xref: /titanic_41/usr/src/uts/common/dtrace/profile.c (revision 5ce5f3670f7934e376808da0d1309924ecf8f9e5)
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 /*
22  * Copyright 2007 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 #pragma ident	"%Z%%M%	%I%	%E% SMI"
27 
28 #include <sys/errno.h>
29 #include <sys/stat.h>
30 #include <sys/modctl.h>
31 #include <sys/conf.h>
32 #include <sys/systm.h>
33 #include <sys/ddi.h>
34 #include <sys/sunddi.h>
35 #include <sys/cpuvar.h>
36 #include <sys/kmem.h>
37 #include <sys/strsubr.h>
38 #include <sys/dtrace.h>
39 #include <sys/cyclic.h>
40 #include <sys/atomic.h>
41 
42 static dev_info_t *profile_devi;
43 static dtrace_provider_id_t profile_id;
44 
45 /*
46  * Regardless of platform, the stack frames look like this in the case of the
47  * profile provider:
48  *
49  *	profile_fire
50  *	cyclic_expire
51  *	cyclic_fire
52  *	[ cbe ]
53  *	[ interrupt code ]
54  *
55  * On x86, there are five frames from the generic interrupt code; further, the
56  * interrupted instruction appears as its own stack frame, giving us a total of
57  * 10.
58  *
59  * On SPARC, the picture is further complicated because the compiler
60  * optimizes away tail-calls -- so the following frames are optimized away:
61  *
62  * 	profile_fire
63  *	cyclic_expire
64  *
65  * This gives three frames.  However, on DEBUG kernels, the cyclic_expire
66  * frame cannot be tail-call eliminated, yielding four frames in this case.
67  *
68  * All of the above constraints lead to the mess below.  Yes, the profile
69  * provider should ideally figure this out on-the-fly by hitting one of its own
70  * probes and then walking its own stack trace.  This is complicated, however,
71  * and the static definition doesn't seem to be overly brittle.  Still, we
72  * allow for a manual override in case we get it completely wrong.
73  */
74 #ifdef __x86
75 #define	PROF_ARTIFICIAL_FRAMES	10
76 #else
77 #ifdef __sparc
78 #ifdef DEBUG
79 #define	PROF_ARTIFICIAL_FRAMES	4
80 #else
81 #define	PROF_ARTIFICIAL_FRAMES	3
82 #endif
83 #endif
84 #endif
85 
86 #define	PROF_NAMELEN		15
87 
88 #define	PROF_PROFILE		0
89 #define	PROF_TICK		1
90 #define	PROF_PREFIX_PROFILE	"profile-"
91 #define	PROF_PREFIX_TICK	"tick-"
92 
93 typedef struct profile_probe {
94 	char		prof_name[PROF_NAMELEN];
95 	dtrace_id_t	prof_id;
96 	int		prof_kind;
97 	hrtime_t	prof_interval;
98 	cyclic_id_t	prof_cyclic;
99 } profile_probe_t;
100 
101 typedef struct profile_probe_percpu {
102 	hrtime_t	profc_expected;
103 	hrtime_t	profc_interval;
104 	profile_probe_t	*profc_probe;
105 } profile_probe_percpu_t;
106 
107 hrtime_t	profile_interval_min = NANOSEC / 5000;		/* 5000 hz */
108 int		profile_aframes = 0;				/* override */
109 
110 static int profile_rates[] = {
111     97, 199, 499, 997, 1999,
112     4001, 4999, 0, 0, 0,
113     0, 0, 0, 0, 0,
114     0, 0, 0, 0, 0
115 };
116 
117 static int profile_ticks[] = {
118     1, 10, 100, 500, 1000,
119     5000, 0, 0, 0, 0,
120     0, 0, 0, 0, 0
121 };
122 
123 /*
124  * profile_max defines the upper bound on the number of profile probes that
125  * can exist (this is to prevent malicious or clumsy users from exhausing
126  * system resources by creating a slew of profile probes). At mod load time,
127  * this gets its value from PROFILE_MAX_DEFAULT or profile-max-probes if it's
128  * present in the profile.conf file.
