xref: /freebsd/sys/kern/kern_intr.c (revision aa64588d28258aef88cc33b8043112e8856948d0)
1 /*-
2  * Copyright (c) 1997, Stefan Esser <se@freebsd.org>
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice unmodified, this list of conditions, and the following
10  *    disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
16  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
17  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
18  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
19  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
20  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
21  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
22  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
23  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
24  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
25  */
26 
27 #include <sys/cdefs.h>
28 __FBSDID("$FreeBSD$");
29 
30 #include "opt_ddb.h"
31 
32 #include <sys/param.h>
33 #include <sys/bus.h>
34 #include <sys/conf.h>
35 #include <sys/cpuset.h>
36 #include <sys/rtprio.h>
37 #include <sys/systm.h>
38 #include <sys/interrupt.h>
39 #include <sys/kernel.h>
40 #include <sys/kthread.h>
41 #include <sys/ktr.h>
42 #include <sys/limits.h>
43 #include <sys/lock.h>
44 #include <sys/malloc.h>
45 #include <sys/mutex.h>
46 #include <sys/priv.h>
47 #include <sys/proc.h>
48 #include <sys/random.h>
49 #include <sys/resourcevar.h>
50 #include <sys/sched.h>
51 #include <sys/smp.h>
52 #include <sys/sysctl.h>
53 #include <sys/syslog.h>
54 #include <sys/unistd.h>
55 #include <sys/vmmeter.h>
56 #include <machine/atomic.h>
57 #include <machine/cpu.h>
58 #include <machine/md_var.h>
59 #include <machine/stdarg.h>
60 #ifdef DDB
61 #include <ddb/ddb.h>
62 #include <ddb/db_sym.h>
63 #endif
64 
65 /*
66  * Describe an interrupt thread.  There is one of these per interrupt event.
67  */
68 struct intr_thread {
69 	struct intr_event *it_event;
70 	struct thread *it_thread;	/* Kernel thread. */
71 	int	it_flags;		/* (j) IT_* flags. */
72 	int	it_need;		/* Needs service. */
73 };
74 
75 /* Interrupt thread flags kept in it_flags */
76 #define	IT_DEAD		0x000001	/* Thread is waiting to exit. */
77 
78 struct	intr_entropy {
79 	struct	thread *td;
80 	uintptr_t event;
81 };
82 
83 struct	intr_event *clk_intr_event;
84 struct	intr_event *tty_intr_event;
85 void	*vm_ih;
86 struct proc *intrproc;
87 
88 static MALLOC_DEFINE(M_ITHREAD, "ithread", "Interrupt Threads");
89 
90 static int intr_storm_threshold = 1000;
91 TUNABLE_INT("hw.intr_storm_threshold", &intr_storm_threshold);
92 SYSCTL_INT(_hw, OID_AUTO, intr_storm_threshold, CTLFLAG_RW,
93     &intr_storm_threshold, 0,
94     "Number of consecutive interrupts before storm protection is enabled");
95 static TAILQ_HEAD(, intr_event) event_list =
96     TAILQ_HEAD_INITIALIZER(event_list);
97 static struct mtx event_lock;
98 MTX_SYSINIT(intr_event_list, &event_lock, "intr event list", MTX_DEF);
99 
100 static void	intr_event_update(struct intr_event *ie);
101 #ifdef INTR_FILTER
102 static int	intr_event_schedule_thread(struct intr_event *ie,
103 		    struct intr_thread *ithd);
104 static int	intr_filter_loop(struct intr_event *ie,
105 		    struct trapframe *frame, struct intr_thread **ithd);
106 static struct intr_thread *ithread_create(const char *name,
107 			      struct intr_handler *ih);
108 #else
109 static int	intr_event_schedule_thread(struct intr_event *ie);
110 static struct intr_thread *ithread_create(const char *name);
111 #endif
112 static void	ithread_destroy(struct intr_thread *ithread);
113 static void	ithread_execute_handlers(struct proc *p,
114 		    struct intr_event *ie);
115 #ifdef INTR_FILTER
116 static void	priv_ithread_execute_handler(struct proc *p,
117 		    struct intr_handler *ih);
118 #endif
119 static void	ithread_loop(void *);
120 static void	ithread_update(struct intr_thread *ithd);
121 static void	start_softintr(void *);
122 
123 /* Map an interrupt type to an ithread priority. */
124 u_char
125 intr_priority(enum intr_type flags)
126 {
127 	u_char pri;
128 
129 	flags &= (INTR_TYPE_TTY | INTR_TYPE_BIO | INTR_TYPE_NET |
130 	    INTR_TYPE_CAM | INTR_TYPE_MISC | INTR_TYPE_CLK | INTR_TYPE_AV);
131 	switch (flags) {
132 	case INTR_TYPE_TTY:
133 		pri = PI_TTYLOW;
134 		break;
135 	case INTR_TYPE_BIO:
136 		/*
137 		 * XXX We need to refine this.  BSD/OS distinguishes
138 		 * between tape and disk priorities.
139 		 */
140 		pri = PI_DISK;
141 		break;
142 	case INTR_TYPE_NET:
143 		pri = PI_NET;
144 		break;
145 	case INTR_TYPE_CAM:
146 		pri = PI_DISK;          /* XXX or PI_CAM? */
147 		break;
148 	case INTR_TYPE_AV:		/* Audio/video */
149 		pri = PI_AV;
150 		break;
151 	case INTR_TYPE_CLK:
152 		pri = PI_REALTIME;
153 		break;
154 	case INTR_TYPE_MISC:
155 		pri = PI_DULL;          /* don't care */
156 		break;
157 	default:
158 		/* We didn't specify an interrupt level. */
159 		panic("intr_priority: no interrupt type in flags");
160 	}
161 
162 	return pri;
163 }
164 
165 /*
166  * Update an ithread based on the associated intr_event.
167  */
168 static void
169 ithread_update(struct intr_thread *ithd)
170 {
171 	struct intr_event *ie;
172 	struct thread *td;
173 	u_char pri;
174 
175 	ie = ithd->it_event;
176 	td = ithd->it_thread;
177 
178 	/* Determine the overall priority of this event. */
179 	if (TAILQ_EMPTY(&ie->ie_handlers))
180 		pri = PRI_MAX_ITHD;
181 	else
182 		pri = TAILQ_FIRST(&ie->ie_handlers)->ih_pri;
183 
184 	/* Update name and priority. */
185 	strlcpy(td->td_name, ie->ie_fullname, sizeof(td->td_name));
186 	thread_lock(td);
187 	sched_prio(td, pri);
188 	thread_unlock(td);
189 }
190 
191 /*
192  * Regenerate the full name of an interrupt event and update its priority.
193  */
194 static void
195 intr_event_update(struct intr_event *ie)
196 {
197 	struct intr_handler *ih;
198 	char *last;
199 	int missed, space;
200 
201 	/* Start off with no entropy and just the name of the event. */
202 	mtx_assert(&ie->ie_lock, MA_OWNED);
203 	strlcpy(ie->ie_fullname, ie->ie_name, sizeof(ie->ie_fullname));
204 	ie->ie_flags &= ~IE_ENTROPY;
205 	missed = 0;
206 	space = 1;
207 
208 	/* Run through all the handlers updating values. */
209 	TAILQ_FOREACH(ih, &ie->ie_handlers, ih_next) {
210 		if (strlen(ie->ie_fullname) + strlen(ih->ih_name) + 1 <
211 		    sizeof(ie->ie_fullname)) {
212 			strcat(ie->ie_fullname, " ");
213 			strcat(ie->ie_fullname, ih->ih_name);
214 			space = 0;
215 		} else
216 			missed++;
217 		if (ih->ih_flags & IH_ENTROPY)
218 			ie->ie_flags |= IE_ENTROPY;
219 	}
220 
221 	/*
222 	 * If the handler names were too long, add +'s to indicate missing
223 	 * names. If we run out of room and still have +'s to add, change
224 	 * the last character from a + to a *.
225 	 */
226 	last = &ie->ie_fullname[sizeof(ie->ie_fullname) - 2];
227 	while (missed-- > 0) {
228 		if (strlen(ie->ie_fullname) + 1 == sizeof(ie->ie_fullname)) {
229 			if (*last == '+') {
230 				*last = '*';
231 				break;
232 			} else
233 				*last = '+';
234 		} else if (space) {
235 			strcat(ie->ie_fullname, " +");
236 			space = 0;
237 		} else
238 			strcat(ie->ie_fullname, "+");
239 	}
240 
241 	/*
242 	 * If this event has an ithread, update it's priority and
243 	 * name.
244 	 */
245 	if (ie->ie_thread != NULL)
246 		ithread_update(ie->ie_thread);
247 	CTR2(KTR_INTR, "%s: updated %s", __func__, ie->ie_fullname);
248 }
249 
250 int
251 intr_event_create(struct intr_event **event, void *source, int flags, int irq,
252     void (*pre_ithread)(void *), void (*post_ithread)(void *),
253     void (*post_filter)(void *), int (*assign_cpu)(void *, u_char),
254     const char *fmt, ...)
