xref: /freebsd/sys/kern/subr_prof.c (revision 0384fff8c5b098545c3db311b0e0aa1ec4c9ae7e)
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  * $FreeBSD$
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/ipl.h>
46 #include <machine/cpu.h>
47 
48 #ifdef GPROF
49 #include <sys/malloc.h>
50 #include <sys/gmon.h>
51 #undef MCOUNT
52 
53 static MALLOC_DEFINE(M_GPROF, "gprof", "kernel profiling buffer");
54 
55 static void kmstartup __P((void *));
56 SYSINIT(kmem, SI_SUB_KPROF, SI_ORDER_FIRST, kmstartup, NULL)
57 
58 struct gmonparam _gmonparam = { GMON_PROF_OFF };
59 
60 #ifdef GUPROF
61 #include <machine/asmacros.h>
62 
63 void
64 nullfunc_loop_profiled()
65 {
66 	int i;
67 
68 	for (i = 0; i < CALIB_SCALE; i++)
69 		nullfunc_profiled();
70 }
71 
72 #define	nullfunc_loop_profiled_end	nullfunc_profiled	/* XXX */
73 
74 void
75 nullfunc_profiled()
76 {
77 }
78 #endif /* GUPROF */
79 
80 static void
81 kmstartup(dummy)
82 	void *dummy;
83 {
84 	char *cp;
85 	struct gmonparam *p = &_gmonparam;
86 #ifdef GUPROF
87 	int cputime_overhead;
88 	int empty_loop_time;
89 	int i;
90 	int mcount_overhead;
91 	int mexitcount_overhead;
92 	int nullfunc_loop_overhead;
93 	int nullfunc_loop_profiled_time;
94 	uintfptr_t tmp_addr;
95 #endif
96 	int intrstate;
97 
98 	/*
99 	 * Round lowpc and highpc to multiples of the density we're using
100 	 * so the rest of the scaling (here and in gprof) stays in ints.
101 	 */
102 	p->lowpc = ROUNDDOWN((u_long)btext, HISTFRACTION * sizeof(HISTCOUNTER));
103 	p->highpc = ROUNDUP((u_long)etext, HISTFRACTION * sizeof(HISTCOUNTER));
104 	p->textsize = p->highpc - p->lowpc;
105 	printf("Profiling kernel, textsize=%lu [%x..%x]\n",
106 	       p->textsize, p->lowpc, p->highpc);
107 	p->kcountsize = p->textsize / HISTFRACTION;
108 	p->hashfraction = HASHFRACTION;
109 	p->fromssize = p->textsize / HASHFRACTION;
110 	p->tolimit = p->textsize * ARCDENSITY / 100;
111 	if (p->tolimit < MINARCS)
112 		p->tolimit = MINARCS;
113 	else if (p->tolimit > MAXARCS)
114 		p->tolimit = MAXARCS;
115 	p->tossize = p->tolimit * sizeof(struct tostruct);
116 	cp = (char *)malloc(p->kcountsize + p->fromssize + p->tossize,
117 	    M_GPROF, M_NOWAIT);
118 	if (cp == 0) {
119 		printf("No memory for profiling.\n");
120 		return;
121 	}
122 	bzero(cp, p->kcountsize + p->tossize + p->fromssize);
123 	p->tos = (struct tostruct *)cp;
124 	cp += p->tossize;
125 	p->kcount = (HISTCOUNTER *)cp;
126 	cp += p->kcountsize;
127 	p->froms = (u_short *)cp;
128 
129 #ifdef GUPROF
130 	/* Initialize pointers to overhead counters. */
131 	p->cputime_count = &KCOUNT(p, PC_TO_I(p, cputime));
132 	p->mcount_count = &KCOUNT(p, PC_TO_I(p, mcount));
133 	p->mexitcount_count = &KCOUNT(p, PC_TO_I(p, mexitcount));
134 
135 	/*
136 	 * Disable interrupts to avoid interference while we calibrate
137 	 * things.
138 	 */
139 	intrstate = save_intr();
140 	disable_intr();
141 
142 	/*
143 	 * Determine overheads.
144 	 * XXX this needs to be repeated for each useful timer/counter.
145 	 */
146 	cputime_overhead = 0;
147 	startguprof(p);
148 	for (i = 0; i < CALIB_SCALE; i++)
149 		cputime_overhead += cputime();
150 
151 	empty_loop();
152 	startguprof(p);
153 	empty_loop();
154 	empty_loop_time = cputime();
155 
156 	nullfunc_loop_profiled();
157 
158 	/*
159 	 * Start profiling.  There won't be any normal function calls since
160 	 * interrupts are disabled, but we will call the profiling routines
161 	 * directly to determine their overheads.