129  */
130 #define	PROFILE_MAX_DEFAULT	1000	/* default max. number of probes */
131 static uint32_t profile_max;		/* maximum number of profile probes */
132 static uint32_t profile_total;	/* current number of profile probes */
133 
134 static void
135 profile_fire(void *arg)
136 {
137 	profile_probe_percpu_t *pcpu = arg;
138 	profile_probe_t *prof = pcpu->profc_probe;
139 	hrtime_t late;
140 
141 	late = dtrace_gethrtime() - pcpu->profc_expected;
142 	pcpu->profc_expected += pcpu->profc_interval;
143 
144 	dtrace_probe(prof->prof_id, CPU->cpu_profile_pc,
145 	    CPU->cpu_profile_upc, late, 0, 0);
146 }
147 
148 static void
149 profile_tick(void *arg)
150 {
151 	profile_probe_t *prof = arg;
152 
153 	dtrace_probe(prof->prof_id, CPU->cpu_profile_pc,
154 	    CPU->cpu_profile_upc, 0, 0, 0);
155 }
156 
157 static void
158 profile_create(hrtime_t interval, const char *name, int kind)
159 {
160 	profile_probe_t *prof;
161 	int nr_frames = PROF_ARTIFICIAL_FRAMES + dtrace_mach_aframes();
162 
163 	if (profile_aframes)
164 		nr_frames = profile_aframes;
165 
166 	if (interval < profile_interval_min)
167 		return;
168 
169 	if (dtrace_probe_lookup(profile_id, NULL, NULL, name) != 0)
170 		return;
171 
172 	atomic_add_32(&profile_total, 1);
173 	if (profile_total > profile_max) {
174 		atomic_add_32(&profile_total, -1);
175 		return;
176 	}
177 
178 	prof = kmem_zalloc(sizeof (profile_probe_t), KM_SLEEP);
179 	(void) strcpy(prof->prof_name, name);
180 	prof->prof_interval = interval;
181 	prof->prof_cyclic = CYCLIC_NONE;
182 	prof->prof_kind = kind;
183 	prof->prof_id = dtrace_probe_create(profile_id,
184 	    NULL, NULL, name, nr_frames, prof);
185 }
186 
187 /*ARGSUSED*/
188 static void
189 profile_provide(void *arg, const dtrace_probedesc_t *desc)
190 {
191 	int i, j, rate, kind;
192 	hrtime_t val = 0, mult = 1, len;
193 	const char *name, *suffix = NULL;
194 
195 	const struct {
196 		char *prefix;
197 		int kind;
198 	} types[] = {
199 		{ PROF_PREFIX_PROFILE, PROF_PROFILE },
200 		{ PROF_PREFIX_TICK, PROF_TICK },
201 		{ NULL, NULL }
202 	};
203 
204 	const struct {
205 		char *name;
206 		hrtime_t mult;
207 	} suffixes[] = {
208 		{ "ns", 	NANOSEC / NANOSEC },
209 		{ "nsec",	NANOSEC / NANOSEC },
210 		{ "us",		NANOSEC / MICROSEC },
211 		{ "usec",	NANOSEC / MICROSEC },
212 		{ "ms",		NANOSEC / MILLISEC },
213 		{ "msec",	NANOSEC / MILLISEC },
214 		{ "s",		NANOSEC / SEC },
215 		{ "sec",	NANOSEC / SEC },
216 		{ "m",		NANOSEC * (hrtime_t)60 },
217 		{ "min",	NANOSEC * (hrtime_t)60 },
218 		{ "h",		NANOSEC * (hrtime_t)(60 * 60) },
219 		{ "hour",	NANOSEC * (hrtime_t)(60 * 60) },
220 		{ "d",		NANOSEC * (hrtime_t)(24 * 60 * 60) },
221 		{ "day",	NANOSEC * (hrtime_t)(24 * 60 * 60) },
222 		{ "hz",		0 },
223 		{ NULL }
224 	};
225 
226 	if (desc == NULL) {
227 		char n[PROF_NAMELEN];
228 
229 		/*
230 		 * If no description was provided, provide all of our probes.
