d5c90663b2
since the branch caches on at least Athlon XP through Athlon 64 CPU's
don't understand such instructions and guarantee a cache miss taking
at least 10 cycles. Use the documented workaround "ret $0" instead
("nop; ret" also works, but "ret $0" is probably faster on old CPUs).
Normal code (even asm code) doesn't branch to "ret", since there is
usually some cleanup to do, but the __mcount, .mcount and .mexitcount
entry points were optimized too well to have the minimum number of
instructions (3 instructions each if profiling is not enabled) and
they did this. I didn't see a significant number of cache misses for
.mexitcount, but for the shared "ret" for __mcount and .mcount I
observed cache misses costing 26 cycles each. For a send(2) syscall
that makes about 70 function calls, the cost of these cache misses
alone increased the syscall time from about 4000 cycles to about 7000
cycles. 4000 is for a profiling (GUPROF) kernel with profiling disabled;
after this fix, configuring profiling only costs about 600 cycles in the
4000, which is consistent with almost perfect branch prediction in the
mcounting calls.
368 lines
10 KiB
C
368 lines
10 KiB
C
/*-
|
|
* Copyright (c) 1996 Bruce D. Evans.
|
|
* All rights reserved.
|
|
*
|
|
* Redistribution and use in source and binary forms, with or without
|
|
* modification, are permitted provided that the following conditions
|
|
* are met:
|
|
* 1. Redistributions of source code must retain the above copyright
|
|
* notice, this list of conditions and the following disclaimer.
|
|
* 2. Redistributions in binary form must reproduce the above copyright
|
|
* notice, this list of conditions and the following disclaimer in the
|
|
* documentation and/or other materials provided with the distribution.
|
|
*
|
|
* THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
|
|
* ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
|
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
|
|
* ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
|
|
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
|
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
|
|
* OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
|
|
* HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
|
|
* LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
|
|
* OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
|
|
* SUCH DAMAGE.
|
|
*/
|
|
|
|
#include <sys/cdefs.h>
|
|
__FBSDID("$FreeBSD$");
|
|
|
|
#ifdef GUPROF
|
|
#include "opt_i586_guprof.h"
|
|
#include "opt_perfmon.h"
|
|
|
|
#include <sys/param.h>
|
|
#include <sys/systm.h>
|
|
#include <sys/bus.h>
|
|
#include <sys/cpu.h>
|
|
#include <sys/eventhandler.h>
|
|
#include <sys/gmon.h>
|
|
#include <sys/kernel.h>
|
|
#include <sys/smp.h>
|
|
#include <sys/sysctl.h>
|
|
|
|
#include <machine/clock.h>
|
|
#include <machine/perfmon.h>
|
|
#include <machine/timerreg.h>
|
|
|
|
#define CPUTIME_CLOCK_UNINITIALIZED 0
|
|
#define CPUTIME_CLOCK_I8254 1
|
|
#define CPUTIME_CLOCK_TSC 2
|
|
#define CPUTIME_CLOCK_I586_PMC 3
|
|
#define CPUTIME_CLOCK_I8254_SHIFT 7
|
|
|
|
int cputime_bias = 1; /* initialize for locality of reference */
|
|
|
|
static int cputime_clock = CPUTIME_CLOCK_UNINITIALIZED;
|
|
#if defined(PERFMON) && defined(I586_PMC_GUPROF)
|
|
static u_int cputime_clock_pmc_conf = I586_PMC_GUPROF;
|
|
static int cputime_clock_pmc_init;
|
|
