xref: /freebsd/sys/kern/subr_prof.c (revision 6e8394b8baa7d5d9153ab90de6824bcd19b3b4e1)
1 /*-
2  * Copyright (c) 1982, 1986, 1993
3  *	The Regents of the University of California.  All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice, this list of conditions and the following disclaimer.
10  * 2. Redistributions in binary form must reproduce the above copyright
11  *    notice, this list of conditions and the following disclaimer in the
12  *    documentation and/or other materials provided with the distribution.
13  * 3. All advertising materials mentioning features or use of this software
14  *    must display the following acknowledgement:
15  *	This product includes software developed by the University of
16  *	California, Berkeley and its contributors.
17  * 4. Neither the name of the University nor the names of its contributors
18  *    may be used to endorse or promote products derived from this software
19  *    without specific prior written permission.
20  *
21  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
22  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
25  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31  * SUCH DAMAGE.
32  *
33  *	@(#)subr_prof.c	8.3 (Berkeley) 9/23/93
34  * $Id: subr_prof.c,v 1.28 1998/09/05 14:30:11 bde Exp $
35  */
36 
37 #include <sys/param.h>
38 #include <sys/systm.h>
39 #include <sys/sysproto.h>
40 #include <sys/kernel.h>
41 #include <sys/proc.h>
42 #include <sys/resourcevar.h>
43 #include <sys/sysctl.h>
44 
45 #include <machine/cpu.h>
46 
47 #ifdef GPROF
48 #include <sys/malloc.h>
49 #include <sys/gmon.h>
50 #undef MCOUNT
51 
52 static MALLOC_DEFINE(M_GPROF, "gprof", "kernel profiling buffer");
53 
54 static void kmstartup __P((void *));
55 SYSINIT(kmem, SI_SUB_KPROF, SI_ORDER_FIRST, kmstartup, NULL)
56 
57 struct gmonparam _gmonparam = { GMON_PROF_OFF };
58 
59 #ifdef GUPROF
60 #include <machine/asmacros.h>
61 
62 void
63 nullfunc_loop_profiled()
64 {
65 	int i;
66 
67 	for (i = 0; i < CALIB_SCALE; i++)
68 		nullfunc_profiled();
69 }
70 
71 #define	nullfunc_loop_profiled_end	nullfunc_profiled	/* XXX */
72 
73 void
74 nullfunc_profiled()
75 {
76 }
77 #endif /* GUPROF */
78 
79 static void
80 kmstartup(dummy)
81 	void *dummy;
82 {
83 	char *cp;
84 	struct gmonparam *p = &_gmonparam;
85 #ifdef GUPROF
86 	int cputime_overhead;
87 	int empty_loop_time;
88 	int i;
89 	int mcount_overhead;
90 	int mexitcount_overhead;
91 	int nullfunc_loop_overhead;
92 	int nullfunc_loop_profiled_time;
93 	uintfptr_t tmp_addr;
94 #endif
95 
96 	/*
97 	 * Round lowpc and highpc to multiples of the density we're using
98 	 * so the rest of the scaling (here and in gprof) stays in ints.
99 	 */
100 	p->lowpc = ROUNDDOWN((u_long)btext, HISTFRACTION * sizeof(HISTCOUNTER));
101 	p->highpc = ROUNDUP((u_long)etext, HISTFRACTION * sizeof(HISTCOUNTER));
102 	p->textsize = p->highpc - p->lowpc;
103 	printf("Profiling kernel, textsize=%lu [%x..%x]\n",
104 	       p->textsize, p->lowpc, p->highpc);
105 	p->kcountsize = p->textsize / HISTFRACTION;
106 	p->hashfraction = HASHFRACTION;
107 	p->fromssize = p->textsize / HASHFRACTION;
108 	p->tolimit = p->textsize * ARCDENSITY / 100;
109 	if (p->tolimit < MINARCS)
110 		p->tolimit = MINARCS;
111 	else if (p->tolimit > MAXARCS)
112 		p->tolimit = MAXARCS;
113 	p->tossize = p->tolimit * sizeof(struct tostruct);
114 	cp = (char *)malloc(p->kcountsize + p->fromssize + p->tossize,
115 	    M_GPROF, M_NOWAIT);
116 	if (cp == 0) {
117 		printf("No memory for profiling.\n");
118 		return;
119 	}
120 	bzero(cp, p->kcountsize + p->tossize + p->fromssize);
121 	p->tos = (struct tostruct *)cp;
122 	cp += p->tossize;
123 	p->kcount = (HISTCOUNTER *)cp;
124 	cp += p->kcountsize;
125 	p->froms = (u_short *)cp;
126 
127 #ifdef GUPROF
128 	/* Initialize pointers to overhead counters. */
129 	p->cputime_count = &KCOUNT(p, PC_TO_I(p, cputime));
130 	p->mcount_count = &KCOUNT(p, PC_TO_I(p, mcount));
131 	p->mexitcount_count = &KCOUNT(p, PC_TO_I(p, mexitcount));
132 
133 	/*
134 	 * Disable interrupts to avoid interference while we calibrate
135 	 * things.
