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         cmpl    $GMON_PROF_OFF,_gmonparam+GM_STATE              \n\
 105         je      .mcount_exit                                    \n\
 106         #                                                       \n\
 107         # The caller's stack frame has already been built, so   \n\
 108         # %rbp is the caller's frame pointer.  The caller's     \n\
 109         # raddr is in the caller's frame following the caller's \n\
 110         # caller's frame pointer.                               \n\
 111         #                                                       \n\
 112         pushq   %rax                                            \n\
 113         pushq   %rdx                                            \n\
 114         pushq   %rcx                                            \n\
 115         pushq   %rsi                                            \n\
 116         pushq   %rdi                                            \n\
 117         pushq   %r8                                             \n\
 118         pushq   %r9                                             \n\
 119         movq    8(%rbp),%rdi                                    \n\
 120 .got_frompc:                                                    \n\
 121         #                                                       \n\
 122         # Our raddr is the caller's pc.                         \n\
 123         #                                                       \n\
 124         movq    7*8(%rsp),%rsi                                  \n\
 125                                                                 \n\
 126         pushfq                                                  \n\
 127         cli                                                     \n\
 128         call    mcount                                          \n\
 129         popfq                                                   \n\
 130         popq    %r9                                             \n\
 131         popq    %r8                                             \n\
 132         popq    %rdi                                            \n\
 133         popq    %rsi                                            \n\
 134         popq    %rcx                                            \n\
 135         popq    %rdx                                            \n\
 136         popq    %rax                                            \n\
 137 .mcount_exit:                                                   \n\
 138         ret     $0                                              \n\
 139 ");
 140 #else /* !__GNUCLIKE_ASM */
 141 #error "this file needs to be ported to your compiler"
 142 #endif /* __GNUCLIKE_ASM */
 143
 144 #ifdef GUPROF
 145 /*
 146  * [.]mexitcount saves the return register(s), loads selfpc and calls
 147  * mexitcount(selfpc) to do the work.  Someday it should be in a machine
 148  * dependent file together with cputime(), __mcount and [.]mcount.  cputime()
 149  * can't just be put in machdep.c because it has to be compiled without -pg.
 150  */
 151 #ifdef __GNUCLIKE_ASM
 152 __asm("                                                         \n\
 153         .text                                                   \n\
 154 #                                                               \n\
 155 # Dummy label to be seen when gprof -u hides [.]mexitcount.     \n\
 156 #                                                               \n\
 157         .p2align 4,0x90                                         \n\
 158         .globl  __mexitcount                                    \n\
 159         .type   __mexitcount,@function                          \n\
 160 __mexitcount:                                                   \n\
 161         nop                                                     \n\
 162                                                                 \n\
 163 GMON_PROF_HIRES =       4                                       \n\
 164                                                                 \n\
 165         .p2align 4,0x90                                         \n\
 166         .globl  .mexitcount                                     \n\
 167 .mexitcount:                                                    \n\
 168         cmpl    $GMON_PROF_HIRES,_gmonparam+GM_STATE            \n\
 169         jne     .mexitcount_exit                                \n\
 170         pushq   %rax                                            \n\
 171         pushq   %rdx                                            \n\
 172         pushq   %rcx                                            \n\
 173         pushq   %rsi                                            \n\
 174         pushq   %rdi                                            \n\
 175         pushq   %r8                                             \n\
 176         pushq   %r9                                             \n\
 177         movq    7*8(%rsp),%rdi                                  \n\
 178         pushfq                                                  \n\
 179         cli                                                     \n\
 180         call    mexitcount                                      \n\
 181         popfq                                                   \n\
 182         popq    %r9                                             \n\
 183         popq    %r8                                             \n\
 184         popq    %rdi                                            \n\
 185         popq    %rsi                                            \n\
 186         popq    %rcx                                            \n\
 187         popq    %rdx                                            \n\
 188         popq    %rax                                            \n\
 189 .mexitcount_exit:                                               \n\
 190         ret     $0                                              \n\
 191 ");
 192 #endif /* __GNUCLIKE_ASM */
 193
 194 /*
 195  * Return the time elapsed since the last call.  The units are machine-
 196  * dependent.
