xref: /freebsd/sys/kern/kern_intr.c (revision 8381e9f49ec733437754a822ef2e8344115289ac)
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, u_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, const void *m)
447 {
448 	struct intr_event *ie;
449 	const 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 	mtx_unlock(&event_lock);
545 	if (ie->ie_thread != NULL)
546 		ithread_destroy(ie->ie_thread);
547 	mtx_unlock(&ie->ie_lock);
548 	mtx_destroy(&ie->ie_lock);
549 	free(ie, M_ITHREAD);
550 	return (0);
551 }
552 
553 static struct intr_thread *
554 ithread_create(const char *name)
555 {
556 	struct intr_thread *ithd;
557 	struct thread *td;
558 	int error;
559 
560 	ithd = malloc(sizeof(struct intr_thread), M_ITHREAD, M_WAITOK | M_ZERO);
561 
562 	error = kproc_kthread_add(ithread_loop, ithd, &intrproc,
563 		    &td, RFSTOPPED | RFHIGHPID,
564 		    0, "intr", "%s", name);
565 	if (error)
566 		panic("kproc_create() failed with %d", error);
567 	thread_lock(td);
568 	sched_class(td, PRI_ITHD);
569 	TD_SET_IWAIT(td);
570 	thread_unlock(td);
571 	td->td_pflags |= TDP_ITHREAD;
572 	ithd->it_thread = td;
573 	CTR2(KTR_INTR, "%s: created %s", __func__, name);
574 	return (ithd);
575 }
576 
577 static void
578 ithread_destroy(struct intr_thread *ithread)
579 {
580 	struct intr_event *ie;
581 	struct thread *td;
582 
583 	td = ithread->it_thread;
584 	ie = ithread->it_event;
585 
586 	mtx_assert(&ie->ie_lock, MA_OWNED);
587 
588 	CTR2(KTR_INTR, "%s: killing %s", __func__, ie->ie_name);
589 
590 	thread_lock(td);
591 	ithread->it_flags |= IT_DEAD;
592 	if (TD_AWAITING_INTR(td)) {
593 		TD_CLR_IWAIT(td);
594 		sched_wakeup(td, SRQ_INTR);
595 	} else
596 		thread_unlock(td);
597 	while (ie->ie_thread != NULL)
598 		msleep(ithread, &ie->ie_lock, 0, "ithd_dth", 0);
599 }
600 
601 int
602 intr_event_add_handler(struct intr_event *ie, const char *name,
603     driver_filter_t filter, driver_intr_t handler, void *arg, u_char pri,
604     enum intr_type flags, void **cookiep)
605 {
606 	struct intr_handler *ih, *temp_ih;
607 	struct intr_handler **prevptr;
608 	struct intr_thread *it;
609 
610 	if (ie == NULL || name == NULL || (handler == NULL && filter == NULL))
611 		return (EINVAL);
612 
613 	/* Allocate and populate an interrupt handler structure. */
614 	ih = malloc(sizeof(struct intr_handler), M_ITHREAD, M_WAITOK | M_ZERO);
615 	ih->ih_filter = filter;
616 	ih->ih_handler = handler;
617 	ih->ih_argument = arg;
618 	strlcpy(ih->ih_name, name, sizeof(ih->ih_name));
619 	ih->ih_event = ie;
620 	ih->ih_pri = pri;
621 	if (flags & INTR_EXCL)
622 		ih->ih_flags = IH_EXCLUSIVE;
623 	if (flags & INTR_MPSAFE)
624 		ih->ih_flags |= IH_MPSAFE;
625 	if (flags & INTR_ENTROPY)
626 		ih->ih_flags |= IH_ENTROPY;
627 	if (flags & INTR_TYPE_NET)
628 		ih->ih_flags |= IH_NET;
629 
630 	/* We can only have one exclusive handler in a event. */
631 	mtx_lock(&ie->ie_lock);
632 	if (!CK_SLIST_EMPTY(&ie->ie_handlers)) {
633 		if ((flags & INTR_EXCL) ||
634 		    (CK_SLIST_FIRST(&ie->ie_handlers)->ih_flags & IH_EXCLUSIVE)) {
635 			mtx_unlock(&ie->ie_lock);
636 			free(ih, M_ITHREAD);
637 			return (EINVAL);
638 		}
639 	}
640 
641 	/* Create a thread if we need one. */
642 	while (ie->ie_thread == NULL && handler != NULL) {
643 		if (ie->ie_flags & IE_ADDING_THREAD)
644 			msleep(ie, &ie->ie_lock, 0, "ithread", 0);
645 		else {
646 			ie->ie_flags |= IE_ADDING_THREAD;
647 			mtx_unlock(&ie->ie_lock);
648 			it = ithread_create("intr: newborn");
649 			mtx_lock(&ie->ie_lock);
650 			ie->ie_flags &= ~IE_ADDING_THREAD;
651 			ie->ie_thread = it;
652 			it->it_event = ie;
653 			ithread_update(it);
654 			wakeup(ie);
655 		}
656 	}
657 
658 	/* Add the new handler to the event in priority order. */
659 	CK_SLIST_FOREACH_PREVPTR(temp_ih, prevptr, &ie->ie_handlers, ih_next) {
660 		if (temp_ih->ih_pri > ih->ih_pri)
661 			break;
662 	}
663 	CK_SLIST_INSERT_PREVPTR(prevptr, temp_ih, ih, ih_next);
664 
665 	intr_event_update(ie);
666 
667 	CTR3(KTR_INTR, "%s: added %s to %s", __func__, ih->ih_name,
668 	    ie->ie_name);
669 	mtx_unlock(&ie->ie_lock);
670 
671 	if (cookiep != NULL)
672 		*cookiep = ih;
673 	return (0);
674 }
675 
676 /*
677  * Append a description preceded by a ':' to the name of the specified
678  * interrupt handler.
