xref: /freebsd/sys/kern/kern_intr.c (revision 732a02b4e77866604a120a275c082bb6221bd2ff)
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, flags;
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 	flags = 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 		flags |= ih->ih_flags;
211 	}
212 	ie->ie_hflags = flags;
213 
214 	/*
215 	 * If there is only one handler and its name is too long, just copy in
216 	 * as much of the end of the name (includes the unit number) as will
217 	 * fit.  Otherwise, we have multiple handlers and not all of the names
218 	 * will fit.  Add +'s to indicate missing names.  If we run out of room
219 	 * and still have +'s to add, change the last character from a + to a *.
220 	 */
221 	if (missed == 1 && space == 1) {
222 		ih = CK_SLIST_FIRST(&ie->ie_handlers);
223 		missed = strlen(ie->ie_fullname) + strlen(ih->ih_name) + 2 -
224 		    sizeof(ie->ie_fullname);
225 		strcat(ie->ie_fullname, (missed == 0) ? " " : "-");
226 		strcat(ie->ie_fullname, &ih->ih_name[missed]);
227 		missed = 0;
228 	}
229 	last = &ie->ie_fullname[sizeof(ie->ie_fullname) - 2];
230 	while (missed-- > 0) {
231 		if (strlen(ie->ie_fullname) + 1 == sizeof(ie->ie_fullname)) {
232 			if (*last == '+') {
233 				*last = '*';
234 				break;
235 			} else
236 				*last = '+';
237 		} else if (space) {
238 			strcat(ie->ie_fullname, " +");
239 			space = 0;
240 		} else
241 			strcat(ie->ie_fullname, "+");
242 	}
243 
244 	/*
245 	 * If this event has an ithread, update it's priority and
246 	 * name.
247 	 */
248 	if (ie->ie_thread != NULL)
249 		ithread_update(ie->ie_thread);
250 	CTR2(KTR_INTR, "%s: updated %s", __func__, ie->ie_fullname);
251 }
252 
253 int
254 intr_event_create(struct intr_event **event, void *source, int flags, int irq,
255     void (*pre_ithread)(void *), void (*post_ithread)(void *),
256     void (*post_filter)(void *), int (*assign_cpu)(void *, int),
257     const char *fmt, ...)
258 {
259 	struct intr_event *ie;
260 	va_list ap;
261 
262 	/* The only valid flag during creation is IE_SOFT. */
263 	if ((flags & ~IE_SOFT) != 0)
264 		return (EINVAL);
265 	ie = malloc(sizeof(struct intr_event), M_ITHREAD, M_WAITOK | M_ZERO);
266 	ie->ie_source = source;
267 	ie->ie_pre_ithread = pre_ithread;
268 	ie->ie_post_ithread = post_ithread;
269 	ie->ie_post_filter = post_filter;
270 	ie->ie_assign_cpu = assign_cpu;
271 	ie->ie_flags = flags;
272 	ie->ie_irq = irq;
273 	ie->ie_cpu = NOCPU;
274 	CK_SLIST_INIT(&ie->ie_handlers);
275 	mtx_init(&ie->ie_lock, "intr event", NULL, MTX_DEF);
276 
277 	va_start(ap, fmt);
278 	vsnprintf(ie->ie_name, sizeof(ie->ie_name), fmt, ap);
279 	va_end(ap);
280 	strlcpy(ie->ie_fullname, ie->ie_name, sizeof(ie->ie_fullname));
281 	mtx_lock(&event_lock);
282 	TAILQ_INSERT_TAIL(&event_list, ie, ie_list);
283 	mtx_unlock(&event_lock);
284 	if (event != NULL)
285 		*event = ie;
286 	CTR2(KTR_INTR, "%s: created %s", __func__, ie->ie_name);
287 	return (0);
288 }
289 
290 /*
291  * Bind an interrupt event to the specified CPU.  Note that not all
292  * platforms support binding an interrupt to a CPU.  For those
293  * platforms this request will fail.  Using a cpu id of NOCPU unbinds
294  * the interrupt event.
295  */
296 static int
297 _intr_event_bind(struct intr_event *ie, int cpu, bool bindirq, bool bindithread)
298 {
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 	if (bindithread) {
318 		mtx_lock(&ie->ie_lock);
319 		if (ie->ie_thread != NULL) {
320 			id = ie->ie_thread->it_thread->td_tid;
321 			mtx_unlock(&ie->ie_lock);
322 			error = cpuset_setithread(id, cpu);
323 			if (error)
324 				return (error);
325 		} else
326 			mtx_unlock(&ie->ie_lock);
327 	}
328 	if (bindirq)
329 		error = ie->ie_assign_cpu(ie->ie_source, cpu);
330 	if (error) {
331 		if (bindithread) {
332 			mtx_lock(&ie->ie_lock);
333 			if (ie->ie_thread != NULL) {
334 				cpu = ie->ie_cpu;
335 				id = ie->ie_thread->it_thread->td_tid;
336 				mtx_unlock(&ie->ie_lock);
337 				(void)cpuset_setithread(id, cpu);
338 			} else
339 				mtx_unlock(&ie->ie_lock);
340 		}
341 		return (error);
342 	}
343 
344 	if (bindirq) {
345 		mtx_lock(&ie->ie_lock);
346 		ie->ie_cpu = cpu;
347 		mtx_unlock(&ie->ie_lock);
348 	}
349 
350 	return (error);
351 }
352 
353 /*
354  * Bind an interrupt event to the specified CPU.  For supported platforms, any
355  * associated ithreads as well as the primary interrupt context will be bound
356  * to the specificed CPU.
357  */
358 int
359 intr_event_bind(struct intr_event *ie, int cpu)
360 {
361 
362 	return (_intr_event_bind(ie, cpu, true, true));
363 }
364 
365 /*
366  * Bind an interrupt event to the specified CPU, but do not bind associated
367  * ithreads.
368  */
369 int
370 intr_event_bind_irqonly(struct intr_event *ie, int cpu)
371 {
372 
373 	return (_intr_event_bind(ie, cpu, true, false));
374 }
375 
376 /*
377  * Bind an interrupt event's ithread to the specified CPU.
378  */
379 int
380 intr_event_bind_ithread(struct intr_event *ie, int cpu)
381 {
382 
383 	return (_intr_event_bind(ie, cpu, false, true));
384 }
385 
386 /*
387  * Bind an interrupt event's ithread to the specified cpuset.
