subr_prof.c revision 1.47 1 1.47 christos /* $NetBSD: subr_prof.c,v 1.47 2014/07/10 21:13:52 christos Exp $ */
2 1.3 cgd
3 1.1 cgd /*-
4 1.1 cgd * Copyright (c) 1982, 1986, 1993
5 1.1 cgd * The Regents of the University of California. All rights reserved.
6 1.1 cgd *
7 1.1 cgd * Redistribution and use in source and binary forms, with or without
8 1.1 cgd * modification, are permitted provided that the following conditions
9 1.1 cgd * are met:
10 1.1 cgd * 1. Redistributions of source code must retain the above copyright
11 1.1 cgd * notice, this list of conditions and the following disclaimer.
12 1.1 cgd * 2. Redistributions in binary form must reproduce the above copyright
13 1.1 cgd * notice, this list of conditions and the following disclaimer in the
14 1.1 cgd * documentation and/or other materials provided with the distribution.
15 1.28 agc * 3. Neither the name of the University nor the names of its contributors
16 1.1 cgd * may be used to endorse or promote products derived from this software
17 1.1 cgd * without specific prior written permission.
18 1.1 cgd *
19 1.1 cgd * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
20 1.1 cgd * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21 1.1 cgd * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22 1.1 cgd * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
23 1.1 cgd * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24 1.1 cgd * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
25 1.1 cgd * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
26 1.1 cgd * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27 1.1 cgd * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28 1.1 cgd * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29 1.1 cgd * SUCH DAMAGE.
30 1.1 cgd *
31 1.17 fvdl * @(#)subr_prof.c 8.4 (Berkeley) 2/14/95
32 1.1 cgd */
33 1.25 lukem
34 1.25 lukem #include <sys/cdefs.h>
35 1.47 christos __KERNEL_RCSID(0, "$NetBSD: subr_prof.c,v 1.47 2014/07/10 21:13:52 christos Exp $");
36 1.1 cgd
37 1.1 cgd #include <sys/param.h>
38 1.1 cgd #include <sys/systm.h>
39 1.1 cgd #include <sys/kernel.h>
40 1.1 cgd #include <sys/proc.h>
41 1.4 cgd #include <sys/mount.h>
42 1.4 cgd #include <sys/syscallargs.h>
43 1.16 jonathan #include <sys/sysctl.h>
44 1.4 cgd
45 1.42 ad #include <sys/cpu.h>
46 1.9 christos
47 1.1 cgd #ifdef GPROF
48 1.1 cgd #include <sys/malloc.h>
49 1.1 cgd #include <sys/gmon.h>
50 1.27 thorpej
51 1.27 thorpej MALLOC_DEFINE(M_GPROF, "gprof", "kernel profiling buffer");
52 1.1 cgd
53 1.1 cgd /*
54 1.1 cgd * Froms is actually a bunch of unsigned shorts indexing tos
55 1.1 cgd */
56 1.37 christos struct gmonparam _gmonparam = { .state = GMON_PROF_OFF };
57 1.1 cgd
58 1.15 gwr /* Actual start of the kernel text segment. */
59 1.15 gwr extern char kernel_text[];
60 1.15 gwr
61 1.1 cgd extern char etext[];
62 1.1 cgd
63 1.9 christos
64 1.9 christos void
65 1.32 thorpej kmstartup(void)
66 1.1 cgd {
67 1.1 cgd char *cp;
68 1.1 cgd struct gmonparam *p = &_gmonparam;
69 1.1 cgd /*
70 1.1 cgd * Round lowpc and highpc to multiples of the density we're using
71 1.1 cgd * so the rest of the scaling (here and in gprof) stays in ints.
72 1.1 cgd */
73 1.34 christos p->lowpc = rounddown(((u_long)kernel_text),
74 1.15 gwr HISTFRACTION * sizeof(HISTCOUNTER));
75 1.34 christos p->highpc = roundup((u_long)etext,
76 1.15 gwr HISTFRACTION * sizeof(HISTCOUNTER));
77 1.1 cgd p->textsize = p->highpc - p->lowpc;
78 1.14 christos printf("Profiling kernel, textsize=%ld [%lx..%lx]\n",
79 1.1 cgd p->textsize, p->lowpc, p->highpc);
80 1.1 cgd p->kcountsize = p->textsize / HISTFRACTION;
81 1.1 cgd p->hashfraction = HASHFRACTION;
82 1.1 cgd p->fromssize = p->textsize / HASHFRACTION;
83 1.1 cgd p->tolimit = p->textsize * ARCDENSITY / 100;
84 1.1 cgd if (p->tolimit < MINARCS)
85 1.1 cgd p->tolimit = MINARCS;
86 1.1 cgd else if (p->tolimit > MAXARCS)
87 1.1 cgd p->tolimit = MAXARCS;
88 1.1 cgd p->tossize = p->tolimit * sizeof(struct tostruct);
89 1.47 christos cp = malloc(p->kcountsize + p->fromssize + p->tossize,
90 1.38 ad M_GPROF, M_NOWAIT | M_ZERO);
91 1.1 cgd if (cp == 0) {
92 1.14 christos printf("No memory for profiling.\n");
93 1.1 cgd return;
94 1.1 cgd }
95 1.1 cgd p->tos = (struct tostruct *)cp;
96 1.1 cgd cp += p->tossize;
97 1.1 cgd p->kcount = (u_short *)cp;
98 1.1 cgd cp += p->kcountsize;
99 1.1 cgd p->froms = (u_short *)cp;
100 1.1 cgd }
101 1.1 cgd
102 1.1 cgd /*
103 1.1 cgd * Return kernel profiling information.