static struct gmonparam saved_gmp;
|
|
#endif
|
|
#if defined(I586_CPU) || defined(I686_CPU)
|
|
static int cputime_prof_active;
|
|
#endif
|
|
#endif /* GUPROF */
|
|
|
|
#ifdef __GNUCLIKE_ASM
|
|
__asm(" \n\
|
|
GM_STATE = 0 \n\
|
|
GMON_PROF_OFF = 3 \n\
|
|
\n\
|
|
.text \n\
|
|
.p2align 4,0x90 \n\
|
|
.globl __mcount \n\
|
|
.type __mcount,@function \n\
|
|
__mcount: \n\
|
|
# \n\
|
|
# Check that we are profiling. Do it early for speed. \n\
|
|
# \n\
|
|
cmpl $GMON_PROF_OFF,_gmonparam+GM_STATE \n\
|
|
je .mcount_exit \n\
|
|
# \n\
|
|
# __mcount is the same as [.]mcount except the caller \n\
|
|
# hasn't changed the stack except to call here, so the \n\
|
|
# caller's raddr is above our raddr. \n\
|
|
# \n\
|
|
movl 4(%esp),%edx \n\
|
|
jmp .got_frompc \n\
|
|
\n\
|
|
.p2align 4,0x90 \n\
|
|
.globl .mcount \n\
|
|
.mcount: \n\
|
|
cmpl $GMON_PROF_OFF,_gmonparam+GM_STATE \n\
|
|
je .mcount_exit \n\
|
|
# \n\
|
|
# The caller's stack frame has already been built, so \n\
|
|
# %ebp is the caller's frame pointer. The caller's \n\
|
|
# raddr is in the caller's frame following the caller's \n\
|
|
# caller's frame pointer. \n\
|
|
# \n\
|
|
movl 4(%ebp),%edx \n\
|
|
.got_frompc: \n\
|
|
# \n\
|
|
# Our raddr is the caller's pc. \n\
|
|
# \n\
|
|
movl (%esp),%eax \n\
|
|
\n\
|
|
pushfl \n\
|
|
pushl %eax \n\
|
|
pushl %edx \n\
|
|
cli \n\
|
|
call mcount \n\
|
|
addl $8,%esp \n\
|
|
popfl \n\
|
|
.mcount_exit: \n\
|
|
ret $0 \n\
|
|
");
|
|
#else /* !__GNUCLIKE_ASM */
|
|
#error "this file needs to be ported to your compiler"
|
|
#endif /* __GNUCLIKE_ASM */
|
|
|
|
#ifdef GUPROF
|
|
/*
|
|
* [.]mexitcount saves the return register(s), loads selfpc and calls
|
|
* mexitcount(selfpc) to do the work. Someday it should be in a machine
|
|
* dependent file together with cputime(), __mcount and [.]mcount. cputime()
|
|
* can't just be put in machdep.c because it has to be compiled without -pg.
|
|
*/
|
|
#ifdef __GNUCLIKE_ASM
|
|
__asm(" \n\
|
|
.text \n\
|
|
# \n\
|
|
# Dummy label to be seen when gprof -u hides [.]mexitcount. \n\
|
|
# \n\
|
|
.p2align 4,0x90 \n\
|
|
.globl __mexitcount \n\
|
|
.type __mexitcount,@function \n\
|
|
__mexitcount: \n\
|
|
nop \n\
|
|
\n\
|
|
GMON_PROF_HIRES = 4 \n\
|
|
\n\
|
|
.p2align 4,0x90 \n\
|
|
.globl .mexitcount \n\
|
|
.mexitcount: \n\
|
|
cmpl $GMON_PROF_HIRES,_gmonparam+GM_STATE \n\
|
|
jne .mexitcount_exit \n\
|
|
pushl %edx \n\
|
|
pushl %eax \n\
|
|
movl 8(%esp),%eax \n\
|
|
pushfl \n\
|
|
pushl %eax \n\
|
|
cli \n\
|
|
call mexitcount \n\
|
|
addl $4,%esp \n\
|
|
popfl \n\
|
|
popl %eax \n\
|
|
popl %edx \n\
|
|
.mexitcount_exit: \n\
|
|
ret $0 \n\
|
|
");
|
|
#endif /* __GNUCLIKE_ASM */
|
|
|
|
/*
|
|
* Return the time elapsed since the last call. The units are machine-
|
|
* dependent.
|
|
*/
|
|
int
|
|
cputime()
|
|
{
|
|
u_int count;
|
|
int delta;
|
|
#if (defined(I586_CPU) || defined(I686_CPU)) && !defined(SMP) && \
|
|
defined(PERFMON) && defined(I586_PMC_GUPROF)
|
|
u_quad_t event_count;
|
|
#endif
|
|
u_char high, low;
|
|
static u_int prev_count;
|
|
|
|
#if defined(I586_CPU) || defined(I686_CPU)
|
|
if (cputime_clock == CPUTIME_CLOCK_TSC) {
|
|
/*
|
|
* Scale the TSC a little to make cputime()'s frequency
|
|
* fit in an int, assuming that the TSC frequency fits
|
|
* in a u_int. Use a fixed scale since dynamic scaling
|
|
* would be slower and we can't really use the low bit
|
|
* of precision.
