xref: /freebsd/sys/kern/kern_intr.c (revision 5bb3134a8c21cb87b30e135ef168483f0333dabb)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 1997, Stefan Esser <se@freebsd.org>
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice unmodified, this list of conditions, and the following
12  *    disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
18  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
19  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
20  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
21  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
22  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
23  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
24  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
25  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
26  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
27  */
28 
29 #include <sys/cdefs.h>
30 __FBSDID("$FreeBSD$");
31 
32 #include "opt_ddb.h"
33 #include "opt_hwpmc_hooks.h"
34 #include "opt_kstack_usage_prof.h"
35 
36 #include <sys/param.h>
37 #include <sys/bus.h>
38 #include <sys/conf.h>
39 #include <sys/cpuset.h>
40 #include <sys/rtprio.h>
41 #include <sys/systm.h>
42 #include <sys/interrupt.h>
43 #include <sys/kernel.h>
44 #include <sys/kthread.h>
45 #include <sys/ktr.h>
46 #include <sys/limits.h>
47 #include <sys/lock.h>
48 #include <sys/malloc.h>
49 #include <sys/mutex.h>
50 #include <sys/priv.h>
51 #include <sys/proc.h>
52 #include <sys/epoch.h>
53 #include <sys/random.h>
54 #include <sys/resourcevar.h>
55 #include <sys/sched.h>
56 #include <sys/smp.h>
57 #include <sys/sysctl.h>
58 #include <sys/syslog.h>
59 #include <sys/unistd.h>
60 #include <sys/vmmeter.h>
61 #include <machine/atomic.h>
62 #include <machine/cpu.h>
63 #include <machine/md_var.h>
64 #include <machine/smp.h>
65 #include <machine/stdarg.h>
66 #ifdef DDB
67 #include <ddb/ddb.h>
68 #include <ddb/db_sym.h>
69 #endif
70 
71 /*
72  * Describe an interrupt thread.  There is one of these per interrupt event.
73  */
74 struct intr_thread {
75 	struct intr_event *it_event;
76 	struct thread *it_thread;	/* Kernel thread. */
77 	int	it_flags;		/* (j) IT_* flags. */
78 	int	it_need;		/* Needs service. */
79 	int	it_waiting;		/* Waiting in the runq. */
80 };
81 
82 /* Interrupt thread flags kept in it_flags */
83 #define	IT_DEAD		0x000001	/* Thread is waiting to exit. */
84 #define	IT_WAIT		0x000002	/* Thread is waiting for completion. */
85 
86 struct	intr_entropy {
87 	struct	thread *td;
88 	uintptr_t event;
89 };
90 
91 struct	intr_event *clk_intr_event;
92 struct	intr_event *tty_intr_event;
93 void	*vm_ih;
94 struct proc *intrproc;
95 
96 static MALLOC_DEFINE(M_ITHREAD, "ithread", "Interrupt Threads");
97 
98 static int intr_storm_threshold = 0;
99 SYSCTL_INT(_hw, OID_AUTO, intr_storm_threshold, CTLFLAG_RWTUN,
100     &intr_storm_threshold, 0,
101     "Number of consecutive interrupts before storm protection is enabled");
102 static int intr_epoch_batch = 1000;
103 SYSCTL_INT(_hw, OID_AUTO, intr_epoch_batch, CTLFLAG_RWTUN, &intr_epoch_batch,
104     0, "Maximum interrupt handler executions without re-entering epoch(9)");
105 #ifdef HWPMC_HOOKS
106 static int intr_hwpmc_waiting_report_threshold = 1;
107 SYSCTL_INT(_hw, OID_AUTO, intr_hwpmc_waiting_report_threshold, CTLFLAG_RWTUN,
108     &intr_hwpmc_waiting_report_threshold, 1,
109     "Threshold for reporting number of events in a workq");
110 #define	PMC_HOOK_INSTALLED_ANY() __predict_false(pmc_hook != NULL)
111 #endif
112 static TAILQ_HEAD(, intr_event) event_list =
113     TAILQ_HEAD_INITIALIZER(event_list);
114 static struct mtx event_lock;
115 MTX_SYSINIT(intr_event_list, &event_lock, "intr event list", MTX_DEF);
116 
117 static void	intr_event_update(struct intr_event *ie);
118 static int	intr_event_schedule_thread(struct intr_event *ie, struct trapframe *frame);
119 static struct intr_thread *ithread_create(const char *name);
120 static void	ithread_destroy(struct intr_thread *ithread);
121 static void	ithread_execute_handlers(struct proc *p,
122 		    struct intr_event *ie);
123 static void	ithread_loop(void *);
124 static void	ithread_update(struct intr_thread *ithd);
125 static void	start_softintr(void *);
126 
127 #ifdef HWPMC_HOOKS
128 #include <sys/pmckern.h>
129 PMC_SOFT_DEFINE( , , intr, all);
130 PMC_SOFT_DEFINE( , , intr, ithread);
131 PMC_SOFT_DEFINE( , , intr, filter);
132 PMC_SOFT_DEFINE( , , intr, stray);
133 PMC_SOFT_DEFINE( , , intr, schedule);
134 PMC_SOFT_DEFINE( , , intr, waiting);
135 
136 #define PMC_SOFT_CALL_INTR_HLPR(event, frame)			\
137 do {					\
138 	if (frame != NULL)					\
139 		PMC_SOFT_CALL_TF( , , intr, event, frame);	\
140 	else							\
141 		PMC_SOFT_CALL( , , intr, event);		\
142 } while (0)
143 #endif
144 
145 /* Map an interrupt type to an ithread priority. */
146 u_char
147 intr_priority(enum intr_type flags)
148 {
149 	u_char pri;
150 
151 	flags &= (INTR_TYPE_TTY | INTR_TYPE_BIO | INTR_TYPE_NET |
152 	    INTR_TYPE_CAM | INTR_TYPE_MISC | INTR_TYPE_CLK | INTR_TYPE_AV);
153 	switch (flags) {
154 	case INTR_TYPE_TTY:
155 		pri = PI_TTY;
156 		break;
157 	case INTR_TYPE_BIO:
158 		pri = PI_DISK;
159 		break;
160 	case INTR_TYPE_NET:
161 		pri = PI_NET;
162 		break;
163 	case INTR_TYPE_CAM:
164 		pri = PI_DISK;
165 		break;
166 	case INTR_TYPE_AV:
167 		pri = PI_AV;
168 		break;
169 	case INTR_TYPE_CLK:
170 		pri = PI_REALTIME;
171 		break;
172 	case INTR_TYPE_MISC:
173 		pri = PI_DULL;          /* don't care */
174 		break;
175 	default:
176 		/* We didn't specify an interrupt level. */
177 		panic("intr_priority: no interrupt type in flags");
178 	}
179 
180 	return pri;
181 }
182 
183 /*
184  * Update an ithread based on the associated intr_event.
185  */
186 static void
187 ithread_update(struct intr_thread *ithd)
188 {
189 	struct intr_event *ie;
190 	struct thread *td;
191 	u_char pri;
192 
193 	ie = ithd->it_event;
194 	td = ithd->it_thread;
195 	mtx_assert(&ie->ie_lock, MA_OWNED);
196 
197 	/* Determine the overall priority of this event. */
198 	if (CK_SLIST_EMPTY(&ie->ie_handlers))
199 		pri = PRI_MAX_ITHD;
200 	else
201 		pri = CK_SLIST_FIRST(&ie->ie_handlers)->ih_pri;
202 
203 	/* Update name and priority. */
204 	strlcpy(td->td_name, ie->ie_fullname, sizeof(td->td_name));
205 #ifdef KTR
206 	sched_clear_tdname(td);
207 #endif
208 	thread_lock(td);
209 	sched_prio(td, pri);
210 	thread_unlock(td);
211 }
212 
213 /*
214  * Regenerate the full name of an interrupt event and update its priority.
