xref: /netbsd-src/sys/kern/subr_prof.c (revision 404fbe5fb94ca1e054339640cabb2801ce52dd30)
1 /*	$NetBSD: subr_prof.c,v 1.43 2007/12/20 23:03:10 dsl 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 1.43 2007/12/20 23:03:10 dsl 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/syscallargs.h>
44 #include <sys/sysctl.h>
45 
46 #include <sys/cpu.h>
47 
48 #ifdef GPROF
49 #include <sys/malloc.h>
50 #include <sys/gmon.h>
51 
52 MALLOC_DEFINE(M_GPROF, "gprof", "kernel profiling buffer");
53 
54 /*
55  * Froms is actually a bunch of unsigned shorts indexing tos
56  */
57 struct gmonparam _gmonparam = { .state = GMON_PROF_OFF };
58 
59 /* Actual start of the kernel text segment. */
60 extern char kernel_text[];
61 
62 extern char etext[];
63 
64 
65 void
66 kmstartup(void)
67 {
68 	char *cp;
69 	struct gmonparam *p = &_gmonparam;
70 	/*
71 	 * Round lowpc and highpc to multiples of the density we're using
72 	 * so the rest of the scaling (here and in gprof) stays in ints.
73 	 */
74 	p->lowpc = rounddown(((u_long)kernel_text),
75 		HISTFRACTION * sizeof(HISTCOUNTER));
76 	p->highpc = roundup((u_long)etext,
77 		HISTFRACTION * sizeof(HISTCOUNTER));
78 	p->textsize = p->highpc - p->lowpc;
79 	printf("Profiling kernel, textsize=%ld [%lx..%lx]\n",
80 	       p->textsize, p->lowpc, p->highpc);
81 	p->kcountsize = p->textsize / HISTFRACTION;
82 	p->hashfraction = HASHFRACTION;
83 	p->fromssize = p->textsize / HASHFRACTION;
84 	p->tolimit = p->textsize * ARCDENSITY / 100;
85 	if (p->tolimit < MINARCS)
86 		p->tolimit = MINARCS;
87 	else if (p->tolimit > MAXARCS)
88 		p->tolimit = MAXARCS;
89 	p->tossize = p->tolimit * sizeof(struct tostruct);
90 	cp = (char *)malloc(p->kcountsize + p->fromssize + p->tossize,
91 	    M_GPROF, M_NOWAIT | M_ZERO);
92 	if (cp == 0) {
93 		printf("No memory for profiling.\n");
94 		return;
95 	}
96 	p->tos = (struct tostruct *)cp;
97 	cp += p->tossize;
98 	p->kcount = (u_short *)cp;
99 	cp += p->kcountsize;
100 	p->froms = (u_short *)cp;
101 }
102 
103 /*
104  * Return kernel profiling information.
105  */
106 /*
107  * sysctl helper routine for kern.profiling subtree.  enables/disables
108  * kernel profiling and gives out copies of the profiling data.
109  */
110 static int
111 sysctl_kern_profiling(SYSCTLFN_ARGS)
112 {
113 	struct gmonparam *gp = &_gmonparam;
114 	int error;
115 	struct sysctlnode node;
116 
117 	node = *rnode;
118 
119 	switch (node.sysctl_num) {
120 	case GPROF_STATE:
121 		node.sysctl_data = &gp->state;
122 		break;
123 	case GPROF_COUNT:
124 		node.sysctl_data = gp->kcount;
125 		node.sysctl_size = gp->kcountsize;
126 		break;
127 	case GPROF_FROMS:
128 		node.sysctl_data = gp->froms;
129 		node.sysctl_size = gp->fromssize;
130 		break;
131 	case GPROF_TOS:
132 		node.sysctl_data = gp->tos;
133 		node.sysctl_size = gp->tossize;
134 		break;
135 	case GPROF_GMONPARAM:
136 		node.sysctl_data = gp;
137 		node.sysctl_size = sizeof(*gp);
138 		break;
139 	default:
140 		return (EOPNOTSUPP);
141 	}
142 
143 	error = sysctl_lookup(SYSCTLFN_CALL(&node));
144 	if (error || newp == NULL)
145 		return (error);
146 
147 	if (node.sysctl_num == GPROF_STATE) {
148 		mutex_spin_enter(&proc0.p_stmutex);
149 		if (gp->state == GMON_PROF_OFF)
150 			stopprofclock(&proc0);
151 		else
152 			startprofclock(&proc0);
153 		mutex_spin_exit(&proc0.p_stmutex);
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(struct lwp *l, const struct sys_profil_args *uap, register_t *retval)
219 {
220 	/* {
221 		syscallarg(char *) samples;
222 		syscallarg(u_int) size;
223 		syscallarg(u_int) offset;
224 		syscallarg(u_int) scale;
225 	} */
226 	struct proc *p = l->l_proc;
227 	struct uprof *upp;
228 
229 	if (SCARG(uap, scale) > (1 << 16))
230 		return (EINVAL);
231 	if (SCARG(uap, scale) == 0) {
232 		mutex_spin_enter(&p->p_stmutex);
233 		stopprofclock(p);
234 		mutex_spin_exit(&p->p_stmutex);
235 		return (0);
236 	}
237 	upp = &p->p_stats->p_prof;
238 
239 	/* Block profile interrupts while changing state. */
240 	mutex_spin_enter(&p->p_stmutex);
241 	upp->pr_off = SCARG(uap, offset);
242 	upp->pr_scale = SCARG(uap, scale);
243 	upp->pr_base = SCARG(uap, samples);
244 	upp->pr_size = SCARG(uap, size);
245 	startprofclock(p);
246 	mutex_spin_exit(&p->p_stmutex);
247 
248 	return (0);
249 }
250 
251 /*
252  * Scale is a fixed-point number with the binary point 16 bits
253  * into the value, and is <= 1.0.  pc is at most 32 bits, so the
254  * intermediate result is at most 48 bits.