 197  */
 198 int
 199 cputime()
 200 {
 201         u_int count;
 202         int delta;
 203 #if (defined(I586_CPU) || defined(I686_CPU)) && !defined(SMP) && \
 204     defined(PERFMON) && defined(I586_PMC_GUPROF)
 205         u_quad_t event_count;
 206 #endif
 207         u_char high, low;
 208         static u_int prev_count;
 209
 210 #if defined(I586_CPU) || defined(I686_CPU)
 211         if (cputime_clock == CPUTIME_CLOCK_TSC) {
 212                 /*
 213                  * Scale the TSC a little to make cputime()'s frequency
 214                  * fit in an int, assuming that the TSC frequency fits
 215                  * in a u_int.  Use a fixed scale since dynamic scaling
 216                  * would be slower and we can't really use the low bit
 217                  * of precision.
 218                  */
 219                 count = (u_int)rdtsc() & ~1u;
 220                 delta = (int)(count - prev_count) >> 1;
 221                 prev_count = count;
 222                 return (delta);
 223         }
 224 #if defined(PERFMON) && defined(I586_PMC_GUPROF) && !defined(SMP)
 225         if (cputime_clock == CPUTIME_CLOCK_I586_PMC) {
 226                 /*
 227                  * XXX permon_read() should be inlined so that the
 228                  * perfmon module doesn't need to be compiled with
 229                  * profiling disabled and so that it is fast.
 230                  */
 231                 perfmon_read(0, &event_count);
 232
 233                 count = (u_int)event_count;
 234                 delta = (int)(count - prev_count);
 235                 prev_count = count;
 236                 return (delta);
 237         }
 238 #endif /* PERFMON && I586_PMC_GUPROF && !SMP */
 239 #endif /* I586_CPU || I686_CPU */
 240
 241         /*
 242          * Read the current value of the 8254 timer counter 0.
 243          */
 244         outb(TIMER_MODE, TIMER_SEL0 | TIMER_LATCH);
 245         low = inb(TIMER_CNTR0);
 246         high = inb(TIMER_CNTR0);
 247         count = ((high << 8) | low) << CPUTIME_CLOCK_I8254_SHIFT;
 248
 249         /*
 250          * The timer counts down from TIMER_CNTR0_MAX to 0 and then resets.
 251          * While profiling is enabled, this routine is called at least twice
 252          * per timer reset (for mcounting and mexitcounting hardclock()),
 253          * so at most one reset has occurred since the last call, and one
 254          * has occurred iff the current count is larger than the previous
 255          * count.  This allows counter underflow to be detected faster
 256          * than in microtime().
 257          */
 258         delta = prev_count - count;
 259         prev_count = count;
 260         if ((int) delta <= 0)
 261                 return (delta + (i8254_max_count << CPUTIME_CLOCK_I8254_SHIFT));
 262         return (delta);
 263 }
 264
 265 static int
 266 sysctl_machdep_cputime_clock(SYSCTL_HANDLER_ARGS)
 267 {
 268         int clock;
 269         int error;
 270 #if defined(PERFMON) && defined(I586_PMC_GUPROF)
 271         int event;
 272         struct pmc pmc;
 273 #endif
 274
 275         clock = cputime_clock;
 276 #if defined(PERFMON) && defined(I586_PMC_GUPROF)
 277         if (clock == CPUTIME_CLOCK_I586_PMC) {
 278                 pmc.pmc_val = cputime_clock_pmc_conf;
 279                 clock += pmc.pmc_event;
 280         }
 281 #endif
 282         error = sysctl_handle_opaque(oidp, &clock, sizeof clock, req);
 283         if (error == 0 && req->newptr != NULL) {
 284 #if defined(PERFMON) && defined(I586_PMC_GUPROF)
 285                 if (clock >= CPUTIME_CLOCK_I586_PMC) {
 286                         event = clock - CPUTIME_CLOCK_I586_PMC;
 287                         if (event >= 256)
 288                                 return (EINVAL);
 289                         pmc.pmc_num = 0;
 290                         pmc.pmc_event = event;
 291                         pmc.pmc_unit = 0;
 292                         pmc.pmc_flags = PMCF_E | PMCF_OS | PMCF_USR;
 293                         pmc.pmc_mask = 0;
 294                         cputime_clock_pmc_conf = pmc.pmc_val;
 295                         cputime_clock = CPUTIME_CLOCK_I586_PMC;
 296                 } else
 297 #endif
 298                 {
 299                         if (clock < 0 || clock >= CPUTIME_CLOCK_I586_PMC)
 300                                 return (EINVAL);
 301                         cputime_clock = clock;
 302                 }
 303         }
 304         return (error);
 305 }
 306
 307 SYSCTL_PROC(_machdep, OID_AUTO, cputime_clock, CTLTYPE_INT | CTLFLAG_RW,
 308             0, sizeof(u_int), sysctl_machdep_cputime_clock, "I", "");
 309
 310 /*
 311  * The start and stop routines need not be here since we turn off profiling
 312  * before calling them.  They are here for convenience.
