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/gmon.h>
  39 #include <sys/kernel.h>
  40 #include <sys/sysctl.h>
  41
  42 #include <machine/clock.h>
  43 #if 0
  44 #include <machine/perfmon.h>
  45 #endif
  46 #include <machine/profile.h>
  47 #undef MCOUNT
  48 #endif
  49
  50 #include <machine/asmacros.h>
  51 #include <machine/timerreg.h>
  52
  53 #ifdef GUPROF
  54 #define CPUTIME_CLOCK_UNINITIALIZED     0
  55 #define CPUTIME_CLOCK_I8254             1
  56 #define CPUTIME_CLOCK_TSC               2
  57 #define CPUTIME_CLOCK_I586_PMC          3
  58 #define CPUTIME_CLOCK_I8254_SHIFT       7
  59
  60 int     cputime_bias = 1;       /* initialize for locality of reference */
  61
  62 static int      cputime_clock = CPUTIME_CLOCK_UNINITIALIZED;
  63 #ifdef I586_PMC_GUPROF
  64 static u_int    cputime_clock_pmc_conf = I586_PMC_GUPROF;
  65 static int      cputime_clock_pmc_init;
  66 static struct gmonparam saved_gmp;
  67 #endif
  68 #endif /* GUPROF */
  69
  70 #ifdef __GNUCLIKE_ASM
  71 __asm("                                                         \n\
  72 GM_STATE        =       0                                       \n\
  73 GMON_PROF_OFF   =       3                                       \n\
  74                                                                 \n\
  75         .text                                                   \n\
  76         .p2align 4,0x90                                         \n\
  77         .globl  __mcount                                        \n\
  78         .type   __mcount,@function                              \n\
  79 __mcount:                                                       \n\
  80         #                                                       \n\
  81         # Check that we are profiling.  Do it early for speed.  \n\
  82         #                                                       \n\
  83         cmpl    $GMON_PROF_OFF," __XSTRING(CNAME(_gmonparam)) "+GM_STATE \n\
  84         je      .mcount_exit                                    \n\
  85         #                                                       \n\
  86         # __mcount is the same as [.]mcount except the caller   \n\
  87         # hasn't changed the stack except to call here, so the  \n\
  88         # caller's raddr is above our raddr.                    \n\
  89         #                                                       \n\
  90         pushq   %rax                                            \n\
  91         pushq   %rdx                                            \n\
  92         pushq   %rcx                                            \n\
  93         pushq   %rsi                                            \n\
  94         pushq   %rdi                                            \n\
  95         pushq   %r8                                             \n\
  96         pushq   %r9                                             \n\
  97         movq    7*8+8(%rsp),%rdi                                \n\
  98         jmp     .got_frompc                                     \n\
  99                                                                 \n\
 100         .p2align 4,0x90                                         \n\
 101         .globl  " __XSTRING(HIDENAME(mcount)) "                 \n\
 102 " __XSTRING(HIDENAME(mcount)) ":                                \n\
 103         .globl  __cyg_profile_func_enter                        \n\
 104 __cyg_profile_func_enter:                                       \n\
 105         cmpl    $GMON_PROF_OFF," __XSTRING(CNAME(_gmonparam)) "+GM_STATE \n\
 106         je      .mcount_exit                                    \n\
 107         #                                                       \n\
 108         # The caller's stack frame has already been built, so   \n\
 109         # %rbp is the caller's frame pointer.  The caller's     \n\
 110         # raddr is in the caller's frame following the caller's \n\
 111         # caller's frame pointer.                               \n\
 112         #                                                       \n\
 113         pushq   %rax                                            \n\
 114         pushq   %rdx                                            \n\
 115         pushq   %rcx                                            \n\
 116         pushq   %rsi                                            \n\
 117         pushq   %rdi                                            \n\
 118         pushq   %r8                                             \n\
 119         pushq   %r9                                             \n\
 120         movq    8(%rbp),%rdi                                    \n\
 121 .got_frompc:                                                    \n\
 122         #                                                       \n\
 123         # Our raddr is the caller's pc.                         \n\