255 {
256 	struct intr_event *ie;
257 	va_list ap;
258 
259 	/* The only valid flag during creation is IE_SOFT. */
260 	if ((flags & ~IE_SOFT) != 0)
261 		return (EINVAL);
262 	ie = malloc(sizeof(struct intr_event), M_ITHREAD, M_WAITOK | M_ZERO);
263 	ie->ie_source = source;
264 	ie->ie_pre_ithread = pre_ithread;
265 	ie->ie_post_ithread = post_ithread;
266 	ie->ie_post_filter = post_filter;
267 	ie->ie_assign_cpu = assign_cpu;
268 	ie->ie_flags = flags;
269 	ie->ie_irq = irq;
270 	ie->ie_cpu = NOCPU;
271 	TAILQ_INIT(&ie->ie_handlers);
272 	mtx_init(&ie->ie_lock, "intr event", NULL, MTX_DEF);
273 
274 	va_start(ap, fmt);
275 	vsnprintf(ie->ie_name, sizeof(ie->ie_name), fmt, ap);
276 	va_end(ap);
277 	strlcpy(ie->ie_fullname, ie->ie_name, sizeof(ie->ie_fullname));
278 	mtx_lock(&event_lock);
279 	TAILQ_INSERT_TAIL(&event_list, ie, ie_list);
280 	mtx_unlock(&event_lock);
281 	if (event != NULL)
282 		*event = ie;
283 	CTR2(KTR_INTR, "%s: created %s", __func__, ie->ie_name);
284 	return (0);
285 }
286 
287 /*
288  * Bind an interrupt event to the specified CPU.  Note that not all
289  * platforms support binding an interrupt to a CPU.  For those
290  * platforms this request will fail.  For supported platforms, any
291  * associated ithreads as well as the primary interrupt context will
292  * be bound to the specificed CPU.  Using a cpu id of NOCPU unbinds
293  * the interrupt event.
294  */
295 int
296 intr_event_bind(struct intr_event *ie, u_char cpu)
297 {
298 	cpuset_t mask;
299 	lwpid_t id;
300 	int error;
301 
302 	/* Need a CPU to bind to. */
303 	if (cpu != NOCPU && CPU_ABSENT(cpu))
304 		return (EINVAL);
305 
306 	if (ie->ie_assign_cpu == NULL)
307 		return (EOPNOTSUPP);
308 
309 	error = priv_check(curthread, PRIV_SCHED_CPUSET_INTR);
310 	if (error)
311 		return (error);
312 
313 	/*
314 	 * If we have any ithreads try to set their mask first to verify
315 	 * permissions, etc.
316 	 */
317 	mtx_lock(&ie->ie_lock);
318 	if (ie->ie_thread != NULL) {
319 		CPU_ZERO(&mask);
320 		if (cpu == NOCPU)
321 			CPU_COPY(cpuset_root, &mask);
322 		else
323 			CPU_SET(cpu, &mask);
324 		id = ie->ie_thread->it_thread->td_tid;
325 		mtx_unlock(&ie->ie_lock);
326 		error = cpuset_setthread(id, &mask);
327 		if (error)
328 			return (error);
329 	} else
330 		mtx_unlock(&ie->ie_lock);
331 	error = ie->ie_assign_cpu(ie->ie_source, cpu);
332 	if (error) {
333 		mtx_lock(&ie->ie_lock);
334 		if (ie->ie_thread != NULL) {
335 			CPU_ZERO(&mask);
336 			if (ie->ie_cpu == NOCPU)
337 				CPU_COPY(cpuset_root, &mask);
338 			else
339 				CPU_SET(cpu, &mask);
340 			id = ie->ie_thread->it_thread->td_tid;
341 			mtx_unlock(&ie->ie_lock);
342 			(void)cpuset_setthread(id, &mask);
343 		} else
344 			mtx_unlock(&ie->ie_lock);
345 		return (error);
346 	}
347 
348 	mtx_lock(&ie->ie_lock);
349 	ie->ie_cpu = cpu;
350 	mtx_unlock(&ie->ie_lock);
351 
352 	return (error);
353 }
354 
355 static struct intr_event *
356 intr_lookup(int irq)
357 {
358 	struct intr_event *ie;
359 
360 	mtx_lock(&event_lock);
361 	TAILQ_FOREACH(ie, &event_list, ie_list)
362 		if (ie->ie_irq == irq &&
363 		    (ie->ie_flags & IE_SOFT) == 0 &&
364 		    TAILQ_FIRST(&ie->ie_handlers) != NULL)
365 			break;
366 	mtx_unlock(&event_lock);
367 	return (ie);
368 }
369 
370 int
371 intr_setaffinity(int irq, void *m)
372 {
373 	struct intr_event *ie;
374 	cpuset_t *mask;
375 	u_char cpu;
376 	int n;
377 
378 	mask = m;
379 	cpu = NOCPU;
380 	/*
381 	 * If we're setting all cpus we can unbind.  Otherwise make sure
382 	 * only one cpu is in the set.
383 	 */
384 	if (CPU_CMP(cpuset_root, mask)) {
385 		for (n = 0; n < CPU_SETSIZE; n++) {
386 			if (!CPU_ISSET(n, mask))
387 				continue;
388 			if (cpu != NOCPU)
389 				return (EINVAL);
390 			cpu = (u_char)n;
391 		}
392 	}
393 	ie = intr_lookup(irq);
394 	if (ie == NULL)
395 		return (ESRCH);
396 	return (intr_event_bind(ie, cpu));
397 }
398 
399 int
400 intr_getaffinity(int irq, void *m)
401 {
402 	struct intr_event *ie;
403 	cpuset_t *mask;
404 
405 	mask = m;
406 	ie = intr_lookup(irq);
407 	if (ie == NULL)
408 		return (ESRCH);
409 	CPU_ZERO(mask);
410 	mtx_lock(&ie->ie_lock);
411 	if (ie->ie_cpu == NOCPU)
412 		CPU_COPY(cpuset_root, mask);
413 	else
414 		CPU_SET(ie->ie_cpu, mask);
415 	mtx_unlock(&ie->ie_lock);
416 	return (0);
417 }
418 
419 int
420 intr_event_destroy(struct intr_event *ie)
421 {
422 
423 	mtx_lock(&event_lock);
424 	mtx_lock(&ie->ie_lock);
425 	if (!TAILQ_EMPTY(&ie->ie_handlers)) {
426 		mtx_unlock(&ie->ie_lock);
427 		mtx_unlock(&event_lock);
428 		return (EBUSY);
429 	}
430 	TAILQ_REMOVE(&event_list, ie, ie_list);
431 #ifndef notyet
432 	if (ie->ie_thread != NULL) {
433 		ithread_destroy(ie->ie_thread);
434 		ie->ie_thread = NULL;
435 	}
436 #endif
437 	mtx_unlock(&ie->ie_lock);
438 	mtx_unlock(&event_lock);
439 	mtx_destroy(&ie->ie_lock);
440 	free(ie, M_ITHREAD);
441 	return (0);
442 }
443 
444 #ifndef INTR_FILTER
445 static struct intr_thread *
446 ithread_create(const char *name)
447 {
448 	struct intr_thread *ithd;
449 	struct thread *td;
450 	int error;
451 
452 	ithd = malloc(sizeof(struct intr_thread), M_ITHREAD, M_WAITOK | M_ZERO);
453 
454 	error = kproc_kthread_add(ithread_loop, ithd, &intrproc,
455 		    &td, RFSTOPPED | RFHIGHPID,
456 	    	    0, "intr", "%s", name);
457 	if (error)
458 		panic("kproc_create() failed with %d", error);
459 	thread_lock(td);
460 	sched_class(td, PRI_ITHD);
461 	TD_SET_IWAIT(td);
462 	thread_unlock(td);
463 	td->td_pflags |= TDP_ITHREAD;
464 	ithd->it_thread = td;
465 	CTR2(KTR_INTR, "%s: created %s", __func__, name);
466 	return (ithd);
467 }
468 #else
469 static struct intr_thread *
470 ithread_create(const char *name, struct intr_handler *ih)
471 {
472 	struct intr_thread *ithd;
473 	struct thread *td;
474 	int error;
475 
476 	ithd = malloc(sizeof(struct intr_thread), M_ITHREAD, M_WAITOK | M_ZERO);
477 
478 	error = kproc_kthread_add(ithread_loop, ih, &intrproc,
479 		    &td, RFSTOPPED | RFHIGHPID,
480 	    	    0, "intr", "%s", name);
481 	if (error)
482 		panic("kproc_create() failed with %d", error);
483 	thread_lock(td);
484 	sched_class(td, PRI_ITHD);
485 	TD_SET_IWAIT(td);
486 	thread_unlock(td);
487 	td->td_pflags |= TDP_ITHREAD;
488 	ithd->it_thread = td;
489 	CTR2(KTR_INTR, "%s: created %s", __func__, name);
490 	return (ithd);
491 }
492 #endif
493 
494 static void