215  */
216 static void
217 intr_event_update(struct intr_event *ie)
218 {
219 	struct intr_handler *ih;
220 	char *last;
221 	int missed, space, flags;
222 
223 	/* Start off with no entropy and just the name of the event. */
224 	mtx_assert(&ie->ie_lock, MA_OWNED);
225 	strlcpy(ie->ie_fullname, ie->ie_name, sizeof(ie->ie_fullname));
226 	flags = 0;
227 	missed = 0;
228 	space = 1;
229 
230 	/* Run through all the handlers updating values. */
231 	CK_SLIST_FOREACH(ih, &ie->ie_handlers, ih_next) {
232 		if (strlen(ie->ie_fullname) + strlen(ih->ih_name) + 1 <
233 		    sizeof(ie->ie_fullname)) {
234 			strcat(ie->ie_fullname, " ");
235 			strcat(ie->ie_fullname, ih->ih_name);
236 			space = 0;
237 		} else
238 			missed++;
239 		flags |= ih->ih_flags;
240 	}
241 	ie->ie_hflags = flags;
242 
243 	/*
244 	 * If there is only one handler and its name is too long, just copy in
245 	 * as much of the end of the name (includes the unit number) as will
246 	 * fit.  Otherwise, we have multiple handlers and not all of the names
247 	 * will fit.  Add +'s to indicate missing names.  If we run out of room
248 	 * and still have +'s to add, change the last character from a + to a *.
249 	 */
250 	if (missed == 1 && space == 1) {
251 		ih = CK_SLIST_FIRST(&ie->ie_handlers);
252 		missed = strlen(ie->ie_fullname) + strlen(ih->ih_name) + 2 -
253 		    sizeof(ie->ie_fullname);
254 		strcat(ie->ie_fullname, (missed == 0) ? " " : "-");
255 		strcat(ie->ie_fullname, &ih->ih_name[missed]);
256 		missed = 0;
257 	}
258 	last = &ie->ie_fullname[sizeof(ie->ie_fullname) - 2];
259 	while (missed-- > 0) {
260 		if (strlen(ie->ie_fullname) + 1 == sizeof(ie->ie_fullname)) {
261 			if (*last == '+') {
262 				*last = '*';
263 				break;
264 			} else
265 				*last = '+';
266 		} else if (space) {
267 			strcat(ie->ie_fullname, " +");
268 			space = 0;
269 		} else
270 			strcat(ie->ie_fullname, "+");
271 	}
272 
273 	/*
274 	 * If this event has an ithread, update it's priority and
275 	 * name.
276 	 */
277 	if (ie->ie_thread != NULL)
278 		ithread_update(ie->ie_thread);
279 	CTR2(KTR_INTR, "%s: updated %s", __func__, ie->ie_fullname);
280 }
281 
282 int
283 intr_event_create(struct intr_event **event, void *source, int flags, int irq,
284     void (*pre_ithread)(void *), void (*post_ithread)(void *),
285     void (*post_filter)(void *), int (*assign_cpu)(void *, int),
286     const char *fmt, ...)
287 {
288 	struct intr_event *ie;
289 	va_list ap;
290 
291 	/* The only valid flag during creation is IE_SOFT. */
292 	if ((flags & ~IE_SOFT) != 0)
293 		return (EINVAL);
294 	ie = malloc(sizeof(struct intr_event), M_ITHREAD, M_WAITOK | M_ZERO);
295 	ie->ie_source = source;
296 	ie->ie_pre_ithread = pre_ithread;
297 	ie->ie_post_ithread = post_ithread;
298 	ie->ie_post_filter = post_filter;
299 	ie->ie_assign_cpu = assign_cpu;
300 	ie->ie_flags = flags;
301 	ie->ie_irq = irq;
302 	ie->ie_cpu = NOCPU;
303 	CK_SLIST_INIT(&ie->ie_handlers);
304 	mtx_init(&ie->ie_lock, "intr event", NULL, MTX_DEF);
305 
306 	va_start(ap, fmt);
307 	vsnprintf(ie->ie_name, sizeof(ie->ie_name), fmt, ap);
308 	va_end(ap);
309 	strlcpy(ie->ie_fullname, ie->ie_name, sizeof(ie->ie_fullname));
310 	mtx_lock(&event_lock);
311 	TAILQ_INSERT_TAIL(&event_list, ie, ie_list);
312 	mtx_unlock(&event_lock);
313 	if (event != NULL)
314 		*event = ie;
315 	CTR2(KTR_INTR, "%s: created %s", __func__, ie->ie_name);
316 	return (0);
317 }
318 
319 /*
320  * Bind an interrupt event to the specified CPU.  Note that not all
321  * platforms support binding an interrupt to a CPU.  For those
322  * platforms this request will fail.  Using a cpu id of NOCPU unbinds
323  * the interrupt event.
324  */
325 static int
326 _intr_event_bind(struct intr_event *ie, int cpu, bool bindirq, bool bindithread)
327 {
328 	lwpid_t id;
329 	int error;
330 
331 	/* Need a CPU to bind to. */
332 	if (cpu != NOCPU && CPU_ABSENT(cpu))
333 		return (EINVAL);
334 
335 	if (ie->ie_assign_cpu == NULL)
336 		return (EOPNOTSUPP);
337 
338 	error = priv_check(curthread, PRIV_SCHED_CPUSET_INTR);
339 	if (error)
340 		return (error);
341 
342 	/*
343 	 * If we have any ithreads try to set their mask first to verify
344 	 * permissions, etc.
345 	 */
346 	if (bindithread) {
347 		mtx_lock(&ie->ie_lock);
348 		if (ie->ie_thread != NULL) {
349 			id = ie->ie_thread->it_thread->td_tid;
350 			mtx_unlock(&ie->ie_lock);
351 			error = cpuset_setithread(id, cpu);
352 			if (error)
353 				return (error);
354 		} else
355 			mtx_unlock(&ie->ie_lock);
356 	}
357 	if (bindirq)
358 		error = ie->ie_assign_cpu(ie->ie_source, cpu);
359 	if (error) {
360 		if (bindithread) {
361 			mtx_lock(&ie->ie_lock);
362 			if (ie->ie_thread != NULL) {
363 				cpu = ie->ie_cpu;
364 				id = ie->ie_thread->it_thread->td_tid;
365 				mtx_unlock(&ie->ie_lock);
366 				(void)cpuset_setithread(id, cpu);
367 			} else
368 				mtx_unlock(&ie->ie_lock);
369 		}
370 		return (error);
371 	}
372 
373 	if (bindirq) {
374 		mtx_lock(&ie->ie_lock);
375 		ie->ie_cpu = cpu;
376 		mtx_unlock(&ie->ie_lock);
377 	}
378 
379 	return (error);
380 }
381 
382 /*
383  * Bind an interrupt event to the specified CPU.  For supported platforms, any
384  * associated ithreads as well as the primary interrupt context will be bound
385  * to the specificed CPU.
386  */
387 int
388 intr_event_bind(struct intr_event *ie, int cpu)
389 {
390 
391 	return (_intr_event_bind(ie, cpu, true, true));
392 }
393 
394 /*
395  * Bind an interrupt event to the specified CPU, but do not bind associated
396  * ithreads.
397  */
398 int
399 intr_event_bind_irqonly(struct intr_event *ie, int cpu)
400 {
401 
402 	return (_intr_event_bind(ie, cpu, true, false));
403 }
404 
405 /*
406  * Bind an interrupt event's ithread to the specified CPU.
407  */
408 int
409 intr_event_bind_ithread(struct intr_event *ie, int cpu)
410 {
411 
412 	return (_intr_event_bind(ie, cpu, false, true));
413 }
414 
415 /*
416  * Bind an interrupt event's ithread to the specified cpuset.