162 	 */
163 	p->state = GMON_PROF_HIRES;
164 
165 	startguprof(p);
166 	nullfunc_loop_profiled();
167 
168 	startguprof(p);
169 	for (i = 0; i < CALIB_SCALE; i++)
170 #if defined(__i386__) && __GNUC__ >= 2
171 		__asm("pushl %0; call __mcount; popl %%ecx"
172 		      :
173 		      : "i" (profil)
174 		      : "ax", "bx", "cx", "dx", "memory");
175 #else
176 #error
177 #endif
178 	mcount_overhead = KCOUNT(p, PC_TO_I(p, profil));
179 
180 	startguprof(p);
181 	for (i = 0; i < CALIB_SCALE; i++)
182 #if defined(__i386__) && __GNUC__ >= 2
183 		    __asm("call " __XSTRING(HIDENAME(mexitcount)) "; 1:"
184 			  : : : "ax", "bx", "cx", "dx", "memory");
185 	__asm("movl $1b,%0" : "=rm" (tmp_addr));
186 #else
187 #error
188 #endif
189 	mexitcount_overhead = KCOUNT(p, PC_TO_I(p, tmp_addr));
190 
191 	p->state = GMON_PROF_OFF;
192 	stopguprof(p);
193 
194 	restore_intr(intrstate);
195 
196 	nullfunc_loop_profiled_time = 0;
197 	for (tmp_addr = (uintfptr_t)nullfunc_loop_profiled;
198 	     tmp_addr < (uintfptr_t)nullfunc_loop_profiled_end;
199 	     tmp_addr += HISTFRACTION * sizeof(HISTCOUNTER))
200 		nullfunc_loop_profiled_time += KCOUNT(p, PC_TO_I(p, tmp_addr));
201 #define CALIB_DOSCALE(count)	(((count) + CALIB_SCALE / 3) / CALIB_SCALE)
202 #define	c2n(count, freq)	((int)((count) * 1000000000LL / freq))
203 	printf("cputime %d, empty_loop %d, nullfunc_loop_profiled %d, mcount %d, mexitcount %d\n",
204 	       CALIB_DOSCALE(c2n(cputime_overhead, p->profrate)),
205 	       CALIB_DOSCALE(c2n(empty_loop_time, p->profrate)),
206 	       CALIB_DOSCALE(c2n(nullfunc_loop_profiled_time, p->profrate)),
207 	       CALIB_DOSCALE(c2n(mcount_overhead, p->profrate)),
208 	       CALIB_DOSCALE(c2n(mexitcount_overhead, p->profrate)));
209 	cputime_overhead -= empty_loop_time;
210 	mcount_overhead -= empty_loop_time;
211 	mexitcount_overhead -= empty_loop_time;
212 
213 	/*-
214 	 * Profiling overheads are determined by the times between the
215 	 * following events:
216 	 *	MC1: mcount() is called
217 	 *	MC2: cputime() (called from mcount()) latches the timer
218 	 *	MC3: mcount() completes
219 	 *	ME1: mexitcount() is called
220 	 *	ME2: cputime() (called from mexitcount()) latches the timer
221 	 *	ME3: mexitcount() completes.
222 	 * The times between the events vary slightly depending on instruction
223 	 * combination and cache misses, etc.  Attempt to determine the
224 	 * minimum times.  These can be subtracted from the profiling times
225 	 * without much risk of reducing the profiling times below what they
226 	 * would be when profiling is not configured.  Abbreviate:
227 	 *	ab = minimum time between MC1 and MC3
228 	 *	a  = minumum time between MC1 and MC2
229 	 *	b  = minimum time between MC2 and MC3
230 	 *	cd = minimum time between ME1 and ME3
231 	 *	c  = minimum time between ME1 and ME2
232 	 *	d  = minimum time between ME2 and ME3.
233 	 * These satisfy the relations:
234 	 *	ab            <= mcount_overhead		(just measured)
235 	 *	a + b         <= ab
236 	 *	        cd    <= mexitcount_overhead		(just measured)
237 	 *	        c + d <= cd
238 	 *	a         + d <= nullfunc_loop_profiled_time	(just measured)
239 	 *	a >= 0, b >= 0, c >= 0, d >= 0.