388  */
389 int
390 intr_event_bind_ithread_cpuset(struct intr_event *ie, cpuset_t *cs)
391 {
392 	lwpid_t id;
393 
394 	mtx_lock(&ie->ie_lock);
395 	if (ie->ie_thread != NULL) {
396 		id = ie->ie_thread->it_thread->td_tid;
397 		mtx_unlock(&ie->ie_lock);
398 		return (cpuset_setthread(id, cs));
399 	} else {
400 		mtx_unlock(&ie->ie_lock);
401 	}
402 	return (ENODEV);
403 }
404 
405 static struct intr_event *
406 intr_lookup(int irq)
407 {
408 	struct intr_event *ie;
409 
410 	mtx_lock(&event_lock);
411 	TAILQ_FOREACH(ie, &event_list, ie_list)
412 		if (ie->ie_irq == irq &&
413 		    (ie->ie_flags & IE_SOFT) == 0 &&
414 		    CK_SLIST_FIRST(&ie->ie_handlers) != NULL)
415 			break;
416 	mtx_unlock(&event_lock);
417 	return (ie);
418 }
419 
420 int
421 intr_setaffinity(int irq, int mode, void *m)
422 {
423 	struct intr_event *ie;
424 	cpuset_t *mask;
425 	int cpu, n;
426 
427 	mask = m;
428 	cpu = NOCPU;
429 	/*
430 	 * If we're setting all cpus we can unbind.  Otherwise make sure
431 	 * only one cpu is in the set.
432 	 */
433 	if (CPU_CMP(cpuset_root, mask)) {
434 		for (n = 0; n < CPU_SETSIZE; n++) {
435 			if (!CPU_ISSET(n, mask))
436 				continue;
437 			if (cpu != NOCPU)
438 				return (EINVAL);
439 			cpu = n;
440 		}
441 	}
442 	ie = intr_lookup(irq);
443 	if (ie == NULL)
444 		return (ESRCH);
445 	switch (mode) {
446 	case CPU_WHICH_IRQ:
447 		return (intr_event_bind(ie, cpu));
448 	case CPU_WHICH_INTRHANDLER:
449 		return (intr_event_bind_irqonly(ie, cpu));
450 	case CPU_WHICH_ITHREAD:
451 		return (intr_event_bind_ithread(ie, cpu));
452 	default:
453 		return (EINVAL);
454 	}
455 }
456 
457 int
458 intr_getaffinity(int irq, int mode, void *m)
459 {
460 	struct intr_event *ie;
461 	struct thread *td;
462 	struct proc *p;
463 	cpuset_t *mask;
464 	lwpid_t id;
465 	int error;
466 
467 	mask = m;
468 	ie = intr_lookup(irq);
469 	if (ie == NULL)
470 		return (ESRCH);
471 
472 	error = 0;
473 	CPU_ZERO(mask);
474 	switch (mode) {
475 	case CPU_WHICH_IRQ:
476 	case CPU_WHICH_INTRHANDLER:
477 		mtx_lock(&ie->ie_lock);
478 		if (ie->ie_cpu == NOCPU)
479 			CPU_COPY(cpuset_root, mask);
480 		else
481 			CPU_SET(ie->ie_cpu, mask);
482 		mtx_unlock(&ie->ie_lock);
483 		break;
484 	case CPU_WHICH_ITHREAD:
485 		mtx_lock(&ie->ie_lock);
486 		if (ie->ie_thread == NULL) {
487 			mtx_unlock(&ie->ie_lock);
488 			CPU_COPY(cpuset_root, mask);
489 		} else {
490 			id = ie->ie_thread->it_thread->td_tid;
491 			mtx_unlock(&ie->ie_lock);
492 			error = cpuset_which(CPU_WHICH_TID, id, &p, &td, NULL);
493 			if (error != 0)
494 				return (error);
495 			CPU_COPY(&td->td_cpuset->cs_mask, mask);
496 			PROC_UNLOCK(p);
497 		}
498 	default:
499 		return (EINVAL);
500 	}
501 	return (0);
502 }
503 
504 int
505 intr_event_destroy(struct intr_event *ie)
506 {
507 
508 	mtx_lock(&event_lock);
509 	mtx_lock(&ie->ie_lock);
510 	if (!CK_SLIST_EMPTY(&ie->ie_handlers)) {
511 		mtx_unlock(&ie->ie_lock);
512 		mtx_unlock(&event_lock);
513 		return (EBUSY);
514 	}
515 	TAILQ_REMOVE(&event_list, ie, ie_list);
516 #ifndef notyet
517 	if (ie->ie_thread != NULL) {
518 		ithread_destroy(ie->ie_thread);
519 		ie->ie_thread = NULL;
520 	}
521 #endif
522 	mtx_unlock(&ie->ie_lock);
523 	mtx_unlock(&event_lock);
524 	mtx_destroy(&ie->ie_lock);
525 	free(ie, M_ITHREAD);
526 	return (0);
527 }
528 
529 static struct intr_thread *
530 ithread_create(const char *name)
531 {
532 	struct intr_thread *ithd;
533 	struct thread *td;
534 	int error;
535 
536 	ithd = malloc(sizeof(struct intr_thread), M_ITHREAD, M_WAITOK | M_ZERO);
537 
538 	error = kproc_kthread_add(ithread_loop, ithd, &intrproc,
539 		    &td, RFSTOPPED | RFHIGHPID,
540 		    0, "intr", "%s", name);
541 	if (error)
542 		panic("kproc_create() failed with %d", error);
543 	thread_lock(td);
544 	sched_class(td, PRI_ITHD);
545 	TD_SET_IWAIT(td);
546 	thread_unlock(td);
547 	td->td_pflags |= TDP_ITHREAD;
548 	ithd->it_thread = td;
549 	CTR2(KTR_INTR, "%s: created %s", __func__, name);
550 	return (ithd);
551 }
552 
553 static void
554 ithread_destroy(struct intr_thread *ithread)
555 {
556 	struct thread *td;
557 
558 	CTR2(KTR_INTR, "%s: killing %s", __func__, ithread->it_event->ie_name);
559 	td = ithread->it_thread;
560 	thread_lock(td);
561 	ithread->it_flags |= IT_DEAD;
562 	if (TD_AWAITING_INTR(td)) {
563 		TD_CLR_IWAIT(td);
564 		sched_add(td, SRQ_INTR);
565 	} else
566 		thread_unlock(td);
567 }
568 
569 int
570 intr_event_add_handler(struct intr_event *ie, const char *name,
571     driver_filter_t filter, driver_intr_t handler, void *arg, u_char pri,
572     enum intr_type flags, void **cookiep)
573 {
574 	struct intr_handler *ih, *temp_ih;
575 	struct intr_handler **prevptr;
576 	struct intr_thread *it;
577 
578 	if (ie == NULL || name == NULL || (handler == NULL && filter == NULL))
579 		return (EINVAL);
580 
581 	/* Allocate and populate an interrupt handler structure. */
582 	ih = malloc(sizeof(struct intr_handler), M_ITHREAD, M_WAITOK | M_ZERO);
583 	ih->ih_filter = filter;
584 	ih->ih_handler = handler;
