sys/i386/isa/prof_machdep.c

   1 /*-
   2  * Copyright (c) 1996 Bruce D. Evans.
   3  * 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  *
  14  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
  15  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  16  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  17  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
  18  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  19  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  20  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  21  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  22  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  23  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  24  * SUCH DAMAGE.
  25  *
  26  * $FreeBSD: src/sys/i386/isa/prof_machdep.c,v 1.14.2.1 2000/08/03 00:09:30 ps Exp $
  27  */
  28
  29 #ifdef GUPROF
  30 #include "opt_i586_guprof.h"
  31 #include "opt_perfmon.h"
  32
  33 #include <sys/param.h>
  34 #include <sys/systm.h>
  35 #include <sys/gmon.h>
  36 #include <sys/kernel.h>
  37 #include <sys/sysctl.h>
  38
  39 #include <machine/clock.h>
  40 #include <machine/perfmon.h>
  41 #include <machine/profile.h>
  42 #undef MCOUNT
  43 #endif
  44
  45 #include <machine/asmacros.h>
  46
  47 #ifdef PC98
  48 #include <pc98/pc98/pc98.h>
  49 #else
  50 #include <i386/isa/isa.h>
  51 #endif
  52 #include <i386/isa/timerreg.h>
  53
  54 #ifdef GUPROF
  55 #define CPUTIME_CLOCK_UNINITIALIZED     0
  56 #define CPUTIME_CLOCK_I8254             1
  57 #define CPUTIME_CLOCK_TSC               2
  58 #define CPUTIME_CLOCK_I586_PMC          3
  59 #define CPUTIME_CLOCK_I8254_SHIFT       7
  60
  61 int     cputime_bias = 1;       /* initialize for locality of reference */
  62
  63 static int      cputime_clock = CPUTIME_CLOCK_UNINITIALIZED;
  64 #ifdef I586_PMC_GUPROF
  65 static u_int    cputime_clock_pmc_conf = I586_PMC_GUPROF;
  66 static int      cputime_clock_pmc_init;
  67 static struct gmonparam saved_gmp;
  68 #endif
  69 #endif /* GUPROF */
  70
  71 #ifdef __GNUC__
  72 __asm("                                                         \n\
  73 GM_STATE        =       0                                       \n\
  74 GMON_PROF_OFF   =       3                                       \n\
  75                                                                 \n\
  76         .text                                                   \n\
  77         .p2align 4,0x90                                         \n\
  78         .globl  __mcount                                        \n\
  79         .type   __mcount,@function                              \n\
  80 __mcount:                                                       \n\
  81         #                                                       \n\
  82         # Check that we are profiling.  Do it early for speed.  \n\
  83         #                                                       \n\
  84         cmpl    $GMON_PROF_OFF," __XSTRING(CNAME(_gmonparam)) "+GM_STATE \n\
  85         je      .mcount_exit                                    \n\
  86         #                                                       \n\
  87         # __mcount is the same as [.]mcount except the caller   \n\
  88         # hasn't changed the stack except to call here, so the  \n\
  89         # caller's raddr is above our raddr.                    \n\
  90         #                                                       \n\
  91         movl    4(%esp),%edx                                    \n\
  92         jmp     .got_frompc                                     \n\
  93                                                                 \n\
  94         .p2align 4,0x90                                         \n\
  95         .globl  " __XSTRING(HIDENAME(mcount)) "                 \n\
  96 " __XSTRING(HIDENAME(mcount)) ":                                \n\
  97         cmpl    $GMON_PROF_OFF," __XSTRING(CNAME(_gmonparam)) "+GM_STATE \n\
  98         je      .mcount_exit                                    \n\
  99         #                                                       \n\
 100         # The caller's stack frame has already been built, so   \n\
 101         # %ebp is the caller's frame pointer.  The caller's     \n\
 102         # raddr is in the caller's frame following the caller's \n\
 103         # caller's frame pointer.                               \n\
 104         #                                                       \n\
 105         movl    4(%ebp),%edx                                    \n\
 106 .got_frompc:                                                    \n\
 107         #                                                       \n\
 108         # Our raddr is the caller's pc.                         \n\
 109         #                                                       \n\
 110         movl    (%esp),%eax                                     \n\
 111                                                                 \n\
 112         pushfl                                                  \n\
 113         pushl   %eax                                            \n\
 114         pushl   %edx                                            \n\
 115         cli                                                     \n\
 116         call    " __XSTRING(CNAME(mcount)) "                    \n\
 117         addl    $8,%esp                                         \n\
 118         popfl                                                   \n\
 119 .mcount_exit:                                                   \n\
 120         ret                                                     \n\
 121 ");
 122 #else /* !__GNUC__ */
 123 #error
 124 #endif /* __GNUC__ */
 125
 126 #ifdef GUPROF
 127 /*
 128  * [.]mexitcount saves the return register(s), loads selfpc and calls
 129  * mexitcount(selfpc) to do the work.  Someday it should be in a machine
 130  * dependent file together with cputime(), __mcount and [.]mcount.  cputime()
 131  * can't just be put in machdep.c because it has to be compiled without -pg.
