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 #if defined(I586_CPU) || defined(I686_CPU)
  67 static int      cputime_prof_active;
  68 #endif
  69 #endif /* GUPROF */
  70
  71 #ifdef __GNUCLIKE_ASM
  72 __asm("                                                         \n\
  73 GM_STATE        =       0                                       \n\
  74 GMON_PROF_OFF   =       3                                       \n\
  75                                                                 \n\
  76         .text                                                   \n\
  77         .p2align 4,0x90                                         \n\
  78         .globl  __mcount                                        \n\
  79         .type   __mcount,@function                              \n\
  80 __mcount:                                                       \n\
  81         #                                                       \n\
  82         # Check that we are profiling.  Do it early for speed.  \n\
  83         #                                                       \n\
  84         cmpl    $GMON_PROF_OFF,_gmonparam+GM_STATE              \n\
  85         je      .mcount_exit                                    \n\
  86         #                                                       \n\
  87         # __mcount is the same as [.]mcount except the caller   \n\
  88         # hasn't changed the stack except to call here, so the  \n\
  89         # caller's raddr is above our raddr.                    \n\
  90         #                                                       \n\
  91         pushq   %rax                                            \n\
  92         pushq   %rdx                                            \n\
  93         pushq   %rcx                                            \n\
  94         pushq   %rsi                                            \n\
  95         pushq   %rdi                                            \n\
  96         pushq   %r8                                             \n\
  97         pushq   %r9                                             \n\
  98         movq    7*8+8(%rsp),%rdi                                \n\
  99         jmp     .got_frompc                                     \n\
 100                                                                 \n\
 101         .p2align 4,0x90                                         \n\
 102         .globl  .mcount                                         \n\
 103 .mcount:                                                        \n\
 104         .globl  __cyg_profile_func_enter                        \n\
 105 __cyg_profile_func_enter:                                       \n\
 106         cmpl    $GMON_PROF_OFF,_gmonparam+GM_STATE              \n\
 107         je      .mcount_exit                                    \n\
 108         #                                                       \n\
 109         # The caller's stack frame has already been built, so   \n\
 110         # %rbp is the caller's frame pointer.  The caller's     \n\
 111         # raddr is in the caller's frame following the caller's \n\
 112         # caller's frame pointer.                               \n\
 113         #                                                       \n\
 114         pushq   %rax                                            \n\
 115         pushq   %rdx                                            \n\
 116         pushq   %rcx                                            \n\
 117         pushq   %rsi                                            \n\
 118         pushq   %rdi                                            \n\
 119         pushq   %r8                                             \n\
 120         pushq   %r9                                             \n\
 121         movq    8(%rbp),%rdi                                    \n\
 122 .got_frompc:                                                    \n\
 123         #                                                       \n\
 124         # Our raddr is the caller's pc.                         \n\
 125         #                                                       \n\
 126         movq    7*8(%rsp),%rsi                                  \n\
 127                                                                 \n\
 128         pushfq                                                  \n\
 129         cli                                                     \n\
 130         call    mcount                                          \n\
 131         popfq                                                   \n\
 132         popq    %r9                                             \n\
 133         popq    %r8                                             \n\
 134         popq    %rdi                                            \n\
 135         popq    %rsi                                            \n\
 136         popq    %rcx                                            \n\
 137         popq    %rdx                                            \n\
 138         popq    %rax                                            \n\
 139 .mcount_exit:                                                   \n\
 140         ret                                                     \n\
 141 ");
 142 #else /* !__GNUCLIKE_ASM */
 143 #error "this file needs to be ported to your compiler"
 144 #endif /* __GNUCLIKE_ASM */
 145
 146 #ifdef GUPROF
 147 /*
 148  * [.]mexitcount saves the return register(s), loads selfpc and calls
 149  * mexitcount(selfpc) to do the work.  Someday it should be in a machine
 150  * dependent file together with cputime(), __mcount and [.]mcount.  cputime()
 151  * can't just be put in machdep.c because it has to be compiled without -pg.