585 	ih->ih_argument = arg;
586 	strlcpy(ih->ih_name, name, sizeof(ih->ih_name));
587 	ih->ih_event = ie;
588 	ih->ih_pri = pri;
589 	if (flags & INTR_EXCL)
590 		ih->ih_flags = IH_EXCLUSIVE;
591 	if (flags & INTR_MPSAFE)
592 		ih->ih_flags |= IH_MPSAFE;
593 	if (flags & INTR_ENTROPY)
594 		ih->ih_flags |= IH_ENTROPY;
595 	if (flags & INTR_TYPE_NET)
596 		ih->ih_flags |= IH_NET;
597 
598 	/* We can only have one exclusive handler in a event. */
599 	mtx_lock(&ie->ie_lock);
600 	if (!CK_SLIST_EMPTY(&ie->ie_handlers)) {
601 		if ((flags & INTR_EXCL) ||
602 		    (CK_SLIST_FIRST(&ie->ie_handlers)->ih_flags & IH_EXCLUSIVE)) {
603 			mtx_unlock(&ie->ie_lock);
604 			free(ih, M_ITHREAD);
605 			return (EINVAL);
606 		}
607 	}
608 
609 	/* Create a thread if we need one. */
610 	while (ie->ie_thread == NULL && handler != NULL) {
611 		if (ie->ie_flags & IE_ADDING_THREAD)
612 			msleep(ie, &ie->ie_lock, 0, "ithread", 0);
613 		else {
614 			ie->ie_flags |= IE_ADDING_THREAD;
615 			mtx_unlock(&ie->ie_lock);
616 			it = ithread_create("intr: newborn");
617 			mtx_lock(&ie->ie_lock);
618 			ie->ie_flags &= ~IE_ADDING_THREAD;
619 			ie->ie_thread = it;
620 			it->it_event = ie;
621 			ithread_update(it);
622 			wakeup(ie);
623 		}
624 	}
625 
626 	/* Add the new handler to the event in priority order. */
627 	CK_SLIST_FOREACH_PREVPTR(temp_ih, prevptr, &ie->ie_handlers, ih_next) {
628 		if (temp_ih->ih_pri > ih->ih_pri)
629 			break;
630 	}
631 	CK_SLIST_INSERT_PREVPTR(prevptr, temp_ih, ih, ih_next);
632 
633 	intr_event_update(ie);
634 
635 	CTR3(KTR_INTR, "%s: added %s to %s", __func__, ih->ih_name,
636 	    ie->ie_name);
637 	mtx_unlock(&ie->ie_lock);
638 
639 	if (cookiep != NULL)
640 		*cookiep = ih;
641 	return (0);
642 }
643 
644 /*
645  * Append a description preceded by a ':' to the name of the specified
646  * interrupt handler.
647  */
648 int
649 intr_event_describe_handler(struct intr_event *ie, void *cookie,
650     const char *descr)
651 {
652 	struct intr_handler *ih;
653 	size_t space;
654 	char *start;
655 
656 	mtx_lock(&ie->ie_lock);
657 #ifdef INVARIANTS
658 	CK_SLIST_FOREACH(ih, &ie->ie_handlers, ih_next) {
659 		if (ih == cookie)
660 			break;
661 	}
662 	if (ih == NULL) {
663 		mtx_unlock(&ie->ie_lock);
664 		panic("handler %p not found in interrupt event %p", cookie, ie);
665 	}
666 #endif
667 	ih = cookie;
668 
669 	/*
670 	 * Look for an existing description by checking for an
671 	 * existing ":".  This assumes device names do not include
672 	 * colons.  If one is found, prepare to insert the new
673 	 * description at that point.  If one is not found, find the
674 	 * end of the name to use as the insertion point.
675 	 */
676 	start = strchr(ih->ih_name, ':');
677 	if (start == NULL)
678 		start = strchr(ih->ih_name, 0);
679 
680 	/*
681 	 * See if there is enough remaining room in the string for the
682 	 * description + ":".  The "- 1" leaves room for the trailing
683 	 * '\0'.  The "+ 1" accounts for the colon.
684 	 */
685 	space = sizeof(ih->ih_name) - (start - ih->ih_name) - 1;
686 	if (strlen(descr) + 1 > space) {
687 		mtx_unlock(&ie->ie_lock);
688 		return (ENOSPC);
689 	}
690 
691 	/* Append a colon followed by the description. */
692 	*start = ':';
693 	strcpy(start + 1, descr);
694 	intr_event_update(ie);
695 	mtx_unlock(&ie->ie_lock);
696 	return (0);
697 }
698 
699 /*
700  * Return the ie_source field from the intr_event an intr_handler is
701  * associated with.
702  */
703 void *
704 intr_handler_source(void *cookie)
705 {
706 	struct intr_handler *ih;
707 	struct intr_event *ie;
708 
709 	ih = (struct intr_handler *)cookie;
710 	if (ih == NULL)
711 		return (NULL);
712 	ie = ih->ih_event;
713 	KASSERT(ie != NULL,
714 	    ("interrupt handler \"%s\" has a NULL interrupt event",
715 	    ih->ih_name));
716 	return (ie->ie_source);
717 }
718 
719 /*
720  * If intr_event_handle() is running in the ISR context at the time of the call,
721  * then wait for it to complete.
722  */
723 static void
724 intr_event_barrier(struct intr_event *ie)
725 {
726 	int phase;
727 
728 	mtx_assert(&ie->ie_lock, MA_OWNED);
729 	phase = ie->ie_phase;
730 
731 	/*
732 	 * Switch phase to direct future interrupts to the other active counter.
733 	 * Make sure that any preceding stores are visible before the switch.
734 	 */
735 	KASSERT(ie->ie_active[!phase] == 0, ("idle phase has activity"));
736 	atomic_store_rel_int(&ie->ie_phase, !phase);
737 
738 	/*
739 	 * This code cooperates with wait-free iteration of ie_handlers
740 	 * in intr_event_handle.
741 	 * Make sure that the removal and the phase update are not reordered
742 	 * with the active count check.
743 	 * Note that no combination of acquire and release fences can provide
744 	 * that guarantee as Store->Load sequences can always be reordered.
745 	 */
746 	atomic_thread_fence_seq_cst();
747 
748 	/*
749 	 * Now wait on the inactive phase.
750 	 * The acquire fence is needed so that that all post-barrier accesses
751 	 * are after the check.