240 	 * Assume that ab and cd are equal to the minimums.
241 	 */
242 	p->cputime_overhead = CALIB_DOSCALE(cputime_overhead);
243 	p->mcount_overhead = CALIB_DOSCALE(mcount_overhead - cputime_overhead);
244 	p->mexitcount_overhead = CALIB_DOSCALE(mexitcount_overhead
245 					       - cputime_overhead);
246 	nullfunc_loop_overhead = nullfunc_loop_profiled_time - empty_loop_time;
247 	p->mexitcount_post_overhead = CALIB_DOSCALE((mcount_overhead
248 						     - nullfunc_loop_overhead)
249 						    / 4);
250 	p->mexitcount_pre_overhead = p->mexitcount_overhead
251 				     + p->cputime_overhead
252 				     - p->mexitcount_post_overhead;
253 	p->mcount_pre_overhead = CALIB_DOSCALE(nullfunc_loop_overhead)
254 				 - p->mexitcount_post_overhead;
255 	p->mcount_post_overhead = p->mcount_overhead
256 				  + p->cputime_overhead
257 				  - p->mcount_pre_overhead;
258 	printf(
259 "Profiling overheads: mcount: %d+%d, %d+%d; mexitcount: %d+%d, %d+%d nsec\n",
260 	       c2n(p->cputime_overhead, p->profrate),
261 	       c2n(p->mcount_overhead, p->profrate),
262 	       c2n(p->mcount_pre_overhead, p->profrate),
263 	       c2n(p->mcount_post_overhead, p->profrate),
264 	       c2n(p->cputime_overhead, p->profrate),
265 	       c2n(p->mexitcount_overhead, p->profrate),
266 	       c2n(p->mexitcount_pre_overhead, p->profrate),
267 	       c2n(p->mexitcount_post_overhead, p->profrate));
268 	printf(
269 "Profiling overheads: mcount: %d+%d, %d+%d; mexitcount: %d+%d, %d+%d cycles\n",
270 	       p->cputime_overhead, p->mcount_overhead,
271 	       p->mcount_pre_overhead, p->mcount_post_overhead,
272 	       p->cputime_overhead, p->mexitcount_overhead,
273 	       p->mexitcount_pre_overhead, p->mexitcount_post_overhead);
274 #endif /* GUPROF */
275 }
276 
277 /*
278  * Return kernel profiling information.
279  */
280 static int
281 sysctl_kern_prof(SYSCTL_HANDLER_ARGS)
282 {
283 	int *name = (int *) arg1;
284 	u_int namelen = arg2;
285 	struct gmonparam *gp = &_gmonparam;
286 	int error;
287 	int state;
288 
289 	/* all sysctl names at this level are terminal */
290 	if (namelen != 1)
291 		return (ENOTDIR);		/* overloaded */
292 
293 	switch (name[0]) {
294 	case GPROF_STATE:
295 		state = gp->state;
296 		error = sysctl_handle_int(oidp, &state, 0, req);
297 		if (error)
298 			return (error);
299 		if (!req->newptr)
300 			return (0);
301 		if (state == GMON_PROF_OFF) {
302 			gp->state = state;
303 			stopprofclock(&proc0);
304 			stopguprof(gp);
305 		} else if (state == GMON_PROF_ON) {
306 			gp->state = GMON_PROF_OFF;
307 			stopguprof(gp);
308 			gp->profrate = profhz;
309 			startprofclock(&proc0);
310 			gp->state = state;
311 #ifdef GUPROF
312 		} else if (state == GMON_PROF_HIRES) {
313 			gp->state = GMON_PROF_OFF;
314 			stopprofclock(&proc0);
315 			startguprof(gp);
316 			gp->state = state;
317 #endif
318 		} else if (state != gp->state)
319 			return (EINVAL);
320 		return (0);
321 	case GPROF_COUNT:
322 		return (sysctl_handle_opaque(oidp,
323 			gp->kcount, gp->kcountsize, req));
324 	case GPROF_FROMS:
325 		return (sysctl_handle_opaque(oidp,
326 			gp->froms, gp->fromssize, req));
327 	case GPROF_TOS:
328 		return (sysctl_handle_opaque(oidp,
329 			gp->tos, gp->tossize, req));
330 	case GPROF_GMONPARAM:
331 		return (sysctl_handle_opaque(oidp, gp, sizeof *gp, req));
332 	default:
333 		return (EOPNOTSUPP);
334 	}
335 	/* NOTREACHED */
336 }
337 
338 SYSCTL_NODE(_kern, KERN_PROF, prof, CTLFLAG_RW, sysctl_kern_prof, "");
339 #endif /* GPROF */
340 
341 /*
342  * Profiling system call.
