xref: /freebsd/sys/kern/kern_intr.c (revision 4c9a0adad18263ec8725d9bfc5f560c6ad1da8bd)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
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 #include "opt_ddb.h"
31 #include "opt_hwpmc_hooks.h"
32 #include "opt_kstack_usage_prof.h"
33 
34 #include <sys/param.h>
35 #include <sys/bus.h>
36 #include <sys/conf.h>
37 #include <sys/cpuset.h>
38 #include <sys/rtprio.h>
39 #include <sys/systm.h>
40 #include <sys/interrupt.h>
41 #include <sys/kernel.h>
42 #include <sys/kthread.h>
43 #include <sys/ktr.h>
44 #include <sys/limits.h>
45 #include <sys/lock.h>
46 #include <sys/malloc.h>
47 #include <sys/mutex.h>
48 #include <sys/priv.h>
49 #include <sys/proc.h>
50 #include <sys/epoch.h>
51 #include <sys/random.h>
52 #include <sys/resourcevar.h>
53 #include <sys/sched.h>
54 #include <sys/smp.h>
55 #include <sys/sysctl.h>
56 #include <sys/syslog.h>
57 #include <sys/unistd.h>
58 #include <sys/vmmeter.h>
59 #include <machine/atomic.h>
60 #include <machine/cpu.h>
61 #include <machine/md_var.h>
62 #include <machine/smp.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 	int	it_waiting;		/* Waiting in the runq. */
78 };
79 
80 /* Interrupt thread flags kept in it_flags */
81 #define	IT_DEAD		0x000001	/* Thread is waiting to exit. */
82 #define	IT_WAIT		0x000002	/* Thread is waiting for completion. */
83 
84 struct	intr_entropy {
85 	struct	thread *td;
86 	uintptr_t event;
87 };
88 
89 struct	intr_event *clk_intr_event;
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 #ifdef HWPMC_HOOKS
102 static int intr_hwpmc_waiting_report_threshold = 1;
103 SYSCTL_INT(_hw, OID_AUTO, intr_hwpmc_waiting_report_threshold, CTLFLAG_RWTUN,
104     &intr_hwpmc_waiting_report_threshold, 1,
105     "Threshold for reporting number of events in a workq");
106 #define	PMC_HOOK_INSTALLED_ANY() __predict_false(pmc_hook != NULL)
107 #endif
108 static TAILQ_HEAD(, intr_event) event_list =
109     TAILQ_HEAD_INITIALIZER(event_list);
110 static struct mtx event_lock;
111 MTX_SYSINIT(intr_event_list, &event_lock, "intr event list", MTX_DEF);
112 
113 static void	intr_event_update(struct intr_event *ie);
114 static int	intr_event_schedule_thread(struct intr_event *ie, struct trapframe *frame);
115 static struct intr_thread *ithread_create(const char *name);
116 static void	ithread_destroy(struct intr_thread *ithread);
117 static void	ithread_execute_handlers(struct proc *p,
118 		    struct intr_event *ie);
119 static void	ithread_loop(void *);
120 static void	ithread_update(struct intr_thread *ithd);
121 static void	start_softintr(void *);
122 
123 #ifdef HWPMC_HOOKS
124 #include <sys/pmckern.h>
125 PMC_SOFT_DEFINE( , , intr, all);
126 PMC_SOFT_DEFINE( , , intr, ithread);
127 PMC_SOFT_DEFINE( , , intr, filter);
128 PMC_SOFT_DEFINE( , , intr, stray);
129 PMC_SOFT_DEFINE( , , intr, schedule);
130 PMC_SOFT_DEFINE( , , intr, waiting);
131 
132 #define PMC_SOFT_CALL_INTR_HLPR(event, frame)			\
133 do {					\
134 	if (frame != NULL)					\
135 		PMC_SOFT_CALL_TF( , , intr, event, frame);	\
136 	else							\
137 		PMC_SOFT_CALL( , , intr, event);		\
138 } while (0)
139 #endif
140 
141 /* Map an interrupt type to an ithread priority. */
142 u_char
143 intr_priority(enum intr_type flags)
144 {
145 	u_char pri;
146 
147 	flags &= (INTR_TYPE_TTY | INTR_TYPE_BIO | INTR_TYPE_NET |
148 	    INTR_TYPE_CAM | INTR_TYPE_MISC | INTR_TYPE_CLK | INTR_TYPE_AV);
149 	switch (flags) {
150 	case INTR_TYPE_TTY:
151 		pri = PI_TTY;
152 		break;
153 	case INTR_TYPE_BIO:
154 		pri = PI_DISK;
155 		break;
156 	case INTR_TYPE_NET:
157 		pri = PI_NET;
158 		break;
159 	case INTR_TYPE_CAM:
160 		pri = PI_DISK;
161 		break;
162 	case INTR_TYPE_AV:
163 		pri = PI_AV;
164 		break;
165 	case INTR_TYPE_CLK:
166 		pri = PI_REALTIME;
167 		break;
168 	case INTR_TYPE_MISC:
169 		pri = PI_DULL;          /* don't care */
170 		break;
171 	default:
172 		/* We didn't specify an interrupt level. */
173 		panic("intr_priority: no interrupt type in flags");
174 	}
175 
176 	return pri;
177 }
178 
179 /*
180  * Update an ithread based on the associated intr_event.
181  */
182 static void
183 ithread_update(struct intr_thread *ithd)
184 {
185 	struct intr_event *ie;
186 	struct thread *td;
187 	u_char pri;
188 
189 	ie = ithd->it_event;
190 	td = ithd->it_thread;
191 	mtx_assert(&ie->ie_lock, MA_OWNED);
192 
193 	/* Determine the overall priority of this event. */
194 	if (CK_SLIST_EMPTY(&ie->ie_handlers))
195 		pri = PRI_MAX_ITHD;
196 	else
197 		pri = CK_SLIST_FIRST(&ie->ie_handlers)->ih_pri;
198 
199 	/* Update name and priority. */
200 	strlcpy(td->td_name, ie->ie_fullname, sizeof(td->td_name));
201 #ifdef KTR
202 	sched_clear_tdname(td);
203 #endif
204 	thread_lock(td);
205 	sched_ithread_prio(td, pri);
206 	thread_unlock(td);
207 }
208 
209 /*
210  * Regenerate the full name of an interrupt event and update its priority.
211  */
212 static void
213 intr_event_update(struct intr_event *ie)
214 {
215 	struct intr_handler *ih;
216 	char *last;
217 	int missed, space, flags;
218 
219 	/* Start off with no entropy and just the name of the event. */
220 	mtx_assert(&ie->ie_lock, MA_OWNED);
221 	strlcpy(ie->ie_fullname, ie->ie_name, sizeof(ie->ie_fullname));
222 	flags = 0;
223 	missed = 0;
224 	space = 1;
225 
226 	/* Run through all the handlers updating values. */
227 	CK_SLIST_FOREACH(ih, &ie->ie_handlers, ih_next) {
228 		if (strlen(ie->ie_fullname) + strlen(ih->ih_name) + 1 <
229 		    sizeof(ie->ie_fullname)) {
230 			strcat(ie->ie_fullname, " ");
231 			strcat(ie->ie_fullname, ih->ih_name);
232 			space = 0;
233 		} else
234 			missed++;
235 		flags |= ih->ih_flags;
236 	}
237 	ie->ie_hflags = flags;
238 
239 	/*
240 	 * If there is only one handler and its name is too long, just copy in
241 	 * as much of the end of the name (includes the unit number) as will
242 	 * fit.  Otherwise, we have multiple handlers and not all of the names
243 	 * will fit.  Add +'s to indicate missing names.  If we run out of room
244 	 * and still have +'s to add, change the last character from a + to a *.
