sys/amd64/amd64/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
  34 #include <sys/param.h>
  35 #include <sys/systm.h>
  36 #include <sys/bus.h>
  37 #include <sys/cpu.h>
  38 #include <sys/eventhandler.h>
  39 #include <sys/gmon.h>
  40 #include <sys/kernel.h>
  41 #include <sys/smp.h>
  42 #include <sys/sysctl.h>
  43
  44 #include <machine/clock.h>
  45 #include <machine/timerreg.h>
  46
  47 #define CPUTIME_CLOCK_UNINITIALIZED     0
  48 #define CPUTIME_CLOCK_I8254             1
  49 #define CPUTIME_CLOCK_TSC               2
  50 #define CPUTIME_CLOCK_I8254_SHIFT       7
  51
  52 int     cputime_bias = 1;       /* initialize for locality of reference */
  53
  54 static int      cputime_clock = CPUTIME_CLOCK_UNINITIALIZED;
  55 static int      cputime_prof_active;
  56 #endif /* GUPROF */
  57
  58 #ifdef __GNUCLIKE_ASM
  59 __asm("                                                         \n\
  60 GM_STATE        =       0                                       \n\
  61 GMON_PROF_OFF   =       3                                       \n\
  62                                                                 \n\
  63         .text                                                   \n\
  64         .p2align 4,0x90                                         \n\
  65         .globl  __mcount                                        \n\
  66         .type   __mcount,@function                              \n\
  67 __mcount:                                                       \n\
  68         #                                                       \n\
  69         # Check that we are profiling.  Do it early for speed.  \n\
  70         #                                                       \n\
  71         cmpl    $GMON_PROF_OFF,_gmonparam+GM_STATE              \n\
  72         je      .mcount_exit                                    \n\
  73         #                                                       \n\
  74         # __mcount is the same as [.]mcount except the caller   \n\
  75         # hasn't changed the stack except to call here, so the  \n\
  76         # caller's raddr is above our raddr.                    \n\
  77         #                                                       \n\
  78         pushq   %rax                                            \n\
  79         pushq   %rdx                                            \n\
  80         pushq   %rcx                                            \n\
  81         pushq   %rsi                                            \n\
  82         pushq   %rdi                                            \n\
  83         pushq   %r8                                             \n\
  84         pushq   %r9                                             \n\
  85         movq    7*8+8(%rsp),%rdi                                \n\
  86         jmp     .got_frompc                                     \n\
  87                                                                 \n\
  88         .p2align 4,0x90                                         \n\
  89         .globl  .mcount                                         \n\
  90 .mcount:                                                        \n\
  91         cmpl    $GMON_PROF_OFF,_gmonparam+GM_STATE              \n\
  92         je      .mcount_exit                                    \n\
  93         #                                                       \n\
  94         # The caller's stack frame has already been built, so   \n\
  95         # %rbp is the caller's frame pointer.  The caller's     \n\
  96         # raddr is in the caller's frame following the caller's \n\
  97         # caller's frame pointer.                               \n\
  98         #                                                       \n\
  99         pushq   %rax                                            \n\
 100         pushq   %rdx                                            \n\
 101         pushq   %rcx                                            \n\
 102         pushq   %rsi                                            \n\
 103         pushq   %rdi                                            \n\
 104         pushq   %r8                                             \n\
 105         pushq   %r9                                             \n\
 106         movq    8(%rbp),%rdi                                    \n\
 107 .got_frompc:                                                    \n\
 108         #                                                       \n\
 109         # Our raddr is the caller's pc.                         \n\
 110         #                                                       \n\
 111         movq    7*8(%rsp),%rsi                                  \n\
 112                                                                 \n\
 113         pushfq                                                  \n\
 114         cli                                                     \n\
 115         call    mcount                                          \n\
 116         popfq                                                   \n\
 117         popq    %r9                                             \n\
 118         popq    %r8                                             \n\
 119         popq    %rdi                                            \n\
 120         popq    %rsi                                            \n\
 121         popq    %rcx                                            \n\
 122         popq    %rdx                                            \n\
 123         popq    %rax                                            \n\
 124 .mcount_exit:                                                   \n\
 125         ret     $0                                              \n\
 126 ");
 127 #else /* !__GNUCLIKE_ASM */
 128 #error "this file needs to be ported to your compiler"
 129 #endif /* __GNUCLIKE_ASM */
 130
 131 #ifdef GUPROF
 132 /*
 133  * [.]mexitcount saves the return register(s), loads selfpc and calls
 134  * mexitcount(selfpc) to do the work.  Someday it should be in a machine
 135  * dependent file together with cputime(), __mcount and [.]mcount.  cputime()
 136  * can't just be put in machdep.c because it has to be compiled without -pg.
