prof_machdep.c revision 46548
119000Sbde/*- 219000Sbde * Copyright (c) 1996 Bruce D. Evans. 319000Sbde * All rights reserved. 415146Swollman * 519000Sbde * Redistribution and use in source and binary forms, with or without 619000Sbde * modification, are permitted provided that the following conditions 719000Sbde * are met: 819000Sbde * 1. Redistributions of source code must retain the above copyright 919000Sbde * notice, this list of conditions and the following disclaimer. 1019000Sbde * 2. Redistributions in binary form must reproduce the above copyright 1119000Sbde * notice, this list of conditions and the following disclaimer in the 1219000Sbde * documentation and/or other materials provided with the distribution. 1319000Sbde * 1419000Sbde * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 1519000Sbde * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 1619000Sbde * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 1719000Sbde * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 1819000Sbde * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 1919000Sbde * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 2019000Sbde * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 2119000Sbde * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 2219000Sbde * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 2319000Sbde * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 2419000Sbde * SUCH DAMAGE. 2519000Sbde * 2646548Sbde * $Id: prof_machdep.c,v 1.11 1998/12/14 18:21:34 bde Exp $ 2715146Swollman */ 2819000Sbde 2919000Sbde#ifdef GUPROF 3019000Sbde#include "opt_i586_guprof.h" 3119000Sbde#include "opt_perfmon.h" 3219000Sbde 3313107Sbde#include <sys/param.h> 3413107Sbde#include <sys/systm.h> 3519000Sbde#include <sys/gmon.h> 3631395Sbde#include <sys/kernel.h> 3731395Sbde#include <sys/sysctl.h> 3819000Sbde 3913107Sbde#include <machine/clock.h> 4019000Sbde#include <machine/perfmon.h> 4119000Sbde#include <machine/profile.h> 4246548Sbde#undef MCOUNT 4319000Sbde#endif 4419000Sbde 4546548Sbde#include <machine/asmacros.h> 4646548Sbde 4719269Sasami#ifdef PC98 4819269Sasami#include <pc98/pc98/pc98.h> 4919269Sasami#else 5013107Sbde#include <i386/isa/isa.h> 5119269Sasami#endif 5213107Sbde#include <i386/isa/timerreg.h> 5313107Sbde 5413107Sbde#ifdef GUPROF 5519000Sbde#define CPUTIME_CLOCK_UNINITIALIZED 0 5619000Sbde#define CPUTIME_CLOCK_I8254 1 5732005Sphk#define CPUTIME_CLOCK_TSC 2 5819000Sbde#define CPUTIME_CLOCK_I586_PMC 3 5919000Sbde#define CPUTIME_CLOCK_I8254_SHIFT 7 6019000Sbde 6119000Sbdeint cputime_bias = 1; /* initialize for locality of reference */ 6219000Sbde 6319000Sbdestatic int cputime_clock = CPUTIME_CLOCK_UNINITIALIZED; 6419000Sbde#ifdef I586_PMC_GUPROF 6519000Sbdestatic u_int cputime_clock_pmc_conf = I586_PMC_GUPROF; 6619000Sbdestatic int cputime_clock_pmc_init; 6719000Sbdestatic struct gmonparam saved_gmp; 6813107Sbde#endif 6919000Sbde#endif /* GUPROF */ 7013107Sbde 7113107Sbde#ifdef __GNUC__ 7235303Sbde__asm(" \n\ 7335303SbdeGM_STATE = 0 \n\ 7435303SbdeGMON_PROF_OFF = 3 \n\ 7535303Sbde \n\ 7635303Sbde .text \n\ 7735303Sbde .align 4,0x90 \n\ 7835303Sbde .globl __mcount \n\ 7946548Sbde .type __mcount,@function \n\ 8035303Sbde__mcount: \n\ 8135303Sbde # \n\ 8235303Sbde # Check that we are profiling. Do it early for speed. \n\ 8335303Sbde # \n\ 8446548Sbde cmpl $GMON_PROF_OFF," __XSTRING(CNAME(_gmonparam)) "+GM_STATE \n\ 8546548Sbde je .mcount_exit \n\ 8635303Sbde # \n\ 8746548Sbde # __mcount is the same as [.]mcount except the caller \n\ 8835303Sbde # hasn't changed the stack except to call here, so the \n\ 8935303Sbde # caller's raddr is above our raddr. \n\ 9035303Sbde # \n\ 9135303Sbde movl 4(%esp),%edx \n\ 9246548Sbde jmp .got_frompc \n\ 9335303Sbde \n\ 9435303Sbde .align 4,0x90 \n\ 9546548Sbde .globl " __XSTRING(HIDENAME(mcount)) " \n\ 9646548Sbde" __XSTRING(HIDENAME(mcount)) ": \n\ 9746548Sbde cmpl $GMON_PROF_OFF," __XSTRING(CNAME(_gmonparam)) "+GM_STATE \n\ 9846548Sbde je .mcount_exit \n\ 9935303Sbde # \n\ 10035303Sbde # The caller's stack frame has already been built, so \n\ 10135303Sbde # %ebp is the caller's frame pointer. The caller's \n\ 10235303Sbde # raddr is in the caller's frame following the caller's \n\ 10335303Sbde # caller's frame pointer. \n\ 10435303Sbde # \n\ 10535303Sbde movl 4(%ebp),%edx \n\ 10646548Sbde.got_frompc: \n\ 10735303Sbde # \n\ 10835303Sbde # Our raddr is the caller's pc. \n\ 10935303Sbde # \n\ 11035303Sbde movl (%esp),%eax \n\ 11135303Sbde \n\ 11235303Sbde pushfl \n\ 11335303Sbde pushl %eax \n\ 11435303Sbde pushl %edx \n\ 11535303Sbde cli \n\ 11646548Sbde call " __XSTRING(CNAME(mcount)) " \n\ 11735303Sbde addl $8,%esp \n\ 11835303Sbde popfl \n\ 11946548Sbde.mcount_exit: \n\ 12035303Sbde ret \n\ 12113107Sbde"); 12213107Sbde#else /* !__GNUC__ */ 12313107Sbde#error 12413107Sbde#endif /* __GNUC__ */ 12513107Sbde 12613107Sbde#ifdef GUPROF 12713107Sbde/* 12846548Sbde * [.]mexitcount saves the return register(s), loads selfpc and calls 12913107Sbde * mexitcount(selfpc) to do the work. Someday it should be in a machine 13046548Sbde * dependent file together with cputime(), __mcount and [.]mcount. cputime() 13113107Sbde * can't just be put in machdep.c because it has to be compiled without -pg. 13213107Sbde */ 13313107Sbde#ifdef __GNUC__ 13435303Sbde__asm(" \n\ 13535303Sbde .text \n\ 13635303Sbde# \n\ 13746548Sbde# Dummy label to be seen when gprof -u hides [.]mexitcount. \n\ 13835303Sbde# \n\ 13935303Sbde .align 4,0x90 \n\ 14035303Sbde .globl __mexitcount \n\ 14146548Sbde .type __mexitcount,@function \n\ 14235303Sbde__mexitcount: \n\ 14335303Sbde nop \n\ 14435303Sbde \n\ 14535303SbdeGMON_PROF_HIRES = 4 \n\ 14635303Sbde \n\ 14735303Sbde .align 4,0x90 \n\ 14846548Sbde .globl " __XSTRING(HIDENAME(mexitcount)) " \n\ 14946548Sbde" __XSTRING(HIDENAME(mexitcount)) ": \n\ 15046548Sbde cmpl $GMON_PROF_HIRES," __XSTRING(CNAME(_gmonparam)) "+GM_STATE \n\ 15146548Sbde jne .mexitcount_exit \n\ 15235303Sbde pushl %edx \n\ 15335303Sbde pushl %eax \n\ 15435303Sbde movl 8(%esp),%eax \n\ 15535303Sbde pushfl \n\ 15635303Sbde pushl %eax \n\ 15735303Sbde cli \n\ 15846548Sbde call " __XSTRING(CNAME(mexitcount)) " \n\ 15935303Sbde addl $4,%esp \n\ 16035303Sbde popfl \n\ 16135303Sbde popl %eax \n\ 16235303Sbde popl %edx \n\ 16346548Sbde.mexitcount_exit: \n\ 16435303Sbde ret \n\ 16513107Sbde"); 16613107Sbde#else /* !__GNUC__ */ 16713107Sbde#error 16813107Sbde#endif /* __GNUC__ */ 16913107Sbde 17013107Sbde/* 17113107Sbde * Return the time elapsed since the last call. The units are machine- 17213107Sbde * dependent. 