679  */
680 int
681 intr_event_describe_handler(struct intr_event *ie, void *cookie,
682     const char *descr)
683 {
684 	struct intr_handler *ih;
685 	size_t space;
686 	char *start;
687 
688 	mtx_lock(&ie->ie_lock);
689 #ifdef INVARIANTS
690 	CK_SLIST_FOREACH(ih, &ie->ie_handlers, ih_next) {
691 		if (ih == cookie)
692 			break;
693 	}
694 	if (ih == NULL) {
695 		mtx_unlock(&ie->ie_lock);
696 		panic("handler %p not found in interrupt event %p", cookie, ie);
697 	}
698 #endif
699 	ih = cookie;
700 
701 	/*
702 	 * Look for an existing description by checking for an
703 	 * existing ":".  This assumes device names do not include
704 	 * colons.  If one is found, prepare to insert the new
705 	 * description at that point.  If one is not found, find the
706 	 * end of the name to use as the insertion point.
707 	 */
708 	start = strchr(ih->ih_name, ':');
709 	if (start == NULL)
710 		start = strchr(ih->ih_name, 0);
711 
712 	/*
713 	 * See if there is enough remaining room in the string for the
714 	 * description + ":".  The "- 1" leaves room for the trailing
715 	 * '\0'.  The "+ 1" accounts for the colon.
716 	 */
717 	space = sizeof(ih->ih_name) - (start - ih->ih_name) - 1;
718 	if (strlen(descr) + 1 > space) {
719 		mtx_unlock(&ie->ie_lock);
720 		return (ENOSPC);
721 	}
722 
723 	/* Append a colon followed by the description. */
724 	*start = ':';
725 	strcpy(start + 1, descr);
726 	intr_event_update(ie);
727 	mtx_unlock(&ie->ie_lock);
728 	return (0);
729 }
730 
731 /*
732  * Return the ie_source field from the intr_event an intr_handler is
733  * associated with.
734  */
735 void *
736 intr_handler_source(void *cookie)
737 {
738 	struct intr_handler *ih;
739 	struct intr_event *ie;
740 
741 	ih = (struct intr_handler *)cookie;
742 	if (ih == NULL)
743 		return (NULL);
744 	ie = ih->ih_event;
745 	KASSERT(ie != NULL,
746 	    ("interrupt handler \"%s\" has a NULL interrupt event",
747 	    ih->ih_name));
748 	return (ie->ie_source);
749 }
750 
751 /*
752  * If intr_event_handle() is running in the ISR context at the time of the call,
753  * then wait for it to complete.
754  */
755 static void
756 intr_event_barrier(struct intr_event *ie)
757 {
758 	int phase;
759 
760 	mtx_assert(&ie->ie_lock, MA_OWNED);
761 	phase = ie->ie_phase;
762 
763 	/*
764 	 * Switch phase to direct future interrupts to the other active counter.
765 	 * Make sure that any preceding stores are visible before the switch.
766 	 */
767 	KASSERT(ie->ie_active[!phase] == 0, ("idle phase has activity"));
768 	atomic_store_rel_int(&ie->ie_phase, !phase);
769 
770 	/*
771 	 * This code cooperates with wait-free iteration of ie_handlers
772 	 * in intr_event_handle.
773 	 * Make sure that the removal and the phase update are not reordered
774 	 * with the active count check.
775 	 * Note that no combination of acquire and release fences can provide
776 	 * that guarantee as Store->Load sequences can always be reordered.
777 	 */
778 	atomic_thread_fence_seq_cst();
779 
780 	/*
781 	 * Now wait on the inactive phase.
782 	 * The acquire fence is needed so that all post-barrier accesses
783 	 * are after the check.
784 	 */
785 	while (ie->ie_active[phase] > 0)
786 		cpu_spinwait();
787 	atomic_thread_fence_acq();
788 }
789 
790 static void
791 intr_handler_barrier(struct intr_handler *handler)
792 {
793 	struct intr_event *ie;
794 
795 	ie = handler->ih_event;
796 	mtx_assert(&ie->ie_lock, MA_OWNED);
797 	KASSERT((handler->ih_flags & IH_DEAD) == 0,
798 	    ("update for a removed handler"));
799 
800 	if (ie->ie_thread == NULL) {
801 		intr_event_barrier(ie);
802 		return;
803 	}
804 	if ((handler->ih_flags & IH_CHANGED) == 0) {
805 		handler->ih_flags |= IH_CHANGED;
806 		intr_event_schedule_thread(ie, NULL);
807 	}
808 	while ((handler->ih_flags & IH_CHANGED) != 0)
809 		msleep(handler, &ie->ie_lock, 0, "ih_barr", 0);
810 }
811 
812 /*
813  * Sleep until an ithread finishes executing an interrupt handler.
814  *
815  * XXX Doesn't currently handle interrupt filters or fast interrupt
816  * handlers. This is intended for LinuxKPI drivers only.
817  * Do not use in BSD code.
818  */
819 void
820 _intr_drain(int irq)
821 {
822 	struct intr_event *ie;
823 	struct intr_thread *ithd;
824 	struct thread *td;
825 
826 	ie = intr_lookup(irq);
827 	if (ie == NULL)
828 		return;
829 	if (ie->ie_thread == NULL)
830 		return;
831 	ithd = ie->ie_thread;
832 	td = ithd->it_thread;
833 	/*
834 	 * We set the flag and wait for it to be cleared to avoid
835 	 * long delays with potentially busy interrupt handlers
836 	 * were we to only sample TD_AWAITING_INTR() every tick.