495 ithread_destroy(struct intr_thread *ithread)
496 {
497 	struct thread *td;
498 
499 	CTR2(KTR_INTR, "%s: killing %s", __func__, ithread->it_event->ie_name);
500 	td = ithread->it_thread;
501 	thread_lock(td);
502 	ithread->it_flags |= IT_DEAD;
503 	if (TD_AWAITING_INTR(td)) {
504 		TD_CLR_IWAIT(td);
505 		sched_add(td, SRQ_INTR);
506 	}
507 	thread_unlock(td);
508 }
509 
510 #ifndef INTR_FILTER
511 int
512 intr_event_add_handler(struct intr_event *ie, const char *name,
513     driver_filter_t filter, driver_intr_t handler, void *arg, u_char pri,
514     enum intr_type flags, void **cookiep)
515 {
516 	struct intr_handler *ih, *temp_ih;
517 	struct intr_thread *it;
518 
519 	if (ie == NULL || name == NULL || (handler == NULL && filter == NULL))
520 		return (EINVAL);
521 
522 	/* Allocate and populate an interrupt handler structure. */
523 	ih = malloc(sizeof(struct intr_handler), M_ITHREAD, M_WAITOK | M_ZERO);
524 	ih->ih_filter = filter;
525 	ih->ih_handler = handler;
526 	ih->ih_argument = arg;
527 	strlcpy(ih->ih_name, name, sizeof(ih->ih_name));
528 	ih->ih_event = ie;
529 	ih->ih_pri = pri;
530 	if (flags & INTR_EXCL)
531 		ih->ih_flags = IH_EXCLUSIVE;
532 	if (flags & INTR_MPSAFE)
533 		ih->ih_flags |= IH_MPSAFE;
534 	if (flags & INTR_ENTROPY)
535 		ih->ih_flags |= IH_ENTROPY;
536 
537 	/* We can only have one exclusive handler in a event. */
538 	mtx_lock(&ie->ie_lock);
539 	if (!TAILQ_EMPTY(&ie->ie_handlers)) {
540 		if ((flags & INTR_EXCL) ||
541 		    (TAILQ_FIRST(&ie->ie_handlers)->ih_flags & IH_EXCLUSIVE)) {
542 			mtx_unlock(&ie->ie_lock);
543 			free(ih, M_ITHREAD);
544 			return (EINVAL);
545 		}
546 	}
547 
548 	/* Add the new handler to the event in priority order. */
549 	TAILQ_FOREACH(temp_ih, &ie->ie_handlers, ih_next) {
550 		if (temp_ih->ih_pri > ih->ih_pri)
551 			break;
552 	}
553 	if (temp_ih == NULL)
554 		TAILQ_INSERT_TAIL(&ie->ie_handlers, ih, ih_next);
555 	else
556 		TAILQ_INSERT_BEFORE(temp_ih, ih, ih_next);
557 	intr_event_update(ie);
558 
559 	/* Create a thread if we need one. */
560 	while (ie->ie_thread == NULL && handler != NULL) {
561 		if (ie->ie_flags & IE_ADDING_THREAD)
562 			msleep(ie, &ie->ie_lock, 0, "ithread", 0);
563 		else {
564 			ie->ie_flags |= IE_ADDING_THREAD;
565 			mtx_unlock(&ie->ie_lock);
566 			it = ithread_create("intr: newborn");
567 			mtx_lock(&ie->ie_lock);
568 			ie->ie_flags &= ~IE_ADDING_THREAD;
569 			ie->ie_thread = it;
570 			it->it_event = ie;
571 			ithread_update(it);
572 			wakeup(ie);
573 		}
574 	}
575 	CTR3(KTR_INTR, "%s: added %s to %s", __func__, ih->ih_name,
576 	    ie->ie_name);
577 	mtx_unlock(&ie->ie_lock);
578 
579 	if (cookiep != NULL)
580 		*cookiep = ih;
581 	return (0);
582 }
583 #else
584 int
585 intr_event_add_handler(struct intr_event *ie, const char *name,
586     driver_filter_t filter, driver_intr_t handler, void *arg, u_char pri,
587     enum intr_type flags, void **cookiep)
588 {
589 	struct intr_handler *ih, *temp_ih;
590 	struct intr_thread *it;
591 
592 	if (ie == NULL || name == NULL || (handler == NULL && filter == NULL))
593 		return (EINVAL);
594 
595 	/* Allocate and populate an interrupt handler structure. */
596 	ih = malloc(sizeof(struct intr_handler), M_ITHREAD, M_WAITOK | M_ZERO);
597 	ih->ih_filter = filter;
598 	ih->ih_handler = handler;
599 	ih->ih_argument = arg;
600 	strlcpy(ih->ih_name, name, sizeof(ih->ih_name));
601 	ih->ih_event = ie;
602 	ih->ih_pri = pri;
603 	if (flags & INTR_EXCL)
604 		ih->ih_flags = IH_EXCLUSIVE;
605 	if (flags & INTR_MPSAFE)
606 		ih->ih_flags |= IH_MPSAFE;
607 	if (flags & INTR_ENTROPY)
608 		ih->ih_flags |= IH_ENTROPY;
609 
610 	/* We can only have one exclusive handler in a event. */
611 	mtx_lock(&ie->ie_lock);
612 	if (!TAILQ_EMPTY(&ie->ie_handlers)) {
613 		if ((flags & INTR_EXCL) ||
614 		    (TAILQ_FIRST(&ie->ie_handlers)->ih_flags & IH_EXCLUSIVE)) {
615 			mtx_unlock(&ie->ie_lock);
616 			free(ih, M_ITHREAD);
617 			return (EINVAL);
618 		}
619 	}
620 
621 	/* Add the new handler to the event in priority order. */
622 	TAILQ_FOREACH(temp_ih, &ie->ie_handlers, ih_next) {
623 		if (temp_ih->ih_pri > ih->ih_pri)
624 			break;
625 	}
626 	if (temp_ih == NULL)
627 		TAILQ_INSERT_TAIL(&ie->ie_handlers, ih, ih_next);
628 	else
629 		TAILQ_INSERT_BEFORE(temp_ih, ih, ih_next);
630 	intr_event_update(ie);
631 
632 	/* For filtered handlers, create a private ithread to run on. */
633 	if (filter != NULL && handler != NULL) {
634 		mtx_unlock(&ie->ie_lock);
635 		it = ithread_create("intr: newborn", ih);
636 		mtx_lock(&ie->ie_lock);
637 		it->it_event = ie;
638 		ih->ih_thread = it;
639 		ithread_update(it); // XXX - do we really need this?!?!?
640 	} else { /* Create the global per-event thread if we need one. */
641 		while (ie->ie_thread == NULL && handler != NULL) {
642 			if (ie->ie_flags & IE_ADDING_THREAD)
643 				msleep(ie, &ie->ie_lock, 0, "ithread", 0);
644 			else {
645 				ie->ie_flags |= IE_ADDING_THREAD;
646 				mtx_unlock(&ie->ie_lock);
647 				it = ithread_create("intr: newborn", ih);
648 				mtx_lock(&ie->ie_lock);
649 				ie->ie_flags &= ~IE_ADDING_THREAD;
650 				ie->ie_thread = it;
651 				it->it_event = ie;
652 				ithread_update(it);
653 				wakeup(ie);
654 			}
655 		}
656 	}
657 	CTR3(KTR_INTR, "%s: added %s to %s", __func__, ih->ih_name,
658 	    ie->ie_name);
659 	mtx_unlock(&ie->ie_lock);
660 
661 	if (cookiep != NULL)
662 		*cookiep = ih;
663 	return (0);
664 }
665 #endif
666 
667 /*
668  * Append a description preceded by a ':' to the name of the specified
669  * interrupt handler.
670  */
671 int
672 intr_event_describe_handler(struct intr_event *ie, void *cookie,
673     const char *descr)
674 {
675 	struct intr_handler *ih;
676 	size_t space;
677 	char *start;
678 
679 	mtx_lock(&ie->ie_lock);
680 #ifdef INVARIANTS
681 	TAILQ_FOREACH(ih, &ie->ie_handlers, ih_next) {
682 		if (ih == cookie)
683 			break;
684 	}
685 	if (ih == NULL) {
686 		mtx_unlock(&ie->ie_lock);
687 		panic("handler %p not found in interrupt event %p", cookie, ie);
688 	}
689 #endif
690 	ih = cookie;
691 
692 	/*
693 	 * Look for an existing description by checking for an
694 	 * existing ":".  This assumes device names do not include
695 	 * colons.  If one is found, prepare to insert the new
696 	 * description at that point.  If one is not found, find the
697 	 * end of the name to use as the insertion point.