17313107Sbde */ 17419000Sbdeint 17513107Sbdecputime() 17613107Sbde{ 17713107Sbde u_int count; 17819000Sbde int delta; 17941794Sbde#if (defined(I586_CPU) || defined(I686_CPU)) && !defined(SMP) && \ 18041794Sbde defined(PERFMON) && defined(I586_PMC_GUPROF) 18119000Sbde u_quad_t event_count; 18219000Sbde#endif 18319000Sbde u_char high, low; 18413107Sbde static u_int prev_count; 18513107Sbde 18631395Sbde#if (defined(I586_CPU) || defined(I686_CPU)) && !defined(SMP) 18732005Sphk if (cputime_clock == CPUTIME_CLOCK_TSC) { 18819000Sbde count = (u_int)rdtsc(); 18919000Sbde delta = (int)(count - prev_count); 19019000Sbde prev_count = count; 19119000Sbde return (delta); 19219000Sbde } 19331395Sbde#if defined(PERFMON) && defined(I586_PMC_GUPROF) 19419000Sbde if (cputime_clock == CPUTIME_CLOCK_I586_PMC) { 19519000Sbde /* 19619000Sbde * XXX permon_read() should be inlined so that the 19719000Sbde * perfmon module doesn't need to be compiled with 19819000Sbde * profiling disabled and so that it is fast. 19919000Sbde */ 20019000Sbde perfmon_read(0, &event_count); 20119000Sbde 20219000Sbde count = (u_int)event_count; 20319000Sbde delta = (int)(count - prev_count); 20419000Sbde prev_count = count; 20519000Sbde return (delta); 20619000Sbde } 20731395Sbde#endif /* PERFMON && I586_PMC_GUPROF */ 20831395Sbde#endif /* (I586_CPU || I686_CPU) && !SMP */ 20919000Sbde 21013107Sbde /* 21113107Sbde * Read the current value of the 8254 timer counter 0. 21213107Sbde */ 21313107Sbde outb(TIMER_MODE, TIMER_SEL0 | TIMER_LATCH); 21413107Sbde low = inb(TIMER_CNTR0); 21519000Sbde high = inb(TIMER_CNTR0); 21619000Sbde count = ((high << 8) | low) << CPUTIME_CLOCK_I8254_SHIFT; 21713107Sbde 21813107Sbde /* 21913107Sbde * The timer counts down from TIMER_CNTR0_MAX to 0 and then resets. 22013107Sbde * While profiling is enabled, this routine is called at least twice 22113107Sbde * per timer reset (for mcounting and mexitcounting hardclock()), 22213107Sbde * so at most one reset has occurred since the last call, and one 22313107Sbde * has occurred iff the current count is larger than the previous 22413107Sbde * count. This allows counter underflow to be detected faster 22513107Sbde * than in microtime(). 22613107Sbde */ 22713107Sbde delta = prev_count - count; 22813107Sbde prev_count = count; 22913107Sbde if ((int) delta <= 0) 23019000Sbde return (delta + (timer0_max_count << CPUTIME_CLOCK_I8254_SHIFT)); 23113107Sbde return (delta); 23213107Sbde} 23319000Sbde 23431395Sbdestatic int 23531395Sbdesysctl_machdep_cputime_clock SYSCTL_HANDLER_ARGS 23631395Sbde{ 23731395Sbde int clock; 23841794Sbde int error; 23941794Sbde#if defined(PERFMON) && defined(I586_PMC_GUPROF) 24031395Sbde int event; 24131395Sbde struct pmc pmc; 24241794Sbde#endif 24331395Sbde 24431395Sbde clock = cputime_clock; 24531395Sbde#if defined(PERFMON) && defined(I586_PMC_GUPROF) 24631395Sbde if (clock == CPUTIME_CLOCK_I586_PMC) { 24731395Sbde pmc.pmc_val = cputime_clock_pmc_conf; 24831395Sbde clock += pmc.pmc_event; 24931395Sbde } 25031395Sbde#endif 25131395Sbde error = sysctl_handle_opaque(oidp, &clock, sizeof clock, req); 25231395Sbde if (error == 0 && req->newptr != NULL) { 25331395Sbde#if defined(PERFMON) && defined(I586_PMC_GUPROF) 25431395Sbde if (clock >= CPUTIME_CLOCK_I586_PMC) { 25531395Sbde event = clock - CPUTIME_CLOCK_I586_PMC; 25631395Sbde if (event >= 256) 25731395Sbde return (EINVAL); 25831395Sbde pmc.pmc_num = 0; 25931395Sbde pmc.pmc_event = event; 26031395Sbde