1 /*- 2 * SPDX-License-Identifier: BSD-3-Clause 3 * 4 * Copyright (c) 1982, 1986, 1993 5 * The Regents of the University of California. 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 * 3. Neither the name of the University nor the names of its contributors 16 * may be used to endorse or promote products derived from this software 17 * without specific prior written permission. 18 * 19 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 20 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 21 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 22 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 23 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 24 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 25 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 26 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 27 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 28 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 29 * SUCH DAMAGE. 30 * 31 * @(#)subr_prof.c 8.3 (Berkeley) 9/23/93 32 */ 33 34 #include <sys/cdefs.h> 35 __FBSDID("$FreeBSD$"); 36 37 #include <sys/param.h> 38 #include <sys/systm.h> 39 #include <sys/sysproto.h> 40 #include <sys/kernel.h> 41 #include <sys/lock.h> 42 #include <sys/mutex.h> 43 #include <sys/proc.h> 44 #include <sys/resourcevar.h> 45 #include <sys/sysctl.h> 46 47 #include <machine/cpu.h> 48 49 #ifdef GPROF 50 #include <sys/malloc.h> 51 #include <sys/gmon.h> 52 #undef MCOUNT 53 54 static MALLOC_DEFINE(M_GPROF, "gprof", "kernel profiling buffer"); 55 56 static void kmstartup(void *); 57 SYSINIT(kmem, SI_SUB_KPROF, SI_ORDER_FIRST, kmstartup, NULL); 58 59 struct gmonparam _gmonparam = { GMON_PROF_OFF }; 60 61 #ifdef GUPROF 62 void 63 nullfunc_loop_profiled() 64 { 65 int i; 66 67 for (i = 0; i < CALIB_SCALE; i++) 68 nullfunc_profiled(); 69 } 70 71 #define nullfunc_loop_profiled_end nullfunc_profiled /* XXX */ 72 73 void 74 nullfunc_profiled() 75 { 76 } 77 #endif /* GUPROF */ 78 79 /* 80 * Update the histograms to support extending the text region arbitrarily. 81 * This is done slightly naively (no sparse regions), so will waste slight 82 * amounts of memory, but will overall work nicely enough to allow profiling 83 * of KLDs. 84 */ 85 void 86 kmupetext(uintfptr_t nhighpc) 87 { 88 struct gmonparam np; /* slightly large */ 89 struct gmonparam *p = &_gmonparam; 90 char *cp; 91 92 GIANT_REQUIRED; 93 bcopy(p, &np, sizeof(*p)); 94 np.highpc = ROUNDUP(nhighpc, HISTFRACTION * sizeof(HISTCOUNTER)); 95 if (np.highpc <= p->highpc) 96 return; 97 np.textsize = np.highpc - p->lowpc; 98 np.kcountsize = np.textsize / HISTFRACTION; 99 np.hashfraction = HASHFRACTION; 100 np.fromssize = np.textsize / HASHFRACTION; 101 np.tolimit = np.textsize * ARCDENSITY / 100; 102 if (np.tolimit < MINARCS) 103 np.tolimit = MINARCS; 104 else if (np.tolimit > MAXARCS) 105 np.tolimit = MAXARCS; 106 np.tossize = np.tolimit * sizeof(struct tostruct); 107 cp = malloc(np.kcountsize + np.fromssize + np.tossize, 108 M_GPROF, M_WAITOK); 109 /* 110 * Check for something else extending highpc while we slept. 