xref: /freebsd/sys/kern/kern_intr.c (revision 56e766af41cd68310f5583bb893b13c006fcb44f)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 1997, Stefan Esser <se@freebsd.org>
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice unmodified, this list of conditions, and the following
12  *    disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
18  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
19  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
20  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
21  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
22  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
23  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
24  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
25  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
26  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
27  */
28 
29 #include <sys/cdefs.h>
30 __FBSDID("$FreeBSD$");
31 
32 #include "opt_ddb.h"
33 #include "opt_kstack_usage_prof.h"
34 
35 #include <sys/param.h>
36 #include <sys/bus.h>
37 #include <sys/conf.h>
38 #include <sys/cpuset.h>
39 #include <sys/rtprio.h>
40 #include <sys/systm.h>
41 #include <sys/interrupt.h>
42 #include <sys/kernel.h>
43 #include <sys/kthread.h>
44 #include <sys/ktr.h>
45 #include <sys/limits.h>
46 #include <sys/lock.h>
47 #include <sys/malloc.h>
48 #include <sys/mutex.h>
49 #include <sys/priv.h>
50 #include <sys/proc.h>
51 #include <sys/epoch.h>
52 #include <sys/random.h>
53 #include <sys/resourcevar.h>
54 #include <sys/sched.h>
55 #include <sys/smp.h>
56 #include <sys/sysctl.h>
57 #include <sys/syslog.h>
58 #include <sys/unistd.h>
59 #include <sys/vmmeter.h>
60 #include <machine/atomic.h>
61 #include <machine/cpu.h>
62 #include <machine/md_var.h>
63 #include <machine/stdarg.h>
64 #ifdef DDB
65 #include <ddb/ddb.h>
66 #include <ddb/db_sym.h>
67 #endif
68 
69 /*
70  * Describe an interrupt thread.  There is one of these per interrupt event.
71  */
72 struct intr_thread {
73 	struct intr_event *it_event;
74 	struct thread *it_thread;	/* Kernel thread. */
75 	int	it_flags;		/* (j) IT_* flags. */
76 	int	it_need;		/* Needs service. */
77 };
78 
79 /* Interrupt thread flags kept in it_flags */
80 #define	IT_DEAD		0x000001	/* Thread is waiting to exit. */
81 #define	IT_WAIT		0x000002	/* Thread is waiting for completion. */
82 
83 struct	intr_entropy {
84 	struct	thread *td;
85 	uintptr_t event;
86 };
87 
88 struct	intr_event *tty_intr_event;
89 void	*vm_ih;
90 struct proc *intrproc;
91 
92 static MALLOC_DEFINE(M_ITHREAD, "ithread", "Interrupt Threads");
93 
94 static int intr_storm_threshold = 0;
95 SYSCTL_INT(_hw, OID_AUTO, intr_storm_threshold, CTLFLAG_RWTUN,
96     &intr_storm_threshold, 0,
97     "Number of consecutive interrupts before storm protection is enabled");
98 static int intr_epoch_batch = 1000;
99 SYSCTL_INT(_hw, OID_AUTO, intr_epoch_batch, CTLFLAG_RWTUN, &intr_epoch_batch,
100     0, "Maximum interrupt handler executions without re-entering epoch(9)");
101 static TAILQ_HEAD(, intr_event) event_list =
102     TAILQ_HEAD_INITIALIZER(event_list);
103 static struct mtx event_lock;
104 MTX_SYSINIT(intr_event_list, &event_lock, "intr event list", MTX_DEF);
105 
106 static void	intr_event_update(struct intr_event *ie);
107 static int	intr_event_schedule_thread(struct intr_event *ie);
108 static struct intr_thread *ithread_create(const char *name);
109 static void	ithread_destroy(struct intr_thread *ithread);
110 static void	ithread_execute_handlers(struct proc *p,
111 		    struct intr_event *ie);
112 static void	ithread_loop(void *);
113 static void	ithread_update(struct intr_thread *ithd);
114 static void	start_softintr(void *);
115 
116 /* Map an interrupt type to an ithread priority. */
117 u_char
118 intr_priority(enum intr_type flags)
119 {
120 	u_char pri;
121 
122 	flags &= (INTR_TYPE_TTY | INTR_TYPE_BIO | INTR_TYPE_NET |
123 	    INTR_TYPE_CAM | INTR_TYPE_MISC | INTR_TYPE_CLK | INTR_TYPE_AV);
124 	switch (flags) {
125 	case INTR_TYPE_TTY:
126 		pri = PI_TTY;
127 		break;
128 	case INTR_TYPE_BIO:
129 		pri = PI_DISK;
130 		break;
131 	case INTR_TYPE_NET:
132 		pri = PI_NET;
133 		break;
134 	case INTR_TYPE_CAM:
135 		pri = PI_DISK;
136 		break;
137 	case INTR_TYPE_AV:
138 		pri = PI_AV;
139 		break;
140 	case INTR_TYPE_CLK:
141 		pri = PI_REALTIME;
142 		break;
143 	case INTR_TYPE_MISC:
144 		pri = PI_DULL;          /* don't care */
145 		break;
146 	default:
147 		/* We didn't specify an interrupt level. */
148 		panic("intr_priority: no interrupt type in flags");
149 	}
150 
151 	return pri;
152 }
153 
154 /*
155  * Update an ithread based on the associated intr_event.
156  */
157 static void
158 ithread_update(struct intr_thread *ithd)
159 {
160 	struct intr_event *ie;
161 	struct thread *td;
162 	u_char pri;
163 
164 	ie = ithd->it_event;
165 	td = ithd->it_thread;
166 	mtx_assert(&ie->ie_lock, MA_OWNED);
167 
168 	/* Determine the overall priority of this event. */
169 	if (CK_SLIST_EMPTY(&ie->ie_handlers))
170 		pri = PRI_MAX_ITHD;
171 	else
172 		pri = CK_SLIST_FIRST(&ie->ie_handlers)->ih_pri;
173 
174 	/* Update name and priority. */
175 	strlcpy(td->td_name, ie->ie_fullname, sizeof(td->td_name));
176 #ifdef KTR
177 	sched_clear_tdname(td);
178 #endif
179 	thread_lock(td);
180 	sched_prio(td, pri);
181 	thread_unlock(td);
182 }
183 
184 /*
185  * Regenerate the full name of an interrupt event and update its priority.
186  */
187 static void
188 intr_event_update(struct intr_event *ie)
189 {
190 	struct intr_handler *ih;
191 	char *last;
192 	int missed, space;
193 
194 	/* Start off with no entropy and just the name of the event. */
195 	mtx_assert(&ie->ie_lock, MA_OWNED);
196 	strlcpy(ie->ie_fullname, ie->ie_name, sizeof(ie->ie_fullname));
197 	ie->ie_hflags = 0;
198 	missed = 0;
199 	space = 1;
200 
201 	/* Run through all the handlers updating values. */
202 	CK_SLIST_FOREACH(ih, &ie->ie_handlers, ih_next) {
203 		if (strlen(ie->ie_fullname) + strlen(ih->ih_name) + 1 <
204 		    sizeof(ie->ie_fullname)) {
205 			strcat(ie->ie_fullname, " ");
206 			strcat(ie->ie_fullname, ih->ih_name);
207 			space = 0;
208 		} else
209 			missed++;
210 		ie->ie_hflags |= ih->ih_flags;
211 	}
212 
213 	/*
214 	 * If there is only one handler and its name is too long, just copy in
215 	 * as much of the end of the name (includes the unit number) as will
216 	 * fit.  Otherwise, we have multiple handlers and not all of the names
217 	 * will fit.  Add +'s to indicate missing names.  If we run out of room
218 	 * and still have +'s to add, change the last character from a + to a *.
