sys/kern/subr_prof.c

   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  * $FreeBSD: src/sys/kern/subr_prof.c,v 1.32.2.2 2000/08/03 00:09:32 ps Exp $
  35  * $DragonFly: src/sys/kern/subr_prof.c,v 1.11 2005/10/08 12:24:26 corecode Exp $
  36  */
  37
  38 #include <sys/param.h>
  39 #include <sys/systm.h>
  40 #include <sys/sysproto.h>
  41 #include <sys/kernel.h>
  42 #include <sys/proc.h>
  43 #include <sys/resourcevar.h>
  44 #include <sys/sysctl.h>
  45 #include <sys/thread2.h>
  46
  47 #include <machine/ipl.h>
  48 #include <machine/cpu.h>
  49
  50 #ifdef GPROF
  51 #include <sys/malloc.h>
  52 #include <sys/gmon.h>
  53 #undef MCOUNT
  54
  55 static MALLOC_DEFINE(M_GPROF, "gprof", "kernel profiling buffer");
  56
  57 static void kmstartup (void *);
  58 SYSINIT(kmem, SI_SUB_KPROF, SI_ORDER_FIRST, kmstartup, NULL)
  59
  60 struct gmonparam _gmonparam = { GMON_PROF_OFF };
  61
  62 #ifdef GUPROF
  63 #include <machine/asmacros.h>
  64
  65 void
  66 nullfunc_loop_profiled()
  67 {
  68         int i;
  69
  70         for (i = 0; i < CALIB_SCALE; i++)
  71                 nullfunc_profiled();
  72 }
  73
  74 #define nullfunc_loop_profiled_end      nullfunc_profiled       /* XXX */
  75
  76 void
  77 nullfunc_profiled()
  78 {
  79 }
  80 #endif /* GUPROF */
  81
  82 static void
  83 kmstartup(void *dummy)
  84 {
  85         char *cp;
  86         struct gmonparam *p = &_gmonparam;
  87 #ifdef GUPROF
  88         int cputime_overhead;
  89         int empty_loop_time;
  90         int i;
  91         int mcount_overhead;
  92         int mexitcount_overhead;
  93         int nullfunc_loop_overhead;
  94         int nullfunc_loop_profiled_time;
  95         uintfptr_t tmp_addr;
  96 #endif
  97
  98         /*
  99          * Round lowpc and highpc to multiples of the density we're using
 100          * so the rest of the scaling (here and in gprof) stays in ints.
 101          */
 102         p->lowpc = ROUNDDOWN((u_long)btext, HISTFRACTION * sizeof(HISTCOUNTER));
 103         p->highpc = ROUNDUP((u_long)etext, HISTFRACTION * sizeof(HISTCOUNTER));
 104         p->textsize = p->highpc - p->lowpc;
 105         printf("Profiling kernel, textsize=%lu [%x..%x]\n",
 106                p->textsize, p->lowpc, p->highpc);
 107         p->kcountsize = p->textsize / HISTFRACTION;
 108         p->hashfraction = HASHFRACTION;
 109         p->fromssize = p->textsize / HASHFRACTION;
 110         p->tolimit = p->textsize * ARCDENSITY / 100;
 111         if (p->tolimit < MINARCS)
 112                 p->tolimit = MINARCS;
 113         else if (p->tolimit > MAXARCS)
 114                 p->tolimit = MAXARCS;
 115         p->tossize = p->tolimit * sizeof(struct tostruct);
 116         cp = (char *)malloc(p->kcountsize + p->fromssize + p->tossize,
 117             M_GPROF, M_NOWAIT);
 118         if (cp == 0) {
 119                 printf("No memory for profiling.\n");
 120                 return;
 121         }
 122         bzero(cp, p->kcountsize + p->tossize + p->fromssize);
 123         p->tos = (struct tostruct *)cp;
 124         cp += p->tossize;
 125         p->kcount = (HISTCOUNTER *)cp;
 126         cp += p->kcountsize;
 127         p->froms = (u_short *)cp;
 128
 129 #ifdef GUPROF
 130         /* Initialize pointers to overhead counters. */
 131         p->cputime_count = &KCOUNT(p, PC_TO_I(p, cputime));
 132         p->mcount_count = &KCOUNT(p, PC_TO_I(p, mcount));
 133         p->mexitcount_count = &KCOUNT(p, PC_TO_I(p, mexitcount));
 134
 135         /*
 136          * Disable interrupts to avoid interference while we calibrate
 137          * things.
