xref: /freebsd/sys/kern/kern_intr.c (revision 657729a89dd578d8cfc70d6616f5c65a48a8b33a)
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 proc *intrproc;
93 
94 static MALLOC_DEFINE(M_ITHREAD, "ithread", "Interrupt Threads");
95 
96 static int intr_storm_threshold = 0;
97 SYSCTL_INT(_hw, OID_AUTO, intr_storm_threshold, CTLFLAG_RWTUN,
98     &intr_storm_threshold, 0,
99     "Number of consecutive interrupts before storm protection is enabled");
100 static int intr_epoch_batch = 1000;
101 SYSCTL_INT(_hw, OID_AUTO, intr_epoch_batch, CTLFLAG_RWTUN, &intr_epoch_batch,
102     0, "Maximum interrupt handler executions without re-entering epoch(9)");
103 #ifdef HWPMC_HOOKS
104 static int intr_hwpmc_waiting_report_threshold = 1;
105 SYSCTL_INT(_hw, OID_AUTO, intr_hwpmc_waiting_report_threshold, CTLFLAG_RWTUN,
106     &intr_hwpmc_waiting_report_threshold, 1,
107     "Threshold for reporting number of events in a workq");
108 #define	PMC_HOOK_INSTALLED_ANY() __predict_false(pmc_hook != NULL)
109 #endif
110 static TAILQ_HEAD(, intr_event) event_list =
111     TAILQ_HEAD_INITIALIZER(event_list);
112 static struct mtx event_lock;
113 MTX_SYSINIT(intr_event_list, &event_lock, "intr event list", MTX_DEF);
114 
115 static void	intr_event_update(struct intr_event *ie);
116 static int	intr_event_schedule_thread(struct intr_event *ie, struct trapframe *frame);
117 static struct intr_thread *ithread_create(const char *name);
118 static void	ithread_destroy(struct intr_thread *ithread);
119 static void	ithread_execute_handlers(struct proc *p,
120 		    struct intr_event *ie);
121 static void	ithread_loop(void *);
122 static void	ithread_update(struct intr_thread *ithd);
123 static void	start_softintr(void *);
124 
125 #ifdef HWPMC_HOOKS
126 #include <sys/pmckern.h>
127 PMC_SOFT_DEFINE( , , intr, all);
128 PMC_SOFT_DEFINE( , , intr, ithread);
129 PMC_SOFT_DEFINE( , , intr, filter);
130 PMC_SOFT_DEFINE( , , intr, stray);
131 PMC_SOFT_DEFINE( , , intr, schedule);
132 PMC_SOFT_DEFINE( , , intr, waiting);
133 
134 #define PMC_SOFT_CALL_INTR_HLPR(event, frame)			\
135 do {					\
136 	if (frame != NULL)					\
137 		PMC_SOFT_CALL_TF( , , intr, event, frame);	\
138 	else							\
139 		PMC_SOFT_CALL( , , intr, event);		\
140 } while (0)
141 #endif
142 
143 /* Map an interrupt type to an ithread priority. */
144 u_char
145 intr_priority(enum intr_type flags)
146 {
147 	u_char pri;
148 
149 	flags &= (INTR_TYPE_TTY | INTR_TYPE_BIO | INTR_TYPE_NET |
150 	    INTR_TYPE_CAM | INTR_TYPE_MISC | INTR_TYPE_CLK | INTR_TYPE_AV);
151 	switch (flags) {
152 	case INTR_TYPE_TTY:
153 		pri = PI_TTY;
154 		break;
155 	case INTR_TYPE_BIO:
156 		pri = PI_DISK;
157 		break;
158 	case INTR_TYPE_NET:
159 		pri = PI_NET;
160 		break;
161 	case INTR_TYPE_CAM:
162 		pri = PI_DISK;
163 		break;
164 	case INTR_TYPE_AV:
165 		pri = PI_AV;
166 		break;
167 	case INTR_TYPE_CLK:
168 		pri = PI_REALTIME;
169 		break;
170 	case INTR_TYPE_MISC:
171 		pri = PI_DULL;          /* don't care */
172 		break;
173 	default:
174 		/* We didn't specify an interrupt level. */
175 		panic("intr_priority: no interrupt type in flags");
176 	}
177 
178 	return pri;
179 }
180 
181 /*
182  * Update an ithread based on the associated intr_event.
183  */
184 static void
185 ithread_update(struct intr_thread *ithd)
186 {
187 	struct intr_event *ie;
188 	struct thread *td;
189 	u_char pri;
190 
191 	ie = ithd->it_event;
192 	td = ithd->it_thread;
193 	mtx_assert(&ie->ie_lock, MA_OWNED);
194 
195 	/* Determine the overall priority of this event. */
196 	if (CK_SLIST_EMPTY(&ie->ie_handlers))
197 		pri = PRI_MAX_ITHD;
198 	else
199 		pri = CK_SLIST_FIRST(&ie->ie_handlers)->ih_pri;
200 
201 	/* Update name and priority. */
202 	strlcpy(td->td_name, ie->ie_fullname, sizeof(td->td_name));
203 #ifdef KTR
204 	sched_clear_tdname(td);
205 #endif
206 	thread_lock(td);
207 	sched_ithread_prio(td, pri);
208 	thread_unlock(td);
209 }
210 
211 /*
212  * Regenerate the full name of an interrupt event and update its priority.