752 	 */
753 	while (ie->ie_active[phase] > 0)
754 		cpu_spinwait();
755 	atomic_thread_fence_acq();
756 }
757 
758 static void
759 intr_handler_barrier(struct intr_handler *handler)
760 {
761 	struct intr_event *ie;
762 
763 	ie = handler->ih_event;
764 	mtx_assert(&ie->ie_lock, MA_OWNED);
765 	KASSERT((handler->ih_flags & IH_DEAD) == 0,
766 	    ("update for a removed handler"));
767 
768 	if (ie->ie_thread == NULL) {
769 		intr_event_barrier(ie);
770 		return;
771 	}
772 	if ((handler->ih_flags & IH_CHANGED) == 0) {
773 		handler->ih_flags |= IH_CHANGED;
774 		intr_event_schedule_thread(ie);
775 	}
776 	while ((handler->ih_flags & IH_CHANGED) != 0)
777 		msleep(handler, &ie->ie_lock, 0, "ih_barr", 0);
778 }
779 
780 /*
781  * Sleep until an ithread finishes executing an interrupt handler.
782  *
783  * XXX Doesn't currently handle interrupt filters or fast interrupt
784  * handlers.  This is intended for compatibility with linux drivers
785  * only.  Do not use in BSD code.
786  */
787 void
788 _intr_drain(int irq)
789 {
790 	struct intr_event *ie;
791 	struct intr_thread *ithd;
792 	struct thread *td;
793 
794 	ie = intr_lookup(irq);
795 	if (ie == NULL)
796 		return;
797 	if (ie->ie_thread == NULL)
798 		return;
799 	ithd = ie->ie_thread;
800 	td = ithd->it_thread;
801 	/*
802 	 * We set the flag and wait for it to be cleared to avoid
803 	 * long delays with potentially busy interrupt handlers
804 	 * were we to only sample TD_AWAITING_INTR() every tick.
805 	 */
806 	thread_lock(td);
807 	if (!TD_AWAITING_INTR(td)) {
808 		ithd->it_flags |= IT_WAIT;
809 		while (ithd->it_flags & IT_WAIT) {
810 			thread_unlock(td);
811 			pause("idrain", 1);
812 			thread_lock(td);
813 		}
814 	}
815 	thread_unlock(td);
816 	return;
817 }
818 
819 int
820 intr_event_remove_handler(void *cookie)
821 {
822 	struct intr_handler *handler = (struct intr_handler *)cookie;
823 	struct intr_event *ie;
824 	struct intr_handler *ih;
825 	struct intr_handler **prevptr;
826 #ifdef notyet
827 	int dead;
828 #endif
829 
830 	if (handler == NULL)
831 		return (EINVAL);
832 	ie = handler->ih_event;
833 	KASSERT(ie != NULL,
834 	    ("interrupt handler \"%s\" has a NULL interrupt event",
835 	    handler->ih_name));
836 
837 	mtx_lock(&ie->ie_lock);
838 	CTR3(KTR_INTR, "%s: removing %s from %s", __func__, handler->ih_name,
839 	    ie->ie_name);
840 	CK_SLIST_FOREACH_PREVPTR(ih, prevptr, &ie->ie_handlers, ih_next) {
841 		if (ih == handler)
842 			break;
843 	}
844 	if (ih == NULL) {
845 		panic("interrupt handler \"%s\" not found in "
846 		    "interrupt event \"%s\"", handler->ih_name, ie->ie_name);
847 	}
848 
849 	/*
850 	 * If there is no ithread, then directly remove the handler.  Note that
851 	 * intr_event_handle() iterates ie_handlers in a lock-less fashion, so
852 	 * care needs to be taken to keep ie_handlers consistent and to free
853 	 * the removed handler only when ie_handlers is quiescent.
854 	 */
855 	if (ie->ie_thread == NULL) {
856 		CK_SLIST_REMOVE_PREVPTR(prevptr, ih, ih_next);
857 		intr_event_barrier(ie);
858 		intr_event_update(ie);
859 		mtx_unlock(&ie->ie_lock);
860 		free(handler, M_ITHREAD);
861 		return (0);
862 	}
863 
864 	/*
865 	 * Let the interrupt thread do the job.
866 	 * The interrupt source is disabled when the interrupt thread is
867 	 * running, so it does not have to worry about interaction with
868 	 * intr_event_handle().
869 	 */
870 	KASSERT((handler->ih_flags & IH_DEAD) == 0,
871 	    ("duplicate handle remove"));
872 	handler->ih_flags |= IH_DEAD;
873 	intr_event_schedule_thread(ie);
874 	while (handler->ih_flags & IH_DEAD)
875 		msleep(handler, &ie->ie_lock, 0, "iev_rmh", 0);
876 	intr_event_update(ie);
877 
878 #ifdef notyet
879 	/*
880 	 * XXX: This could be bad in the case of ppbus(8).  Also, I think
881 	 * this could lead to races of stale data when servicing an
882 	 * interrupt.
883 	 */
884 	dead = 1;
885 	CK_SLIST_FOREACH(ih, &ie->ie_handlers, ih_next) {
886 		if (ih->ih_handler != NULL) {
887 			dead = 0;
888 			break;
889 		}
890 	}
891 	if (dead) {
892 		ithread_destroy(ie->ie_thread);
893 		ie->ie_thread = NULL;
894 	}
895 #endif
896 	mtx_unlock(&ie->ie_lock);
897 	free(handler, M_ITHREAD);
898 	return (0);
899 }
900 
901 int
902 intr_event_suspend_handler(void *cookie)
903 {
904 	struct intr_handler *handler = (struct intr_handler *)cookie;
905 	struct intr_event *ie;
906 
907 	if (handler == NULL)
908 		return (EINVAL);
909 	ie = handler->ih_event;
910 	KASSERT(ie != NULL,
911 	    ("interrupt handler \"%s\" has a NULL interrupt event",
912 	    handler->ih_name));
913 	mtx_lock(&ie->ie_lock);
914 	handler->ih_flags |= IH_SUSP;
915 	intr_handler_barrier(handler);
916 	mtx_unlock(&ie->ie_lock);
917 	return (0);
918 }
919 
920 int
921 intr_event_resume_handler(void *cookie)
922 {
923 	struct intr_handler *handler = (struct intr_handler *)cookie;
924 	struct intr_event *ie;
925 
926 	if (handler == NULL)
927 		return (EINVAL);
928 	ie = handler->ih_event;
929 	KASSERT(ie != NULL,
930 	    ("interrupt handler \"%s\" has a NULL interrupt event",
931 	    handler->ih_name));
932 
933 	/*
934 	 * intr_handler_barrier() acts not only as a barrier,
935 	 * it also allows to check for any pending interrupts.
936 	 */
937 	mtx_lock(&ie->ie_lock);
938 	handler->ih_flags &= ~IH_SUSP;
939 	intr_handler_barrier(handler);
940 	mtx_unlock(&ie->ie_lock);
941 	return (0);
942 }
943 
944 static int
945 intr_event_schedule_thread(struct intr_event *ie)
946 {
947 	struct intr_entropy entropy;
948 	struct intr_thread *it;
949 	struct thread *td;
950 	struct thread *ctd;
951 
952 	/*
953 	 * If no ithread or no handlers, then we have a stray interrupt.
