sys/i386/i386/prof_machdep.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
   3  *
   4  * Copyright (c) 1996 Bruce D. Evans.
   5  * All rights reserved.
   6  *
   7  * Redistribution and use in source and binary forms, with or without
   8  * modification, are permitted provided that the following conditions
   9  * are met:
  10  * 1. Redistributions of source code must retain the above copyright
  11  *    notice, this list of conditions and the following disclaimer.
  12  * 2. Redistributions in binary form must reproduce the above copyright
  13  *    notice, this list of conditions and the following disclaimer in the
  14  *    documentation and/or other materials provided with the distribution.
  15  *
  16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
  17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  19  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
  20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  26  * SUCH DAMAGE.
  27  */
  28
  29 #include <sys/cdefs.h>
  30 __FBSDID("$FreeBSD$");
  31
  32 #ifdef GUPROF
  33 #include "opt_i586_guprof.h"
  34 #include "opt_perfmon.h"
  35
  36 #include <sys/param.h>
  37 #include <sys/systm.h>
  38 #include <sys/bus.h>
  39 #include <sys/cpu.h>
  40 #include <sys/eventhandler.h>
  41 #include <sys/gmon.h>
  42 #include <sys/kernel.h>
  43 #include <sys/smp.h>
  44 #include <sys/sysctl.h>
  45
  46 #include <machine/clock.h>
  47 #include <machine/perfmon.h>
  48 #include <machine/timerreg.h>
  49
  50 #define CPUTIME_CLOCK_UNINITIALIZED     0
  51 #define CPUTIME_CLOCK_I8254             1
  52 #define CPUTIME_CLOCK_TSC               2
  53 #define CPUTIME_CLOCK_I586_PMC          3
  54 #define CPUTIME_CLOCK_I8254_SHIFT       7
  55
  56 int     cputime_bias = 1;       /* initialize for locality of reference */
  57
  58 static int      cputime_clock = CPUTIME_CLOCK_UNINITIALIZED;
  59 #if defined(PERFMON) && defined(I586_PMC_GUPROF)
  60 static u_int    cputime_clock_pmc_conf = I586_PMC_GUPROF;
  61 static int      cputime_clock_pmc_init;
  62 static struct gmonparam saved_gmp;
  63 #endif
  64 #if defined(I586_CPU) || defined(I686_CPU)
  65 static int      cputime_prof_active;
  66 #endif
  67 #endif /* GUPROF */
  68
  69 #ifdef __GNUCLIKE_ASM
  70 __asm("                                                         \n\
  71 GM_STATE        =       0                                       \n\
  72 GMON_PROF_OFF   =       3                                       \n\
  73                                                                 \n\
  74         .text                                                   \n\
  75         .p2align 4,0x90                                         \n\
  76         .globl  __mcount                                        \n\
  77         .type   __mcount,@function                              \n\
  78 __mcount:                                                       \n\
  79         #                                                       \n\
  80         # Check that we are profiling.  Do it early for speed.  \n\
  81         #                                                       \n\
  82         cmpl    $GMON_PROF_OFF,_gmonparam+GM_STATE              \n\
  83         je      .mcount_exit                                    \n\
  84         #                                                       \n\
  85         # __mcount is the same as [.]mcount except the caller   \n\
  86         # hasn't changed the stack except to call here, so the  \n\
  87         # caller's raddr is above our raddr.                    \n\
  88         #                                                       \n\
  89         movl    4(%esp),%edx                                    \n\
  90         jmp     .got_frompc                                     \n\
  91                                                                 \n\
  92         .p2align 4,0x90                                         \n\
  93         .globl  .mcount                                         \n\
  94 .mcount:                                                        \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     $0                                              \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         cmpl    $GMON_PROF_HIRES,_gmonparam+GM_STATE            \n\
 149         jne     .mexitcount_exit                                \n\
 150         pushl   %edx                                            \n\
 151         pushl   %eax                                            \n\
 152         movl    8(%esp),%eax                                    \n\
 153         pushfl                                                  \n\
 154         pushl   %eax                                            \n\
 155         cli                                                     \n\
 156         call    mexitcount                                      \n\
 157         addl    $4,%esp                                         \n\
 158         popfl                                                   \n\
 159         popl    %eax                                            \n\
 160         popl    %edx                                            \n\
 161 .mexitcount_exit:                                               \n\
 162         ret     $0                                              \n\
 163 ");
 164 #endif /* __GNUCLIKE_ASM */
 165
 166 /*
 167  * Return the time elapsed since the last call.  The units are machine-
 168  * dependent.