698 	 */
699 	start = index(ih->ih_name, ':');
700 	if (start == NULL)
701 		start = index(ih->ih_name, 0);
702 
703 	/*
704 	 * See if there is enough remaining room in the string for the
705 	 * description + ":".  The "- 1" leaves room for the trailing
706 	 * '\0'.  The "+ 1" accounts for the colon.
707 	 */
708 	space = sizeof(ih->ih_name) - (start - ih->ih_name) - 1;
709 	if (strlen(descr) + 1 > space) {
710 		mtx_unlock(&ie->ie_lock);
711 		return (ENOSPC);
712 	}
713 
714 	/* Append a colon followed by the description. */
715 	*start = ':';
716 	strcpy(start + 1, descr);
717 	intr_event_update(ie);
718 	mtx_unlock(&ie->ie_lock);
719 	return (0);
720 }
721 
722 /*
723  * Return the ie_source field from the intr_event an intr_handler is
724  * associated with.
725  */
726 void *
727 intr_handler_source(void *cookie)
728 {
729 	struct intr_handler *ih;
730 	struct intr_event *ie;
731 
732 	ih = (struct intr_handler *)cookie;
733 	if (ih == NULL)
734 		return (NULL);
735 	ie = ih->ih_event;
736 	KASSERT(ie != NULL,
737 	    ("interrupt handler \"%s\" has a NULL interrupt event",
738 	    ih->ih_name));
739 	return (ie->ie_source);
740 }
741 
742 #ifndef INTR_FILTER
743 int
744 intr_event_remove_handler(void *cookie)
745 {
746 	struct intr_handler *handler = (struct intr_handler *)cookie;
747 	struct intr_event *ie;
748 #ifdef INVARIANTS
749 	struct intr_handler *ih;
750 #endif
751 #ifdef notyet
752 	int dead;
753 #endif
754 
755 	if (handler == NULL)
756 		return (EINVAL);
757 	ie = handler->ih_event;
758 	KASSERT(ie != NULL,
759 	    ("interrupt handler \"%s\" has a NULL interrupt event",
760 	    handler->ih_name));
761 	mtx_lock(&ie->ie_lock);
762 	CTR3(KTR_INTR, "%s: removing %s from %s", __func__, handler->ih_name,
763 	    ie->ie_name);
764 #ifdef INVARIANTS
765 	TAILQ_FOREACH(ih, &ie->ie_handlers, ih_next)
766 		if (ih == handler)
767 			goto ok;
768 	mtx_unlock(&ie->ie_lock);
769 	panic("interrupt handler \"%s\" not found in interrupt event \"%s\"",
770 	    ih->ih_name, ie->ie_name);
771 ok:
772 #endif
773 	/*
774 	 * If there is no ithread, then just remove the handler and return.
775 	 * XXX: Note that an INTR_FAST handler might be running on another
776 	 * CPU!
777 	 */
778 	if (ie->ie_thread == NULL) {
779 		TAILQ_REMOVE(&ie->ie_handlers, handler, ih_next);
780 		mtx_unlock(&ie->ie_lock);
781 		free(handler, M_ITHREAD);
782 		return (0);
783 	}
784 
785 	/*
786 	 * If the interrupt thread is already running, then just mark this
787 	 * handler as being dead and let the ithread do the actual removal.
788 	 *
789 	 * During a cold boot while cold is set, msleep() does not sleep,
790 	 * so we have to remove the handler here rather than letting the
791 	 * thread do it.
792 	 */
793 	thread_lock(ie->ie_thread->it_thread);
794 	if (!TD_AWAITING_INTR(ie->ie_thread->it_thread) && !cold) {
795 		handler->ih_flags |= IH_DEAD;
796 
797 		/*
798 		 * Ensure that the thread will process the handler list
799 		 * again and remove this handler if it has already passed
800 		 * it on the list.
801 		 */
802 		ie->ie_thread->it_need = 1;
803 	} else
804 		TAILQ_REMOVE(&ie->ie_handlers, handler, ih_next);
805 	thread_unlock(ie->ie_thread->it_thread);
806 	while (handler->ih_flags & IH_DEAD)
807 		msleep(handler, &ie->ie_lock, 0, "iev_rmh", 0);
808 	intr_event_update(ie);
809 #ifdef notyet
810 	/*
811 	 * XXX: This could be bad in the case of ppbus(8).  Also, I think
812 	 * this could lead to races of stale data when servicing an
813 	 * interrupt.
814 	 */
815 	dead = 1;
816 	TAILQ_FOREACH(ih, &ie->ie_handlers, ih_next) {
817 		if (!(ih->ih_flags & IH_FAST)) {
818 			dead = 0;
819 			break;
820 		}
821 	}
822 	if (dead) {
823 		ithread_destroy(ie->ie_thread);
824 		ie->ie_thread = NULL;
825 	}
826 #endif
827 	mtx_unlock(&ie->ie_lock);
828 	free(handler, M_ITHREAD);
829 	return (0);
830 }
831 
832 static int
833 intr_event_schedule_thread(struct intr_event *ie)
834 {
835 	struct intr_entropy entropy;
836 	struct intr_thread *it;
837 	struct thread *td;
838 	struct thread *ctd;
839 	struct proc *p;
840 
841 	/*
842 	 * If no ithread or no handlers, then we have a stray interrupt.
843 	 */
844 	if (ie == NULL || TAILQ_EMPTY(&ie->ie_handlers) ||
845 	    ie->ie_thread == NULL)
846 		return (EINVAL);
847 
848 	ctd = curthread;
849 	it = ie->ie_thread;
850 	td = it->it_thread;
851 	p = td->td_proc;
852 
853 	/*
854 	 * If any of the handlers for this ithread claim to be good
855 	 * sources of entropy, then gather some.
856 	 */
857 	if (harvest.interrupt && ie->ie_flags & IE_ENTROPY) {
858 		CTR3(KTR_INTR, "%s: pid %d (%s) gathering entropy", __func__,
859 		    p->p_pid, td->td_name);
860 		entropy.event = (uintptr_t)ie;
861 		entropy.td = ctd;
862 		random_harvest(&entropy, sizeof(entropy), 2, 0,
863 		    RANDOM_INTERRUPT);
864 	}
865 
866 	KASSERT(p != NULL, ("ithread %s has no process", ie->ie_name));
867 
868 	/*
869 	 * Set it_need to tell the thread to keep running if it is already
870 	 * running.  Then, lock the thread and see if we actually need to
871 	 * put it on the runqueue.
872 	 */
873 	it->it_need = 1;
874 	thread_lock(td);
875 	if (TD_AWAITING_INTR(td)) {
876 		CTR3(KTR_INTR, "%s: schedule pid %d (%s)", __func__, p->p_pid,
877 		    td->td_name);
878 		TD_CLR_IWAIT(td);
879 		sched_add(td, SRQ_INTR);
880 	} else {
881 		CTR5(KTR_INTR, "%s: pid %d (%s): it_need %d, state %d",
882 		    __func__, p->p_pid, td->td_name, it->it_need, td->td_state);
883 	}
884 	thread_unlock(td);
885 
886 	return (0);
887 }
888 #else
889 int
890 intr_event_remove_handler(void *cookie)
891 {
892 	struct intr_handler *handler = (struct intr_handler *)cookie;
893 	struct intr_event *ie;
894 	struct intr_thread *it;
895 #ifdef INVARIANTS
896 	struct intr_handler *ih;
897 #endif
898 #ifdef notyet
899 	int dead;
900 #endif
901 
902 	if (handler == NULL)
903 		return (EINVAL);
904 	ie = handler->ih_event;
905 	KASSERT(ie != NULL,
906 	    ("interrupt handler \"%s\" has a NULL interrupt event",
907 	    handler->ih_name));
908 	mtx_lock(&ie->ie_lock);
909 	CTR3(KTR_INTR, "%s: removing %s from %s", __func__, handler->ih_name,
910 	    ie->ie_name);
911 #ifdef INVARIANTS
912 	TAILQ_FOREACH(ih, &ie->ie_handlers, ih_next)
913 		if (ih == handler)
914 			goto ok;
915 	mtx_unlock(&ie->ie_lock);
916 	panic("interrupt handler \"%s\" not found in interrupt event \"%s\"",
917 	    ih->ih_name, ie->ie_name);
918 ok:
919 #endif
920 	/*
921 	 * If there are no ithreads (per event and per handler), then
922 	 * just remove the handler and return.
923 	 * XXX: Note that an INTR_FAST handler might be running on another CPU!
