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. Neither the name of the University nor the names of its contributors
  14  *    may be used to endorse or promote products derived from this software
  15  *    without specific prior written permission.
  16  *
  17  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  27  * SUCH DAMAGE.
  28  *
  29  *      @(#)subr_prof.c 8.3 (Berkeley) 9/23/93
  30  * $FreeBSD: src/sys/kern/subr_prof.c,v 1.32.2.2 2000/08/03 00:09:32 ps Exp $
  31  */
  32
  33 #include <sys/param.h>
  34 #include <sys/systm.h>
  35 #include <sys/sysproto.h>
  36 #include <sys/kernel.h>
  37 #include <sys/proc.h>
  38 #include <sys/resourcevar.h>
  39 #include <sys/sysctl.h>
  40
  41 #include <sys/thread2.h>
  42
  43 #include <machine/cpu.h>
  44
  45 #ifdef GPROF
  46 #include <sys/malloc.h>
  47 #include <sys/gmon.h>
  48 #undef MCOUNT
  49
  50 static MALLOC_DEFINE(M_GPROF, "gprof", "kernel profiling buffer");
  51
  52 static void kmstartup (void *);
  53 SYSINIT(kmem, SI_SUB_KPROF, SI_ORDER_FIRST, kmstartup, NULL);
  54
  55 struct gmonparam _gmonparam = { GMON_PROF_OFF };
  56
  57 #ifdef GUPROF
  58 #include <machine/asmacros.h>
  59
  60 void
  61 nullfunc_loop_profiled(void)
  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(void)
  73 {
  74 }
  75 #endif /* GUPROF */
  76
  77 static void
  78 kmstartup(void *dummy)
  79 {
  80         char *cp;
  81         struct gmonparam *p = &_gmonparam;
  82 #ifdef GUPROF
  83         int cputime_overhead;
  84         int empty_loop_time;
  85         int i;
  86         int mcount_overhead;
  87         int mexitcount_overhead;
  88         int nullfunc_loop_overhead;
  89         int nullfunc_loop_profiled_time;
  90         uintfptr_t tmp_addr;
  91 #endif
  92
  93         /*
  94          * Round lowpc and highpc to multiples of the density we're using
  95          * so the rest of the scaling (here and in gprof) stays in ints.
  96          */
  97         p->lowpc = ROUNDDOWN((u_long)btext, HISTFRACTION * sizeof(HISTCOUNTER));
  98         p->highpc = ROUNDUP((u_long)etext, HISTFRACTION * sizeof(HISTCOUNTER));
  99         p->textsize = p->highpc - p->lowpc;
 100         kprintf("Profiling kernel, textsize=%lu [%jx..%jx]\n",
 101             p->textsize, (uintmax_t)p->lowpc, (uintmax_t)p->highpc);
 102         p->kcountsize = p->textsize / HISTFRACTION;
 103         p->hashfraction = HASHFRACTION;
 104         p->fromssize = p->textsize / HASHFRACTION;
 105         p->tolimit = p->textsize * ARCDENSITY / 100;
 106         if (p->tolimit < MINARCS)
 107                 p->tolimit = MINARCS;
 108         else if (p->tolimit > MAXARCS)
 109                 p->tolimit = MAXARCS;
 110         p->tossize = p->tolimit * sizeof(struct tostruct);
 111         cp = (char *)kmalloc(p->kcountsize + p->fromssize + p->tossize,
 112             M_GPROF, M_NOWAIT | M_ZERO);
 113         if (cp == NULL) {
 114                 kprintf("No memory for profiling.\n");
 115                 return;
 116         }
 117         p->tos = (struct tostruct *)cp;
 118         cp += p->tossize;
 119         p->kcount = (HISTCOUNTER *)cp;
 120         cp += p->kcountsize;
 121         p->froms = (u_short *)cp;
 122
 123 #ifdef GUPROF
 124         /* Initialize pointers to overhead counters. */
 125         p->cputime_count = &KCOUNT(p, PC_TO_I(p, cputime));
 126         p->mcount_count = &KCOUNT(p, PC_TO_I(p, mcount));
 127         p->mexitcount_count = &KCOUNT(p, PC_TO_I(p, mexitcount));
 128
 129         /*
 130          * Disable interrupts to avoid interference while we calibrate
 131          * things.
