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 #include <sys/mplock2.h>
  43
  44 #include <machine/cpu.h>
  45
  46 #ifdef GPROF
  47 #include <sys/malloc.h>
  48 #include <sys/gmon.h>
  49 #undef MCOUNT
  50
  51 static MALLOC_DEFINE(M_GPROF, "gprof", "kernel profiling buffer");
  52
  53 static void kmstartup (void *);
  54 SYSINIT(kmem, SI_SUB_KPROF, SI_ORDER_FIRST, kmstartup, NULL);
  55
  56 struct gmonparam _gmonparam = { GMON_PROF_OFF };
  57
  58 #ifdef GUPROF
  59 #include <machine/asmacros.h>
  60
  61 void
  62 nullfunc_loop_profiled(void)
  63 {
  64         int i;
  65
  66         for (i = 0; i < CALIB_SCALE; i++)
  67                 nullfunc_profiled();
  68 }
  69
  70 #define nullfunc_loop_profiled_end      nullfunc_profiled       /* XXX */
  71
  72 void
  73 nullfunc_profiled(void)
  74 {
  75 }
  76 #endif /* GUPROF */
  77
  78 static void
  79 kmstartup(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         uintfptr_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         kprintf("Profiling kernel, textsize=%lu [%jx..%jx]\n",
 102             p->textsize, (uintmax_t)p->lowpc, (uintmax_t)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 *)kmalloc(p->kcountsize + p->fromssize + p->tossize,
 113             M_GPROF, M_NOWAIT | M_ZERO);
 114         if (cp == NULL) {
 115                 kprintf("No memory for profiling.\n");
 116                 return;
 117         }
 118         p->tos = (struct tostruct *)cp;
 119         cp += p->tossize;
 120         p->kcount = (HISTCOUNTER *)cp;
 121         cp += p->kcountsize;
 122         p->froms = (u_short *)cp;
 123
 124 #ifdef GUPROF
 125         /* Initialize pointers to overhead counters. */
 126         p->cputime_count = &KCOUNT(p, PC_TO_I(p, cputime));
 127         p->mcount_count = &KCOUNT(p, PC_TO_I(p, mcount));
 128         p->mexitcount_count = &KCOUNT(p, PC_TO_I(p, mexitcount));
 129
 130         /*
 131          * Disable interrupts to avoid interference while we calibrate
 132          * things.
 133          */
 134         cpu_disable_intr();
 135
 136         /*
 137          * Determine overheads.
 138          * XXX this needs to be repeated for each useful timer/counter.
 139          */
 140         cputime_overhead = 0;
 141         startguprof(p);
 142         for (i = 0; i < CALIB_SCALE; i++)
 143                 cputime_overhead += cputime();
 144
 145         empty_loop();
 146         startguprof(p);
 147         empty_loop();
 148         empty_loop_time = cputime();
 149
 150         nullfunc_loop_profiled();
 151
 152         /*
 153          * Start profiling.  There won't be any normal function calls since
 154          * interrupts are disabled, but we will call the profiling routines
 155          * directly to determine their overheads.
 156          */
 157         p->state = GMON_PROF_HIRES;
 158
 159         startguprof(p);
 160         nullfunc_loop_profiled();
 161
 162         startguprof(p);
 163         for (i = 0; i < CALIB_SCALE; i++)
 164 #if defined(__i386__) && __GNUC__ >= 2
 165                 __asm("pushl %0; call __mcount; popl %%ecx"
 166                       :
 167                       : "i" (profil)
 168                       : "ax", "bx", "cx", "dx", "memory");
 169 #else
 170 #error
 171 #endif
 172         mcount_overhead = KCOUNT(p, PC_TO_I(p, profil));
 173
 174         startguprof(p);
 175         for (i = 0; i < CALIB_SCALE; i++)
 176 #if defined(__i386__) && __GNUC__ >= 2
 177                     __asm("call " __XSTRING(HIDENAME(mexitcount)) "; 1:"
 178                           : : : "ax", "bx", "cx", "dx", "memory");
 179         __asm("movl $1b,%0" : "=rm" (tmp_addr));
 180 #else
 181 #error
 182 #endif
 183         mexitcount_overhead = KCOUNT(p, PC_TO_I(p, tmp_addr));
 184
 185         p->state = GMON_PROF_OFF;
 186         stopguprof(p);
 187
 188         cpu_enable_intr();
 189
 190         nullfunc_loop_profiled_time = 0;
 191         for (tmp_addr = (uintfptr_t)nullfunc_loop_profiled;
 192              tmp_addr < (uintfptr_t)nullfunc_loop_profiled_end;
 193              tmp_addr += HISTFRACTION * sizeof(HISTCOUNTER))
 194                 nullfunc_loop_profiled_time += KCOUNT(p, PC_TO_I(p, tmp_addr));
 195 #define CALIB_DOSCALE(count)    (((count) + CALIB_SCALE / 3) / CALIB_SCALE)
 196 #define c2n(count, freq)        ((int)((count) * 1000000000LL / freq))
 197         kprintf("cputime %d, empty_loop %d, nullfunc_loop_profiled %d, mcount %d, mexitcount %d\n",
 198                CALIB_DOSCALE(c2n(cputime_overhead, p->profrate)),
 199                CALIB_DOSCALE(c2n(empty_loop_time, p->profrate)),
 200                CALIB_DOSCALE(c2n(nullfunc_loop_profiled_time, p->profrate)),
 201                CALIB_DOSCALE(c2n(mcount_overhead, p->profrate)),
 202                CALIB_DOSCALE(c2n(mexitcount_overhead, p->profrate)));
 203         cputime_overhead -= empty_loop_time;
 204         mcount_overhead -= empty_loop_time;
 205         mexitcount_overhead -= empty_loop_time;
 206
 207         /*-
 208          * Profiling overheads are determined by the times between the
 209          * following events:
 210          *      MC1: mcount() is called
 211          *      MC2: cputime() (called from mcount()) latches the timer
 212          *      MC3: mcount() completes
 213          *      ME1: mexitcount() is called
 214          *      ME2: cputime() (called from mexitcount()) latches the timer
 215          *      ME3: mexitcount() completes.