219 	 */
220 	if (missed == 1 && space == 1) {
221 		ih = CK_SLIST_FIRST(&ie->ie_handlers);
222 		missed = strlen(ie->ie_fullname) + strlen(ih->ih_name) + 2 -
223 		    sizeof(ie->ie_fullname);
224 		strcat(ie->ie_fullname, (missed == 0) ? " " : "-");
225 		strcat(ie->ie_fullname, &ih->ih_name[missed]);
226 		missed = 0;
227 	}
228 	last = &ie->ie_fullname[sizeof(ie->ie_fullname) - 2];
229 	while (missed-- > 0) {
230 		if (strlen(ie->ie_fullname) + 1 == sizeof(ie->ie_fullname)) {
231 			if (*last == '+') {
232 				*last = '*';
233 				break;
234 			} else
235 				*last = '+';
236 		} else if (space) {
237 			strcat(ie->ie_fullname, " +");
238 			space = 0;
239 		} else
240 			strcat(ie->ie_fullname, "+");
241 	}
242 
243 	/*
244 	 * If this event has an ithread, update it's priority and
245 	 * name.
246 	 */
247 	if (ie->ie_thread != NULL)
248 		ithread_update(ie->ie_thread);
249 	CTR2(KTR_INTR, "%s: updated %s", __func__, ie->ie_fullname);
250 }
251 
252 int
253 intr_event_create(struct intr_event **event, void *source, int flags, int irq,
254     void (*pre_ithread)(void *), void (*post_ithread)(void *),
255     void (*post_filter)(void *), int (*assign_cpu)(void *, int),
256     const char *fmt, ...)
257 {
258 	struct intr_event *ie;
259 	va_list ap;
260 
261 	/* The only valid flag during creation is IE_SOFT. */
262 	if ((flags & ~IE_SOFT) != 0)
263 		return (EINVAL);
264 	ie = malloc(sizeof(struct intr_event), M_ITHREAD, M_WAITOK | M_ZERO);
265 	ie->ie_source = source;
266 	ie->ie_pre_ithread = pre_ithread;
267 	ie->ie_post_ithread = post_ithread;
268 	ie->ie_post_filter = post_filter;
269 	ie->ie_assign_cpu = assign_cpu;
270 	ie->ie_flags = flags;
271 	ie->ie_irq = irq;
272 	ie->ie_cpu = NOCPU;
273 	CK_SLIST_INIT(&ie->ie_handlers);
274 	mtx_init(&ie->ie_lock, "intr event", NULL, MTX_DEF);
275 
276 	va_start(ap, fmt);
277 	vsnprintf(ie->ie_name, sizeof(ie->ie_name), fmt, ap);
278 	va_end(ap);
279 	strlcpy(ie->ie_fullname, ie->ie_name, sizeof(ie->ie_fullname));
280 	mtx_lock(&event_lock);
281 	TAILQ_INSERT_TAIL(&event_list, ie, ie_list);
282 	mtx_unlock(&event_lock);
283 	if (event != NULL)
284 		*event = ie;
285 	CTR2(KTR_INTR, "%s: created %s", __func__, ie->ie_name);
286 	return (0);
287 }
288 
289 /*
290  * Bind an interrupt event to the specified CPU.  Note that not all
291  * platforms support binding an interrupt to a CPU.  For those
292  * platforms this request will fail.  Using a cpu id of NOCPU unbinds
293  * the interrupt event.
294  */
295 static int
296 _intr_event_bind(struct intr_event *ie, int cpu, bool bindirq, bool bindithread)
297 {
298 	lwpid_t id;
299 	int error;
300 
301 	/* Need a CPU to bind to. */
302 	if (cpu != NOCPU && CPU_ABSENT(cpu))
303 		return (EINVAL);
304 
305 	if (ie->ie_assign_cpu == NULL)
306 		return (EOPNOTSUPP);
307 
308 	error = priv_check(curthread, PRIV_SCHED_CPUSET_INTR);
309 	if (error)
310 		return (error);
311 
312 	/*
313 	 * If we have any ithreads try to set their mask first to verify
314 	 * permissions, etc.
315 	 */
316 	if (bindithread) {
317 		mtx_lock(&ie->ie_lock);
318 		if (ie->ie_thread != NULL) {
319 			id = ie->ie_thread->it_thread->td_tid;
320 			mtx_unlock(&ie->ie_lock);
321 			error = cpuset_setithread(id, cpu);
322 			if (error)
323 				return (error);
324 		} else
325 			mtx_unlock(&ie->ie_lock);
326 	}
327 	if (bindirq)
328 		error = ie->ie_assign_cpu(ie->ie_source, cpu);
329 	if (error) {
330 		if (bindithread) {
331 			mtx_lock(&ie->ie_lock);
332 			if (ie->ie_thread != NULL) {
333 				cpu = ie->ie_cpu;
334 				id = ie->ie_thread->it_thread->td_tid;
335 				mtx_unlock(&ie->ie_lock);
336 				(void)cpuset_setithread(id, cpu);
337 			} else
338 				mtx_unlock(&ie->ie_lock);
339 		}
340 		return (error);
341 	}
342 
343 	if (bindirq) {
344 		mtx_lock(&ie->ie_lock);
345 		ie->ie_cpu = cpu;
346 		mtx_unlock(&ie->ie_lock);
347 	}
348 
349 	return (error);
350 }
351 
352 /*
353  * Bind an interrupt event to the specified CPU.  For supported platforms, any
354  * associated ithreads as well as the primary interrupt context will be bound
355  * to the specificed CPU.
356  */
357 int
358 intr_event_bind(struct intr_event *ie, int cpu)
359 {
360 
361 	return (_intr_event_bind(ie, cpu, true, true));
362 }
363 
364 /*
365  * Bind an interrupt event to the specified CPU, but do not bind associated
366  * ithreads.
367  */
368 int
369 intr_event_bind_irqonly(struct intr_event *ie, int cpu)
370 {
371 
372 	return (_intr_event_bind(ie, cpu, true, false));
373 }
374 
375 /*
376  * Bind an interrupt event's ithread to the specified CPU.
377  */
378 int
379 intr_event_bind_ithread(struct intr_event *ie, int cpu)
380 {
381 
382 	return (_intr_event_bind(ie, cpu, false, true));
383 }
384 
385 /*
386  * Bind an interrupt event's ithread to the specified cpuset.
387  */
388 int
389 intr_event_bind_ithread_cpuset(struct intr_event *ie, cpuset_t *cs)
390 {
391 	lwpid_t id;
392 
393 	mtx_lock(&ie->ie_lock);
394 	if (ie->ie_thread != NULL) {
395 		id = ie->ie_thread->it_thread->td_tid;
396 		mtx_unlock(&ie->ie_lock);
397 		return (cpuset_setthread(id, cs));
398 	} else {
399 		mtx_unlock(&ie->ie_lock);
400 	}
401 	return (ENODEV);
402 }
403 
404 static struct intr_event *
405 intr_lookup(int irq)
406 {
407 	struct intr_event *ie;
408 
409 	mtx_lock(&event_lock);
410 	TAILQ_FOREACH(ie, &event_list, ie_list)
411 		if (ie->ie_irq == irq &&
412 		    (ie->ie_flags & IE_SOFT) == 0 &&
413 		    CK_SLIST_FIRST(&ie->ie_handlers) != NULL)
414 			break;
415 	mtx_unlock(&event_lock);
416 	return (ie);
417 }
418 
419 int
420 intr_setaffinity(int irq, int mode, void *m)
421 {
422 	struct intr_event *ie;
423 	cpuset_t *mask;
424 	int cpu, n;
425 
426 	mask = m;
427 	cpu = NOCPU;
428 	/*
429 	 * If we're setting all cpus we can unbind.  Otherwise make sure
430 	 * only one cpu is in the set.