111 */ 112 if (np.highpc <= p->highpc) { 113 free(cp, M_GPROF); 114 return; 115 } 116 np.tos = (struct tostruct *)cp; 117 cp += np.tossize; 118 np.kcount = (HISTCOUNTER *)cp; 119 cp += np.kcountsize; 120 np.froms = (u_short *)cp; 121 #ifdef GUPROF 122 /* Reinitialize pointers to overhead counters. */ 123 np.cputime_count = &KCOUNT(&np, PC_TO_I(&np, cputime)); 124 np.mcount_count = &KCOUNT(&np, PC_TO_I(&np, mcount)); 125 np.mexitcount_count = &KCOUNT(&np, PC_TO_I(&np, mexitcount)); 126 #endif 127 critical_enter(); 128 bcopy(p->tos, np.tos, p->tossize); 129 bzero((char *)np.tos + p->tossize, np.tossize - p->tossize); 130 bcopy(p->kcount, np.kcount, p->kcountsize); 131 bzero((char *)np.kcount + p->kcountsize, np.kcountsize - 132 p->kcountsize); 133 bcopy(p->froms, np.froms, p->fromssize); 134 bzero((char *)np.froms + p->fromssize, np.fromssize - p->fromssize); 135 cp = (char *)p->tos; 136 bcopy(&np, p, sizeof(*p)); 137 critical_exit(); 138 free(cp, M_GPROF); 139 } 140 141 static void 142 kmstartup(dummy) 143 void *dummy; 144 { 145 char *cp; 146 struct gmonparam *p = &_gmonparam; 147 #ifdef GUPROF 148 int cputime_overhead; 149 int empty_loop_time; 150 int i; 151 int mcount_overhead; 152 int mexitcount_overhead; 153 int nullfunc_loop_overhead; 154 int nullfunc_loop_profiled_time; 155 uintfptr_t tmp_addr; 156 #endif 157 158 /* 159 * Round lowpc and highpc to multiples of the density we're using 160 * so the rest of the scaling (here and in gprof) stays in ints. 161 */ 162 p->lowpc = ROUNDDOWN((u_long)btext, HISTFRACTION * sizeof(HISTCOUNTER)); 163 p->highpc = ROUNDUP((u_long)etext, HISTFRACTION * sizeof(HISTCOUNTER)); 164 p->textsize = p->highpc - p->lowpc; 165 printf("Profiling kernel, textsize=%lu [%jx..%jx]\n", 166 p->textsize, (uintmax_t)p->lowpc, (uintmax_t)p->highpc); 167 p->kcountsize = p->textsize / HISTFRACTION; 168 p->hashfraction = HASHFRACTION; 169 p->fromssize = p->textsize / HASHFRACTION; 170 p->tolimit = p->textsize * ARCDENSITY / 100; 171 if (p->tolimit < MINARCS) 172 p->tolimit = MINARCS; 173 else if (p->tolimit > MAXARCS) 174 p->tolimit = MAXARCS; 175 p->tossize = p->tolimit * sizeof(struct tostruct); 176 cp = (char *)malloc(p->kcountsize + p->fromssize + p->tossize, 177 M_GPROF, M_WAITOK | M_ZERO); 178 p->tos = (struct tostruct *)cp; 179 cp += p->tossize; 180 p->kcount = (HISTCOUNTER *)cp; 181 cp += p->kcountsize; 182 p->froms = (u_short *)cp; 183 p->histcounter_type = FUNCTION_ALIGNMENT / HISTFRACTION * NBBY; 184 185 #ifdef GUPROF 186 /* Signed counters. */ 187 p->histcounter_type = -p->histcounter_type; 188 189 /* 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 MCOUNT_OVERHEAD(sys_profil); 229 mcount_overhead = KCOUNT(p, PC_TO_I(p, sys_profil)); 230 231 startguprof(p); 232 for (i = 0; i < CALIB_SCALE; i++) 233 MEXITCOUNT_OVERHEAD(); 234 MEXITCOUNT_OVERHEAD_GETLABEL(tmp_addr); 235 mexitcount_overhead = KCOUNT(p, PC_TO_I(p, tmp_addr)); 236 237 p->state = GMON_PROF_OFF; 238 stopguprof(p); 239 240 critical_exit(); 241 242 nullfunc_loop_profiled_time = 0; 243 for (tmp_addr = (uintfptr_t)nullfunc_loop_profiled; 244 tmp_addr < (uintfptr_t)nullfunc_loop_profiled_end; 