sys/amd64/amd64/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 #if 0
  32 #include "opt_i586_guprof.h"
  33 #include "opt_perfmon.h"
  34 #endif
  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 #if 0
  48 #include <machine/perfmon.h>
  49 #endif
  50 #include <machine/timerreg.h>
  51
  52 #define CPUTIME_CLOCK_UNINITIALIZED     0
  53 #define CPUTIME_CLOCK_I8254             1
  54 #define CPUTIME_CLOCK_TSC               2
  55 #define CPUTIME_CLOCK_I586_PMC          3
  56 #define CPUTIME_CLOCK_I8254_SHIFT       7
  57
  58 int     cputime_bias = 1;       /* initialize for locality of reference */
  59
  60 static int      cputime_clock = CPUTIME_CLOCK_UNINITIALIZED;
  61 #if defined(PERFMON) && defined(I586_PMC_GUPROF)
  62 static u_int    cputime_clock_pmc_conf = I586_PMC_GUPROF;
  63 static int      cputime_clock_pmc_init;
  64 static struct gmonparam saved_gmp;
  65 #endif
  66 static int      cputime_prof_active;
  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         pushq   %rax                                            \n\
  90         pushq   %rdx                                            \n\
  91         pushq   %rcx                                            \n\
  92         pushq   %rsi                                            \n\
  93         pushq   %rdi                                            \n\
  94         pushq   %r8                                             \n\
  95         pushq   %r9                                             \n\
  96         movq    7*8+8(%rsp),%rdi                                \n\
  97         jmp     .got_frompc                                     \n\
  98                                                                 \n\
  99         .p2align 4,0x90                                         \n\
 100         .globl  .mcount                                         \n\
 101 .mcount:                                                        \n\
 102         cmpl    $GMON_PROF_OFF,_gmonparam+GM_STATE              \n\
 103         je      .mcount_exit                                    \n\
 104         #                                                       \n\
 105         # The caller's stack frame has already been built, so   \n\
 106         # %rbp is the caller's frame pointer.  The caller's     \n\
 107         # raddr is in the caller's frame following the caller's \n\
 108         # caller's frame pointer.                               \n\
 109         #                                                       \n\
 110         pushq   %rax                                            \n\
 111         pushq   %rdx                                            \n\
 112         pushq   %rcx                                            \n\
 113         pushq   %rsi                                            \n\
 114         pushq   %rdi                                            \n\
 115         pushq   %r8                                             \n\
 116         pushq   %r9                                             \n\
 117         movq    8(%rbp),%rdi                                    \n\
 118 .got_frompc:                                                    \n\
 119         #                                                       \n\
 120         # Our raddr is the caller's pc.                         \n\
 121         #                                                       \n\
 122         movq    7*8(%rsp),%rsi                                  \n\
 123                                                                 \n\
 124         pushfq                                                  \n\
 125         cli                                                     \n\
 126         call    mcount                                          \n\
 127         popfq                                                   \n\
 128         popq    %r9                                             \n\
 129         popq    %r8                                             \n\
 130         popq    %rdi                                            \n\
 131         popq    %rsi                                            \n\
 132         popq    %rcx                                            \n\
 133         popq    %rdx                                            \n\
 134         popq    %rax                                            \n\
 135 .mcount_exit:                                                   \n\
 136         ret     $0                                              \n\
 137 ");
 138 #else /* !__GNUCLIKE_ASM */
 139 #error "this file needs to be ported to your compiler"
 140 #endif /* __GNUCLIKE_ASM */
 141
 142 #ifdef GUPROF
 143 /*
 144  * [.]mexitcount saves the return register(s), loads selfpc and calls
 145  * mexitcount(selfpc) to do the work.  Someday it should be in a machine
 146  * dependent file together with cputime(), __mcount and [.]mcount.  cputime()
 147  * can't just be put in machdep.c because it has to be compiled without -pg.
