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         cmpl    $GMON_PROF_OFF,_gmonparam+GM_STATE              \n\
  94         je      .mcount_exit                                    \n\
  95         #                                                       \n\
  96         # The caller's stack frame has already been built, so   \n\
  97         # %ebp is the caller's frame pointer.  The caller's     \n\
  98         # raddr is in the caller's frame following the caller's \n\
  99         # caller's frame pointer.                               \n\
 100         #                                                       \n\
 101         movl    4(%ebp),%edx                                    \n\
 102 .got_frompc:                                                    \n\
 103         #                                                       \n\
 104         # Our raddr is the caller's pc.                         \n\
 105         #                                                       \n\
 106         movl    (%esp),%eax                                     \n\
 107                                                                 \n\
 108         pushfl                                                  \n\
 109         pushl   %eax                                            \n\
 110         pushl   %edx                                            \n\
 111         cli                                                     \n\
 112         call    mcount                                          \n\
 113         addl    $8,%esp                                         \n\
 114         popfl                                                   \n\
 115 .mcount_exit:                                                   \n\
 116         ret     $0                                              \n\
 117 ");
 118 #else /* !__GNUCLIKE_ASM */
 119 #error "this file needs to be ported to your compiler"
 120 #endif /* __GNUCLIKE_ASM */
 121
 122 #ifdef GUPROF
 123 /*
 124  * [.]mexitcount saves the return register(s), loads selfpc and calls
 125  * mexitcount(selfpc) to do the work.  Someday it should be in a machine
 126  * dependent file together with cputime(), __mcount and [.]mcount.  cputime()
 127  * can't just be put in machdep.c because it has to be compiled without -pg.
 128  */
 129 #ifdef __GNUCLIKE_ASM
 130 __asm("                                                         \n\
 131         .text                                                   \n\
 132 #                                                               \n\
 133 # Dummy label to be seen when gprof -u hides [.]mexitcount.     \n\
 134 #                                                               \n\
 135         .p2align 4,0x90                                         \n\
 136         .globl  __mexitcount                                    \n\
 137         .type   __mexitcount,@function                          \n\
 138 __mexitcount:                                                   \n\
 139         nop                                                     \n\
 140                                                                 \n\
 141 GMON_PROF_HIRES =       4                                       \n\
 142                                                                 \n\
 143         .p2align 4,0x90                                         \n\
 144         .globl  .mexitcount                                     \n\
 145 .mexitcount:                                                    \n\
 146         cmpl    $GMON_PROF_HIRES,_gmonparam+GM_STATE            \n\
 147         jne     .mexitcount_exit                                \n\
 148         pushl   %edx                                            \n\
 149         pushl   %eax                                            \n\
 150         movl    8(%esp),%eax                                    \n\
 151         pushfl                                                  \n\
 152         pushl   %eax                                            \n\
 153         cli                                                     \n\
 154         call    mexitcount                                      \n\
 155         addl    $4,%esp                                         \n\
 156         popfl                                                   \n\
 157         popl    %eax                                            \n\
 158         popl    %edx                                            \n\
 159 .mexitcount_exit:                                               \n\
 160         ret     $0                                              \n\
 161 ");
 162 #endif /* __GNUCLIKE_ASM */
 163
 164 /*
 165  * Return the time elapsed since the last call.  The units are machine-
 166  * dependent.
 167  */
 168 int
 169 cputime()
 170 {
 171         u_int count;
 172         int delta;
 173 #if (defined(I586_CPU) || defined(I686_CPU)) && \
 174     defined(PERFMON) && defined(I586_PMC_GUPROF) && !defined(SMP)
 175         u_quad_t event_count;
 176 #endif
 177         u_char high, low;
 178         static u_int prev_count;
 179
 180 #if defined(I586_CPU) || defined(I686_CPU)
 181         if (cputime_clock == CPUTIME_CLOCK_TSC) {
 182                 /*
 183                  * Scale the TSC a little to make cputime()'s frequency
 184                  * fit in an int, assuming that the TSC frequency fits
 185                  * in a u_int.  Use a fixed scale since dynamic scaling
 186                  * would be slower and we can't really use the low bit
 187                  * of precision.
