1/*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 *
21 * Portions Copyright 2006-2008 John Birrell jb@freebsd.org
22 *
23 * $FreeBSD$
24 *
25 */
26
27/*
28 * Copyright 2006 Sun Microsystems, Inc.  All rights reserved.
29 * Use is subject to license terms.
30 */
31
32#include <sys/cdefs.h>
33#include <sys/param.h>
34#include <sys/systm.h>
35#include <sys/conf.h>
36#include <sys/cpuvar.h>
37#include <sys/fcntl.h>
38#include <sys/filio.h>
39#include <sys/kdb.h>
40#include <sys/kernel.h>
41#include <sys/kmem.h>
42#include <sys/kthread.h>
43#include <sys/limits.h>
44#include <sys/linker.h>
45#include <sys/lock.h>
46#include <sys/malloc.h>
47#include <sys/module.h>
48#include <sys/mutex.h>
49#include <sys/poll.h>
50#include <sys/proc.h>
51#include <sys/selinfo.h>
52#include <sys/smp.h>
53#include <sys/uio.h>
54#include <sys/unistd.h>
55#include <machine/stdarg.h>
56
57#include <sys/cyclic.h>
58#include <sys/dtrace.h>
59#include <sys/dtrace_bsd.h>
60
61#define	PROF_NAMELEN		15
62
63#define	PROF_PROFILE		0
64#define	PROF_TICK		1
65#define	PROF_PREFIX_PROFILE	"profile-"
66#define	PROF_PREFIX_TICK	"tick-"
67
68/*
69 * Regardless of platform, there are five artificial frames in the case of the
70 * profile provider:
71 *
72 *	profile_fire
73 *	cyclic_expire
74 *	cyclic_fire
75 *	[ cbe ]
76 *	[ locore ]
77 *
78 * On amd64, there are two frames associated with locore:  one in locore, and
79 * another in common interrupt dispatch code.  (i386 has not been modified to
80 * use this common layer.)  Further, on i386, the interrupted instruction
81 * appears as its own stack frame.  All of this means that we need to add one
82 * frame for amd64, and then take one away for both amd64 and i386.
83 *
84 * On SPARC, the picture is further complicated because the compiler
85 * optimizes away tail-calls -- so the following frames are optimized away:
86 *
87 * 	profile_fire
88 *	cyclic_expire
89 *
90 * This gives three frames.  However, on DEBUG kernels, the cyclic_expire
91 * frame cannot be tail-call eliminated, yielding four frames in this case.
92 *
93 * All of the above constraints lead to the mess below.  Yes, the profile
94 * provider should ideally figure this out on-the-fly by hiting one of its own
95 * probes and then walking its own stack trace.  This is complicated, however,
96 * and the static definition doesn't seem to be overly brittle.  Still, we
97 * allow for a manual override in case we get it completely wrong.
98 */
99#ifdef __amd64
100#define	PROF_ARTIFICIAL_FRAMES	7
101#else
102#ifdef __i386
103#define	PROF_ARTIFICIAL_FRAMES	6
104#else
105#ifdef __sparc
106#ifdef DEBUG
107#define	PROF_ARTIFICIAL_FRAMES	4
108#else
109#define	PROF_ARTIFICIAL_FRAMES	3
110#endif
111#endif
112#endif
113#endif
114
115#ifdef __mips
116/*
117 * This value is bogus just to make module compilable on mips
118 */
119#define	PROF_ARTIFICIAL_FRAMES	3
120#endif
121
122#ifdef __powerpc__
123/*
124 * This value is bogus just to make module compilable on powerpc
125 */
126#define	PROF_ARTIFICIAL_FRAMES	3
127#endif
128
129typedef struct profile_probe {
130	char		prof_name[PROF_NAMELEN];
131	dtrace_id_t	prof_id;
132	int		prof_kind;
133	hrtime_t	prof_interval;
134	cyclic_id_t	prof_cyclic;
135} profile_probe_t;
136
137typedef struct profile_probe_percpu {
138	hrtime_t	profc_expected;
139	hrtime_t	profc_interval;
140	profile_probe_t	*profc_probe;
141} profile_probe_percpu_t;
142
143static d_open_t	profile_open;
144static int	profile_unload(void);
145static void	profile_create(hrtime_t, char *, int);
146static void	profile_destroy(void *, dtrace_id_t, void *);
147static void	profile_enable(void *, dtrace_id_t, void *);
148static void	profile_disable(void *, dtrace_id_t, void *);
149static void	profile_load(void *);
150static void	profile_provide(void *, dtrace_probedesc_t *);
151
152static int profile_rates[] = {
153    97, 199, 499, 997, 1999,
154    4001, 4999, 0, 0, 0,
155    0, 0, 0, 0, 0,
156    0, 0, 0, 0, 0
157};
158
159static int profile_ticks[] = {
160    1, 10, 100, 500, 1000,
161    5000, 0, 0, 0, 0,
162    0, 0, 0, 0, 0
163};
164
165/*
166 * profile_max defines the upper bound on the number of profile probes that
167 * can exist (this is to prevent malicious or clumsy users from exhausing
168 * system resources by creating a slew of profile probes). At mod load time,
169 * this gets its value from PROFILE_MAX_DEFAULT or profile-max-probes if it's
170 * present in the profile.conf file.
