xref: /freebsd-12.1/sys/libkern/mcount.c (revision dbe30617)
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  */
_MCOUNT_DECL(uintfptr_t frompc,uintfptr_t selfpc)61 _MCOUNT_DECL(uintfptr_t frompc, uintfptr_t selfpc)	/* _mcount; may be static, inline, etc */
62 {
63 #ifdef GUPROF
64 	int delta;
65 #endif
66 	fptrdiff_t frompci;
67 	u_short *frompcindex;
68 	struct tostruct *top, *prevtop;
69 	struct gmonparam *p;
70 	long toindex;
71 #ifdef _KERNEL
72 	MCOUNT_DECL(s)
73 #endif
74 
75 	p = &_gmonparam;
76 #ifndef GUPROF			/* XXX */
77 	/*
78 	 * check that we are profiling
79 	 * and that we aren't recursively invoked.
80 	 */
81 	if (p->state != GMON_PROF_ON)
82 		return;
83 #endif
84 #ifdef _KERNEL
85 	MCOUNT_ENTER(s);
86 #else
87 	p->state = GMON_PROF_BUSY;
88 #endif
89 
90 #ifdef _KERNEL
91 	/* De-relocate any addresses in a (single) trampoline. */
92 #ifdef MCOUNT_DETRAMP
93 	MCOUNT_DETRAMP(frompc);
94 	MCOUNT_DETRAMP(selfpc);
95 #endif
96 	/*
97 	 * When we are called from an exception handler, frompc may be
98 	 * a user address.  Convert such frompc's to some representation
99 	 * in kernel address space.
100 	 */
101 #ifdef MCOUNT_FROMPC_USER
102 	frompc = MCOUNT_FROMPC_USER(frompc);
103 #elif defined(MCOUNT_USERPC)
104 	/*
105 	 * For separate address spaces, we can only guess that addresses
106 	 * in the range known to us are actually kernel addresses.  Outside
107 	 * of this range, conerting to the user address is fail-safe.
108 	 */
109 	if (frompc < p->lowpc || frompc - p->lowpc >= p->textsize)
110 		frompc = MCOUNT_USERPC;
111 #endif
112 #endif /* _KERNEL */
113 
114 	frompci = frompc - p->lowpc;
115 	if (frompci >= p->textsize)
116 		goto done;
117 
118 #ifdef GUPROF
119 	if (p->state == GMON_PROF_HIRES) {
120 		/*
121 		 * Count the time since cputime() was previously called
122 		 * against `frompc'.  Compensate for overheads.
123 		 *
124 		 * cputime() sets its prev_count variable to the count when
125 		 * it is called.  This in effect starts a counter for
126 		 * the next period of execution (normally from now until
127 		 * the next call to mcount() or mexitcount()).  We set
128 		 * cputime_bias to compensate for our own overhead.
129 		 *
130 		 * We use the usual sampling counters since they can be
131 		 * located efficiently.  4-byte counters are usually
132 		 * necessary.  gprof will add up the scattered counts
133 		 * just like it does for statistical profiling.  All
134 		 * counts are signed so that underflow in the subtractions
135 		 * doesn't matter much (negative counts are normally
136 		 * compensated for by larger counts elsewhere).  Underflow
137 		 * shouldn't occur, but may be caused by slightly wrong
138 		 * calibrations or from not clearing cputime_bias.
139 		 */
140 		delta = cputime() - cputime_bias - p->mcount_pre_overhead;
141 		cputime_bias = p->mcount_post_overhead;
142 		KCOUNT(p, frompci) += delta;
143 		*p->cputime_count += p->cputime_overhead;
144 		*p->mcount_count += p->mcount_overhead;
145 	}
146 #endif /* GUPROF */
147 
148 #ifdef _KERNEL
149 	/*
150 	 * When we are called from an exception handler, frompc is faked
151 	 * to be for where the exception occurred.  We've just solidified
152 	 * the count for there.  Now convert frompci to an index that
153 	 * represents the kind of exception so that interruptions appear
154 	 * in the call graph as calls from those index instead of calls
155 	 * from all over.
156 	 */
157 	frompc = MCOUNT_FROMPC_INTR(selfpc);
158 	if ((frompc - p->lowpc) < p->textsize)
159 		frompci = frompc - p->lowpc;
160 #endif
161 
162 	/*
163 	 * check that frompc is a reasonable pc value.
164 	 * for example:	signal catchers get called from the stack,
165 	 *		not from text space.  too bad.
166 	 */
167 	if (frompci >= p->textsize)
168 		goto done;
169 
170 	frompcindex = &p->froms[frompci / (p->hashfraction * sizeof(*p->froms))];
171 	toindex = *frompcindex;
172 	if (toindex == 0) {
173 		/*
174 		 *	first time traversing this arc
175 		 */
176 		toindex = ++p->tos[0].link;
177 		if (toindex >= p->tolimit)
178 			/* halt further profiling */
179 			goto overflow;
180 
181 		*frompcindex = toindex;
182 		top = &p->tos[toindex];
183 		top->selfpc = selfpc;
184 		top->count = 1;
185 		top->link = 0;
186 		goto done;
187 	}
188 	top = &p->tos[toindex];
189 	if (top->selfpc == selfpc) {
190 		/*
191 		 * arc at front of chain; usual case.
192 		 */
193 		top->count++;
194 		goto done;
195 	}
196 	/*
197 	 * have to go looking down chain for it.
198 	 * top points to what we are looking at,
199 	 * prevtop points to previous top.
200 	 * we know it is not at the head of the chain.
201 	 */
202 	for (; /* goto done */; ) {
203 		if (top->link == 0) {
204 			/*
205 			 * top is end of the chain and none of the chain
206 			 * had top->selfpc == selfpc.
207 			 * so we allocate a new tostruct
208 			 * and link it to the head of the chain.
209 			 */
210 			toindex = ++p->tos[0].link;
211 			if (toindex >= p->tolimit)
212 				goto overflow;
213 
214 			top = &p->tos[toindex];
215 			top->selfpc = selfpc;
216 			top->count = 1;
217 			top->link = *frompcindex;
218 			*frompcindex = toindex;
219 			goto done;
220 		}
221 		/*
222 		 * otherwise, check the next arc on the chain.
223 		 */
224 		prevtop = top;
225 		top = &p->tos[top->link];
226 		if (top->selfpc == selfpc) {
227 			/*
228 			 * there it is.
229 			 * increment its count
230 			 * move it to the head of the chain.
231 			 */
232 			top->count++;
233 			toindex = prevtop->link;
234 			prevtop->link = top->link;
235 			top->link = *frompcindex;
236 			*frompcindex = toindex;
237 			goto done;
238 		}
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()294 empty_loop()
295 {
296 	int i;
297 
298 	for (i = 0; i < CALIB_SCALE; i++)
299 		__asm __volatile("");
300 }
301 
302 void
nullfunc()303 nullfunc()
304 {
305 	__asm __volatile("");
306 }
307 
308 void
nullfunc_loop()309 nullfunc_loop()
310 {
311 	int i;
312 
313 	for (i = 0; i < CALIB_SCALE; i++)
314 		nullfunc();
315 }
316 #endif /* GUPROF */
317