 188                  */
 189                 count = (u_int)rdtsc() & ~1u;
 190                 delta = (int)(count - prev_count) >> 1;
 191                 prev_count = count;
 192                 return (delta);
 193         }
 194 #if defined(PERFMON) && defined(I586_PMC_GUPROF) && !defined(SMP)
 195         if (cputime_clock == CPUTIME_CLOCK_I586_PMC) {
 196                 /*
 197                  * XXX permon_read() should be inlined so that the
 198                  * perfmon module doesn't need to be compiled with
 199                  * profiling disabled and so that it is fast.
 200                  */
 201                 perfmon_read(0, &event_count);
 202
 203                 count = (u_int)event_count;
 204                 delta = (int)(count - prev_count);
 205                 prev_count = count;
 206                 return (delta);
 207         }
 208 #endif /* PERFMON && I586_PMC_GUPROF && !SMP */
 209 #endif /* I586_CPU || I686_CPU */
 210
 211         /*
 212          * Read the current value of the 8254 timer counter 0.
 213          */
 214         outb(TIMER_MODE, TIMER_SEL0 | TIMER_LATCH);
 215         low = inb(TIMER_CNTR0);
 216         high = inb(TIMER_CNTR0);
 217         count = ((high << 8) | low) << CPUTIME_CLOCK_I8254_SHIFT;
 218
 219         /*
 220          * The timer counts down from TIMER_CNTR0_MAX to 0 and then resets.
 221          * While profiling is enabled, this routine is called at least twice
 222          * per timer reset (for mcounting and mexitcounting hardclock()),
 223          * so at most one reset has occurred since the last call, and one
 224          * has occurred iff the current count is larger than the previous
 225          * count.  This allows counter underflow to be detected faster
 226          * than in microtime().
 227          */
 228         delta = prev_count - count;
 229         prev_count = count;
 230         if ((int) delta <= 0)
 231                 return (delta + (i8254_max_count << CPUTIME_CLOCK_I8254_SHIFT));
 232         return (delta);
 233 }
 234
 235 static int
 236 sysctl_machdep_cputime_clock(SYSCTL_HANDLER_ARGS)
 237 {
 238         int clock;
 239         int error;
 240 #if defined(PERFMON) && defined(I586_PMC_GUPROF)
 241         int event;
 242         struct pmc pmc;
 243 #endif
 244
 245         clock = cputime_clock;
 246 #if defined(PERFMON) && defined(I586_PMC_GUPROF)
 247         if (clock == CPUTIME_CLOCK_I586_PMC) {
 248                 pmc.pmc_val = cputime_clock_pmc_conf;
 249                 clock += pmc.pmc_event;
 250         }
 251 #endif
 252         error = sysctl_handle_opaque(oidp, &clock, sizeof clock, req);
 253         if (error == 0 && req->newptr != NULL) {
 254 #if defined(PERFMON) && defined(I586_PMC_GUPROF)
 255                 if (clock >= CPUTIME_CLOCK_I586_PMC) {
 256                         event = clock - CPUTIME_CLOCK_I586_PMC;
 257                         if (event >= 256)
 258                                 return (EINVAL);
 259                         pmc.pmc_num = 0;
 260                         pmc.pmc_event = event;
 261                         pmc.pmc_unit = 0;
 262                         pmc.pmc_flags = PMCF_E | PMCF_OS | PMCF_USR;
 263                         pmc.pmc_mask = 0;
 264                         cputime_clock_pmc_conf = pmc.pmc_val;
 265                         cputime_clock = CPUTIME_CLOCK_I586_PMC;
 266                 } else
 267 #endif
 268                 {
 269                         if (clock < 0 || clock >= CPUTIME_CLOCK_I586_PMC)
 270                                 return (EINVAL);
 271                         cputime_clock = clock;
 272                 }
 273         }
 274         return (error);
 275 }
 276
 277 SYSCTL_PROC(_machdep, OID_AUTO, cputime_clock, CTLTYPE_INT | CTLFLAG_RW,
 278             0, sizeof(u_int), sysctl_machdep_cputime_clock, "I", "");
 279
 280 /*
 281  * The start and stop routines need not be here since we turn off profiling
 282  * before calling them.  They are here for convenience.