837 	 */
838 	thread_lock(td);
839 	if (!TD_AWAITING_INTR(td)) {
840 		ithd->it_flags |= IT_WAIT;
841 		while (ithd->it_flags & IT_WAIT) {
842 			thread_unlock(td);
843 			pause("idrain", 1);
844 			thread_lock(td);
845 		}
846 	}
847 	thread_unlock(td);
848 	return;
849 }
850 
851 int
852 intr_event_remove_handler(void *cookie)
853 {
854 	struct intr_handler *handler = (struct intr_handler *)cookie;
855 	struct intr_event *ie;
856 	struct intr_handler *ih;
857 	struct intr_handler **prevptr;
858 
859 	if (handler == NULL)
860 		return (EINVAL);
861 	ie = handler->ih_event;
862 	KASSERT(ie != NULL,
863 	    ("interrupt handler \"%s\" has a NULL interrupt event",
864 	    handler->ih_name));
865 
866 	mtx_lock(&ie->ie_lock);
867 	CTR3(KTR_INTR, "%s: removing %s from %s", __func__, handler->ih_name,
868 	    ie->ie_name);
869 	CK_SLIST_FOREACH_PREVPTR(ih, prevptr, &ie->ie_handlers, ih_next) {
870 		if (ih == handler)
871 			break;
872 	}
873 	if (ih == NULL) {
874 		panic("interrupt handler \"%s\" not found in "
875 		    "interrupt event \"%s\"", handler->ih_name, ie->ie_name);
876 	}
877 
878 	if (ie->ie_thread == NULL) {
879 		/*
880 		 * If there is no ithread, then directly remove the handler.
881 		 * Note that intr_event_handle() iterates ie_handlers in a
882 		 * lock-less fashion, so care needs to be taken to keep
883 		 * ie_handlers consistent and to free the removed handler only
884 		 * when ie_handlers is quiescent.
885 		 */
886 		CK_SLIST_REMOVE_PREVPTR(prevptr, ih, ih_next);
887 		intr_event_barrier(ie);
888 	} else {
889 		/*
890 		 * Let the interrupt thread do the job.  The interrupt source is
891 		 * disabled when the interrupt thread is running, so it does not
892 		 * have to worry about interaction with intr_event_handle().
893 		 */
894 		KASSERT((handler->ih_flags & IH_DEAD) == 0,
895 		    ("duplicate handle remove"));
896 		handler->ih_flags |= IH_DEAD;
897 		intr_event_schedule_thread(ie, NULL);
898 		while (handler->ih_flags & IH_DEAD)
899 			msleep(handler, &ie->ie_lock, 0, "iev_rmh", 0);
900 	}
901 	intr_event_update(ie);
902 	mtx_unlock(&ie->ie_lock);
903 	free(handler, M_ITHREAD);
904 	return (0);
905 }
906 
907 int
908 intr_event_suspend_handler(void *cookie)
909 {
910 	struct intr_handler *handler = (struct intr_handler *)cookie;
911 	struct intr_event *ie;
912 
913 	if (handler == NULL)
914 		return (EINVAL);
915 	ie = handler->ih_event;
916 	KASSERT(ie != NULL,
917 	    ("interrupt handler \"%s\" has a NULL interrupt event",
918 	    handler->ih_name));
919 	mtx_lock(&ie->ie_lock);
920 	handler->ih_flags |= IH_SUSP;
921 	intr_handler_barrier(handler);
922 	mtx_unlock(&ie->ie_lock);
923 	return (0);
924 }
925 
926 int
927 intr_event_resume_handler(void *cookie)
928 {
929 	struct intr_handler *handler = (struct intr_handler *)cookie;
930 	struct intr_event *ie;
931 
932 	if (handler == NULL)
933 		return (EINVAL);
934 	ie = handler->ih_event;
935 	KASSERT(ie != NULL,
936 	    ("interrupt handler \"%s\" has a NULL interrupt event",
937 	    handler->ih_name));
938 
939 	/*
940 	 * intr_handler_barrier() acts not only as a barrier,
941 	 * it also allows to check for any pending interrupts.
942 	 */
943 	mtx_lock(&ie->ie_lock);
944 	handler->ih_flags &= ~IH_SUSP;
945 	intr_handler_barrier(handler);
946 	mtx_unlock(&ie->ie_lock);
947 	return (0);
948 }
949 
950 static int
951 intr_event_schedule_thread(struct intr_event *ie, struct trapframe *frame)
952 {
953 	struct intr_entropy entropy;
954 	struct intr_thread *it;
955 	struct thread *td;
956 	struct thread *ctd;
957 
958 	/*
959 	 * If no ithread or no handlers, then we have a stray interrupt.
960 	 */
961 	if (ie == NULL || CK_SLIST_EMPTY(&ie->ie_handlers) ||
962 	    ie->ie_thread == NULL)
963 		return (EINVAL);
964 
965 	ctd = curthread;
966 	it = ie->ie_thread;
967 	td = it->it_thread;
968 
969 	/*
970 	 * If any of the handlers for this ithread claim to be good
971 	 * sources of entropy, then gather some.
972 	 */
973 	if (ie->ie_hflags & IH_ENTROPY) {
974 		entropy.event = (uintptr_t)ie;
975 		entropy.td = ctd;
976 		random_harvest_queue(&entropy, sizeof(entropy), RANDOM_INTERRUPT);
977 	}
978 
979 	KASSERT(td->td_proc != NULL, ("ithread %s has no process", ie->ie_name));
980 
981 	/*
982 	 * Set it_need to tell the thread to keep running if it is already
983 	 * running.  Then, lock the thread and see if we actually need to
984 	 * put it on the runqueue.