|
|
*/
|
|
count = (u_int)rdtsc() & ~1u;
|
|
delta = (int)(count - prev_count) >> 1;
|
|
prev_count = count;
|
|
return (delta);
|
|
}
|
|
#if defined(PERFMON) && defined(I586_PMC_GUPROF) && !defined(SMP)
|
|
if (cputime_clock == CPUTIME_CLOCK_I586_PMC) {
|
|
/*
|
|
* XXX permon_read() should be inlined so that the
|
|
* perfmon module doesn't need to be compiled with
|
|
* profiling disabled and so that it is fast.
|
|
*/
|
|
perfmon_read(0, &event_count);
|
|
|
|
count = (u_int)event_count;
|
|
delta = (int)(count - prev_count);
|
|
prev_count = count;
|
|
return (delta);
|
|
}
|
|
#endif /* PERFMON && I586_PMC_GUPROF && !SMP */
|
|
#endif /* I586_CPU || I686_CPU */
|
|
|
|
/*
|
|
* Read the current value of the 8254 timer counter 0.
|
|
*/
|
|
outb(TIMER_MODE, TIMER_SEL0 | TIMER_LATCH);
|
|
low = inb(TIMER_CNTR0);
|
|
high = inb(TIMER_CNTR0);
|
|
count = ((high << 8) | low) << CPUTIME_CLOCK_I8254_SHIFT;
|
|
|
|
/*
|
|
* The timer counts down from TIMER_CNTR0_MAX to 0 and then resets.
|
|
* While profiling is enabled, this routine is called at least twice
|
|
* per timer reset (for mcounting and mexitcounting hardclock()),
|
|
* so at most one reset has occurred since the last call, and one
|
|
* has occurred iff the current count is larger than the previous
|
|
* count. This allows counter underflow to be detected faster
|
|
* than in microtime().
|
|
*/
|
|
delta = prev_count - count;
|
|
prev_count = count;
|
|
if ((int) delta <= 0)
|
|
return (delta + (timer0_max_count << CPUTIME_CLOCK_I8254_SHIFT));
|
|
return (delta);
|
|
}
|
|
|
|
static int
|
|
sysctl_machdep_cputime_clock(SYSCTL_HANDLER_ARGS)
|
|
{
|
|
int clock;
|
|
int error;
|
|
#if defined(PERFMON) && defined(I586_PMC_GUPROF)
|
|
int event;
|
|
struct pmc pmc;
|
|
#endif
|
|
|
|
clock = cputime_clock;
|
|
#if defined(PERFMON) && defined(I586_PMC_GUPROF)
|
|
if (clock == CPUTIME_CLOCK_I586_PMC) {
|
|
pmc.pmc_val = cputime_clock_pmc_conf;
|
|
clock += pmc.pmc_event;
|
|
}
|
|
#endif
|
|
error = sysctl_handle_opaque(oidp, &clock, sizeof clock, req);
|
|
if (error == 0 && req->newptr != NULL) {
|
|
#if defined(PERFMON) && defined(I586_PMC_GUPROF)
|
|
if (clock >= CPUTIME_CLOCK_I586_PMC) {
|
|
event = clock - CPUTIME_CLOCK_I586_PMC;
|
|
if (event >= 256)
|
|
return (EINVAL);
|
|
pmc.pmc_num = 0;
|
|
pmc.pmc_event = event;
|
|
pmc.pmc_unit = 0;
|
|
pmc.pmc_flags = PMCF_E | PMCF_OS | PMCF_USR;
|
|
pmc.pmc_mask = 0;
|
|
cputime_clock_pmc_conf = pmc.pmc_val;
|
|
cputime_clock = CPUTIME_CLOCK_I586_PMC;
|
|
} else
|
|
#endif
|
|
{
|
|
if (clock < 0 || clock >= CPUTIME_CLOCK_I586_PMC)
|
|
return (EINVAL);
|
|
cputime_clock = clock;
|
|
}
|
|
}
|
|
return (error);
|
|
}
|
|
|
|
SYSCTL_PROC(_machdep, OID_AUTO, cputime_clock, CTLTYPE_INT | CTLFLAG_RW,
|
|
0, sizeof(u_int), sysctl_machdep_cputime_clock, "I", "");
|
|
|
|
/*
|
|
* The start and stop routines need not be here since we turn off profiling
|
|
* before calling them. They are here for convenience.
