xref: /netbsd-src/sys/kern/subr_prof.c (revision 23c8222edbfb0f0932d88a8351d3a0cf817dfb9e)
1 /*	$NetBSD: subr_prof.c,v 1.31 2004/04/08 06:20:30 atatat Exp $	*/
2 
3 /*-
4  * Copyright (c) 1982, 1986, 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  *	@(#)subr_prof.c	8.4 (Berkeley) 2/14/95
32  */
33 
34 #include <sys/cdefs.h>
35 __KERNEL_RCSID(0, "$NetBSD: subr_prof.c,v 1.31 2004/04/08 06:20:30 atatat Exp $");
36 
37 #include <sys/param.h>
38 #include <sys/systm.h>
39 #include <sys/kernel.h>
40 #include <sys/proc.h>
41 #include <sys/user.h>
42 #include <sys/mount.h>
43 #include <sys/sa.h>
44 #include <sys/syscallargs.h>
45 #include <sys/sysctl.h>
46 
47 #include <machine/cpu.h>
48 
49 #ifdef GPROF
50 #include <sys/malloc.h>
51 #include <sys/gmon.h>
52 
53 MALLOC_DEFINE(M_GPROF, "gprof", "kernel profiling buffer");
54 
55 /*
56  * Froms is actually a bunch of unsigned shorts indexing tos
57  */
58 struct gmonparam _gmonparam = { GMON_PROF_OFF };
59 
60 /* Actual start of the kernel text segment. */
61 extern char kernel_text[];
62 
63 extern char etext[];
64 
65 
66 void
67 kmstartup()
68 {
69 	char *cp;
70 	struct gmonparam *p = &_gmonparam;
71 	/*
72 	 * Round lowpc and highpc to multiples of the density we're using
73 	 * so the rest of the scaling (here and in gprof) stays in ints.
74 	 */
75 	p->lowpc = ROUNDDOWN(((u_long)kernel_text),
76 		HISTFRACTION * sizeof(HISTCOUNTER));
77 	p->highpc = ROUNDUP((u_long)etext,
78 		HISTFRACTION * sizeof(HISTCOUNTER));
79 	p->textsize = p->highpc - p->lowpc;
80 	printf("Profiling kernel, textsize=%ld [%lx..%lx]\n",
81 	       p->textsize, p->lowpc, p->highpc);
82 	p->kcountsize = p->textsize / HISTFRACTION;
83 	p->hashfraction = HASHFRACTION;
84 	p->fromssize = p->textsize / HASHFRACTION;
85 	p->tolimit = p->textsize * ARCDENSITY / 100;
86 	if (p->tolimit < MINARCS)
87 		p->tolimit = MINARCS;
88 	else if (p->tolimit > MAXARCS)
89 		p->tolimit = MAXARCS;
90 	p->tossize = p->tolimit * sizeof(struct tostruct);
91 	cp = (char *)malloc(p->kcountsize + p->fromssize + p->tossize,
92 	    M_GPROF, M_NOWAIT);
93 	if (cp == 0) {
94 		printf("No memory for profiling.\n");
95 		return;
96 	}
97 	memset(cp, 0, p->kcountsize + p->tossize + p->fromssize);
98 	p->tos = (struct tostruct *)cp;
99 	cp += p->tossize;
100 	p->kcount = (u_short *)cp;
101 	cp += p->kcountsize;
102 	p->froms = (u_short *)cp;
103 }
104 
105 /*
106  * Return kernel profiling information.
107  */
108 /*
109  * sysctl helper routine for kern.profiling subtree.  enables/disables
110  * kernel profiling and gives out copies of the profiling data.