985 	 *
986 	 * Use store_rel to arrange that the store to ih_need in
987 	 * swi_sched() is before the store to it_need and prepare for
988 	 * transfer of this order to loads in the ithread.
989 	 */
990 	atomic_store_rel_int(&it->it_need, 1);
991 	thread_lock(td);
992 	if (TD_AWAITING_INTR(td)) {
993 #ifdef HWPMC_HOOKS
994 		it->it_waiting = 0;
995 		if (PMC_HOOK_INSTALLED_ANY())
996 			PMC_SOFT_CALL_INTR_HLPR(schedule, frame);
997 #endif
998 		CTR3(KTR_INTR, "%s: schedule pid %d (%s)", __func__, td->td_proc->p_pid,
999 		    td->td_name);
1000 		TD_CLR_IWAIT(td);
1001 		sched_wakeup(td, SRQ_INTR);
1002 	} else {
1003 #ifdef HWPMC_HOOKS
1004 		it->it_waiting++;
1005 		if (PMC_HOOK_INSTALLED_ANY() &&
1006 		    (it->it_waiting >= intr_hwpmc_waiting_report_threshold))
1007 			PMC_SOFT_CALL_INTR_HLPR(waiting, frame);
1008 #endif
1009 		CTR5(KTR_INTR, "%s: pid %d (%s): it_need %d, state %d",
1010 		    __func__, td->td_proc->p_pid, td->td_name, it->it_need, TD_GET_STATE(td));
1011 		thread_unlock(td);
1012 	}
1013 
1014 	return (0);
1015 }
1016 
1017 /*
1018  * Allow interrupt event binding for software interrupt handlers -- a no-op,
1019  * since interrupts are generated in software rather than being directed by
1020  * a PIC.
1021  */
1022 static int
1023 swi_assign_cpu(void *arg, int cpu)
1024 {
1025 
1026 	return (0);
1027 }
1028 
1029 /*
1030  * Add a software interrupt handler to a specified event.  If a given event
1031  * is not specified, then a new event is created.
1032  */
1033 int
1034 swi_add(struct intr_event **eventp, const char *name, driver_intr_t handler,
1035 	    void *arg, int pri, enum intr_type flags, void **cookiep)
1036 {
1037 	struct intr_event *ie;
1038 	int error = 0;
1039 
1040 	if (flags & INTR_ENTROPY)
1041 		return (EINVAL);
1042 
1043 	ie = (eventp != NULL) ? *eventp : NULL;
1044 
1045 	if (ie != NULL) {
1046 		if (!(ie->ie_flags & IE_SOFT))
1047 			return (EINVAL);
1048 	} else {
1049 		error = intr_event_create(&ie, NULL, IE_SOFT, 0,
1050 		    NULL, NULL, NULL, swi_assign_cpu, "swi%d:", pri);
1051 		if (error)
1052 			return (error);
1053 		if (eventp != NULL)
1054 			*eventp = ie;
1055 	}
1056 	if (handler != NULL) {
1057 		error = intr_event_add_handler(ie, name, NULL, handler, arg,
1058 		    PI_SWI(pri), flags, cookiep);
1059 	}
1060 	return (error);
1061 }
1062 
1063 /*
1064  * Schedule a software interrupt thread.
1065  */
1066 void
1067 swi_sched(void *cookie, int flags)
1068 {
1069 	struct intr_handler *ih = (struct intr_handler *)cookie;
1070 	struct intr_event *ie = ih->ih_event;
1071 	struct intr_entropy entropy;
1072 	int error __unused;
1073 
1074 	CTR3(KTR_INTR, "swi_sched: %s %s need=%d", ie->ie_name, ih->ih_name,
1075 	    ih->ih_need);
1076 
1077 	if ((flags & SWI_FROMNMI) == 0) {
1078 		entropy.event = (uintptr_t)ih;
1079 		entropy.td = curthread;
1080 		random_harvest_queue(&entropy, sizeof(entropy), RANDOM_SWI);
1081 	}
1082 
1083 	/*
1084 	 * Set ih_need for this handler so that if the ithread is already
1085 	 * running it will execute this handler on the next pass.  Otherwise,
1086 	 * it will execute it the next time it runs.
1087 	 */
1088 	ih->ih_need = 1;
1089 
1090 	if (flags & SWI_DELAY)
1091 		return;
1092 
1093 	if (flags & SWI_FROMNMI) {
1094 #if defined(SMP) && (defined(__i386__) || defined(__amd64__))
1095 		KASSERT(ie == clk_intr_event,
1096 		    ("SWI_FROMNMI used not with clk_intr_event"));
1097 		ipi_self_from_nmi(IPI_SWI);
1098 #endif
1099 	} else {
1100 		VM_CNT_INC(v_soft);
1101 		error = intr_event_schedule_thread(ie, NULL);
1102 		KASSERT(error == 0, ("stray software interrupt"));
1103 	}
1104 }
1105 
1106 /*
1107  * Remove a software interrupt handler.  Currently this code does not
1108  * remove the associated interrupt event if it becomes empty.  Calling code
1109  * may do so manually via intr_event_destroy(), but that's not really
1110  * an optimal interface.