213  */
214 static void
215 intr_event_update(struct intr_event *ie)
216 {
217 	struct intr_handler *ih;
218 	char *last;
219 	int missed, space, flags;
220 
221 	/* Start off with no entropy and just the name of the event. */
222 	mtx_assert(&ie->ie_lock, MA_OWNED);
223 	strlcpy(ie->ie_fullname, ie->ie_name, sizeof(ie->ie_fullname));
224 	flags = 0;
225 	missed = 0;
226 	space = 1;
227 
228 	/* Run through all the handlers updating values. */
229 	CK_SLIST_FOREACH(ih, &ie->ie_handlers, ih_next) {
230 		if (strlen(ie->ie_fullname) + strlen(ih->ih_name) + 1 <
231 		    sizeof(ie->ie_fullname)) {
232 			strcat(ie->ie_fullname, " ");
233 			strcat(ie->ie_fullname, ih->ih_name);
234 			space = 0;
235 		} else
236 			missed++;
237 		flags |= ih->ih_flags;
238 	}
239 	ie->ie_hflags = flags;
240 
241 	/*
242 	 * If there is only one handler and its name is too long, just copy in
243 	 * as much of the end of the name (includes the unit number) as will
244 	 * fit.  Otherwise, we have multiple handlers and not all of the names
245 	 * will fit.  Add +'s to indicate missing names.  If we run out of room
246 	 * and still have +'s to add, change the last character from a + to a *.
247 	 */
248 	if (missed == 1 && space == 1) {
249 		ih = CK_SLIST_FIRST(&ie->ie_handlers);
250 		missed = strlen(ie->ie_fullname) + strlen(ih->ih_name) + 2 -
251 		    sizeof(ie->ie_fullname);
252 		strcat(ie->ie_fullname, (missed == 0) ? " " : "-");
253 		strcat(ie->ie_fullname, &ih->ih_name[missed]);
254 		missed = 0;
255 	}
256 	last = &ie->ie_fullname[sizeof(ie->ie_fullname) - 2];
257 	while (missed-- > 0) {
258 		if (strlen(ie->ie_fullname) + 1 == sizeof(ie->ie_fullname)) {
259 			if (*last == '+') {
260 				*last = '*';
261 				break;
262 			} else
263 				*last = '+';
264 		} else if (space) {
265 			strcat(ie->ie_fullname, " +");
266 			space = 0;
267 		} else
268 			strcat(ie->ie_fullname, "+");
269 	}
270 
271 	/*
272 	 * If this event has an ithread, update it's priority and
273 	 * name.
274 	 */
275 	if (ie->ie_thread != NULL)
276 		ithread_update(ie->ie_thread);
277 	CTR2(KTR_INTR, "%s: updated %s", __func__, ie->ie_fullname);
278 }
279 
280 int
281 intr_event_create(struct intr_event **event, void *source, int flags, int irq,
282     void (*pre_ithread)(void *), void (*post_ithread)(void *),
283     void (*post_filter)(void *), int (*assign_cpu)(void *, int),
284     const char *fmt, ...)
285 {
286 	struct intr_event *ie;
287 	va_list ap;
288 
289 	/* The only valid flag during creation is IE_SOFT. */
290 	if ((flags & ~IE_SOFT) != 0)
291 		return (EINVAL);
292 	ie = malloc(sizeof(struct intr_event), M_ITHREAD, M_WAITOK | M_ZERO);
293 	ie->ie_source = source;
294 	ie->ie_pre_ithread = pre_ithread;
295 	ie->ie_post_ithread = post_ithread;
296 	ie->ie_post_filter = post_filter;
297 	ie->ie_assign_cpu = assign_cpu;
298 	ie->ie_flags = flags;
299 	ie->ie_irq = irq;
300 	ie->ie_cpu = NOCPU;
301 	CK_SLIST_INIT(&ie->ie_handlers);
302 	mtx_init(&ie->ie_lock, "intr event", NULL, MTX_DEF);
303 
304 	va_start(ap, fmt);
305 	vsnprintf(ie->ie_name, sizeof(ie->ie_name), fmt, ap);
306 	va_end(ap);
307 	strlcpy(ie->ie_fullname, ie->ie_name, sizeof(ie->ie_fullname));
308 	mtx_lock(&event_lock);
309 	TAILQ_INSERT_TAIL(&event_list, ie, ie_list);
310 	mtx_unlock(&event_lock);
311 	if (event != NULL)
312 		*event = ie;
313 	CTR2(KTR_INTR, "%s: created %s", __func__, ie->ie_name);
314 	return (0);
315 }
316 
317 /*
318  * Bind an interrupt event to the specified CPU.  Note that not all
319  * platforms support binding an interrupt to a CPU.  For those
320  * platforms this request will fail.  Using a cpu id of NOCPU unbinds
321  * the interrupt event.