924 	 */
925 	if (ie->ie_thread == NULL && handler->ih_thread == NULL) {
926 		TAILQ_REMOVE(&ie->ie_handlers, handler, ih_next);
927 		mtx_unlock(&ie->ie_lock);
928 		free(handler, M_ITHREAD);
929 		return (0);
930 	}
931 
932 	/* Private or global ithread? */
933 	it = (handler->ih_thread) ? handler->ih_thread : ie->ie_thread;
934 	/*
935 	 * If the interrupt thread is already running, then just mark this
936 	 * handler as being dead and let the ithread do the actual removal.
937 	 *
938 	 * During a cold boot while cold is set, msleep() does not sleep,
939 	 * so we have to remove the handler here rather than letting the
940 	 * thread do it.
941 	 */
942 	thread_lock(it->it_thread);
943 	if (!TD_AWAITING_INTR(it->it_thread) && !cold) {
944 		handler->ih_flags |= IH_DEAD;
945 
946 		/*
947 		 * Ensure that the thread will process the handler list
948 		 * again and remove this handler if it has already passed
949 		 * it on the list.
950 		 */
951 		it->it_need = 1;
952 	} else
953 		TAILQ_REMOVE(&ie->ie_handlers, handler, ih_next);
954 	thread_unlock(it->it_thread);
955 	while (handler->ih_flags & IH_DEAD)
956 		msleep(handler, &ie->ie_lock, 0, "iev_rmh", 0);
957 	/*
958 	 * At this point, the handler has been disconnected from the event,
959 	 * so we can kill the private ithread if any.
960 	 */
961 	if (handler->ih_thread) {
962 		ithread_destroy(handler->ih_thread);
963 		handler->ih_thread = NULL;
964 	}
965 	intr_event_update(ie);
966 #ifdef notyet
967 	/*
968 	 * XXX: This could be bad in the case of ppbus(8).  Also, I think
969 	 * this could lead to races of stale data when servicing an
970 	 * interrupt.
971 	 */
972 	dead = 1;
973 	TAILQ_FOREACH(ih, &ie->ie_handlers, ih_next) {
974 		if (handler != NULL) {
975 			dead = 0;
976 			break;
977 		}
978 	}
979 	if (dead) {
980 		ithread_destroy(ie->ie_thread);
981 		ie->ie_thread = NULL;
982 	}
983 #endif
984 	mtx_unlock(&ie->ie_lock);
985 	free(handler, M_ITHREAD);
986 	return (0);
987 }
988 
989 static int
990 intr_event_schedule_thread(struct intr_event *ie, struct intr_thread *it)
991 {
992 	struct intr_entropy entropy;
993 	struct thread *td;
994 	struct thread *ctd;
995 	struct proc *p;
996 
997 	/*
998 	 * If no ithread or no handlers, then we have a stray interrupt.
999 	 */
1000 	if (ie == NULL || TAILQ_EMPTY(&ie->ie_handlers) || it == NULL)
1001 		return (EINVAL);
1002 
1003 	ctd = curthread;
1004 	td = it->it_thread;
1005 	p = td->td_proc;
1006 
1007 	/*
1008 	 * If any of the handlers for this ithread claim to be good
1009 	 * sources of entropy, then gather some.
1010 	 */
1011 	if (harvest.interrupt && ie->ie_flags & IE_ENTROPY) {
1012 		CTR3(KTR_INTR, "%s: pid %d (%s) gathering entropy", __func__,
1013 		    p->p_pid, td->td_name);
1014 		entropy.event = (uintptr_t)ie;
1015 		entropy.td = ctd;
1016 		random_harvest(&entropy, sizeof(entropy), 2, 0,
1017 		    RANDOM_INTERRUPT);
1018 	}
1019 
1020 	KASSERT(p != NULL, ("ithread %s has no process", ie->ie_name));
1021 
1022 	/*
1023 	 * Set it_need to tell the thread to keep running if it is already
1024 	 * running.  Then, lock the thread and see if we actually need to
1025 	 * put it on the runqueue.
1026 	 */
1027 	it->it_need = 1;
1028 	thread_lock(td);
1029 	if (TD_AWAITING_INTR(td)) {
1030 		CTR3(KTR_INTR, "%s: schedule pid %d (%s)", __func__, p->p_pid,
1031 		    td->td_name);
1032 		TD_CLR_IWAIT(td);
1033 		sched_add(td, SRQ_INTR);
1034 	} else {
1035 		CTR5(KTR_INTR, "%s: pid %d (%s): it_need %d, state %d",
1036 		    __func__, p->p_pid, td->td_name, it->it_need, td->td_state);
1037 	}
1038 	thread_unlock(td);
1039 
1040 	return (0);
1041 }
1042 #endif
1043 
1044 /*
1045  * Allow interrupt event binding for software interrupt handlers -- a no-op,
1046  * since interrupts are generated in software rather than being directed by
1047  * a PIC.
1048  */
1049 static int
1050 swi_assign_cpu(void *arg, u_char cpu)
1051 {
1052 
1053 	return (0);
1054 }
1055 
1056 /*
1057  * Add a software interrupt handler to a specified event.  If a given event
1058  * is not specified, then a new event is created.
1059  */
1060 int
1061 swi_add(struct intr_event **eventp, const char *name, driver_intr_t handler,
1062 	    void *arg, int pri, enum intr_type flags, void **cookiep)
1063 {
1064 	struct thread *td;
1065 	struct intr_event *ie;
1066 	int error;
1067 
1068 	if (flags & INTR_ENTROPY)
1069 		return (EINVAL);
1070 
1071 	ie = (eventp != NULL) ? *eventp : NULL;
1072 
1073 	if (ie != NULL) {
1074 		if (!(ie->ie_flags & IE_SOFT))
1075 			return (EINVAL);
1076 	} else {
1077 		error = intr_event_create(&ie, NULL, IE_SOFT, 0,
1078 		    NULL, NULL, NULL, swi_assign_cpu, "swi%d:", pri);
1079 		if (error)
1080 			return (error);
1081 		if (eventp != NULL)
1082 			*eventp = ie;
1083 	}
1084 	error = intr_event_add_handler(ie, name, NULL, handler, arg,
1085 	    (pri * RQ_PPQ) + PI_SOFT, flags, cookiep);
1086 	if (error)
1087 		return (error);
1088 	if (pri == SWI_CLOCK) {
1089 		td = ie->ie_thread->it_thread;
1090 		thread_lock(td);
1091 		td->td_flags |= TDF_NOLOAD;
1092 		thread_unlock(td);
1093 	}
1094 	return (0);
1095 }
1096 
1097 /*
1098  * Schedule a software interrupt thread.
1099  */
1100 void
1101 swi_sched(void *cookie, int flags)
1102 {
1103 	struct intr_handler *ih = (struct intr_handler *)cookie;
1104 	struct intr_event *ie = ih->ih_event;
1105 	int error;
1106 
1107 	CTR3(KTR_INTR, "swi_sched: %s %s need=%d", ie->ie_name, ih->ih_name,
1108 	    ih->ih_need);
1109 
1110 	/*
1111 	 * Set ih_need for this handler so that if the ithread is already
1112 	 * running it will execute this handler on the next pass.  Otherwise,
1113 	 * it will execute it the next time it runs.
1114 	 */
1115 	atomic_store_rel_int(&ih->ih_need, 1);
1116 
1117 	if (!(flags & SWI_DELAY)) {
1118 		PCPU_INC(cnt.v_soft);
1119 #ifdef INTR_FILTER
1120 		error = intr_event_schedule_thread(ie, ie->ie_thread);
1121 #else
1122 		error = intr_event_schedule_thread(ie);
1123 #endif
1124 		KASSERT(error == 0, ("stray software interrupt"));
1125 	}
1126 }
1127 
1128 /*
1129  * Remove a software interrupt handler.  Currently this code does not
1130  * remove the associated interrupt event if it becomes empty.  Calling code
1131  * may do so manually via intr_event_destroy(), but that's not really
1132  * an optimal interface.
1133  */
1134 int
1135 swi_remove(void *cookie)
1136 {
1137 
1138 	return (intr_event_remove_handler(cookie));
1139 }
1140 
1141 #ifdef INTR_FILTER
1142 static void
1143 priv_ithread_execute_handler(struct proc *p, struct intr_handler *ih)
1144 {
1145 	struct intr_event *ie;
1146 
1147 	ie = ih->ih_event;
1148 	/*
1149 	 * If this handler is marked for death, remove it from
1150 	 * the list of handlers and wake up the sleeper.