 132          */
 133         cpu_disable_intr();
 134
 135         /*
 136          * Determine overheads.
 137          * XXX this needs to be repeated for each useful timer/counter.
 138          */
 139         cputime_overhead = 0;
 140         startguprof(p);
 141         for (i = 0; i < CALIB_SCALE; i++)
 142                 cputime_overhead += cputime();
 143
 144         empty_loop();
 145         startguprof(p);
 146         empty_loop();
 147         empty_loop_time = cputime();
 148
 149         nullfunc_loop_profiled();
 150
 151         /*
 152          * Start profiling.  There won't be any normal function calls since
 153          * interrupts are disabled, but we will call the profiling routines
 154          * directly to determine their overheads.
 155          */
 156         p->state = GMON_PROF_HIRES;
 157
 158         startguprof(p);
 159         nullfunc_loop_profiled();
 160
 161         startguprof(p);
 162         for (i = 0; i < CALIB_SCALE; i++)
 163 #if defined(__i386__) && __GNUC__ >= 2
 164                 __asm("pushl %0; call __mcount; popl %%ecx"
 165                       :
 166                       : "i" (profil)
 167                       : "ax", "bx", "cx", "dx", "memory");
 168 #else
 169 #error
 170 #endif
 171         mcount_overhead = KCOUNT(p, PC_TO_I(p, profil));
 172
 173         startguprof(p);
 174         for (i = 0; i < CALIB_SCALE; i++)
 175 #if defined(__i386__) && __GNUC__ >= 2
 176                     __asm("call " __XSTRING(HIDENAME(mexitcount)) "; 1:"
 177                           : : : "ax", "bx", "cx", "dx", "memory");
 178         __asm("movl $1b,%0" : "=rm" (tmp_addr));
 179 #else
 180 #error
 181 #endif
 182         mexitcount_overhead = KCOUNT(p, PC_TO_I(p, tmp_addr));
 183
 184         p->state = GMON_PROF_OFF;
 185         stopguprof(p);
 186
 187         cpu_enable_intr();
 188
 189         nullfunc_loop_profiled_time = 0;
 190         for (tmp_addr = (uintfptr_t)nullfunc_loop_profiled;
 191              tmp_addr < (uintfptr_t)nullfunc_loop_profiled_end;
 192              tmp_addr += HISTFRACTION * sizeof(HISTCOUNTER))
 193                 nullfunc_loop_profiled_time += KCOUNT(p, PC_TO_I(p, tmp_addr));
 194 #define CALIB_DOSCALE(count)    (((count) + CALIB_SCALE / 3) / CALIB_SCALE)
 195 #define c2n(count, freq)        ((int)((count) * 1000000000LL / freq))
 196         kprintf("cputime %d, empty_loop %d, nullfunc_loop_profiled %d, mcount %d, mexitcount %d\n",
 197                CALIB_DOSCALE(c2n(cputime_overhead, p->profrate)),
 198                CALIB_DOSCALE(c2n(empty_loop_time, p->profrate)),
 199                CALIB_DOSCALE(c2n(nullfunc_loop_profiled_time, p->profrate)),
 200                CALIB_DOSCALE(c2n(mcount_overhead, p->profrate)),
 201                CALIB_DOSCALE(c2n(mexitcount_overhead, p->profrate)));
 202         cputime_overhead -= empty_loop_time;
 203         mcount_overhead -= empty_loop_time;
 204         mexitcount_overhead -= empty_loop_time;
 205
 206         /*-
 207          * Profiling overheads are determined by the times between the
 208          * following events:
 209          *      MC1: mcount() is called
 210          *      MC2: cputime() (called from mcount()) latches the timer
 211          *      MC3: mcount() completes
 212          *      ME1: mexitcount() is called
 213          *      ME2: cputime() (called from mexitcount()) latches the timer
 214          *      ME3: mexitcount() completes.
 215          * The times between the events vary slightly depending on instruction
 216          * combination and cache misses, etc.  Attempt to determine the
 217          * minimum times.  These can be subtracted from the profiling times
 218          * without much risk of reducing the profiling times below what they
 219          * would be when profiling is not configured.  Abbreviate:
 220          *      ab = minimum time between MC1 and MC3
 221          *      a  = minumum time between MC1 and MC2
 222          *      b  = minimum time between MC2 and MC3
 223          *      cd = minimum time between ME1 and ME3
 224          *      c  = minimum time between ME1 and ME2
 225          *      d  = minimum time between ME2 and ME3.