171 */
172#define	PROFILE_MAX_DEFAULT	1000	/* default max. number of probes */
173static uint32_t profile_max = PROFILE_MAX_DEFAULT;
174					/* maximum number of profile probes */
175static uint32_t profile_total;		/* current number of profile probes */
176
177static struct cdevsw profile_cdevsw = {
178	.d_version	= D_VERSION,
179	.d_open		= profile_open,
180	.d_name		= "profile",
181};
182
183static dtrace_pattr_t profile_attr = {
184{ DTRACE_STABILITY_EVOLVING, DTRACE_STABILITY_EVOLVING, DTRACE_CLASS_COMMON },
185{ DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
186{ DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_ISA },
187{ DTRACE_STABILITY_EVOLVING, DTRACE_STABILITY_EVOLVING, DTRACE_CLASS_COMMON },
188{ DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_ISA },
189};
190
191static dtrace_pops_t profile_pops = {
192	profile_provide,
193	NULL,
194	profile_enable,
195	profile_disable,
196	NULL,
197	NULL,
198	NULL,
199	NULL,
200	NULL,
201	profile_destroy
202};
203
204static struct cdev		*profile_cdev;
205static dtrace_provider_id_t	profile_id;
206static hrtime_t			profile_interval_min = NANOSEC / 5000;	/* 5000 hz */
207static int			profile_aframes = 0;			/* override */
208
209static void
210profile_fire(void *arg)
211{
212	profile_probe_percpu_t *pcpu = arg;
213	profile_probe_t *prof = pcpu->profc_probe;
214	hrtime_t late;
215	solaris_cpu_t *c = &solaris_cpu[curcpu];
216
217	late = gethrtime() - pcpu->profc_expected;
218	pcpu->profc_expected += pcpu->profc_interval;
219
220	dtrace_probe(prof->prof_id, c->cpu_profile_pc,
221	    c->cpu_profile_upc, late, 0, 0);
222}
223
224static void
225profile_tick(void *arg)
226{
227	profile_probe_t *prof = arg;
228	solaris_cpu_t *c = &solaris_cpu[curcpu];
229
230	dtrace_probe(prof->prof_id, c->cpu_profile_pc,
231	    c->cpu_profile_upc, 0, 0, 0);
232}
233
234static void
235profile_create(hrtime_t interval, char *name, int kind)
236{
237	profile_probe_t *prof;
238
239	if (interval < profile_interval_min)
240		return;
241
242	if (dtrace_probe_lookup(profile_id, NULL, NULL, name) != 0)
243		return;
244
245	atomic_add_32(&profile_total, 1);
246	if (profile_total > profile_max) {
247		atomic_add_32(&profile_total, -1);
248		return;
249	}
250
251	prof = kmem_zalloc(sizeof (profile_probe_t), KM_SLEEP);
252	(void) strcpy(prof->prof_name, name);
253	prof->prof_interval = interval;
254	prof->prof_cyclic = CYCLIC_NONE;
255	prof->prof_kind = kind;
256	prof->prof_id = dtrace_probe_create(profile_id,
257	    NULL, NULL, name,
258	    profile_aframes ? profile_aframes : PROF_ARTIFICIAL_FRAMES, prof);
259}
260
261/*ARGSUSED*/
262static void
263profile_provide(void *arg, dtrace_probedesc_t *desc)
264{
265	int i, j, rate, kind;
266	hrtime_t val = 0, mult = 1, len = 0;
267	char *name, *suffix = NULL;