245 	 */
246 	if (missed == 1 && space == 1) {
247 		ih = CK_SLIST_FIRST(&ie->ie_handlers);
248 		missed = strlen(ie->ie_fullname) + strlen(ih->ih_name) + 2 -
249 		    sizeof(ie->ie_fullname);
250 		strcat(ie->ie_fullname, (missed == 0) ? " " : "-");
251 		strcat(ie->ie_fullname, &ih->ih_name[missed]);
252 		missed = 0;
253 	}
254 	last = &ie->ie_fullname[sizeof(ie->ie_fullname) - 2];
255 	while (missed-- > 0) {
256 		if (strlen(ie->ie_fullname) + 1 == sizeof(ie->ie_fullname)) {
257 			if (*last == '+') {
258 				*last = '*';
259 				break;
260 			} else
261 				*last = '+';
262 		} else if (space) {
263 			strcat(ie->ie_fullname, " +");
264 			space = 0;
265 		} else
266 			strcat(ie->ie_fullname, "+");
267 	}
268 
269 	/*
270 	 * If this event has an ithread, update it's priority and
271 	 * name.
272 	 */
273 	if (ie->ie_thread != NULL)
274 		ithread_update(ie->ie_thread);
275 	CTR2(KTR_INTR, "%s: updated %s", __func__, ie->ie_fullname);
276 }
277 
278 int
279 intr_event_create(struct intr_event **event, void *source, int flags, int irq,
280     void (*pre_ithread)(void *), void (*post_ithread)(void *),
281     void (*post_filter)(void *), int (*assign_cpu)(void *, int),
282     const char *fmt, ...)
283 {
284 	struct intr_event *ie;
285 	va_list ap;
286 
287 	/* The only valid flag during creation is IE_SOFT. */
288 	if ((flags & ~IE_SOFT) != 0)
289 		return (EINVAL);
290 	ie = malloc(sizeof(struct intr_event), M_ITHREAD, M_WAITOK | M_ZERO);
291 	ie->ie_source = source;
292 	ie->ie_pre_ithread = pre_ithread;
293 	ie->ie_post_ithread = post_ithread;
294 	ie->ie_post_filter = post_filter;
295 	ie->ie_assign_cpu = assign_cpu;
296 	ie->ie_flags = flags;
297 	ie->ie_irq = irq;
298 	ie->ie_cpu = NOCPU;
299 	CK_SLIST_INIT(&ie->ie_handlers);
300 	mtx_init(&ie->ie_lock, "intr event", NULL, MTX_DEF);
301 
302 	va_start(ap, fmt);
303 	vsnprintf(ie->ie_name, sizeof(ie->ie_name), fmt, ap);
304 	va_end(ap);
305 	strlcpy(ie->ie_fullname, ie->ie_name, sizeof(ie->ie_fullname));
306 	mtx_lock(&event_lock);
307 	TAILQ_INSERT_TAIL(&event_list, ie, ie_list);
308 	mtx_unlock(&event_lock);
309 	if (event != NULL)
310 		*event = ie;
311 	CTR2(KTR_INTR, "%s: created %s", __func__, ie->ie_name);
312 	return (0);
313 }
314 
315 /*
316  * Bind an interrupt event to the specified CPU.  Note that not all
317  * platforms support binding an interrupt to a CPU.  For those
318  * platforms this request will fail.  Using a cpu id of NOCPU unbinds
319  * the interrupt event.
320  */
321 static int
322 _intr_event_bind(struct intr_event *ie, int cpu, bool bindirq, bool bindithread)
323 {
324 	lwpid_t id;
325 	int error;
326 
327 	/* Need a CPU to bind to. */
328 	if (cpu != NOCPU && CPU_ABSENT(cpu))
329 		return (EINVAL);
330 
331 	if (ie->ie_assign_cpu == NULL)
332 		return (EOPNOTSUPP);
333 
334 	error = priv_check(curthread, PRIV_SCHED_CPUSET_INTR);
335 	if (error)
336 		return (error);
337 
338 	/*
339 	 * If we have any ithreads try to set their mask first to verify
340 	 * permissions, etc.
341 	 */
342 	if (bindithread) {
343 		mtx_lock(&ie->ie_lock);
344 		if (ie->ie_thread != NULL) {
345 			id = ie->ie_thread->it_thread->td_tid;
346 			mtx_unlock(&ie->ie_lock);
347 			error = cpuset_setithread(id, cpu);
348 			if (error)
349 				return (error);
350 		} else
351 			mtx_unlock(&ie->ie_lock);
352 	}
353 	if (bindirq)
354 		error = ie->ie_assign_cpu(ie->ie_source, cpu);
355 	if (error) {
356 		if (bindithread) {
357 			mtx_lock(&ie->ie_lock);
358 			if (ie->ie_thread != NULL) {
359 				cpu = ie->ie_cpu;
360 				id = ie->ie_thread->it_thread->td_tid;
361 				mtx_unlock(&ie->ie_lock);
362 				(void)cpuset_setithread(id, cpu);
363 			} else
364 				mtx_unlock(&ie->ie_lock);
365 		}
366 		return (error);
367 	}
368 
369 	if (bindirq) {
370 		mtx_lock(&ie->ie_lock);
371 		ie->ie_cpu = cpu;
372 		mtx_unlock(&ie->ie_lock);
373 	}
374 
375 	return (error);
376 }
377 
378 /*
379  * Bind an interrupt event to the specified CPU.  For supported platforms, any
380  * associated ithreads as well as the primary interrupt context will be bound
381  * to the specificed CPU.
382  */
383 int
384 intr_event_bind(struct intr_event *ie, int cpu)
385 {
386 
387 	return (_intr_event_bind(ie, cpu, true, true));
388 }
389 
390 /*
391  * Bind an interrupt event to the specified CPU, but do not bind associated
392  * ithreads.
393  */
394 int
395 intr_event_bind_irqonly(struct intr_event *ie, int cpu)
396 {
397 
398 	return (_intr_event_bind(ie, cpu, true, false));
399 }
400 
401 /*
402  * Bind an interrupt event's ithread to the specified CPU.
403  */
404 int
405 intr_event_bind_ithread(struct intr_event *ie, int cpu)
406 {
407 
408 	return (_intr_event_bind(ie, cpu, false, true));
409 }
410 
411 /*
412  * Bind an interrupt event's ithread to the specified cpuset.
413  */
414 int
415 intr_event_bind_ithread_cpuset(struct intr_event *ie, cpuset_t *cs)
416 {
417 	lwpid_t id;
418 
419 	mtx_lock(&ie->ie_lock);
420 	if (ie->ie_thread != NULL) {
421 		id = ie->ie_thread->it_thread->td_tid;
422 		mtx_unlock(&ie->ie_lock);
423 		return (cpuset_setthread(id, cs));
424 	} else {
425 		mtx_unlock(&ie->ie_lock);
426 	}
427 	return (ENODEV);
428 }
429 
430 static struct intr_event *
431 intr_lookup(int irq)
432 {
433 	struct intr_event *ie;
434 
435 	mtx_lock(&event_lock);
436 	TAILQ_FOREACH(ie, &event_list, ie_list)
437 		if (ie->ie_irq == irq &&
438 		    (ie->ie_flags & IE_SOFT) == 0 &&
439 		    CK_SLIST_FIRST(&ie->ie_handlers) != NULL)
440 			break;
441 	mtx_unlock(&event_lock);
442 	return (ie);
443 }
444 
445 int
446 intr_setaffinity(int irq, int mode, void *m)
447 {
448 	struct intr_event *ie;
449 	cpuset_t *mask;
450 	int cpu, n;
451 
452 	mask = m;
453 	cpu = NOCPU;
454 	/*
455 	 * If we're setting all cpus we can unbind.  Otherwise make sure
456 	 * only one cpu is in the set.