417  */
418 int
419 intr_event_bind_ithread_cpuset(struct intr_event *ie, cpuset_t *cs)
420 {
421 	lwpid_t id;
422 
423 	mtx_lock(&ie->ie_lock);
424 	if (ie->ie_thread != NULL) {
425 		id = ie->ie_thread->it_thread->td_tid;
426 		mtx_unlock(&ie->ie_lock);
427 		return (cpuset_setthread(id, cs));
428 	} else {
429 		mtx_unlock(&ie->ie_lock);
430 	}
431 	return (ENODEV);
432 }
433 
434 static struct intr_event *
435 intr_lookup(int irq)
436 {
437 	struct intr_event *ie;
438 
439 	mtx_lock(&event_lock);
440 	TAILQ_FOREACH(ie, &event_list, ie_list)
441 		if (ie->ie_irq == irq &&
442 		    (ie->ie_flags & IE_SOFT) == 0 &&
443 		    CK_SLIST_FIRST(&ie->ie_handlers) != NULL)
444 			break;
445 	mtx_unlock(&event_lock);
446 	return (ie);
447 }
448 
449 int
450 intr_setaffinity(int irq, int mode, void *m)
451 {
452 	struct intr_event *ie;
453 	cpuset_t *mask;
454 	int cpu, n;
455 
456 	mask = m;
457 	cpu = NOCPU;
458 	/*
459 	 * If we're setting all cpus we can unbind.  Otherwise make sure
460 	 * only one cpu is in the set.
461 	 */
462 	if (CPU_CMP(cpuset_root, mask)) {
463 		for (n = 0; n < CPU_SETSIZE; n++) {
464 			if (!CPU_ISSET(n, mask))
465 				continue;
466 			if (cpu != NOCPU)
467 				return (EINVAL);
468 			cpu = n;
469 		}
470 	}
471 	ie = intr_lookup(irq);
472 	if (ie == NULL)
473 		return (ESRCH);
474 	switch (mode) {
475 	case CPU_WHICH_IRQ:
476 		return (intr_event_bind(ie, cpu));
477 	case CPU_WHICH_INTRHANDLER:
478 		return (intr_event_bind_irqonly(ie, cpu));
479 	case CPU_WHICH_ITHREAD:
480 		return (intr_event_bind_ithread(ie, cpu));
481 	default:
482 		return (EINVAL);
483 	}
484 }
485 
486 int
487 intr_getaffinity(int irq, int mode, void *m)
488 {
489 	struct intr_event *ie;
490 	struct thread *td;
491 	struct proc *p;
492 	cpuset_t *mask;
493 	lwpid_t id;
494 	int error;
495 
496 	mask = m;
497 	ie = intr_lookup(irq);
498 	if (ie == NULL)
499 		return (ESRCH);
500 
501 	error = 0;
502 	CPU_ZERO(mask);
503 	switch (mode) {
504 	case CPU_WHICH_IRQ:
505 	case CPU_WHICH_INTRHANDLER:
506 		mtx_lock(&ie->ie_lock);
507 		if (ie->ie_cpu == NOCPU)
508 			CPU_COPY(cpuset_root, mask);
509 		else
510 			CPU_SET(ie->ie_cpu, mask);
511 		mtx_unlock(&ie->ie_lock);
512 		break;
513 	case CPU_WHICH_ITHREAD:
514 		mtx_lock(&ie->ie_lock);
515 		if (ie->ie_thread == NULL) {
516 			mtx_unlock(&ie->ie_lock);
517 			CPU_COPY(cpuset_root, mask);
518 		} else {
519 			id = ie->ie_thread->it_thread->td_tid;
520 			mtx_unlock(&ie->ie_lock);
521 			error = cpuset_which(CPU_WHICH_TID, id, &p, &td, NULL);
522 			if (error != 0)
523 				return (error);
524 			CPU_COPY(&td->td_cpuset->cs_mask, mask);
525 			PROC_UNLOCK(p);
526 		}
527 	default:
528 		return (EINVAL);
529 	}
530 	return (0);
531 }
532 
533 int
534 intr_event_destroy(struct intr_event *ie)
535 {
536 
537 	mtx_lock(&event_lock);
538 	mtx_lock(&ie->ie_lock);
539 	if (!CK_SLIST_EMPTY(&ie->ie_handlers)) {
540 		mtx_unlock(&ie->ie_lock);
541 		mtx_unlock(&event_lock);
542 		return (EBUSY);
543 	}
544 	TAILQ_REMOVE(&event_list, ie, ie_list);
545 #ifndef notyet
546 	if (ie->ie_thread != NULL) {
547 		ithread_destroy(ie->ie_thread);
548 		ie->ie_thread = NULL;
549 	}
550 #endif
551 	mtx_unlock(&ie->ie_lock);
552 	mtx_unlock(&event_lock);
553 	mtx_destroy(&ie->ie_lock);
554 	free(ie, M_ITHREAD);
555 	return (0);
556 }
557 
558 static struct intr_thread *
559 ithread_create(const char *name)
560 {
561 	struct intr_thread *ithd;
562 	struct thread *td;
563 	int error;
564 
565 	ithd = malloc(sizeof(struct intr_thread), M_ITHREAD, M_WAITOK | M_ZERO);
566 
567 	error = kproc_kthread_add(ithread_loop, ithd, &intrproc,
568 		    &td, RFSTOPPED | RFHIGHPID,
569 		    0, "intr", "%s", name);
570 	if (error)
571 		panic("kproc_create() failed with %d", error);
572 	thread_lock(td);
573 	sched_class(td, PRI_ITHD);
574 	TD_SET_IWAIT(td);
575 	thread_unlock(td);
576 	td->td_pflags |= TDP_ITHREAD;
577 	ithd->it_thread = td;
578 	CTR2(KTR_INTR, "%s: created %s", __func__, name);
579 	return (ithd);
580 }
581 
582 static void
583 ithread_destroy(struct intr_thread *ithread)
584 {
585 	struct thread *td;
586 
587 	CTR2(KTR_INTR, "%s: killing %s", __func__, ithread->it_event->ie_name);
588 	td = ithread->it_thread;
589 	thread_lock(td);
590 	ithread->it_flags |= IT_DEAD;
591 	if (TD_AWAITING_INTR(td)) {
592 		TD_CLR_IWAIT(td);
593 		sched_add(td, SRQ_INTR);
594 	} else
595 		thread_unlock(td);
596 }
597 
598 int
599 intr_event_add_handler(struct intr_event *ie, const char *name,
600     driver_filter_t filter, driver_intr_t handler, void *arg, u_char pri,
601     enum intr_type flags, void **cookiep)
602 {
603 	struct intr_handler *ih, *temp_ih;
604 	struct intr_handler **prevptr;
605 	struct intr_thread *it;
606 
607 	if (ie == NULL || name == NULL || (handler == NULL && filter == NULL))
608 		return (EINVAL);
609 
610 	/* Allocate and populate an interrupt handler structure. */
611 	ih = malloc(sizeof(struct intr_handler), M_ITHREAD, M_WAITOK | M_ZERO);
612 	ih->ih_filter = filter;
613 	ih->ih_handler = handler;
614 	ih->ih_argument = arg;
615 	strlcpy(ih->ih_name, name, sizeof(ih->ih_name));
616 	ih->ih_event = ie;
617 	ih->ih_pri = pri;
618 	if (flags & INTR_EXCL)
619 		ih->ih_flags = IH_EXCLUSIVE;
620 	if (flags & INTR_MPSAFE)
621 		ih->ih_flags |= IH_MPSAFE;
622 	if (flags & INTR_ENTROPY)
623 		ih->ih_flags |= IH_ENTROPY;
624 	if (flags & INTR_TYPE_NET)
625 		ih->ih_flags |= IH_NET;
626 
627 	/* We can only have one exclusive handler in a event. */
628 	mtx_lock(&ie->ie_lock);
629 	if (!CK_SLIST_EMPTY(&ie->ie_handlers)) {
630 		if ((flags & INTR_EXCL) ||
631 		    (CK_SLIST_FIRST(&ie->ie_handlers)->ih_flags & IH_EXCLUSIVE)) {
632 			mtx_unlock(&ie->ie_lock);
633 			free(ih, M_ITHREAD);
634 			return (EINVAL);
635 		}
636 	}
637 
638 	/* Create a thread if we need one. */
639 	while (ie->ie_thread == NULL && handler != NULL) {
640 		if (ie->ie_flags & IE_ADDING_THREAD)
641 			msleep(ie, &ie->ie_lock, 0, "ithread", 0);
642 		else {
643 			ie->ie_flags |= IE_ADDING_THREAD;
644 			mtx_unlock(&ie->ie_lock);
645 			it = ithread_create("intr: newborn");
646 			mtx_lock(&ie->ie_lock);
647 			ie->ie_flags &= ~IE_ADDING_THREAD;
648 			ie->ie_thread = it;
649 			it->it_event = ie;
650 			ithread_update(it);
651 			wakeup(ie);
652 		}
653 	}
654 
655 	/* Add the new handler to the event in priority order. */
656 	CK_SLIST_FOREACH_PREVPTR(temp_ih, prevptr, &ie->ie_handlers, ih_next) {
657 		if (temp_ih->ih_pri > ih->ih_pri)
658 			break;
659 	}
660 	CK_SLIST_INSERT_PREVPTR(prevptr, temp_ih, ih, ih_next);
661 
662 	intr_event_update(ie);
663 
664 	CTR3(KTR_INTR, "%s: added %s to %s", __func__, ih->ih_name,
665 	    ie->ie_name);
666 	mtx_unlock(&ie->ie_lock);
667 
668 	if (cookiep != NULL)
669 		*cookiep = ih;
670 	return (0);
671 }
672 
673 /*
674  * Append a description preceded by a ':' to the name of the specified
675  * interrupt handler.