 169  */
 170 int
 171 cputime()
 172 {
 173         u_int count;
 174         int delta;
 175 #if (defined(I586_CPU) || defined(I686_CPU)) && \
 176     defined(PERFMON) && defined(I586_PMC_GUPROF) && !defined(SMP)
 177         u_quad_t event_count;
 178 #endif
 179         u_char high, low;
 180         static u_int prev_count;
 181
 182 #if defined(I586_CPU) || defined(I686_CPU)
 183         if (cputime_clock == CPUTIME_CLOCK_TSC) {
 184                 /*
 185                  * Scale the TSC a little to make cputime()'s frequency
 186                  * fit in an int, assuming that the TSC frequency fits
 187                  * in a u_int.  Use a fixed scale since dynamic scaling
 188                  * would be slower and we can't really use the low bit
 189                  * of precision.
 190                  */
 191                 count = (u_int)rdtsc() & ~1u;
 192                 delta = (int)(count - prev_count) >> 1;
 193                 prev_count = count;
 194                 return (delta);
 195         }
 196 #if defined(PERFMON) && defined(I586_PMC_GUPROF) && !defined(SMP)
 197         if (cputime_clock == CPUTIME_CLOCK_I586_PMC) {
 198                 /*
 199                  * XXX permon_read() should be inlined so that the
 200                  * perfmon module doesn't need to be compiled with
 201                  * profiling disabled and so that it is fast.
 202                  */
 203                 perfmon_read(0, &event_count);
 204
 205                 count = (u_int)event_count;
 206                 delta = (int)(count - prev_count);
 207                 prev_count = count;
 208                 return (delta);
 209         }
 210 #endif /* PERFMON && I586_PMC_GUPROF && !SMP */
 211 #endif /* I586_CPU || I686_CPU */
 212
 213         /*
 214          * Read the current value of the 8254 timer counter 0.
 215          */
 216         outb(TIMER_MODE, TIMER_SEL0 | TIMER_LATCH);
 217         low = inb(TIMER_CNTR0);
 218         high = inb(TIMER_CNTR0);
 219         count = ((high << 8) | low) << CPUTIME_CLOCK_I8254_SHIFT;
 220
 221         /*
 222          * The timer counts down from TIMER_CNTR0_MAX to 0 and then resets.
 223          * While profiling is enabled, this routine is called at least twice
 224          * per timer reset (for mcounting and mexitcounting hardclock()),
 225          * so at most one reset has occurred since the last call, and one
 226          * has occurred iff the current count is larger than the previous
 227          * count.  This allows counter underflow to be detected faster
 228          * than in microtime().
 229          */
 230         delta = prev_count - count;
 231         prev_count = count;
 232         if ((int) delta <= 0)
 233                 return (delta + (i8254_max_count << CPUTIME_CLOCK_I8254_SHIFT));
 234         return (delta);
 235 }
 236
 237 static int
 238 sysctl_machdep_cputime_clock(SYSCTL_HANDLER_ARGS)
 239 {
 240         int clock;
 241         int error;
 242 #if defined(PERFMON) && defined(I586_PMC_GUPROF)
 243         int event;
 244         struct pmc pmc;
 245 #endif
 246
 247         clock = cputime_clock;
 248 #if defined(PERFMON) && defined(I586_PMC_GUPROF)
 249         if (clock == CPUTIME_CLOCK_I586_PMC) {
 250                 pmc.pmc_val = cputime_clock_pmc_conf;
 251                 clock += pmc.pmc_event;
 252         }
 253 #endif
 254         error = sysctl_handle_opaque(oidp, &clock, sizeof clock, req);
 255         if (error == 0 && req->newptr != NULL) {
 256 #if defined(PERFMON) && defined(I586_PMC_GUPROF)
 257                 if (clock >= CPUTIME_CLOCK_I586_PMC) {
 258                         event = clock - CPUTIME_CLOCK_I586_PMC;
 259                         if (event >= 256)
 260                                 return (EINVAL);
 261                         pmc.pmc_num = 0;
 262                         pmc.pmc_event = event;
 263                         pmc.pmc_unit = 0;
 264                         pmc.pmc_flags = PMCF_E | PMCF_OS | PMCF_USR;
 265                         pmc.pmc_mask = 0;
 266                         cputime_clock_pmc_conf = pmc.pmc_val;
 267                         cputime_clock = CPUTIME_CLOCK_I586_PMC;
 268                 } else
 269 #endif
 270                 {
 271                         if (clock < 0 || clock >= CPUTIME_CLOCK_I586_PMC)
 272                                 return (EINVAL);
 273                         cputime_clock = clock;
 274                 }
 275         }
 276         return (error);
 277 }
 278
 279 SYSCTL_PROC(_machdep, OID_AUTO, cputime_clock, CTLTYPE_INT | CTLFLAG_RW,
 280             0, sizeof(u_int), sysctl_machdep_cputime_clock, "I", "");
 281
 282 /*
 283  * The start and stop routines need not be here since we turn off profiling
 284  * before calling them.  They are here for convenience.