 148  */
 149 #ifdef __GNUCLIKE_ASM
 150 __asm("                                                         \n\
 151         .text                                                   \n\
 152 #                                                               \n\
 153 # Dummy label to be seen when gprof -u hides [.]mexitcount.     \n\
 154 #                                                               \n\
 155         .p2align 4,0x90                                         \n\
 156         .globl  __mexitcount                                    \n\
 157         .type   __mexitcount,@function                          \n\
 158 __mexitcount:                                                   \n\
 159         nop                                                     \n\
 160                                                                 \n\
 161 GMON_PROF_HIRES =       4                                       \n\
 162                                                                 \n\
 163         .p2align 4,0x90                                         \n\
 164         .globl  .mexitcount                                     \n\
 165 .mexitcount:                                                    \n\
 166         cmpl    $GMON_PROF_HIRES,_gmonparam+GM_STATE            \n\
 167         jne     .mexitcount_exit                                \n\
 168         pushq   %rax                                            \n\
 169         pushq   %rdx                                            \n\
 170         pushq   %rcx                                            \n\
 171         pushq   %rsi                                            \n\
 172         pushq   %rdi                                            \n\
 173         pushq   %r8                                             \n\
 174         pushq   %r9                                             \n\
 175         movq    7*8(%rsp),%rdi                                  \n\
 176         pushfq                                                  \n\
 177         cli                                                     \n\
 178         call    mexitcount                                      \n\
 179         popfq                                                   \n\
 180         popq    %r9                                             \n\
 181         popq    %r8                                             \n\
 182         popq    %rdi                                            \n\
 183         popq    %rsi                                            \n\
 184         popq    %rcx                                            \n\
 185         popq    %rdx                                            \n\
 186         popq    %rax                                            \n\
 187 .mexitcount_exit:                                               \n\
 188         ret     $0                                              \n\
 189 ");
 190 #endif /* __GNUCLIKE_ASM */
 191
 192 /*
 193  * Return the time elapsed since the last call.  The units are machine-
 194  * dependent.
 195  */
 196 int
 197 cputime()
 198 {
 199         u_int count;
 200         int delta;
 201 #if defined(PERFMON) && defined(I586_PMC_GUPROF) && !defined(SMP)
 202         u_quad_t event_count;
 203 #endif
 204         u_char high, low;
 205         static u_int prev_count;
 206
 207         if (cputime_clock == CPUTIME_CLOCK_TSC) {
 208                 /*
 209                  * Scale the TSC a little to make cputime()'s frequency
 210                  * fit in an int, assuming that the TSC frequency fits
 211                  * in a u_int.  Use a fixed scale since dynamic scaling
 212                  * would be slower and we can't really use the low bit
 213                  * of precision.
 214                  */
 215                 count = (u_int)rdtsc() & ~1u;
 216                 delta = (int)(count - prev_count) >> 1;
 217                 prev_count = count;
 218                 return (delta);
 219         }
 220 #if defined(PERFMON) && defined(I586_PMC_GUPROF) && !defined(SMP)
 221         if (cputime_clock == CPUTIME_CLOCK_I586_PMC) {
 222                 /*
 223                  * XXX permon_read() should be inlined so that the
 224                  * perfmon module doesn't need to be compiled with
 225                  * profiling disabled and so that it is fast.
 226                  */
 227                 perfmon_read(0, &event_count);
 228
 229                 count = (u_int)event_count;
 230                 delta = (int)(count - prev_count);
 231                 prev_count = count;
 232                 return (delta);
 233         }
 234 #endif /* PERFMON && I586_PMC_GUPROF && !SMP */
 235
 236         /*
 237          * Read the current value of the 8254 timer counter 0.
 238          */
 239         outb(TIMER_MODE, TIMER_SEL0 | TIMER_LATCH);
 240         low = inb(TIMER_CNTR0);
 241         high = inb(TIMER_CNTR0);
 242         count = ((high << 8) | low) << CPUTIME_CLOCK_I8254_SHIFT;
 243
 244         /*
 245          * The timer counts down from TIMER_CNTR0_MAX to 0 and then resets.
 246          * While profiling is enabled, this routine is called at least twice
 247          * per timer reset (for mcounting and mexitcounting hardclock()),
 248          * so at most one reset has occurred since the last call, and one
 249          * has occurred iff the current count is larger than the previous
 250          * count.  This allows counter underflow to be detected faster
 251          * than in microtime().