1111  */
1112 int
1113 swi_remove(void *cookie)
1114 {
1115 
1116 	return (intr_event_remove_handler(cookie));
1117 }
1118 
1119 static void
1120 intr_event_execute_handlers(struct proc *p, struct intr_event *ie)
1121 {
1122 	struct intr_handler *ih, *ihn, *ihp;
1123 
1124 	ihp = NULL;
1125 	CK_SLIST_FOREACH_SAFE(ih, &ie->ie_handlers, ih_next, ihn) {
1126 		/*
1127 		 * If this handler is marked for death, remove it from
1128 		 * the list of handlers and wake up the sleeper.
1129 		 */
1130 		if (ih->ih_flags & IH_DEAD) {
1131 			mtx_lock(&ie->ie_lock);
1132 			if (ihp == NULL)
1133 				CK_SLIST_REMOVE_HEAD(&ie->ie_handlers, ih_next);
1134 			else
1135 				CK_SLIST_REMOVE_AFTER(ihp, ih_next);
1136 			ih->ih_flags &= ~IH_DEAD;
1137 			wakeup(ih);
1138 			mtx_unlock(&ie->ie_lock);
1139 			continue;
1140 		}
1141 
1142 		/*
1143 		 * Now that we know that the current element won't be removed
1144 		 * update the previous element.
1145 		 */
1146 		ihp = ih;
1147 
1148 		if ((ih->ih_flags & IH_CHANGED) != 0) {
1149 			mtx_lock(&ie->ie_lock);
1150 			ih->ih_flags &= ~IH_CHANGED;
1151 			wakeup(ih);
1152 			mtx_unlock(&ie->ie_lock);
1153 		}
1154 
1155 		/* Skip filter only handlers */
1156 		if (ih->ih_handler == NULL)
1157 			continue;
1158 
1159 		/* Skip suspended handlers */
1160 		if ((ih->ih_flags & IH_SUSP) != 0)
1161 			continue;
1162 
1163 		/*
1164 		 * For software interrupt threads, we only execute
1165 		 * handlers that have their need flag set.  Hardware
1166 		 * interrupt threads always invoke all of their handlers.
1167 		 *
1168 		 * ih_need can only be 0 or 1.  Failed cmpset below
1169 		 * means that there is no request to execute handlers,
1170 		 * so a retry of the cmpset is not needed.
1171 		 */
1172 		if ((ie->ie_flags & IE_SOFT) != 0 &&
1173 		    atomic_cmpset_int(&ih->ih_need, 1, 0) == 0)
1174 			continue;
1175 
1176 		/* Execute this handler. */
1177 		CTR6(KTR_INTR, "%s: pid %d exec %p(%p) for %s flg=%x",
1178 		    __func__, p->p_pid, (void *)ih->ih_handler,
1179 		    ih->ih_argument, ih->ih_name, ih->ih_flags);
1180 
1181 		if (!(ih->ih_flags & IH_MPSAFE))
1182 			mtx_lock(&Giant);
1183 		ih->ih_handler(ih->ih_argument);
1184 		if (!(ih->ih_flags & IH_MPSAFE))
1185 			mtx_unlock(&Giant);
1186 	}
1187 }
1188 
1189 static void
1190 ithread_execute_handlers(struct proc *p, struct intr_event *ie)
1191 {
1192 
1193 	/* Interrupt handlers should not sleep. */
1194 	if (!(ie->ie_flags & IE_SOFT))
1195 		THREAD_NO_SLEEPING();
1196 	intr_event_execute_handlers(p, ie);
1197 	if (!(ie->ie_flags & IE_SOFT))
1198 		THREAD_SLEEPING_OK();
1199 
1200 	/*
1201 	 * Interrupt storm handling:
1202 	 *
1203 	 * If this interrupt source is currently storming, then throttle
1204 	 * it to only fire the handler once  per clock tick.
1205 	 *
1206 	 * If this interrupt source is not currently storming, but the
1207 	 * number of back to back interrupts exceeds the storm threshold,
1208 	 * then enter storming mode.
1209 	 */
1210 	if (__predict_false(intr_storm_threshold != 0 &&
1211 	    ie->ie_count >= intr_storm_threshold &&
1212 	    (ie->ie_flags & IE_SOFT) == 0)) {
1213 		/* Report the message only once every second. */
1214 		if (ppsratecheck(&ie->ie_warntm, &ie->ie_warncnt, 1)) {
1215 			printf(
1216 	"interrupt storm detected on \"%s\"; throttling interrupt source\n",
1217 			    ie->ie_name);
1218 		}
1219 		pause("istorm", 1);
1220 	} else
1221 		ie->ie_count++;
1222 
1223 	/*
1224 	 * Now that all the handlers have had a chance to run, reenable
1225 	 * the interrupt source.
1226 	 */
1227 	if (ie->ie_post_ithread != NULL)
1228 		ie->ie_post_ithread(ie->ie_source);
1229 }
1230 
1231 /*
1232  * This is the main code for interrupt threads.
1233  */
1234 static void
1235 ithread_loop(void *arg)
1236 {
1237 	struct epoch_tracker et;
1238 	struct intr_thread *ithd;
1239 	struct intr_event *ie;
1240 	struct thread *td;
1241 	struct proc *p;
1242 	int epoch_count;
1243 	bool needs_epoch;
1244 
1245 	td = curthread;
1246 	p = td->td_proc;
1247 	ithd = (struct intr_thread *)arg;
1248 	KASSERT(ithd->it_thread == td,
1249 	    ("%s: ithread and proc linkage out of sync", __func__));
1250 	ie = ithd->it_event;
1251 	ie->ie_count = 0;
1252 
1253 	/*
1254 	 * As long as we have interrupts outstanding, go through the
1255 	 * list of handlers, giving each one a go at it.