104 1.1 cgd */
105 1.29 atatat /*
106 1.29 atatat * sysctl helper routine for kern.profiling subtree. enables/disables
107 1.29 atatat * kernel profiling and gives out copies of the profiling data.
108 1.29 atatat */
109 1.29 atatat static int
110 1.29 atatat sysctl_kern_profiling(SYSCTLFN_ARGS)
111 1.1 cgd {
112 1.1 cgd struct gmonparam *gp = &_gmonparam;
113 1.1 cgd int error;
114 1.29 atatat struct sysctlnode node;
115 1.1 cgd
116 1.29 atatat node = *rnode;
117 1.23 bjh21
118 1.29 atatat switch (node.sysctl_num) {
119 1.1 cgd case GPROF_STATE:
120 1.29 atatat node.sysctl_data = &gp->state;
121 1.29 atatat break;
122 1.1 cgd case GPROF_COUNT:
123 1.29 atatat node.sysctl_data = gp->kcount;
124 1.29 atatat node.sysctl_size = gp->kcountsize;
125 1.29 atatat break;
126 1.1 cgd case GPROF_FROMS:
127 1.29 atatat node.sysctl_data = gp->froms;
128 1.29 atatat node.sysctl_size = gp->fromssize;
129 1.29 atatat break;
130 1.1 cgd case GPROF_TOS:
131 1.29 atatat node.sysctl_data = gp->tos;
132 1.29 atatat node.sysctl_size = gp->tossize;
133 1.29 atatat break;
134 1.1 cgd case GPROF_GMONPARAM:
135 1.29 atatat node.sysctl_data = gp;
136 1.29 atatat node.sysctl_size = sizeof(*gp);
137 1.29 atatat break;
138 1.1 cgd default:
139 1.1 cgd return (EOPNOTSUPP);
140 1.1 cgd }
141 1.29 atatat
142 1.29 atatat error = sysctl_lookup(SYSCTLFN_CALL(&node));
143 1.29 atatat if (error || newp == NULL)
144 1.29 atatat return (error);
145 1.29 atatat
146 1.29 atatat if (node.sysctl_num == GPROF_STATE) {
147 1.38 ad mutex_spin_enter(&proc0.p_stmutex);
148 1.29 atatat if (gp->state == GMON_PROF_OFF)
149 1.29 atatat stopprofclock(&proc0);
150 1.29 atatat else
151 1.29 atatat startprofclock(&proc0);
152 1.38 ad mutex_spin_exit(&proc0.p_stmutex);
153 1.29 atatat }
154 1.29 atatat
155 1.29 atatat return (0);
156 1.29 atatat }
157 1.29 atatat
158 1.29 atatat SYSCTL_SETUP(sysctl_kern_gprof_setup, "sysctl kern.profiling subtree setup")
159 1.29 atatat {
160 1.29 atatat
161 1.30 atatat sysctl_createv(clog, 0, NULL, NULL,
162 1.30 atatat CTLFLAG_PERMANENT,
163 1.31 atatat CTLTYPE_NODE, "profiling",
164 1.31 atatat SYSCTL_DESCR("Profiling information (available)"),
165 1.29 atatat NULL, 0, NULL, 0,
166 1.29 atatat CTL_KERN, KERN_PROF, CTL_EOL);
167 1.29 atatat
168 1.30 atatat sysctl_createv(clog, 0, NULL, NULL,
169 1.30 atatat CTLFLAG_PERMANENT|CTLFLAG_READWRITE,
170 1.31 atatat CTLTYPE_INT, "state",
171 1.31 atatat SYSCTL_DESCR("Profiling state"),
172 1.29 atatat sysctl_kern_profiling, 0, NULL, 0,
173 1.29 atatat CTL_KERN, KERN_PROF, GPROF_STATE, CTL_EOL);
174 1.30 atatat sysctl_createv(clog, 0, NULL, NULL,
175 1.30 atatat CTLFLAG_PERMANENT|CTLFLAG_READWRITE,
176 1.31 atatat CTLTYPE_STRUCT, "count",
177 1.31 atatat SYSCTL_DESCR("Array of statistical program counters"),
178 1.29 atatat sysctl_kern_profiling, 0, NULL, 0,