 137  */
 138 #ifdef __GNUCLIKE_ASM
 139 __asm("                                                         \n\
 140         .text                                                   \n\
 141 #                                                               \n\
 142 # Dummy label to be seen when gprof -u hides [.]mexitcount.     \n\
 143 #                                                               \n\
 144         .p2align 4,0x90                                         \n\
 145         .globl  __mexitcount                                    \n\
 146         .type   __mexitcount,@function                          \n\
 147 __mexitcount:                                                   \n\
 148         nop                                                     \n\
 149                                                                 \n\
 150 GMON_PROF_HIRES =       4                                       \n\
 151                                                                 \n\
 152         .p2align 4,0x90                                         \n\
 153         .globl  .mexitcount                                     \n\
 154 .mexitcount:                                                    \n\
 155         cmpl    $GMON_PROF_HIRES,_gmonparam+GM_STATE            \n\
 156         jne     .mexitcount_exit                                \n\
 157         pushq   %rax                                            \n\
 158         pushq   %rdx                                            \n\
 159         pushq   %rcx                                            \n\
 160         pushq   %rsi                                            \n\
 161         pushq   %rdi                                            \n\
 162         pushq   %r8                                             \n\
 163         pushq   %r9                                             \n\
 164         movq    7*8(%rsp),%rdi                                  \n\
 165         pushfq                                                  \n\
 166         cli                                                     \n\
 167         call    mexitcount                                      \n\
 168         popfq                                                   \n\
 169         popq    %r9                                             \n\
 170         popq    %r8                                             \n\
 171         popq    %rdi                                            \n\
 172         popq    %rsi                                            \n\
 173         popq    %rcx                                            \n\
 174         popq    %rdx                                            \n\
 175         popq    %rax                                            \n\
 176 .mexitcount_exit:                                               \n\
 177         ret     $0                                              \n\
 178 ");
 179 #endif /* __GNUCLIKE_ASM */
 180
 181 /*
 182  * Return the time elapsed since the last call.  The units are machine-
 183  * dependent.
 184  */
 185 int
 186 cputime()
 187 {
 188         u_int count;
 189         int delta;
 190         u_char high, low;
 191         static u_int prev_count;
 192
 193         if (cputime_clock == CPUTIME_CLOCK_TSC) {
 194                 /*
 195                  * Scale the TSC a little to make cputime()'s frequency
 196                  * fit in an int, assuming that the TSC frequency fits
 197                  * in a u_int.  Use a fixed scale since dynamic scaling
 198                  * would be slower and we can't really use the low bit
 199                  * of precision.
 200                  */
 201                 count = (u_int)rdtsc() & ~1u;
 202                 delta = (int)(count - prev_count) >> 1;
 203                 prev_count = count;
 204                 return (delta);
 205         }
 206
 207         /*
 208          * Read the current value of the 8254 timer counter 0.
 209          */
 210         outb(TIMER_MODE, TIMER_SEL0 | TIMER_LATCH);
 211         low = inb(TIMER_CNTR0);
 212         high = inb(TIMER_CNTR0);
 213         count = ((high << 8) | low) << CPUTIME_CLOCK_I8254_SHIFT;
 214
 215         /*
 216          * The timer counts down from TIMER_CNTR0_MAX to 0 and then resets.
 217          * While profiling is enabled, this routine is called at least twice
 218          * per timer reset (for mcounting and mexitcounting hardclock()),
 219          * so at most one reset has occurred since the last call, and one
 220          * has occurred iff the current count is larger than the previous
 221          * count.  This allows counter underflow to be detected faster
 222          * than in microtime().
 223          */
 224         delta = prev_count - count;
 225         prev_count = count;
 226         if ((int) delta <= 0)
 227                 return (delta + (i8254_max_count << CPUTIME_CLOCK_I8254_SHIFT));
 228         return (delta);
 229 }
 230
 231 static int
 232 sysctl_machdep_cputime_clock(SYSCTL_HANDLER_ARGS)
 233 {
 234         int clock;
 235         int error;
 236
 237         clock = cputime_clock;
 238         error = sysctl_handle_opaque(oidp, &clock, sizeof clock, req);
 239         if (error == 0 && req->newptr != NULL) {
 240                 if (clock < 0 || clock > CPUTIME_CLOCK_TSC)
 241                         return (EINVAL);
 242                 cputime_clock = clock;
 243         }
 244         return (error);
 245 }
 246
 247 SYSCTL_PROC(_machdep, OID_AUTO, cputime_clock, CTLTYPE_INT | CTLFLAG_RW,
 248             0, sizeof(u_int), sysctl_machdep_cputime_clock, "I", "");
 249
 250 /*
 251  * The start and stop routines need not be here since we turn off profiling
 252  * before calling them.  They are here for convenience.
 253  */
 254
 255 void
 256 startguprof(gp)
 257         struct gmonparam *gp;
 258 {
 259         uint64_t freq;
 260
 261         freq = atomic_load_acq_64(&tsc_freq);
 262         if (cputime_clock == CPUTIME_CLOCK_UNINITIALIZED) {
 263                 if (freq != 0 && mp_ncpus == 1)
 264                         cputime_clock = CPUTIME_CLOCK_TSC;
 265                 else
 266                         cputime_clock = CPUTIME_CLOCK_I8254;
 267         }
 268         if (cputime_clock == CPUTIME_CLOCK_TSC) {
 269                 gp->profrate = freq >> 1;
 270                 cputime_prof_active = 1;
 271         } else
 272                 gp->profrate = i8254_freq << CPUTIME_CLOCK_I8254_SHIFT;
 273         cputime_bias = 0;
 274         cputime();
 275 }
 276
 277 void
 278 stopguprof(gp)
 279         struct gmonparam *gp;
 280 {
 281         if (cputime_clock == CPUTIME_CLOCK_TSC)
 282                 cputime_prof_active = 0;
 283 }
 284
 285 /* If the cpu frequency changed while profiling, report a warning. */
 286 static void
 287 tsc_freq_changed(void *arg, const struct cf_level *level, int status)
 288 {
 289
 290         /*
 291          * If there was an error during the transition or
 292          * TSC is P-state invariant, don't do anything.
 293          */
 294         if (status != 0 || tsc_is_invariant)
 295                 return;
 296         if (cputime_prof_active && cputime_clock == CPUTIME_CLOCK_TSC)
 297                 printf("warning: cpu freq changed while profiling active\n");
 298 }
 299
 300 EVENTHANDLER_DEFINE(cpufreq_post_change, tsc_freq_changed, NULL,
 301     EVENTHANDLER_PRI_ANY);
 302
 303 #endif /* GUPROF */