 124         #                                                       \n\
 125         movq    7*8(%rsp),%rsi                                  \n\
 126                                                                 \n\
 127         pushfq                                                  \n\
 128         cli                                                     \n\
 129         call    " __XSTRING(CNAME(mcount)) "                    \n\
 130         popfq                                                   \n\
 131         popq    %r9                                             \n\
 132         popq    %r8                                             \n\
 133         popq    %rdi                                            \n\
 134         popq    %rsi                                            \n\
 135         popq    %rcx                                            \n\
 136         popq    %rdx                                            \n\
 137         popq    %rax                                            \n\
 138 .mcount_exit:                                                   \n\
 139         ret                                                     \n\
 140 ");
 141 #else /* !__GNUCLIKE_ASM */
 142 #error this file needs to be ported to your compiler
 143 #endif /* __GNUCLIKE_ASM */
 144
 145 #ifdef GUPROF
 146 /*
 147  * [.]mexitcount saves the return register(s), loads selfpc and calls
 148  * mexitcount(selfpc) to do the work.  Someday it should be in a machine
 149  * dependent file together with cputime(), __mcount and [.]mcount.  cputime()
 150  * can't just be put in machdep.c because it has to be compiled without -pg.
 151  */
 152 #ifdef __GNUCLIKE_ASM
 153 __asm("                                                         \n\
 154         .text                                                   \n\
 155 #                                                               \n\
 156 # Dummy label to be seen when gprof -u hides [.]mexitcount.     \n\
 157 #                                                               \n\
 158         .p2align 4,0x90                                         \n\
 159         .globl  __mexitcount                                    \n\
 160         .type   __mexitcount,@function                          \n\
 161 __mexitcount:                                                   \n\
 162         nop                                                     \n\
 163                                                                 \n\
 164 GMON_PROF_HIRES =       4                                       \n\
 165                                                                 \n\
 166         .p2align 4,0x90                                         \n\
 167         .globl  " __XSTRING(HIDENAME(mexitcount)) "             \n\
 168 " __XSTRING(HIDENAME(mexitcount)) ":                            \n\
 169         .globl  __cyg_profile_func_exit                         \n\
 170 __cyg_profile_func_exit:                                        \n\
 171         cmpl    $GMON_PROF_HIRES," __XSTRING(CNAME(_gmonparam)) "+GM_STATE \n\
 172         jne     .mexitcount_exit                                \n\
 173         pushq   %rax                                            \n\
 174         pushq   %rdx                                            \n\
 175         pushq   %rcx                                            \n\
 176         pushq   %rsi                                            \n\
 177         pushq   %rdi                                            \n\
 178         pushq   %r8                                             \n\
 179         pushq   %r9                                             \n\
 180         movq    7*8(%rsp),%rdi                                  \n\
 181         pushfq                                                  \n\
 182         cli                                                     \n\
 183         call    " __XSTRING(CNAME(mexitcount)) "                \n\
 184         popfq                                                   \n\
 185         popq    %r9                                             \n\
 186         popq    %r8                                             \n\
 187         popq    %rdi                                            \n\
 188         popq    %rsi                                            \n\
 189         popq    %rcx                                            \n\
 190         popq    %rdx                                            \n\
 191         popq    %rax                                            \n\
 192 .mexitcount_exit:                                               \n\
 193         ret                                                     \n\
 194 ");
 195 #else /* !__GNUCLIKE_ASM */
 196 #error this file needs to be ported to your compiler
 197 #endif /* __GNUCLIKE_ASM */
 198
 199 /*
 200  * Return the time elapsed since the last call.  The units are machine-
 201  * dependent.