179 1.29 atatat CTL_KERN, KERN_PROF, GPROF_COUNT, CTL_EOL);
180 1.30 atatat sysctl_createv(clog, 0, NULL, NULL,
181 1.30 atatat CTLFLAG_PERMANENT|CTLFLAG_READWRITE,
182 1.31 atatat CTLTYPE_STRUCT, "froms",
183 1.31 atatat SYSCTL_DESCR("Array indexed by program counter of "
184 1.31 atatat "call-from points"),
185 1.29 atatat sysctl_kern_profiling, 0, NULL, 0,
186 1.29 atatat CTL_KERN, KERN_PROF, GPROF_FROMS, CTL_EOL);
187 1.30 atatat sysctl_createv(clog, 0, NULL, NULL,
188 1.30 atatat CTLFLAG_PERMANENT|CTLFLAG_READWRITE,
189 1.31 atatat CTLTYPE_STRUCT, "tos",
190 1.31 atatat SYSCTL_DESCR("Array of structures describing "
191 1.31 atatat "destination of calls and their counts"),
192 1.29 atatat sysctl_kern_profiling, 0, NULL, 0,
193 1.29 atatat CTL_KERN, KERN_PROF, GPROF_TOS, CTL_EOL);
194 1.30 atatat sysctl_createv(clog, 0, NULL, NULL,
195 1.30 atatat CTLFLAG_PERMANENT,
196 1.31 atatat CTLTYPE_STRUCT, "gmonparam",
197 1.31 atatat SYSCTL_DESCR("Structure giving the sizes of the above "
198 1.31 atatat "arrays"),
199 1.29 atatat sysctl_kern_profiling, 0, NULL, 0,
200 1.29 atatat CTL_KERN, KERN_PROF, GPROF_GMONPARAM, CTL_EOL);
201 1.1 cgd }
202 1.1 cgd #endif /* GPROF */
203 1.1 cgd
204 1.1 cgd /*
205 1.1 cgd * Profiling system call.
206 1.1 cgd *
207 1.1 cgd * The scale factor is a fixed point number with 16 bits of fraction, so that
208 1.1 cgd * 1.0 is represented as 0x10000. A scale factor of 0 turns off profiling.
209 1.1 cgd */
210 1.1 cgd /* ARGSUSED */
211 1.9 christos int
212 1.43 dsl sys_profil(struct lwp *l, const struct sys_profil_args *uap, register_t *retval)
213 1.6 thorpej {
214 1.43 dsl /* {
215 1.40 drochner syscallarg(char *) samples;
216 1.44 dsl syscallarg(size_t) size;
217 1.44 dsl syscallarg(u_long) offset;
218 1.4 cgd syscallarg(u_int) scale;
219 1.43 dsl } */
220 1.26 thorpej struct proc *p = l->l_proc;
221 1.20 augustss struct uprof *upp;
222 1.1 cgd
223 1.4 cgd if (SCARG(uap, scale) > (1 << 16))
224 1.1 cgd return (EINVAL);
225 1.4 cgd if (SCARG(uap, scale) == 0) {
226 1.38 ad mutex_spin_enter(&p->p_stmutex);
227 1.1 cgd stopprofclock(p);
228 1.38 ad mutex_spin_exit(&p->p_stmutex);
229 1.1 cgd return (0);
230 1.1 cgd }
231 1.1 cgd upp = &p->p_stats->p_prof;
232 1.1 cgd
233 1.1 cgd /* Block profile interrupts while changing state. */
234 1.38 ad mutex_spin_enter(&p->p_stmutex);
235 1.4 cgd upp->pr_off = SCARG(uap, offset);
236 1.4 cgd upp->pr_scale = SCARG(uap, scale);
237 1.4 cgd upp->pr_base = SCARG(uap, samples);
238 1.4 cgd upp->pr_size = SCARG(uap, size);
239 1.1 cgd startprofclock(p);
240 1.38 ad mutex_spin_exit(&p->p_stmutex);
241 1.1 cgd
242 1.1 cgd return (0);
243 1.1 cgd }
244 1.1 cgd
245 1.1 cgd /*
246 1.1 cgd * Scale is a fixed-point number with the binary point 16 bits
247 1.1 cgd * into the value, and is <= 1.0. pc is at most 32 bits, so the
248 1.1 cgd * intermediate result is at most 48 bits.
