = GMON_PROF_OFF; 378 PROC_LOCK(&proc0); 379 stopprofclock(&proc0); 380 PROC_UNLOCK(&proc0); 381 startguprof(gp); 382 gp->state = state; 383 #endif 384 } else if (state != gp->state) 385 return (EINVAL); 386 return (0); 387 case GPROF_COUNT: 388 return (sysctl_handle_opaque(oidp, 389 gp->kcount, gp->kcountsize, req)); 390 case GPROF_FROMS: 391 return (sysctl_handle_opaque(oidp, 392 gp->froms, gp->fromssize, req)); 393 case GPROF_TOS: 394 return (sysctl_handle_opaque(oidp, 395 gp->tos, gp->tossize, req)); 396 case GPROF_GMONPARAM: 397 return (sysctl_handle_opaque(oidp, gp, sizeof *gp, req)); 398 default: 399 return (EOPNOTSUPP); 400 } 401 /* NOTREACHED */ 402 } 403 404 SYSCTL_NODE(_kern, KERN_PROF, prof, CTLFLAG_RW, sysctl_kern_prof, ""); 405 #endif /* GPROF */ 406 407 /* 408 * Profiling system call. 409 * 410 * The scale factor is a fixed point number with 16 bits of fraction, so that 411 * 1.0 is represented as 0x10000. A scale factor of 0 turns off profiling. 412 */ 413 #ifndef _SYS_SYSPROTO_H_ 414 struct profil_args { 415 caddr_t samples; 416 size_t size; 417 size_t offset; 418 u_int scale; 419 }; 420 #endif 421 /* 422 * MPSAFE 423 */ 424 /* ARGSUSED */ 425 int 426 profil(td, uap) 427 struct thread *td; 428 register struct profil_args *uap; 429 { 430 struct uprof *upp; 431 struct proc *p; 432 433 if (uap->scale > (1 << 16)) 434 return (EINVAL); 435 436 p = td->td_proc; 437 if (uap->scale == 0) { 438 PROC_LOCK(td->td_proc); 439 stopprofclock(td->td_proc); 440 PROC_UNLOCK(td->td_proc); 441 return (0); 442 } 443 upp = &td->td_proc->p_stats->p_prof; 444 upp->pr_off = uap->offset; 445 upp->pr_scale = uap->scale; 446 upp->pr_base = uap->samples; 447 upp->pr_size = uap->size; 448 PROC_LOCK(p); 449 startprofclock(p); 450 PROC_UNLOCK(p); 451 452 return (0); 453 } 454 455 /* 456 * Scale is a fixed-point number with the binary point 16 bits 457 * into the value, and is <= 1.0. pc is at most 32 bits, so the 458 * intermediate result is at most 48 bits. 459 */ 460 #define PC_TO_INDEX(pc, prof) \ 461 ((int)(((u_quad_t)((pc) - (prof)->pr_off) * \ 462 (u_quad_t)((prof)->pr_scale)) >> 16) & ~1) 463 464 /* 465 * Collect user-level profiling statistics; called on a profiling tick, 466 * when a process is running in user-mode. This routine may be called 467 * from an interrupt context. We try to update the user profiling buffers 468 * cheaply with fuswintr() and suswintr(). If that fails, we revert to 469 * an AST that will vector us to trap() with a context in which copyin 470 * and copyout will work. Trap will then call addupc_task(). 471 * 472 * Note that we may (rarely) not get around to the AST soon enough, and 473 * lose profile ticks when the next tick overwrites this one, but in this 474 * case the system is overloaded and the profile is probably already 475 * inaccurate. 476 */ 477 void 478 addupc_intr(struct thread *td, uintptr_t pc, u_int ticks) 479 { 480 struct uprof *prof; 481 caddr_t addr; 482 u_int i; 483 int v; 484 485 if (ticks == 0) 486 return; 487 prof = &td->td_proc->p_stats->p_prof; 488 if (pc < prof->pr_off || 489 (i = PC_TO_INDEX(pc, prof)) >= prof->pr_size) 490 return; /* out of range; ignore */ 491 492 addr = prof->pr_base + i; 493 if ((v = fuswintr(addr)) == -1 || suswintr(addr, v + ticks) == -1) { 494 mtx_lock_spin(&sched_lock); 495 prof->pr_addr = pc; 496 prof->pr_ticks = ticks; 497 td->td_flags |= TDF_OWEUPC | TDF_ASTPENDING ; 498 mtx_unlock_spin(&sched_lock); 499 } 500 } 501 502 /* 503 * Much like before, but we can afford to take faults here. If the 504 * update fails, we simply turn off profiling. 505 */ 506 void 507 addupc_task(struct thread *td, uintptr_t pc, u_int ticks) 508 { 509 struct proc *p = td->td_proc; 510 struct uprof *prof; 511 caddr_t addr; 512 u_int i; 513 u_short v; 514 int stop = 0; 515 516 if (ticks == 0) 517 return; 518 519 PROC_LOCK(p); 520 if (!(p->p_flag & P_PROFIL)) { 521 PROC_UNLOCK(p); 522 return; 523 } 524 p->p_profthreads++; 525 PROC_UNLOCK(p); 526 prof = &p->p_stats->p_prof; 527 if (pc < prof->pr_off || 528 (i = PC_TO_INDEX(pc, prof)) >= prof->pr_size) { 529 goto out; 530 } 531 532 addr = prof->pr_base + i; 533 if (copyin(addr, &v, sizeof(v)) == 0) { 534 v += ticks; 535 if (copyout(&v, addr, sizeof(v)) == 0) 536 goto out; 537 } 538 stop = 1; 539 540 out: 541 PROC_LOCK(p); 542 if (--p->p_profthreads == 0) { 543 if (p->p_flag & P_STOPPROF) { 544 wakeup(&p->p_profthreads); 545 stop = 0; 546 } 547 } 548 if (stop) 549 stopprofclock(p); 550 PROC_UNLOCK(p); 551 } 552 553 #if defined(__i386__) && __GNUC__ >= 2 554 /* 555 * Support for "--test-coverage --profile-arcs" in GCC. 556 * 557 * We need to call all the functions in the .ctor section, in order 558 * to get all the counter-arrays strung into a list. 559 * 560 * XXX: the .ctors call __bb_init_func which is located in over in 561 * XXX: i386/i386/support.s for historical reasons. There is probably 562 * XXX: no reason for that to be assembler anymore, but doing it right 563 * XXX: in MI C code requires one to reverse-engineer the type-selection 564 * XXX: inside GCC. Have fun. 565 * 566 * XXX: Worrisome perspective: Calling the .ctors may make C++ in the 567 * XXX: kernel feasible. Don't. 568 */ 569 typedef void (*ctor_t)(void); 570 extern ctor_t _start_ctors, _stop_ctors; 571 572 static void 573 tcov_init(void *foo __unused) 574 { 575 ctor_t *p, q; 576 577 for (p = &_start_ctors; p < &_stop_ctors; p++) { 578 q = *p; 579 q(); 580 } 581 } 582 583 SYSINIT(tcov_init, SI_SUB_KPROF, SI_ORDER_SECOND, tcov_init, NULL) 584 585 /* 586 * GCC contains magic to recognize calls to for instance execve() and 587 * puts in calls to this function to preserve the profile counters. 588 * XXX: Put zinging punchline here. 589 */ 590 void __bb_fork_func(void); 591 void 592 __bb_fork_func(void) 593 { 594 } 595 596 #endif 597 598