245 tmp_addr += HISTFRACTION * sizeof(HISTCOUNTER)) 246 nullfunc_loop_profiled_time += KCOUNT(p, PC_TO_I(p, tmp_addr)); 247 #define CALIB_DOSCALE(count) (((count) + CALIB_SCALE / 3) / CALIB_SCALE) 248 #define c2n(count, freq) ((int)((count) * 1000000000LL / freq)) 249 printf("cputime %d, empty_loop %d, nullfunc_loop_profiled %d, mcount %d, mexitcount %d\n", 250 CALIB_DOSCALE(c2n(cputime_overhead, p->profrate)), 251 CALIB_DOSCALE(c2n(empty_loop_time, p->profrate)), 252 CALIB_DOSCALE(c2n(nullfunc_loop_profiled_time, p->profrate)), 253 CALIB_DOSCALE(c2n(mcount_overhead, p->profrate)), 254 CALIB_DOSCALE(c2n(mexitcount_overhead, p->profrate))); 255 cputime_overhead -= empty_loop_time; 256 mcount_overhead -= empty_loop_time; 257 mexitcount_overhead -= empty_loop_time; 258 259 /*- 260 * Profiling overheads are determined by the times between the 261 * following events: 262 * MC1: mcount() is called 263 * MC2: cputime() (called from mcount()) latches the timer 264 * MC3: mcount() completes 265 * ME1: mexitcount() is called 266 * ME2: cputime() (called from mexitcount()) latches the timer 267 * ME3: mexitcount() completes. 268 * The times between the events vary slightly depending on instruction 269 * combination and cache misses, etc. Attempt to determine the 270 * minimum times. These can be subtracted from the profiling times 271 * without much risk of reducing the profiling times below what they 272 * would be when profiling is not configured. Abbreviate: 273 * ab = minimum time between MC1 and MC3 274 * a = minimum time between MC1 and MC2 275 * b = minimum time between MC2 and MC3 276 * cd = minimum time between ME1 and ME3 277 * c = minimum time between ME1 and ME2 278 * d = minimum time between ME2 and ME3. 279 * These satisfy the relations: 280 * ab <= mcount_overhead (just measured) 281 * a + b <= ab 282 * cd <= mexitcount_overhead (just measured) 283 * c + d <= cd 284 * a + d <= nullfunc_loop_profiled_time (just measured) 285 * a >= 0, b >= 0, c >= 0, d >= 0. 286 * Assume that ab and cd are equal to the minimums. 287 */ 288 p->cputime_overhead = CALIB_DOSCALE(cputime_overhead); 289 p->mcount_overhead = CALIB_DOSCALE(mcount_overhead - cputime_overhead); 290 p->mexitcount_overhead = CALIB_DOSCALE(mexitcount_overhead 291 - cputime_overhead); 292 nullfunc_loop_overhead = nullfunc_loop_profiled_time - empty_loop_time; 293 p->mexitcount_post_overhead = CALIB_DOSCALE((mcount_overhead 294 - nullfunc_loop_overhead) 295 / 4); 296 p->mexitcount_pre_overhead = p->mexitcount_overhead 297 + p->cputime_overhead 298 - p->mexitcount_post_overhead; 299 p->mcount_pre_overhead = CALIB_DOSCALE(nullfunc_loop_overhead) 300 - p->mexitcount_post_overhead; 301 p->mcount_post_overhead = p->mcount_overhead 302 + p->cputime_overhead 303 - p->mcount_pre_overhead; 304 printf( 305 "Profiling overheads: mcount: %d+%d, %d+%d; mexitcount: %d+%d, %d+%d nsec\n", 306 c2n(p->cputime_overhead, p->profrate), 307 c2n(p->mcount_overhead, p->profrate), 308 c2n(p->mcount_pre_overhead, p->profrate), 309 c2n(p->mcount_post_overhead, p->profrate), 310 c2n(p->cputime_overhead, p->profrate), 