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
 121 void    __mcount(void);
 122 void    (*__mcountp)(void) = __mcount;
 123 #else /* !__GNUCLIKE_ASM */
 124 #error "this file needs to be ported to your compiler"
 125 #endif /* __GNUCLIKE_ASM */
 126
 127 #ifdef GUPROF
 128 /*
 129  * [.]mexitcount saves the return register(s), loads selfpc and calls
 130  * mexitcount(selfpc) to do the work.  Someday it should be in a machine
 131  * dependent file together with cputime(), __mcount and [.]mcount.  cputime()
 132  * can't just be put in machdep.c because it has to be compiled without -pg.
 133  */
 134 #ifdef __GNUCLIKE_ASM
 135 __asm("                                                         \n\
 136         .text                                                   \n\
 137 #                                                               \n\
 138 # Dummy label to be seen when gprof -u hides [.]mexitcount.     \n\
 139 #                                                               \n\
 140         .p2align 4,0x90                                         \n\
 141         .globl  __mexitcount                                    \n\
 142         .type   __mexitcount,@function                          \n\
 143 __mexitcount:                                                   \n\
 144         nop                                                     \n\
 145                                                                 \n\
 146 GMON_PROF_HIRES =       4                                       \n\
 147                                                                 \n\
 148         .p2align 4,0x90                                         \n\
 149         .globl  .mexitcount                                     \n\
 150 .mexitcount:                                                    \n\
 151         cmpl    $GMON_PROF_HIRES,_gmonparam+GM_STATE            \n\
 152         jne     .mexitcount_exit                                \n\
 153         pushl   %edx                                            \n\
 154         pushl   %eax                                            \n\
 155         movl    8(%esp),%eax                                    \n\
 156         pushfl                                                  \n\
 157         pushl   %eax                                            \n\
 158         cli                                                     \n\
 159         call    mexitcount                                      \n\
 160         addl    $4,%esp                                         \n\
 161         popfl                                                   \n\
 162         popl    %eax                                            \n\
 163         popl    %edx                                            \n\
 164 .mexitcount_exit:                                               \n\
 165         ret     $0                                              \n\
 166 ");
 167 #endif /* __GNUCLIKE_ASM */
 168
 169 void    __mexitcount(void);
 170 void    (*__mexitcountp)(void) = __mexitcount;
 171
 172 /*
 173  * Return the time elapsed since the last call.  The units are machine-
 174  * dependent.
 175  */
 176 int
 177 cputime()
 178 {
 179         u_int count;
 180         int delta;
 181 #if (defined(I586_CPU) || defined(I686_CPU)) && \
 182     defined(PERFMON) && defined(I586_PMC_GUPROF) && !defined(SMP)
 183         u_quad_t event_count;
 184 #endif
 185         u_char high, low;
 186         static u_int prev_count;
 187
 188 #if defined(I586_CPU) || defined(I686_CPU)
 189         if (cputime_clock == CPUTIME_CLOCK_TSC) {
 190                 /*
 191                  * Scale the TSC a little to make cputime()'s frequency
 192                  * fit in an int, assuming that the TSC frequency fits
 193                  * in a u_int.  Use a fixed scale since dynamic scaling
 194                  * would be slower and we can't really use the low bit
 195                  * of precision.
 196                  */
 197                 count = (u_int)rdtsc() & ~1u;
 198                 delta = (int)(count - prev_count) >> 1;
 199                 prev_count = count;
 200                 return (delta);
 201         }
 202 #if defined(PERFMON) && defined(I586_PMC_GUPROF) && !defined(SMP)
 203         if (cputime_clock == CPUTIME_CLOCK_I586_PMC) {
 204                 /*
 205                  * XXX permon_read() should be inlined so that the
 206                  * perfmon module doesn't need to be compiled with
 207                  * profiling disabled and so that it is fast.
 208                  */
 209                 perfmon_read(0, &event_count);
 210
 211                 count = (u_int)event_count;
 212                 delta = (int)(count - prev_count);
 213                 prev_count = count;
 214                 return (delta);
 215         }
 216 #endif /* PERFMON && I586_PMC_GUPROF && !SMP */
 217 #endif /* I586_CPU || I686_CPU */
 218
 219         /*
 220          * Read the current value of the 8254 timer counter 0.
 221          */
 222         outb(TIMER_MODE, TIMER_SEL0 | TIMER_LATCH);
 223         low = inb(TIMER_CNTR0);
 224         high = inb(TIMER_CNTR0);
 225         count = ((high << 8) | low) << CPUTIME_CLOCK_I8254_SHIFT;
 226
 227         /*
 228          * The timer counts down from TIMER_CNTR0_MAX to 0 and then resets.
 229          * While profiling is enabled, this routine is called at least twice
 230          * per timer reset (for mcounting and mexitcounting hardclock()),
 231          * so at most one reset has occurred since the last call, and one
 232          * has occurred iff the current count is larger than the previous
 233          * count.  This allows counter underflow to be detected faster
 234          * than in microtime().