431 	 */
432 	if (CPU_CMP(cpuset_root, mask)) {
433 		for (n = 0; n < CPU_SETSIZE; n++) {
434 			if (!CPU_ISSET(n, mask))
435 				continue;
436 			if (cpu != NOCPU)
437 				return (EINVAL);
438 			cpu = n;
439 		}
440 	}
441 	ie = intr_lookup(irq);
442 	if (ie == NULL)
443 		return (ESRCH);
444 	switch (mode) {
445 	case CPU_WHICH_IRQ:
446 		return (intr_event_bind(ie, cpu));
447 	case CPU_WHICH_INTRHANDLER:
448 		return (intr_event_bind_irqonly(ie, cpu));
449 	case CPU_WHICH_ITHREAD:
450 		return (intr_event_bind_ithread(ie, cpu));
451 	default:
452 		return (EINVAL);
453 	}
454 }
455 
456 int
457 intr_getaffinity(int irq, int mode, void *m)
458 {
459 	struct intr_event *ie;
460 	struct thread *td;
461 	struct proc *p;
462 	cpuset_t *mask;
463 	lwpid_t id;
464 	int error;
465 
466 	mask = m;
467 	ie = intr_lookup(irq);
468 	if (ie == NULL)
469 		return (ESRCH);
470 
471 	error = 0;
472 	CPU_ZERO(mask);
473 	switch (mode) {
474 	case CPU_WHICH_IRQ:
475 	case CPU_WHICH_INTRHANDLER:
476 		mtx_lock(&ie->ie_lock);
477 		if (ie->ie_cpu == NOCPU)
478 			CPU_COPY(cpuset_root, mask);
479 		else
480 			CPU_SET(ie->ie_cpu, mask);
481 		mtx_unlock(&ie->ie_lock);
482 		break;
483 	case CPU_WHICH_ITHREAD:
484 		mtx_lock(&ie->ie_lock);
485 		if (ie->ie_thread == NULL) {
486 			mtx_unlock(&ie->ie_lock);
487 			CPU_COPY(cpuset_root, mask);
488 		} else {
489 			id = ie->ie_thread->it_thread->td_tid;
490 			mtx_unlock(&ie->ie_lock);
491 			error = cpuset_which(CPU_WHICH_TID, id, &p, &td, NULL);
492 			if (error != 0)
493 				return (error);
494 			CPU_COPY(&td->td_cpuset->cs_mask, mask);
495 			PROC_UNLOCK(p);
496 		}
497 	default:
498 		return (EINVAL);
499 	}
500 	return (0);
501 }
502 
503 int
504 intr_event_destroy(struct intr_event *ie)
505 {
506 
507 	mtx_lock(&event_lock);
508 	mtx_lock(&ie->ie_lock);
509 	if (!CK_SLIST_EMPTY(&ie->ie_handlers)) {
510 		mtx_unlock(&ie->ie_lock);
511 		mtx_unlock(&event_lock);
512 		return (EBUSY);
513 	}
514 	TAILQ_REMOVE(&event_list, ie, ie_list);
515 #ifndef notyet
516 	if (ie->ie_thread != NULL) {
517 		ithread_destroy(ie->ie_thread);
518 		ie->ie_thread = NULL;
519 	}
520 #endif
521 	mtx_unlock(&ie->ie_lock);
522 	mtx_unlock(&event_lock);
523 	mtx_destroy(&ie->ie_lock);
524 	free(ie, M_ITHREAD);
525 	return (0);
526 }
527 
528 static struct intr_thread *
529 ithread_create(const char *name)
530 {
531 	struct intr_thread *ithd;
532 	struct thread *td;
533 	int error;
534 
535 	ithd = malloc(sizeof(struct intr_thread), M_ITHREAD, M_WAITOK | M_ZERO);
536 
537 	error = kproc_kthread_add(ithread_loop, ithd, &intrproc,
538 		    &td, RFSTOPPED | RFHIGHPID,
539 		    0, "intr", "%s", name);
540 	if (error)
541 		panic("kproc_create() failed with %d", error);
542 	thread_lock(td);
543 	sched_class(td, PRI_ITHD);
544 	TD_SET_IWAIT(td);
545 	thread_unlock(td);
546 	td->td_pflags |= TDP_ITHREAD;
547 	ithd->it_thread = td;
548 	CTR2(KTR_INTR, "%s: created %s", __func__, name);
549 	return (ithd);
550 }
551 
552 static void
553 ithread_destroy(struct intr_thread *ithread)
554 {
555 	struct thread *td;
556 
557 	CTR2(KTR_INTR, "%s: killing %s", __func__, ithread->it_event->ie_name);
558 	td = ithread->it_thread;
559 	thread_lock(td);
560 	ithread->it_flags |= IT_DEAD;
561 	if (TD_AWAITING_INTR(td)) {
562 		TD_CLR_IWAIT(td);
563 		sched_add(td, SRQ_INTR);
564 	} else
565 		thread_unlock(td);
566 }
567 
568 int
569 intr_event_add_handler(struct intr_event *ie, const char *name,
570     driver_filter_t filter, driver_intr_t handler, void *arg, u_char pri,
571     enum intr_type flags, void **cookiep)
572 {
573 	struct intr_handler *ih, *temp_ih;
574 	struct intr_handler **prevptr;
575 	struct intr_thread *it;
576 
577 	if (ie == NULL || name == NULL || (handler == NULL && filter == NULL))
578 		return (EINVAL);
579 
580 	/* Allocate and populate an interrupt handler structure. */
581 	ih = malloc(sizeof(struct intr_handler), M_ITHREAD, M_WAITOK | M_ZERO);
582 	ih->ih_filter = filter;
583 	ih->ih_handler = handler;
584 	ih->ih_argument = arg;
585 	strlcpy(ih->ih_name, name, sizeof(ih->ih_name));
586 	ih->ih_event = ie;
587 	ih->ih_pri = pri;
588 	if (flags & INTR_EXCL)
589 		ih->ih_flags = IH_EXCLUSIVE;
590 	if (flags & INTR_MPSAFE)
591 		ih->ih_flags |= IH_MPSAFE;
592 	if (flags & INTR_ENTROPY)
593 		ih->ih_flags |= IH_ENTROPY;
594 	if (flags & INTR_TYPE_NET)
595 		ih->ih_flags |= IH_NET;
596 
597 	/* We can only have one exclusive handler in a event. */
598 	mtx_lock(&ie->ie_lock);
599 	if (!CK_SLIST_EMPTY(&ie->ie_handlers)) {
600 		if ((flags & INTR_EXCL) ||
601 		    (CK_SLIST_FIRST(&ie->ie_handlers)->ih_flags & IH_EXCLUSIVE)) {
602 			mtx_unlock(&ie->ie_lock);
603 			free(ih, M_ITHREAD);
604 			return (EINVAL);
605 		}
606 	}
607 
608 	/* Create a thread if we need one. */
609 	while (ie->ie_thread == NULL && handler != NULL) {
610 		if (ie->ie_flags & IE_ADDING_THREAD)
611 			msleep(ie, &ie->ie_lock, 0, "ithread", 0);
612 		else {
613 			ie->ie_flags |= IE_ADDING_THREAD;
614 			mtx_unlock(&ie->ie_lock);
615 			it = ithread_create("intr: newborn");
616 			mtx_lock(&ie->ie_lock);
617 			ie->ie_flags &= ~IE_ADDING_THREAD;
618 			ie->ie_thread = it;
619 			it->it_event = ie;
620 			ithread_update(it);
621 			wakeup(ie);
622 		}
623 	}
624 
625 	/* Add the new handler to the event in priority order. */
626 	CK_SLIST_FOREACH_PREVPTR(temp_ih, prevptr, &ie->ie_handlers, ih_next) {
627 		if (temp_ih->ih_pri > ih->ih_pri)
628 			break;
629 	}
630 	CK_SLIST_INSERT_PREVPTR(prevptr, temp_ih, ih, ih_next);
631 
632 	intr_event_update(ie);
633 
634 	CTR3(KTR_INTR, "%s: added %s to %s", __func__, ih->ih_name,
635 	    ie->ie_name);
636 	mtx_unlock(&ie->ie_lock);
637 
638 	if (cookiep != NULL)
639 		*cookiep = ih;
640 	return (0);
641 }
642 
643 /*
644  * Append a description preceded by a ':' to the name of the specified
645  * interrupt handler.