111  */
112 static int
113 sysctl_kern_profiling(SYSCTLFN_ARGS)
114 {
115 	struct gmonparam *gp = &_gmonparam;
116 	int error;
117 	struct sysctlnode node;
118 
119 	node = *rnode;
120 
121 	switch (node.sysctl_num) {
122 	case GPROF_STATE:
123 		node.sysctl_data = &gp->state;
124 		break;
125 	case GPROF_COUNT:
126 		node.sysctl_data = gp->kcount;
127 		node.sysctl_size = gp->kcountsize;
128 		break;
129 	case GPROF_FROMS:
130 		node.sysctl_data = gp->froms;
131 		node.sysctl_size = gp->fromssize;
132 		break;
133 	case GPROF_TOS:
134 		node.sysctl_data = gp->tos;
135 		node.sysctl_size = gp->tossize;
136 		break;
137 	case GPROF_GMONPARAM:
138 		node.sysctl_data = gp;
139 		node.sysctl_size = sizeof(*gp);
140 		break;
141 	default:
142 		return (EOPNOTSUPP);
143 	}
144 
145 	error = sysctl_lookup(SYSCTLFN_CALL(&node));
146 	if (error || newp == NULL)
147 		return (error);
148 
149 	if (node.sysctl_num == GPROF_STATE) {
150 		if (gp->state == GMON_PROF_OFF)
151 			stopprofclock(&proc0);
152 		else
153 			startprofclock(&proc0);
154 	}
155 
156 	return (0);
157 }
158 
159 SYSCTL_SETUP(sysctl_kern_gprof_setup, "sysctl kern.profiling subtree setup")
160 {
161 
162 	sysctl_createv(clog, 0, NULL, NULL,
163 		       CTLFLAG_PERMANENT,
164 		       CTLTYPE_NODE, "kern", NULL,
165 		       NULL, 0, NULL, 0,
166 		       CTL_KERN, CTL_EOL);
167 	sysctl_createv(clog, 0, NULL, NULL,
168 		       CTLFLAG_PERMANENT,
169 		       CTLTYPE_NODE, "profiling",
170 		       SYSCTL_DESCR("Profiling information (available)"),
171 		       NULL, 0, NULL, 0,
172 		       CTL_KERN, KERN_PROF, CTL_EOL);
173 
174 	sysctl_createv(clog, 0, NULL, NULL,
175 		       CTLFLAG_PERMANENT|CTLFLAG_READWRITE,
176 		       CTLTYPE_INT, "state",
177 		       SYSCTL_DESCR("Profiling state"),
178 		       sysctl_kern_profiling, 0, NULL, 0,
179 		       CTL_KERN, KERN_PROF, GPROF_STATE, CTL_EOL);
180 	sysctl_createv(clog, 0, NULL, NULL,
181 		       CTLFLAG_PERMANENT|CTLFLAG_READWRITE,
182 		       CTLTYPE_STRUCT, "count",
183 		       SYSCTL_DESCR("Array of statistical program counters"),
184 		       sysctl_kern_profiling, 0, NULL, 0,
185 		       CTL_KERN, KERN_PROF, GPROF_COUNT, CTL_EOL);
186 	sysctl_createv(clog, 0, NULL, NULL,
187 		       CTLFLAG_PERMANENT|CTLFLAG_READWRITE,
188 		       CTLTYPE_STRUCT, "froms",
189 		       SYSCTL_DESCR("Array indexed by program counter of "
190 				    "call-from points"),
191 		       sysctl_kern_profiling, 0, NULL, 0,
192 		       CTL_KERN, KERN_PROF, GPROF_FROMS, CTL_EOL);
193 	sysctl_createv(clog, 0, NULL, NULL,
194 		       CTLFLAG_PERMANENT|CTLFLAG_READWRITE,
195 		       CTLTYPE_STRUCT, "tos",
196 		       SYSCTL_DESCR("Array of structures describing "
197 				    "destination of calls and their counts"),
198 		       sysctl_kern_profiling, 0, NULL, 0,
199 		       CTL_KERN, KERN_PROF, GPROF_TOS, CTL_EOL);
200 	sysctl_createv(clog, 0, NULL, NULL,
201 		       CTLFLAG_PERMANENT,
202 		       CTLTYPE_STRUCT, "gmonparam",
203 		       SYSCTL_DESCR("Structure giving the sizes of the above "
204 				    "arrays"),
205 		       sysctl_kern_profiling, 0, NULL, 0,
206 		       CTL_KERN, KERN_PROF, GPROF_GMONPARAM, CTL_EOL);
207 }
208 #endif /* GPROF */
209 
210 /*
211  * Profiling system call.
212  *
213  * The scale factor is a fixed point number with 16 bits of fraction, so that
214  * 1.0 is represented as 0x10000.  A scale factor of 0 turns off profiling.