676  */
677 int
678 intr_event_describe_handler(struct intr_event *ie, void *cookie,
679     const char *descr)
680 {
681 	struct intr_handler *ih;
682 	size_t space;
683 	char *start;
684 
685 	mtx_lock(&ie->ie_lock);
686 #ifdef INVARIANTS
687 	CK_SLIST_FOREACH(ih, &ie->ie_handlers, ih_next) {
688 		if (ih == cookie)
689 			break;
690 	}
691 	if (ih == NULL) {
692 		mtx_unlock(&ie->ie_lock);
693 		panic("handler %p not found in interrupt event %p", cookie, ie);
694 	}
695 #endif
696 	ih = cookie;
697 
698 	/*
699 	 * Look for an existing description by checking for an
700 	 * existing ":".  This assumes device names do not include
701 	 * colons.  If one is found, prepare to insert the new
702 	 * description at that point.  If one is not found, find the
703 	 * end of the name to use as the insertion point.
704 	 */
705 	start = strchr(ih->ih_name, ':');
706 	if (start == NULL)
707 		start = strchr(ih->ih_name, 0);
708 
709 	/*
710 	 * See if there is enough remaining room in the string for the
711 	 * description + ":".  The "- 1" leaves room for the trailing
712 	 * '\0'.  The "+ 1" accounts for the colon.
713 	 */
714 	space = sizeof(ih->ih_name) - (start - ih->ih_name) - 1;
715 	if (strlen(descr) + 1 > space) {
716 		mtx_unlock(&ie->ie_lock);
717 		return (ENOSPC);
718 	}
719 
720 	/* Append a colon followed by the description. */
721 	*start = ':';
722 	strcpy(start + 1, descr);
723 	intr_event_update(ie);
724 	mtx_unlock(&ie->ie_lock);
725 	return (0);
726 }
727 
728 /*
729  * Return the ie_source field from the intr_event an intr_handler is
730  * associated with.
731  */
732 void *
733 intr_handler_source(void *cookie)
734 {
735 	struct intr_handler *ih;
736 	struct intr_event *ie;
737 
738 	ih = (struct intr_handler *)cookie;
739 	if (ih == NULL)
740 		return (NULL);
741 	ie = ih->ih_event;
742 	KASSERT(ie != NULL,
743 	    ("interrupt handler \"%s\" has a NULL interrupt event",
744 	    ih->ih_name));
745 	return (ie->ie_source);
746 }
747 
748 /*
749  * If intr_event_handle() is running in the ISR context at the time of the call,
750  * then wait for it to complete.
751  */
752 static void
753 intr_event_barrier(struct intr_event *ie)
754 {
755 	int phase;
756 
757 	mtx_assert(&ie->ie_lock, MA_OWNED);
758 	phase = ie->ie_phase;
759 
760 	/*
761 	 * Switch phase to direct future interrupts to the other active counter.
762 	 * Make sure that any preceding stores are visible before the switch.
763 	 */
764 	KASSERT(ie->ie_active[!phase] == 0, ("idle phase has activity"));
765 	atomic_store_rel_int(&ie->ie_phase, !phase);
766 
767 	/*
768 	 * This code cooperates with wait-free iteration of ie_handlers
769 	 * in intr_event_handle.
770 	 * Make sure that the removal and the phase update are not reordered
771 	 * with the active count check.
772 	 * Note that no combination of acquire and release fences can provide
773 	 * that guarantee as Store->Load sequences can always be reordered.
774 	 */
775 	atomic_thread_fence_seq_cst();
776 
777 	/*
778 	 * Now wait on the inactive phase.
779 	 * The acquire fence is needed so that that all post-barrier accesses
780 	 * are after the check.
781 	 */
782 	while (ie->ie_active[phase] > 0)
783 		cpu_spinwait();
784 	atomic_thread_fence_acq();
785 }
786 
787 static void
788 intr_handler_barrier(struct intr_handler *handler)
789 {
790 	struct intr_event *ie;
791 
792 	ie = handler->ih_event;
793 	mtx_assert(&ie->ie_lock, MA_OWNED);
794 	KASSERT((handler->ih_flags & IH_DEAD) == 0,
795 	    ("update for a removed handler"));
796 
797 	if (ie->ie_thread == NULL) {
798 		intr_event_barrier(ie);
799 		return;
800 	}
801 	if ((handler->ih_flags & IH_CHANGED) == 0) {
802 		handler->ih_flags |= IH_CHANGED;
803 		intr_event_schedule_thread(ie, NULL);
804 	}
805 	while ((handler->ih_flags & IH_CHANGED) != 0)
806 		msleep(handler, &ie->ie_lock, 0, "ih_barr", 0);
807 }
808 
809 /*
810  * Sleep until an ithread finishes executing an interrupt handler.
811  *
812  * XXX Doesn't currently handle interrupt filters or fast interrupt
813  * handlers. This is intended for LinuxKPI drivers only.
814  * Do not use in BSD code.
815  */
816 void
817 _intr_drain(int irq)
818 {
819 	struct intr_event *ie;
820 	struct intr_thread *ithd;
821 	struct thread *td;
822 
823 	ie = intr_lookup(irq);
824 	if (ie == NULL)
825 		return;
826 	if (ie->ie_thread == NULL)
827 		return;
828 	ithd = ie->ie_thread;
829 	td = ithd->it_thread;
830 	/*
831 	 * We set the flag and wait for it to be cleared to avoid
832 	 * long delays with potentially busy interrupt handlers
833 	 * were we to only sample TD_AWAITING_INTR() every tick.
834 	 */
835 	thread_lock(td);
836 	if (!TD_AWAITING_INTR(td)) {
837 		ithd->it_flags |= IT_WAIT;
838 		while (ithd->it_flags & IT_WAIT) {
839 			thread_unlock(td);
840 			pause("idrain", 1);
841 			thread_lock(td);
842 		}
843 	}
844 	thread_unlock(td);
845 	return;
846 }
847 
848 int
849 intr_event_remove_handler(void *cookie)
850 {
851 	struct intr_handler *handler = (struct intr_handler *)cookie;
852 	struct intr_event *ie;
853 	struct intr_handler *ih;
854 	struct intr_handler **prevptr;
855 #ifdef notyet
856 	int dead;
857 #endif
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 	/*
879 	 * If there is no ithread, then directly remove the handler.  Note that
880 	 * intr_event_handle() iterates ie_handlers in a lock-less fashion, so
881 	 * care needs to be taken to keep ie_handlers consistent and to free
882 	 * the removed handler only when ie_handlers is quiescent.
883 	 */
884 	if (ie->ie_thread == NULL) {
885 		CK_SLIST_REMOVE_PREVPTR(prevptr, ih, ih_next);
886 		intr_event_barrier(ie);
887 		intr_event_update(ie);
888 		mtx_unlock(&ie->ie_lock);
889 		free(handler, M_ITHREAD);
890 		return (0);
891 	}
892 
893 	/*
894 	 * Let the interrupt thread do the job.