1256 	 */
1257 	for (;;) {
1258 		/*
1259 		 * If we are an orphaned thread, then just die.
1260 		 */
1261 		if (__predict_false((ithd->it_flags & IT_DEAD) != 0)) {
1262 			CTR3(KTR_INTR, "%s: pid %d (%s) exiting", __func__,
1263 			    p->p_pid, td->td_name);
1264 			mtx_lock(&ie->ie_lock);
1265 			ie->ie_thread = NULL;
1266 			wakeup(ithd);
1267 			mtx_unlock(&ie->ie_lock);
1268 
1269 			free(ithd, M_ITHREAD);
1270 			kthread_exit();
1271 		}
1272 
1273 		/*
1274 		 * Service interrupts.  If another interrupt arrives while
1275 		 * we are running, it will set it_need to note that we
1276 		 * should make another pass.
1277 		 *
1278 		 * The load_acq part of the following cmpset ensures
1279 		 * that the load of ih_need in ithread_execute_handlers()
1280 		 * is ordered after the load of it_need here.
1281 		 */
1282 		needs_epoch =
1283 		    (atomic_load_int(&ie->ie_hflags) & IH_NET) != 0;
1284 		if (needs_epoch) {
1285 			epoch_count = 0;
1286 			NET_EPOCH_ENTER(et);
1287 		}
1288 		while (atomic_cmpset_acq_int(&ithd->it_need, 1, 0) != 0) {
1289 			ithread_execute_handlers(p, ie);
1290 			if (needs_epoch &&
1291 			    ++epoch_count >= intr_epoch_batch) {
1292 				NET_EPOCH_EXIT(et);
1293 				epoch_count = 0;
1294 				NET_EPOCH_ENTER(et);
1295 			}
1296 		}
1297 		if (needs_epoch)
1298 			NET_EPOCH_EXIT(et);
1299 		WITNESS_WARN(WARN_PANIC, NULL, "suspending ithread");
1300 		mtx_assert(&Giant, MA_NOTOWNED);
1301 
1302 		/*
1303 		 * Processed all our interrupts.  Now get the sched
1304 		 * lock.  This may take a while and it_need may get
1305 		 * set again, so we have to check it again.
1306 		 */
1307 		thread_lock(td);
1308 		if (atomic_load_acq_int(&ithd->it_need) == 0 &&
1309 		    (ithd->it_flags & (IT_DEAD | IT_WAIT)) == 0) {
1310 			TD_SET_IWAIT(td);
1311 			ie->ie_count = 0;
1312 			mi_switch(SW_VOL | SWT_IWAIT);
1313 		} else if ((ithd->it_flags & IT_WAIT) != 0) {
1314 			ithd->it_flags &= ~IT_WAIT;
1315 			thread_unlock(td);
1316 			wakeup(ithd);
1317 		} else
1318 			thread_unlock(td);
1319 	}
1320 }
1321 
1322 /*
1323  * Main interrupt handling body.
1324  *
1325  * Input:
1326  * o ie:                        the event connected to this interrupt.
1327 --------------------------------------------------------------------------------
1328  * o frame:                     the current trap frame. If the client interrupt
1329  *				handler needs this frame, they should get it
1330  *				via curthread->td_intr_frame.
1331  *
1332  * Return value:
1333  * o 0:                         everything ok.
1334  * o EINVAL:                    stray interrupt.
1335  */
1336 int
1337 intr_event_handle(struct intr_event *ie, struct trapframe *frame)
1338 {
1339 	struct intr_handler *ih;
1340 	struct trapframe *oldframe;
1341 	struct thread *td;
1342 	int phase;
1343 	int ret;
1344 	bool filter, thread;
1345 
1346 	td = curthread;
1347 
1348 #ifdef KSTACK_USAGE_PROF
1349 	intr_prof_stack_use(td, frame);
1350 #endif
1351 
1352 	/* An interrupt with no event or handlers is a stray interrupt. */
1353 	if (ie == NULL || CK_SLIST_EMPTY(&ie->ie_handlers))
1354 		return (EINVAL);
1355 
1356 	/*
1357 	 * Execute fast interrupt handlers directly.
1358 	 */
1359 	td->td_intr_nesting_level++;
1360 	filter = false;
1361 	thread = false;
1362 	ret = 0;
1363 	critical_enter();
1364 	oldframe = td->td_intr_frame;
1365 	td->td_intr_frame = frame;
1366 
1367 	phase = ie->ie_phase;
1368 	atomic_add_int(&ie->ie_active[phase], 1);
1369 
1370 	/*
1371 	 * This fence is required to ensure that no later loads are
1372 	 * re-ordered before the ie_active store.