646  */
647 int
648 intr_event_describe_handler(struct intr_event *ie, void *cookie,
649     const char *descr)
650 {
651 	struct intr_handler *ih;
652 	size_t space;
653 	char *start;
654 
655 	mtx_lock(&ie->ie_lock);
656 #ifdef INVARIANTS
657 	CK_SLIST_FOREACH(ih, &ie->ie_handlers, ih_next) {
658 		if (ih == cookie)
659 			break;
660 	}
661 	if (ih == NULL) {
662 		mtx_unlock(&ie->ie_lock);
663 		panic("handler %p not found in interrupt event %p", cookie, ie);
664 	}
665 #endif
666 	ih = cookie;
667 
668 	/*
669 	 * Look for an existing description by checking for an
670 	 * existing ":".  This assumes device names do not include
671 	 * colons.  If one is found, prepare to insert the new
672 	 * description at that point.  If one is not found, find the
673 	 * end of the name to use as the insertion point.
674 	 */
675 	start = strchr(ih->ih_name, ':');
676 	if (start == NULL)
677 		start = strchr(ih->ih_name, 0);
678 
679 	/*
680 	 * See if there is enough remaining room in the string for the
681 	 * description + ":".  The "- 1" leaves room for the trailing
682 	 * '\0'.  The "+ 1" accounts for the colon.
683 	 */
684 	space = sizeof(ih->ih_name) - (start - ih->ih_name) - 1;
685 	if (strlen(descr) + 1 > space) {
686 		mtx_unlock(&ie->ie_lock);
687 		return (ENOSPC);
688 	}
689 
690 	/* Append a colon followed by the description. */
691 	*start = ':';
692 	strcpy(start + 1, descr);
693 	intr_event_update(ie);
694 	mtx_unlock(&ie->ie_lock);
695 	return (0);
696 }
697 
698 /*
699  * Return the ie_source field from the intr_event an intr_handler is
700  * associated with.
701  */
702 void *
703 intr_handler_source(void *cookie)
704 {
705 	struct intr_handler *ih;
706 	struct intr_event *ie;
707 
708 	ih = (struct intr_handler *)cookie;
709 	if (ih == NULL)
710 		return (NULL);
711 	ie = ih->ih_event;
712 	KASSERT(ie != NULL,
713 	    ("interrupt handler \"%s\" has a NULL interrupt event",
714 	    ih->ih_name));
715 	return (ie->ie_source);
716 }
717 
718 /*
719  * If intr_event_handle() is running in the ISR context at the time of the call,
720  * then wait for it to complete.
721  */
722 static void
723 intr_event_barrier(struct intr_event *ie)
724 {
725 	int phase;
726 
727 	mtx_assert(&ie->ie_lock, MA_OWNED);
728 	phase = ie->ie_phase;
729 
730 	/*
731 	 * Switch phase to direct future interrupts to the other active counter.
732 	 * Make sure that any preceding stores are visible before the switch.
733 	 */
734 	KASSERT(ie->ie_active[!phase] == 0, ("idle phase has activity"));
735 	atomic_store_rel_int(&ie->ie_phase, !phase);
736 
737 	/*
738 	 * This code cooperates with wait-free iteration of ie_handlers
739 	 * in intr_event_handle.
740 	 * Make sure that the removal and the phase update are not reordered
741 	 * with the active count check.
742 	 * Note that no combination of acquire and release fences can provide
743 	 * that guarantee as Store->Load sequences can always be reordered.
744 	 */
745 	atomic_thread_fence_seq_cst();
746 
747 	/*
748 	 * Now wait on the inactive phase.
749 	 * The acquire fence is needed so that that all post-barrier accesses
750 	 * are after the check.
751 	 */
752 	while (ie->ie_active[phase] > 0)
753 		cpu_spinwait();
754 	atomic_thread_fence_acq();
755 }
756 
757 static void
758 intr_handler_barrier(struct intr_handler *handler)
759 {
760 	struct intr_event *ie;
761 
762 	ie = handler->ih_event;
763 	mtx_assert(&ie->ie_lock, MA_OWNED);
764 	KASSERT((handler->ih_flags & IH_DEAD) == 0,
765 	    ("update for a removed handler"));
766 
767 	if (ie->ie_thread == NULL) {
768 		intr_event_barrier(ie);
769 		return;
770 	}
771 	if ((handler->ih_flags & IH_CHANGED) == 0) {
772 		handler->ih_flags |= IH_CHANGED;
773 		intr_event_schedule_thread(ie);
774 	}
775 	while ((handler->ih_flags & IH_CHANGED) != 0)
776 		msleep(handler, &ie->ie_lock, 0, "ih_barr", 0);
777 }
778 
779 /*
780  * Sleep until an ithread finishes executing an interrupt handler.
781  *
782  * XXX Doesn't currently handle interrupt filters or fast interrupt
783  * handlers.  This is intended for compatibility with linux drivers
784  * only.  Do not use in BSD code.
785  */
786 void
787 _intr_drain(int irq)
788 {
789 	struct intr_event *ie;
790 	struct intr_thread *ithd;
791 	struct thread *td;
792 
793 	ie = intr_lookup(irq);
794 	if (ie == NULL)
795 		return;
796 	if (ie->ie_thread == NULL)
797 		return;
798 	ithd = ie->ie_thread;
799 	td = ithd->it_thread;
800 	/*
801 	 * We set the flag and wait for it to be cleared to avoid
802 	 * long delays with potentially busy interrupt handlers
803 	 * were we to only sample TD_AWAITING_INTR() every tick.
804 	 */
805 	thread_lock(td);
806 	if (!TD_AWAITING_INTR(td)) {
807 		ithd->it_flags |= IT_WAIT;
808 		while (ithd->it_flags & IT_WAIT) {
809 			thread_unlock(td);
810 			pause("idrain", 1);
811 			thread_lock(td);
812 		}
813 	}
814 	thread_unlock(td);
815 	return;
816 }
817 
818 int
819 intr_event_remove_handler(void *cookie)
820 {
821 	struct intr_handler *handler = (struct intr_handler *)cookie;
822 	struct intr_event *ie;
823 	struct intr_handler *ih;
824 	struct intr_handler **prevptr;
825 #ifdef notyet
826 	int dead;
827 #endif
828 
829 	if (handler == NULL)
830 		return (EINVAL);
831 	ie = handler->ih_event;
832 	KASSERT(ie != NULL,
833 	    ("interrupt handler \"%s\" has a NULL interrupt event",
834 	    handler->ih_name));
835 
836 	mtx_lock(&ie->ie_lock);
837 	CTR3(KTR_INTR, "%s: removing %s from %s", __func__, handler->ih_name,
838 	    ie->ie_name);
839 	CK_SLIST_FOREACH_PREVPTR(ih, prevptr, &ie->ie_handlers, ih_next) {
840 		if (ih == handler)
841 			break;
842 	}
843 	if (ih == NULL) {
844 		panic("interrupt handler \"%s\" not found in "
845 		    "interrupt event \"%s\"", handler->ih_name, ie->ie_name);
846 	}
847 
848 	/*
849 	 * If there is no ithread, then directly remove the handler.  Note that
850 	 * intr_event_handle() iterates ie_handlers in a lock-less fashion, so
851 	 * care needs to be taken to keep ie_handlers consistent and to free
852 	 * the removed handler only when ie_handlers is quiescent.