895 	 * The interrupt source is disabled when the interrupt thread is
896 	 * running, so it does not have to worry about interaction with
897 	 * intr_event_handle().
898 	 */
899 	KASSERT((handler->ih_flags & IH_DEAD) == 0,
900 	    ("duplicate handle remove"));
901 	handler->ih_flags |= IH_DEAD;
902 	intr_event_schedule_thread(ie, NULL);
903 	while (handler->ih_flags & IH_DEAD)
904 		msleep(handler, &ie->ie_lock, 0, "iev_rmh", 0);
905 	intr_event_update(ie);
906 
907 #ifdef notyet
908 	/*
909 	 * XXX: This could be bad in the case of ppbus(8).  Also, I think
910 	 * this could lead to races of stale data when servicing an
911 	 * interrupt.
912 	 */
913 	dead = 1;
914 	CK_SLIST_FOREACH(ih, &ie->ie_handlers, ih_next) {
915 		if (ih->ih_handler != NULL) {
916 			dead = 0;
917 			break;
918 		}
919 	}
920 	if (dead) {
921 		ithread_destroy(ie->ie_thread);
922 		ie->ie_thread = NULL;
923 	}
924 #endif
925 	mtx_unlock(&ie->ie_lock);
926 	free(handler, M_ITHREAD);
927 	return (0);
928 }
929 
930 int
931 intr_event_suspend_handler(void *cookie)
932 {
933 	struct intr_handler *handler = (struct intr_handler *)cookie;
934 	struct intr_event *ie;
935 
936 	if (handler == NULL)
937 		return (EINVAL);
938 	ie = handler->ih_event;
939 	KASSERT(ie != NULL,
940 	    ("interrupt handler \"%s\" has a NULL interrupt event",
941 	    handler->ih_name));
942 	mtx_lock(&ie->ie_lock);
943 	handler->ih_flags |= IH_SUSP;
944 	intr_handler_barrier(handler);
945 	mtx_unlock(&ie->ie_lock);
946 	return (0);
947 }
948 
949 int
950 intr_event_resume_handler(void *cookie)
951 {
952 	struct intr_handler *handler = (struct intr_handler *)cookie;
953 	struct intr_event *ie;
954 
955 	if (handler == NULL)
956 		return (EINVAL);
957 	ie = handler->ih_event;
958 	KASSERT(ie != NULL,
959 	    ("interrupt handler \"%s\" has a NULL interrupt event",
960 	    handler->ih_name));
961 
962 	/*
963 	 * intr_handler_barrier() acts not only as a barrier,
964 	 * it also allows to check for any pending interrupts.
965 	 */
966 	mtx_lock(&ie->ie_lock);
967 	handler->ih_flags &= ~IH_SUSP;
968 	intr_handler_barrier(handler);
969 	mtx_unlock(&ie->ie_lock);
970 	return (0);
971 }
972 
973 static int
974 intr_event_schedule_thread(struct intr_event *ie, struct trapframe *frame)
975 {
976 	struct intr_entropy entropy;
977 	struct intr_thread *it;
978 	struct thread *td;
979 	struct thread *ctd;
980 
981 	/*
982 	 * If no ithread or no handlers, then we have a stray interrupt.
983 	 */
984 	if (ie == NULL || CK_SLIST_EMPTY(&ie->ie_handlers) ||
985 	    ie->ie_thread == NULL)
986 		return (EINVAL);
987 
988 	ctd = curthread;
989 	it = ie->ie_thread;
990 	td = it->it_thread;
991 
992 	/*
993 	 * If any of the handlers for this ithread claim to be good
994 	 * sources of entropy, then gather some.
995 	 */
996 	if (ie->ie_hflags & IH_ENTROPY) {
997 		entropy.event = (uintptr_t)ie;
998 		entropy.td = ctd;
999 		random_harvest_queue(&entropy, sizeof(entropy), RANDOM_INTERRUPT);
1000 	}
1001 
1002 	KASSERT(td->td_proc != NULL, ("ithread %s has no process", ie->ie_name));
1003 
1004 	/*
1005 	 * Set it_need to tell the thread to keep running if it is already
1006 	 * running.  Then, lock the thread and see if we actually need to
1007 	 * put it on the runqueue.
1008 	 *
1009 	 * Use store_rel to arrange that the store to ih_need in
1010 	 * swi_sched() is before the store to it_need and prepare for
1011 	 * transfer of this order to loads in the ithread.
1012 	 */
1013 	atomic_store_rel_int(&it->it_need, 1);
1014 	thread_lock(td);
1015 	if (TD_AWAITING_INTR(td)) {
1016 #ifdef HWPMC_HOOKS
1017 		it->it_waiting = 0;
1018 		if (PMC_HOOK_INSTALLED_ANY())
1019 			PMC_SOFT_CALL_INTR_HLPR(schedule, frame);
1020 #endif
1021 		CTR3(KTR_INTR, "%s: schedule pid %d (%s)", __func__, td->td_proc->p_pid,
1022 		    td->td_name);
1023 		TD_CLR_IWAIT(td);
1024 		sched_add(td, SRQ_INTR);
1025 	} else {
1026 #ifdef HWPMC_HOOKS
1027 		it->it_waiting++;
1028 		if (PMC_HOOK_INSTALLED_ANY() &&
1029 		    (it->it_waiting >= intr_hwpmc_waiting_report_threshold))
1030 			PMC_SOFT_CALL_INTR_HLPR(waiting, frame);
1031 #endif
1032 		CTR5(KTR_INTR, "%s: pid %d (%s): it_need %d, state %d",
1033 		    __func__, td->td_proc->p_pid, td->td_name, it->it_need, TD_GET_STATE(td));
1034 		thread_unlock(td);
1035 	}
1036 
1037 	return (0);
1038 }
1039 
1040 /*
1041  * Allow interrupt event binding for software interrupt handlers -- a no-op,
1042  * since interrupts are generated in software rather than being directed by
1043  * a PIC.
1044  */
1045 static int
1046 swi_assign_cpu(void *arg, int cpu)
1047 {
1048 
1049 	return (0);
1050 }
1051 
1052 /*
1053  * Add a software interrupt handler to a specified event.  If a given event
1054  * is not specified, then a new event is created.
1055  */
1056 int
1057 swi_add(struct intr_event **eventp, const char *name, driver_intr_t handler,
1058 	    void *arg, int pri, enum intr_type flags, void **cookiep)
1059 {
1060 	struct intr_event *ie;
1061 	int error = 0;
1062 
1063 	if (flags & INTR_ENTROPY)
1064 		return (EINVAL);
1065 
1066 	ie = (eventp != NULL) ? *eventp : NULL;
1067 
1068 	if (ie != NULL) {
1069 		if (!(ie->ie_flags & IE_SOFT))
1070 			return (EINVAL);
1071 	} else {
1072 		error = intr_event_create(&ie, NULL, IE_SOFT, 0,
1073 		    NULL, NULL, NULL, swi_assign_cpu, "swi%d:", pri);
1074 		if (error)
1075 			return (error);
1076 		if (eventp != NULL)
1077 			*eventp = ie;
1078 	}
1079 	if (handler != NULL) {
1080 		error = intr_event_add_handler(ie, name, NULL, handler, arg,
1081 		    PI_SWI(pri), flags, cookiep);
1082 	}
1083 	return (error);
1084 }
1085 
1086 /*
1087  * Schedule a software interrupt thread.
1088  */
1089 void
1090 swi_sched(void *cookie, int flags)
1091 {
1092 	struct intr_handler *ih = (struct intr_handler *)cookie;
1093 	struct intr_event *ie = ih->ih_event;
1094 	struct intr_entropy entropy;
1095 	int error __unused;
1096 
1097 	CTR3(KTR_INTR, "swi_sched: %s %s need=%d", ie->ie_name, ih->ih_name,
1098 	    ih->ih_need);
1099 
1100 	if ((flags & SWI_FROMNMI) == 0) {
1101 		entropy.event = (uintptr_t)ih;
1102 		entropy.td = curthread;
1103 		random_harvest_queue(&entropy, sizeof(entropy), RANDOM_SWI);
1104 	}
1105 
1106 	/*
1107 	 * Set ih_need for this handler so that if the ithread is already
1108 	 * running it will execute this handler on the next pass.  Otherwise,
1109 	 * it will execute it the next time it runs.
