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 * $Id: subr_prof.c,v 1.17 1996/10/17 19:32:18 bde Exp $ 35 */ 36 37 #include <sys/param.h> 38 #include <sys/systm.h> 39 #include <sys/sysproto.h> 40 #include <sys/kernel.h> 41 #include <sys/proc.h> 42 #include <sys/resourcevar.h> 43 #include <sys/sysctl.h> 44 45 #include <machine/cpu.h> 46 47 #ifdef GPROF 48 #include <sys/malloc.h> 49 #include <sys/gmon.h> 50 51 static void kmstartup __P((void *)); 52 SYSINIT(kmem, SI_SUB_KPROF, SI_ORDER_FIRST, kmstartup, NULL) 53 54 struct gmonparam _gmonparam = { GMON_PROF_OFF }; 55 56 extern char btext[]; 57 extern char etext[]; 58 59 #ifdef GUPROF 60 void 61 nullfunc_loop_profiled() 62 { 63 int i; 64 65 for (i = 0; i < CALIB_SCALE; i++) 66 nullfunc_profiled(); 67 } 68 69 #define nullfunc_loop_profiled_end nullfunc_profiled /* XXX */ 70 71 void 72 nullfunc_profiled() 73 { 74 } 75 #endif /* GUPROF */ 76 77 static void 78 kmstartup(dummy) 79 void *dummy; 80 { 81 char *cp; 82 struct gmonparam *p = &_gmonparam; 83 #ifdef GUPROF 84 int cputime_overhead; 85 int empty_loop_time; 86 int i; 87 int mcount_overhead; 88 int mexitcount_overhead; 89 int nullfunc_loop_overhead; 90 int nullfunc_loop_profiled_time; 91 fptrint_t tmp_addr; 92 #endif 93 94 /* 95 * Round lowpc and highpc to multiples of the density we're using 96 * so the rest of the scaling (here and in gprof) stays in ints. 97 */ 98 p->lowpc = ROUNDDOWN((u_long)btext, HISTFRACTION * sizeof(HISTCOUNTER)); 99 p->highpc = ROUNDUP((u_long)etext, HISTFRACTION * sizeof(HISTCOUNTER)); 100 p->textsize = p->highpc - p->lowpc; 101 printf("Profiling kernel, textsize=%lu [%x..%x]\n", 102 p->textsize, p->lowpc, p->highpc); 103 p->kcountsize = p->textsize / HISTFRACTION; 104 p->hashfraction = HASHFRACTION; 105 p->fromssize = p->textsize / HASHFRACTION; 106 p->tolimit = p->textsize * ARCDENSITY / 100; 107 if (p->tolimit < MINARCS) 108 p->tolimit = MINARCS; 109 else if (p->tolimit > MAXARCS) 110 p->tolimit = MAXARCS; 111 p->tossize = p->tolimit * sizeof(struct tostruct); 112 cp = (char *)malloc(p->kcountsize + p->fromssize + p->tossize, 113 M_GPROF, M_NOWAIT); 114 if (cp == 0) { 115 printf("No memory for profiling.\n"); 116 return; 117 } 118 bzero(cp, p->kcountsize + p->tossize + p->fromssize); 119 p->tos = (struct tostruct *)cp; 120 cp += p->tossize; 121 p->kcount = (HISTCOUNTER *)cp; 122 cp += p->kcountsize; 123 p->froms = (u_short *)cp; 124 125 #ifdef GUPROF 126 /* Initialize pointers to overhead counters. */ 127 p->cputime_count = &KCOUNT(p, PC_TO_I(p, cputime)); 128 p->mcount_count = &KCOUNT(p, PC_TO_I(p, mcount)); 129 p->mexitcount_count = &KCOUNT(p, PC_TO_I(p, mexitcount)); 130 131 /* 132 * Disable interrupts to avoid interference while we calibrate 133 * things. 134 */ 135 disable_intr(); 136 137 /* 138 * Determine overheads. 139 * XXX this needs to be repeated for each useful timer/counter. 140 */ 141 cputime_overhead = 0; 142 startguprof(p); 143 for (i = 0; i < CALIB_SCALE; i++) 144 cputime_overhead += cputime(); 145 146 empty_loop(); 147 startguprof(p); 148 empty_loop(); 149 empty_loop_time = cputime(); 150 151 nullfunc_loop_profiled(); 152 153 /* 154 * Start profiling. There won't be any normal function calls since 155 * interrupts are disabled, but we will call the profiling routines 156 * directly to determine their overheads. 