853 	 */
854 	if (ie->ie_thread == NULL) {
855 		CK_SLIST_REMOVE_PREVPTR(prevptr, ih, ih_next);
856 		intr_event_barrier(ie);
857 		intr_event_update(ie);
858 		mtx_unlock(&ie->ie_lock);
859 		free(handler, M_ITHREAD);
860 		return (0);
861 	}
862 
863 	/*
864 	 * Let the interrupt thread do the job.
865 	 * The interrupt source is disabled when the interrupt thread is
866 	 * running, so it does not have to worry about interaction with
867 	 * intr_event_handle().
868 	 */
869 	KASSERT((handler->ih_flags & IH_DEAD) == 0,
870 	    ("duplicate handle remove"));
871 	handler->ih_flags |= IH_DEAD;
872 	intr_event_schedule_thread(ie);
873 	while (handler->ih_flags & IH_DEAD)
874 		msleep(handler, &ie->ie_lock, 0, "iev_rmh", 0);
875 	intr_event_update(ie);
876 
877 #ifdef notyet
878 	/*
879 	 * XXX: This could be bad in the case of ppbus(8).  Also, I think
880 	 * this could lead to races of stale data when servicing an
881 	 * interrupt.
882 	 */
883 	dead = 1;
884 	CK_SLIST_FOREACH(ih, &ie->ie_handlers, ih_next) {
885 		if (ih->ih_handler != NULL) {
886 			dead = 0;
887 			break;
888 		}
889 	}
890 	if (dead) {
891 		ithread_destroy(ie->ie_thread);
892 		ie->ie_thread = NULL;
893 	}
894 #endif
895 	mtx_unlock(&ie->ie_lock);
896 	free(handler, M_ITHREAD);
897 	return (0);
898 }
899 
900 int
901 intr_event_suspend_handler(void *cookie)
902 {
903 	struct intr_handler *handler = (struct intr_handler *)cookie;
904 	struct intr_event *ie;
905 
906 	if (handler == NULL)
907 		return (EINVAL);
908 	ie = handler->ih_event;
909 	KASSERT(ie != NULL,
910 	    ("interrupt handler \"%s\" has a NULL interrupt event",
911 	    handler->ih_name));
912 	mtx_lock(&ie->ie_lock);
913 	handler->ih_flags |= IH_SUSP;
914 	intr_handler_barrier(handler);
915 	mtx_unlock(&ie->ie_lock);
916 	return (0);
917 }
918 
919 int
920 intr_event_resume_handler(void *cookie)
921 {
922 	struct intr_handler *handler = (struct intr_handler *)cookie;
923 	struct intr_event *ie;
924 
925 	if (handler == NULL)
926 		return (EINVAL);
927 	ie = handler->ih_event;
928 	KASSERT(ie != NULL,
929 	    ("interrupt handler \"%s\" has a NULL interrupt event",
930 	    handler->ih_name));
931 
932 	/*
933 	 * intr_handler_barrier() acts not only as a barrier,
934 	 * it also allows to check for any pending interrupts.
935 	 */
936 	mtx_lock(&ie->ie_lock);
937 	handler->ih_flags &= ~IH_SUSP;
938 	intr_handler_barrier(handler);
939 	mtx_unlock(&ie->ie_lock);
940 	return (0);
941 }
942 
943 static int
944 intr_event_schedule_thread(struct intr_event *ie)
945 {
946 	struct intr_entropy entropy;
947 	struct intr_thread *it;
948 	struct thread *td;
949 	struct thread *ctd;
950 
951 	/*
952 	 * If no ithread or no handlers, then we have a stray interrupt.
953 	 */
954 	if (ie == NULL || CK_SLIST_EMPTY(&ie->ie_handlers) ||
955 	    ie->ie_thread == NULL)
956 		return (EINVAL);
957 
958 	ctd = curthread;
959 	it = ie->ie_thread;
960 	td = it->it_thread;
961 
962 	/*
963 	 * If any of the handlers for this ithread claim to be good
964 	 * sources of entropy, then gather some.
965 	 */
966 	if (ie->ie_hflags & IH_ENTROPY) {
967 		entropy.event = (uintptr_t)ie;
968 		entropy.td = ctd;
969 		random_harvest_queue(&entropy, sizeof(entropy), RANDOM_INTERRUPT);
970 	}
971 
972 	KASSERT(td->td_proc != NULL, ("ithread %s has no process", ie->ie_name));
973 
974 	/*
975 	 * Set it_need to tell the thread to keep running if it is already
976 	 * running.  Then, lock the thread and see if we actually need to
977 	 * put it on the runqueue.
978 	 *
979 	 * Use store_rel to arrange that the store to ih_need in
980 	 * swi_sched() is before the store to it_need and prepare for
981 	 * transfer of this order to loads in the ithread.
982 	 */
983 	atomic_store_rel_int(&it->it_need, 1);
984 	thread_lock(td);
985 	if (TD_AWAITING_INTR(td)) {
986 		CTR3(KTR_INTR, "%s: schedule pid %d (%s)", __func__, td->td_proc->p_pid,
987 		    td->td_name);
988 		TD_CLR_IWAIT(td);
989 		sched_add(td, SRQ_INTR);
990 	} else {
991 		CTR5(KTR_INTR, "%s: pid %d (%s): it_need %d, state %d",
992 		    __func__, td->td_proc->p_pid, td->td_name, it->it_need, td->td_state);
993 		thread_unlock(td);
994 	}
995 
996 	return (0);
997 }
998 
999 /*
1000  * Allow interrupt event binding for software interrupt handlers -- a no-op,
1001  * since interrupts are generated in software rather than being directed by
1002  * a PIC.
1003  */
1004 static int
1005 swi_assign_cpu(void *arg, int cpu)
1006 {
1007 
1008 	return (0);
1009 }
1010 
1011 /*
1012  * Add a software interrupt handler to a specified event.  If a given event
1013  * is not specified, then a new event is created.
1014  */
1015 int
1016 swi_add(struct intr_event **eventp, const char *name, driver_intr_t handler,
1017 	    void *arg, int pri, enum intr_type flags, void **cookiep)
1018 {
1019 	struct intr_event *ie;
1020 	int error;
1021 
1022 	if (flags & INTR_ENTROPY)
1023 		return (EINVAL);
1024 
1025 	ie = (eventp != NULL) ? *eventp : NULL;
1026 
1027 	if (ie != NULL) {
1028 		if (!(ie->ie_flags & IE_SOFT))
1029 			return (EINVAL);
1030 	} else {
1031 		error = intr_event_create(&ie, NULL, IE_SOFT, 0,
1032 		    NULL, NULL, NULL, swi_assign_cpu, "swi%d:", pri);
1033 		if (error)
1034 			return (error);
1035 		if (eventp != NULL)
1036 			*eventp = ie;
1037 	}
1038 	error = intr_event_add_handler(ie, name, NULL, handler, arg,
1039 	    PI_SWI(pri), flags, cookiep);
1040 	return (error);
1041 }
1042 
1043 /*
1044  * Schedule a software interrupt thread.