 152  */
 153 #ifdef __GNUCLIKE_ASM
 154 __asm("                                                         \n\
 155         .text                                                   \n\
 156 #                                                               \n\
 157 # Dummy label to be seen when gprof -u hides [.]mexitcount.     \n\
 158 #                                                               \n\
 159         .p2align 4,0x90                                         \n\
 160         .globl  __mexitcount                                    \n\
 161         .type   __mexitcount,@function                          \n\
 162 __mexitcount:                                                   \n\
 163         nop                                                     \n\
 164                                                                 \n\
 165 GMON_PROF_HIRES =       4                                       \n\
 166                                                                 \n\
 167         .p2align 4,0x90                                         \n\
 168         .globl  .mexitcount                                     \n\
 169 .mexitcount:                                                    \n\
 170         .globl  __cyg_profile_func_exit                         \n\
 171 __cyg_profile_func_exit:                                        \n\
 172         cmpl    $GMON_PROF_HIRES,_gmonparam+GM_STATE            \n\
 173         jne     .mexitcount_exit                                \n\
 174         pushq   %rax                                            \n\
 175         pushq   %rdx                                            \n\
 176         pushq   %rcx                                            \n\
 177         pushq   %rsi                                            \n\
 178         pushq   %rdi                                            \n\
 179         pushq   %r8                                             \n\
 180         pushq   %r9                                             \n\
 181         movq    7*8(%rsp),%rdi                                  \n\
 182         pushfq                                                  \n\
 183         cli                                                     \n\
 184         call    mexitcount                                      \n\
 185         popfq                                                   \n\
 186         popq    %r9                                             \n\
 187         popq    %r8                                             \n\
 188         popq    %rdi                                            \n\
 189         popq    %rsi                                            \n\
 190         popq    %rcx                                            \n\
 191         popq    %rdx                                            \n\
 192         popq    %rax                                            \n\
 193 .mexitcount_exit:                                               \n\
 194         ret                                                     \n\
 195 ");
 196 #endif /* __GNUCLIKE_ASM */
 197
 198 /*
 199  * Return the time elapsed since the last call.  The units are machine-
 200  * dependent.
 201  */
 202 int
 203 cputime()
 204 {
 205         u_int count;
 206         int delta;
 207 #if (defined(I586_CPU) || defined(I686_CPU)) && !defined(SMP) && \
 208     defined(PERFMON) && defined(I586_PMC_GUPROF)
 209         u_quad_t event_count;
 210 #endif
 211         u_char high, low;
 212         static u_int prev_count;
 213
 214 #if defined(I586_CPU) || defined(I686_CPU)
 215         if (cputime_clock == CPUTIME_CLOCK_TSC) {
 216                 /*
 217                  * Scale the TSC a little to make cputime()'s frequency
 218                  * fit in an int, assuming that the TSC frequency fits
 219                  * in a u_int.  Use a fixed scale since dynamic scaling
 220                  * would be slower and we can't really use the low bit
 221                  * of precision.
 222                  */
 223                 count = (u_int)rdtsc() & ~1u;
 224                 delta = (int)(count - prev_count) >> 1;
 225                 prev_count = count;
 226                 return (delta);
 227         }
 228 #if defined(PERFMON) && defined(I586_PMC_GUPROF) && !defined(SMP)
 229         if (cputime_clock == CPUTIME_CLOCK_I586_PMC) {
 230                 /*
 231                  * XXX permon_read() should be inlined so that the
 232                  * perfmon module doesn't need to be compiled with
 233                  * profiling disabled and so that it is fast.
 234                  */
 235                 perfmon_read(0, &event_count);
 236
 237                 count = (u_int)event_count;
 238                 delta = (int)(count - prev_count);
 239                 prev_count = count;
 240                 return (delta);
 241         }
 242 #endif /* PERFMON && I586_PMC_GUPROF && !SMP */
 243 #endif /* I586_CPU || I686_CPU */
 244
 245         /*
 246          * Read the current value of the 8254 timer counter 0.
 247          */
 248         outb(TIMER_MODE, TIMER_SEL0 | TIMER_LATCH);
 249         low = inb(TIMER_CNTR0);
 250         high = inb(TIMER_CNTR0);
 251         count = ((high << 8) | low) << CPUTIME_CLOCK_I8254_SHIFT;
 252
 253         /*
 254          * The timer counts down from TIMER_CNTR0_MAX to 0 and then resets.
 255          * While profiling is enabled, this routine is called at least twice
 256          * per timer reset (for mcounting and mexitcounting hardclock()),
 257          * so at most one reset has occurred since the last call, and one
 258          * has occurred iff the current count is larger than the previous
 259          * count.  This allows counter underflow to be detected faster
 260          * than in microtime().