 283  */
 284
 285 void
 286 startguprof(gp)
 287         struct gmonparam *gp;
 288 {
 289 #if defined(I586_CPU) || defined(I686_CPU)
 290         uint64_t freq;
 291
 292         freq = atomic_load_acq_64(&tsc_freq);
 293         if (cputime_clock == CPUTIME_CLOCK_UNINITIALIZED) {
 294                 if (freq != 0 && mp_ncpus == 1)
 295                         cputime_clock = CPUTIME_CLOCK_TSC;
 296                 else
 297                         cputime_clock = CPUTIME_CLOCK_I8254;
 298         }
 299         if (cputime_clock == CPUTIME_CLOCK_TSC) {
 300                 gp->profrate = freq >> 1;
 301                 cputime_prof_active = 1;
 302         } else
 303                 gp->profrate = i8254_freq << CPUTIME_CLOCK_I8254_SHIFT;
 304 #if defined(PERFMON) && defined(I586_PMC_GUPROF)
 305         if (cputime_clock == CPUTIME_CLOCK_I586_PMC) {
 306                 if (perfmon_avail() &&
 307                     perfmon_setup(0, cputime_clock_pmc_conf) == 0) {
 308                         if (perfmon_start(0) != 0)
 309                                 perfmon_fini(0);
 310                         else {
 311                                 /* XXX 1 event == 1 us. */
 312                                 gp->profrate = 1000000;
 313
 314                                 saved_gmp = *gp;
 315
 316                                 /* Zap overheads.  They are invalid. */
 317                                 gp->cputime_overhead = 0;
 318                                 gp->mcount_overhead = 0;
 319                                 gp->mcount_post_overhead = 0;
 320                                 gp->mcount_pre_overhead = 0;
 321                                 gp->mexitcount_overhead = 0;
 322                                 gp->mexitcount_post_overhead = 0;
 323                                 gp->mexitcount_pre_overhead = 0;
 324
 325                                 cputime_clock_pmc_init = TRUE;
 326                         }
 327                 }
 328         }
 329 #endif /* PERFMON && I586_PMC_GUPROF */
 330 #else /* !(I586_CPU || I686_CPU) */
 331         if (cputime_clock == CPUTIME_CLOCK_UNINITIALIZED)
 332                 cputime_clock = CPUTIME_CLOCK_I8254;
 333         gp->profrate = i8254_freq << CPUTIME_CLOCK_I8254_SHIFT;
 334 #endif /* I586_CPU || I686_CPU */
 335         cputime_bias = 0;
 336         cputime();
 337 }
 338
 339 void
 340 stopguprof(gp)
 341         struct gmonparam *gp;
 342 {
 343 #if defined(PERFMON) && defined(I586_PMC_GUPROF)
 344         if (cputime_clock_pmc_init) {
 345                 *gp = saved_gmp;
 346                 perfmon_fini(0);
 347                 cputime_clock_pmc_init = FALSE;
 348         }
 349 #endif
 350 #if defined(I586_CPU) || defined(I686_CPU)
 351         if (cputime_clock == CPUTIME_CLOCK_TSC)
 352                 cputime_prof_active = 0;
 353 #endif
 354 }
 355
 356 #if defined(I586_CPU) || defined(I686_CPU)
 357 /* If the cpu frequency changed while profiling, report a warning. */
 358 static void
 359 tsc_freq_changed(void *arg, const struct cf_level *level, int status)
 360 {
 361
 362         /*
 363          * If there was an error during the transition or
 364          * TSC is P-state invariant, don't do anything.
 365          */
 366         if (status != 0 || tsc_is_invariant)
 367                 return;
 368         if (cputime_prof_active && cputime_clock == CPUTIME_CLOCK_TSC)
 369                 printf("warning: cpu freq changed while profiling active\n");
 370 }
 371
 372 EVENTHANDLER_DEFINE(cpufreq_post_change, tsc_freq_changed, NULL,
 373     EVENTHANDLER_PRI_ANY);
 374 #endif /* I586_CPU || I686_CPU */
 375
 376 #endif /* GUPROF */