255  */
256 #define	PC_TO_INDEX(pc, prof) \
257 	((int)(((u_quad_t)((pc) - (prof)->pr_off) * \
258 	    (u_quad_t)((prof)->pr_scale)) >> 16) & ~1)
259 
260 /*
261  * Collect user-level profiling statistics; called on a profiling tick,
262  * when a process is running in user-mode.  This routine may be called
263  * from an interrupt context.  We try to update the user profiling buffers
264  * cheaply with fuswintr() and suswintr().  If that fails, we revert to
265  * an AST that will vector us to trap() with a context in which copyin
266  * and copyout will work.  Trap will then call addupc_task().
267  *
268  * Note that we may (rarely) not get around to the AST soon enough, and
269  * lose profile ticks when the next tick overwrites this one, but in this
270  * case the system is overloaded and the profile is probably already
271  * inaccurate.
272  */
273 void
274 addupc_intr(struct lwp *l, u_long pc)
275 {
276 	struct uprof *prof;
277 	struct proc *p;
278 	void *addr;
279 	u_int i;
280 	int v;
281 
282 	p = l->l_proc;
283 
284 	KASSERT(mutex_owned(&p->p_stmutex));
285 
286 	prof = &p->p_stats->p_prof;
287 	if (pc < prof->pr_off ||
288 	    (i = PC_TO_INDEX(pc, prof)) >= prof->pr_size)
289 		return;			/* out of range; ignore */
290 
291 	addr = prof->pr_base + i;
292 	mutex_spin_exit(&p->p_stmutex);
293 	if ((v = fuswintr(addr)) == -1 || suswintr(addr, v + 1) == -1) {
294 		/* XXXSMP */
295 		prof->pr_addr = pc;
296 		prof->pr_ticks++;
297 		cpu_need_proftick(l);
298 	}
299 	mutex_spin_enter(&p->p_stmutex);
300 }
301 
302 /*
303  * Much like before, but we can afford to take faults here.  If the
304  * update fails, we simply turn off profiling.
305  */
306 void
307 addupc_task(struct lwp *l, u_long pc, u_int ticks)
308 {
309 	struct uprof *prof;
310 	struct proc *p;
311 	void *addr;
312 	int error;
313 	u_int i;
314 	u_short v;
315 
316 	p = l->l_proc;
317 
318 	if (ticks == 0)
319 		return;
320 
321 	mutex_spin_enter(&p->p_stmutex);
322 	prof = &p->p_stats->p_prof;
323 
324 	/* Testing P_PROFIL may be unnecessary, but is certainly safe. */
325 	if ((p->p_stflag & PST_PROFIL) == 0 || pc < prof->pr_off ||
326 	    (i = PC_TO_INDEX(pc, prof)) >= prof->pr_size) {
327 		mutex_spin_exit(&p->p_stmutex);
328 		return;
329 	}
330 
331 	addr = prof->pr_base + i;
332 	mutex_spin_exit(&p->p_stmutex);
333 	if ((error = copyin(addr, (void *)&v, sizeof(v))) == 0) {
334 		v += ticks;
335 		error = copyout((void *)&v, addr, sizeof(v));
336 	}
337 	if (error != 0) {
338 		mutex_spin_enter(&p->p_stmutex);
339 		stopprofclock(p);
340 		mutex_spin_exit(&p->p_stmutex);
341 	}
342 }
343