1151 	 */
1152 	if (ih->ih_flags & IH_DEAD) {
1153 		mtx_lock(&ie->ie_lock);
1154 		TAILQ_REMOVE(&ie->ie_handlers, ih, ih_next);
1155 		ih->ih_flags &= ~IH_DEAD;
1156 		wakeup(ih);
1157 		mtx_unlock(&ie->ie_lock);
1158 		return;
1159 	}
1160 
1161 	/* Execute this handler. */
1162 	CTR6(KTR_INTR, "%s: pid %d exec %p(%p) for %s flg=%x",
1163 	     __func__, p->p_pid, (void *)ih->ih_handler, ih->ih_argument,
1164 	     ih->ih_name, ih->ih_flags);
1165 
1166 	if (!(ih->ih_flags & IH_MPSAFE))
1167 		mtx_lock(&Giant);
1168 	ih->ih_handler(ih->ih_argument);
1169 	if (!(ih->ih_flags & IH_MPSAFE))
1170 		mtx_unlock(&Giant);
1171 }
1172 #endif
1173 
1174 /*
1175  * This is a public function for use by drivers that mux interrupt
1176  * handlers for child devices from their interrupt handler.
1177  */
1178 void
1179 intr_event_execute_handlers(struct proc *p, struct intr_event *ie)
1180 {
1181 	struct intr_handler *ih, *ihn;
1182 
1183 	TAILQ_FOREACH_SAFE(ih, &ie->ie_handlers, ih_next, ihn) {
1184 		/*
1185 		 * If this handler is marked for death, remove it from
1186 		 * the list of handlers and wake up the sleeper.
1187 		 */
1188 		if (ih->ih_flags & IH_DEAD) {
1189 			mtx_lock(&ie->ie_lock);
1190 			TAILQ_REMOVE(&ie->ie_handlers, ih, ih_next);
1191 			ih->ih_flags &= ~IH_DEAD;
1192 			wakeup(ih);
1193 			mtx_unlock(&ie->ie_lock);
1194 			continue;
1195 		}
1196 
1197 		/* Skip filter only handlers */
1198 		if (ih->ih_handler == NULL)
1199 			continue;
1200 
1201 		/*
1202 		 * For software interrupt threads, we only execute
1203 		 * handlers that have their need flag set.  Hardware
1204 		 * interrupt threads always invoke all of their handlers.
1205 		 */
1206 		if (ie->ie_flags & IE_SOFT) {
1207 			if (!ih->ih_need)
1208 				continue;
1209 			else
1210 				atomic_store_rel_int(&ih->ih_need, 0);
1211 		}
1212 
1213 		/* Execute this handler. */
1214 		CTR6(KTR_INTR, "%s: pid %d exec %p(%p) for %s flg=%x",
1215 		    __func__, p->p_pid, (void *)ih->ih_handler,
1216 		    ih->ih_argument, ih->ih_name, ih->ih_flags);
1217 
1218 		if (!(ih->ih_flags & IH_MPSAFE))
1219 			mtx_lock(&Giant);
1220 		ih->ih_handler(ih->ih_argument);
1221 		if (!(ih->ih_flags & IH_MPSAFE))
1222 			mtx_unlock(&Giant);
1223 	}
1224 }
1225 
1226 static void
1227 ithread_execute_handlers(struct proc *p, struct intr_event *ie)
1228 {
1229 
1230 	/* Interrupt handlers should not sleep. */
1231 	if (!(ie->ie_flags & IE_SOFT))
1232 		THREAD_NO_SLEEPING();
1233 	intr_event_execute_handlers(p, ie);
1234 	if (!(ie->ie_flags & IE_SOFT))
1235 		THREAD_SLEEPING_OK();
1236 
1237 	/*
1238 	 * Interrupt storm handling:
1239 	 *
1240 	 * If this interrupt source is currently storming, then throttle
1241 	 * it to only fire the handler once  per clock tick.
1242 	 *
1243 	 * If this interrupt source is not currently storming, but the
1244 	 * number of back to back interrupts exceeds the storm threshold,
1245 	 * then enter storming mode.
1246 	 */
1247 	if (intr_storm_threshold != 0 && ie->ie_count >= intr_storm_threshold &&
1248 	    !(ie->ie_flags & IE_SOFT)) {
1249 		/* Report the message only once every second. */
1250 		if (ppsratecheck(&ie->ie_warntm, &ie->ie_warncnt, 1)) {
1251 			printf(
1252 	"interrupt storm detected on \"%s\"; throttling interrupt source\n",
1253 			    ie->ie_name);
1254 		}
1255 		pause("istorm", 1);
1256 	} else
1257 		ie->ie_count++;
1258 
1259 	/*
1260 	 * Now that all the handlers have had a chance to run, reenable
1261 	 * the interrupt source.
1262 	 */
1263 	if (ie->ie_post_ithread != NULL)
1264 		ie->ie_post_ithread(ie->ie_source);
1265 }
1266 
1267 #ifndef INTR_FILTER
1268 /*
1269  * This is the main code for interrupt threads.
1270  */
1271 static void
1272 ithread_loop(void *arg)
1273 {
1274 	struct intr_thread *ithd;
1275 	struct intr_event *ie;
1276 	struct thread *td;
1277 	struct proc *p;
1278 
1279 	td = curthread;
1280 	p = td->td_proc;
1281 	ithd = (struct intr_thread *)arg;
1282 	KASSERT(ithd->it_thread == td,
1283 	    ("%s: ithread and proc linkage out of sync", __func__));
1284 	ie = ithd->it_event;
1285 	ie->ie_count = 0;
1286 
1287 	/*
1288 	 * As long as we have interrupts outstanding, go through the
1289 	 * list of handlers, giving each one a go at it.
1290 	 */
1291 	for (;;) {
1292 		/*
1293 		 * If we are an orphaned thread, then just die.
1294 		 */
1295 		if (ithd->it_flags & IT_DEAD) {
1296 			CTR3(KTR_INTR, "%s: pid %d (%s) exiting", __func__,
1297 			    p->p_pid, td->td_name);
1298 			free(ithd, M_ITHREAD);
1299 			kthread_exit();
1300 		}
1301 
1302 		/*
1303 		 * Service interrupts.  If another interrupt arrives while
1304 		 * we are running, it will set it_need to note that we
1305 		 * should make another pass.
1306 		 */
1307 		while (ithd->it_need) {
1308 			/*
1309 			 * This might need a full read and write barrier
1310 			 * to make sure that this write posts before any
1311 			 * of the memory or device accesses in the
1312 			 * handlers.
1313 			 */
1314 			atomic_store_rel_int(&ithd->it_need, 0);
1315 			ithread_execute_handlers(p, ie);
1316 		}
1317 		WITNESS_WARN(WARN_PANIC, NULL, "suspending ithread");
1318 		mtx_assert(&Giant, MA_NOTOWNED);
1319 
1320 		/*
1321 		 * Processed all our interrupts.  Now get the sched
1322 		 * lock.  This may take a while and it_need may get
1323 		 * set again, so we have to check it again.
1324 		 */
1325 		thread_lock(td);
1326 		if (!ithd->it_need && !(ithd->it_flags & IT_DEAD)) {
1327 			TD_SET_IWAIT(td);
1328 			ie->ie_count = 0;
1329 			mi_switch(SW_VOL | SWT_IWAIT, NULL);
1330 		}
1331 		thread_unlock(td);
1332 	}
1333 }
1334 
1335 /*
1336  * Main interrupt handling body.
1337  *
1338  * Input:
1339  * o ie:                        the event connected to this interrupt.
1340  * o frame:                     some archs (i.e. i386) pass a frame to some.
1341  *                              handlers as their main argument.
1342  * Return value:
1343  * o 0:                         everything ok.
1344  * o EINVAL:                    stray interrupt.
1345  */
1346 int
1347 intr_event_handle(struct intr_event *ie, struct trapframe *frame)
1348 {
1349 	struct intr_handler *ih;
1350 	struct thread *td;
1351 	int error, ret, thread;
1352 
1353 	td = curthread;
1354 
1355 	/* An interrupt with no event or handlers is a stray interrupt. */
1356 	if (ie == NULL || TAILQ_EMPTY(&ie->ie_handlers))
1357 		return (EINVAL);
1358 
1359 	/*
1360 	 * Execute fast interrupt handlers directly.
1361 	 * To support clock handlers, if a handler registers
1362 	 * with a NULL argument, then we pass it a pointer to
1363 	 * a trapframe as its argument.
1364 	 */
1365 	td->td_intr_nesting_level++;
1366 	thread = 0;
1367 	ret = 0;
1368 	critical_enter();
1369 	TAILQ_FOREACH(ih, &ie->ie_handlers, ih_next) {
1370 		if (ih->ih_filter == NULL) {
1371 			thread = 1;
1372 			continue;
1373 		}
1374 		CTR4(KTR_INTR, "%s: exec %p(%p) for %s", __func__,
1375 		    ih->ih_filter, ih->ih_argument == NULL ? frame :
1376 		    ih->ih_argument, ih->ih_name);
1377 		if (ih->ih_argument == NULL)
1378 			ret = ih->ih_filter(frame);
1379 		else
1380 			ret = ih->ih_filter(ih->ih_argument);
1381 		KASSERT(ret == FILTER_STRAY ||
1382 		    ((ret & (FILTER_SCHEDULE_THREAD | FILTER_HANDLED)) != 0 &&
1383 		    (ret & ~(FILTER_SCHEDULE_THREAD | FILTER_HANDLED)) == 0),
1384 		    ("%s: incorrect return value %#x from %s", __func__, ret,
1385 		    ih->ih_name));
1386 
1387 		/*
1388 		 * Wrapper handler special handling:
1389 		 *
1390 		 * in some particular cases (like pccard and pccbb),
1391 		 * the _real_ device handler is wrapped in a couple of
1392 		 * functions - a filter wrapper and an ithread wrapper.