311 c2n(p->mexitcount_overhead, p->profrate), 312 c2n(p->mexitcount_pre_overhead, p->profrate), 313 c2n(p->mexitcount_post_overhead, p->profrate)); 314 printf( 315 "Profiling overheads: mcount: %d+%d, %d+%d; mexitcount: %d+%d, %d+%d cycles\n", 316 p->cputime_overhead, p->mcount_overhead, 317 p->mcount_pre_overhead, p->mcount_post_overhead, 318 p->cputime_overhead, p->mexitcount_overhead, 319 p->mexitcount_pre_overhead, p->mexitcount_post_overhead); 320 #endif /* GUPROF */ 321 } 322 323 /* 324 * Return kernel profiling information. 325 */ 326 static int 327 sysctl_kern_prof(SYSCTL_HANDLER_ARGS) 328 { 329 int *name = (int *) arg1; 330 u_int namelen = arg2; 331 struct gmonparam *gp = &_gmonparam; 332 int error; 333 int state; 334 335 /* all sysctl names at this level are terminal */ 336 if (namelen != 1) 337 return (ENOTDIR); /* overloaded */ 338 339 switch (name[0]) { 340 case GPROF_STATE: 341 state = gp->state; 342 error = sysctl_handle_int(oidp, &state, 0, req); 343 if (error) 344 return (error); 345 if (!req->newptr) 346 return (0); 347 if (state == GMON_PROF_OFF) { 348 gp->state = state; 349 PROC_LOCK(&proc0); 350 stopprofclock(&proc0); 351 PROC_UNLOCK(&proc0); 352 stopguprof(gp); 353 } else if (state == GMON_PROF_ON) { 354 gp->state = GMON_PROF_OFF; 355 stopguprof(gp); 356 gp->profrate = profhz; 357 PROC_LOCK(&proc0); 358 startprofclock(&proc0); 359 PROC_UNLOCK(&proc0); 360 gp->state = state; 361 #ifdef GUPROF 362 } else if (state == GMON_PROF_HIRES) { 363 gp->state = GMON_PROF_OFF; 364 PROC_LOCK(&proc0); 365 stopprofclock(&proc0); 366 PROC_UNLOCK(&proc0); 367 startguprof(gp); 368 gp->state = state; 369 #endif 370 } else if (state != gp->state) 371 return (EINVAL); 372 return (0); 373 case GPROF_COUNT: 374 return (sysctl_handle_opaque(oidp, 375 gp->kcount, gp->kcountsize, req)); 376 case GPROF_FROMS: 377 return (sysctl_handle_opaque(oidp, 378 gp->froms, gp->fromssize, req)); 379 case GPROF_TOS: 380 return (sysctl_handle_opaque(oidp, 381 gp->tos, gp->tossize, req)); 382 case GPROF_GMONPARAM: 383 return (sysctl_handle_opaque(oidp, gp, sizeof *gp, req)); 384 default: 385 return (EOPNOTSUPP); 386 } 387 /* NOTREACHED */ 388 } 389 390 static SYSCTL_NODE(_kern, KERN_PROF, prof, CTLFLAG_RW, sysctl_kern_prof, ""); 391 #endif /* GPROF */ 392 393 /* 394 * Profiling system call. 395 * 396 * The scale factor is a fixed point number with 16 bits of fraction, so that 397 * 1.0 is represented as 0x10000. A scale factor of 0 turns off profiling. 398 */ 399 #ifndef _SYS_SYSPROTO_H_ 400 struct profil_args { 401 caddr_t samples; 402 size_t size; 403 size_t offset; 404 u_int scale; 405 }; 406 #endif 407 /* ARGSUSED */ 408 int 409 sys_profil(struct thread *td, struct profil_args *uap) 410 { 411 struct uprof *upp; 412 struct proc *p; 413 414 if (uap->scale > (1 << 16)) 415 return (EINVAL); 416 417 p = td->td_proc; 418 if (uap->scale == 0) { 419 PROC_LOCK(p); 420 stopprofclock(p); 421 PROC_UNLOCK(p); 422 return (0); 423 } 424 PROC_LOCK(p); 425 upp = &td->td_proc->p_stats->p_prof; 426 PROC_PROFLOCK(p); 427 upp->pr_off = uap->offset; 428 upp->pr_scale = uap->scale; 