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
  27 #include <sys/cdefs.h>
  28 __FBSDID("$FreeBSD$");
  29
  30 #ifdef GUPROF
  31 #include "opt_i586_guprof.h"
  32 #include "opt_perfmon.h"
  33
  34 #include <sys/param.h>
  35 #include <sys/systm.h>
  36 #include <sys/bus.h>
  37 #include <sys/cpu.h>
  38 #include <sys/eventhandler.h>
  39 #include <sys/gmon.h>
  40 #include <sys/kernel.h>
  41 #include <sys/smp.h>
  42 #include <sys/sysctl.h>
  43
  44 #include <machine/clock.h>
  45 #include <machine/perfmon.h>
  46 #include <machine/timerreg.h>
  47
  48 #define CPUTIME_CLOCK_UNINITIALIZED     0
  49 #define CPUTIME_CLOCK_I8254             1
  50 #define CPUTIME_CLOCK_TSC               2
  51 #define CPUTIME_CLOCK_I586_PMC          3
  52 #define CPUTIME_CLOCK_I8254_SHIFT       7
  53
  54 int     cputime_bias = 1;       /* initialize for locality of reference */
  55
  56 static int      cputime_clock = CPUTIME_CLOCK_UNINITIALIZED;
  57 #if defined(PERFMON) && defined(I586_PMC_GUPROF)
  58 static u_int    cputime_clock_pmc_conf = I586_PMC_GUPROF;
  59 static int      cputime_clock_pmc_init;
  60 static struct gmonparam saved_gmp;
  61 #endif
  62 #if defined(I586_CPU) || defined(I686_CPU)
  63 static int      cputime_prof_active;
  64 #endif
  65 #endif /* GUPROF */
  66
  67 #ifdef __GNUCLIKE_ASM
  68 __asm("                                                         \n\
  69 GM_STATE        =       0                                       \n\
  70 GMON_PROF_OFF   =       3                                       \n\
  71                                                                 \n\
  72         .text                                                   \n\
  73         .p2align 4,0x90                                         \n\
  74         .globl  __mcount                                        \n\
  75         .type   __mcount,@function                              \n\
  76 __mcount:                                                       \n\
  77         #                                                       \n\
  78         # Check that we are profiling.  Do it early for speed.  \n\
  79         #                                                       \n\
  80         cmpl    $GMON_PROF_OFF,_gmonparam+GM_STATE              \n\
  81         je      .mcount_exit                                    \n\
  82         #                                                       \n\
  83         # __mcount is the same as [.]mcount except the caller   \n\
  84         # hasn't changed the stack except to call here, so the  \n\
  85         # caller's raddr is above our raddr.                    \n\
  86         #                                                       \n\
  87         movl    4(%esp),%edx                                    \n\
  88         jmp     .got_frompc                                     \n\
  89                                                                 \n\
  90         .p2align 4,0x90                                         \n\
  91         .globl  .mcount                                         \n\
  92 .mcount:                                                        \n\
  93         .globl  __cyg_profile_func_enter                        \n\
  94 __cyg_profile_func_enter:                                       \n\
  95         cmpl    $GMON_PROF_OFF,_gmonparam+GM_STATE              \n\
  96         je      .mcount_exit                                    \n\
  97         #                                                       \n\
  98         # The caller's stack frame has already been built, so   \n\
  99         # %ebp is the caller's frame pointer.  The caller's     \n\
 100         # raddr is in the caller's frame following the caller's \n\
 101         # caller's frame pointer.                               \n\
 102         #                                                       \n\
 103         movl    4(%ebp),%edx                                    \n\
 104 .got_frompc:                                                    \n\
 105         #                                                       \n\
 106         # Our raddr is the caller's pc.                         \n\
 107         #                                                       \n\
 108         movl    (%esp),%eax                                     \n\
 109                                                                 \n\
 110         pushfl                                                  \n\
 111         pushl   %eax                                            \n\
 112         pushl   %edx                                            \n\
 113         cli                                                     \n\
 114         call    mcount                                          \n\
 115         addl    $8,%esp                                         \n\
 116         popfl                                                   \n\
 117 .mcount_exit:                                                   \n\
 118         ret                                                     \n\
 119 ");
 120 #else /* !__GNUCLIKE_ASM */
 121 #error "this file needs to be ported to your compiler"
 122 #endif /* __GNUCLIKE_ASM */
 123
 124 #ifdef GUPROF
 125 /*
 126  * [.]mexitcount saves the return register(s), loads selfpc and calls
 127  * mexitcount(selfpc) to do the work.  Someday it should be in a machine
 128  * dependent file together with cputime(), __mcount and [.]mcount.  cputime()
 129  * can't just be put in machdep.c because it has to be compiled without -pg.
