xref: /f-stack/freebsd/libkern/mcount.c (revision 22ce4aff)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 1983, 1992, 1993
5  *	The Regents of the University of California.  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  * 3. Neither the name of the University nor the names of its contributors
16  *    may be used to endorse or promote products derived from this software
17  *    without specific prior written permission.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
20  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
23  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
25  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
26  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29  * SUCH DAMAGE.
30  */
31 
32 #include <sys/cdefs.h>
33 __FBSDID("$FreeBSD$");
34 
35 #include <sys/param.h>
36 #include <sys/gmon.h>
37 #ifdef _KERNEL
38 #ifndef GUPROF
39 #include <sys/systm.h>
40 #endif
41 #include <vm/vm.h>
42 #include <vm/vm_param.h>
43 #include <vm/pmap.h>
44 #endif
45 
46 /*
47  * mcount is called on entry to each function compiled with the profiling
48  * switch set.  _mcount(), which is declared in a machine-dependent way
49  * with _MCOUNT_DECL, does the actual work and is either inlined into a
50  * C routine or called by an assembly stub.  In any case, this magic is
51  * taken care of by the MCOUNT definition in <machine/profile.h>.
52  *
53  * _mcount updates data structures that represent traversals of the
54  * program's call graph edges.  frompc and selfpc are the return
55  * address and function address that represents the given call graph edge.
56  *
57  * Note: the original BSD code used the same variable (frompcindex) for
58  * both frompcindex and frompc.  Any reasonable, modern compiler will
59  * perform this optimization.
60  */
61 /* _mcount; may be static, inline, etc */
_MCOUNT_DECL(uintfptr_t frompc,uintfptr_t selfpc)62 _MCOUNT_DECL(uintfptr_t frompc, uintfptr_t selfpc)
63 {
64 #ifdef GUPROF
65 	int delta;
66 #endif
67 	fptrdiff_t frompci;
68 	u_short *frompcindex;
69 	struct tostruct *top, *prevtop;
70 	struct gmonparam *p;
71 	long toindex;
72 #ifdef _KERNEL
73 	MCOUNT_DECL(s)
74 #endif
75 
76 	p = &_gmonparam;
77 #ifndef GUPROF			/* XXX */
78 	/*
79 	 * check that we are profiling
80 	 * and that we aren't recursively invoked.
81 	 */
82 	if (p->state != GMON_PROF_ON)
83 		return;
84 #endif
85 #ifdef _KERNEL
86 	MCOUNT_ENTER(s);
87 #else
88 	p->state = GMON_PROF_BUSY;
89 #endif
90 
91 #ifdef _KERNEL
92 	/* De-relocate any addresses in a (single) trampoline. */
93 #ifdef MCOUNT_DETRAMP
94 	MCOUNT_DETRAMP(frompc);
95 	MCOUNT_DETRAMP(selfpc);
96 #endif
97 	/*
98 	 * When we are called from an exception handler, frompc may be
99 	 * a user address.  Convert such frompc's to some representation
100 	 * in kernel address space.
101 	 */
102 #ifdef MCOUNT_FROMPC_USER
103 	frompc = MCOUNT_FROMPC_USER(frompc);
104 #elif defined(MCOUNT_USERPC)
105 	/*
106 	 * For separate address spaces, we can only guess that addresses
107 	 * in the range known to us are actually kernel addresses.  Outside
108 	 * of this range, conerting to the user address is fail-safe.
109 	 */
110 	if (frompc < p->lowpc || frompc - p->lowpc >= p->textsize)
111 		frompc = MCOUNT_USERPC;
112 #endif
113 #endif /* _KERNEL */
114 
115 	frompci = frompc - p->lowpc;
116 	if (frompci >= p->textsize)
117 		goto done;
118 
119 #ifdef GUPROF
120 	if (p->state == GMON_PROF_HIRES) {
121 		/*
122 		 * Count the time since cputime() was previously called
123 		 * against `frompc'.  Compensate for overheads.
124 		 *
125 		 * cputime() sets its prev_count variable to the count when
126 		 * it is called.  This in effect starts a counter for
127 		 * the next period of execution (normally from now until
128 		 * the next call to mcount() or mexitcount()).  We set
129 		 * cputime_bias to compensate for our own overhead.
130 		 *
131 		 * We use the usual sampling counters since they can be
132 		 * located efficiently.  4-byte counters are usually
133 		 * necessary.  gprof will add up the scattered counts
134 		 * just like it does for statistical profiling.  All
135 		 * counts are signed so that underflow in the subtractions
136 		 * doesn't matter much (negative counts are normally
137 		 * compensated for by larger counts elsewhere).  Underflow
138 		 * shouldn't occur, but may be caused by slightly wrong
139 		 * calibrations or from not clearing cputime_bias.
140 		 */
141 		delta = cputime() - cputime_bias - p->mcount_pre_overhead;
142 		cputime_bias = p->mcount_post_overhead;
143 		KCOUNT(p, frompci) += delta;
144 		*p->cputime_count += p->cputime_overhead;
145 		*p->mcount_count += p->mcount_overhead;
146 	}
147 #endif /* GUPROF */
148 
149 #ifdef _KERNEL
150 	/*
151 	 * When we are called from an exception handler, frompc is faked
152 	 * to be for where the exception occurred.  We've just solidified
153 	 * the count for there.  Now convert frompci to an index that
154 	 * represents the kind of exception so that interruptions appear
155 	 * in the call graph as calls from those index instead of calls
156 	 * from all over.