 132  */
 133 #ifdef __GNUC__
 134 __asm("                                                         \n\
 135         .text                                                   \n\
 136 #                                                               \n\
 137 # Dummy label to be seen when gprof -u hides [.]mexitcount.     \n\
 138 #                                                               \n\
 139         .p2align 4,0x90                                         \n\
 140         .globl  __mexitcount                                    \n\
 141         .type   __mexitcount,@function                          \n\
 142 __mexitcount:                                                   \n\
 143         nop                                                     \n\
 144                                                                 \n\
 145 GMON_PROF_HIRES =       4                                       \n\
 146                                                                 \n\
 147         .p2align 4,0x90                                         \n\
 148         .globl  " __XSTRING(HIDENAME(mexitcount)) "             \n\
 149 " __XSTRING(HIDENAME(mexitcount)) ":                            \n\
 150         cmpl    $GMON_PROF_HIRES," __XSTRING(CNAME(_gmonparam)) "+GM_STATE \n\
 151         jne     .mexitcount_exit                                \n\
 152         pushl   %edx                                            \n\
 153         pushl   %eax                                            \n\
 154         movl    8(%esp),%eax                                    \n\
 155         pushfl                                                  \n\
 156         pushl   %eax                                            \n\
 157         cli                                                     \n\
 158         call    " __XSTRING(CNAME(mexitcount)) "                \n\
 159         addl    $4,%esp                                         \n\
 160         popfl                                                   \n\
 161         popl    %eax                                            \n\
 162         popl    %edx                                            \n\
 163 .mexitcount_exit:                                               \n\
 164         ret                                                     \n\
 165 ");
 166 #else /* !__GNUC__ */
 167 #error
 168 #endif /* __GNUC__ */
 169
 170 /*
 171  * Return the time elapsed since the last call.  The units are machine-
 172  * dependent.
 173  */
 174 int
 175 cputime()
 176 {
 177         u_int count;
 178         int delta;
 179 #if (defined(I586_CPU) || defined(I686_CPU)) && !defined(SMP) && \
 180     defined(PERFMON) && defined(I586_PMC_GUPROF)
 181         u_quad_t event_count;
 182 #endif
 183         u_char high, low;
 184         static u_int prev_count;
 185
 186 #if (defined(I586_CPU) || defined(I686_CPU)) && !defined(SMP)
 187         if (cputime_clock == CPUTIME_CLOCK_TSC) {
 188                 count = (u_int)rdtsc();
 189                 delta = (int)(count - prev_count);
 190                 prev_count = count;
 191                 return (delta);
 192         }
 193 #if defined(PERFMON) && defined(I586_PMC_GUPROF)
 194         if (cputime_clock == CPUTIME_CLOCK_I586_PMC) {
 195                 /*
 196                  * XXX permon_read() should be inlined so that the
 197                  * perfmon module doesn't need to be compiled with
 198                  * profiling disabled and so that it is fast.
 199                  */
 200                 perfmon_read(0, &event_count);
 201
 202                 count = (u_int)event_count;
 203                 delta = (int)(count - prev_count);
 204                 prev_count = count;
 205                 return (delta);
 206         }
 207 #endif /* PERFMON && I586_PMC_GUPROF */
 208 #endif /* (I586_CPU || I686_CPU) && !SMP */
 209
 210         /*
 211          * Read the current value of the 8254 timer counter 0.
 212          */
 213         outb(TIMER_MODE, TIMER_SEL0 | TIMER_LATCH);
 214         low = inb(TIMER_CNTR0);
 215         high = inb(TIMER_CNTR0);
 216         count = ((high << 8) | low) << CPUTIME_CLOCK_I8254_SHIFT;
 217
 218         /*
 219          * The timer counts down from TIMER_CNTR0_MAX to 0 and then resets.
 220          * While profiling is enabled, this routine is called at least twice
 221          * per timer reset (for mcounting and mexitcounting hardclock()),
 222          * so at most one reset has occurred since the last call, and one
 223          * has occurred iff the current count is larger than the previous
 224          * count.  This allows counter underflow to be detected faster
 225          * than in microtime().