954 	 */
955 	if (ie == NULL || CK_SLIST_EMPTY(&ie->ie_handlers) ||
956 	    ie->ie_thread == NULL)
957 		return (EINVAL);
958 
959 	ctd = curthread;
960 	it = ie->ie_thread;
961 	td = it->it_thread;
962 
963 	/*
964 	 * If any of the handlers for this ithread claim to be good
965 	 * sources of entropy, then gather some.
966 	 */
967 	if (ie->ie_hflags & IH_ENTROPY) {
968 		entropy.event = (uintptr_t)ie;
969 		entropy.td = ctd;
970 		random_harvest_queue(&entropy, sizeof(entropy), RANDOM_INTERRUPT);
971 	}
972 
973 	KASSERT(td->td_proc != NULL, ("ithread %s has no process", ie->ie_name));
974 
975 	/*
976 	 * Set it_need to tell the thread to keep running if it is already
977 	 * running.  Then, lock the thread and see if we actually need to
978 	 * put it on the runqueue.
979 	 *
980 	 * Use store_rel to arrange that the store to ih_need in
981 	 * swi_sched() is before the store to it_need and prepare for
982 	 * transfer of this order to loads in the ithread.
983 	 */
984 	atomic_store_rel_int(&it->it_need, 1);
985 	thread_lock(td);
986 	if (TD_AWAITING_INTR(td)) {
987 		CTR3(KTR_INTR, "%s: schedule pid %d (%s)", __func__, td->td_proc->p_pid,
988 		    td->td_name);
989 		TD_CLR_IWAIT(td);
990 		sched_add(td, SRQ_INTR);
991 	} else {
992 		CTR5(KTR_INTR, "%s: pid %d (%s): it_need %d, state %d",
993 		    __func__, td->td_proc->p_pid, td->td_name, it->it_need, td->td_state);
994 		thread_unlock(td);
995 	}
996 
997 	return (0);
998 }
999 
1000 /*
1001  * Allow interrupt event binding for software interrupt handlers -- a no-op,
1002  * since interrupts are generated in software rather than being directed by
1003  * a PIC.
1004  */
1005 static int
1006 swi_assign_cpu(void *arg, int cpu)
1007 {
1008 
1009 	return (0);
1010 }
1011 
1012 /*
1013  * Add a software interrupt handler to a specified event.  If a given event
1014  * is not specified, then a new event is created.
1015  */
1016 int
1017 swi_add(struct intr_event **eventp, const char *name, driver_intr_t handler,
1018 	    void *arg, int pri, enum intr_type flags, void **cookiep)
1019 {
1020 	struct intr_event *ie;
1021 	int error;
1022 
1023 	if (flags & INTR_ENTROPY)
1024 		return (EINVAL);
1025 
1026 	ie = (eventp != NULL) ? *eventp : NULL;
1027 
1028 	if (ie != NULL) {
1029 		if (!(ie->ie_flags & IE_SOFT))
1030 			return (EINVAL);
1031 	} else {
1032 		error = intr_event_create(&ie, NULL, IE_SOFT, 0,
1033 		    NULL, NULL, NULL, swi_assign_cpu, "swi%d:", pri);
1034 		if (error)
1035 			return (error);
1036 		if (eventp != NULL)
1037 			*eventp = ie;
1038 	}
1039 	error = intr_event_add_handler(ie, name, NULL, handler, arg,
1040 	    PI_SWI(pri), flags, cookiep);
1041 	return (error);
1042 }
1043 
1044 /*
1045  * Schedule a software interrupt thread.
1046  */
1047 void
1048 swi_sched(void *cookie, int flags)
1049 {
1050 	struct intr_handler *ih = (struct intr_handler *)cookie;
1051 	struct intr_event *ie = ih->ih_event;
1052 	struct intr_entropy entropy;
1053 	int error __unused;
1054 
1055 	CTR3(KTR_INTR, "swi_sched: %s %s need=%d", ie->ie_name, ih->ih_name,
1056 	    ih->ih_need);
1057 
1058 	entropy.event = (uintptr_t)ih;
1059 	entropy.td = curthread;
1060 	random_harvest_queue(&entropy, sizeof(entropy), RANDOM_SWI);
1061 
1062 	/*
1063 	 * Set ih_need for this handler so that if the ithread is already
1064 	 * running it will execute this handler on the next pass.  Otherwise,
1065 	 * it will execute it the next time it runs.
1066 	 */
1067 	ih->ih_need = 1;
1068 
1069 	if (!(flags & SWI_DELAY)) {
1070 		VM_CNT_INC(v_soft);
1071 		error = intr_event_schedule_thread(ie);
1072 		KASSERT(error == 0, ("stray software interrupt"));
1073 	}
1074 }
1075 
1076 /*
1077  * Remove a software interrupt handler.  Currently this code does not
1078  * remove the associated interrupt event if it becomes empty.  Calling code
1079  * may do so manually via intr_event_destroy(), but that's not really
1080  * an optimal interface.
1081  */
1082 int
1083 swi_remove(void *cookie)
1084 {
1085 
1086 	return (intr_event_remove_handler(cookie));
1087 }
1088 
1089 static void
1090 intr_event_execute_handlers(struct proc *p, struct intr_event *ie)
1091 {
1092 	struct intr_handler *ih, *ihn, *ihp;
1093 
1094 	ihp = NULL;
1095 	CK_SLIST_FOREACH_SAFE(ih, &ie->ie_handlers, ih_next, ihn) {
1096 		/*
1097 		 * If this handler is marked for death, remove it from
1098 		 * the list of handlers and wake up the sleeper.
1099 		 */
1100 		if (ih->ih_flags & IH_DEAD) {
1101 			mtx_lock(&ie->ie_lock);
1102 			if (ihp == NULL)
1103 				CK_SLIST_REMOVE_HEAD(&ie->ie_handlers, ih_next);
1104 			else
1105 				CK_SLIST_REMOVE_AFTER(ihp, ih_next);
1106 			ih->ih_flags &= ~IH_DEAD;
1107 			wakeup(ih);
1108 			mtx_unlock(&ie->ie_lock);
1109 			continue;
1110 		}
1111 
1112 		/*
1113 		 * Now that we know that the current element won't be removed
1114 		 * update the previous element.
1115 		 */
1116 		ihp = ih;
1117 
1118 		if ((ih->ih_flags & IH_CHANGED) != 0) {
1119 			mtx_lock(&ie->ie_lock);
1120 			ih->ih_flags &= ~IH_CHANGED;
1121 			wakeup(ih);
1122 			mtx_unlock(&ie->ie_lock);
1123 		}
1124 
1125 		/* Skip filter only handlers */
1126 		if (ih->ih_handler == NULL)
1127 			continue;
1128 
1129 		/* Skip suspended handlers */
1130 		if ((ih->ih_flags & IH_SUSP) != 0)
1131 			continue;
1132 
1133 		/*
1134 		 * For software interrupt threads, we only execute
1135 		 * handlers that have their need flag set.  Hardware
1136 		 * interrupt threads always invoke all of their handlers.