 202  */
 203 int
 204 cputime()
 205 {
 206         u_int count;
 207         int delta;
 208 #if (defined(I586_CPU) || defined(I686_CPU)) && !defined(SMP) && \
 209     defined(PERFMON) && defined(I586_PMC_GUPROF)
 210         u_quad_t event_count;
 211 #endif
 212         u_char high, low;
 213         static u_int prev_count;
 214
 215 #ifndef SMP
 216         if (cputime_clock == CPUTIME_CLOCK_TSC) {
 217                 /*
 218                  * Scale the TSC a little to make cputime()'s frequency
 219                  * fit in an int, assuming that the TSC frequency fits
 220                  * in a u_int.  Use a fixed scale since dynamic scaling
 221                  * would be slower and we can't really use the low bit
 222                  * of precision.
 223                  */
 224                 count = (u_int)rdtsc() & ~1u;
 225                 delta = (int)(count - prev_count) >> 1;
 226                 prev_count = count;
 227                 return (delta);
 228         }
 229 #if defined(PERFMON) && defined(I586_PMC_GUPROF)
 230         if (cputime_clock == CPUTIME_CLOCK_I586_PMC) {
 231                 /*
 232                  * XXX permon_read() should be inlined so that the
 233                  * perfmon module doesn't need to be compiled with
 234                  * profiling disabled and so that it is fast.
 235                  */
 236                 perfmon_read(0, &event_count);
 237
 238                 count = (u_int)event_count;
 239                 delta = (int)(count - prev_count);
 240                 prev_count = count;
 241                 return (delta);
 242         }
 243 #endif /* PERFMON && I586_PMC_GUPROF */
 244 #endif /* !SMP */
 245
 246         /*
 247          * Read the current value of the 8254 timer counter 0.
 248          */
 249         outb(TIMER_MODE, TIMER_SEL0 | TIMER_LATCH);
 250         low = inb(TIMER_CNTR0);
 251         high = inb(TIMER_CNTR0);
 252         count = ((high << 8) | low) << CPUTIME_CLOCK_I8254_SHIFT;
 253
 254         /*
 255          * The timer counts down from TIMER_CNTR0_MAX to 0 and then resets.
 256          * While profiling is enabled, this routine is called at least twice
 257          * per timer reset (for mcounting and mexitcounting hardclock()),
 258          * so at most one reset has occurred since the last call, and one
 259          * has occurred iff the current count is larger than the previous
 260          * count.  This allows counter underflow to be detected faster
 261          * than in microtime().
 262          */
 263         delta = prev_count - count;
 264         prev_count = count;
 265         if ((int) delta <= 0)
 266                 return (delta + (timer0_max_count << CPUTIME_CLOCK_I8254_SHIFT));
 267         return (delta);
 268 }
 269
 270 static int
 271 sysctl_machdep_cputime_clock(SYSCTL_HANDLER_ARGS)
 272 {
 273         int clock;
 274         int error;
 275 #if defined(PERFMON) && defined(I586_PMC_GUPROF)
 276         int event;
 277         struct pmc pmc;
 278 #endif
 279
 280         clock = cputime_clock;
 281 #if defined(PERFMON) && defined(I586_PMC_GUPROF)
 282         if (clock == CPUTIME_CLOCK_I586_PMC) {
 283                 pmc.pmc_val = cputime_clock_pmc_conf;
 284                 clock += pmc.pmc_event;
 285         }
 286 #endif
 287         error = sysctl_handle_opaque(oidp, &clock, sizeof clock, req);
 288         if (error == 0 && req->newptr != NULL) {
 289 #if defined(PERFMON) && defined(I586_PMC_GUPROF)
 290                 if (clock >= CPUTIME_CLOCK_I586_PMC) {
 291                         event = clock - CPUTIME_CLOCK_I586_PMC;
 292                         if (event >= 256)
 293                                 return (EINVAL);
 294                         pmc.pmc_num = 0;
 295                         pmc.pmc_event = event;
 296                         pmc.pmc_unit = 0;
 297                         pmc.pmc_flags = PMCF_E | PMCF_OS | PMCF_USR;
 298                         pmc.pmc_mask = 0;
 299                         cputime_clock_pmc_conf = pmc.pmc_val;
 300                         cputime_clock = CPUTIME_CLOCK_I586_PMC;
 301                 } else
 302 #endif
 303                 {
 304                         if (clock < 0 || clock >= CPUTIME_CLOCK_I586_PMC)
 305                                 return (EINVAL);
 306                         cputime_clock = clock;
 307                 }
 308         }
 309         return (error);
 310 }
 311
 312 SYSCTL_PROC(_machdep, OID_AUTO, cputime_clock, CTLTYPE_INT | CTLFLAG_RW,
 313             0, sizeof(u_int), sysctl_machdep_cputime_clock, "I", "");
 314
 315 /*
 316  * The start and stop routines need not be here since we turn off profiling
 317  * before calling them.  They are here for convenience.