1110 	 */
1111 	ih->ih_need = 1;
1112 
1113 	if (flags & SWI_DELAY)
1114 		return;
1115 
1116 	if (flags & SWI_FROMNMI) {
1117 #if defined(SMP) && (defined(__i386__) || defined(__amd64__))
1118 		KASSERT(ie == clk_intr_event,
1119 		    ("SWI_FROMNMI used not with clk_intr_event"));
1120 		ipi_self_from_nmi(IPI_SWI);
1121 #endif
1122 	} else {
1123 		VM_CNT_INC(v_soft);
1124 		error = intr_event_schedule_thread(ie, NULL);
1125 		KASSERT(error == 0, ("stray software interrupt"));
1126 	}
1127 }
1128 
1129 /*
1130  * Remove a software interrupt handler.  Currently this code does not
1131  * remove the associated interrupt event if it becomes empty.  Calling code
1132  * may do so manually via intr_event_destroy(), but that's not really
1133  * an optimal interface.
1134  */
1135 int
1136 swi_remove(void *cookie)
1137 {
1138 
1139 	return (intr_event_remove_handler(cookie));
1140 }
1141 
1142 static void
1143 intr_event_execute_handlers(struct proc *p, struct intr_event *ie)
1144 {
1145 	struct intr_handler *ih, *ihn, *ihp;
1146 
1147 	ihp = NULL;
1148 	CK_SLIST_FOREACH_SAFE(ih, &ie->ie_handlers, ih_next, ihn) {
1149 		/*
1150 		 * If this handler is marked for death, remove it from
1151 		 * the list of handlers and wake up the sleeper.
1152 		 */
1153 		if (ih->ih_flags & IH_DEAD) {
1154 			mtx_lock(&ie->ie_lock);
1155 			if (ihp == NULL)
1156 				CK_SLIST_REMOVE_HEAD(&ie->ie_handlers, ih_next);
1157 			else
1158 				CK_SLIST_REMOVE_AFTER(ihp, ih_next);
1159 			ih->ih_flags &= ~IH_DEAD;
1160 			wakeup(ih);
1161 			mtx_unlock(&ie->ie_lock);
1162 			continue;
1163 		}
1164 
1165 		/*
1166 		 * Now that we know that the current element won't be removed
1167 		 * update the previous element.
1168 		 */
1169 		ihp = ih;
1170 
1171 		if ((ih->ih_flags & IH_CHANGED) != 0) {
1172 			mtx_lock(&ie->ie_lock);
1173 			ih->ih_flags &= ~IH_CHANGED;
1174 			wakeup(ih);
1175 			mtx_unlock(&ie->ie_lock);
1176 		}
1177 
1178 		/* Skip filter only handlers */
1179 		if (ih->ih_handler == NULL)
1180 			continue;
1181 
1182 		/* Skip suspended handlers */
1183 		if ((ih->ih_flags & IH_SUSP) != 0)
1184 			continue;
1185 
1186 		/*
1187 		 * For software interrupt threads, we only execute
1188 		 * handlers that have their need flag set.  Hardware
1189 		 * interrupt threads always invoke all of their handlers.
1190 		 *
1191 		 * ih_need can only be 0 or 1.  Failed cmpset below
1192 		 * means that there is no request to execute handlers,
1193 		 * so a retry of the cmpset is not needed.
1194 		 */
1195 		if ((ie->ie_flags & IE_SOFT) != 0 &&
1196 		    atomic_cmpset_int(&ih->ih_need, 1, 0) == 0)
1197 			continue;
1198 
1199 		/* Execute this handler. */
1200 		CTR6(KTR_INTR, "%s: pid %d exec %p(%p) for %s flg=%x",
1201 		    __func__, p->p_pid, (void *)ih->ih_handler,
1202 		    ih->ih_argument, ih->ih_name, ih->ih_flags);
1203 
1204 		if (!(ih->ih_flags & IH_MPSAFE))
1205 			mtx_lock(&Giant);
1206 		ih->ih_handler(ih->ih_argument);
1207 		if (!(ih->ih_flags & IH_MPSAFE))
1208 			mtx_unlock(&Giant);
1209 	}
1210 }
1211 
1212 static void
1213 ithread_execute_handlers(struct proc *p, struct intr_event *ie)
1214 {
1215 
1216 	/* Interrupt handlers should not sleep. */
1217 	if (!(ie->ie_flags & IE_SOFT))
1218 		THREAD_NO_SLEEPING();
1219 	intr_event_execute_handlers(p, ie);
1220 	if (!(ie->ie_flags & IE_SOFT))
1221 		THREAD_SLEEPING_OK();
1222 
1223 	/*
1224 	 * Interrupt storm handling:
1225 	 *
1226 	 * If this interrupt source is currently storming, then throttle
1227 	 * it to only fire the handler once  per clock tick.
1228 	 *
1229 	 * If this interrupt source is not currently storming, but the
1230 	 * number of back to back interrupts exceeds the storm threshold,
1231 	 * then enter storming mode.
1232 	 */
1233 	if (intr_storm_threshold != 0 && ie->ie_count >= intr_storm_threshold &&
1234 	    !(ie->ie_flags & IE_SOFT)) {
1235 		/* Report the message only once every second. */
1236 		if (ppsratecheck(&ie->ie_warntm, &ie->ie_warncnt, 1)) {
1237 			printf(
1238 	"interrupt storm detected on \"%s\"; throttling interrupt source\n",
1239 			    ie->ie_name);
1240 		}
1241 		pause("istorm", 1);
1242 	} else
1243 		ie->ie_count++;
1244 
1245 	/*
1246 	 * Now that all the handlers have had a chance to run, reenable
1247 	 * the interrupt source.
1248 	 */
1249 	if (ie->ie_post_ithread != NULL)
1250 		ie->ie_post_ithread(ie->ie_source);
1251 }
1252 
1253 /*
1254  * This is the main code for interrupt threads.
1255  */
1256 static void
1257 ithread_loop(void *arg)
1258 {
1259 	struct epoch_tracker et;
1260 	struct intr_thread *ithd;
1261 	struct intr_event *ie;
1262 	struct thread *td;
1263 	struct proc *p;
1264 	int wake, epoch_count;
1265 	bool needs_epoch;
1266 
1267 	td = curthread;
1268 	p = td->td_proc;
1269 	ithd = (struct intr_thread *)arg;
1270 	KASSERT(ithd->it_thread == td,
1271 	    ("%s: ithread and proc linkage out of sync", __func__));
1272 	ie = ithd->it_event;
1273 	ie->ie_count = 0;
1274 	wake = 0;
1275 
1276 	/*
1277 	 * As long as we have interrupts outstanding, go through the
1278 	 * list of handlers, giving each one a go at it.
1279 	 */
1280 	for (;;) {
1281 		/*
1282 		 * If we are an orphaned thread, then just die.
1283 		 */
1284 		if (ithd->it_flags & IT_DEAD) {
1285 			CTR3(KTR_INTR, "%s: pid %d (%s) exiting", __func__,
1286 			    p->p_pid, td->td_name);
1287 			free(ithd, M_ITHREAD);
1288 			kthread_exit();
1289 		}
1290 
1291 		/*
1292 		 * Service interrupts.  If another interrupt arrives while
1293 		 * we are running, it will set it_need to note that we
1294 		 * should make another pass.
1295 		 *
1296 		 * The load_acq part of the following cmpset ensures
1297 		 * that the load of ih_need in ithread_execute_handlers()
1298 		 * is ordered after the load of it_need here.