231 		 */
232 		for (i = 0; i < sizeof (profile_rates) / sizeof (int); i++) {
233 			if ((rate = profile_rates[i]) == 0)
234 				continue;
235 
236 			(void) snprintf(n, PROF_NAMELEN, "%s%d",
237 			    PROF_PREFIX_PROFILE, rate);
238 			profile_create(NANOSEC / rate, n, PROF_PROFILE);
239 		}
240 
241 		for (i = 0; i < sizeof (profile_ticks) / sizeof (int); i++) {
242 			if ((rate = profile_ticks[i]) == 0)
243 				continue;
244 
245 			(void) snprintf(n, PROF_NAMELEN, "%s%d",
246 			    PROF_PREFIX_TICK, rate);
247 			profile_create(NANOSEC / rate, n, PROF_TICK);
248 		}
249 
250 		return;
251 	}
252 
253 	name = desc->dtpd_name;
254 
255 	for (i = 0; types[i].prefix != NULL; i++) {
256 		len = strlen(types[i].prefix);
257 
258 		if (strncmp(name, types[i].prefix, len) != 0)
259 			continue;
260 		break;
261 	}
262 
263 	if (types[i].prefix == NULL)
264 		return;
265 
266 	kind = types[i].kind;
267 	j = strlen(name) - len;
268 
269 	/*
270 	 * We need to start before any time suffix.
271 	 */
272 	for (j = strlen(name); j >= len; j--) {
273 		if (name[j] >= '0' && name[j] <= '9')
274 			break;
275 		suffix = &name[j];
276 	}
277 
278 	ASSERT(suffix != NULL);
279 
280 	/*
281 	 * Now determine the numerical value present in the probe name.
282 	 */
283 	for (; j >= len; j--) {
284 		if (name[j] < '0' || name[j] > '9')
285 			return;
286 
287 		val += (name[j] - '0') * mult;
288 		mult *= (hrtime_t)10;
289 	}
290 
291 	if (val == 0)
292 		return;
293 
294 	/*
295 	 * Look-up the suffix to determine the multiplier.
296 	 */
297 	for (i = 0, mult = 0; suffixes[i].name != NULL; i++) {
298 		if (strcasecmp(suffixes[i].name, suffix) == 0) {
299 			mult = suffixes[i].mult;
300 			break;
301 		}
302 	}
303 
304 	if (suffixes[i].name == NULL && *suffix != '\0')
305 		return;
306 
307 	if (mult == 0) {
308 		/*
309 		 * The default is frequency-per-second.
310 		 */
311 		val = NANOSEC / val;
312 	} else {
313 		val *= mult;
314 	}
315 
316 	profile_create(val, name, kind);
317 }
318 
319 /*ARGSUSED*/
320 static void
321 profile_destroy(void *arg, dtrace_id_t id, void *parg)
322 {
323 	profile_probe_t *prof = parg;
324 
325 	ASSERT(prof->prof_cyclic == CYCLIC_NONE);
326 	kmem_free(prof, sizeof (profile_probe_t));
327 
328 	ASSERT(profile_total >= 1);
329 	atomic_add_32(&profile_total, -1);
330 }
331 
332 /*ARGSUSED*/
333 static void
334 profile_online(void *arg, cpu_t *cpu, cyc_handler_t *hdlr, cyc_time_t *when)
335 {
336 	profile_probe_t *prof = arg;
337 	profile_probe_percpu_t *pcpu;
338 
339 	pcpu = kmem_zalloc(sizeof (profile_probe_percpu_t), KM_SLEEP);
340 	pcpu->profc_probe = prof;
341 
342 	hdlr->cyh_func = profile_fire;
343 	hdlr->cyh_arg = pcpu;
344 	hdlr->cyh_level = CY_HIGH_LEVEL;
345 
346 	when->cyt_interval = prof->prof_interval;
347 	when->cyt_when = dtrace_gethrtime() + when->cyt_interval;
348 