1373 	 */
1374 	atomic_thread_fence_seq_cst();
1375 
1376 	CK_SLIST_FOREACH(ih, &ie->ie_handlers, ih_next) {
1377 		if ((ih->ih_flags & IH_SUSP) != 0)
1378 			continue;
1379 		if ((ie->ie_flags & IE_SOFT) != 0 && ih->ih_need == 0)
1380 			continue;
1381 		if (ih->ih_filter == NULL) {
1382 			thread = true;
1383 			continue;
1384 		}
1385 		CTR4(KTR_INTR, "%s: exec %p(%p) for %s", __func__,
1386 		    ih->ih_filter, ih->ih_argument, ih->ih_name);
1387 		ret = ih->ih_filter(ih->ih_argument);
1388 #ifdef HWPMC_HOOKS
1389 		PMC_SOFT_CALL_TF( , , intr, all, frame);
1390 #endif
1391 		KASSERT(ret == FILTER_STRAY ||
1392 		    ((ret & (FILTER_SCHEDULE_THREAD | FILTER_HANDLED)) != 0 &&
1393 		    (ret & ~(FILTER_SCHEDULE_THREAD | FILTER_HANDLED)) == 0),
1394 		    ("%s: incorrect return value %#x from %s", __func__, ret,
1395 		    ih->ih_name));
1396 		filter = filter || ret == FILTER_HANDLED;
1397 #ifdef HWPMC_HOOKS
1398 		if (ret & FILTER_SCHEDULE_THREAD)
1399 			PMC_SOFT_CALL_TF( , , intr, ithread, frame);
1400 		else if (ret & FILTER_HANDLED)
1401 			PMC_SOFT_CALL_TF( , , intr, filter, frame);
1402 		else if (ret == FILTER_STRAY)
1403 			PMC_SOFT_CALL_TF( , , intr, stray, frame);
1404 #endif
1405 
1406 		/*
1407 		 * Wrapper handler special handling:
1408 		 *
1409 		 * in some particular cases (like pccard and pccbb),
1410 		 * the _real_ device handler is wrapped in a couple of
1411 		 * functions - a filter wrapper and an ithread wrapper.
1412 		 * In this case (and just in this case), the filter wrapper
1413 		 * could ask the system to schedule the ithread and mask
1414 		 * the interrupt source if the wrapped handler is composed
1415 		 * of just an ithread handler.
1416 		 *
1417 		 * TODO: write a generic wrapper to avoid people rolling
1418 		 * their own.
1419 		 */
1420 		if (!thread) {
1421 			if (ret == FILTER_SCHEDULE_THREAD)
1422 				thread = true;
1423 		}
1424 	}
1425 	atomic_add_rel_int(&ie->ie_active[phase], -1);
1426 
1427 	td->td_intr_frame = oldframe;
1428 
1429 	if (thread) {
1430 		if (ie->ie_pre_ithread != NULL)
1431 			ie->ie_pre_ithread(ie->ie_source);
1432 	} else {
1433 		if (ie->ie_post_filter != NULL)
1434 			ie->ie_post_filter(ie->ie_source);
1435 	}
1436 
1437 	/* Schedule the ithread if needed. */
1438 	if (thread) {
1439 		int error __unused;
1440 
1441 		error =  intr_event_schedule_thread(ie, frame);
1442 		KASSERT(error == 0, ("bad stray interrupt"));
1443 	}
1444 	critical_exit();
1445 	td->td_intr_nesting_level--;
1446 #ifdef notyet
1447 	/* The interrupt is not aknowledged by any filter and has no ithread. */
1448 	if (!thread && !filter)
1449 		return (EINVAL);
1450 #endif
1451 	return (0);
1452 }
1453 
1454 #ifdef DDB
1455 /*
1456  * Dump details about an interrupt handler
1457  */
1458 static void
1459 db_dump_intrhand(struct intr_handler *ih)
1460 {
1461 	int comma;
1462 
1463 	db_printf("\t%-10s ", ih->ih_name);
1464 	switch (ih->ih_pri) {
1465 	case PI_REALTIME:
1466 		db_printf("CLK ");
1467 		break;
1468 	case PI_INTR:
1469 		db_printf("INTR");
1470 		break;
1471 	default:
1472 		if (ih->ih_pri >= PI_SOFT)
1473 			db_printf("SWI ");
1474 		else
1475 			db_printf("%4u", ih->ih_pri);
1476 		break;
1477 	}
1478 	db_printf(" ");
1479 	if (ih->ih_filter != NULL) {
1480 		db_printf("[F]");
1481 		db_printsym((uintptr_t)ih->ih_filter, DB_STGY_PROC);
1482 	}
1483 	if (ih->ih_handler != NULL) {
1484 		if (ih->ih_filter != NULL)
1485 			db_printf(",");
1486 		db_printf("[H]");
1487 		db_printsym((uintptr_t)ih->ih_handler, DB_STGY_PROC);
1488 	}
1489 	db_printf("(%p)", ih->ih_argument);
1490 	if (ih->ih_need ||
1491 	    (ih->ih_flags & (IH_EXCLUSIVE | IH_ENTROPY | IH_DEAD |
1492 	    IH_MPSAFE)) != 0) {
1493 		db_printf(" {");
1494 		comma = 0;
1495 		if (ih->ih_flags & IH_EXCLUSIVE) {
1496 			if (comma)
1497 				db_printf(", ");
1498 			db_printf("EXCL");
1499 			comma = 1;
1500 		}
1501 		if (ih->ih_flags & IH_ENTROPY) {
1502 			if (comma)
1503 				db_printf(", ");
1504 			db_printf("ENTROPY");
1505 			comma = 1;
1506 		}
1507 		if (ih->ih_flags & IH_DEAD) {
1508 			if (comma)
1509 				db_printf(", ");
1510 			db_printf("DEAD");
1511 			comma = 1;
1512 		}
1513 		if (ih->ih_flags & IH_MPSAFE) {
1514 			if (comma)
1515 				db_printf(", ");
1516 			db_printf("MPSAFE");
1517 			comma = 1;
1518 		}
1519 		if (ih->ih_need) {
1520 			if (comma)
1521 				db_printf(", ");
1522 			db_printf("NEED");
1523 		}
1524 		db_printf("}");
1525 	}
1526 	db_printf("\n");
1527 }
1528 
1529 /*
1530  * Dump details about a event.