pmc.pmc_unit = 0; 26131395Sbde pmc.pmc_flags = PMCF_E | PMCF_OS | PMCF_USR; 26231395Sbde pmc.pmc_mask = 0; 26331395Sbde cputime_clock_pmc_conf = pmc.pmc_val; 26431395Sbde cputime_clock = CPUTIME_CLOCK_I586_PMC; 26531395Sbde } else 26631395Sbde#endif 26731395Sbde { 26831395Sbde if (clock < 0 || clock >= CPUTIME_CLOCK_I586_PMC) 26931395Sbde return (EINVAL); 27031395Sbde cputime_clock = clock; 27131395Sbde } 27231395Sbde } 27331395Sbde return (error); 27431395Sbde} 27531395Sbde 27631395SbdeSYSCTL_PROC(_machdep, OID_AUTO, cputime_clock, CTLTYPE_INT | CTLFLAG_RW, 27731395Sbde 0, sizeof(u_int), sysctl_machdep_cputime_clock, "I", ""); 27831395Sbde 27919000Sbde/* 28019000Sbde * The start and stop routines need not be here since we turn off profiling 28119000Sbde * before calling them. They are here for convenience. 28219000Sbde */ 28319000Sbde 28419000Sbdevoid 28519000Sbdestartguprof(gp) 28619000Sbde struct gmonparam *gp; 28719000Sbde{ 28819000Sbde if (cputime_clock == CPUTIME_CLOCK_UNINITIALIZED) { 28919000Sbde cputime_clock = CPUTIME_CLOCK_I8254; 29031395Sbde#if (defined(I586_CPU) || defined(I686_CPU)) && !defined(SMP) 29132005Sphk if (tsc_freq != 0) 29232005Sphk cputime_clock = CPUTIME_CLOCK_TSC; 29319000Sbde#endif 29419000Sbde } 29519000Sbde gp->profrate = timer_freq << CPUTIME_CLOCK_I8254_SHIFT; 29631395Sbde#if (defined(I586_CPU) || defined(I686_CPU)) && !defined(SMP) 29732005Sphk if (cputime_clock == CPUTIME_CLOCK_TSC) 29832005Sphk gp->profrate = tsc_freq; 29931395Sbde#if defined(PERFMON) && defined(I586_PMC_GUPROF) 30019000Sbde else if (cputime_clock == CPUTIME_CLOCK_I586_PMC) { 30119000Sbde if (perfmon_avail() && 30219000Sbde perfmon_setup(0, cputime_clock_pmc_conf) == 0) { 30319000Sbde if (perfmon_start(0) != 0) 30419000Sbde perfmon_fini(0); 30519000Sbde else { 30619000Sbde /* XXX 1 event == 1 us. */ 30719000Sbde gp->profrate = 1000000; 30819000Sbde 30919000Sbde saved_gmp = *gp; 31019000Sbde 31119000Sbde /* Zap overheads. They are invalid. */ 31219000Sbde gp->cputime_overhead = 0; 31319000Sbde gp->mcount_overhead = 0; 31419000Sbde gp->mcount_post_overhead = 0; 31519000Sbde gp->mcount_pre_overhead = 0; 31619000Sbde gp->mexitcount_overhead = 0; 31719000Sbde gp->mexitcount_post_overhead = 0; 31819000Sbde gp->mexitcount_pre_overhead = 0; 31919000Sbde 32019000Sbde cputime_clock_pmc_init = TRUE; 32119000Sbde } 32219000Sbde } 32319000Sbde } 32431395Sbde#endif /* PERFMON && I586_PMC_GUPROF */ 32531395Sbde#endif /* (I586_CPU || I686_CPU) && !SMP */ 32619000Sbde cputime_bias = 0; 32719000Sbde cputime(); 32819000Sbde} 32919000Sbde 33019000Sbdevoid 33119000Sbdestopguprof(gp) 33219000Sbde struct gmonparam *gp; 33319000Sbde{ 33419000Sbde#if defined(PERFMON) && defined(I586_PMC_GUPROF) 33519000Sbde if (cputime_clock_pmc_init) { 33619000Sbde *gp = saved_gmp; 33719000Sbde perfmon_fini(0); 33819000Sbde cputime_clock_pmc_init = FALSE; 33919000Sbde } 34019000Sbde#endif 34119000Sbde} 34219000Sbde 34319000Sbde#else /* !GUPROF */ 34413107Sbde#ifdef __GNUC__ 34535303Sbde__asm(" \n\ 34635303Sbde .text \n\ 34735303Sbde .align 4,0x90 \n\ 34846548Sbde .globl " __XSTRING(HIDENAME(mexitcount)) " \n\ 34946548Sbde" __XSTRING(HIDENAME(mexitcount)) ": \n\ 35035303Sbde ret \n\ 35113107Sbde"); 35213107Sbde#else /* !__GNUC__ */ 35313107Sbde#error 35413107Sbde#endif /* __GNUC__ */ 35513107Sbde#endif /* GUPROF */ 356