349 	pcpu->profc_expected = when->cyt_when;
350 	pcpu->profc_interval = when->cyt_interval;
351 }
352 
353 /*ARGSUSED*/
354 static void
355 profile_offline(void *arg, cpu_t *cpu, void *oarg)
356 {
357 	profile_probe_percpu_t *pcpu = oarg;
358 
359 	ASSERT(pcpu->profc_probe == arg);
360 	kmem_free(pcpu, sizeof (profile_probe_percpu_t));
361 }
362 
363 /*ARGSUSED*/
364 static void
365 profile_enable(void *arg, dtrace_id_t id, void *parg)
366 {
367 	profile_probe_t *prof = parg;
368 	cyc_omni_handler_t omni;
369 	cyc_handler_t hdlr;
370 	cyc_time_t when;
371 
372 	ASSERT(prof->prof_interval != 0);
373 	ASSERT(MUTEX_HELD(&cpu_lock));
374 
375 	if (prof->prof_kind == PROF_TICK) {
376 		hdlr.cyh_func = profile_tick;
377 		hdlr.cyh_arg = prof;
378 		hdlr.cyh_level = CY_HIGH_LEVEL;
379 
380 		when.cyt_interval = prof->prof_interval;
381 		when.cyt_when = dtrace_gethrtime() + when.cyt_interval;
382 	} else {
383 		ASSERT(prof->prof_kind == PROF_PROFILE);
384 		omni.cyo_online = profile_online;
385 		omni.cyo_offline = profile_offline;
386 		omni.cyo_arg = prof;
387 	}
388 
389 	if (prof->prof_kind == PROF_TICK) {
390 		prof->prof_cyclic = cyclic_add(&hdlr, &when);
391 	} else {
392 		prof->prof_cyclic = cyclic_add_omni(&omni);
393 	}
394 }
395 
396 /*ARGSUSED*/
397 static void
398 profile_disable(void *arg, dtrace_id_t id, void *parg)
399 {
400 	profile_probe_t *prof = parg;
401 
402 	ASSERT(prof->prof_cyclic != CYCLIC_NONE);
403 	ASSERT(MUTEX_HELD(&cpu_lock));
404 
405 	cyclic_remove(prof->prof_cyclic);
406 	prof->prof_cyclic = CYCLIC_NONE;
407 }
408 
409 /*ARGSUSED*/
410 static int
411 profile_usermode(void *arg, dtrace_id_t id, void *parg)
412 {
413 	return (CPU->cpu_profile_pc == 0);
414 }
415 
416 static dtrace_pattr_t profile_attr = {
417 { DTRACE_STABILITY_EVOLVING, DTRACE_STABILITY_EVOLVING, DTRACE_CLASS_COMMON },
418 { DTRACE_STABILITY_UNSTABLE, DTRACE_STABILITY_UNSTABLE, DTRACE_CLASS_UNKNOWN },
419 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
420 { DTRACE_STABILITY_EVOLVING, DTRACE_STABILITY_EVOLVING, DTRACE_CLASS_COMMON },
421 { DTRACE_STABILITY_EVOLVING, DTRACE_STABILITY_EVOLVING, DTRACE_CLASS_COMMON },
422 };
423 
424 static dtrace_pops_t profile_pops = {
425 	profile_provide,
426 	NULL,
427 	profile_enable,
428 	profile_disable,
429 	NULL,
430 	NULL,
431 	NULL,
432 	NULL,
433 	profile_usermode,
434 	profile_destroy
435 };
436 
437 static int
438 profile_attach(dev_info_t *devi, ddi_attach_cmd_t cmd)
439 {
440 	switch (cmd) {
441 	case DDI_ATTACH:
442 		break;
443 	case DDI_RESUME:
444 		return (DDI_SUCCESS);
445 	default:
446 		return (DDI_FAILURE);
447 	}
448 
449 	if (ddi_create_minor_node(devi, "profile", S_IFCHR, 0,
450 	    DDI_PSEUDO, NULL) == DDI_FAILURE ||
451 	    dtrace_register("profile", &profile_attr,
452 	    DTRACE_PRIV_KERNEL | DTRACE_PRIV_USER, NULL,
453 	    &profile_pops, NULL, &profile_id) != 0) {