1045  */
1046 void
1047 swi_sched(void *cookie, int flags)
1048 {
1049 	struct intr_handler *ih = (struct intr_handler *)cookie;
1050 	struct intr_event *ie = ih->ih_event;
1051 	struct intr_entropy entropy;
1052 	int error __unused;
1053 
1054 	CTR3(KTR_INTR, "swi_sched: %s %s need=%d", ie->ie_name, ih->ih_name,
1055 	    ih->ih_need);
1056 
1057 	entropy.event = (uintptr_t)ih;
1058 	entropy.td = curthread;
1059 	random_harvest_queue(&entropy, sizeof(entropy), RANDOM_SWI);
1060 
1061 	/*
1062 	 * Set ih_need for this handler so that if the ithread is already
1063 	 * running it will execute this handler on the next pass.  Otherwise,
1064 	 * it will execute it the next time it runs.
1065 	 */
1066 	ih->ih_need = 1;
1067 
1068 	if (!(flags & SWI_DELAY)) {
1069 		VM_CNT_INC(v_soft);
1070 		error = intr_event_schedule_thread(ie);
1071 		KASSERT(error == 0, ("stray software interrupt"));
1072 	}
1073 }
1074 
1075 /*
1076  * Remove a software interrupt handler.  Currently this code does not
1077  * remove the associated interrupt event if it becomes empty.  Calling code
1078  * may do so manually via intr_event_destroy(), but that's not really
1079  * an optimal interface.
1080  */
1081 int
1082 swi_remove(void *cookie)
1083 {
1084 
1085 	return (intr_event_remove_handler(cookie));
1086 }
1087 
1088 static void
1089 intr_event_execute_handlers(struct proc *p, struct intr_event *ie)
1090 {
1091 	struct intr_handler *ih, *ihn, *ihp;
1092 
1093 	ihp = NULL;
1094 	CK_SLIST_FOREACH_SAFE(ih, &ie->ie_handlers, ih_next, ihn) {
1095 		/*
1096 		 * If this handler is marked for death, remove it from
1097 		 * the list of handlers and wake up the sleeper.
1098 		 */
1099 		if (ih->ih_flags & IH_DEAD) {
1100 			mtx_lock(&ie->ie_lock);
1101 			if (ihp == NULL)
1102 				CK_SLIST_REMOVE_HEAD(&ie->ie_handlers, ih_next);
1103 			else
1104 				CK_SLIST_REMOVE_AFTER(ihp, ih_next);
1105 			ih->ih_flags &= ~IH_DEAD;
1106 			wakeup(ih);
1107 			mtx_unlock(&ie->ie_lock);
1108 			continue;
1109 		}
1110 
1111 		/*
1112 		 * Now that we know that the current element won't be removed
1113 		 * update the previous element.
1114 		 */
1115 		ihp = ih;
1116 
1117 		if ((ih->ih_flags & IH_CHANGED) != 0) {
1118 			mtx_lock(&ie->ie_lock);
1119 			ih->ih_flags &= ~IH_CHANGED;
1120 			wakeup(ih);
1121 			mtx_unlock(&ie->ie_lock);
1122 		}
1123 
1124 		/* Skip filter only handlers */
1125 		if (ih->ih_handler == NULL)
1126 			continue;
1127 
1128 		/* Skip suspended handlers */
1129 		if ((ih->ih_flags & IH_SUSP) != 0)
1130 			continue;
1131 
1132 		/*
1133 		 * For software interrupt threads, we only execute
1134 		 * handlers that have their need flag set.  Hardware
1135 		 * interrupt threads always invoke all of their handlers.
1136 		 *
1137 		 * ih_need can only be 0 or 1.  Failed cmpset below
1138 		 * means that there is no request to execute handlers,
1139 		 * so a retry of the cmpset is not needed.
1140 		 */
1141 		if ((ie->ie_flags & IE_SOFT) != 0 &&
1142 		    atomic_cmpset_int(&ih->ih_need, 1, 0) == 0)
1143 			continue;
1144 
1145 		/* Execute this handler. */
1146 		CTR6(KTR_INTR, "%s: pid %d exec %p(%p) for %s flg=%x",
1147 		    __func__, p->p_pid, (void *)ih->ih_handler,
1148 		    ih->ih_argument, ih->ih_name, ih->ih_flags);
1149 
1150 		if (!(ih->ih_flags & IH_MPSAFE))
1151 			mtx_lock(&Giant);
1152 		ih->ih_handler(ih->ih_argument);
1153 		if (!(ih->ih_flags & IH_MPSAFE))
1154 			mtx_unlock(&Giant);
1155 	}
1156 }
1157 
1158 static void
1159 ithread_execute_handlers(struct proc *p, struct intr_event *ie)
1160 {
1161 
1162 	/* Interrupt handlers should not sleep. */
1163 	if (!(ie->ie_flags & IE_SOFT))
1164 		THREAD_NO_SLEEPING();
1165 	intr_event_execute_handlers(p, ie);
1166 	if (!(ie->ie_flags & IE_SOFT))
1167 		THREAD_SLEEPING_OK();
1168 
1169 	/*
1170 	 * Interrupt storm handling:
1171 	 *
1172 	 * If this interrupt source is currently storming, then throttle
1173 	 * it to only fire the handler once  per clock tick.
1174 	 *
1175 	 * If this interrupt source is not currently storming, but the
1176 	 * number of back to back interrupts exceeds the storm threshold,
1177 	 * then enter storming mode.
1178 	 */
1179 	if (intr_storm_threshold != 0 && ie->ie_count >= intr_storm_threshold &&
1180 	    !(ie->ie_flags & IE_SOFT)) {
1181 		/* Report the message only once every second. */
1182 		if (ppsratecheck(&ie->ie_warntm, &ie->ie_warncnt, 1)) {
1183 			printf(
1184 	"interrupt storm detected on \"%s\"; throttling interrupt source\n",
1185 			    ie->ie_name);
1186 		}
1187 		pause("istorm", 1);
1188 	} else
1189 		ie->ie_count++;
1190 
1191 	/*
1192 	 * Now that all the handlers have had a chance to run, reenable
1193 	 * the interrupt source.
1194 	 */
1195 	if (ie->ie_post_ithread != NULL)
1196 		ie->ie_post_ithread(ie->ie_source);
1197 }
1198 
1199 /*
1200  * This is the main code for interrupt threads.
1201  */
1202 static void
1203 ithread_loop(void *arg)
1204 {
1205 	struct epoch_tracker et;
1206 	struct intr_thread *ithd;
1207 	struct intr_event *ie;
1208 	struct thread *td;
1209 	struct proc *p;
1210 	int wake, epoch_count;
1211 
1212 	td = curthread;
1213 	p = td->td_proc;
1214 	ithd = (struct intr_thread *)arg;
1215 	KASSERT(ithd->it_thread == td,
1216 	    ("%s: ithread and proc linkage out of sync", __func__));
1217 	ie = ithd->it_event;
1218 	ie->ie_count = 0;
1219 	wake = 0;
1220 
1221 	/*
1222 	 * As long as we have interrupts outstanding, go through the
1223 	 * list of handlers, giving each one a go at it.
1224 	 */
1225 	for (;;) {
1226 		/*
1227 		 * If we are an orphaned thread, then just die.
1228 		 */
1229 		if (ithd->it_flags & IT_DEAD) {
1230 			CTR3(KTR_INTR, "%s: pid %d (%s) exiting", __func__,
1231 			    p->p_pid, td->td_name);
1232 			free(ithd, M_ITHREAD);
1233 			kthread_exit();
1234 		}
1235 
1236 		/*
1237 		 * Service interrupts.  If another interrupt arrives while
1238 		 * we are running, it will set it_need to note that we
1239 		 * should make another pass.
1240 		 *
1241 		 * The load_acq part of the following cmpset ensures
1242 		 * that the load of ih_need in ithread_execute_handlers()
1243 		 * is ordered after the load of it_need here.