 226          * These satisfy the relations:
 227          *      ab            <= mcount_overhead                (just measured)
 228          *      a + b         <= ab
 229          *              cd    <= mexitcount_overhead            (just measured)
 230          *              c + d <= cd
 231          *      a         + d <= nullfunc_loop_profiled_time    (just measured)
 232          *      a >= 0, b >= 0, c >= 0, d >= 0.
 233          * Assume that ab and cd are equal to the minimums.
 234          */
 235         p->cputime_overhead = CALIB_DOSCALE(cputime_overhead);
 236         p->mcount_overhead = CALIB_DOSCALE(mcount_overhead - cputime_overhead);
 237         p->mexitcount_overhead = CALIB_DOSCALE(mexitcount_overhead
 238                                                - cputime_overhead);
 239         nullfunc_loop_overhead = nullfunc_loop_profiled_time - empty_loop_time;
 240         p->mexitcount_post_overhead = CALIB_DOSCALE((mcount_overhead
 241                                                      - nullfunc_loop_overhead)
 242                                                     / 4);
 243         p->mexitcount_pre_overhead = p->mexitcount_overhead
 244                                      + p->cputime_overhead
 245                                      - p->mexitcount_post_overhead;
 246         p->mcount_pre_overhead = CALIB_DOSCALE(nullfunc_loop_overhead)
 247                                  - p->mexitcount_post_overhead;
 248         p->mcount_post_overhead = p->mcount_overhead
 249                                   + p->cputime_overhead
 250                                   - p->mcount_pre_overhead;
 251         kprintf(
 252 "Profiling overheads: mcount: %d+%d, %d+%d; mexitcount: %d+%d, %d+%d nsec\n",
 253                c2n(p->cputime_overhead, p->profrate),
 254                c2n(p->mcount_overhead, p->profrate),
 255                c2n(p->mcount_pre_overhead, p->profrate),
 256                c2n(p->mcount_post_overhead, p->profrate),
 257                c2n(p->cputime_overhead, p->profrate),
 258                c2n(p->mexitcount_overhead, p->profrate),
 259                c2n(p->mexitcount_pre_overhead, p->profrate),
 260                c2n(p->mexitcount_post_overhead, p->profrate));
 261         kprintf(
 262 "Profiling overheads: mcount: %d+%d, %d+%d; mexitcount: %d+%d, %d+%d cycles\n",
 263                p->cputime_overhead, p->mcount_overhead,
 264                p->mcount_pre_overhead, p->mcount_post_overhead,
 265                p->cputime_overhead, p->mexitcount_overhead,
 266                p->mexitcount_pre_overhead, p->mexitcount_post_overhead);
 267 #endif /* GUPROF */
 268 }
 269
 270 /*
 271  * Return kernel profiling information.
 272  */
 273 static int
 274 sysctl_kern_prof(SYSCTL_HANDLER_ARGS)
 275 {
 276         int *name = (int *) arg1;
 277         u_int namelen = arg2;
 278         struct gmonparam *gp = &_gmonparam;
 279         int error;
 280         int state;
 281
 282         /* all sysctl names at this level are terminal */
 283         if (namelen != 1)
 284                 return (ENOTDIR);               /* overloaded */
 285
 286         switch (name[0]) {
 287         case GPROF_STATE:
 288                 state = gp->state;
 289                 error = sysctl_handle_int(oidp, &state, 0, req);
 290                 if (error)
 291                         return (error);
 292                 if (!req->newptr)
 293                         return (0);
 294                 lwkt_gettoken(&proc0.p_token);
 295                 if (state == GMON_PROF_OFF) {
 296                         gp->state = state;
 297                         stopprofclock(&proc0);
 298                         stopguprof(gp);
 299                 } else if (state == GMON_PROF_ON) {
 300                         gp->state = GMON_PROF_OFF;
 301                         stopguprof(gp);
 302                         gp->profrate = profhz;
 303                         startprofclock(&proc0);
 304                         gp->state = state;
 305 #ifdef GUPROF
 306                 } else if (state == GMON_PROF_HIRES) {