457 	 */
458 	if (CPU_CMP(cpuset_root, mask)) {
459 		for (n = 0; n < CPU_SETSIZE; n++) {
460 			if (!CPU_ISSET(n, mask))
461 				continue;
462 			if (cpu != NOCPU)
463 				return (EINVAL);
464 			cpu = n;
465 		}
466 	}
467 	ie = intr_lookup(irq);
468 	if (ie == NULL)
469 		return (ESRCH);
470 	switch (mode) {
471 	case CPU_WHICH_IRQ:
472 		return (intr_event_bind(ie, cpu));
473 	case CPU_WHICH_INTRHANDLER:
474 		return (intr_event_bind_irqonly(ie, cpu));
475 	case CPU_WHICH_ITHREAD:
476 		return (intr_event_bind_ithread(ie, cpu));
477 	default:
478 		return (EINVAL);
479 	}
480 }
481 
482 int
483 intr_getaffinity(int irq, int mode, void *m)
484 {
485 	struct intr_event *ie;
486 	struct thread *td;
487 	struct proc *p;
488 	cpuset_t *mask;
489 	lwpid_t id;
490 	int error;
491 
492 	mask = m;
493 	ie = intr_lookup(irq);
494 	if (ie == NULL)
495 		return (ESRCH);
496 
497 	error = 0;
498 	CPU_ZERO(mask);
499 	switch (mode) {
500 	case CPU_WHICH_IRQ:
501 	case CPU_WHICH_INTRHANDLER:
502 		mtx_lock(&ie->ie_lock);
503 		if (ie->ie_cpu == NOCPU)
504 			CPU_COPY(cpuset_root, mask);
505 		else
506 			CPU_SET(ie->ie_cpu, mask);
507 		mtx_unlock(&ie->ie_lock);
508 		break;
509 	case CPU_WHICH_ITHREAD:
510 		mtx_lock(&ie->ie_lock);
511 		if (ie->ie_thread == NULL) {
512 			mtx_unlock(&ie->ie_lock);
513 			CPU_COPY(cpuset_root, mask);
514 		} else {
515 			id = ie->ie_thread->it_thread->td_tid;
516 			mtx_unlock(&ie->ie_lock);
517 			error = cpuset_which(CPU_WHICH_TID, id, &p, &td, NULL);
518 			if (error != 0)
519 				return (error);
520 			CPU_COPY(&td->td_cpuset->cs_mask, mask);
521 			PROC_UNLOCK(p);
522 		}
523 	default:
524 		return (EINVAL);
525 	}
526 	return (0);
527 }
528 
529 int
530 intr_event_destroy(struct intr_event *ie)
531 {
532 
533 	if (ie == NULL)
534 		return (EINVAL);
535 
536 	mtx_lock(&event_lock);
537 	mtx_lock(&ie->ie_lock);
538 	if (!CK_SLIST_EMPTY(&ie->ie_handlers)) {
539 		mtx_unlock(&ie->ie_lock);
540 		mtx_unlock(&event_lock);
541 		return (EBUSY);
542 	}
543 	TAILQ_REMOVE(&event_list, ie, ie_list);
544 #ifndef notyet
545 	if (ie->ie_thread != NULL) {
546 		ithread_destroy(ie->ie_thread);
547 		ie->ie_thread = NULL;
548 	}
549 #endif
550 	mtx_unlock(&ie->ie_lock);
551 	mtx_unlock(&event_lock);
552 	mtx_destroy(&ie->ie_lock);
553 	free(ie, M_ITHREAD);
554 	return (0);
555 }
556 
557 static struct intr_thread *
558 ithread_create(const char *name)
559 {
560 	struct intr_thread *ithd;
561 	struct thread *td;
562 	int error;
563 
564 	ithd = malloc(sizeof(struct intr_thread), M_ITHREAD, M_WAITOK | M_ZERO);
565 
566 	error = kproc_kthread_add(ithread_loop, ithd, &intrproc,
567 		    &td, RFSTOPPED | RFHIGHPID,
568 		    0, "intr", "%s", name);
569 	if (error)
570 		panic("kproc_create() failed with %d", error);
571 	thread_lock(td);
572 	sched_class(td, PRI_ITHD);
573 	TD_SET_IWAIT(td);
574 	thread_unlock(td);
575 	td->td_pflags |= TDP_ITHREAD;
576 	ithd->it_thread = td;
577 	CTR2(KTR_INTR, "%s: created %s", __func__, name);
578 	return (ithd);
579 }
580 
581 static void
582 ithread_destroy(struct intr_thread *ithread)
583 {
584 	struct thread *td;
585 
586 	CTR2(KTR_INTR, "%s: killing %s", __func__, ithread->it_event->ie_name);
587 	td = ithread->it_thread;
588 	thread_lock(td);
589 	ithread->it_flags |= IT_DEAD;
590 	if (TD_AWAITING_INTR(td)) {
591 		TD_CLR_IWAIT(td);
592 		sched_wakeup(td, SRQ_INTR);
593 	} else
594 		thread_unlock(td);
595 }
596 
597 int
598 intr_event_add_handler(struct intr_event *ie, const char *name,
599     driver_filter_t filter, driver_intr_t handler, void *arg, u_char pri,
600     enum intr_type flags, void **cookiep)
601 {
602 	struct intr_handler *ih, *temp_ih;
603 	struct intr_handler **prevptr;
604 	struct intr_thread *it;
605 
606 	if (ie == NULL || name == NULL || (handler == NULL && filter == NULL))
607 		return (EINVAL);
608 
609 	/* Allocate and populate an interrupt handler structure. */
610 	ih = malloc(sizeof(struct intr_handler), M_ITHREAD, M_WAITOK | M_ZERO);
611 	ih->ih_filter = filter;
612 	ih->ih_handler = handler;
613 	ih->ih_argument = arg;
614 	strlcpy(ih->ih_name, name, sizeof(ih->ih_name));
615 	ih->ih_event = ie;
616 	ih->ih_pri = pri;
617 	if (flags & INTR_EXCL)
618 		ih->ih_flags = IH_EXCLUSIVE;
619 	if (flags & INTR_MPSAFE)
620 		ih->ih_flags |= IH_MPSAFE;
621 	if (flags & INTR_ENTROPY)
622 		ih->ih_flags |= IH_ENTROPY;
623 	if (flags & INTR_TYPE_NET)
624 		ih->ih_flags |= IH_NET;
625 
626 	/* We can only have one exclusive handler in a event. */
627 	mtx_lock(&ie->ie_lock);
628 	if (!CK_SLIST_EMPTY(&ie->ie_handlers)) {
629 		if ((flags & INTR_EXCL) ||
630 		    (CK_SLIST_FIRST(&ie->ie_handlers)->ih_flags & IH_EXCLUSIVE)) {
631 			mtx_unlock(&ie->ie_lock);
632 			free(ih, M_ITHREAD);
633 			return (EINVAL);
634 		}
635 	}
636 
637 	/* Create a thread if we need one. */
638 	while (ie->ie_thread == NULL && handler != NULL) {
639 		if (ie->ie_flags & IE_ADDING_THREAD)
640 			msleep(ie, &ie->ie_lock, 0, "ithread", 0);
641 		else {
642 			ie->ie_flags |= IE_ADDING_THREAD;
643 			mtx_unlock(&ie->ie_lock);
644 			it = ithread_create("intr: newborn");
645 			mtx_lock(&ie->ie_lock);
646 			ie->ie_flags &= ~IE_ADDING_THREAD;
647 			ie->ie_thread = it;
648 			it->it_event = ie;
649 			ithread_update(it);
650 			wakeup(ie);
651 		}
652 	}
653 
654 	/* Add the new handler to the event in priority order. */
655 	CK_SLIST_FOREACH_PREVPTR(temp_ih, prevptr, &ie->ie_handlers, ih_next) {
656 		if (temp_ih->ih_pri > ih->ih_pri)
657 			break;
658 	}
659 	CK_SLIST_INSERT_PREVPTR(prevptr, temp_ih, ih, ih_next);
660 
661 	intr_event_update(ie);
662 
663 	CTR3(KTR_INTR, "%s: added %s to %s", __func__, ih->ih_name,
664 	    ie->ie_name);
665 	mtx_unlock(&ie->ie_lock);
666 
667 	if (cookiep != NULL)
668 		*cookiep = ih;
669 	return (0);
670 }
671 
672 /*
673  * Append a description preceded by a ':' to the name of the specified
674  * interrupt handler.
