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