1531  */
1532 void
1533 db_dump_intr_event(struct intr_event *ie, int handlers)
1534 {
1535 	struct intr_handler *ih;
1536 	struct intr_thread *it;
1537 	int comma;
1538 
1539 	db_printf("%s ", ie->ie_fullname);
1540 	it = ie->ie_thread;
1541 	if (it != NULL)
1542 		db_printf("(pid %d)", it->it_thread->td_proc->p_pid);
1543 	else
1544 		db_printf("(no thread)");
1545 	if ((ie->ie_flags & (IE_SOFT | IE_ADDING_THREAD)) != 0 ||
1546 	    (it != NULL && it->it_need)) {
1547 		db_printf(" {");
1548 		comma = 0;
1549 		if (ie->ie_flags & IE_SOFT) {
1550 			db_printf("SOFT");
1551 			comma = 1;
1552 		}
1553 		if (ie->ie_flags & IE_ADDING_THREAD) {
1554 			if (comma)
1555 				db_printf(", ");
1556 			db_printf("ADDING_THREAD");
1557 			comma = 1;
1558 		}
1559 		if (it != NULL && it->it_need) {
1560 			if (comma)
1561 				db_printf(", ");
1562 			db_printf("NEED");
1563 		}
1564 		db_printf("}");
1565 	}
1566 	db_printf("\n");
1567 
1568 	if (handlers)
1569 		CK_SLIST_FOREACH(ih, &ie->ie_handlers, ih_next)
1570 		    db_dump_intrhand(ih);
1571 }
1572 
1573 /*
1574  * Dump data about interrupt handlers
1575  */
1576 DB_SHOW_COMMAND_FLAGS(intr, db_show_intr, DB_CMD_MEMSAFE)
1577 {
1578 	struct intr_event *ie;
1579 	int all, verbose;
1580 
1581 	verbose = strchr(modif, 'v') != NULL;
1582 	all = strchr(modif, 'a') != NULL;
1583 	TAILQ_FOREACH(ie, &event_list, ie_list) {
1584 		if (!all && CK_SLIST_EMPTY(&ie->ie_handlers))
1585 			continue;
1586 		db_dump_intr_event(ie, verbose);
1587 		if (db_pager_quit)
1588 			break;
1589 	}
1590 }
1591 #endif /* DDB */
1592 
1593 /*
1594  * Start standard software interrupt threads
1595  */
1596 static void
1597 start_softintr(void *dummy)
1598 {
1599 
1600 	if (swi_add(&clk_intr_event, "clk", NULL, NULL, SWI_CLOCK,
1601 	    INTR_MPSAFE, NULL))
1602 		panic("died while creating clk swi ithread");
1603 }
1604 SYSINIT(start_softintr, SI_SUB_SOFTINTR, SI_ORDER_FIRST, start_softintr,
1605     NULL);
1606 
1607 /*
1608  * Sysctls used by systat and others: hw.intrnames and hw.intrcnt.
1609  * The data for this machine dependent, and the declarations are in machine
1610  * dependent code.  The layout of intrnames and intrcnt however is machine
1611  * independent.
1612  *
1613  * We do not know the length of intrcnt and intrnames at compile time, so
1614  * calculate things at run time.
1615  */
1616 static int
1617 sysctl_intrnames(SYSCTL_HANDLER_ARGS)
1618 {
1619 	return (sysctl_handle_opaque(oidp, intrnames, sintrnames, req));
1620 }
1621 
1622 SYSCTL_PROC(_hw, OID_AUTO, intrnames,
1623     CTLTYPE_OPAQUE | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
1624     sysctl_intrnames, "",
1625     "Interrupt Names");
1626 
1627 static int
1628 sysctl_intrcnt(SYSCTL_HANDLER_ARGS)
1629 {
1630 #ifdef SCTL_MASK32
1631 	uint32_t *intrcnt32;
1632 	unsigned i;
1633 	int error;
1634 
1635 	if (req->flags & SCTL_MASK32) {
1636 		if (!req->oldptr)
1637 			return (sysctl_handle_opaque(oidp, NULL, sintrcnt / 2, req));
1638 		intrcnt32 = malloc(sintrcnt / 2, M_TEMP, M_NOWAIT);
1639 		if (intrcnt32 == NULL)
1640 			return (ENOMEM);
1641 		for (i = 0; i < sintrcnt / sizeof (u_long); i++)
1642 			intrcnt32[i] = intrcnt[i];
1643 		error = sysctl_handle_opaque(oidp, intrcnt32, sintrcnt / 2, req);
1644 		free(intrcnt32, M_TEMP);
1645 		return (error);
1646 	}
1647 #endif
1648 	return (sysctl_handle_opaque(oidp, intrcnt, sintrcnt, req));
1649 }
1650 
1651 SYSCTL_PROC(_hw, OID_AUTO, intrcnt,
1652     CTLTYPE_OPAQUE | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
1653     sysctl_intrcnt, "",
1654     "Interrupt Counts");
1655 
1656 #ifdef DDB
1657 /*
1658  * DDB command to dump the interrupt statistics.
1659  */
1660 DB_SHOW_COMMAND_FLAGS(intrcnt, db_show_intrcnt, DB_CMD_MEMSAFE)
1661 {
1662 	u_long *i;
1663 	char *cp;
1664 	u_int j;
1665 
1666 	cp = intrnames;
1667 	j = 0;
1668 	for (i = intrcnt; j < (sintrcnt / sizeof(u_long)) && !db_pager_quit;
1669 	    i++, j++) {
1670 		if (*cp == '\0')
1671 			break;
1672 		if (*i != 0)
1673 			db_printf("%s\t%lu\n", cp, *i);
1674 		cp += strlen(cp) + 1;
1675 	}
1676 }
1677 #endif
1678