157 */ 158 p->state = GMON_PROF_HIRES; 159 160 startguprof(p); 161 nullfunc_loop_profiled(); 162 163 startguprof(p); 164 for (i = 0; i < CALIB_SCALE; i++) 165 #if defined(i386) && __GNUC__ >= 2 166 asm("pushl %0; call __mcount; popl %%ecx" 167 : 168 : "i" (profil) 169 : "ax", "bx", "cx", "dx", "memory"); 170 #else 171 #error 172 #endif 173 mcount_overhead = KCOUNT(p, PC_TO_I(p, profil)); 174 175 startguprof(p); 176 for (i = 0; i < CALIB_SCALE; i++) 177 #if defined(i386) && __GNUC__ >= 2 178 asm("call mexitcount; 1:" 179 : : : "ax", "bx", "cx", "dx", "memory"); 180 asm("movl $1b,%0" : "=rm" (tmp_addr)); 181 #else 182 #error 183 #endif 184 mexitcount_overhead = KCOUNT(p, PC_TO_I(p, tmp_addr)); 185 186 p->state = GMON_PROF_OFF; 187 stopguprof(p); 188 189 enable_intr(); 190 191 nullfunc_loop_profiled_time = 0; 192 for (tmp_addr = (fptrint_t)nullfunc_loop_profiled; 193 tmp_addr < (fptrint_t)nullfunc_loop_profiled_end; 194 tmp_addr += HISTFRACTION * sizeof(HISTCOUNTER)) 195 nullfunc_loop_profiled_time += KCOUNT(p, PC_TO_I(p, tmp_addr)); 196 #define CALIB_DOSCALE(count) (((count) + CALIB_SCALE / 3) / CALIB_SCALE) 197 #define c2n(count, freq) ((int)((count) * 1000000000LL / freq)) 198 printf("cputime %d, empty_loop %d, nullfunc_loop_profiled %d, mcount %d, mexitcount %d\n", 199 CALIB_DOSCALE(c2n(cputime_overhead, p->profrate)), 200 CALIB_DOSCALE(c2n(empty_loop_time, p->profrate)), 201 CALIB_DOSCALE(c2n(nullfunc_loop_profiled_time, p->profrate)), 202 CALIB_DOSCALE(c2n(mcount_overhead, p->profrate)), 203 CALIB_DOSCALE(c2n(mexitcount_overhead, p->profrate))); 204 cputime_overhead -= empty_loop_time; 205 mcount_overhead -= empty_loop_time; 206 mexitcount_overhead -= empty_loop_time; 207 208 /*- 209 * Profiling overheads are determined by the times between the 210 * following events: 211 * MC1: mcount() is called 212 * MC2: cputime() (called from mcount()) latches the timer 213 * MC3: mcount() completes 214 * ME1: mexitcount() is called 215 * ME2: cputime() (called from mexitcount()) latches the timer 216 * ME3: mexitcount() completes. 217 * The times between the events vary slightly depending on instruction 218 * combination and cache misses, etc. Attempt to determine the 219 * minimum times. These can be subtracted from the profiling times 220 * without much risk of reducing the profiling times below what they 221 * would be when profiling is not configured. Abbreviate: 222 * ab = minimum time between MC1 and MC3 223 * a = minumum time between MC1 and MC2 224 * b = minimum time between MC2 and MC3 225 * cd = minimum time between ME1 and ME3 226 * c = minimum time between ME1 and ME2 227 * d = minimum time between ME2 and ME3. 228 * These satisfy the relations: 229 * ab <= mcount_overhead (just measured) 230 * a + b <= ab 231 * cd <= mexitcount_overhead (just measured) 232 * c + d <= cd 233 * a + d <= nullfunc_loop_profiled_time (just measured) 234 * a >= 0, b >= 0, c >= 0, d >= 0. 235 * Assume that ab and cd are equal to the minimums. 