136 	 */
137 	disable_intr();
138 
139 	/*
140 	 * Determine overheads.
141 	 * XXX this needs to be repeated for each useful timer/counter.
142 	 */
143 	cputime_overhead = 0;
144 	startguprof(p);
145 	for (i = 0; i < CALIB_SCALE; i++)
146 		cputime_overhead += cputime();
147 
148 	empty_loop();
149 	startguprof(p);
150 	empty_loop();
151 	empty_loop_time = cputime();
152 
153 	nullfunc_loop_profiled();
154 
155 	/*
156 	 * Start profiling.  There won't be any normal function calls since
157 	 * interrupts are disabled, but we will call the profiling routines
158 	 * directly to determine their overheads.
159 	 */
160 	p->state = GMON_PROF_HIRES;
161 
162 	startguprof(p);
163 	nullfunc_loop_profiled();
164 
165 	startguprof(p);
166 	for (i = 0; i < CALIB_SCALE; i++)
167 #if defined(__i386__) && __GNUC__ >= 2
168 		__asm("pushl %0; call __mcount; popl %%ecx"
169 		      :
170 		      : "i" (profil)
171 		      : "ax", "bx", "cx", "dx", "memory");
172 #else
173 #error
174 #endif
175 	mcount_overhead = KCOUNT(p, PC_TO_I(p, profil));
176 
177 	startguprof(p);
178 	for (i = 0; i < CALIB_SCALE; i++)
179 #if defined(__i386__) && __GNUC__ >= 2
180 		    __asm("call " __XSTRING(HIDENAME(mexitcount)) "; 1:"
181 			  : : : "ax", "bx", "cx", "dx", "memory");
182 	__asm("movl $1b,%0" : "=rm" (tmp_addr));
183 #else
184 #error
185 #endif
186 	mexitcount_overhead = KCOUNT(p, PC_TO_I(p, tmp_addr));
187 
188 	p->state = GMON_PROF_OFF;
189 	stopguprof(p);
190 
191 	enable_intr();
192 
193 	nullfunc_loop_profiled_time = 0;
194 	for (tmp_addr = (uintfptr_t)nullfunc_loop_profiled;
195 	     tmp_addr < (uintfptr_t)nullfunc_loop_profiled_end;
196 	     tmp_addr += HISTFRACTION * sizeof(HISTCOUNTER))
197 		nullfunc_loop_profiled_time += KCOUNT(p, PC_TO_I(p, tmp_addr));
198 #define CALIB_DOSCALE(count)	(((count) + CALIB_SCALE / 3) / CALIB_SCALE)
199 #define	c2n(count, freq)	((int)((count) * 1000000000LL / freq))
200 	printf("cputime %d, empty_loop %d, nullfunc_loop_profiled %d, mcount %d, mexitcount %d\n",
201 	       CALIB_DOSCALE(c2n(cputime_overhead, p->profrate)),
202 	       CALIB_DOSCALE(c2n(empty_loop_time, p->profrate)),
203 	       CALIB_DOSCALE(c2n(nullfunc_loop_profiled_time, p->profrate)),
204 	       CALIB_DOSCALE(c2n(mcount_overhead, p->profrate)),
205 	       CALIB_DOSCALE(c2n(mexitcount_overhead, p->profrate)));
206 	cputime_overhead -= empty_loop_time;
207 	mcount_overhead -= empty_loop_time;
208 	mexitcount_overhead -= empty_loop_time;
209 
210 	/*-
211 	 * Profiling overheads are determined by the times between the
212 	 * following events:
213 	 *	MC1: mcount() is called
214 	 *	MC2: cputime() (called from mcount()) latches the timer
215 	 *	MC3: mcount() completes
216 	 *	ME1: mexitcount() is called
217 	 *	ME2: cputime() (called from mexitcount()) latches the timer
218 	 *	ME3: mexitcount() completes.
219 	 * The times between the events vary slightly depending on instruction
220 	 * combination and cache misses, etc.  Attempt to determine the
221 	 * minimum times.  These can be subtracted from the profiling times
222 	 * without much risk of reducing the profiling times below what they
223 	 * would be when profiling is not configured.  Abbreviate:
224 	 *	ab = minimum time between MC1 and MC3
225 	 *	a  = minumum time between MC1 and MC2
226 	 *	b  = minimum time between MC2 and MC3
227 	 *	cd = minimum time between ME1 and ME3
228 	 *	c  = minimum time between ME1 and ME2
229 	 *	d  = minimum time between ME2 and ME3.