 138          */
 139         cpu_disable_intr();
 140
 141         /*
 142          * Determine overheads.
 143          * XXX this needs to be repeated for each useful timer/counter.
 144          */
 145         cputime_overhead = 0;
 146         startguprof(p);
 147         for (i = 0; i < CALIB_SCALE; i++)
 148                 cputime_overhead += cputime();
 149
 150         empty_loop();
 151         startguprof(p);
 152         empty_loop();
 153         empty_loop_time = cputime();
 154
 155         nullfunc_loop_profiled();
 156
 157         /*
 158          * Start profiling.  There won't be any normal function calls since
 159          * interrupts are disabled, but we will call the profiling routines
 160          * directly to determine their overheads.
 161          */
 162         p->state = GMON_PROF_HIRES;
 163
 164         startguprof(p);
 165         nullfunc_loop_profiled();
 166
 167         startguprof(p);
 168         for (i = 0; i < CALIB_SCALE; i++)
 169 #if defined(__i386__) && __GNUC__ >= 2
 170                 __asm("pushl %0; call __mcount; popl %%ecx"
 171                       :
 172                       : "i" (profil)
 173                       : "ax", "bx", "cx", "dx", "memory");
 174 #else
 175 #error
 176 #endif
 177         mcount_overhead = KCOUNT(p, PC_TO_I(p, profil));
 178
 179         startguprof(p);
 180         for (i = 0; i < CALIB_SCALE; i++)
 181 #if defined(__i386__) && __GNUC__ >= 2
 182                     __asm("call " __XSTRING(HIDENAME(mexitcount)) "; 1:"
 183                           : : : "ax", "bx", "cx", "dx", "memory");
 184         __asm("movl $1b,%0" : "=rm" (tmp_addr));
 185 #else
 186 #error
 187 #endif
 188         mexitcount_overhead = KCOUNT(p, PC_TO_I(p, tmp_addr));
 189
 190         p->state = GMON_PROF_OFF;
 191         stopguprof(p);
 192
 193         cpu_enable_intr();
 194
 195         nullfunc_loop_profiled_time = 0;
 196         for (tmp_addr = (uintfptr_t)nullfunc_loop_profiled;
 197              tmp_addr < (uintfptr_t)nullfunc_loop_profiled_end;
 198              tmp_addr += HISTFRACTION * sizeof(HISTCOUNTER))
 199                 nullfunc_loop_profiled_time += KCOUNT(p, PC_TO_I(p, tmp_addr));
 200 #define CALIB_DOSCALE(count)    (((count) + CALIB_SCALE / 3) / CALIB_SCALE)
 201 #define c2n(count, freq)        ((int)((count) * 1000000000LL / freq))
 202         printf("cputime %d, empty_loop %d, nullfunc_loop_profiled %d, mcount %d, mexitcount %d\n",
 203                CALIB_DOSCALE(c2n(cputime_overhead, p->profrate)),
 204                CALIB_DOSCALE(c2n(empty_loop_time, p->profrate)),
 205                CALIB_DOSCALE(c2n(nullfunc_loop_profiled_time, p->profrate)),
 206                CALIB_DOSCALE(c2n(mcount_overhead, p->profrate)),
 207                CALIB_DOSCALE(c2n(mexitcount_overhead, p->profrate)));
 208         cputime_overhead -= empty_loop_time;
 209         mcount_overhead -= empty_loop_time;
 210         mexitcount_overhead -= empty_loop_time;
 211
 212         /*-
 213          * Profiling overheads are determined by the times between the
 214          * following events:
 215          *      MC1: mcount() is called
 216          *      MC2: cputime() (called from mcount()) latches the timer
 217          *      MC3: mcount() completes
 218          *      ME1: mexitcount() is called
 219          *      ME2: cputime() (called from mexitcount()) latches the timer
 220          *      ME3: mexitcount() completes.
 221          * The times between the events vary slightly depending on instruction
 222          * combination and cache misses, etc.  Attempt to determine the
 223          * minimum times.  These can be subtracted from the profiling times
 224          * without much risk of reducing the profiling times below what they
 225          * would be when profiling is not configured.  Abbreviate:
 226          *      ab = minimum time between MC1 and MC3
 227          *      a  = minumum time between MC1 and MC2
 228          *      b  = minimum time between MC2 and MC3
 229          *      cd = minimum time between ME1 and ME3
 230          *      c  = minimum time between ME1 and ME2
 231          *      d  = minimum time between ME2 and ME3.