1393 		 * In this case (and just in this case), the filter wrapper
1394 		 * could ask the system to schedule the ithread and mask
1395 		 * the interrupt source if the wrapped handler is composed
1396 		 * of just an ithread handler.
1397 		 *
1398 		 * TODO: write a generic wrapper to avoid people rolling
1399 		 * their own
1400 		 */
1401 		if (!thread) {
1402 			if (ret == FILTER_SCHEDULE_THREAD)
1403 				thread = 1;
1404 		}
1405 	}
1406 
1407 	if (thread) {
1408 		if (ie->ie_pre_ithread != NULL)
1409 			ie->ie_pre_ithread(ie->ie_source);
1410 	} else {
1411 		if (ie->ie_post_filter != NULL)
1412 			ie->ie_post_filter(ie->ie_source);
1413 	}
1414 
1415 	/* Schedule the ithread if needed. */
1416 	if (thread) {
1417 		error = intr_event_schedule_thread(ie);
1418 #ifndef XEN
1419 		KASSERT(error == 0, ("bad stray interrupt"));
1420 #else
1421 		if (error != 0)
1422 			log(LOG_WARNING, "bad stray interrupt");
1423 #endif
1424 	}
1425 	critical_exit();
1426 	td->td_intr_nesting_level--;
1427 	return (0);
1428 }
1429 #else
1430 /*
1431  * This is the main code for interrupt threads.
1432  */
1433 static void
1434 ithread_loop(void *arg)
1435 {
1436 	struct intr_thread *ithd;
1437 	struct intr_handler *ih;
1438 	struct intr_event *ie;
1439 	struct thread *td;
1440 	struct proc *p;
1441 	int priv;
1442 
1443 	td = curthread;
1444 	p = td->td_proc;
1445 	ih = (struct intr_handler *)arg;
1446 	priv = (ih->ih_thread != NULL) ? 1 : 0;
1447 	ithd = (priv) ? ih->ih_thread : ih->ih_event->ie_thread;
1448 	KASSERT(ithd->it_thread == td,
1449 	    ("%s: ithread and proc linkage out of sync", __func__));
1450 	ie = ithd->it_event;
1451 	ie->ie_count = 0;
1452 
1453 	/*
1454 	 * As long as we have interrupts outstanding, go through the
1455 	 * list of handlers, giving each one a go at it.
1456 	 */
1457 	for (;;) {
1458 		/*
1459 		 * If we are an orphaned thread, then just die.
1460 		 */
1461 		if (ithd->it_flags & IT_DEAD) {
1462 			CTR3(KTR_INTR, "%s: pid %d (%s) exiting", __func__,
1463 			    p->p_pid, td->td_name);
1464 			free(ithd, M_ITHREAD);
1465 			kthread_exit();
1466 		}
1467 
1468 		/*
1469 		 * Service interrupts.  If another interrupt arrives while
1470 		 * we are running, it will set it_need to note that we
1471 		 * should make another pass.
1472 		 */
1473 		while (ithd->it_need) {
1474 			/*
1475 			 * This might need a full read and write barrier
1476 			 * to make sure that this write posts before any
1477 			 * of the memory or device accesses in the
1478 			 * handlers.
1479 			 */
1480 			atomic_store_rel_int(&ithd->it_need, 0);
1481 			if (priv)
1482 				priv_ithread_execute_handler(p, ih);
1483 			else
1484 				ithread_execute_handlers(p, ie);
1485 		}
1486 		WITNESS_WARN(WARN_PANIC, NULL, "suspending ithread");
1487 		mtx_assert(&Giant, MA_NOTOWNED);
1488 
1489 		/*
1490 		 * Processed all our interrupts.  Now get the sched
1491 		 * lock.  This may take a while and it_need may get
1492 		 * set again, so we have to check it again.
1493 		 */
1494 		thread_lock(td);
1495 		if (!ithd->it_need && !(ithd->it_flags & IT_DEAD)) {
1496 			TD_SET_IWAIT(td);
1497 			ie->ie_count = 0;
1498 			mi_switch(SW_VOL | SWT_IWAIT, NULL);
1499 		}
1500 		thread_unlock(td);
1501 	}
1502 }
1503 
1504 /*
1505  * Main loop for interrupt filter.
1506  *
1507  * Some architectures (i386, amd64 and arm) require the optional frame
1508  * parameter, and use it as the main argument for fast handler execution
1509  * when ih_argument == NULL.
1510  *
1511  * Return value:
1512  * o FILTER_STRAY:              No filter recognized the event, and no
1513  *                              filter-less handler is registered on this
1514  *                              line.
1515  * o FILTER_HANDLED:            A filter claimed the event and served it.
1516  * o FILTER_SCHEDULE_THREAD:    No filter claimed the event, but there's at
1517  *                              least one filter-less handler on this line.
1518  * o FILTER_HANDLED |
1519  *   FILTER_SCHEDULE_THREAD:    A filter claimed the event, and asked for
1520  *                              scheduling the per-handler ithread.
1521  *
1522  * In case an ithread has to be scheduled, in *ithd there will be a
1523  * pointer to a struct intr_thread containing the thread to be
1524  * scheduled.
1525  */
1526 
1527 static int
1528 intr_filter_loop(struct intr_event *ie, struct trapframe *frame,
1529 		 struct intr_thread **ithd)
1530 {
1531 	struct intr_handler *ih;
1532 	void *arg;
1533 	int ret, thread_only;
1534 
1535 	ret = 0;
1536 	thread_only = 0;
1537 	TAILQ_FOREACH(ih, &ie->ie_handlers, ih_next) {
1538 		/*
1539 		 * Execute fast interrupt handlers directly.
1540 		 * To support clock handlers, if a handler registers
1541 		 * with a NULL argument, then we pass it a pointer to
1542 		 * a trapframe as its argument.
1543 		 */
1544 		arg = ((ih->ih_argument == NULL) ? frame : ih->ih_argument);
1545 
1546 		CTR5(KTR_INTR, "%s: exec %p/%p(%p) for %s", __func__,
1547 		     ih->ih_filter, ih->ih_handler, arg, ih->ih_name);
1548 
1549 		if (ih->ih_filter != NULL)
1550 			ret = ih->ih_filter(arg);
1551 		else {
1552 			thread_only = 1;
1553 			continue;
1554 		}
1555 		KASSERT(ret == FILTER_STRAY ||
1556 		    ((ret & (FILTER_SCHEDULE_THREAD | FILTER_HANDLED)) != 0 &&
1557 		    (ret & ~(FILTER_SCHEDULE_THREAD | FILTER_HANDLED)) == 0),
1558 		    ("%s: incorrect return value %#x from %s", __func__, ret,
1559 		    ih->ih_name));
1560 		if (ret & FILTER_STRAY)
1561 			continue;
1562 		else {
1563 			*ithd = ih->ih_thread;
1564 			return (ret);
1565 		}
1566 	}
1567 
1568 	/*
1569 	 * No filters handled the interrupt and we have at least
1570 	 * one handler without a filter.  In this case, we schedule
1571 	 * all of the filter-less handlers to run in the ithread.
1572 	 */
1573 	if (thread_only) {
1574 		*ithd = ie->ie_thread;
1575 		return (FILTER_SCHEDULE_THREAD);
1576 	}
1577 	return (FILTER_STRAY);
1578 }
1579 
1580 /*
1581  * Main interrupt handling body.
1582  *
1583  * Input:
1584  * o ie:                        the event connected to this interrupt.
1585  * o frame:                     some archs (i.e. i386) pass a frame to some.
1586  *                              handlers as their main argument.
1587  * Return value:
1588  * o 0:                         everything ok.
1589  * o EINVAL:                    stray interrupt.