249 1.1 cgd */
250 1.1 cgd #define PC_TO_INDEX(pc, prof) \
251 1.1 cgd ((int)(((u_quad_t)((pc) - (prof)->pr_off) * \
252 1.1 cgd (u_quad_t)((prof)->pr_scale)) >> 16) & ~1)
253 1.1 cgd
254 1.1 cgd /*
255 1.1 cgd * Collect user-level profiling statistics; called on a profiling tick,
256 1.1 cgd * when a process is running in user-mode. This routine may be called
257 1.1 cgd * from an interrupt context. We try to update the user profiling buffers
258 1.1 cgd * cheaply with fuswintr() and suswintr(). If that fails, we revert to
259 1.1 cgd * an AST that will vector us to trap() with a context in which copyin
260 1.1 cgd * and copyout will work. Trap will then call addupc_task().
261 1.1 cgd *
262 1.1 cgd * Note that we may (rarely) not get around to the AST soon enough, and
263 1.1 cgd * lose profile ticks when the next tick overwrites this one, but in this
264 1.1 cgd * case the system is overloaded and the profile is probably already
265 1.1 cgd * inaccurate.
266 1.1 cgd */
267 1.1 cgd void
268 1.38 ad addupc_intr(struct lwp *l, u_long pc)
269 1.1 cgd {
270 1.20 augustss struct uprof *prof;
271 1.38 ad struct proc *p;
272 1.39 christos void *addr;
273 1.20 augustss u_int i;
274 1.20 augustss int v;
275 1.1 cgd
276 1.38 ad p = l->l_proc;
277 1.38 ad
278 1.41 ad KASSERT(mutex_owned(&p->p_stmutex));
279 1.38 ad
280 1.1 cgd prof = &p->p_stats->p_prof;
281 1.1 cgd if (pc < prof->pr_off ||
282 1.1 cgd (i = PC_TO_INDEX(pc, prof)) >= prof->pr_size)
283 1.1 cgd return; /* out of range; ignore */
284 1.1 cgd
285 1.40 drochner addr = prof->pr_base + i;
286 1.38 ad mutex_spin_exit(&p->p_stmutex);
287 1.22 mycroft if ((v = fuswintr(addr)) == -1 || suswintr(addr, v + 1) == -1) {
288 1.38 ad /* XXXSMP */
289 1.1 cgd prof->pr_addr = pc;
290 1.22 mycroft prof->pr_ticks++;
291 1.38 ad cpu_need_proftick(l);
292 1.1 cgd }
293 1.38 ad mutex_spin_enter(&p->p_stmutex);
294 1.1 cgd }
295 1.1 cgd
296 1.1 cgd /*
297 1.1 cgd * Much like before, but we can afford to take faults here. If the
298 1.1 cgd * update fails, we simply turn off profiling.
299 1.1 cgd */
300 1.1 cgd void
301 1.38 ad addupc_task(struct lwp *l, u_long pc, u_int ticks)
302 1.1 cgd {
303 1.20 augustss struct uprof *prof;
304 1.38 ad struct proc *p;
305 1.39 christos void *addr;
306 1.38 ad int error;
307 1.20 augustss u_int i;
308 1.1 cgd u_short v;
309 1.1 cgd
310 1.38 ad p = l->l_proc;
311 1.38 ad
312 1.38 ad if (ticks == 0)
313 1.1 cgd return;
314 1.1 cgd
315 1.38 ad mutex_spin_enter(&p->p_stmutex);
316 1.1 cgd prof = &p->p_stats->p_prof;
317 1.38 ad
318 1.38 ad /* Testing P_PROFIL may be unnecessary, but is certainly safe. */
319 1.38 ad if ((p->p_stflag & PST_PROFIL) == 0 || pc < prof->pr_off ||
320 1.38 ad (i = PC_TO_INDEX(pc, prof)) >= prof->pr_size) {
321 1.38 ad mutex_spin_exit(&p->p_stmutex);
322 1.1 cgd return;
323 1.38 ad }
324 1.1 cgd
325 1.40 drochner addr = prof->pr_base + i;
326 1.38 ad mutex_spin_exit(&p->p_stmutex);
327 1.39 christos if ((error = copyin(addr, (void *)&v, sizeof(v))) == 0) {
328 1.1 cgd v += ticks;
329 1.39 christos error = copyout((void *)&v, addr, sizeof(v));
330 1.38 ad }
331 1.38 ad if (error != 0) {
332 1.38 ad mutex_spin_enter(&p->p_stmutex);
333 1.38 ad stopprofclock(p);
334 1.38 ad mutex_spin_exit(&p->p_stmutex);
335 1.1 cgd }
336 1.1 cgd }
337