 226          */
 227         delta = prev_count - count;
 228         prev_count = count;
 229         if ((int) delta <= 0)
 230                 return (delta + (timer0_max_count << CPUTIME_CLOCK_I8254_SHIFT));
 231         return (delta);
 232 }
 233
 234 static int
 235 sysctl_machdep_cputime_clock(SYSCTL_HANDLER_ARGS)
 236 {
 237         int clock;
 238         int error;
 239 #if defined(PERFMON) && defined(I586_PMC_GUPROF)
 240         int event;
 241         struct pmc pmc;
 242 #endif
 243
 244         clock = cputime_clock;
 245 #if defined(PERFMON) && defined(I586_PMC_GUPROF)
 246         if (clock == CPUTIME_CLOCK_I586_PMC) {
 247                 pmc.pmc_val = cputime_clock_pmc_conf;
 248                 clock += pmc.pmc_event;
 249         }
 250 #endif
 251         error = sysctl_handle_opaque(oidp, &clock, sizeof clock, req);
 252         if (error == 0 && req->newptr != NULL) {
 253 #if defined(PERFMON) && defined(I586_PMC_GUPROF)
 254                 if (clock >= CPUTIME_CLOCK_I586_PMC) {
 255                         event = clock - CPUTIME_CLOCK_I586_PMC;
 256                         if (event >= 256)
 257                                 return (EINVAL);
 258                         pmc.pmc_num = 0;
 259                         pmc.pmc_event = event;
 260                         pmc.pmc_unit = 0;
 261                         pmc.pmc_flags = PMCF_E | PMCF_OS | PMCF_USR;
 262                         pmc.pmc_mask = 0;
 263                         cputime_clock_pmc_conf = pmc.pmc_val;
 264                         cputime_clock = CPUTIME_CLOCK_I586_PMC;
 265                 } else
 266 #endif
 267                 {
 268                         if (clock < 0 || clock >= CPUTIME_CLOCK_I586_PMC)
 269                                 return (EINVAL);
 270                         cputime_clock = clock;
 271                 }
 272         }
 273         return (error);
 274 }
 275
 276 SYSCTL_PROC(_machdep, OID_AUTO, cputime_clock, CTLTYPE_INT | CTLFLAG_RW,
 277             0, sizeof(u_int), sysctl_machdep_cputime_clock, "I", "");
 278
 279 /*
 280  * The start and stop routines need not be here since we turn off profiling
 281  * before calling them.  They are here for convenience.
 282  */
 283
 284 void
 285 startguprof(gp)
 286         struct gmonparam *gp;
 287 {
 288         if (cputime_clock == CPUTIME_CLOCK_UNINITIALIZED) {
 289                 cputime_clock = CPUTIME_CLOCK_I8254;
 290 #if (defined(I586_CPU) || defined(I686_CPU)) && !defined(SMP)
 291                 if (tsc_freq != 0)
 292                         cputime_clock = CPUTIME_CLOCK_TSC;
 293 #endif
 294         }
 295         gp->profrate = timer_freq << CPUTIME_CLOCK_I8254_SHIFT;
 296 #if (defined(I586_CPU) || defined(I686_CPU)) && !defined(SMP)
 297         if (cputime_clock == CPUTIME_CLOCK_TSC)
 298                 gp->profrate = tsc_freq;
 299 #if defined(PERFMON) && defined(I586_PMC_GUPROF)
 300         else if (cputime_clock == CPUTIME_CLOCK_I586_PMC) {
 301                 if (perfmon_avail() &&
 302                     perfmon_setup(0, cputime_clock_pmc_conf) == 0) {
 303                         if (perfmon_start(0) != 0)
 304                                 perfmon_fini(0);
 305                         else {
 306                                 /* XXX 1 event == 1 us. */
 307                                 gp->profrate = 1000000;
 308
 309                                 saved_gmp = *gp;
 310
 311                                 /* Zap overheads.  They are invalid. */
 312                                 gp->cputime_overhead = 0;
 313                                 gp->mcount_overhead = 0;
 314                                 gp->mcount_post_overhead = 0;
 315                                 gp->mcount_pre_overhead = 0;
 316                                 gp->mexitcount_overhead = 0;
 317                                 gp->mexitcount_post_overhead = 0;
 318                                 gp->mexitcount_pre_overhead = 0;
 319
 320                                 cputime_clock_pmc_init = TRUE;
 321                         }
 322                 }
 323         }
 324 #endif /* PERFMON && I586_PMC_GUPROF */
 325 #endif /* (I586_CPU || I686_CPU) && !SMP */
 326         cputime_bias = 0;
 327         cputime();
 328 }
 329
 330 void
 331 stopguprof(gp)
 332         struct gmonparam *gp;
 333 {
 334 #if defined(PERFMON) && defined(I586_PMC_GUPROF)
 335         if (cputime_clock_pmc_init) {
 336                 *gp = saved_gmp;
 337                 perfmon_fini(0);
 338                 cputime_clock_pmc_init = FALSE;
 339         }
 340 #endif
 341 }
 342
 343 #else /* !GUPROF */
 344 #ifdef __GNUC__
 345 __asm("                                                         \n\
 346         .text                                                   \n\
 347         .p2align 4,0x90                                         \n\
 348         .globl  " __XSTRING(HIDENAME(mexitcount)) "             \n\
 349 " __XSTRING(HIDENAME(mexitcount)) ":                            \n\
 350         ret                                                     \n\
 351 ");
 352 #else /* !__GNUC__ */
 353 #error
 354 #endif /* __GNUC__ */
 355 #endif /* GUPROF */