113116Sbde/*- 213116Sbde * Copyright (c) 1983, 1992, 1993 313116Sbde * The Regents of the University of California. All rights reserved. 413116Sbde * 513116Sbde * Redistribution and use in source and binary forms, with or without 613116Sbde * modification, are permitted provided that the following conditions 713116Sbde * are met: 813116Sbde * 1. Redistributions of source code must retain the above copyright 913116Sbde * notice, this list of conditions and the following disclaimer. 1013116Sbde * 2. Redistributions in binary form must reproduce the above copyright 1113116Sbde * notice, this list of conditions and the following disclaimer in the 1213116Sbde * documentation and/or other materials provided with the distribution. 1313116Sbde * 4. Neither the name of the University nor the names of its contributors 1413116Sbde * may be used to endorse or promote products derived from this software 1513116Sbde * without specific prior written permission. 1613116Sbde * 1713116Sbde * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 1813116Sbde * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 1913116Sbde * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 2013116Sbde * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 2113116Sbde * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 2213116Sbde * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 2313116Sbde * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 2413116Sbde * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 2513116Sbde * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 2613116Sbde * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 2713116Sbde * SUCH DAMAGE. 2813116Sbde */ 291541Srgrimes 30116189Sobrien#include <sys/cdefs.h> 31116189Sobrien__FBSDID("$FreeBSD$"); 32116189Sobrien 3313116Sbde#include <sys/param.h> 3413116Sbde#include <sys/gmon.h> 3555206Speter#ifdef _KERNEL 3619169Sbde#ifndef GUPROF 3719169Sbde#include <sys/systm.h> 3819169Sbde#endif 3913116Sbde#include <vm/vm.h> 4013116Sbde#include <vm/vm_param.h> 4113116Sbde#include <vm/pmap.h> 4213116Sbde#endif 4313116Sbde 4413116Sbde/* 4513116Sbde * mcount is called on entry to each function compiled with the profiling 4613116Sbde * switch set. _mcount(), which is declared in a machine-dependent way 4713116Sbde * with _MCOUNT_DECL, does the actual work and is either inlined into a 4813116Sbde * C routine or called by an assembly stub. In any case, this magic is 4913116Sbde * taken care of by the MCOUNT definition in <machine/profile.h>. 5013116Sbde * 5113116Sbde * _mcount updates data structures that represent traversals of the 5213116Sbde * program's call graph edges. frompc and selfpc are the return 5313116Sbde * address and function address that represents the given call graph edge. 5413116Sbde * 5513116Sbde * Note: the original BSD code used the same variable (frompcindex) for 5613116Sbde * both frompcindex and frompc. Any reasonable, modern compiler will 5713116Sbde * perform this optimization. 5813116Sbde */ 5913116Sbde_MCOUNT_DECL(frompc, selfpc) /* _mcount; may be static, inline, etc */ 60134398Smarcel uintfptr_t frompc, selfpc; 6113116Sbde{ 6213116Sbde#ifdef GUPROF 6319000Sbde int delta; 6413116Sbde#endif 65134398Smarcel fptrdiff_t frompci; 66134398Smarcel u_short *frompcindex; 67134398Smarcel struct tostruct *top, *prevtop; 68134398Smarcel struct gmonparam *p; 69134398Smarcel long toindex; 7055206Speter#ifdef _KERNEL 7117879Sbde MCOUNT_DECL(s) 7213116Sbde#endif 7313116Sbde 7413116Sbde p = &_gmonparam; 7513116Sbde#ifndef GUPROF /* XXX */ 7613116Sbde /* 7713116Sbde * check that we are profiling 7813116Sbde * and that we aren't recursively invoked. 