 318  */
 319
 320 void
 321 startguprof(gp)
 322         struct gmonparam *gp;
 323 {
 324         if (cputime_clock == CPUTIME_CLOCK_UNINITIALIZED) {
 325                 cputime_clock = CPUTIME_CLOCK_I8254;
 326 #ifndef SMP
 327                 if (tsc_freq != 0)
 328                         cputime_clock = CPUTIME_CLOCK_TSC;
 329 #endif
 330         }
 331         gp->profrate = timer_freq << CPUTIME_CLOCK_I8254_SHIFT;
 332 #ifndef SMP
 333         if (cputime_clock == CPUTIME_CLOCK_TSC)
 334                 gp->profrate = tsc_freq >> 1;
 335 #if defined(PERFMON) && defined(I586_PMC_GUPROF)
 336         else if (cputime_clock == CPUTIME_CLOCK_I586_PMC) {
 337                 if (perfmon_avail() &&
 338                     perfmon_setup(0, cputime_clock_pmc_conf) == 0) {
 339                         if (perfmon_start(0) != 0)
 340                                 perfmon_fini(0);
 341                         else {
 342                                 /* XXX 1 event == 1 us. */
 343                                 gp->profrate = 1000000;
 344
 345                                 saved_gmp = *gp;
 346
 347                                 /* Zap overheads.  They are invalid. */
 348                                 gp->cputime_overhead = 0;
 349                                 gp->mcount_overhead = 0;
 350                                 gp->mcount_post_overhead = 0;
 351                                 gp->mcount_pre_overhead = 0;
 352                                 gp->mexitcount_overhead = 0;
 353                                 gp->mexitcount_post_overhead = 0;
 354                                 gp->mexitcount_pre_overhead = 0;
 355
 356                                 cputime_clock_pmc_init = TRUE;
 357                         }
 358                 }
 359         }
 360 #endif /* PERFMON && I586_PMC_GUPROF */
 361 #endif /* !SMP */
 362         cputime_bias = 0;
 363         cputime();
 364 }
 365
 366 void
 367 stopguprof(gp)
 368         struct gmonparam *gp;
 369 {
 370 #if defined(PERFMON) && defined(I586_PMC_GUPROF)
 371         if (cputime_clock_pmc_init) {
 372                 *gp = saved_gmp;
 373                 perfmon_fini(0);
 374                 cputime_clock_pmc_init = FALSE;
 375         }
 376 #endif
 377 }
 378
 379 #else /* !GUPROF */
 380 #ifdef __GNUCLIKE_ASM
 381 __asm("                                                         \n\
 382         .text                                                   \n\
 383         .p2align 4,0x90                                         \n\
 384         .globl  " __XSTRING(HIDENAME(mexitcount)) "             \n\
 385 " __XSTRING(HIDENAME(mexitcount)) ":                            \n\
 386         ret                                                     \n\
 387 ");
 388 #else /* !__GNUCLIKE_ASM */
 389 #error this file needs to be ported to your compiler
 390 #endif /* __GNUCLIKE_ASM */
 391 #endif /* GUPROF */