 232          * These satisfy the relations:
 233          *      ab            <= mcount_overhead                (just measured)
 234          *      a + b         <= ab
 235          *              cd    <= mexitcount_overhead            (just measured)
 236          *              c + d <= cd
 237          *      a         + d <= nullfunc_loop_profiled_time    (just measured)
 238          *      a >= 0, b >= 0, c >= 0, d >= 0.
 239          * Assume that ab and cd are equal to the minimums.
 240          */
 241         p->cputime_overhead = CALIB_DOSCALE(cputime_overhead);
 242         p->mcount_overhead = CALIB_DOSCALE(mcount_overhead - cputime_overhead);
 243         p->mexitcount_overhead = CALIB_DOSCALE(mexitcount_overhead
 244                                                - cputime_overhead);
 245         nullfunc_loop_overhead = nullfunc_loop_profiled_time - empty_loop_time;
 246         p->mexitcount_post_overhead = CALIB_DOSCALE((mcount_overhead
 247                                                      - nullfunc_loop_overhead)
 248                                                     / 4);
 249         p->mexitcount_pre_overhead = p->mexitcount_overhead
 250                                      + p->cputime_overhead
 251                                      - p->mexitcount_post_overhead;
 252         p->mcount_pre_overhead = CALIB_DOSCALE(nullfunc_loop_overhead)
 253                                  - p->mexitcount_post_overhead;
 254         p->mcount_post_overhead = p->mcount_overhead
 255                                   + p->cputime_overhead
 256                                   - p->mcount_pre_overhead;
 257         printf(
 258 "Profiling overheads: mcount: %d+%d, %d+%d; mexitcount: %d+%d, %d+%d nsec\n",
 259                c2n(p->cputime_overhead, p->profrate),
 260                c2n(p->mcount_overhead, p->profrate),
 261                c2n(p->mcount_pre_overhead, p->profrate),
 262                c2n(p->mcount_post_overhead, p->profrate),
 263                c2n(p->cputime_overhead, p->profrate),
 264                c2n(p->mexitcount_overhead, p->profrate),
 265                c2n(p->mexitcount_pre_overhead, p->profrate),
 266                c2n(p->mexitcount_post_overhead, p->profrate));
 267         printf(
 268 "Profiling overheads: mcount: %d+%d, %d+%d; mexitcount: %d+%d, %d+%d cycles\n",
 269                p->cputime_overhead, p->mcount_overhead,
 270                p->mcount_pre_overhead, p->mcount_post_overhead,
 271                p->cputime_overhead, p->mexitcount_overhead,
 272                p->mexitcount_pre_overhead, p->mexitcount_post_overhead);
 273 #endif /* GUPROF */
 274 }
 275
 276 /*
 277  * Return kernel profiling information.
 278  */
 279 static int
 280 sysctl_kern_prof(SYSCTL_HANDLER_ARGS)
 281 {
 282         int *name = (int *) arg1;
 283         u_int namelen = arg2;
 284         struct gmonparam *gp = &_gmonparam;
 285         int error;
 286         int state;
 287
 288         /* all sysctl names at this level are terminal */
 289         if (namelen != 1)
 290                 return (ENOTDIR);               /* overloaded */
 291
 292         switch (name[0]) {
 293         case GPROF_STATE:
 294                 state = gp->state;
 295                 error = sysctl_handle_int(oidp, &state, 0, req);
 296                 if (error)
 297                         return (error);
 298                 if (!req->newptr)
 299                         return (0);
 300                 if (state == GMON_PROF_OFF) {
 301                         gp->state = state;
 302                         stopprofclock(&proc0);
 303                         stopguprof(gp);
 304                 } else if (state == GMON_PROF_ON) {
 305                         gp->state = GMON_PROF_OFF;
 306                         stopguprof(gp);
 307                         gp->profrate = profhz;
 308                         startprofclock(&proc0);
 309                         gp->state = state;
 310 #ifdef GUPROF
 311                 } else if (state == GMON_PROF_HIRES) {
 312                         gp->state = GMON_PROF_OFF;
 313                         stopprofclock(&proc0);
 314                         startguprof(gp);
 315                         gp->state = state;
 316 #endif
 317                 } else if (state != gp->state)
 318                         return (EINVAL);
 319                 return (0);
 320         case GPROF_COUNT:
 321                 return (sysctl_handle_opaque(oidp,
 322                         gp->kcount, gp->kcountsize, req));
 323         case GPROF_FROMS:
 324                 return (sysctl_handle_opaque(oidp,
 325                         gp->froms, gp->fromssize, req));
 326         case GPROF_TOS:
 327                 return (sysctl_handle_opaque(oidp,
 328                         gp->tos, gp->tossize, req));
 329         case GPROF_GMONPARAM:
 330                 return (sysctl_handle_opaque(oidp, gp, sizeof *gp, req));
 331         default:
 332                 return (EOPNOTSUPP);
 333         }
 334         /* NOTREACHED */
 335 }
 336
 337 SYSCTL_NODE(_kern, KERN_PROF, prof, CTLFLAG_RW, sysctl_kern_prof, "");
 338 #endif /* GPROF */
 339
 340 /*
 341  * Profiling system call.