454 		ddi_remove_minor_node(devi, NULL);
455 		return (DDI_FAILURE);
456 	}
457 
458 	profile_max = ddi_getprop(DDI_DEV_T_ANY, devi, DDI_PROP_DONTPASS,
459 	    "profile-max-probes", PROFILE_MAX_DEFAULT);
460 
461 	ddi_report_dev(devi);
462 	profile_devi = devi;
463 	return (DDI_SUCCESS);
464 }
465 
466 static int
467 profile_detach(dev_info_t *devi, ddi_detach_cmd_t cmd)
468 {
469 	switch (cmd) {
470 	case DDI_DETACH:
471 		break;
472 	case DDI_SUSPEND:
473 		return (DDI_SUCCESS);
474 	default:
475 		return (DDI_FAILURE);
476 	}
477 
478 	if (dtrace_unregister(profile_id) != 0)
479 		return (DDI_FAILURE);
480 
481 	ddi_remove_minor_node(devi, NULL);
482 	return (DDI_SUCCESS);
483 }
484 
485 /*ARGSUSED*/
486 static int
487 profile_info(dev_info_t *dip, ddi_info_cmd_t infocmd, void *arg, void **result)
488 {
489 	int error;
490 
491 	switch (infocmd) {
492 	case DDI_INFO_DEVT2DEVINFO:
493 		*result = (void *)profile_devi;
494 		error = DDI_SUCCESS;
495 		break;
496 	case DDI_INFO_DEVT2INSTANCE:
497 		*result = (void *)0;
498 		error = DDI_SUCCESS;
499 		break;
500 	default:
501 		error = DDI_FAILURE;
502 	}
503 	return (error);
504 }
505 
506 /*ARGSUSED*/
507 static int
508 profile_open(dev_t *devp, int flag, int otyp, cred_t *cred_p)
509 {
510 	return (0);
511 }
512 
513 static struct cb_ops profile_cb_ops = {
514 	profile_open,		/* open */
515 	nodev,			/* close */
516 	nulldev,		/* strategy */
517 	nulldev,		/* print */
518 	nodev,			/* dump */
519 	nodev,			/* read */
520 	nodev,			/* write */
521 	nodev,			/* ioctl */
522 	nodev,			/* devmap */
523 	nodev,			/* mmap */
524 	nodev,			/* segmap */
525 	nochpoll,		/* poll */
526 	ddi_prop_op,		/* cb_prop_op */
527 	0,			/* streamtab  */
528 	D_NEW | D_MP		/* Driver compatibility flag */
529 };
530 
531 static struct dev_ops profile_ops = {
532 	DEVO_REV,		/* devo_rev, */
533 	0,			/* refcnt  */
534 	profile_info,		/* get_dev_info */
535 	nulldev,		/* identify */
536 	nulldev,		/* probe */
537 	profile_attach,		/* attach */
538 	profile_detach,		/* detach */
539 	nodev,			/* reset */
540 	&profile_cb_ops,	/* driver operations */
541 	NULL,			/* bus operations */
542 	nodev			/* dev power */
543 };
544 
545 /*
546  * Module linkage information for the kernel.
547  */
548 static struct modldrv modldrv = {
549 	&mod_driverops,		/* module type (this is a pseudo driver) */
550 	"Profile Interrupt Tracing",	/* name of module */
551 	&profile_ops,		/* driver ops */
552 };
553 
554 static struct modlinkage modlinkage = {
555 	MODREV_1,
556 	(void *)&modldrv,
557 	NULL
558 };
559 
560 int
561 _init(void)
562 {
563 	return (mod_install(&modlinkage));
564 }
565 
566 int
567 _info(struct modinfo *modinfop)
568 {
569 	return (mod_info(&modlinkage, modinfop));
570 }
571 
572 int
573 _fini(void)
574 {
575 	return (mod_remove(&modlinkage));
576 }
577