236 */ 237 p->cputime_overhead = CALIB_DOSCALE(cputime_overhead); 238 p->mcount_overhead = CALIB_DOSCALE(mcount_overhead - cputime_overhead); 239 p->mexitcount_overhead = CALIB_DOSCALE(mexitcount_overhead 240 - cputime_overhead); 241 nullfunc_loop_overhead = nullfunc_loop_profiled_time - empty_loop_time; 242 p->mexitcount_post_overhead = CALIB_DOSCALE((mcount_overhead 243 - nullfunc_loop_overhead) 244 / 4); 245 p->mexitcount_pre_overhead = p->mexitcount_overhead 246 + p->cputime_overhead 247 - p->mexitcount_post_overhead; 248 p->mcount_pre_overhead = CALIB_DOSCALE(nullfunc_loop_overhead) 249 - p->mexitcount_post_overhead; 250 p->mcount_post_overhead = p->mcount_overhead 251 + p->cputime_overhead 252 - p->mcount_pre_overhead; 253 printf( 254 "Profiling overheads: mcount: %d+%d, %d+%d; mexitcount: %d+%d, %d+%d nsec\n", 255 c2n(p->cputime_overhead, p->profrate), 256 c2n(p->mcount_overhead, p->profrate), 257 c2n(p->mcount_pre_overhead, p->profrate), 258 c2n(p->mcount_post_overhead, p->profrate), 259 c2n(p->cputime_overhead, p->profrate), 260 c2n(p->mexitcount_overhead, p->profrate), 261 c2n(p->mexitcount_pre_overhead, p->profrate), 262 c2n(p->mexitcount_post_overhead, p->profrate)); 263 printf( 264 "Profiling overheads: mcount: %d+%d, %d+%d; mexitcount: %d+%d, %d+%d cycles\n", 265 p->cputime_overhead, p->mcount_overhead, 266 p->mcount_pre_overhead, p->mcount_post_overhead, 267 p->cputime_overhead, p->mexitcount_overhead, 268 p->mexitcount_pre_overhead, p->mexitcount_post_overhead); 269 #endif /* GUPROF */ 270 } 271 272 /* 273 * Return kernel profiling information. 274 */ 275 static int 276 sysctl_kern_prof SYSCTL_HANDLER_ARGS 277 { 278 int *name = (int *) arg1; 279 u_int namelen = arg2; 280 struct gmonparam *gp = &_gmonparam; 281 int error; 282 int state; 283 284 /* all sysctl names at this level are terminal */ 285 if (namelen != 1) 286 return (ENOTDIR); /* overloaded */ 287 288 switch (name[0]) { 289 case GPROF_STATE: 290 state = gp->state; 291 error = sysctl_handle_int(oidp, &state, 0, req); 292 if (error) 293 return (error); 294 if (!req->newptr) 295 return (0); 296 if (state == GMON_PROF_OFF) { 297 gp->state = state; 298 stopprofclock(&proc0); 299 stopguprof(gp); 300 } else if (state == GMON_PROF_ON) { 301 gp->state = GMON_PROF_OFF; 302 stopguprof(gp); 303 gp->profrate = profhz; 304 startprofclock(&proc0); 305 gp->state = state; 306 #ifdef GUPROF 307 } else if (state == GMON_PROF_HIRES) { 308 gp->state = GMON_PROF_OFF; 309 stopprofclock(&proc0); 310 startguprof(gp); 311 gp->state = state; 312 #endif 313 } else if (state != gp->state) 314 return (EINVAL); 315 return (0); 316 case GPROF_COUNT: 317 return (sysctl_handle_opaque(oidp, 318 gp->kcount, gp->kcountsize, req)); 319 case GPROF_FROMS: 320 return (sysctl_handle_opaque(oidp, 321 gp->froms, gp->fromssize, req)); 322 case GPROF_TOS: 323 return (sysctl_handle_opaque(oidp, 324 gp->tos, gp->tossize, req)); 325 case GPROF_GMONPARAM: 326 return (sysctl_handle_opaque(oidp, gp, sizeof *gp, req)); 327 default: 328 return (EOPNOTSUPP); 329 } 330 /* NOTREACHED */ 331 } 332 333 SYSCTL_NODE(_kern, KERN_PROF, prof, CTLFLAG_RW, sysctl_kern_prof, ""); 334 #endif /* GPROF */ 335 336 /* 337 * Profiling system call. 338 * 339 * The scale factor is a fixed point number with 16 bits of fraction, so that 340 * 1.0 is represented as 0x10000. A scale factor of 0 turns off profiling. 