675  */
676 int
677 intr_event_describe_handler(struct intr_event *ie, void *cookie,
678     const char *descr)
679 {
680 	struct intr_handler *ih;
681 	size_t space;
682 	char *start;
683 
684 	mtx_lock(&ie->ie_lock);
685 #ifdef INVARIANTS
686 	CK_SLIST_FOREACH(ih, &ie->ie_handlers, ih_next) {
687 		if (ih == cookie)
688 			break;
689 	}
690 	if (ih == NULL) {
691 		mtx_unlock(&ie->ie_lock);
692 		panic("handler %p not found in interrupt event %p", cookie, ie);
693 	}
694 #endif
695 	ih = cookie;
696 
697 	/*
698 	 * Look for an existing description by checking for an
699 	 * existing ":".  This assumes device names do not include
700 	 * colons.  If one is found, prepare to insert the new
701 	 * description at that point.  If one is not found, find the
702 	 * end of the name to use as the insertion point.
703 	 */
704 	start = strchr(ih->ih_name, ':');
705 	if (start == NULL)
706 		start = strchr(ih->ih_name, 0);
707 
708 	/*
709 	 * See if there is enough remaining room in the string for the
710 	 * description + ":".  The "- 1" leaves room for the trailing
711 	 * '\0'.  The "+ 1" accounts for the colon.
712 	 */
713 	space = sizeof(ih->ih_name) - (start - ih->ih_name) - 1;
714 	if (strlen(descr) + 1 > space) {
715 		mtx_unlock(&ie->ie_lock);
716 		return (ENOSPC);
717 	}
718 
719 	/* Append a colon followed by the description. */
720 	*start = ':';
721 	strcpy(start + 1, descr);
722 	intr_event_update(ie);
723 	mtx_unlock(&ie->ie_lock);
724 	return (0);
725 }
726 
727 /*
728  * Return the ie_source field from the intr_event an intr_handler is
729  * associated with.
730  */
731 void *
732 intr_handler_source(void *cookie)
733 {
734 	struct intr_handler *ih;
735 	struct intr_event *ie;
736 
737 	ih = (struct intr_handler *)cookie;
738 	if (ih == NULL)
739 		return (NULL);
740 	ie = ih->ih_event;
741 	KASSERT(ie != NULL,
742 	    ("interrupt handler \"%s\" has a NULL interrupt event",
743 	    ih->ih_name));
744 	return (ie->ie_source);
745 }
746 
747 /*
748  * If intr_event_handle() is running in the ISR context at the time of the call,
749  * then wait for it to complete.
750  */
751 static void
752 intr_event_barrier(struct intr_event *ie)
753 {
754 	int phase;
755 
756 	mtx_assert(&ie->ie_lock, MA_OWNED);
757 	phase = ie->ie_phase;
758 
759 	/*
760 	 * Switch phase to direct future interrupts to the other active counter.
761 	 * Make sure that any preceding stores are visible before the switch.
762 	 */
763 	KASSERT(ie->ie_active[!phase] == 0, ("idle phase has activity"));
764 	atomic_store_rel_int(&ie->ie_phase, !phase);
765 
766 	/*
767 	 * This code cooperates with wait-free iteration of ie_handlers
768 	 * in intr_event_handle.
769 	 * Make sure that the removal and the phase update are not reordered
770 	 * with the active count check.
771 	 * Note that no combination of acquire and release fences can provide
772 	 * that guarantee as Store->Load sequences can always be reordered.
773 	 */
774 	atomic_thread_fence_seq_cst();
775 
776 	/*
777 	 * Now wait on the inactive phase.
778 	 * The acquire fence is needed so that all post-barrier accesses
779 	 * are after the check.
780 	 */
781 	while (ie->ie_active[phase] > 0)
782 		cpu_spinwait();
783 	atomic_thread_fence_acq();
784 }
785 
786 static void
787 intr_handler_barrier(struct intr_handler *handler)
788 {
789 	struct intr_event *ie;
790 
791 	ie = handler->ih_event;
792 	mtx_assert(&ie->ie_lock, MA_OWNED);
793 	KASSERT((handler->ih_flags & IH_DEAD) == 0,
794 	    ("update for a removed handler"));
795 
796 	if (ie->ie_thread == NULL) {
797 		intr_event_barrier(ie);
798 		return;
799 	}
800 	if ((handler->ih_flags & IH_CHANGED) == 0) {
801 		handler->ih_flags |= IH_CHANGED;
802 		intr_event_schedule_thread(ie, NULL);
803 	}
804 	while ((handler->ih_flags & IH_CHANGED) != 0)
805 		msleep(handler, &ie->ie_lock, 0, "ih_barr", 0);
806 }
807 
808 /*
809  * Sleep until an ithread finishes executing an interrupt handler.
810  *
811  * XXX Doesn't currently handle interrupt filters or fast interrupt
812  * handlers. This is intended for LinuxKPI drivers only.
813  * Do not use in BSD code.
814  */
815 void
816 _intr_drain(int irq)
817 {
818 	struct intr_event *ie;
819 	struct intr_thread *ithd;
820 	struct thread *td;
821 
822 	ie = intr_lookup(irq);
823 	if (ie == NULL)
824 		return;
825 	if (ie->ie_thread == NULL)
826 		return;
827 	ithd = ie->ie_thread;
828 	td = ithd->it_thread;
829 	/*
830 	 * We set the flag and wait for it to be cleared to avoid
831 	 * long delays with potentially busy interrupt handlers
832 	 * were we to only sample TD_AWAITING_INTR() every tick.
833 	 */
834 	thread_lock(td);
835 	if (!TD_AWAITING_INTR(td)) {
836 		ithd->it_flags |= IT_WAIT;
837 		while (ithd->it_flags & IT_WAIT) {
838 			thread_unlock(td);
839 			pause("idrain", 1);
840 			thread_lock(td);
841 		}
842 	}
843 	thread_unlock(td);
844 	return;
845 }
846 
847 int
848 intr_event_remove_handler(void *cookie)
849 {
850 	struct intr_handler *handler = (struct intr_handler *)cookie;
851 	struct intr_event *ie;
852 	struct intr_handler *ih;
853 	struct intr_handler **prevptr;
854 #ifdef notyet
855 	int dead;
856 #endif
857 
858 	if (handler == NULL)
859 		return (EINVAL);
860 	ie = handler->ih_event;
861 	KASSERT(ie != NULL,
862 	    ("interrupt handler \"%s\" has a NULL interrupt event",
863 	    handler->ih_name));
864 
865 	mtx_lock(&ie->ie_lock);
866 	CTR3(KTR_INTR, "%s: removing %s from %s", __func__, handler->ih_name,
867 	    ie->ie_name);
868 	CK_SLIST_FOREACH_PREVPTR(ih, prevptr, &ie->ie_handlers, ih_next) {
869 		if (ih == handler)
870 			break;
871 	}
872 	if (ih == NULL) {
873 		panic("interrupt handler \"%s\" not found in "
874 		    "interrupt event \"%s\"", handler->ih_name, ie->ie_name);
875 	}
876 
877 	/*
878 	 * If there is no ithread, then directly remove the handler.  Note that
879 	 * intr_event_handle() iterates ie_handlers in a lock-less fashion, so
880 	 * care needs to be taken to keep ie_handlers consistent and to free
881 	 * the removed handler only when ie_handlers is quiescent.
882 	 */
883 	if (ie->ie_thread == NULL) {
884 		CK_SLIST_REMOVE_PREVPTR(prevptr, ih, ih_next);
885 		intr_event_barrier(ie);
886 		intr_event_update(ie);
887 		mtx_unlock(&ie->ie_lock);
888 		free(handler, M_ITHREAD);
889 		return (0);
890 	}
891 
892 	/*
893 	 * Let the interrupt thread do the job.