1299 		 */
1300 		needs_epoch =
1301 		    (atomic_load_int(&ie->ie_hflags) & IH_NET) != 0;
1302 		if (needs_epoch) {
1303 			epoch_count = 0;
1304 			NET_EPOCH_ENTER(et);
1305 		}
1306 		while (atomic_cmpset_acq_int(&ithd->it_need, 1, 0) != 0) {
1307 			ithread_execute_handlers(p, ie);
1308 			if (needs_epoch &&
1309 			    ++epoch_count >= intr_epoch_batch) {
1310 				NET_EPOCH_EXIT(et);
1311 				epoch_count = 0;
1312 				NET_EPOCH_ENTER(et);
1313 			}
1314 		}
1315 		if (needs_epoch)
1316 			NET_EPOCH_EXIT(et);
1317 		WITNESS_WARN(WARN_PANIC, NULL, "suspending ithread");
1318 		mtx_assert(&Giant, MA_NOTOWNED);
1319 
1320 		/*
1321 		 * Processed all our interrupts.  Now get the sched
1322 		 * lock.  This may take a while and it_need may get
1323 		 * set again, so we have to check it again.
1324 		 */
1325 		thread_lock(td);
1326 		if (atomic_load_acq_int(&ithd->it_need) == 0 &&
1327 		    (ithd->it_flags & (IT_DEAD | IT_WAIT)) == 0) {
1328 			TD_SET_IWAIT(td);
1329 			ie->ie_count = 0;
1330 			mi_switch(SW_VOL | SWT_IWAIT);
1331 		} else {
1332 			if (ithd->it_flags & IT_WAIT) {
1333 				wake = 1;
1334 				ithd->it_flags &= ~IT_WAIT;
1335 			}
1336 			thread_unlock(td);
1337 		}
1338 		if (wake) {
1339 			wakeup(ithd);
1340 			wake = 0;
1341 		}
1342 	}
1343 }
1344 
1345 /*
1346  * Main interrupt handling body.
1347  *
1348  * Input:
1349  * o ie:                        the event connected to this interrupt.
1350  * o frame:                     some archs (i.e. i386) pass a frame to some.
1351  *                              handlers as their main argument.
1352  * Return value:
1353  * o 0:                         everything ok.
1354  * o EINVAL:                    stray interrupt.
1355  */
1356 int
1357 intr_event_handle(struct intr_event *ie, struct trapframe *frame)
1358 {
1359 	struct intr_handler *ih;
1360 	struct trapframe *oldframe;
1361 	struct thread *td;
1362 	int phase;
1363 	int ret;
1364 	bool filter, thread;
1365 
1366 	td = curthread;
1367 
1368 #ifdef KSTACK_USAGE_PROF
1369 	intr_prof_stack_use(td, frame);
1370 #endif
1371 
1372 	/* An interrupt with no event or handlers is a stray interrupt. */
1373 	if (ie == NULL || CK_SLIST_EMPTY(&ie->ie_handlers))
1374 		return (EINVAL);
1375 
1376 	/*
1377 	 * Execute fast interrupt handlers directly.
1378 	 * To support clock handlers, if a handler registers
1379 	 * with a NULL argument, then we pass it a pointer to
1380 	 * a trapframe as its argument.
1381 	 */
1382 	td->td_intr_nesting_level++;
1383 	filter = false;
1384 	thread = false;
1385 	ret = 0;
1386 	critical_enter();
1387 	oldframe = td->td_intr_frame;
1388 	td->td_intr_frame = frame;
1389 
1390 	phase = ie->ie_phase;
1391 	atomic_add_int(&ie->ie_active[phase], 1);
1392 
1393 	/*
1394 	 * This fence is required to ensure that no later loads are
1395 	 * re-ordered before the ie_active store.
1396 	 */
1397 	atomic_thread_fence_seq_cst();
1398 
1399 	CK_SLIST_FOREACH(ih, &ie->ie_handlers, ih_next) {
1400 		if ((ih->ih_flags & IH_SUSP) != 0)
1401 			continue;
1402 		if ((ie->ie_flags & IE_SOFT) != 0 && ih->ih_need == 0)
1403 			continue;
1404 		if (ih->ih_filter == NULL) {
1405 			thread = true;
1406 			continue;
1407 		}
1408 		CTR4(KTR_INTR, "%s: exec %p(%p) for %s", __func__,
1409 		    ih->ih_filter, ih->ih_argument == NULL ? frame :
1410 		    ih->ih_argument, ih->ih_name);
1411 		if (ih->ih_argument == NULL)
1412 			ret = ih->ih_filter(frame);
1413 		else
1414 			ret = ih->ih_filter(ih->ih_argument);
1415 #ifdef HWPMC_HOOKS
1416 		PMC_SOFT_CALL_TF( , , intr, all, frame);
1417 #endif
1418 		KASSERT(ret == FILTER_STRAY ||
1419 		    ((ret & (FILTER_SCHEDULE_THREAD | FILTER_HANDLED)) != 0 &&
1420 		    (ret & ~(FILTER_SCHEDULE_THREAD | FILTER_HANDLED)) == 0),
1421 		    ("%s: incorrect return value %#x from %s", __func__, ret,
1422 		    ih->ih_name));
1423 		filter = filter || ret == FILTER_HANDLED;
1424 #ifdef HWPMC_HOOKS
1425 		if (ret & FILTER_SCHEDULE_THREAD)
1426 			PMC_SOFT_CALL_TF( , , intr, ithread, frame);
1427 		else if (ret & FILTER_HANDLED)
1428 			PMC_SOFT_CALL_TF( , , intr, filter, frame);
1429 		else if (ret == FILTER_STRAY)
1430 			PMC_SOFT_CALL_TF( , , intr, stray, frame);
1431 #endif
1432 
1433 		/*
1434 		 * Wrapper handler special handling:
1435 		 *
1436 		 * in some particular cases (like pccard and pccbb),
1437 		 * the _real_ device handler is wrapped in a couple of
1438 		 * functions - a filter wrapper and an ithread wrapper.
1439 		 * In this case (and just in this case), the filter wrapper
1440 		 * could ask the system to schedule the ithread and mask
1441 		 * the interrupt source if the wrapped handler is composed
1442 		 * of just an ithread handler.
1443 		 *
1444 		 * TODO: write a generic wrapper to avoid people rolling
1445 		 * their own.