|
|
*/
|
|
|
|
void
|
|
startguprof(gp)
|
|
struct gmonparam *gp;
|
|
{
|
|
if (cputime_clock == CPUTIME_CLOCK_UNINITIALIZED) {
|
|
cputime_clock = CPUTIME_CLOCK_I8254;
|
|
#if defined(I586_CPU) || defined(I686_CPU)
|
|
if (tsc_freq != 0 && !tsc_is_broken && mp_ncpus == 1)
|
|
cputime_clock = CPUTIME_CLOCK_TSC;
|
|
#endif
|
|
}
|
|
gp->profrate = timer_freq << CPUTIME_CLOCK_I8254_SHIFT;
|
|
#if defined(I586_CPU) || defined(I686_CPU)
|
|
if (cputime_clock == CPUTIME_CLOCK_TSC) {
|
|
gp->profrate = tsc_freq >> 1;
|
|
cputime_prof_active = 1;
|
|
}
|
|
#if defined(PERFMON) && defined(I586_PMC_GUPROF)
|
|
else if (cputime_clock == CPUTIME_CLOCK_I586_PMC) {
|
|
if (perfmon_avail() &&
|
|
perfmon_setup(0, cputime_clock_pmc_conf) == 0) {
|
|
if (perfmon_start(0) != 0)
|
|
perfmon_fini(0);
|
|
else {
|
|
/* XXX 1 event == 1 us. */
|
|
gp->profrate = 1000000;
|
|
|
|
saved_gmp = *gp;
|
|
|
|
/* Zap overheads. They are invalid. */
|
|
gp->cputime_overhead = 0;
|
|
gp->mcount_overhead = 0;
|
|
gp->mcount_post_overhead = 0;
|
|
gp->mcount_pre_overhead = 0;
|
|
gp->mexitcount_overhead = 0;
|
|
gp->mexitcount_post_overhead = 0;
|
|
gp->mexitcount_pre_overhead = 0;
|
|
|
|
cputime_clock_pmc_init = TRUE;
|
|
}
|
|
}
|
|
}
|
|
#endif /* PERFMON && I586_PMC_GUPROF */
|
|
#endif /* I586_CPU || I686_CPU */
|
|
cputime_bias = 0;
|
|
cputime();
|
|
}
|
|
|
|
void
|
|
stopguprof(gp)
|
|
struct gmonparam *gp;
|
|
{
|
|
#if defined(PERFMON) && defined(I586_PMC_GUPROF)
|
|
if (cputime_clock_pmc_init) {
|
|
*gp = saved_gmp;
|
|
perfmon_fini(0);
|
|
cputime_clock_pmc_init = FALSE;
|
|
}
|
|
#endif
|
|
#if defined(I586_CPU) || defined(I686_CPU)
|
|
if (cputime_clock == CPUTIME_CLOCK_TSC)
|
|
cputime_prof_active = 0;
|
|
#endif
|
|
}
|
|
|
|
#if defined(I586_CPU) || defined(I686_CPU)
|
|
/* If the cpu frequency changed while profiling, report a warning. */
|
|
static void
|
|
tsc_freq_changed(void *arg, const struct cf_level *level, int status)
|
|
{
|
|
|
|
/* If there was an error during the transition, don't do anything. */
|
|
if (status != 0)
|
|
return;
|
|
if (cputime_prof_active && cputime_clock == CPUTIME_CLOCK_TSC)
|
|
printf("warning: cpu freq changed while profiling active\n");
|
|
}
|
|
|
|
EVENTHANDLER_DEFINE(cpufreq_post_change, tsc_freq_changed, NULL,
|
|
EVENTHANDLER_PRI_ANY);
|
|
#endif /* I586_CPU || I686_CPU */
|
|
|
|
#endif /* GUPROF */
|