894 	 * The interrupt source is disabled when the interrupt thread is
895 	 * running, so it does not have to worry about interaction with
896 	 * intr_event_handle().
897 	 */
898 	KASSERT((handler->ih_flags & IH_DEAD) == 0,
899 	    ("duplicate handle remove"));
900 	handler->ih_flags |= IH_DEAD;
901 	intr_event_schedule_thread(ie, NULL);
902 	while (handler->ih_flags & IH_DEAD)
903 		msleep(handler, &ie->ie_lock, 0, "iev_rmh", 0);
904 	intr_event_update(ie);
905 
906 #ifdef notyet
907 	/*
908 	 * XXX: This could be bad in the case of ppbus(8).  Also, I think
909 	 * this could lead to races of stale data when servicing an
910 	 * interrupt.
911 	 */
912 	dead = 1;
913 	CK_SLIST_FOREACH(ih, &ie->ie_handlers, ih_next) {
914 		if (ih->ih_handler != NULL) {
915 			dead = 0;
916 			break;
917 		}
918 	}
919 	if (dead) {
920 		ithread_destroy(ie->ie_thread);
921 		ie->ie_thread = NULL;
922 	}
923 #endif
924 	mtx_unlock(&ie->ie_lock);
925 	free(handler, M_ITHREAD);
926 	return (0);
927 }
928 
929 int
930 intr_event_suspend_handler(void *cookie)
931 {
932 	struct intr_handler *handler = (struct intr_handler *)cookie;
933 	struct intr_event *ie;
934 
935 	if (handler == NULL)
936 		return (EINVAL);
937 	ie = handler->ih_event;
938 	KASSERT(ie != NULL,
939 	    ("interrupt handler \"%s\" has a NULL interrupt event",
940 	    handler->ih_name));
941 	mtx_lock(&ie->ie_lock);
942 	handler->ih_flags |= IH_SUSP;
943 	intr_handler_barrier(handler);
944 	mtx_unlock(&ie->ie_lock);
945 	return (0);
946 }
947 
948 int
949 intr_event_resume_handler(void *cookie)
950 {
951 	struct intr_handler *handler = (struct intr_handler *)cookie;
952 	struct intr_event *ie;
953 
954 	if (handler == NULL)
955 		return (EINVAL);
956 	ie = handler->ih_event;
957 	KASSERT(ie != NULL,
958 	    ("interrupt handler \"%s\" has a NULL interrupt event",
959 	    handler->ih_name));
960 
961 	/*
962 	 * intr_handler_barrier() acts not only as a barrier,
963 	 * it also allows to check for any pending interrupts.
964 	 */
965 	mtx_lock(&ie->ie_lock);
966 	handler->ih_flags &= ~IH_SUSP;
967 	intr_handler_barrier(handler);
968 	mtx_unlock(&ie->ie_lock);
969 	return (0);
970 }
971 
972 static int
973 intr_event_schedule_thread(struct intr_event *ie, struct trapframe *frame)
974 {
975 	struct intr_entropy entropy;
976 	struct intr_thread *it;
977 	struct thread *td;
978 	struct thread *ctd;
979 
980 	/*
981 	 * If no ithread or no handlers, then we have a stray interrupt.
982 	 */
983 	if (ie == NULL || CK_SLIST_EMPTY(&ie->ie_handlers) ||
984 	    ie->ie_thread == NULL)
985 		return (EINVAL);
986 
987 	ctd = curthread;
988 	it = ie->ie_thread;
989 	td = it->it_thread;
990 
991 	/*
992 	 * If any of the handlers for this ithread claim to be good
993 	 * sources of entropy, then gather some.
994 	 */
995 	if (ie->ie_hflags & IH_ENTROPY) {
996 		entropy.event = (uintptr_t)ie;
997 		entropy.td = ctd;
998 		random_harvest_queue(&entropy, sizeof(entropy), RANDOM_INTERRUPT);
999 	}
1000 
1001 	KASSERT(td->td_proc != NULL, ("ithread %s has no process", ie->ie_name));
1002 
1003 	/*
1004 	 * Set it_need to tell the thread to keep running if it is already
1005 	 * running.  Then, lock the thread and see if we actually need to
1006 	 * put it on the runqueue.
1007 	 *
1008 	 * Use store_rel to arrange that the store to ih_need in
1009 	 * swi_sched() is before the store to it_need and prepare for
1010 	 * transfer of this order to loads in the ithread.
1011 	 */
1012 	atomic_store_rel_int(&it->it_need, 1);
1013 	thread_lock(td);
1014 	if (TD_AWAITING_INTR(td)) {
1015 #ifdef HWPMC_HOOKS
1016 		it->it_waiting = 0;
1017 		if (PMC_HOOK_INSTALLED_ANY())
1018 			PMC_SOFT_CALL_INTR_HLPR(schedule, frame);
1019 #endif
1020 		CTR3(KTR_INTR, "%s: schedule pid %d (%s)", __func__, td->td_proc->p_pid,
1021 		    td->td_name);
1022 		TD_CLR_IWAIT(td);
1023 		sched_wakeup(td, SRQ_INTR);
1024 	} else {
1025 #ifdef HWPMC_HOOKS
1026 		it->it_waiting++;
1027 		if (PMC_HOOK_INSTALLED_ANY() &&
1028 		    (it->it_waiting >= intr_hwpmc_waiting_report_threshold))
1029 			PMC_SOFT_CALL_INTR_HLPR(waiting, frame);
1030 #endif
1031 		CTR5(KTR_INTR, "%s: pid %d (%s): it_need %d, state %d",
1032 		    __func__, td->td_proc->p_pid, td->td_name, it->it_need, TD_GET_STATE(td));
1033 		thread_unlock(td);
1034 	}
1035 
1036 	return (0);
1037 }
1038 
1039 /*
1040  * Allow interrupt event binding for software interrupt handlers -- a no-op,
1041  * since interrupts are generated in software rather than being directed by
1042  * a PIC.
1043  */
1044 static int
1045 swi_assign_cpu(void *arg, int cpu)
1046 {
1047 
1048 	return (0);
1049 }
1050 
1051 /*
1052  * Add a software interrupt handler to a specified event.  If a given event
1053  * is not specified, then a new event is created.
1054  */
1055 int
1056 swi_add(struct intr_event **eventp, const char *name, driver_intr_t handler,
1057 	    void *arg, int pri, enum intr_type flags, void **cookiep)
1058 {
1059 	struct intr_event *ie;
1060 	int error = 0;
1061 
1062 	if (flags & INTR_ENTROPY)
1063 		return (EINVAL);
1064 
1065 	ie = (eventp != NULL) ? *eventp : NULL;
1066 
1067 	if (ie != NULL) {
1068 		if (!(ie->ie_flags & IE_SOFT))
1069 			return (EINVAL);
1070 	} else {
1071 		error = intr_event_create(&ie, NULL, IE_SOFT, 0,
1072 		    NULL, NULL, NULL, swi_assign_cpu, "swi%d:", pri);
1073 		if (error)
1074 			return (error);
1075 		if (eventp != NULL)
1076 			*eventp = ie;
1077 	}
1078 	if (handler != NULL) {
1079 		error = intr_event_add_handler(ie, name, NULL, handler, arg,
1080 		    PI_SWI(pri), flags, cookiep);
1081 	}
1082 	return (error);
1083 }
1084 
1085 /*
1086  * Schedule a software interrupt thread.
1087  */
1088 void
1089 swi_sched(void *cookie, int flags)
1090 {
1091 	struct intr_handler *ih = (struct intr_handler *)cookie;
1092 	struct intr_event *ie = ih->ih_event;
1093 	struct intr_entropy entropy;
1094 	int error __unused;
1095 
1096 	CTR3(KTR_INTR, "swi_sched: %s %s need=%d", ie->ie_name, ih->ih_name,
1097 	    ih->ih_need);
1098 
1099 	if ((flags & SWI_FROMNMI) == 0) {
1100 		entropy.event = (uintptr_t)ih;
1101 		entropy.td = curthread;
1102 		random_harvest_queue(&entropy, sizeof(entropy), RANDOM_SWI);
1103 	}
1104 
1105 	/*
1106 	 * Set ih_need for this handler so that if the ithread is already
1107 	 * running it will execute this handler on the next pass.  Otherwise,
1108 	 * it will execute it the next time it runs.