7913116Sbde */ 8013116Sbde if (p->state != GMON_PROF_ON) 8113116Sbde return; 8213116Sbde#endif 8355206Speter#ifdef _KERNEL 8417879Sbde MCOUNT_ENTER(s); 8513116Sbde#else 8613116Sbde p->state = GMON_PROF_BUSY; 8713116Sbde#endif 8813116Sbde 8955206Speter#ifdef _KERNEL 9013116Sbde /* 91134398Smarcel * When we are called from an exception handler, frompc may be 92134398Smarcel * a user address. Convert such frompc's to some representation 93134398Smarcel * in kernel address space. 9413116Sbde */ 95134398Smarcel frompc = MCOUNT_FROMPC_USER(frompc); 9655206Speter#endif 9713116Sbde 98134398Smarcel frompci = frompc - p->lowpc; 99134398Smarcel if (frompci >= p->textsize) 100134398Smarcel goto done; 101134398Smarcel 10213116Sbde#ifdef GUPROF 10319000Sbde if (p->state == GMON_PROF_HIRES) { 10419000Sbde /* 10519000Sbde * Count the time since cputime() was previously called 10619000Sbde * against `frompc'. Compensate for overheads. 10719000Sbde * 10819000Sbde * cputime() sets its prev_count variable to the count when 10919000Sbde * it is called. This in effect starts a counter for 11019000Sbde * the next period of execution (normally from now until 11119000Sbde * the next call to mcount() or mexitcount()). We set 11219000Sbde * cputime_bias to compensate for our own overhead. 11319000Sbde * 11419000Sbde * We use the usual sampling counters since they can be 11519000Sbde * located efficiently. 4-byte counters are usually 11619000Sbde * necessary. gprof will add up the scattered counts 11719000Sbde * just like it does for statistical profiling. All 11819000Sbde * counts are signed so that underflow in the subtractions 11919000Sbde * doesn't matter much (negative counts are normally 12019000Sbde * compensated for by larger counts elsewhere). Underflow 12119000Sbde * shouldn't occur, but may be caused by slightly wrong 12219000Sbde * calibrations or from not clearing cputime_bias. 12319000Sbde */ 12419000Sbde delta = cputime() - cputime_bias - p->mcount_pre_overhead; 12519000Sbde cputime_bias = p->mcount_post_overhead; 12619000Sbde KCOUNT(p, frompci) += delta; 12719000Sbde *p->cputime_count += p->cputime_overhead; 12819000Sbde *p->mcount_count += p->mcount_overhead; 12913116Sbde } 13013116Sbde#endif /* GUPROF */ 13113116Sbde 13255206Speter#ifdef _KERNEL 13313116Sbde /* 13413116Sbde * When we are called from an exception handler, frompc is faked 13513116Sbde * to be for where the exception occurred. We've just solidified 136134398Smarcel * the count for there. Now convert frompci to an index that 137134398Smarcel * represents the kind of exception so that interruptions appear 138134398Smarcel * in the call graph as calls from those index instead of calls 139134398Smarcel * from all over. 14013116Sbde */ 141134398Smarcel frompc = MCOUNT_FROMPC_INTR(selfpc); 142134398Smarcel if ((frompc - p->lowpc) < p->textsize) 143134398Smarcel frompci = frompc - p->lowpc; 14455206Speter#endif 14513116Sbde 14613116Sbde /* 14713116Sbde * check that frompc is a reasonable pc value. 14813116Sbde * for example: signal catchers get called from the stack, 14913116Sbde * not from text space. too bad. 15013116Sbde */ 15113116Sbde if (frompci >= p->textsize) 15213116Sbde goto done; 15313116Sbde 15413116Sbde frompcindex = &p->froms[frompci / (p->hashfraction * sizeof(*p->froms))]; 15513116Sbde toindex = *frompcindex; 15613116Sbde if (toindex == 0) { 15713116Sbde /* 15813116Sbde * first time traversing this arc 15913116Sbde */ 16013116Sbde toindex = ++p->tos[0].link; 16113116Sbde if (toindex >= p->tolimit) 16213116Sbde /* halt further profiling */ 16313116Sbde goto overflow; 16413116Sbde 16513116Sbde *frompcindex = toindex; 16613116Sbde top = &p->tos[toindex]; 16713116Sbde top->selfpc = selfpc; 16813116Sbde top->count = 1; 16913116Sbde top->link = 0; 17013116Sbde goto done; 17113116Sbde } 17213116Sbde top = &p->tos[toindex]; 17313116Sbde if (top->selfpc == selfpc) { 17413116Sbde /* 17513116Sbde * arc at front of chain; usual case. 17613116Sbde */ 17713116Sbde top->count++; 17813116Sbde goto done; 17913116Sbde } 18013116Sbde /* 18113116Sbde * have to go looking down chain for it. 18213116Sbde * top points to what we are looking at, 18313116Sbde * prevtop points to previous top. 18413116Sbde * we know it is not at the head of the chain. 