215  */
216 /* ARGSUSED */
217 int
218 sys_profil(l, v, retval)
219 	struct lwp *l;
220 	void *v;
221 	register_t *retval;
222 {
223 	struct sys_profil_args /* {
224 		syscallarg(caddr_t) samples;
225 		syscallarg(u_int) size;
226 		syscallarg(u_int) offset;
227 		syscallarg(u_int) scale;
228 	} */ *uap = v;
229 	struct proc *p = l->l_proc;
230 	struct uprof *upp;
231 	int s;
232 
233 	if (SCARG(uap, scale) > (1 << 16))
234 		return (EINVAL);
235 	if (SCARG(uap, scale) == 0) {
236 		stopprofclock(p);
237 		return (0);
238 	}
239 	upp = &p->p_stats->p_prof;
240 
241 	/* Block profile interrupts while changing state. */
242 	s = splstatclock();
243 	upp->pr_off = SCARG(uap, offset);
244 	upp->pr_scale = SCARG(uap, scale);
245 	upp->pr_base = SCARG(uap, samples);
246 	upp->pr_size = SCARG(uap, size);
247 	startprofclock(p);
248 	splx(s);
249 
250 	return (0);
251 }
252 
253 /*
254  * Scale is a fixed-point number with the binary point 16 bits
255  * into the value, and is <= 1.0.  pc is at most 32 bits, so the
256  * intermediate result is at most 48 bits.
257  */
258 #define	PC_TO_INDEX(pc, prof) \
259 	((int)(((u_quad_t)((pc) - (prof)->pr_off) * \
260 	    (u_quad_t)((prof)->pr_scale)) >> 16) & ~1)
261 
262 /*
263  * Collect user-level profiling statistics; called on a profiling tick,
264  * when a process is running in user-mode.  This routine may be called
265  * from an interrupt context.  We try to update the user profiling buffers
266  * cheaply with fuswintr() and suswintr().  If that fails, we revert to
267  * an AST that will vector us to trap() with a context in which copyin
268  * and copyout will work.  Trap will then call addupc_task().
269  *
270  * Note that we may (rarely) not get around to the AST soon enough, and
271  * lose profile ticks when the next tick overwrites this one, but in this
272  * case the system is overloaded and the profile is probably already
273  * inaccurate.
274  */
275 void
276 addupc_intr(p, pc)
277 	struct proc *p;
278 	u_long pc;
279 {
280 	struct uprof *prof;
281 	caddr_t addr;
282 	u_int i;
283 	int v;
284 
285 	prof = &p->p_stats->p_prof;
286 	if (pc < prof->pr_off ||
287 	    (i = PC_TO_INDEX(pc, prof)) >= prof->pr_size)
288 		return;			/* out of range; ignore */
289 
290 	addr = prof->pr_base + i;
291 	if ((v = fuswintr(addr)) == -1 || suswintr(addr, v + 1) == -1) {
292 		prof->pr_addr = pc;
293 		prof->pr_ticks++;
294 		need_proftick(p);
295 	}
296 }
297 
298 /*
299  * Much like before, but we can afford to take faults here.  If the
300  * update fails, we simply turn off profiling.
301  */
302 void
303 addupc_task(p, pc, ticks)
304 	struct proc *p;
305 	u_long pc;
306 	u_int ticks;
307 {
308 	struct uprof *prof;
309 	caddr_t addr;
310 	u_int i;
311 	u_short v;
312 
313 	/* Testing P_PROFIL may be unnecessary, but is certainly safe. */
314 	if ((p->p_flag & P_PROFIL) == 0 || ticks == 0)
315 		return;
316 
317 	prof = &p->p_stats->p_prof;
318 	if (pc < prof->pr_off ||
319 	    (i = PC_TO_INDEX(pc, prof)) >= prof->pr_size)
320 		return;
321 
322 	addr = prof->pr_base + i;
323 	if (copyin(addr, (caddr_t)&v, sizeof(v)) == 0) {
324 		v += ticks;
325 		if (copyout((caddr_t)&v, addr, sizeof(v)) == 0)
326 			return;
327 	}
328 	stopprofclock(p);
329 }
330