1109 	 */
1110 	ih->ih_need = 1;
1111 
1112 	if (flags & SWI_DELAY)
1113 		return;
1114 
1115 	if (flags & SWI_FROMNMI) {
1116 #if defined(SMP) && (defined(__i386__) || defined(__amd64__))
1117 		KASSERT(ie == clk_intr_event,
1118 		    ("SWI_FROMNMI used not with clk_intr_event"));
1119 		ipi_self_from_nmi(IPI_SWI);
1120 #endif
1121 	} else {
1122 		VM_CNT_INC(v_soft);
1123 		error = intr_event_schedule_thread(ie, NULL);
1124 		KASSERT(error == 0, ("stray software interrupt"));
1125 	}
1126 }
1127 
1128 /*
1129  * Remove a software interrupt handler.  Currently this code does not
1130  * remove the associated interrupt event if it becomes empty.  Calling code
1131  * may do so manually via intr_event_destroy(), but that's not really
1132  * an optimal interface.
1133  */
1134 int
1135 swi_remove(void *cookie)
1136 {
1137 
1138 	return (intr_event_remove_handler(cookie));
1139 }
1140 
1141 static void
1142 intr_event_execute_handlers(struct proc *p, struct intr_event *ie)
1143 {
1144 	struct intr_handler *ih, *ihn, *ihp;
1145 
1146 	ihp = NULL;
1147 	CK_SLIST_FOREACH_SAFE(ih, &ie->ie_handlers, ih_next, ihn) {
1148 		/*
1149 		 * If this handler is marked for death, remove it from
1150 		 * the list of handlers and wake up the sleeper.
1151 		 */
1152 		if (ih->ih_flags & IH_DEAD) {
1153 			mtx_lock(&ie->ie_lock);
1154 			if (ihp == NULL)
1155 				CK_SLIST_REMOVE_HEAD(&ie->ie_handlers, ih_next);
1156 			else
1157 				CK_SLIST_REMOVE_AFTER(ihp, ih_next);
1158 			ih->ih_flags &= ~IH_DEAD;
1159 			wakeup(ih);
1160 			mtx_unlock(&ie->ie_lock);
1161 			continue;
1162 		}
1163 
1164 		/*
1165 		 * Now that we know that the current element won't be removed
1166 		 * update the previous element.
1167 		 */
1168 		ihp = ih;
1169 
1170 		if ((ih->ih_flags & IH_CHANGED) != 0) {
1171 			mtx_lock(&ie->ie_lock);
1172 			ih->ih_flags &= ~IH_CHANGED;
1173 			wakeup(ih);
1174 			mtx_unlock(&ie->ie_lock);
1175 		}
1176 
1177 		/* Skip filter only handlers */
1178 		if (ih->ih_handler == NULL)
1179 			continue;
1180 
1181 		/* Skip suspended handlers */
1182 		if ((ih->ih_flags & IH_SUSP) != 0)
1183 			continue;
1184 
1185 		/*
1186 		 * For software interrupt threads, we only execute
1187 		 * handlers that have their need flag set.  Hardware
1188 		 * interrupt threads always invoke all of their handlers.
1189 		 *
1190 		 * ih_need can only be 0 or 1.  Failed cmpset below
1191 		 * means that there is no request to execute handlers,
1192 		 * so a retry of the cmpset is not needed.
1193 		 */
1194 		if ((ie->ie_flags & IE_SOFT) != 0 &&
1195 		    atomic_cmpset_int(&ih->ih_need, 1, 0) == 0)
1196 			continue;
1197 
1198 		/* Execute this handler. */
1199 		CTR6(KTR_INTR, "%s: pid %d exec %p(%p) for %s flg=%x",
1200 		    __func__, p->p_pid, (void *)ih->ih_handler,
1201 		    ih->ih_argument, ih->ih_name, ih->ih_flags);
1202 
1203 		if (!(ih->ih_flags & IH_MPSAFE))
1204 			mtx_lock(&Giant);
1205 		ih->ih_handler(ih->ih_argument);
1206 		if (!(ih->ih_flags & IH_MPSAFE))
1207 			mtx_unlock(&Giant);
1208 	}
1209 }
1210 
1211 static void
1212 ithread_execute_handlers(struct proc *p, struct intr_event *ie)
1213 {
1214 
1215 	/* Interrupt handlers should not sleep. */
1216 	if (!(ie->ie_flags & IE_SOFT))
1217 		THREAD_NO_SLEEPING();
1218 	intr_event_execute_handlers(p, ie);
1219 	if (!(ie->ie_flags & IE_SOFT))
1220 		THREAD_SLEEPING_OK();
1221 
1222 	/*
1223 	 * Interrupt storm handling:
1224 	 *
1225 	 * If this interrupt source is currently storming, then throttle
1226 	 * it to only fire the handler once  per clock tick.
1227 	 *
1228 	 * If this interrupt source is not currently storming, but the
1229 	 * number of back to back interrupts exceeds the storm threshold,
1230 	 * then enter storming mode.
1231 	 */
1232 	if (intr_storm_threshold != 0 && ie->ie_count >= intr_storm_threshold &&
1233 	    !(ie->ie_flags & IE_SOFT)) {
1234 		/* Report the message only once every second. */
1235 		if (ppsratecheck(&ie->ie_warntm, &ie->ie_warncnt, 1)) {
1236 			printf(
1237 	"interrupt storm detected on \"%s\"; throttling interrupt source\n",
1238 			    ie->ie_name);
1239 		}
1240 		pause("istorm", 1);
1241 	} else
1242 		ie->ie_count++;
1243 
1244 	/*
1245 	 * Now that all the handlers have had a chance to run, reenable
1246 	 * the interrupt source.
1247 	 */
1248 	if (ie->ie_post_ithread != NULL)
1249 		ie->ie_post_ithread(ie->ie_source);
1250 }
1251 
1252 /*
1253  * This is the main code for interrupt threads.
1254  */
1255 static void
1256 ithread_loop(void *arg)
1257 {
1258 	struct epoch_tracker et;
1259 	struct intr_thread *ithd;
1260 	struct intr_event *ie;
1261 	struct thread *td;
1262 	struct proc *p;
1263 	int wake, epoch_count;
1264 	bool needs_epoch;
1265 
1266 	td = curthread;
1267 	p = td->td_proc;
1268 	ithd = (struct intr_thread *)arg;
1269 	KASSERT(ithd->it_thread == td,
1270 	    ("%s: ithread and proc linkage out of sync", __func__));
1271 	ie = ithd->it_event;
1272 	ie->ie_count = 0;
1273 	wake = 0;
1274 
1275 	/*
1276 	 * As long as we have interrupts outstanding, go through the
1277 	 * list of handlers, giving each one a go at it.
1278 	 */
1279 	for (;;) {
1280 		/*
1281 		 * If we are an orphaned thread, then just die.
1282 		 */
1283 		if (ithd->it_flags & IT_DEAD) {
1284 			CTR3(KTR_INTR, "%s: pid %d (%s) exiting", __func__,
1285 			    p->p_pid, td->td_name);
1286 			free(ithd, M_ITHREAD);
1287 			kthread_exit();
1288 		}
1289 
1290 		/*
1291 		 * Service interrupts.  If another interrupt arrives while
1292 		 * we are running, it will set it_need to note that we
1293 		 * should make another pass.
1294 		 *
1295 		 * The load_acq part of the following cmpset ensures
1296 		 * that the load of ih_need in ithread_execute_handlers()
1297 		 * is ordered after the load of it_need here.