1590  */
1591 int
1592 intr_event_handle(struct intr_event *ie, struct trapframe *frame)
1593 {
1594 	struct intr_thread *ithd;
1595 	struct thread *td;
1596 	int thread;
1597 
1598 	ithd = NULL;
1599 	td = curthread;
1600 
1601 	if (ie == NULL || TAILQ_EMPTY(&ie->ie_handlers))
1602 		return (EINVAL);
1603 
1604 	td->td_intr_nesting_level++;
1605 	thread = 0;
1606 	critical_enter();
1607 	thread = intr_filter_loop(ie, frame, &ithd);
1608 	if (thread & FILTER_HANDLED) {
1609 		if (ie->ie_post_filter != NULL)
1610 			ie->ie_post_filter(ie->ie_source);
1611 	} else {
1612 		if (ie->ie_pre_ithread != NULL)
1613 			ie->ie_pre_ithread(ie->ie_source);
1614 	}
1615 	critical_exit();
1616 
1617 	/* Interrupt storm logic */
1618 	if (thread & FILTER_STRAY) {
1619 		ie->ie_count++;
1620 		if (ie->ie_count < intr_storm_threshold)
1621 			printf("Interrupt stray detection not present\n");
1622 	}
1623 
1624 	/* Schedule an ithread if needed. */
1625 	if (thread & FILTER_SCHEDULE_THREAD) {
1626 		if (intr_event_schedule_thread(ie, ithd) != 0)
1627 			panic("%s: impossible stray interrupt", __func__);
1628 	}
1629 	td->td_intr_nesting_level--;
1630 	return (0);
1631 }
1632 #endif
1633 
1634 #ifdef DDB
1635 /*
1636  * Dump details about an interrupt handler
1637  */
1638 static void
1639 db_dump_intrhand(struct intr_handler *ih)
1640 {
1641 	int comma;
1642 
1643 	db_printf("\t%-10s ", ih->ih_name);
1644 	switch (ih->ih_pri) {
1645 	case PI_REALTIME:
1646 		db_printf("CLK ");
1647 		break;
1648 	case PI_AV:
1649 		db_printf("AV  ");
1650 		break;
1651 	case PI_TTYHIGH:
1652 	case PI_TTYLOW:
1653 		db_printf("TTY ");
1654 		break;
1655 	case PI_TAPE:
1656 		db_printf("TAPE");
1657 		break;
1658 	case PI_NET:
1659 		db_printf("NET ");
1660 		break;
1661 	case PI_DISK:
1662 	case PI_DISKLOW:
1663 		db_printf("DISK");
1664 		break;
1665 	case PI_DULL:
1666 		db_printf("DULL");
1667 		break;
1668 	default:
1669 		if (ih->ih_pri >= PI_SOFT)
1670 			db_printf("SWI ");
1671 		else
1672 			db_printf("%4u", ih->ih_pri);
1673 		break;
1674 	}
1675 	db_printf(" ");
1676 	db_printsym((uintptr_t)ih->ih_handler, DB_STGY_PROC);
1677 	db_printf("(%p)", ih->ih_argument);
1678 	if (ih->ih_need ||
1679 	    (ih->ih_flags & (IH_EXCLUSIVE | IH_ENTROPY | IH_DEAD |
1680 	    IH_MPSAFE)) != 0) {
1681 		db_printf(" {");
1682 		comma = 0;
1683 		if (ih->ih_flags & IH_EXCLUSIVE) {
1684 			if (comma)
1685 				db_printf(", ");
1686 			db_printf("EXCL");
1687 			comma = 1;
1688 		}
1689 		if (ih->ih_flags & IH_ENTROPY) {
1690 			if (comma)
1691 				db_printf(", ");
1692 			db_printf("ENTROPY");
1693 			comma = 1;
1694 		}
1695 		if (ih->ih_flags & IH_DEAD) {
1696 			if (comma)
1697 				db_printf(", ");
1698 			db_printf("DEAD");
1699 			comma = 1;
1700 		}
1701 		if (ih->ih_flags & IH_MPSAFE) {
1702 			if (comma)
1703 				db_printf(", ");
1704 			db_printf("MPSAFE");
1705 			comma = 1;
1706 		}
1707 		if (ih->ih_need) {
1708 			if (comma)
1709 				db_printf(", ");
1710 			db_printf("NEED");
1711 		}
1712 		db_printf("}");
1713 	}
1714 	db_printf("\n");
1715 }
1716 
1717 /*
1718  * Dump details about a event.
1719  */
1720 void
1721 db_dump_intr_event(struct intr_event *ie, int handlers)
1722 {
1723 	struct intr_handler *ih;
1724 	struct intr_thread *it;
1725 	int comma;
1726 
1727 	db_printf("%s ", ie->ie_fullname);
1728 	it = ie->ie_thread;
1729 	if (it != NULL)
1730 		db_printf("(pid %d)", it->it_thread->td_proc->p_pid);
1731 	else
1732 		db_printf("(no thread)");
1733 	if ((ie->ie_flags & (IE_SOFT | IE_ENTROPY | IE_ADDING_THREAD)) != 0 ||
1734 	    (it != NULL && it->it_need)) {
1735 		db_printf(" {");
1736 		comma = 0;
1737 		if (ie->ie_flags & IE_SOFT) {
1738 			db_printf("SOFT");
1739 			comma = 1;
1740 		}
1741 		if (ie->ie_flags & IE_ENTROPY) {
1742 			if (comma)
1743 				db_printf(", ");
1744 			db_printf("ENTROPY");
1745 			comma = 1;
1746 		}
1747 		if (ie->ie_flags & IE_ADDING_THREAD) {
1748 			if (comma)
1749 				db_printf(", ");
1750 			db_printf("ADDING_THREAD");
1751 			comma = 1;
1752 		}
1753 		if (it != NULL && it->it_need) {
1754 			if (comma)
1755 				db_printf(", ");
1756 			db_printf("NEED");
1757 		}
1758 		db_printf("}");
1759 	}
1760 	db_printf("\n");
1761 
1762 	if (handlers)
1763 		TAILQ_FOREACH(ih, &ie->ie_handlers, ih_next)
1764 		    db_dump_intrhand(ih);
1765 }
1766 
1767 /*
1768  * Dump data about interrupt handlers
1769  */
1770 DB_SHOW_COMMAND(intr, db_show_intr)
1771 {
1772 	struct intr_event *ie;
1773 	int all, verbose;
1774 
1775 	verbose = index(modif, 'v') != NULL;
1776 	all = index(modif, 'a') != NULL;
1777 	TAILQ_FOREACH(ie, &event_list, ie_list) {
1778 		if (!all && TAILQ_EMPTY(&ie->ie_handlers))
1779 			continue;
1780 		db_dump_intr_event(ie, verbose);
1781 		if (db_pager_quit)
1782 			break;
1783 	}
1784 }
1785 #endif /* DDB */
1786 
1787 /*
1788  * Start standard software interrupt threads
1789  */
1790 static void
1791 start_softintr(void *dummy)
1792 {
1793 
1794 	if (swi_add(NULL, "vm", swi_vm, NULL, SWI_VM, INTR_MPSAFE, &vm_ih))
1795 		panic("died while creating vm swi ithread");
1796 }
1797 SYSINIT(start_softintr, SI_SUB_SOFTINTR, SI_ORDER_FIRST, start_softintr,
1798     NULL);
1799 
1800 /*
1801  * Sysctls used by systat and others: hw.intrnames and hw.intrcnt.
1802  * The data for this machine dependent, and the declarations are in machine
1803  * dependent code.  The layout of intrnames and intrcnt however is machine
1804  * independent.
1805  *
1806  * We do not know the length of intrcnt and intrnames at compile time, so
1807  * calculate things at run time.
1808  */
1809 static int
1810 sysctl_intrnames(SYSCTL_HANDLER_ARGS)
1811 {
1812 	return (sysctl_handle_opaque(oidp, intrnames, eintrnames - intrnames,
1813 	   req));
1814 }
1815 
1816 SYSCTL_PROC(_hw, OID_AUTO, intrnames, CTLTYPE_OPAQUE | CTLFLAG_RD,
1817     NULL, 0, sysctl_intrnames, "", "Interrupt Names");
1818 
1819 static int
1820 sysctl_intrcnt(SYSCTL_HANDLER_ARGS)
1821 {
1822 	return (sysctl_handle_opaque(oidp, intrcnt,
1823 	    (char *)eintrcnt - (char *)intrcnt, req));
1824 }
1825 
1826 SYSCTL_PROC(_hw, OID_AUTO, intrcnt, CTLTYPE_OPAQUE | CTLFLAG_RD,
1827     NULL, 0, sysctl_intrcnt, "", "Interrupt Counts");
1828 
1829 #ifdef DDB
1830 /*
1831  * DDB command to dump the interrupt statistics.
1832  */
1833 DB_SHOW_COMMAND(intrcnt, db_show_intrcnt)
1834 {
1835 	u_long *i;
1836 	char *cp;
1837 
1838 	cp = intrnames;
1839 	for (i = intrcnt; i != eintrcnt && !db_pager_quit; i++) {
1840 		if (*cp == '\0')
1841 			break;
1842 		if (*i != 0)
1843 			db_printf("%s\t%lu\n", cp, *i);
1844 		cp += strlen(cp) + 1;
1845 	}
1846 }
1847 #endif
1848