343  *
344  * The scale factor is a fixed point number with 16 bits of fraction, so that
345  * 1.0 is represented as 0x10000.  A scale factor of 0 turns off profiling.
346  */
347 #ifndef _SYS_SYSPROTO_H_
348 struct profil_args {
349 	caddr_t	samples;
350 	size_t	size;
351 	size_t	offset;
352 	u_int	scale;
353 };
354 #endif
355 /* ARGSUSED */
356 int
357 profil(p, uap)
358 	struct proc *p;
359 	register struct profil_args *uap;
360 {
361 	register struct uprof *upp;
362 	int s;
363 
364 	if (uap->scale > (1 << 16))
365 		return (EINVAL);
366 	if (uap->scale == 0) {
367 		stopprofclock(p);
368 		return (0);
369 	}
370 	upp = &p->p_stats->p_prof;
371 
372 	/* Block profile interrupts while changing state. */
373 	s = splstatclock();
374 	upp->pr_off = uap->offset;
375 	upp->pr_scale = uap->scale;
376 	upp->pr_base = uap->samples;
377 	upp->pr_size = uap->size;
378 	startprofclock(p);
379 	splx(s);
380 
381 	return (0);
382 }
383 
384 /*
385  * Scale is a fixed-point number with the binary point 16 bits
386  * into the value, and is <= 1.0.  pc is at most 32 bits, so the
387  * intermediate result is at most 48 bits.
388  */
389 #define	PC_TO_INDEX(pc, prof) \
390 	((int)(((u_quad_t)((pc) - (prof)->pr_off) * \
391 	    (u_quad_t)((prof)->pr_scale)) >> 16) & ~1)
392 
393 /*
394  * Collect user-level profiling statistics; called on a profiling tick,
395  * when a process is running in user-mode.  This routine may be called
396  * from an interrupt context.  We try to update the user profiling buffers
397  * cheaply with fuswintr() and suswintr().  If that fails, we revert to
398  * an AST that will vector us to trap() with a context in which copyin
399  * and copyout will work.  Trap will then call addupc_task().
400  *
401  * Note that we may (rarely) not get around to the AST soon enough, and
402  * lose profile ticks when the next tick overwrites this one, but in this
403  * case the system is overloaded and the profile is probably already
404  * inaccurate.
405  */
406 void
407 addupc_intr(p, pc, ticks)
408 	register struct proc *p;
409 	register u_long pc;
410 	u_int ticks;
411 {
412 	register struct uprof *prof;
413 	register caddr_t addr;
414 	register u_int i;
415 	register int v;
416 
417 	if (ticks == 0)
418 		return;
419 	prof = &p->p_stats->p_prof;
420 	if (pc < prof->pr_off ||
421 	    (i = PC_TO_INDEX(pc, prof)) >= prof->pr_size)
422 		return;			/* out of range; ignore */
423 
424 	addr = prof->pr_base + i;
425 	if ((v = fuswintr(addr)) == -1 || suswintr(addr, v + ticks) == -1) {
426 		prof->pr_addr = pc;
427 		prof->pr_ticks = ticks;
428 		need_proftick(p);
429 	}
430 }
431 
432 /*
433  * Much like before, but we can afford to take faults here.  If the
434  * update fails, we simply turn off profiling.
435  */
436 void
437 addupc_task(p, pc, ticks)
438 	register struct proc *p;
439 	register u_long pc;
440 	u_int ticks;
441 {
442 	register struct uprof *prof;
443 	register caddr_t addr;
444 	register u_int i;
445 	u_short v;
446 
447 	/* Testing P_PROFIL may be unnecessary, but is certainly safe. */
448 	if ((p->p_flag & P_PROFIL) == 0 || ticks == 0)
449 		return;
450 
451 	prof = &p->p_stats->p_prof;
452 	if (pc < prof->pr_off ||
453 	    (i = PC_TO_INDEX(pc, prof)) >= prof->pr_size)
454 		return;
455 
456 	addr = prof->pr_base + i;
457 	if (copyin(addr, (caddr_t)&v, sizeof(v)) == 0) {
458 		v += ticks;
459 		if (copyout((caddr_t)&v, addr, sizeof(v)) == 0)
460 			return;
461 	}
462 	stopprofclock(p);
463 }
464