 261          */
 262         delta = prev_count - count;
 263         prev_count = count;
 264         if ((int) delta <= 0)
 265                 return (delta + (timer0_max_count << CPUTIME_CLOCK_I8254_SHIFT));
 266         return (delta);
 267 }
 268
 269 static int
 270 sysctl_machdep_cputime_clock(SYSCTL_HANDLER_ARGS)
 271 {
 272         int clock;
 273         int error;
 274 #if defined(PERFMON) && defined(I586_PMC_GUPROF)
 275         int event;
 276         struct pmc pmc;
 277 #endif
 278
 279         clock = cputime_clock;
 280 #if defined(PERFMON) && defined(I586_PMC_GUPROF)
 281         if (clock == CPUTIME_CLOCK_I586_PMC) {
 282                 pmc.pmc_val = cputime_clock_pmc_conf;
 283                 clock += pmc.pmc_event;
 284         }
 285 #endif
 286         error = sysctl_handle_opaque(oidp, &clock, sizeof clock, req);
 287         if (error == 0 && req->newptr != NULL) {
 288 #if defined(PERFMON) && defined(I586_PMC_GUPROF)
 289                 if (clock >= CPUTIME_CLOCK_I586_PMC) {
 290                         event = clock - CPUTIME_CLOCK_I586_PMC;
 291                         if (event >= 256)
 292                                 return (EINVAL);
 293                         pmc.pmc_num = 0;
 294                         pmc.pmc_event = event;
 295                         pmc.pmc_unit = 0;
 296                         pmc.pmc_flags = PMCF_E | PMCF_OS | PMCF_USR;
 297                         pmc.pmc_mask = 0;
 298                         cputime_clock_pmc_conf = pmc.pmc_val;
 299                         cputime_clock = CPUTIME_CLOCK_I586_PMC;
 300                 } else
 301 #endif
 302                 {
 303                         if (clock < 0 || clock >= CPUTIME_CLOCK_I586_PMC)
 304                                 return (EINVAL);
 305                         cputime_clock = clock;
 306                 }
 307         }
 308         return (error);
 309 }
 310
 311 SYSCTL_PROC(_machdep, OID_AUTO, cputime_clock, CTLTYPE_INT | CTLFLAG_RW,
 312             0, sizeof(u_int), sysctl_machdep_cputime_clock, "I", "");
 313
 314 /*
 315  * The start and stop routines need not be here since we turn off profiling
 316  * before calling them.  They are here for convenience.
 317  */
 318
 319 void
 320 startguprof(gp)
 321         struct gmonparam *gp;
 322 {
 323         if (cputime_clock == CPUTIME_CLOCK_UNINITIALIZED) {
 324                 cputime_clock = CPUTIME_CLOCK_I8254;
 325 #if defined(I586_CPU) || defined(I686_CPU)
 326                 if (tsc_freq != 0 && !tsc_is_broken && mp_ncpus == 1)
 327                         cputime_clock = CPUTIME_CLOCK_TSC;
 328 #endif
 329         }
 330         gp->profrate = timer_freq << CPUTIME_CLOCK_I8254_SHIFT;
 331 #if defined(I586_CPU) || defined(I686_CPU)
 332         if (cputime_clock == CPUTIME_CLOCK_TSC) {
 333                 gp->profrate = tsc_freq >> 1;
 334                 cputime_prof_active = 1;
 335         }
 336 #if defined(PERFMON) && defined(I586_PMC_GUPROF)
 337         else if (cputime_clock == CPUTIME_CLOCK_I586_PMC) {
 338                 if (perfmon_avail() &&
 339                     perfmon_setup(0, cputime_clock_pmc_conf) == 0) {
 340                         if (perfmon_start(0) != 0)
 341                                 perfmon_fini(0);
 342                         else {
 343                                 /* XXX 1 event == 1 us. */
 344                                 gp->profrate = 1000000;
 345
 346                                 saved_gmp = *gp;
 347
 348                                 /* Zap overheads.  They are invalid. */
 349                                 gp->cputime_overhead = 0;
 350                                 gp->mcount_overhead = 0;
 351                                 gp->mcount_post_overhead = 0;
 352                                 gp->mcount_pre_overhead = 0;
 353                                 gp->mexitcount_overhead = 0;
 354                                 gp->mexitcount_post_overhead = 0;
 355                                 gp->mexitcount_pre_overhead = 0;
 356
 357                                 cputime_clock_pmc_init = TRUE;
 358                         }
 359                 }
 360         }
 361 #endif /* PERFMON && I586_PMC_GUPROF */
 362 #endif /* I586_CPU || I686_CPU */
 363         cputime_bias = 0;
 364         cputime();
 365 }
 366
 367 void
 368 stopguprof(gp)
 369         struct gmonparam *gp;
 370 {
 371 #if defined(PERFMON) && defined(I586_PMC_GUPROF)
 372         if (cputime_clock_pmc_init) {
 373                 *gp = saved_gmp;
 374                 perfmon_fini(0);
 375                 cputime_clock_pmc_init = FALSE;
 376         }
 377 #endif
 378 #if defined(I586_CPU) || defined(I686_CPU)
 379         if (cputime_clock == CPUTIME_CLOCK_TSC)
 380                 cputime_prof_active = 0;
 381 #endif
 382 }
 383
 384 #if defined(I586_CPU) || defined(I686_CPU)
 385 /* If the cpu frequency changed while profiling, report a warning. */
 386 static void
 387 tsc_freq_changed(void *arg, const struct cf_level *level, int status)
 388 {
 389
 390         /*
 391          * If there was an error during the transition or
 392          * TSC is P-state invariant, don't do anything.
 393          */
 394         if (status != 0 || tsc_is_invariant)
 395                 return;
 396         if (cputime_prof_active && cputime_clock == CPUTIME_CLOCK_TSC)
 397                 printf("warning: cpu freq changed while profiling active\n");
 398 }
 399
 400 EVENTHANDLER_DEFINE(cpufreq_post_change, tsc_freq_changed, NULL,
 401     EVENTHANDLER_PRI_ANY);
 402 #endif /* I586_CPU || I686_CPU */
 403
 404 #endif /* GUPROF */