 252          */
 253         delta = prev_count - count;
 254         prev_count = count;
 255         if ((int) delta <= 0)
 256                 return (delta + (i8254_max_count << CPUTIME_CLOCK_I8254_SHIFT));
 257         return (delta);
 258 }
 259
 260 static int
 261 sysctl_machdep_cputime_clock(SYSCTL_HANDLER_ARGS)
 262 {
 263         int clock;
 264         int error;
 265 #if defined(PERFMON) && defined(I586_PMC_GUPROF)
 266         int event;
 267         struct pmc pmc;
 268 #endif
 269
 270         clock = cputime_clock;
 271 #if defined(PERFMON) && defined(I586_PMC_GUPROF)
 272         if (clock == CPUTIME_CLOCK_I586_PMC) {
 273                 pmc.pmc_val = cputime_clock_pmc_conf;
 274                 clock += pmc.pmc_event;
 275         }
 276 #endif
 277         error = sysctl_handle_opaque(oidp, &clock, sizeof clock, req);
 278         if (error == 0 && req->newptr != NULL) {
 279 #if defined(PERFMON) && defined(I586_PMC_GUPROF)
 280                 if (clock >= CPUTIME_CLOCK_I586_PMC) {
 281                         event = clock - CPUTIME_CLOCK_I586_PMC;
 282                         if (event >= 256)
 283                                 return (EINVAL);
 284                         pmc.pmc_num = 0;
 285                         pmc.pmc_event = event;
 286                         pmc.pmc_unit = 0;
 287                         pmc.pmc_flags = PMCF_E | PMCF_OS | PMCF_USR;
 288                         pmc.pmc_mask = 0;
 289                         cputime_clock_pmc_conf = pmc.pmc_val;
 290                         cputime_clock = CPUTIME_CLOCK_I586_PMC;
 291                 } else
 292 #endif
 293                 {
 294                         if (clock < 0 || clock >= CPUTIME_CLOCK_I586_PMC)
 295                                 return (EINVAL);
 296                         cputime_clock = clock;
 297                 }
 298         }
 299         return (error);
 300 }
 301
 302 SYSCTL_PROC(_machdep, OID_AUTO, cputime_clock, CTLTYPE_INT | CTLFLAG_RW,
 303             0, sizeof(u_int), sysctl_machdep_cputime_clock, "I", "");
 304
 305 /*
 306  * The start and stop routines need not be here since we turn off profiling
 307  * before calling them.  They are here for convenience.
 308  */
 309
 310 void
 311 startguprof(gp)
 312         struct gmonparam *gp;
 313 {
 314         uint64_t freq;
 315
 316         freq = atomic_load_acq_64(&tsc_freq);
 317         if (cputime_clock == CPUTIME_CLOCK_UNINITIALIZED) {
 318                 if (freq != 0 && mp_ncpus == 1)
 319                         cputime_clock = CPUTIME_CLOCK_TSC;
 320                 else
 321                         cputime_clock = CPUTIME_CLOCK_I8254;
 322         }
 323         if (cputime_clock == CPUTIME_CLOCK_TSC) {
 324                 gp->profrate = freq >> 1;
 325                 cputime_prof_active = 1;
 326         } else
 327                 gp->profrate = i8254_freq << CPUTIME_CLOCK_I8254_SHIFT;
 328 #if defined(PERFMON) && defined(I586_PMC_GUPROF)
 329         if (cputime_clock == CPUTIME_CLOCK_I586_PMC) {
 330                 if (perfmon_avail() &&
 331                     perfmon_setup(0, cputime_clock_pmc_conf) == 0) {
 332                         if (perfmon_start(0) != 0)
 333                                 perfmon_fini(0);
 334                         else {
 335                                 /* XXX 1 event == 1 us. */
 336                                 gp->profrate = 1000000;
 337
 338                                 saved_gmp = *gp;
 339
 340                                 /* Zap overheads.  They are invalid. */
 341                                 gp->cputime_overhead = 0;
 342                                 gp->mcount_overhead = 0;
 343                                 gp->mcount_post_overhead = 0;
 344                                 gp->mcount_pre_overhead = 0;
 345                                 gp->mexitcount_overhead = 0;
 346                                 gp->mexitcount_post_overhead = 0;
 347                                 gp->mexitcount_pre_overhead = 0;
 348
 349                                 cputime_clock_pmc_init = TRUE;
 350                         }
 351                 }
 352         }
 353 #endif /* PERFMON && I586_PMC_GUPROF */
 354         cputime_bias = 0;
 355         cputime();
 356 }
 357
 358 void
 359 stopguprof(gp)
 360         struct gmonparam *gp;
 361 {
 362 #if defined(PERFMON) && defined(I586_PMC_GUPROF)
 363         if (cputime_clock_pmc_init) {
 364                 *gp = saved_gmp;
 365                 perfmon_fini(0);
 366                 cputime_clock_pmc_init = FALSE;
 367         }
 368 #endif
 369         if (cputime_clock == CPUTIME_CLOCK_TSC)
 370                 cputime_prof_active = 0;
 371 }
 372
 373 /* If the cpu frequency changed while profiling, report a warning. */
 374 static void
 375 tsc_freq_changed(void *arg, const struct cf_level *level, int status)
 376 {
 377
 378         /*
 379          * If there was an error during the transition or
 380          * TSC is P-state invariant, don't do anything.
 381          */
 382         if (status != 0 || tsc_is_invariant)
 383                 return;
 384         if (cputime_prof_active && cputime_clock == CPUTIME_CLOCK_TSC)
 385                 printf("warning: cpu freq changed while profiling active\n");
 386 }
 387
 388 EVENTHANDLER_DEFINE(cpufreq_post_change, tsc_freq_changed, NULL,
 389     EVENTHANDLER_PRI_ANY);
 390
 391 #endif /* GUPROF */