157 	 */
158 	frompc = MCOUNT_FROMPC_INTR(selfpc);
159 	if ((frompc - p->lowpc) < p->textsize)
160 		frompci = frompc - p->lowpc;
161 #endif
162 
163 	/*
164 	 * check that frompc is a reasonable pc value.
165 	 * for example:	signal catchers get called from the stack,
166 	 *		not from text space.  too bad.
167 	 */
168 	if (frompci >= p->textsize)
169 		goto done;
170 
171 	frompcindex = &p->froms[frompci / (p->hashfraction * sizeof(*p->froms))];
172 	toindex = *frompcindex;
173 	if (toindex == 0) {
174 		/*
175 		 *	first time traversing this arc
176 		 */
177 		toindex = ++p->tos[0].link;
178 		if (toindex >= p->tolimit)
179 			/* halt further profiling */
180 			goto overflow;
181 
182 		*frompcindex = toindex;
183 		top = &p->tos[toindex];
184 		top->selfpc = selfpc;
185 		top->count = 1;
186 		top->link = 0;
187 		goto done;
188 	}
189 	top = &p->tos[toindex];
190 	if (top->selfpc == selfpc) {
191 		/*
192 		 * arc at front of chain; usual case.
193 		 */
194 		top->count++;
195 		goto done;
196 	}
197 	/*
198 	 * have to go looking down chain for it.
199 	 * top points to what we are looking at,
200 	 * prevtop points to previous top.
201 	 * we know it is not at the head of the chain.
202 	 */
203 	for (; /* goto done */; ) {
204 		if (top->link == 0) {
205 			/*
206 			 * top is end of the chain and none of the chain
207 			 * had top->selfpc == selfpc.
208 			 * so we allocate a new tostruct
209 			 * and link it to the head of the chain.
210 			 */
211 			toindex = ++p->tos[0].link;
212 			if (toindex >= p->tolimit)
213 				goto overflow;
214 
215 			top = &p->tos[toindex];
216 			top->selfpc = selfpc;
217 			top->count = 1;
218 			top->link = *frompcindex;
219 			*frompcindex = toindex;
220 			goto done;
221 		}
222 		/*
223 		 * otherwise, check the next arc on the chain.
224 		 */
225 		prevtop = top;
226 		top = &p->tos[top->link];
227 		if (top->selfpc == selfpc) {
228 			/*
229 			 * there it is.
230 			 * increment its count
231 			 * move it to the head of the chain.
232 			 */
233 			top->count++;
234 			toindex = prevtop->link;
235 			prevtop->link = top->link;
236 			top->link = *frompcindex;
237 			*frompcindex = toindex;
238 			goto done;
239 		}
240 	}
241 done:
242 #ifdef _KERNEL
243 	MCOUNT_EXIT(s);
244 #else
245 	p->state = GMON_PROF_ON;
246 #endif
247 	return;
248 overflow:
249 	p->state = GMON_PROF_ERROR;
250 #ifdef _KERNEL
251 	MCOUNT_EXIT(s);
252 #endif
253 	return;
254 }
255 
256 /*
257  * Actual definition of mcount function.  Defined in <machine/profile.h>,
258  * which is included by <sys/gmon.h>.
259  */
260 MCOUNT
261 
262 #ifdef GUPROF
263 void
mexitcount(uintfptr_t selfpc)264 mexitcount(uintfptr_t selfpc)
265 {
266 	struct gmonparam *p;
267 	uintfptr_t selfpcdiff;
268 
269 	p = &_gmonparam;
270 #ifdef MCOUNT_DETRAMP
271 	MCOUNT_DETRAMP(selfpc);
272 #endif
273 	selfpcdiff = selfpc - (uintfptr_t)p->lowpc;
274 	if (selfpcdiff < p->textsize) {
275 		int delta;
276 
277 		/*
278 		 * Count the time since cputime() was previously called
279 		 * against `selfpc'.  Compensate for overheads.
280 		 */
281 		delta = cputime() - cputime_bias - p->mexitcount_pre_overhead;
282 		cputime_bias = p->mexitcount_post_overhead;
283 		KCOUNT(p, selfpcdiff) += delta;
284 		*p->cputime_count += p->cputime_overhead;
285 		*p->mexitcount_count += p->mexitcount_overhead;
286 	}
287 }
288 
289 #ifndef __GNUCLIKE_ASM
290 #error "This file uses null asms to prevent timing loops being optimized away."
291 #endif
292 
293 void
empty_loop(void)294 empty_loop(void)
295 {
296 	int i;
297 
298 	for (i = 0; i < CALIB_SCALE; i++)
299 		__asm __volatile("");
300 }
301 
302 void
nullfunc(void)303 nullfunc(void)
304 {
305 	__asm __volatile("");
306 }
307 
308 void
nullfunc_loop(void)309 nullfunc_loop(void)
310 {
311 	int i;
312 
313 	for (i = 0; i < CALIB_SCALE; i++)
314 		nullfunc();
315 }
316 #endif /* GUPROF */
317