1137 		 *
1138 		 * ih_need can only be 0 or 1.  Failed cmpset below
1139 		 * means that there is no request to execute handlers,
1140 		 * so a retry of the cmpset is not needed.
1141 		 */
1142 		if ((ie->ie_flags & IE_SOFT) != 0 &&
1143 		    atomic_cmpset_int(&ih->ih_need, 1, 0) == 0)
1144 			continue;
1145 
1146 		/* Execute this handler. */
1147 		CTR6(KTR_INTR, "%s: pid %d exec %p(%p) for %s flg=%x",
1148 		    __func__, p->p_pid, (void *)ih->ih_handler,
1149 		    ih->ih_argument, ih->ih_name, ih->ih_flags);
1150 
1151 		if (!(ih->ih_flags & IH_MPSAFE))
1152 			mtx_lock(&Giant);
1153 		ih->ih_handler(ih->ih_argument);
1154 		if (!(ih->ih_flags & IH_MPSAFE))
1155 			mtx_unlock(&Giant);
1156 	}
1157 }
1158 
1159 static void
1160 ithread_execute_handlers(struct proc *p, struct intr_event *ie)
1161 {
1162 
1163 	/* Interrupt handlers should not sleep. */
1164 	if (!(ie->ie_flags & IE_SOFT))
1165 		THREAD_NO_SLEEPING();
1166 	intr_event_execute_handlers(p, ie);
1167 	if (!(ie->ie_flags & IE_SOFT))
1168 		THREAD_SLEEPING_OK();
1169 
1170 	/*
1171 	 * Interrupt storm handling:
1172 	 *
1173 	 * If this interrupt source is currently storming, then throttle
1174 	 * it to only fire the handler once  per clock tick.
1175 	 *
1176 	 * If this interrupt source is not currently storming, but the
1177 	 * number of back to back interrupts exceeds the storm threshold,
1178 	 * then enter storming mode.
1179 	 */
1180 	if (intr_storm_threshold != 0 && ie->ie_count >= intr_storm_threshold &&
1181 	    !(ie->ie_flags & IE_SOFT)) {
1182 		/* Report the message only once every second. */
1183 		if (ppsratecheck(&ie->ie_warntm, &ie->ie_warncnt, 1)) {
1184 			printf(
1185 	"interrupt storm detected on \"%s\"; throttling interrupt source\n",
1186 			    ie->ie_name);
1187 		}
1188 		pause("istorm", 1);
1189 	} else
1190 		ie->ie_count++;
1191 
1192 	/*
1193 	 * Now that all the handlers have had a chance to run, reenable
1194 	 * the interrupt source.
1195 	 */
1196 	if (ie->ie_post_ithread != NULL)
1197 		ie->ie_post_ithread(ie->ie_source);
1198 }
1199 
1200 /*
1201  * This is the main code for interrupt threads.
1202  */
1203 static void
1204 ithread_loop(void *arg)
1205 {
1206 	struct epoch_tracker et;
1207 	struct intr_thread *ithd;
1208 	struct intr_event *ie;
1209 	struct thread *td;
1210 	struct proc *p;
1211 	int wake, epoch_count;
1212 	bool needs_epoch;
1213 
1214 	td = curthread;
1215 	p = td->td_proc;
1216 	ithd = (struct intr_thread *)arg;
1217 	KASSERT(ithd->it_thread == td,
1218 	    ("%s: ithread and proc linkage out of sync", __func__));
1219 	ie = ithd->it_event;
1220 	ie->ie_count = 0;
1221 	wake = 0;
1222 
1223 	/*
1224 	 * As long as we have interrupts outstanding, go through the
1225 	 * list of handlers, giving each one a go at it.
1226 	 */
1227 	for (;;) {
1228 		/*
1229 		 * If we are an orphaned thread, then just die.
1230 		 */
1231 		if (ithd->it_flags & IT_DEAD) {
1232 			CTR3(KTR_INTR, "%s: pid %d (%s) exiting", __func__,
1233 			    p->p_pid, td->td_name);
1234 			free(ithd, M_ITHREAD);
1235 			kthread_exit();
1236 		}
1237 
1238 		/*
1239 		 * Service interrupts.  If another interrupt arrives while
1240 		 * we are running, it will set it_need to note that we
1241 		 * should make another pass.
1242 		 *
1243 		 * The load_acq part of the following cmpset ensures
1244 		 * that the load of ih_need in ithread_execute_handlers()
1245 		 * is ordered after the load of it_need here.
1246 		 */
1247 		needs_epoch =
1248 		    (atomic_load_int(&ie->ie_hflags) & IH_NET) != 0;
1249 		if (needs_epoch) {
1250 			epoch_count = 0;
1251 			NET_EPOCH_ENTER(et);
1252 		}
1253 		while (atomic_cmpset_acq_int(&ithd->it_need, 1, 0) != 0) {
1254 			ithread_execute_handlers(p, ie);
1255 			if (needs_epoch &&
1256 			    ++epoch_count >= intr_epoch_batch) {
1257 				NET_EPOCH_EXIT(et);
1258 				epoch_count = 0;
1259 				NET_EPOCH_ENTER(et);
1260 			}
1261 		}
1262 		if (needs_epoch)
1263 			NET_EPOCH_EXIT(et);
1264 		WITNESS_WARN(WARN_PANIC, NULL, "suspending ithread");
1265 		mtx_assert(&Giant, MA_NOTOWNED);
1266 
1267 		/*
1268 		 * Processed all our interrupts.  Now get the sched
1269 		 * lock.  This may take a while and it_need may get
1270 		 * set again, so we have to check it again.
1271 		 */
1272 		thread_lock(td);
1273 		if (atomic_load_acq_int(&ithd->it_need) == 0 &&
1274 		    (ithd->it_flags & (IT_DEAD | IT_WAIT)) == 0) {
1275 			TD_SET_IWAIT(td);
1276 			ie->ie_count = 0;
1277 			mi_switch(SW_VOL | SWT_IWAIT);
1278 		} else {
1279 			if (ithd->it_flags & IT_WAIT) {
1280 				wake = 1;
1281 				ithd->it_flags &= ~IT_WAIT;
1282 			}
1283 			thread_unlock(td);
1284 		}
1285 		if (wake) {
1286 			wakeup(ithd);
1287 			wake = 0;
1288 		}
1289 	}
1290 }
1291 
1292 /*
1293  * Main interrupt handling body.
1294  *
1295  * Input:
1296  * o ie:                        the event connected to this interrupt.