268
269	const struct {
270		char *prefix;
271		int kind;
272	} types[] = {
273		{ PROF_PREFIX_PROFILE, PROF_PROFILE },
274		{ PROF_PREFIX_TICK, PROF_TICK },
275		{ 0, 0 }
276	};
277
278	const struct {
279		char *name;
280		hrtime_t mult;
281	} suffixes[] = {
282		{ "ns", 	NANOSEC / NANOSEC },
283		{ "nsec",	NANOSEC / NANOSEC },
284		{ "us",		NANOSEC / MICROSEC },
285		{ "usec",	NANOSEC / MICROSEC },
286		{ "ms",		NANOSEC / MILLISEC },
287		{ "msec",	NANOSEC / MILLISEC },
288		{ "s",		NANOSEC / SEC },
289		{ "sec",	NANOSEC / SEC },
290		{ "m",		NANOSEC * (hrtime_t)60 },
291		{ "min",	NANOSEC * (hrtime_t)60 },
292		{ "h",		NANOSEC * (hrtime_t)(60 * 60) },
293		{ "hour",	NANOSEC * (hrtime_t)(60 * 60) },
294		{ "d",		NANOSEC * (hrtime_t)(24 * 60 * 60) },
295		{ "day",	NANOSEC * (hrtime_t)(24 * 60 * 60) },
296		{ "hz",		0 },
297		{ NULL }
298	};
299
300	if (desc == NULL) {
301		char n[PROF_NAMELEN];
302
303		/*
304		 * If no description was provided, provide all of our probes.
305		 */
306		for (i = 0; i < sizeof (profile_rates) / sizeof (int); i++) {
307			if ((rate = profile_rates[i]) == 0)
308				continue;
309
310			(void) snprintf(n, PROF_NAMELEN, "%s%d",
311			    PROF_PREFIX_PROFILE, rate);
312			profile_create(NANOSEC / rate, n, PROF_PROFILE);
313		}
314
315		for (i = 0; i < sizeof (profile_ticks) / sizeof (int); i++) {
316			if ((rate = profile_ticks[i]) == 0)
317				continue;
318
319			(void) snprintf(n, PROF_NAMELEN, "%s%d",
320			    PROF_PREFIX_TICK, rate);
321			profile_create(NANOSEC / rate, n, PROF_TICK);
322		}
323
324		return;
325	}
326
327	name = desc->dtpd_name;
328
329	for (i = 0; types[i].prefix != NULL; i++) {
330		len = strlen(types[i].prefix);
331
332		if (strncmp(name, types[i].prefix, len) != 0)
333			continue;
334		break;
335	}
336
337	if (types[i].prefix == NULL)
338		return;
339
340	kind = types[i].kind;
341	j = strlen(name) - len;
342
343	/*
344	 * We need to start before any time suffix.
345	 */
346	for (j = strlen(name); j >= len; j--) {
347		if (name[j] >= '0' && name[j] <= '9')
348			break;
349		suffix = &name[j];
350	}
351
352	ASSERT(suffix != NULL);
353
354	/*
355	 * Now determine the numerical value present in the probe name.
356	 */
357	for (; j >= len; j--) {
358		if (name[j] < '0' || name[j] > '9')
359			return;
360
361		val += (name[j] - '0') * mult;
362		mult *= (hrtime_t)10;
363	}
364
365	if (val == 0)
366		return;
367
368	/*
369	 * Look-up the suffix to determine the multiplier.
370	 */
371	for (i = 0, mult = 0; suffixes[i].name != NULL; i++) {
372		if (strcasecmp(suffixes[i].name, suffix) == 0) {
373			mult = suffixes[i].mult;
374			break;
375		}
376	}
377
378	if (suffixes[i].name == NULL && *suffix != '\0')
379		return;
380
381	if (mult == 0) {
382		/*
383		 * The default is frequency-per-second.