 130  */
 131 #ifdef __GNUCLIKE_ASM
 132 __asm("                                                         \n\
 133         .text                                                   \n\
 134 #                                                               \n\
 135 # Dummy label to be seen when gprof -u hides [.]mexitcount.     \n\
 136 #                                                               \n\
 137         .p2align 4,0x90                                         \n\
 138         .globl  __mexitcount                                    \n\
 139         .type   __mexitcount,@function                          \n\
 140 __mexitcount:                                                   \n\
 141         nop                                                     \n\
 142                                                                 \n\
 143 GMON_PROF_HIRES =       4                                       \n\
 144                                                                 \n\
 145         .p2align 4,0x90                                         \n\
 146         .globl  .mexitcount                                     \n\
 147 .mexitcount:                                                    \n\
 148         .globl  __cyg_profile_func_exit                         \n\
 149 __cyg_profile_func_exit:                                        \n\
 150         cmpl    $GMON_PROF_HIRES,_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    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 #endif /* __GNUCLIKE_ASM */
 167
 168 /*
 169  * Return the time elapsed since the last call.  The units are machine-
 170  * dependent.
 171  */
 172 int
 173 cputime()
 174 {
 175         u_int count;
 176         int delta;
 177 #if (defined(I586_CPU) || defined(I686_CPU)) && !defined(SMP) && \
 178     defined(PERFMON) && defined(I586_PMC_GUPROF)
 179         u_quad_t event_count;
 180 #endif
 181         u_char high, low;
 182         static u_int prev_count;
 183
 184 #if defined(I586_CPU) || defined(I686_CPU)
 185         if (cputime_clock == CPUTIME_CLOCK_TSC) {
 186                 /*
 187                  * Scale the TSC a little to make cputime()'s frequency
 188                  * fit in an int, assuming that the TSC frequency fits
 189                  * in a u_int.  Use a fixed scale since dynamic scaling
 190                  * would be slower and we can't really use the low bit
 191                  * of precision.
 192                  */
 193                 count = (u_int)rdtsc() & ~1u;
 194                 delta = (int)(count - prev_count) >> 1;
 195                 prev_count = count;
 196                 return (delta);
 197         }
 198 #if defined(PERFMON) && defined(I586_PMC_GUPROF) && !defined(SMP)
 199         if (cputime_clock == CPUTIME_CLOCK_I586_PMC) {
 200                 /*
 201                  * XXX permon_read() should be inlined so that the
 202                  * perfmon module doesn't need to be compiled with
 203                  * profiling disabled and so that it is fast.
 204                  */
 205                 perfmon_read(0, &event_count);
 206
 207                 count = (u_int)event_count;
 208                 delta = (int)(count - prev_count);
 209                 prev_count = count;
 210                 return (delta);
 211         }
 212 #endif /* PERFMON && I586_PMC_GUPROF && !SMP */
 213 #endif /* I586_CPU || I686_CPU */
 214
 215         /*
 216          * Read the current value of the 8254 timer counter 0.
 217          */
 218         outb(TIMER_MODE, TIMER_SEL0 | TIMER_LATCH);
 219         low = inb(TIMER_CNTR0);
 220         high = inb(TIMER_CNTR0);
 221         count = ((high << 8) | low) << CPUTIME_CLOCK_I8254_SHIFT;
 222
 223         /*
 224          * The timer counts down from TIMER_CNTR0_MAX to 0 and then resets.
 225          * While profiling is enabled, this routine is called at least twice
 226          * per timer reset (for mcounting and mexitcounting hardclock()),
 227          * so at most one reset has occurred since the last call, and one
 228          * has occurred iff the current count is larger than the previous
 229          * count.  This allows counter underflow to be detected faster
 230          * than in microtime().
 231          */
 232         delta = prev_count - count;
 233         prev_count = count;
 234         if ((int) delta <= 0)
 235                 return (delta + (timer0_max_count << CPUTIME_CLOCK_I8254_SHIFT));
 236         return (delta);
 237 }
 238
 239 static int
 240 sysctl_machdep_cputime_clock(SYSCTL_HANDLER_ARGS)
 241 {
 242         int clock;
 243         int error;
 244 #if defined(PERFMON) && defined(I586_PMC_GUPROF)
 245         int event;
 246         struct pmc pmc;
 247 #endif
 248
 249         clock = cputime_clock;
 250 #if defined(PERFMON) && defined(I586_PMC_GUPROF)
 251         if (clock == CPUTIME_CLOCK_I586_PMC) {
 252                 pmc.pmc_val = cputime_clock_pmc_conf;
 253                 clock += pmc.pmc_event;
 254         }
 255 #endif
 256         error = sysctl_handle_opaque(oidp, &clock, sizeof clock, req);
 257         if (error == 0 && req->newptr != NULL) {
 258 #if defined(PERFMON) && defined(I586_PMC_GUPROF)
 259                 if (clock >= CPUTIME_CLOCK_I586_PMC) {
 260                         event = clock - CPUTIME_CLOCK_I586_PMC;
 261                         if (event >= 256)
 262                                 return (EINVAL);
 263                         pmc.pmc_num = 0;
 264                         pmc.pmc_event = event;
 265                         pmc.pmc_unit = 0;
 266                         pmc.pmc_flags = PMCF_E | PMCF_OS | PMCF_USR;
 267                         pmc.pmc_mask = 0;
 268                         cputime_clock_pmc_conf = pmc.pmc_val;
 269                         cputime_clock = CPUTIME_CLOCK_I586_PMC;
 270                 } else
 271 #endif
 272                 {
 273                         if (clock < 0 || clock >= CPUTIME_CLOCK_I586_PMC)
 274                                 return (EINVAL);
 275                         cputime_clock = clock;
 276                 }
 277         }
 278         return (error);
 279 }
 280
 281 SYSCTL_PROC(_machdep, OID_AUTO, cputime_clock, CTLTYPE_INT | CTLFLAG_RW,
 282             0, sizeof(u_int), sysctl_machdep_cputime_clock, "I", "");
 283
 284 /*
 285  * The start and stop routines need not be here since we turn off profiling
 286  * before calling them.  They are here for convenience.