1297  * o frame:                     some archs (i.e. i386) pass a frame to some.
1298  *                              handlers as their main argument.
1299  * Return value:
1300  * o 0:                         everything ok.
1301  * o EINVAL:                    stray interrupt.
1302  */
1303 int
1304 intr_event_handle(struct intr_event *ie, struct trapframe *frame)
1305 {
1306 	struct intr_handler *ih;
1307 	struct trapframe *oldframe;
1308 	struct thread *td;
1309 	int phase;
1310 	int ret;
1311 	bool filter, thread;
1312 
1313 	td = curthread;
1314 
1315 #ifdef KSTACK_USAGE_PROF
1316 	intr_prof_stack_use(td, frame);
1317 #endif
1318 
1319 	/* An interrupt with no event or handlers is a stray interrupt. */
1320 	if (ie == NULL || CK_SLIST_EMPTY(&ie->ie_handlers))
1321 		return (EINVAL);
1322 
1323 	/*
1324 	 * Execute fast interrupt handlers directly.
1325 	 * To support clock handlers, if a handler registers
1326 	 * with a NULL argument, then we pass it a pointer to
1327 	 * a trapframe as its argument.
1328 	 */
1329 	td->td_intr_nesting_level++;
1330 	filter = false;
1331 	thread = false;
1332 	ret = 0;
1333 	critical_enter();
1334 	oldframe = td->td_intr_frame;
1335 	td->td_intr_frame = frame;
1336 
1337 	phase = ie->ie_phase;
1338 	atomic_add_int(&ie->ie_active[phase], 1);
1339 
1340 	/*
1341 	 * This fence is required to ensure that no later loads are
1342 	 * re-ordered before the ie_active store.
1343 	 */
1344 	atomic_thread_fence_seq_cst();
1345 
1346 	CK_SLIST_FOREACH(ih, &ie->ie_handlers, ih_next) {
1347 		if ((ih->ih_flags & IH_SUSP) != 0)
1348 			continue;
1349 		if (ih->ih_filter == NULL) {
1350 			thread = true;
1351 			continue;
1352 		}
1353 		CTR4(KTR_INTR, "%s: exec %p(%p) for %s", __func__,
1354 		    ih->ih_filter, ih->ih_argument == NULL ? frame :
1355 		    ih->ih_argument, ih->ih_name);
1356 		if (ih->ih_argument == NULL)
1357 			ret = ih->ih_filter(frame);
1358 		else
1359 			ret = ih->ih_filter(ih->ih_argument);
1360 		KASSERT(ret == FILTER_STRAY ||
1361 		    ((ret & (FILTER_SCHEDULE_THREAD | FILTER_HANDLED)) != 0 &&
1362 		    (ret & ~(FILTER_SCHEDULE_THREAD | FILTER_HANDLED)) == 0),
1363 		    ("%s: incorrect return value %#x from %s", __func__, ret,
1364 		    ih->ih_name));
1365 		filter = filter || ret == FILTER_HANDLED;
1366 
1367 		/*
1368 		 * Wrapper handler special handling:
1369 		 *
1370 		 * in some particular cases (like pccard and pccbb),
1371 		 * the _real_ device handler is wrapped in a couple of
1372 		 * functions - a filter wrapper and an ithread wrapper.
1373 		 * In this case (and just in this case), the filter wrapper
1374 		 * could ask the system to schedule the ithread and mask
1375 		 * the interrupt source if the wrapped handler is composed
1376 		 * of just an ithread handler.
1377 		 *
1378 		 * TODO: write a generic wrapper to avoid people rolling
1379 		 * their own.
1380 		 */
1381 		if (!thread) {
1382 			if (ret == FILTER_SCHEDULE_THREAD)
1383 				thread = true;
1384 		}
1385 	}
1386 	atomic_add_rel_int(&ie->ie_active[phase], -1);
1387 
1388 	td->td_intr_frame = oldframe;
1389 
1390 	if (thread) {
1391 		if (ie->ie_pre_ithread != NULL)
1392 			ie->ie_pre_ithread(ie->ie_source);
1393 	} else {
1394 		if (ie->ie_post_filter != NULL)
1395 			ie->ie_post_filter(ie->ie_source);
1396 	}
1397 
1398 	/* Schedule the ithread if needed. */
1399 	if (thread) {
1400 		int error __unused;
1401 
1402 		error =  intr_event_schedule_thread(ie);
1403 		KASSERT(error == 0, ("bad stray interrupt"));
1404 	}
1405 	critical_exit();
1406 	td->td_intr_nesting_level--;
1407 #ifdef notyet
1408 	/* The interrupt is not aknowledged by any filter and has no ithread. */
1409 	if (!thread && !filter)
1410 		return (EINVAL);
1411 #endif
1412 	return (0);
1413 }
1414 
1415 #ifdef DDB
1416 /*
1417  * Dump details about an interrupt handler
1418  */
1419 static void
1420 db_dump_intrhand(struct intr_handler *ih)
1421 {
1422 	int comma;
1423 
1424 	db_printf("\t%-10s ", ih->ih_name);
1425 	switch (ih->ih_pri) {
1426 	case PI_REALTIME:
1427 		db_printf("CLK ");
1428 		break;
1429 	case PI_AV:
1430 		db_printf("AV  ");
1431 		break;
1432 	case PI_TTY:
1433 		db_printf("TTY ");
1434 		break;
1435 	case PI_NET:
1436 		db_printf("NET ");
1437 		break;
1438 	case PI_DISK:
1439 		db_printf("DISK");
1440 		break;
1441 	case PI_DULL:
1442 		db_printf("DULL");
1443 		break;
1444 	default:
1445 		if (ih->ih_pri >= PI_SOFT)
1446 			db_printf("SWI ");
1447 		else
1448 			db_printf("%4u", ih->ih_pri);
1449 		break;
1450 	}
1451 	db_printf(" ");
1452 	if (ih->ih_filter != NULL) {
1453 		db_printf("[F]");
1454 		db_printsym((uintptr_t)ih->ih_filter, DB_STGY_PROC);
1455 	}
1456 	if (ih->ih_handler != NULL) {
1457 		if (ih->ih_filter != NULL)
1458 			db_printf(",");
1459 		db_printf("[H]");
1460 		db_printsym((uintptr_t)ih->ih_handler, DB_STGY_PROC);
1461 	}
1462 	db_printf("(%p)", ih->ih_argument);
1463 	if (ih->ih_need ||
1464 	    (ih->ih_flags & (IH_EXCLUSIVE | IH_ENTROPY | IH_DEAD |
1465 	    IH_MPSAFE)) != 0) {
1466 		db_printf(" {");
1467 		comma = 0;
1468 		if (ih->ih_flags & IH_EXCLUSIVE) {
1469 			if (comma)
1470 				db_printf(", ");
1471 			db_printf("EXCL");
1472 			comma = 1;
1473 		}
1474 		if (ih->ih_flags & IH_ENTROPY) {
1475 			if (comma)
1476 				db_printf(", ");
1477 			db_printf("ENTROPY");
1478 			comma = 1;
1479 		}
1480 		if (ih->ih_flags & IH_DEAD) {
1481 			if (comma)
1482 				db_printf(", ");
1483 			db_printf("DEAD");
1484 			comma = 1;
1485 		}
1486 		if (ih->ih_flags & IH_MPSAFE) {
1487 			if (comma)
1488 				db_printf(", ");
1489 			db_printf("MPSAFE");
1490 			comma = 1;
1491 		}
1492 		if (ih->ih_need) {
1493 			if (comma)
1494 				db_printf(", ");
1495 			db_printf("NEED");
1496 		}
1497 		db_printf("}");
1498 	}
1499 	db_printf("\n");
1500 }
1501 
1502 /*
1503  * Dump details about a event.