 216          * The times between the events vary slightly depending on instruction
 217          * combination and cache misses, etc.  Attempt to determine the
 218          * minimum times.  These can be subtracted from the profiling times
 219          * without much risk of reducing the profiling times below what they
 220          * would be when profiling is not configured.  Abbreviate:
 221          *      ab = minimum time between MC1 and MC3
 222          *      a  = minumum time between MC1 and MC2
 223          *      b  = minimum time between MC2 and MC3
 224          *      cd = minimum time between ME1 and ME3
 225          *      c  = minimum time between ME1 and ME2
 226          *      d  = minimum time between ME2 and ME3.
 227          * These satisfy the relations:
 228          *      ab            <= mcount_overhead                (just measured)
 229          *      a + b         <= ab
 230          *              cd    <= mexitcount_overhead            (just measured)
 231          *              c + d <= cd
 232          *      a         + d <= nullfunc_loop_profiled_time    (just measured)
 233          *      a >= 0, b >= 0, c >= 0, d >= 0.
 234          * Assume that ab and cd are equal to the minimums.
 235          */
 236         p->cputime_overhead = CALIB_DOSCALE(cputime_overhead);
 237         p->mcount_overhead = CALIB_DOSCALE(mcount_overhead - cputime_overhead);
 238         p->mexitcount_overhead = CALIB_DOSCALE(mexitcount_overhead
 239                                                - cputime_overhead);
 240         nullfunc_loop_overhead = nullfunc_loop_profiled_time - empty_loop_time;
 241         p->mexitcount_post_overhead = CALIB_DOSCALE((mcount_overhead
 242                                                      - nullfunc_loop_overhead)
 243                                                     / 4);
 244         p->mexitcount_pre_overhead = p->mexitcount_overhead
 245                                      + p->cputime_overhead
 246                                      - p->mexitcount_post_overhead;
 247         p->mcount_pre_overhead = CALIB_DOSCALE(nullfunc_loop_overhead)
 248                                  - p->mexitcount_post_overhead;
 249         p->mcount_post_overhead = p->mcount_overhead
 250                                   + p->cputime_overhead
 251                                   - p->mcount_pre_overhead;
 252         kprintf(
 253 "Profiling overheads: mcount: %d+%d, %d+%d; mexitcount: %d+%d, %d+%d nsec\n",
 254                c2n(p->cputime_overhead, p->profrate),
 255                c2n(p->mcount_overhead, p->profrate),
 256                c2n(p->mcount_pre_overhead, p->profrate),
 257                c2n(p->mcount_post_overhead, p->profrate),
 258                c2n(p->cputime_overhead, p->profrate),
 259                c2n(p->mexitcount_overhead, p->profrate),
 260                c2n(p->mexitcount_pre_overhead, p->profrate),
 261                c2n(p->mexitcount_post_overhead, p->profrate));
 262         kprintf(
 263 "Profiling overheads: mcount: %d+%d, %d+%d; mexitcount: %d+%d, %d+%d cycles\n",
 264                p->cputime_overhead, p->mcount_overhead,
 265                p->mcount_pre_overhead, p->mcount_post_overhead,
 266                p->cputime_overhead, p->mexitcount_overhead,
 267                p->mexitcount_pre_overhead, p->mexitcount_post_overhead);
 268 #endif /* GUPROF */
 269 }
 270
 271 /*
 272  * Return kernel profiling information.