1446 		 */
1447 		if (!thread) {
1448 			if (ret == FILTER_SCHEDULE_THREAD)
1449 				thread = true;
1450 		}
1451 	}
1452 	atomic_add_rel_int(&ie->ie_active[phase], -1);
1453 
1454 	td->td_intr_frame = oldframe;
1455 
1456 	if (thread) {
1457 		if (ie->ie_pre_ithread != NULL)
1458 			ie->ie_pre_ithread(ie->ie_source);
1459 	} else {
1460 		if (ie->ie_post_filter != NULL)
1461 			ie->ie_post_filter(ie->ie_source);
1462 	}
1463 
1464 	/* Schedule the ithread if needed. */
1465 	if (thread) {
1466 		int error __unused;
1467 
1468 		error =  intr_event_schedule_thread(ie, frame);
1469 		KASSERT(error == 0, ("bad stray interrupt"));
1470 	}
1471 	critical_exit();
1472 	td->td_intr_nesting_level--;
1473 #ifdef notyet
1474 	/* The interrupt is not aknowledged by any filter and has no ithread. */
1475 	if (!thread && !filter)
1476 		return (EINVAL);
1477 #endif
1478 	return (0);
1479 }
1480 
1481 #ifdef DDB
1482 /*
1483  * Dump details about an interrupt handler
1484  */
1485 static void
1486 db_dump_intrhand(struct intr_handler *ih)
1487 {
1488 	int comma;
1489 
1490 	db_printf("\t%-10s ", ih->ih_name);
1491 	switch (ih->ih_pri) {
1492 	case PI_REALTIME:
1493 		db_printf("CLK ");
1494 		break;
1495 	case PI_AV:
1496 		db_printf("AV  ");
1497 		break;
1498 	case PI_TTY:
1499 		db_printf("TTY ");
1500 		break;
1501 	case PI_NET:
1502 		db_printf("NET ");
1503 		break;
1504 	case PI_DISK:
1505 		db_printf("DISK");
1506 		break;
1507 	case PI_DULL:
1508 		db_printf("DULL");
1509 		break;
1510 	default:
1511 		if (ih->ih_pri >= PI_SOFT)
1512 			db_printf("SWI ");
1513 		else
1514 			db_printf("%4u", ih->ih_pri);
1515 		break;
1516 	}
1517 	db_printf(" ");
1518 	if (ih->ih_filter != NULL) {
1519 		db_printf("[F]");
1520 		db_printsym((uintptr_t)ih->ih_filter, DB_STGY_PROC);
1521 	}
1522 	if (ih->ih_handler != NULL) {
1523 		if (ih->ih_filter != NULL)
1524 			db_printf(",");
1525 		db_printf("[H]");
1526 		db_printsym((uintptr_t)ih->ih_handler, DB_STGY_PROC);
1527 	}
1528 	db_printf("(%p)", ih->ih_argument);
1529 	if (ih->ih_need ||
1530 	    (ih->ih_flags & (IH_EXCLUSIVE | IH_ENTROPY | IH_DEAD |
1531 	    IH_MPSAFE)) != 0) {
1532 		db_printf(" {");
1533 		comma = 0;
1534 		if (ih->ih_flags & IH_EXCLUSIVE) {
1535 			if (comma)
1536 				db_printf(", ");
1537 			db_printf("EXCL");
1538 			comma = 1;
1539 		}
1540 		if (ih->ih_flags & IH_ENTROPY) {
1541 			if (comma)
1542 				db_printf(", ");
1543 			db_printf("ENTROPY");
1544 			comma = 1;
1545 		}
1546 		if (ih->ih_flags & IH_DEAD) {
1547 			if (comma)
1548 				db_printf(", ");
1549 			db_printf("DEAD");
1550 			comma = 1;
1551 		}
1552 		if (ih->ih_flags & IH_MPSAFE) {
1553 			if (comma)
1554 				db_printf(", ");
1555 			db_printf("MPSAFE");
1556 			comma = 1;
1557 		}
1558 		if (ih->ih_need) {
1559 			if (comma)
1560 				db_printf(", ");
1561 			db_printf("NEED");
1562 		}
1563 		db_printf("}");
1564 	}
1565 	db_printf("\n");
1566 }
1567 
1568 /*
1569  * Dump details about a event.
1570  */
1571 void
1572 db_dump_intr_event(struct intr_event *ie, int handlers)
1573 {
1574 	struct intr_handler *ih;
1575 	struct intr_thread *it;
1576 	int comma;
1577 
1578 	db_printf("%s ", ie->ie_fullname);
1579 	it = ie->ie_thread;
1580 	if (it != NULL)
1581 		db_printf("(pid %d)", it->it_thread->td_proc->p_pid);
1582 	else
1583 		db_printf("(no thread)");
1584 	if ((ie->ie_flags & (IE_SOFT | IE_ADDING_THREAD)) != 0 ||
1585 	    (it != NULL && it->it_need)) {
1586 		db_printf(" {");
1587 		comma = 0;
1588 		if (ie->ie_flags & IE_SOFT) {
1589 			db_printf("SOFT");
1590 			comma = 1;
1591 		}
1592 		if (ie->ie_flags & IE_ADDING_THREAD) {
1593 			if (comma)
1594 				db_printf(", ");
1595 			db_printf("ADDING_THREAD");
1596 			comma = 1;
1597 		}
1598 		if (it != NULL && it->it_need) {
1599 			if (comma)
1600 				db_printf(", ");
1601 			db_printf("NEED");
1602 		}
1603 		db_printf("}");
1604 	}
1605 	db_printf("\n");
1606 
1607 	if (handlers)
1608 		CK_SLIST_FOREACH(ih, &ie->ie_handlers, ih_next)
1609 		    db_dump_intrhand(ih);
1610 }
1611 
1612 /*
1613  * Dump data about interrupt handlers
1614  */
1615 DB_SHOW_COMMAND(intr, db_show_intr)
1616 {
1617 	struct intr_event *ie;
1618 	int all, verbose;
1619 
1620 	verbose = strchr(modif, 'v') != NULL;
1621 	all = strchr(modif, 'a') != NULL;
1622 	TAILQ_FOREACH(ie, &event_list, ie_list) {
1623 		if (!all && CK_SLIST_EMPTY(&ie->ie_handlers))
1624 			continue;
1625 		db_dump_intr_event(ie, verbose);
1626 		if (db_pager_quit)
1627 			break;
1628 	}
1629 }
1630 #endif /* DDB */
1631 
1632 /*
1633  * Start standard software interrupt threads
1634  */
1635 static void
1636 start_softintr(void *dummy)
1637 {
1638 
1639 	if (swi_add(&clk_intr_event, "clk", NULL, NULL, SWI_CLOCK,
1640 	    INTR_MPSAFE, NULL))
1641 		panic("died while creating clk swi ithread");
1642 	if (swi_add(NULL, "vm", swi_vm, NULL, SWI_VM, INTR_MPSAFE, &vm_ih))
1643 		panic("died while creating vm swi ithread");
1644 }
1645 SYSINIT(start_softintr, SI_SUB_SOFTINTR, SI_ORDER_FIRST, start_softintr,
1646     NULL);
1647 
1648 /*
1649  * Sysctls used by systat and others: hw.intrnames and hw.intrcnt.
1650  * The data for this machine dependent, and the declarations are in machine
1651  * dependent code.  The layout of intrnames and intrcnt however is machine
1652  * independent.
1653  *
1654  * We do not know the length of intrcnt and intrnames at compile time, so
1655  * calculate things at run time.
1656  */
1657 static int
1658 sysctl_intrnames(SYSCTL_HANDLER_ARGS)
1659 {
1660 	return (sysctl_handle_opaque(oidp, intrnames, sintrnames, req));
1661 }
1662 
1663 SYSCTL_PROC(_hw, OID_AUTO, intrnames,
1664     CTLTYPE_OPAQUE | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
1665     sysctl_intrnames, "",
1666     "Interrupt Names");
1667 
1668 static int
1669 sysctl_intrcnt(SYSCTL_HANDLER_ARGS)
1670 {
1671 #ifdef SCTL_MASK32
1672 	uint32_t *intrcnt32;
1673 	unsigned i;
1674 	int error;
1675 
1676 	if (req->flags & SCTL_MASK32) {
1677 		if (!req->oldptr)
1678 			return (sysctl_handle_opaque(oidp, NULL, sintrcnt / 2, req));
1679 		intrcnt32 = malloc(sintrcnt / 2, M_TEMP, M_NOWAIT);
1680 		if (intrcnt32 == NULL)
1681 			return (ENOMEM);
1682 		for (i = 0; i < sintrcnt / sizeof (u_long); i++)
1683 			intrcnt32[i] = intrcnt[i];
1684 		error = sysctl_handle_opaque(oidp, intrcnt32, sintrcnt / 2, req);
1685 		free(intrcnt32, M_TEMP);
1686 		return (error);
1687 	}
1688 #endif
1689 	return (sysctl_handle_opaque(oidp, intrcnt, sintrcnt, req));
1690 }
1691 
1692 SYSCTL_PROC(_hw, OID_AUTO, intrcnt,
1693     CTLTYPE_OPAQUE | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
1694     sysctl_intrcnt, "",
1695     "Interrupt Counts");
1696 
1697 #ifdef DDB
1698 /*
1699  * DDB command to dump the interrupt statistics.
1700  */
1701 DB_SHOW_COMMAND(intrcnt, db_show_intrcnt)
1702 {
1703 	u_long *i;
1704 	char *cp;
1705 	u_int j;
1706 
1707 	cp = intrnames;
1708 	j = 0;
1709 	for (i = intrcnt; j < (sintrcnt / sizeof(u_long)) && !db_pager_quit;
1710 	    i++, j++) {
1711 		if (*cp == '\0')
1712 			break;
1713 		if (*i != 0)
1714 			db_printf("%s\t%lu\n", cp, *i);
1715 		cp += strlen(cp) + 1;
1716 	}
1717 }
1718 #endif
1719