1504  */
1505 void
1506 db_dump_intr_event(struct intr_event *ie, int handlers)
1507 {
1508 	struct intr_handler *ih;
1509 	struct intr_thread *it;
1510 	int comma;
1511 
1512 	db_printf("%s ", ie->ie_fullname);
1513 	it = ie->ie_thread;
1514 	if (it != NULL)
1515 		db_printf("(pid %d)", it->it_thread->td_proc->p_pid);
1516 	else
1517 		db_printf("(no thread)");
1518 	if ((ie->ie_flags & (IE_SOFT | IE_ADDING_THREAD)) != 0 ||
1519 	    (it != NULL && it->it_need)) {
1520 		db_printf(" {");
1521 		comma = 0;
1522 		if (ie->ie_flags & IE_SOFT) {
1523 			db_printf("SOFT");
1524 			comma = 1;
1525 		}
1526 		if (ie->ie_flags & IE_ADDING_THREAD) {
1527 			if (comma)
1528 				db_printf(", ");
1529 			db_printf("ADDING_THREAD");
1530 			comma = 1;
1531 		}
1532 		if (it != NULL && it->it_need) {
1533 			if (comma)
1534 				db_printf(", ");
1535 			db_printf("NEED");
1536 		}
1537 		db_printf("}");
1538 	}
1539 	db_printf("\n");
1540 
1541 	if (handlers)
1542 		CK_SLIST_FOREACH(ih, &ie->ie_handlers, ih_next)
1543 		    db_dump_intrhand(ih);
1544 }
1545 
1546 /*
1547  * Dump data about interrupt handlers
1548  */
1549 DB_SHOW_COMMAND(intr, db_show_intr)
1550 {
1551 	struct intr_event *ie;
1552 	int all, verbose;
1553 
1554 	verbose = strchr(modif, 'v') != NULL;
1555 	all = strchr(modif, 'a') != NULL;
1556 	TAILQ_FOREACH(ie, &event_list, ie_list) {
1557 		if (!all && CK_SLIST_EMPTY(&ie->ie_handlers))
1558 			continue;
1559 		db_dump_intr_event(ie, verbose);
1560 		if (db_pager_quit)
1561 			break;
1562 	}
1563 }
1564 #endif /* DDB */
1565 
1566 /*
1567  * Start standard software interrupt threads
1568  */
1569 static void
1570 start_softintr(void *dummy)
1571 {
1572 
1573 	if (swi_add(NULL, "vm", swi_vm, NULL, SWI_VM, INTR_MPSAFE, &vm_ih))
1574 		panic("died while creating vm swi ithread");
1575 }
1576 SYSINIT(start_softintr, SI_SUB_SOFTINTR, SI_ORDER_FIRST, start_softintr,
1577     NULL);
1578 
1579 /*
1580  * Sysctls used by systat and others: hw.intrnames and hw.intrcnt.
1581  * The data for this machine dependent, and the declarations are in machine
1582  * dependent code.  The layout of intrnames and intrcnt however is machine
1583  * independent.
1584  *
1585  * We do not know the length of intrcnt and intrnames at compile time, so
1586  * calculate things at run time.
1587  */
1588 static int
1589 sysctl_intrnames(SYSCTL_HANDLER_ARGS)
1590 {
1591 	return (sysctl_handle_opaque(oidp, intrnames, sintrnames, req));
1592 }
1593 
1594 SYSCTL_PROC(_hw, OID_AUTO, intrnames,
1595     CTLTYPE_OPAQUE | CTLFLAG_RD | CTLFLAG_NEEDGIANT, NULL, 0,
1596     sysctl_intrnames, "",
1597     "Interrupt Names");
1598 
1599 static int
1600 sysctl_intrcnt(SYSCTL_HANDLER_ARGS)
1601 {
1602 #ifdef SCTL_MASK32
1603 	uint32_t *intrcnt32;
1604 	unsigned i;
1605 	int error;
1606 
1607 	if (req->flags & SCTL_MASK32) {
1608 		if (!req->oldptr)
1609 			return (sysctl_handle_opaque(oidp, NULL, sintrcnt / 2, req));
1610 		intrcnt32 = malloc(sintrcnt / 2, M_TEMP, M_NOWAIT);
1611 		if (intrcnt32 == NULL)
1612 			return (ENOMEM);
1613 		for (i = 0; i < sintrcnt / sizeof (u_long); i++)
1614 			intrcnt32[i] = intrcnt[i];
1615 		error = sysctl_handle_opaque(oidp, intrcnt32, sintrcnt / 2, req);
1616 		free(intrcnt32, M_TEMP);
1617 		return (error);
1618 	}
1619 #endif
1620 	return (sysctl_handle_opaque(oidp, intrcnt, sintrcnt, req));
1621 }
1622 
1623 SYSCTL_PROC(_hw, OID_AUTO, intrcnt,
1624     CTLTYPE_OPAQUE | CTLFLAG_RD | CTLFLAG_NEEDGIANT, NULL, 0,
1625     sysctl_intrcnt, "",
1626     "Interrupt Counts");
1627 
1628 #ifdef DDB
1629 /*
1630  * DDB command to dump the interrupt statistics.
1631  */
1632 DB_SHOW_COMMAND(intrcnt, db_show_intrcnt)
1633 {
1634 	u_long *i;
1635 	char *cp;
1636 	u_int j;
1637 
1638 	cp = intrnames;
1639 	j = 0;
1640 	for (i = intrcnt; j < (sintrcnt / sizeof(u_long)) && !db_pager_quit;
1641 	    i++, j++) {
1642 		if (*cp == '\0')
1643 			break;
1644 		if (*i != 0)
1645 			db_printf("%s\t%lu\n", cp, *i);
1646 		cp += strlen(cp) + 1;
1647 	}
1648 }
1649 #endif
1650