 273  */
 274 static int
 275 sysctl_kern_prof(SYSCTL_HANDLER_ARGS)
 276 {
 277         int *name = (int *) arg1;
 278         u_int namelen = arg2;
 279         struct gmonparam *gp = &_gmonparam;
 280         int error;
 281         int state;
 282
 283         /* all sysctl names at this level are terminal */
 284         if (namelen != 1)
 285                 return (ENOTDIR);               /* overloaded */
 286
 287         switch (name[0]) {
 288         case GPROF_STATE:
 289                 state = gp->state;
 290                 error = sysctl_handle_int(oidp, &state, 0, req);
 291                 if (error)
 292                         return (error);
 293                 if (!req->newptr)
 294                         return (0);
 295                 lwkt_gettoken(&proc0.p_token);
 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                         error = EINVAL;
 315                 }
 316                 lwkt_reltoken(&proc0.p_token);
 317                 return (error);
 318         case GPROF_COUNT:
 319                 return (sysctl_handle_opaque(oidp,
 320                         gp->kcount, gp->kcountsize, req));
 321         case GPROF_FROMS:
 322                 return (sysctl_handle_opaque(oidp,
 323                         gp->froms, gp->fromssize, req));
 324         case GPROF_TOS:
 325                 return (sysctl_handle_opaque(oidp,
 326                         gp->tos, gp->tossize, req));
 327         case GPROF_GMONPARAM:
 328                 return (sysctl_handle_opaque(oidp, gp, sizeof *gp, req));
 329         default:
 330                 return (EOPNOTSUPP);
 331         }
 332         /* NOTREACHED */
 333 }
 334
 335 SYSCTL_NODE(_kern, KERN_PROF, prof, CTLFLAG_RW, sysctl_kern_prof, "");
 336 #endif /* GPROF */
 337
 338 /*
 339  * Profiling system call.
 340  *
 341  * The scale factor is a fixed point number with 16 bits of fraction, so that
 342  * 1.0 is represented as 0x10000.  A scale factor of 0 turns off profiling.
 343  *
 344  * MPALMOSTSAFE
 345  */
 346 int
 347 sys_profil(struct profil_args *uap)
 348 {
 349         struct proc *p = curproc;
 350         struct uprof *upp;
 351
 352         if (uap->scale > (1 << 16))
 353                 return (EINVAL);
 354         get_mplock();
 355         if (uap->scale == 0) {
 356                 stopprofclock(p);
 357         } else {
 358                 upp = &p->p_prof;
 359
 360                 /* Block profile interrupts while changing state. */
 361                 crit_enter();
 362                 upp->pr_off = uap->offset;
 363                 upp->pr_scale = uap->scale;
 364                 upp->pr_base = uap->samples;
 365                 upp->pr_size = uap->size;
 366                 startprofclock(p);
 367                 crit_exit();
 368         }
 369         rel_mplock();
 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  * XXX fuswintr() and suswintr() never worked (always returnde -1), remove
 391  * them.  It's just a bad idea to try to do this from a hard interrupt.
 392  *
 393  * Note that we may (rarely) not get around to the AST soon enough, and
 394  * lose profile ticks when the next tick overwrites this one, but in this
 395  * case the system is overloaded and the profile is probably already
 396  * inaccurate.
 397  */
 398 void
 399 addupc_intr(struct proc *p, u_long pc, u_int ticks)
 400 {
 401         struct uprof *prof;
 402         u_int i;
 403
 404         if (ticks == 0)
 405                 return;
 406         prof = &p->p_prof;
 407         if (pc < prof->pr_off ||
 408             (i = PC_TO_INDEX(pc, prof)) >= prof->pr_size)
 409                 return;                 /* out of range; ignore */
 410
 411         prof->pr_addr = pc;
 412         prof->pr_ticks = ticks;
 413         need_proftick();
 414 }
 415
 416 /*
 417  * Much like before, but we can afford to take faults here.  If the
 418  * update fails, we simply turn off profiling.
 419  */
 420 void
 421 addupc_task(struct proc *p, u_long pc, u_int ticks)
 422 {
 423         struct uprof *prof;
 424         caddr_t addr;
 425         u_int i;
 426         u_short v;
 427
 428         /* Testing P_PROFIL may be unnecessary, but is certainly safe. */
 429         if ((p->p_flags & P_PROFIL) == 0 || ticks == 0)
 430                 return;
 431
 432         prof = &p->p_prof;
 433         if (pc < prof->pr_off ||
 434             (i = PC_TO_INDEX(pc, prof)) >= prof->pr_size)
 435                 return;
 436
 437         addr = prof->pr_base + i;
 438         if (copyin(addr, (caddr_t)&v, sizeof(v)) == 0) {
 439                 v += ticks;
 440                 if (copyout((caddr_t)&v, addr, sizeof(v)) == 0)
 441                         return;
 442         }
 443         stopprofclock(p);
 444 }