 313  */
 314
 315 void
 316 startguprof(gp)
 317         struct gmonparam *gp;
 318 {
 319         if (cputime_clock == CPUTIME_CLOCK_UNINITIALIZED) {
 320                 cputime_clock = CPUTIME_CLOCK_I8254;
 321 #if defined(I586_CPU) || defined(I686_CPU)
 322                 if (tsc_freq != 0 && !tsc_is_broken && mp_ncpus == 1)
 323                         cputime_clock = CPUTIME_CLOCK_TSC;
 324 #endif
 325         }
 326         gp->profrate = i8254_freq << CPUTIME_CLOCK_I8254_SHIFT;
 327 #if defined(I586_CPU) || defined(I686_CPU)
 328         if (cputime_clock == CPUTIME_CLOCK_TSC) {
 329                 gp->profrate = tsc_freq >> 1;
 330                 cputime_prof_active = 1;
 331         }
 332 #if defined(PERFMON) && defined(I586_PMC_GUPROF)
 333         else if (cputime_clock == CPUTIME_CLOCK_I586_PMC) {
 334                 if (perfmon_avail() &&
 335                     perfmon_setup(0, cputime_clock_pmc_conf) == 0) {
 336                         if (perfmon_start(0) != 0)
 337                                 perfmon_fini(0);
 338                         else {
 339                                 /* XXX 1 event == 1 us. */
 340                                 gp->profrate = 1000000;
 341
 342                                 saved_gmp = *gp;
 343
 344                                 /* Zap overheads.  They are invalid. */
 345                                 gp->cputime_overhead = 0;
 346                                 gp->mcount_overhead = 0;
 347                                 gp->mcount_post_overhead = 0;
 348                                 gp->mcount_pre_overhead = 0;
 349                                 gp->mexitcount_overhead = 0;
 350                                 gp->mexitcount_post_overhead = 0;
 351                                 gp->mexitcount_pre_overhead = 0;
 352
 353                                 cputime_clock_pmc_init = TRUE;
 354                         }
 355                 }
 356         }
 357 #endif /* PERFMON && I586_PMC_GUPROF */
 358 #endif /* I586_CPU || I686_CPU */
 359         cputime_bias = 0;
 360         cputime();
 361 }
 362
 363 void
 364 stopguprof(gp)
 365         struct gmonparam *gp;
 366 {
 367 #if defined(PERFMON) && defined(I586_PMC_GUPROF)
 368         if (cputime_clock_pmc_init) {
 369                 *gp = saved_gmp;
 370                 perfmon_fini(0);
 371                 cputime_clock_pmc_init = FALSE;
 372         }
 373 #endif
 374 #if defined(I586_CPU) || defined(I686_CPU)
 375         if (cputime_clock == CPUTIME_CLOCK_TSC)
 376                 cputime_prof_active = 0;
 377 #endif
 378 }
 379
 380 #if defined(I586_CPU) || defined(I686_CPU)
 381 /* If the cpu frequency changed while profiling, report a warning. */
 382 static void
 383 tsc_freq_changed(void *arg, const struct cf_level *level, int status)
 384 {
 385
 386         /*
 387          * If there was an error during the transition or
 388          * TSC is P-state invariant, don't do anything.
 389          */
 390         if (status != 0 || tsc_is_invariant)
 391                 return;
 392         if (cputime_prof_active && cputime_clock == CPUTIME_CLOCK_TSC)
 393                 printf("warning: cpu freq changed while profiling active\n");
 394 }
 395
 396 EVENTHANDLER_DEFINE(cpufreq_post_change, tsc_freq_changed, NULL,
 397     EVENTHANDLER_PRI_ANY);
 398 #endif /* I586_CPU || I686_CPU */
 399
 400 #endif /* GUPROF */