 285  */
 286
 287 void
 288 startguprof(gp)
 289         struct gmonparam *gp;
 290 {
 291 #if defined(I586_CPU) || defined(I686_CPU)
 292         uint64_t freq;
 293
 294         freq = atomic_load_acq_64(&tsc_freq);
 295         if (cputime_clock == CPUTIME_CLOCK_UNINITIALIZED) {
 296                 if (freq != 0 && mp_ncpus == 1)
 297                         cputime_clock = CPUTIME_CLOCK_TSC;
 298                 else
 299                         cputime_clock = CPUTIME_CLOCK_I8254;
 300         }
 301         if (cputime_clock == CPUTIME_CLOCK_TSC) {
 302                 gp->profrate = freq >> 1;
 303                 cputime_prof_active = 1;
 304         } else
 305                 gp->profrate = i8254_freq << CPUTIME_CLOCK_I8254_SHIFT;
 306 #if defined(PERFMON) && defined(I586_PMC_GUPROF)
 307         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 #else /* !(I586_CPU || I686_CPU) */
 333         if (cputime_clock == CPUTIME_CLOCK_UNINITIALIZED)
 334                 cputime_clock = CPUTIME_CLOCK_I8254;
 335         gp->profrate = i8254_freq << CPUTIME_CLOCK_I8254_SHIFT;
 336 #endif /* I586_CPU || I686_CPU */
 337         cputime_bias = 0;
 338         cputime();
 339 }
 340
 341 void
 342 stopguprof(gp)
 343         struct gmonparam *gp;
 344 {
 345 #if defined(PERFMON) && defined(I586_PMC_GUPROF)
 346         if (cputime_clock_pmc_init) {
 347                 *gp = saved_gmp;
 348                 perfmon_fini(0);
 349                 cputime_clock_pmc_init = FALSE;
 350         }
 351 #endif
 352 #if defined(I586_CPU) || defined(I686_CPU)
 353         if (cputime_clock == CPUTIME_CLOCK_TSC)
 354                 cputime_prof_active = 0;
 355 #endif
 356 }
 357
 358 #if defined(I586_CPU) || defined(I686_CPU)
 359 /* If the cpu frequency changed while profiling, report a warning. */
 360 static void
 361 tsc_freq_changed(void *arg, const struct cf_level *level, int status)
 362 {
 363
 364         /*
 365          * If there was an error during the transition or
 366          * TSC is P-state invariant, don't do anything.
 367          */
 368         if (status != 0 || tsc_is_invariant)
 369                 return;
 370         if (cputime_prof_active && cputime_clock == CPUTIME_CLOCK_TSC)
 371                 printf("warning: cpu freq changed while profiling active\n");
 372 }
 373
 374 EVENTHANDLER_DEFINE(cpufreq_post_change, tsc_freq_changed, NULL,
 375     EVENTHANDLER_PRI_ANY);
 376 #endif /* I586_CPU || I686_CPU */
 377
 378 #endif /* GUPROF */