 235          */
 236         delta = prev_count - count;
 237         prev_count = count;
 238         if ((int) delta <= 0)
 239                 return (delta + (i8254_max_count << CPUTIME_CLOCK_I8254_SHIFT));
 240         return (delta);
 241 }
 242
 243 static int
 244 sysctl_machdep_cputime_clock(SYSCTL_HANDLER_ARGS)
 245 {
 246         int clock;
 247         int error;
 248 #if defined(PERFMON) && defined(I586_PMC_GUPROF)
 249         int event;
 250         struct pmc pmc;
 251 #endif
 252
 253         clock = cputime_clock;
 254 #if defined(PERFMON) && defined(I586_PMC_GUPROF)
 255         if (clock == CPUTIME_CLOCK_I586_PMC) {
 256                 pmc.pmc_val = cputime_clock_pmc_conf;
 257                 clock += pmc.pmc_event;
 258         }
 259 #endif
 260         error = sysctl_handle_opaque(oidp, &clock, sizeof clock, req);
 261         if (error == 0 && req->newptr != NULL) {
 262 #if defined(PERFMON) && defined(I586_PMC_GUPROF)
 263                 if (clock >= CPUTIME_CLOCK_I586_PMC) {
 264                         event = clock - CPUTIME_CLOCK_I586_PMC;
 265                         if (event >= 256)
 266                                 return (EINVAL);
 267                         pmc.pmc_num = 0;
 268                         pmc.pmc_event = event;
 269                         pmc.pmc_unit = 0;
 270                         pmc.pmc_flags = PMCF_E | PMCF_OS | PMCF_USR;
 271                         pmc.pmc_mask = 0;
 272                         cputime_clock_pmc_conf = pmc.pmc_val;
 273                         cputime_clock = CPUTIME_CLOCK_I586_PMC;
 274                 } else
 275 #endif
 276                 {
 277                         if (clock < 0 || clock >= CPUTIME_CLOCK_I586_PMC)
 278                                 return (EINVAL);
 279                         cputime_clock = clock;
 280                 }
 281         }
 282         return (error);
 283 }
 284
 285 SYSCTL_PROC(_machdep, OID_AUTO, cputime_clock, CTLTYPE_INT | CTLFLAG_RW,
 286             0, sizeof(u_int), sysctl_machdep_cputime_clock, "I", "");
 287
 288 /*
 289  * The start and stop routines need not be here since we turn off profiling
 290  * before calling them.  They are here for convenience.
 291  */
 292
 293 void
 294 startguprof(gp)
 295         struct gmonparam *gp;
 296 {
 297 #if defined(I586_CPU) || defined(I686_CPU)
 298         uint64_t freq;
 299
 300         freq = atomic_load_acq_64(&tsc_freq);
 301         if (cputime_clock == CPUTIME_CLOCK_UNINITIALIZED) {
 302                 if (freq != 0 && mp_ncpus == 1)
 303                         cputime_clock = CPUTIME_CLOCK_TSC;
 304                 else
 305                         cputime_clock = CPUTIME_CLOCK_I8254;
 306         }
 307         if (cputime_clock == CPUTIME_CLOCK_TSC) {
 308                 gp->profrate = freq >> 1;
 309                 cputime_prof_active = 1;
 310         } else
 311                 gp->profrate = i8254_freq << CPUTIME_CLOCK_I8254_SHIFT;
 312 #if defined(PERFMON) && defined(I586_PMC_GUPROF)
 313         if (cputime_clock == CPUTIME_CLOCK_I586_PMC) {
 314                 if (perfmon_avail() &&
 315                     perfmon_setup(0, cputime_clock_pmc_conf) == 0) {
 316                         if (perfmon_start(0) != 0)
 317                                 perfmon_fini(0);
 318                         else {
 319                                 /* XXX 1 event == 1 us. */
 320                                 gp->profrate = 1000000;
 321
 322                                 saved_gmp = *gp;
 323
 324                                 /* Zap overheads.  They are invalid. */
 325                                 gp->cputime_overhead = 0;
 326                                 gp->mcount_overhead = 0;
 327                                 gp->mcount_post_overhead = 0;
 328                                 gp->mcount_pre_overhead = 0;
 329                                 gp->mexitcount_overhead = 0;
 330                                 gp->mexitcount_post_overhead = 0;
 331                                 gp->mexitcount_pre_overhead = 0;
 332
 333                                 cputime_clock_pmc_init = TRUE;
 334                         }
 335                 }
 336         }
 337 #endif /* PERFMON && I586_PMC_GUPROF */
 338 #else /* !(I586_CPU || I686_CPU) */
 339         if (cputime_clock == CPUTIME_CLOCK_UNINITIALIZED)
 340                 cputime_clock = CPUTIME_CLOCK_I8254;
 341         gp->profrate = i8254_freq << CPUTIME_CLOCK_I8254_SHIFT;
 342 #endif /* I586_CPU || I686_CPU */
 343         cputime_bias = 0;
 344         cputime();
 345 }
 346
 347 void
 348 stopguprof(gp)
 349         struct gmonparam *gp;
 350 {
 351 #if defined(PERFMON) && defined(I586_PMC_GUPROF)
 352         if (cputime_clock_pmc_init) {
 353                 *gp = saved_gmp;
 354                 perfmon_fini(0);
 355                 cputime_clock_pmc_init = FALSE;
 356         }
 357 #endif
 358 #if defined(I586_CPU) || defined(I686_CPU)
 359         if (cputime_clock == CPUTIME_CLOCK_TSC)
 360                 cputime_prof_active = 0;
 361 #endif
 362 }
 363
 364 #if defined(I586_CPU) || defined(I686_CPU)
 365 /* If the cpu frequency changed while profiling, report a warning. */
 366 static void
 367 tsc_freq_changed(void *arg, const struct cf_level *level, int status)
 368 {
 369
 370         /*
 371          * If there was an error during the transition or
 372          * TSC is P-state invariant, don't do anything.
 373          */
 374         if (status != 0 || tsc_is_invariant)
 375                 return;
 376         if (cputime_prof_active && cputime_clock == CPUTIME_CLOCK_TSC)
 377                 printf("warning: cpu freq changed while profiling active\n");
 378 }
 379
 380 EVENTHANDLER_DEFINE(cpufreq_post_change, tsc_freq_changed, NULL,
 381     EVENTHANDLER_PRI_ANY);
 382 #endif /* I586_CPU || I686_CPU */
 383
 384 #endif /* GUPROF */