18513116Sbde */ 18613116Sbde for (; /* goto done */; ) { 18713116Sbde if (top->link == 0) { 18813116Sbde /* 18913116Sbde * top is end of the chain and none of the chain 19013116Sbde * had top->selfpc == selfpc. 19113116Sbde * so we allocate a new tostruct 19213116Sbde * and link it to the head of the chain. 19313116Sbde */ 19413116Sbde toindex = ++p->tos[0].link; 19513116Sbde if (toindex >= p->tolimit) 19613116Sbde goto overflow; 19713116Sbde 19813116Sbde top = &p->tos[toindex]; 19913116Sbde top->selfpc = selfpc; 20013116Sbde top->count = 1; 20113116Sbde top->link = *frompcindex; 20213116Sbde *frompcindex = toindex; 20313116Sbde goto done; 20413116Sbde } 20513116Sbde /* 20613116Sbde * otherwise, check the next arc on the chain. 20713116Sbde */ 20813116Sbde prevtop = top; 20913116Sbde top = &p->tos[top->link]; 21013116Sbde if (top->selfpc == selfpc) { 21113116Sbde /* 21213116Sbde * there it is. 21313116Sbde * increment its count 21413116Sbde * move it to the head of the chain. 21513116Sbde */ 21613116Sbde top->count++; 21713116Sbde toindex = prevtop->link; 21813116Sbde prevtop->link = top->link; 21913116Sbde top->link = *frompcindex; 22013116Sbde *frompcindex = toindex; 22113116Sbde goto done; 22213116Sbde } 22313116Sbde 22413116Sbde } 22513116Sbdedone: 22655206Speter#ifdef _KERNEL 22717879Sbde MCOUNT_EXIT(s); 22813116Sbde#else 22913116Sbde p->state = GMON_PROF_ON; 23013116Sbde#endif 23113116Sbde return; 23213116Sbdeoverflow: 23313116Sbde p->state = GMON_PROF_ERROR; 23455206Speter#ifdef _KERNEL 23517879Sbde MCOUNT_EXIT(s); 23613116Sbde#endif 23713116Sbde return; 23813116Sbde} 23913116Sbde 24013116Sbde/* 24113116Sbde * Actual definition of mcount function. Defined in <machine/profile.h>, 24213116Sbde * which is included by <sys/gmon.h>. 24313116Sbde */ 24413116SbdeMCOUNT 24513116Sbde 24613116Sbde#ifdef GUPROF 24713116Sbdevoid 24813116Sbdemexitcount(selfpc) 24937629Sbde uintfptr_t selfpc; 25013116Sbde{ 25113116Sbde struct gmonparam *p; 25237629Sbde uintfptr_t selfpcdiff; 25313116Sbde 25413116Sbde p = &_gmonparam; 25537629Sbde selfpcdiff = selfpc - (uintfptr_t)p->lowpc; 25613116Sbde if (selfpcdiff < p->textsize) { 25719000Sbde int delta; 25813116Sbde 25913116Sbde /* 26019000Sbde * Count the time since cputime() was previously called 26119000Sbde * against `selfpc'. Compensate for overheads. 26213116Sbde */ 26319000Sbde delta = cputime() - cputime_bias - p->mexitcount_pre_overhead; 26419000Sbde cputime_bias = p->mexitcount_post_overhead; 26519000Sbde KCOUNT(p, selfpcdiff) += delta; 26613116Sbde *p->cputime_count += p->cputime_overhead; 26719000Sbde *p->mexitcount_count += p->mexitcount_overhead; 26813116Sbde } 26913116Sbde} 27019000Sbde 271170659Sbde#ifndef __GNUCLIKE_ASM 272170659Sbde#error "This file uses null asms to prevent timing loops being optimized away." 273170659Sbde#endif 274170659Sbde 27519000Sbdevoid 27619000Sbdeempty_loop() 27719000Sbde{ 27819000Sbde int i; 27919000Sbde 28019000Sbde for (i = 0; i < CALIB_SCALE; i++) 281170659Sbde __asm __volatile(""); 28219000Sbde} 28319000Sbde 28419000Sbdevoid 28519000Sbdenullfunc() 28619000Sbde{ 287170659Sbde __asm __volatile(""); 28819000Sbde} 28919000Sbde 29019000Sbdevoid 29119000Sbdenullfunc_loop() 29219000Sbde{ 29319000Sbde int i; 29419000Sbde 29519000Sbde for (i = 0; i < CALIB_SCALE; i++) 29619000Sbde nullfunc(); 29719000Sbde} 29813116Sbde#endif /* GUPROF */ 299