429 upp->pr_base = uap->samples; 430 upp->pr_size = uap->size; 431 PROC_PROFUNLOCK(p); 432 startprofclock(p); 433 PROC_UNLOCK(p); 434 435 return (0); 436 } 437 438 /* 439 * Scale is a fixed-point number with the binary point 16 bits 440 * into the value, and is <= 1.0. pc is at most 32 bits, so the 441 * intermediate result is at most 48 bits. 442 */ 443 #define PC_TO_INDEX(pc, prof) \ 444 ((int)(((u_quad_t)((pc) - (prof)->pr_off) * \ 445 (u_quad_t)((prof)->pr_scale)) >> 16) & ~1) 446 447 /* 448 * Collect user-level profiling statistics; called on a profiling tick, 449 * when a process is running in user-mode. This routine may be called 450 * from an interrupt context. We try to update the user profiling buffers 451 * cheaply with fuswintr() and suswintr(). If that fails, we revert to 452 * an AST that will vector us to trap() with a context in which copyin 453 * and copyout will work. Trap will then call addupc_task(). 454 * 455 * Note that we may (rarely) not get around to the AST soon enough, and 456 * lose profile ticks when the next tick overwrites this one, but in this 457 * case the system is overloaded and the profile is probably already 458 * inaccurate. 459 */ 460 void 461 addupc_intr(struct thread *td, uintfptr_t pc, u_int ticks) 462 { 463 struct uprof *prof; 464 caddr_t addr; 465 u_int i; 466 int v; 467 468 if (ticks == 0) 469 return; 470 prof = &td->td_proc->p_stats->p_prof; 471 PROC_PROFLOCK(td->td_proc); 472 if (pc < prof->pr_off || 473 (i = PC_TO_INDEX(pc, prof)) >= prof->pr_size) { 474 PROC_PROFUNLOCK(td->td_proc); 475 return; /* out of range; ignore */ 476 } 477 478 addr = prof->pr_base + i; 479 PROC_PROFUNLOCK(td->td_proc); 480 if ((v = fuswintr(addr)) == -1 || suswintr(addr, v + ticks) == -1) { 481 td->td_profil_addr = pc; 482 td->td_profil_ticks = ticks; 483 td->td_pflags |= TDP_OWEUPC; 484 thread_lock(td); 485 td->td_flags |= TDF_ASTPENDING; 486 thread_unlock(td); 487 } 488 } 489 490 /* 491 * Much like before, but we can afford to take faults here. If the 492 * update fails, we simply turn off profiling. 493 */ 494 void 495 addupc_task(struct thread *td, uintfptr_t pc, u_int ticks) 496 { 497 struct proc *p = td->td_proc; 498 struct uprof *prof; 499 caddr_t addr; 500 u_int i; 501 u_short v; 502 int stop = 0; 503 504 if (ticks == 0) 505 return; 506 507 PROC_LOCK(p); 508 if (!(p->p_flag & P_PROFIL)) { 509 PROC_UNLOCK(p); 510 return; 511 } 512 p->p_profthreads++; 513 prof = &p->p_stats->p_prof; 514 PROC_PROFLOCK(p); 515 if (pc < prof->pr_off || 516 (i = PC_TO_INDEX(pc, prof)) >= prof->pr_size) { 517 PROC_PROFUNLOCK(p); 518 goto out; 519 } 520 521 addr = prof->pr_base + i; 522 PROC_PROFUNLOCK(p); 523 PROC_UNLOCK(p); 524 if (copyin(addr, &v, sizeof(v)) == 0) { 525 v += ticks; 526 if (copyout(&v, addr, sizeof(v)) == 0) { 527 PROC_LOCK(p); 528 goto out; 529 } 530 } 531 stop = 1; 532 PROC_LOCK(p); 533 534 out: 535 if (--p->p_profthreads == 0) { 536 if (p->p_flag & P_STOPPROF) { 537 wakeup(&p->p_profthreads); 538 p->p_flag &= ~P_STOPPROF; 539 stop = 0; 540 } 541 } 542 if (stop) 543 stopprofclock(p); 544 PROC_UNLOCK(p); 545 } 546