341 */ 342 #ifndef _SYS_SYSPROTO_H_ 343 struct profil_args { 344 caddr_t samples; 345 u_int size; 346 u_int offset; 347 u_int scale; 348 }; 349 #endif 350 /* ARGSUSED */ 351 int 352 profil(p, uap, retval) 353 struct proc *p; 354 register struct profil_args *uap; 355 int *retval; 356 { 357 register struct uprof *upp; 358 int s; 359 360 if (uap->scale > (1 << 16)) 361 return (EINVAL); 362 if (uap->scale == 0) { 363 stopprofclock(p); 364 return (0); 365 } 366 upp = &p->p_stats->p_prof; 367 368 /* Block profile interrupts while changing state. */ 369 s = splstatclock(); 370 upp->pr_off = uap->offset; 371 upp->pr_scale = uap->scale; 372 upp->pr_base = uap->samples; 373 upp->pr_size = uap->size; 374 startprofclock(p); 375 splx(s); 376 377 return (0); 378 } 379 380 /* 381 * Scale is a fixed-point number with the binary point 16 bits 382 * into the value, and is <= 1.0. pc is at most 32 bits, so the 383 * intermediate result is at most 48 bits. 384 */ 385 #define PC_TO_INDEX(pc, prof) \ 386 ((int)(((u_quad_t)((pc) - (prof)->pr_off) * \ 387 (u_quad_t)((prof)->pr_scale)) >> 16) & ~1) 388 389 /* 390 * Collect user-level profiling statistics; called on a profiling tick, 391 * when a process is running in user-mode. This routine may be called 392 * from an interrupt context. We try to update the user profiling buffers 393 * cheaply with fuswintr() and suswintr(). If that fails, we revert to 394 * an AST that will vector us to trap() with a context in which copyin 395 * and copyout will work. Trap will then call addupc_task(). 396 * 397 * Note that we may (rarely) not get around to the AST soon enough, and 398 * lose profile ticks when the next tick overwrites this one, but in this 399 * case the system is overloaded and the profile is probably already 400 * inaccurate. 401 */ 402 void 403 addupc_intr(p, pc, ticks) 404 register struct proc *p; 405 register u_long pc; 406 u_int ticks; 407 { 408 register struct uprof *prof; 409 register caddr_t addr; 410 register u_int i; 411 register int v; 412 413 if (ticks == 0) 414 return; 415 prof = &p->p_stats->p_prof; 416 if (pc < prof->pr_off || 417 (i = PC_TO_INDEX(pc, prof)) >= prof->pr_size) 418 return; /* out of range; ignore */ 419 420 addr = prof->pr_base + i; 421 if ((v = fuswintr(addr)) == -1 || suswintr(addr, v + ticks) == -1) { 422 prof->pr_addr = pc; 423 prof->pr_ticks = ticks; 424 need_proftick(p); 425 } 426 } 427 428 /* 429 * Much like before, but we can afford to take faults here. If the 430 * update fails, we simply turn off profiling. 431 */ 432 void 433 addupc_task(p, pc, ticks) 434 register struct proc *p; 435 register u_long pc; 436 u_int ticks; 437 { 438 register struct uprof *prof; 439 register caddr_t addr; 440 register u_int i; 441 u_short v; 442 443 /* Testing P_PROFIL may be unnecessary, but is certainly safe. */ 444 if ((p->p_flag & P_PROFIL) == 0 || ticks == 0) 445 return; 446 447 prof = &p->p_stats->p_prof; 448 if (pc < prof->pr_off || 449 (i = PC_TO_INDEX(pc, prof)) >= prof->pr_size) 450 return; 451 452 addr = prof->pr_base + i; 453 if (copyin(addr, (caddr_t)&v, sizeof(v)) == 0) { 454 v += ticks; 455 if (copyout((caddr_t)&v, addr, sizeof(v)) == 0) 456 return; 457 } 458 stopprofclock(p); 459 } 460