230 	 * These satisfy the relations:
231 	 *	ab            <= mcount_overhead		(just measured)
232 	 *	a + b         <= ab
233 	 *	        cd    <= mexitcount_overhead		(just measured)
234 	 *	        c + d <= cd
235 	 *	a         + d <= nullfunc_loop_profiled_time	(just measured)
236 	 *	a >= 0, b >= 0, c >= 0, d >= 0.
237 	 * Assume that ab and cd are equal to the minimums.
238 	 */
239 	p->cputime_overhead = CALIB_DOSCALE(cputime_overhead);
240 	p->mcount_overhead = CALIB_DOSCALE(mcount_overhead - cputime_overhead);
241 	p->mexitcount_overhead = CALIB_DOSCALE(mexitcount_overhead
242 					       - cputime_overhead);
243 	nullfunc_loop_overhead = nullfunc_loop_profiled_time - empty_loop_time;
244 	p->mexitcount_post_overhead = CALIB_DOSCALE((mcount_overhead
245 						     - nullfunc_loop_overhead)
246 						    / 4);
247 	p->mexitcount_pre_overhead = p->mexitcount_overhead
248 				     + p->cputime_overhead
249 				     - p->mexitcount_post_overhead;
250 	p->mcount_pre_overhead = CALIB_DOSCALE(nullfunc_loop_overhead)
251 				 - p->mexitcount_post_overhead;
252 	p->mcount_post_overhead = p->mcount_overhead
253 				  + p->cputime_overhead
254 				  - p->mcount_pre_overhead;
255 	printf(
256 "Profiling overheads: mcount: %d+%d, %d+%d; mexitcount: %d+%d, %d+%d nsec\n",
257 	       c2n(p->cputime_overhead, p->profrate),
258 	       c2n(p->mcount_overhead, p->profrate),
259 	       c2n(p->mcount_pre_overhead, p->profrate),
260 	       c2n(p->mcount_post_overhead, p->profrate),
261 	       c2n(p->cputime_overhead, p->profrate),
262 	       c2n(p->mexitcount_overhead, p->profrate),
263 	       c2n(p->mexitcount_pre_overhead, p->profrate),
264 	       c2n(p->mexitcount_post_overhead, p->profrate));
265 	printf(
266 "Profiling overheads: mcount: %d+%d, %d+%d; mexitcount: %d+%d, %d+%d cycles\n",
267 	       p->cputime_overhead, p->mcount_overhead,
268 	       p->mcount_pre_overhead, p->mcount_post_overhead,
269 	       p->cputime_overhead, p->mexitcount_overhead,
270 	       p->mexitcount_pre_overhead, p->mexitcount_post_overhead);
271 #endif /* GUPROF */
272 }
273 
274 /*
275  * Return kernel profiling information.
276  */
277 static int
278 sysctl_kern_prof SYSCTL_HANDLER_ARGS
279 {
280 	int *name = (int *) arg1;
281 	u_int namelen = arg2;
282 	struct gmonparam *gp = &_gmonparam;
283 	int error;
284 	int state;
285 
286 	/* all sysctl names at this level are terminal */
287 	if (namelen != 1)
288 		return (ENOTDIR);		/* overloaded */
289 
290 	switch (name[0]) {
291 	case GPROF_STATE:
292 		state = gp->state;
293 		error = sysctl_handle_int(oidp, &state, 0, req);
294 		if (error)
295 			return (error);
296 		if (!req->newptr)
297 			return (0);
298 		if (state == GMON_PROF_OFF) {
299 			gp->state = state;
300 			stopprofclock(&proc0);
301 			stopguprof(gp);
302 		} else if (state == GMON_PROF_ON) {
303 			gp->state = GMON_PROF_OFF;
304 			stopguprof(gp);
305 			gp->profrate = profhz;
306 			startprofclock(&proc0);
307 			gp->state = state;
308 #ifdef GUPROF
309 		} else if (state == GMON_PROF_HIRES) {
310 			gp->state = GMON_PROF_OFF;
311 			stopprofclock(&proc0);
312 			startguprof(gp);
313 			gp->state = state;
314 #endif
315 		} else if (state != gp->state)
316 			return (EINVAL);
317 		return (0);
318 	case GPROF_COUNT:
319 		return (sysctl_handle_opaque(oidp,
320 			gp->kcount, gp->kcountsize, req));
321 	case GPROF_FROMS:
322 		return (sysctl_handle_opaque(oidp,
323 			gp->froms, gp->fromssize, req));
324 	case GPROF_TOS:
325 		return (sysctl_handle_opaque(oidp,
326 			gp->tos, gp->tossize, req));
327 	case GPROF_GMONPARAM:
328 		return (sysctl_handle_opaque(oidp, gp, sizeof *gp, req));
329 	default:
330 		return (EOPNOTSUPP);
331 	}
332 	/* NOTREACHED */
333 }
334 
335 SYSCTL_NODE(_kern, KERN_PROF, prof, CTLFLAG_RW, sysctl_kern_prof, "");
336 #endif /* GPROF */
337 
338 /*
339  * Profiling system call.