384		 */
385		val = NANOSEC / val;
386	} else {
387		val *= mult;
388	}
389
390	profile_create(val, name, kind);
391}
392
393/* ARGSUSED */
394static void
395profile_destroy(void *arg, dtrace_id_t id, void *parg)
396{
397	profile_probe_t *prof = parg;
398
399	ASSERT(prof->prof_cyclic == CYCLIC_NONE);
400	kmem_free(prof, sizeof (profile_probe_t));
401
402	ASSERT(profile_total >= 1);
403	atomic_add_32(&profile_total, -1);
404}
405
406/*ARGSUSED*/
407static void
408profile_online(void *arg, cpu_t *cpu, cyc_handler_t *hdlr, cyc_time_t *when)
409{
410	profile_probe_t *prof = arg;
411	profile_probe_percpu_t *pcpu;
412
413	pcpu = kmem_zalloc(sizeof (profile_probe_percpu_t), KM_SLEEP);
414	pcpu->profc_probe = prof;
415
416	hdlr->cyh_func = profile_fire;
417	hdlr->cyh_arg = pcpu;
418
419	when->cyt_interval = prof->prof_interval;
420	when->cyt_when = gethrtime() + when->cyt_interval;
421
422	pcpu->profc_expected = when->cyt_when;
423	pcpu->profc_interval = when->cyt_interval;
424}
425
426/*ARGSUSED*/
427static void
428profile_offline(void *arg, cpu_t *cpu, void *oarg)
429{
430	profile_probe_percpu_t *pcpu = oarg;
431
432	ASSERT(pcpu->profc_probe == arg);
433	kmem_free(pcpu, sizeof (profile_probe_percpu_t));
434}
435
436/* ARGSUSED */
437static void
438profile_enable(void *arg, dtrace_id_t id, void *parg)
439{
440	profile_probe_t *prof = parg;
441	cyc_omni_handler_t omni;
442	cyc_handler_t hdlr;
443	cyc_time_t when;
444
445	ASSERT(prof->prof_interval != 0);
446	ASSERT(MUTEX_HELD(&cpu_lock));
447
448	if (prof->prof_kind == PROF_TICK) {
449		hdlr.cyh_func = profile_tick;
450		hdlr.cyh_arg = prof;
451
452		when.cyt_interval = prof->prof_interval;
453		when.cyt_when = gethrtime() + when.cyt_interval;
454	} else {
455		ASSERT(prof->prof_kind == PROF_PROFILE);
456		omni.cyo_online = profile_online;
457		omni.cyo_offline = profile_offline;
458		omni.cyo_arg = prof;
459	}
460
461	if (prof->prof_kind == PROF_TICK) {
462		prof->prof_cyclic = cyclic_add(&hdlr, &when);
463	} else {
464		prof->prof_cyclic = cyclic_add_omni(&omni);
465	}
466}
467
468/* ARGSUSED */
469static void
470profile_disable(void *arg, dtrace_id_t id, void *parg)
471{
472	profile_probe_t *prof = parg;
473
474	ASSERT(prof->prof_cyclic != CYCLIC_NONE);
475	ASSERT(MUTEX_HELD(&cpu_lock));
476
477	cyclic_remove(prof->prof_cyclic);
478	prof->prof_cyclic = CYCLIC_NONE;
479}
480
481static void
482profile_load(void *dummy)
483{
484	/* Create the /dev/dtrace/profile entry. */
485	profile_cdev = make_dev(&profile_cdevsw, 0, UID_ROOT, GID_WHEEL, 0600,
486	    "dtrace/profile");
487
488	if (dtrace_register("profile", &profile_attr, DTRACE_PRIV_USER,
489	    NULL, &profile_pops, NULL, &profile_id) != 0)
490		return;
491}
492
493
494static int
495profile_unload()
496{
497	int error = 0;
498
499	if ((error = dtrace_unregister(profile_id)) != 0)
500		return (error);
501
502	destroy_dev(profile_cdev);
503
504	return (error);
505}
506
507/* ARGSUSED */
508static int
509profile_modevent(module_t mod __unused, int type, void *data __unused)
510{
511	int error = 0;
512
513	switch (type) {
514	case MOD_LOAD:
515		break;
516
517	case MOD_UNLOAD:
518		break;
519
520	case MOD_SHUTDOWN:
521		break;
522
523	default:
524		error = EOPNOTSUPP;
525		break;
526
527	}
528	return (error);
529}
530
531/* ARGSUSED */
532static int
533profile_open(struct cdev *dev __unused, int oflags __unused, int devtype __unused, struct thread *td __unused)
534{
535	return (0);
536}
537
538SYSINIT(profile_load, SI_SUB_DTRACE_PROVIDER, SI_ORDER_ANY, profile_load, NULL);
539SYSUNINIT(profile_unload, SI_SUB_DTRACE_PROVIDER, SI_ORDER_ANY, profile_unload, NULL);
540
541DEV_MODULE(profile, profile_modevent, NULL);
542MODULE_VERSION(profile, 1);
543MODULE_DEPEND(profile, dtrace, 1, 1, 1);
544MODULE_DEPEND(profile, cyclic, 1, 1, 1);
545MODULE_DEPEND(profile, opensolaris, 1, 1, 1);
546