1298 		 */
1299 		needs_epoch =
1300 		    (atomic_load_int(&ie->ie_hflags) & IH_NET) != 0;
1301 		if (needs_epoch) {
1302 			epoch_count = 0;
1303 			NET_EPOCH_ENTER(et);
1304 		}
1305 		while (atomic_cmpset_acq_int(&ithd->it_need, 1, 0) != 0) {
1306 			ithread_execute_handlers(p, ie);
1307 			if (needs_epoch &&
1308 			    ++epoch_count >= intr_epoch_batch) {
1309 				NET_EPOCH_EXIT(et);
1310 				epoch_count = 0;
1311 				NET_EPOCH_ENTER(et);
1312 			}
1313 		}
1314 		if (needs_epoch)
1315 			NET_EPOCH_EXIT(et);
1316 		WITNESS_WARN(WARN_PANIC, NULL, "suspending ithread");
1317 		mtx_assert(&Giant, MA_NOTOWNED);
1318 
1319 		/*
1320 		 * Processed all our interrupts.  Now get the sched
1321 		 * lock.  This may take a while and it_need may get
1322 		 * set again, so we have to check it again.
1323 		 */
1324 		thread_lock(td);
1325 		if (atomic_load_acq_int(&ithd->it_need) == 0 &&
1326 		    (ithd->it_flags & (IT_DEAD | IT_WAIT)) == 0) {
1327 			TD_SET_IWAIT(td);
1328 			ie->ie_count = 0;
1329 			mi_switch(SW_VOL | SWT_IWAIT);
1330 		} else {
1331 			if (ithd->it_flags & IT_WAIT) {
1332 				wake = 1;
1333 				ithd->it_flags &= ~IT_WAIT;
1334 			}
1335 			thread_unlock(td);
1336 		}
1337 		if (wake) {
1338 			wakeup(ithd);
1339 			wake = 0;
1340 		}
1341 	}
1342 }
1343 
1344 /*
1345  * Main interrupt handling body.
1346  *
1347  * Input:
1348  * o ie:                        the event connected to this interrupt.
1349  * o frame:                     some archs (i.e. i386) pass a frame to some.
1350  *                              handlers as their main argument.
1351  * Return value:
1352  * o 0:                         everything ok.
1353  * o EINVAL:                    stray interrupt.
1354  */
1355 int
1356 intr_event_handle(struct intr_event *ie, struct trapframe *frame)
1357 {
1358 	struct intr_handler *ih;
1359 	struct trapframe *oldframe;
1360 	struct thread *td;
1361 	int phase;
1362 	int ret;
1363 	bool filter, thread;
1364 
1365 	td = curthread;
1366 
1367 #ifdef KSTACK_USAGE_PROF
1368 	intr_prof_stack_use(td, frame);
1369 #endif
1370 
1371 	/* An interrupt with no event or handlers is a stray interrupt. */
1372 	if (ie == NULL || CK_SLIST_EMPTY(&ie->ie_handlers))
1373 		return (EINVAL);
1374 
1375 	/*
1376 	 * Execute fast interrupt handlers directly.
1377 	 * To support clock handlers, if a handler registers
1378 	 * with a NULL argument, then we pass it a pointer to
1379 	 * a trapframe as its argument.
1380 	 */
1381 	td->td_intr_nesting_level++;
1382 	filter = false;
1383 	thread = false;
1384 	ret = 0;
1385 	critical_enter();
1386 	oldframe = td->td_intr_frame;
1387 	td->td_intr_frame = frame;
1388 
1389 	phase = ie->ie_phase;
1390 	atomic_add_int(&ie->ie_active[phase], 1);
1391 
1392 	/*
1393 	 * This fence is required to ensure that no later loads are
1394 	 * re-ordered before the ie_active store.
1395 	 */
1396 	atomic_thread_fence_seq_cst();
1397 
1398 	CK_SLIST_FOREACH(ih, &ie->ie_handlers, ih_next) {
1399 		if ((ih->ih_flags & IH_SUSP) != 0)
1400 			continue;
1401 		if ((ie->ie_flags & IE_SOFT) != 0 && ih->ih_need == 0)
1402 			continue;
1403 		if (ih->ih_filter == NULL) {
1404 			thread = true;
1405 			continue;
1406 		}
1407 		CTR4(KTR_INTR, "%s: exec %p(%p) for %s", __func__,
1408 		    ih->ih_filter, ih->ih_argument == NULL ? frame :
1409 		    ih->ih_argument, ih->ih_name);
1410 		if (ih->ih_argument == NULL)
1411 			ret = ih->ih_filter(frame);
1412 		else
1413 			ret = ih->ih_filter(ih->ih_argument);
1414 #ifdef HWPMC_HOOKS
1415 		PMC_SOFT_CALL_TF( , , intr, all, frame);
1416 #endif
1417 		KASSERT(ret == FILTER_STRAY ||
1418 		    ((ret & (FILTER_SCHEDULE_THREAD | FILTER_HANDLED)) != 0 &&
1419 		    (ret & ~(FILTER_SCHEDULE_THREAD | FILTER_HANDLED)) == 0),
1420 		    ("%s: incorrect return value %#x from %s", __func__, ret,
1421 		    ih->ih_name));
1422 		filter = filter || ret == FILTER_HANDLED;
1423 #ifdef HWPMC_HOOKS
1424 		if (ret & FILTER_SCHEDULE_THREAD)
1425 			PMC_SOFT_CALL_TF( , , intr, ithread, frame);
1426 		else if (ret & FILTER_HANDLED)
1427 			PMC_SOFT_CALL_TF( , , intr, filter, frame);
1428 		else if (ret == FILTER_STRAY)
1429 			PMC_SOFT_CALL_TF( , , intr, stray, frame);
1430 #endif
1431 
1432 		/*
1433 		 * Wrapper handler special handling:
1434 		 *
1435 		 * in some particular cases (like pccard and pccbb),
1436 		 * the _real_ device handler is wrapped in a couple of
1437 		 * functions - a filter wrapper and an ithread wrapper.
1438 		 * In this case (and just in this case), the filter wrapper
1439 		 * could ask the system to schedule the ithread and mask
1440 		 * the interrupt source if the wrapped handler is composed
1441 		 * of just an ithread handler.
1442 		 *
1443 		 * TODO: write a generic wrapper to avoid people rolling
1444 		 * their own.