340  *
341  * The scale factor is a fixed point number with 16 bits of fraction, so that
342  * 1.0 is represented as 0x10000.  A scale factor of 0 turns off profiling.
343  */
344 #ifndef _SYS_SYSPROTO_H_
345 struct profil_args {
346 	caddr_t	samples;
347 	size_t	size;
348 	size_t	offset;
349 	u_int	scale;
350 };
351 #endif
352 /* ARGSUSED */
353 int
354 profil(p, uap)
355 	struct proc *p;
356 	register struct profil_args *uap;
357 {
358 	register struct uprof *upp;
359 	int s;
360 
361 	if (uap->scale > (1 << 16))
362 		return (EINVAL);
363 	if (uap->scale == 0) {
364 		stopprofclock(p);
365 		return (0);
366 	}
367 	upp = &p->p_stats->p_prof;
368 
369 	/* Block profile interrupts while changing state. */
370 	s = splstatclock();
371 	upp->pr_off = uap->offset;
372 	upp->pr_scale = uap->scale;
373 	upp->pr_base = uap->samples;
374 	upp->pr_size = uap->size;
375 	startprofclock(p);
376 	splx(s);
377 
378 	return (0);
379 }
380 
381 /*
382  * Scale is a fixed-point number with the binary point 16 bits
383  * into the value, and is <= 1.0.  pc is at most 32 bits, so the
384  * intermediate result is at most 48 bits.
385  */
386 #define	PC_TO_INDEX(pc, prof) \
387 	((int)(((u_quad_t)((pc) - (prof)->pr_off) * \
388 	    (u_quad_t)((prof)->pr_scale)) >> 16) & ~1)
389 
390 /*
391  * Collect user-level profiling statistics; called on a profiling tick,
392  * when a process is running in user-mode.  This routine may be called
393  * from an interrupt context.  We try to update the user profiling buffers
394  * cheaply with fuswintr() and suswintr().  If that fails, we revert to
395  * an AST that will vector us to trap() with a context in which copyin
396  * and copyout will work.  Trap will then call addupc_task().
397  *
398  * Note that we may (rarely) not get around to the AST soon enough, and
399  * lose profile ticks when the next tick overwrites this one, but in this
400  * case the system is overloaded and the profile is probably already
401  * inaccurate.
402  */
403 void
404 addupc_intr(p, pc, ticks)
405 	register struct proc *p;
406 	register u_long pc;
407 	u_int ticks;
408 {
409 	register struct uprof *prof;
410 	register caddr_t addr;
411 	register u_int i;
412 	register int v;
413 
414 	if (ticks == 0)
415 		return;
416 	prof = &p->p_stats->p_prof;
417 	if (pc < prof->pr_off ||
418 	    (i = PC_TO_INDEX(pc, prof)) >= prof->pr_size)
419 		return;			/* out of range; ignore */
420 
421 	addr = prof->pr_base + i;
422 	if ((v = fuswintr(addr)) == -1 || suswintr(addr, v + ticks) == -1) {
423 		prof->pr_addr = pc;
424 		prof->pr_ticks = ticks;
425 		need_proftick(p);
426 	}
427 }
428 
429 /*
430  * Much like before, but we can afford to take faults here.  If the
431  * update fails, we simply turn off profiling.
432  */
433 void
434 addupc_task(p, pc, ticks)
435 	register struct proc *p;
436 	register u_long pc;
437 	u_int ticks;
438 {
439 	register struct uprof *prof;
440 	register caddr_t addr;
441 	register u_int i;
442 	u_short v;
443 
444 	/* Testing P_PROFIL may be unnecessary, but is certainly safe. */
445 	if ((p->p_flag & P_PROFIL) == 0 || ticks == 0)
446 		return;
447 
448 	prof = &p->p_stats->p_prof;
449 	if (pc < prof->pr_off ||
450 	    (i = PC_TO_INDEX(pc, prof)) >= prof->pr_size)
451 		return;
452 
453 	addr = prof->pr_base + i;
454 	if (copyin(addr, (caddr_t)&v, sizeof(v)) == 0) {
455 		v += ticks;
456 		if (copyout((caddr_t)&v, addr, sizeof(v)) == 0)
457 			return;
458 	}
459 	stopprofclock(p);
460 }
461