 342  *
 343  * The scale factor is a fixed point number with 16 bits of fraction, so that
 344  * 1.0 is represented as 0x10000.  A scale factor of 0 turns off profiling.
 345  */
 346 /* ARGSUSED */
 347 int
 348 profil(struct profil_args *uap)
 349 {
 350         struct proc *p = curproc;
 351         struct uprof *upp;
 352
 353         if (uap->scale > (1 << 16))
 354                 return (EINVAL);
 355         if (uap->scale == 0) {
 356                 stopprofclock(p);
 357                 return (0);
 358         }
 359         upp = &p->p_prof;
 360
 361         /* Block profile interrupts while changing state. */
 362         crit_enter();
 363         upp->pr_off = uap->offset;
 364         upp->pr_scale = uap->scale;
 365         upp->pr_base = uap->samples;
 366         upp->pr_size = uap->size;
 367         startprofclock(p);
 368         crit_exit();
 369
 370         return (0);
 371 }
 372
 373 /*
 374  * Scale is a fixed-point number with the binary point 16 bits
 375  * into the value, and is <= 1.0.  pc is at most 32 bits, so the
 376  * intermediate result is at most 48 bits.
 377  */
 378 #define PC_TO_INDEX(pc, prof) \
 379         ((int)(((u_quad_t)((pc) - (prof)->pr_off) * \
 380             (u_quad_t)((prof)->pr_scale)) >> 16) & ~1)
 381
 382 /*
 383  * Collect user-level profiling statistics; called on a profiling tick,
 384  * when a process is running in user-mode.  This routine may be called
 385  * from an interrupt context.  We try to update the user profiling buffers
 386  * cheaply with fuswintr() and suswintr().  If that fails, we revert to
 387  * an AST that will vector us to trap() with a context in which copyin
 388  * and copyout will work.  Trap will then call addupc_task().
 389  *
 390  * Note that we may (rarely) not get around to the AST soon enough, and
 391  * lose profile ticks when the next tick overwrites this one, but in this
 392  * case the system is overloaded and the profile is probably already
 393  * inaccurate.
 394  */
 395 void
 396 addupc_intr(struct proc *p, u_long pc, u_int ticks)
 397 {
 398         struct uprof *prof;
 399         caddr_t addr;
 400         u_int i;
 401         int v;
 402
 403         if (ticks == 0)
 404                 return;
 405         prof = &p->p_prof;
 406         if (pc < prof->pr_off ||
 407             (i = PC_TO_INDEX(pc, prof)) >= prof->pr_size)
 408                 return;                 /* out of range; ignore */
 409
 410         addr = prof->pr_base + i;
 411         if ((v = fuswintr(addr)) == -1 || suswintr(addr, v + ticks) == -1) {
 412                 prof->pr_addr = pc;
 413                 prof->pr_ticks = ticks;
 414                 need_proftick();
 415         }
 416 }
 417
 418 /*
 419  * Much like before, but we can afford to take faults here.  If the
 420  * update fails, we simply turn off profiling.
 421  */
 422 void
 423 addupc_task(struct proc *p, u_long pc, u_int ticks)
 424 {
 425         struct uprof *prof;
 426         caddr_t addr;
 427         u_int i;
 428         u_short v;
 429
 430         /* Testing P_PROFIL may be unnecessary, but is certainly safe. */
 431         if ((p->p_flag & P_PROFIL) == 0 || ticks == 0)
 432                 return;
 433
 434         prof = &p->p_prof;
 435         if (pc < prof->pr_off ||
 436             (i = PC_TO_INDEX(pc, prof)) >= prof->pr_size)
 437                 return;
 438
 439         addr = prof->pr_base + i;
 440         if (copyin(addr, (caddr_t)&v, sizeof(v)) == 0) {
 441                 v += ticks;
 442                 if (copyout((caddr_t)&v, addr, sizeof(v)) == 0)
 443                         return;
 444         }
 445         stopprofclock(p);
 446 }