1445 		 */
1446 		if (!thread) {
1447 			if (ret == FILTER_SCHEDULE_THREAD)
1448 				thread = true;
1449 		}
1450 	}
1451 	atomic_add_rel_int(&ie->ie_active[phase], -1);
1452 
1453 	td->td_intr_frame = oldframe;
1454 
1455 	if (thread) {
1456 		if (ie->ie_pre_ithread != NULL)
1457 			ie->ie_pre_ithread(ie->ie_source);
1458 	} else {
1459 		if (ie->ie_post_filter != NULL)
1460 			ie->ie_post_filter(ie->ie_source);
1461 	}
1462 
1463 	/* Schedule the ithread if needed. */
1464 	if (thread) {
1465 		int error __unused;
1466 
1467 		error =  intr_event_schedule_thread(ie, frame);
1468 		KASSERT(error == 0, ("bad stray interrupt"));
1469 	}
1470 	critical_exit();
1471 	td->td_intr_nesting_level--;
1472 #ifdef notyet
1473 	/* The interrupt is not aknowledged by any filter and has no ithread. */
1474 	if (!thread && !filter)
1475 		return (EINVAL);
1476 #endif
1477 	return (0);
1478 }
1479 
1480 #ifdef DDB
1481 /*
1482  * Dump details about an interrupt handler
1483  */
1484 static void
1485 db_dump_intrhand(struct intr_handler *ih)
1486 {
1487 	int comma;
1488 
1489 	db_printf("\t%-10s ", ih->ih_name);
1490 	switch (ih->ih_pri) {
1491 	case PI_REALTIME:
1492 		db_printf("CLK ");
1493 		break;
1494 	case PI_INTR:
1495 		db_printf("INTR");
1496 		break;
1497 	default:
1498 		if (ih->ih_pri >= PI_SOFT)
1499 			db_printf("SWI ");
1500 		else
1501 			db_printf("%4u", ih->ih_pri);
1502 		break;
1503 	}
1504 	db_printf(" ");
1505 	if (ih->ih_filter != NULL) {
1506 		db_printf("[F]");
1507 		db_printsym((uintptr_t)ih->ih_filter, DB_STGY_PROC);
1508 	}
1509 	if (ih->ih_handler != NULL) {
1510 		if (ih->ih_filter != NULL)
1511 			db_printf(",");
1512 		db_printf("[H]");
1513 		db_printsym((uintptr_t)ih->ih_handler, DB_STGY_PROC);
1514 	}
1515 	db_printf("(%p)", ih->ih_argument);
1516 	if (ih->ih_need ||
1517 	    (ih->ih_flags & (IH_EXCLUSIVE | IH_ENTROPY | IH_DEAD |
1518 	    IH_MPSAFE)) != 0) {
1519 		db_printf(" {");
1520 		comma = 0;
1521 		if (ih->ih_flags & IH_EXCLUSIVE) {
1522 			if (comma)
1523 				db_printf(", ");
1524 			db_printf("EXCL");
1525 			comma = 1;
1526 		}
1527 		if (ih->ih_flags & IH_ENTROPY) {
1528 			if (comma)
1529 				db_printf(", ");
1530 			db_printf("ENTROPY");
1531 			comma = 1;
1532 		}
1533 		if (ih->ih_flags & IH_DEAD) {
1534 			if (comma)
1535 				db_printf(", ");
1536 			db_printf("DEAD");
1537 			comma = 1;
1538 		}
1539 		if (ih->ih_flags & IH_MPSAFE) {
1540 			if (comma)
1541 				db_printf(", ");
1542 			db_printf("MPSAFE");
1543 			comma = 1;
1544 		}
1545 		if (ih->ih_need) {
1546 			if (comma)
1547 				db_printf(", ");
1548 			db_printf("NEED");
1549 		}
1550 		db_printf("}");
1551 	}
1552 	db_printf("\n");
1553 }
1554 
1555 /*
1556  * Dump details about a event.
1557  */
1558 void
1559 db_dump_intr_event(struct intr_event *ie, int handlers)
1560 {
1561 	struct intr_handler *ih;
1562 	struct intr_thread *it;
1563 	int comma;
1564 
1565 	db_printf("%s ", ie->ie_fullname);
1566 	it = ie->ie_thread;
1567 	if (it != NULL)
1568 		db_printf("(pid %d)", it->it_thread->td_proc->p_pid);
1569 	else
1570 		db_printf("(no thread)");
1571 	if ((ie->ie_flags & (IE_SOFT | IE_ADDING_THREAD)) != 0 ||
1572 	    (it != NULL && it->it_need)) {
1573 		db_printf(" {");
1574 		comma = 0;
1575 		if (ie->ie_flags & IE_SOFT) {
1576 			db_printf("SOFT");
1577 			comma = 1;
1578 		}
1579 		if (ie->ie_flags & IE_ADDING_THREAD) {
1580 			if (comma)
1581 				db_printf(", ");
1582 			db_printf("ADDING_THREAD");
1583 			comma = 1;
1584 		}
1585 		if (it != NULL && it->it_need) {
1586 			if (comma)
1587 				db_printf(", ");
1588 			db_printf("NEED");
1589 		}
1590 		db_printf("}");
1591 	}
1592 	db_printf("\n");
1593 
1594 	if (handlers)
1595 		CK_SLIST_FOREACH(ih, &ie->ie_handlers, ih_next)
1596 		    db_dump_intrhand(ih);
1597 }
1598 
1599 /*
1600  * Dump data about interrupt handlers
1601  */
1602 DB_SHOW_COMMAND_FLAGS(intr, db_show_intr, DB_CMD_MEMSAFE)
1603 {
1604 	struct intr_event *ie;
1605 	int all, verbose;
1606 
1607 	verbose = strchr(modif, 'v') != NULL;
1608 	all = strchr(modif, 'a') != NULL;
1609 	TAILQ_FOREACH(ie, &event_list, ie_list) {
1610 		if (!all && CK_SLIST_EMPTY(&ie->ie_handlers))
1611 			continue;
1612 		db_dump_intr_event(ie, verbose);
1613 		if (db_pager_quit)
1614 			break;
1615 	}
1616 }
1617 #endif /* DDB */
1618 
1619 /*
1620  * Start standard software interrupt threads
1621  */
1622 static void
1623 start_softintr(void *dummy)
1624 {
1625 
1626 	if (swi_add(&clk_intr_event, "clk", NULL, NULL, SWI_CLOCK,
1627 	    INTR_MPSAFE, NULL))
1628 		panic("died while creating clk swi ithread");
1629 }
1630 SYSINIT(start_softintr, SI_SUB_SOFTINTR, SI_ORDER_FIRST, start_softintr,
1631     NULL);
1632 
1633 /*
1634  * Sysctls used by systat and others: hw.intrnames and hw.intrcnt.
1635  * The data for this machine dependent, and the declarations are in machine
1636  * dependent code.  The layout of intrnames and intrcnt however is machine
1637  * independent.
1638  *
1639  * We do not know the length of intrcnt and intrnames at compile time, so
1640  * calculate things at run time.
1641  */
1642 static int
1643 sysctl_intrnames(SYSCTL_HANDLER_ARGS)
1644 {
1645 	return (sysctl_handle_opaque(oidp, intrnames, sintrnames, req));
1646 }
1647 
1648 SYSCTL_PROC(_hw, OID_AUTO, intrnames,
1649     CTLTYPE_OPAQUE | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
1650     sysctl_intrnames, "",
1651     "Interrupt Names");
1652 
1653 static int
1654 sysctl_intrcnt(SYSCTL_HANDLER_ARGS)
1655 {
1656 #ifdef SCTL_MASK32
1657 	uint32_t *intrcnt32;
1658 	unsigned i;
1659 	int error;
1660 
1661 	if (req->flags & SCTL_MASK32) {
1662 		if (!req->oldptr)
1663 			return (sysctl_handle_opaque(oidp, NULL, sintrcnt / 2, req));
1664 		intrcnt32 = malloc(sintrcnt / 2, M_TEMP, M_NOWAIT);
1665 		if (intrcnt32 == NULL)
1666 			return (ENOMEM);
1667 		for (i = 0; i < sintrcnt / sizeof (u_long); i++)
1668 			intrcnt32[i] = intrcnt[i];
1669 		error = sysctl_handle_opaque(oidp, intrcnt32, sintrcnt / 2, req);
1670 		free(intrcnt32, M_TEMP);
1671 		return (error);
1672 	}
1673 #endif
1674 	return (sysctl_handle_opaque(oidp, intrcnt, sintrcnt, req));
1675 }
1676 
1677 SYSCTL_PROC(_hw, OID_AUTO, intrcnt,
1678     CTLTYPE_OPAQUE | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
1679     sysctl_intrcnt, "",
1680     "Interrupt Counts");
1681 
1682 #ifdef DDB
1683 /*
1684  * DDB command to dump the interrupt statistics.
1685  */
1686 DB_SHOW_COMMAND_FLAGS(intrcnt, db_show_intrcnt, DB_CMD_MEMSAFE)
1687 {
1688 	u_long *i;
1689 	char *cp;
1690 	u_int j;
1691 
1692 	cp = intrnames;
1693 	j = 0;
1694 	for (i = intrcnt; j < (sintrcnt / sizeof(u_long)) && !db_pager_quit;
1695 	    i++, j++) {
1696 		if (*cp == '\0')
1697 			break;
1698 		if (*i != 0)
1699 			db_printf("%s\t%lu\n", cp, *i);
1700 		cp += strlen(cp) + 1;
1701 	}
1702 }
1703 #endif
1704