 287  */
 288
 289 void
 290 startguprof(gp)
 291         struct gmonparam *gp;
 292 {
 293         if (cputime_clock == CPUTIME_CLOCK_UNINITIALIZED) {
 294                 cputime_clock = CPUTIME_CLOCK_I8254;
 295 #if defined(I586_CPU) || defined(I686_CPU)
 296                 if (tsc_freq != 0 && !tsc_is_broken && mp_ncpus == 1)
 297                         cputime_clock = CPUTIME_CLOCK_TSC;
 298 #endif
 299         }
 300         gp->profrate = timer_freq << CPUTIME_CLOCK_I8254_SHIFT;
 301 #if defined(I586_CPU) || defined(I686_CPU)
 302         if (cputime_clock == CPUTIME_CLOCK_TSC) {
 303                 gp->profrate = tsc_freq >> 1;
 304                 cputime_prof_active = 1;
 305         }
 306 #if defined(PERFMON) && defined(I586_PMC_GUPROF)
 307         else if (cputime_clock == CPUTIME_CLOCK_I586_PMC) {
 308                 if (perfmon_avail() &&
 309                     perfmon_setup(0, cputime_clock_pmc_conf) == 0) {
 310                         if (perfmon_start(0) != 0)
 311                                 perfmon_fini(0);
 312                         else {
 313                                 /* XXX 1 event == 1 us. */
 314                                 gp->profrate = 1000000;
 315
 316                                 saved_gmp = *gp;
 317
 318                                 /* Zap overheads.  They are invalid. */
 319                                 gp->cputime_overhead = 0;
 320                                 gp->mcount_overhead = 0;
 321                                 gp->mcount_post_overhead = 0;
 322                                 gp->mcount_pre_overhead = 0;
 323                                 gp->mexitcount_overhead = 0;
 324                                 gp->mexitcount_post_overhead = 0;
 325                                 gp->mexitcount_pre_overhead = 0;
 326
 327                                 cputime_clock_pmc_init = TRUE;
 328                         }
 329                 }
 330         }
 331 #endif /* PERFMON && I586_PMC_GUPROF */
 332 #endif /* I586_CPU || I686_CPU */
 333         cputime_bias = 0;
 334         cputime();
 335 }
 336
 337 void
 338 stopguprof(gp)
 339         struct gmonparam *gp;
 340 {
 341 #if defined(PERFMON) && defined(I586_PMC_GUPROF)
 342         if (cputime_clock_pmc_init) {
 343                 *gp = saved_gmp;
 344                 perfmon_fini(0);
 345                 cputime_clock_pmc_init = FALSE;
 346         }
 347 #endif
 348 #if defined(I586_CPU) || defined(I686_CPU)
 349         if (cputime_clock == CPUTIME_CLOCK_TSC)
 350                 cputime_prof_active = 0;
 351 #endif
 352 }
 353
 354 #if defined(I586_CPU) || defined(I686_CPU)
 355 /* If the cpu frequency changed while profiling, report a warning. */
 356 static void
 357 tsc_freq_changed(void *arg, const struct cf_level *level, int status)
 358 {
 359
 360         /*
 361          * If there was an error during the transition or
 362          * TSC is P-state invariant, don't do anything.
 363          */
 364         if (status != 0 || tsc_is_invariant)
 365                 return;
 366         if (cputime_prof_active && cputime_clock == CPUTIME_CLOCK_TSC)
 367                 printf("warning: cpu freq changed while profiling active\n");
 368 }
 369
 370 EVENTHANDLER_DEFINE(cpufreq_post_change, tsc_freq_changed, NULL,
 371     EVENTHANDLER_PRI_ANY);
 372 #endif /* I586_CPU || I686_CPU */
 373
 374 #endif /* GUPROF */