1244 		 */
1245 		if (ie->ie_hflags & IH_NET) {
1246 			epoch_count = 0;
1247 			NET_EPOCH_ENTER(et);
1248 		}
1249 		while (atomic_cmpset_acq_int(&ithd->it_need, 1, 0) != 0) {
1250 			ithread_execute_handlers(p, ie);
1251 			if ((ie->ie_hflags & IH_NET) &&
1252 			    ++epoch_count >= intr_epoch_batch) {
1253 				NET_EPOCH_EXIT(et);
1254 				epoch_count = 0;
1255 				NET_EPOCH_ENTER(et);
1256 			}
1257 		}
1258 		if (ie->ie_hflags & IH_NET)
1259 			NET_EPOCH_EXIT(et);
1260 		WITNESS_WARN(WARN_PANIC, NULL, "suspending ithread");
1261 		mtx_assert(&Giant, MA_NOTOWNED);
1262 
1263 		/*
1264 		 * Processed all our interrupts.  Now get the sched
1265 		 * lock.  This may take a while and it_need may get
1266 		 * set again, so we have to check it again.
1267 		 */
1268 		thread_lock(td);
1269 		if (atomic_load_acq_int(&ithd->it_need) == 0 &&
1270 		    (ithd->it_flags & (IT_DEAD | IT_WAIT)) == 0) {
1271 			TD_SET_IWAIT(td);
1272 			ie->ie_count = 0;
1273 			mi_switch(SW_VOL | SWT_IWAIT);
1274 		} else {
1275 			if (ithd->it_flags & IT_WAIT) {
1276 				wake = 1;
1277 				ithd->it_flags &= ~IT_WAIT;
1278 			}
1279 			thread_unlock(td);
1280 		}
1281 		if (wake) {
1282 			wakeup(ithd);
1283 			wake = 0;
1284 		}
1285 	}
1286 }
1287 
1288 /*
1289  * Main interrupt handling body.
1290  *
1291  * Input:
1292  * o ie:                        the event connected to this interrupt.
1293  * o frame:                     some archs (i.e. i386) pass a frame to some.
1294  *                              handlers as their main argument.
1295  * Return value:
1296  * o 0:                         everything ok.
1297  * o EINVAL:                    stray interrupt.
1298  */
1299 int
1300 intr_event_handle(struct intr_event *ie, struct trapframe *frame)
1301 {
1302 	struct intr_handler *ih;
1303 	struct trapframe *oldframe;
1304 	struct thread *td;
1305 	int phase;
1306 	int ret;
1307 	bool filter, thread;
1308 
1309 	td = curthread;
1310 
1311 #ifdef KSTACK_USAGE_PROF
1312 	intr_prof_stack_use(td, frame);
1313 #endif
1314 
1315 	/* An interrupt with no event or handlers is a stray interrupt. */
1316 	if (ie == NULL || CK_SLIST_EMPTY(&ie->ie_handlers))
1317 		return (EINVAL);
1318 
1319 	/*
1320 	 * Execute fast interrupt handlers directly.
1321 	 * To support clock handlers, if a handler registers
1322 	 * with a NULL argument, then we pass it a pointer to
1323 	 * a trapframe as its argument.
1324 	 */
1325 	td->td_intr_nesting_level++;
1326 	filter = false;
1327 	thread = false;
1328 	ret = 0;
1329 	critical_enter();
1330 	oldframe = td->td_intr_frame;
1331 	td->td_intr_frame = frame;
1332 
1333 	phase = ie->ie_phase;
1334 	atomic_add_int(&ie->ie_active[phase], 1);
1335 
1336 	/*
1337 	 * This fence is required to ensure that no later loads are
1338 	 * re-ordered before the ie_active store.
1339 	 */
1340 	atomic_thread_fence_seq_cst();
1341 
1342 	CK_SLIST_FOREACH(ih, &ie->ie_handlers, ih_next) {
1343 		if ((ih->ih_flags & IH_SUSP) != 0)
1344 			continue;
1345 		if (ih->ih_filter == NULL) {
1346 			thread = true;
1347 			continue;
1348 		}
1349 		CTR4(KTR_INTR, "%s: exec %p(%p) for %s", __func__,
1350 		    ih->ih_filter, ih->ih_argument == NULL ? frame :
1351 		    ih->ih_argument, ih->ih_name);
1352 		if (ih->ih_argument == NULL)
1353 			ret = ih->ih_filter(frame);
1354 		else
1355 			ret = ih->ih_filter(ih->ih_argument);
1356 		KASSERT(ret == FILTER_STRAY ||
1357 		    ((ret & (FILTER_SCHEDULE_THREAD | FILTER_HANDLED)) != 0 &&
1358 		    (ret & ~(FILTER_SCHEDULE_THREAD | FILTER_HANDLED)) == 0),
1359 		    ("%s: incorrect return value %#x from %s", __func__, ret,
1360 		    ih->ih_name));
1361 		filter = filter || ret == FILTER_HANDLED;
1362 
1363 		/*
1364 		 * Wrapper handler special handling:
1365 		 *
1366 		 * in some particular cases (like pccard and pccbb),
1367 		 * the _real_ device handler is wrapped in a couple of
1368 		 * functions - a filter wrapper and an ithread wrapper.
1369 		 * In this case (and just in this case), the filter wrapper
1370 		 * could ask the system to schedule the ithread and mask
1371 		 * the interrupt source if the wrapped handler is composed
1372 		 * of just an ithread handler.
1373 		 *
1374 		 * TODO: write a generic wrapper to avoid people rolling
1375 		 * their own.
1376 		 */
1377 		if (!thread) {
1378 			if (ret == FILTER_SCHEDULE_THREAD)
1379 				thread = true;
1380 		}
1381 	}
1382 	atomic_add_rel_int(&ie->ie_active[phase], -1);
1383 
1384 	td->td_intr_frame = oldframe;
1385 
1386 	if (thread) {
1387 		if (ie->ie_pre_ithread != NULL)
1388 			ie->ie_pre_ithread(ie->ie_source);
1389 	} else {
1390 		if (ie->ie_post_filter != NULL)
1391 			ie->ie_post_filter(ie->ie_source);
1392 	}
1393 
1394 	/* Schedule the ithread if needed. */
1395 	if (thread) {
1396 		int error __unused;
1397 
1398 		error =  intr_event_schedule_thread(ie);
1399 		KASSERT(error == 0, ("bad stray interrupt"));
1400 	}
1401 	critical_exit();
1402 	td->td_intr_nesting_level--;
1403 #ifdef notyet
1404 	/* The interrupt is not aknowledged by any filter and has no ithread. */
1405 	if (!thread && !filter)
1406 		return (EINVAL);
1407 #endif
1408 	return (0);
1409 }
1410 
1411 #ifdef DDB
1412 /*
1413  * Dump details about an interrupt handler
1414  */
1415 static void
1416 db_dump_intrhand(struct intr_handler *ih)
1417 {
1418 	int comma;
1419 
1420 	db_printf("\t%-10s ", ih->ih_name);
1421 	switch (ih->ih_pri) {
1422 	case PI_REALTIME:
1423 		db_printf("CLK ");
1424 		break;
1425 	case PI_AV:
1426 		db_printf("AV  ");
1427 		break;
1428 	case PI_TTY:
1429 		db_printf("TTY ");
1430 		break;
1431 	case PI_NET:
1432 		db_printf("NET ");
1433 		break;
1434 	case PI_DISK:
1435 		db_printf("DISK");
1436 		break;
1437 	case PI_DULL:
1438 		db_printf("DULL");
1439 		break;
1440 	default:
1441 		if (ih->ih_pri >= PI_SOFT)
1442 			db_printf("SWI ");
1443 		else
1444 			db_printf("%4u", ih->ih_pri);
1445 		break;
1446 	}
1447 	db_printf(" ");
1448 	if (ih->ih_filter != NULL) {
1449 		db_printf("[F]");
1450 		db_printsym((uintptr_t)ih->ih_filter, DB_STGY_PROC);
1451 	}
1452 	if (ih->ih_handler != NULL) {
1453 		if (ih->ih_filter != NULL)
1454 			db_printf(",");
1455 		db_printf("[H]");
1456 		db_printsym((uintptr_t)ih->ih_handler, DB_STGY_PROC);
1457 	}
1458 	db_printf("(%p)", ih->ih_argument);
1459 	if (ih->ih_need ||
1460 	    (ih->ih_flags & (IH_EXCLUSIVE | IH_ENTROPY | IH_DEAD |
1461 	    IH_MPSAFE)) != 0) {
1462 		db_printf(" {");
1463 		comma = 0;
1464 		if (ih->ih_flags & IH_EXCLUSIVE) {
1465 			if (comma)
1466 				db_printf(", ");
1467 			db_printf("EXCL");
1468 			comma = 1;
1469 		}
1470 		if (ih->ih_flags & IH_ENTROPY) {
1471 			if (comma)
1472 				db_printf(", ");
1473 			db_printf("ENTROPY");
1474 			comma = 1;
1475 		}
1476 		if (ih->ih_flags & IH_DEAD) {
1477 			if (comma)
1478 				db_printf(", ");
1479 			db_printf("DEAD");
1480 			comma = 1;
1481 		}
1482 		if (ih->ih_flags & IH_MPSAFE) {
1483 			if (comma)
1484 				db_printf(", ");
1485 			db_printf("MPSAFE");
1486 			comma = 1;
1487 		}
1488 		if (ih->ih_need) {
1489 			if (comma)
1490 				db_printf(", ");
1491 			db_printf("NEED");
1492 		}
1493 		db_printf("}");
1494 	}
1495 	db_printf("\n");
1496 }
1497 
1498 /*
1499  * Dump details about a event.