 307                         gp->state = GMON_PROF_OFF;
 308                         stopprofclock(&proc0);
 309                         startguprof(gp);
 310                         gp->state = state;
 311 #endif
 312                 } else if (state != gp->state) {
 313                         error = EINVAL;
 314                 }
 315                 lwkt_reltoken(&proc0.p_token);
 316                 return (error);
 317         case GPROF_COUNT:
 318                 return (sysctl_handle_opaque(oidp,
 319                         gp->kcount, gp->kcountsize, req));
 320         case GPROF_FROMS:
 321                 return (sysctl_handle_opaque(oidp,
 322                         gp->froms, gp->fromssize, req));
 323         case GPROF_TOS:
 324                 return (sysctl_handle_opaque(oidp,
 325                         gp->tos, gp->tossize, req));
 326         case GPROF_GMONPARAM:
 327                 return (sysctl_handle_opaque(oidp, gp, sizeof *gp, req));
 328         default:
 329                 return (EOPNOTSUPP);
 330         }
 331         /* NOTREACHED */
 332 }
 333
 334 SYSCTL_NODE(_kern, KERN_PROF, prof, CTLFLAG_RW, sysctl_kern_prof, "");
 335 #endif /* GPROF */
 336
 337 /*
 338  * Profiling system call.
 339  *
 340  * The scale factor is a fixed point number with 16 bits of fraction, so that
 341  * 1.0 is represented as 0x10000.  A scale factor of 0 turns off profiling.
 342  *
 343  * MPALMOSTSAFE
 344  */
 345 int
 346 sys_profil(struct profil_args *uap)
 347 {
 348         struct proc *p = curproc;
 349         struct uprof *upp;
 350
 351         if (uap->scale > (1 << 16))
 352                 return (EINVAL);
 353         lwkt_gettoken(&p->p_token);
 354         if (uap->scale == 0) {
 355                 stopprofclock(p);
 356         } else {
 357                 upp = &p->p_prof;
 358
 359                 /* Block profile interrupts while changing state. */
 360                 crit_enter();
 361                 upp->pr_off = uap->offset;
 362                 upp->pr_scale = uap->scale;
 363                 upp->pr_base = uap->samples;
 364                 upp->pr_size = uap->size;
 365                 startprofclock(p);
 366                 crit_exit();
 367         }
 368         lwkt_reltoken(&p->p_token);
 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.
 386  *
 387  * Note that we may (rarely) not get around to the AST soon enough, and
 388  * lose profile ticks when the next tick overwrites this one, but in this
 389  * case the system is overloaded and the profile is probably already
 390  * inaccurate.
 391  */
 392 void
 393 addupc_intr(struct proc *p, u_long pc, u_int ticks)
 394 {
 395         struct uprof *prof;
 396         u_int i;
 397
 398         if (ticks == 0)
 399                 return;
 400         prof = &p->p_prof;
 401         if (pc < prof->pr_off ||
 402             (i = PC_TO_INDEX(pc, prof)) >= prof->pr_size)
 403                 return;                 /* out of range; ignore */
 404
 405         prof->pr_addr = pc;
 406         prof->pr_ticks = ticks;
 407         need_proftick();
 408 }
 409
 410 /*
 411  * Much like before, but we can afford to take faults here.  If the
 412  * update fails, we simply turn off profiling.
 413  */
 414 void
 415 addupc_task(struct proc *p, u_long pc, u_int ticks)
 416 {
 417         struct uprof *prof;
 418         caddr_t addr;
 419         u_int i;
 420         u_short v;
 421
 422         /* Testing P_PROFIL may be unnecessary, but is certainly safe. */
 423         if ((p->p_flags & P_PROFIL) == 0 || ticks == 0)
 424                 return;
 425
 426         prof = &p->p_prof;
 427         if (pc < prof->pr_off ||
 428             (i = PC_TO_INDEX(pc, prof)) >= prof->pr_size)
 429                 return;
 430
 431         addr = prof->pr_base + i;
 432         if (copyin(addr, (caddr_t)&v, sizeof(v)) == 0) {
 433                 v += ticks;
 434                 if (copyout((caddr_t)&v, addr, sizeof(v)) == 0)
 435                         return;
 436         }
 437         stopprofclock(p);
 438 }