1500  */
1501 void
1502 db_dump_intr_event(struct intr_event *ie, int handlers)
1503 {
1504 	struct intr_handler *ih;
1505 	struct intr_thread *it;
1506 	int comma;
1507 
1508 	db_printf("%s ", ie->ie_fullname);
1509 	it = ie->ie_thread;
1510 	if (it != NULL)
1511 		db_printf("(pid %d)", it->it_thread->td_proc->p_pid);
1512 	else
1513 		db_printf("(no thread)");
1514 	if ((ie->ie_flags & (IE_SOFT | IE_ADDING_THREAD)) != 0 ||
1515 	    (it != NULL && it->it_need)) {
1516 		db_printf(" {");
1517 		comma = 0;
1518 		if (ie->ie_flags & IE_SOFT) {
1519 			db_printf("SOFT");
1520 			comma = 1;
1521 		}
1522 		if (ie->ie_flags & IE_ADDING_THREAD) {
1523 			if (comma)
1524 				db_printf(", ");
1525 			db_printf("ADDING_THREAD");
1526 			comma = 1;
1527 		}
1528 		if (it != NULL && it->it_need) {
1529 			if (comma)
1530 				db_printf(", ");
1531 			db_printf("NEED");
1532 		}
1533 		db_printf("}");
1534 	}
1535 	db_printf("\n");
1536 
1537 	if (handlers)
1538 		CK_SLIST_FOREACH(ih, &ie->ie_handlers, ih_next)
1539 		    db_dump_intrhand(ih);
1540 }
1541 
1542 /*
1543  * Dump data about interrupt handlers
1544  */
1545 DB_SHOW_COMMAND(intr, db_show_intr)
1546 {
1547 	struct intr_event *ie;
1548 	int all, verbose;
1549 
1550 	verbose = strchr(modif, 'v') != NULL;
1551 	all = strchr(modif, 'a') != NULL;
1552 	TAILQ_FOREACH(ie, &event_list, ie_list) {
1553 		if (!all && CK_SLIST_EMPTY(&ie->ie_handlers))
1554 			continue;
1555 		db_dump_intr_event(ie, verbose);
1556 		if (db_pager_quit)
1557 			break;
1558 	}
1559 }
1560 #endif /* DDB */
1561 
1562 /*
1563  * Start standard software interrupt threads
1564  */
1565 static void
1566 start_softintr(void *dummy)
1567 {
1568 
1569 	if (swi_add(NULL, "vm", swi_vm, NULL, SWI_VM, INTR_MPSAFE, &vm_ih))
1570 		panic("died while creating vm swi ithread");
1571 }
1572 SYSINIT(start_softintr, SI_SUB_SOFTINTR, SI_ORDER_FIRST, start_softintr,
1573     NULL);
1574 
1575 /*
1576  * Sysctls used by systat and others: hw.intrnames and hw.intrcnt.
1577  * The data for this machine dependent, and the declarations are in machine
1578  * dependent code.  The layout of intrnames and intrcnt however is machine
1579  * independent.
1580  *
1581  * We do not know the length of intrcnt and intrnames at compile time, so
1582  * calculate things at run time.
1583  */
1584 static int
1585 sysctl_intrnames(SYSCTL_HANDLER_ARGS)
1586 {
1587 	return (sysctl_handle_opaque(oidp, intrnames, sintrnames, req));
1588 }
1589 
1590 SYSCTL_PROC(_hw, OID_AUTO, intrnames, CTLTYPE_OPAQUE | CTLFLAG_RD,
1591     NULL, 0, sysctl_intrnames, "", "Interrupt Names");
1592 
1593 static int
1594 sysctl_intrcnt(SYSCTL_HANDLER_ARGS)
1595 {
1596 #ifdef SCTL_MASK32
1597 	uint32_t *intrcnt32;
1598 	unsigned i;
1599 	int error;
1600 
1601 	if (req->flags & SCTL_MASK32) {
1602 		if (!req->oldptr)
1603 			return (sysctl_handle_opaque(oidp, NULL, sintrcnt / 2, req));
1604 		intrcnt32 = malloc(sintrcnt / 2, M_TEMP, M_NOWAIT);
1605 		if (intrcnt32 == NULL)
1606 			return (ENOMEM);
1607 		for (i = 0; i < sintrcnt / sizeof (u_long); i++)
1608 			intrcnt32[i] = intrcnt[i];
1609 		error = sysctl_handle_opaque(oidp, intrcnt32, sintrcnt / 2, req);
1610 		free(intrcnt32, M_TEMP);
1611 		return (error);
1612 	}
1613 #endif
1614 	return (sysctl_handle_opaque(oidp, intrcnt, sintrcnt, req));
1615 }
1616 
1617 SYSCTL_PROC(_hw, OID_AUTO, intrcnt, CTLTYPE_OPAQUE | CTLFLAG_RD,
1618     NULL, 0, sysctl_intrcnt, "", "Interrupt Counts");
1619 
1620 #ifdef DDB
1621 /*
1622  * DDB command to dump the interrupt statistics.
1623  */
1624 DB_SHOW_COMMAND(intrcnt, db_show_intrcnt)
1625 {
1626 	u_long *i;
1627 	char *cp;
1628 	u_int j;
1629 
1630 	cp = intrnames;
1631 	j = 0;
1632 	for (i = intrcnt; j < (sintrcnt / sizeof(u_long)) && !db_pager_quit;
1633 	    i++, j++) {
1634 		if (*cp == '\0')
1635 			break;
1636 		if (*i != 0)
1637 			db_printf("%s\t%lu\n", cp, *i);
1638 		cp += strlen(cp) + 1;
1639 	}
1640 }
1641 #endif
1642