322  */
323 static int
324 _intr_event_bind(struct intr_event *ie, int cpu, bool bindirq, bool bindithread)
325 {
326 	lwpid_t id;
327 	int error;
328 
329 	/* Need a CPU to bind to. */
330 	if (cpu != NOCPU && CPU_ABSENT(cpu))
331 		return (EINVAL);
332 
333 	if (ie->ie_assign_cpu == NULL)
334 		return (EOPNOTSUPP);
335 
336 	error = priv_check(curthread, PRIV_SCHED_CPUSET_INTR);
337 	if (error)
338 		return (error);
339 
340 	/*
341 	 * If we have any ithreads try to set their mask first to verify
342 	 * permissions, etc.
343 	 */
344 	if (bindithread) {
345 		mtx_lock(&ie->ie_lock);
346 		if (ie->ie_thread != NULL) {
347 			id = ie->ie_thread->it_thread->td_tid;
348 			mtx_unlock(&ie->ie_lock);
349 			error = cpuset_setithread(id, cpu);
350 			if (error)
351 				return (error);
352 		} else
353 			mtx_unlock(&ie->ie_lock);
354 	}
355 	if (bindirq)
356 		error = ie->ie_assign_cpu(ie->ie_source, cpu);
357 	if (error) {
358 		if (bindithread) {
359 			mtx_lock(&ie->ie_lock);
360 			if (ie->ie_thread != NULL) {
361 				cpu = ie->ie_cpu;
362 				id = ie->ie_thread->it_thread->td_tid;
363 				mtx_unlock(&ie->ie_lock);
364 				(void)cpuset_setithread(id, cpu);
365 			} else
366 				mtx_unlock(&ie->ie_lock);
367 		}
368 		return (error);
369 	}
370 
371 	if (bindirq) {
372 		mtx_lock(&ie->ie_lock);
373 		ie->ie_cpu = cpu;
374 		mtx_unlock(&ie->ie_lock);
375 	}
376 
377 	return (error);
378 }
379 
380 /*
381  * Bind an interrupt event to the specified CPU.  For supported platforms, any
382  * associated ithreads as well as the primary interrupt context will be bound
383  * to the specificed CPU.
384  */
385 int
386 intr_event_bind(struct intr_event *ie, int cpu)
387 {
388 
389 	return (_intr_event_bind(ie, cpu, true, true));
390 }
391 
392 /*
393  * Bind an interrupt event to the specified CPU, but do not bind associated
394  * ithreads.
395  */
396 int
397 intr_event_bind_irqonly(struct intr_event *ie, int cpu)
398 {
399 
400 	return (_intr_event_bind(ie, cpu, true, false));
401 }
402 
403 /*
404  * Bind an interrupt event's ithread to the specified CPU.
405  */
406 int
407 intr_event_bind_ithread(struct intr_event *ie, int cpu)
408 {
409 
410 	return (_intr_event_bind(ie, cpu, false, true));
411 }
412 
413 /*
414  * Bind an interrupt event's ithread to the specified cpuset.
415  */
416 int
417 intr_event_bind_ithread_cpuset(struct intr_event *ie, cpuset_t *cs)
418 {
419 	lwpid_t id;
420 
421 	mtx_lock(&ie->ie_lock);
422 	if (ie->ie_thread != NULL) {
423 		id = ie->ie_thread->it_thread->td_tid;
424 		mtx_unlock(&ie->ie_lock);
425 		return (cpuset_setthread(id, cs));
426 	} else {
427 		mtx_unlock(&ie->ie_lock);
428 	}
429 	return (ENODEV);
430 }
431 
432 static struct intr_event *
433 intr_lookup(int irq)
434 {
435 	struct intr_event *ie;
436 
437 	mtx_lock(&event_lock);
438 	TAILQ_FOREACH(ie, &event_list, ie_list)
439 		if (ie->ie_irq == irq &&
440 		    (ie->ie_flags & IE_SOFT) == 0 &&
441 		    CK_SLIST_FIRST(&ie->ie_handlers) != NULL)
442 			break;
443 	mtx_unlock(&event_lock);
444 	return (ie);
445 }
446 
447 int
448 intr_setaffinity(int irq, int mode, void *m)
449 {
450 	struct intr_event *ie;
451 	cpuset_t *mask;
452 	int cpu, n;
453 
454 	mask = m;
455 	cpu = NOCPU;
456 	/*
457 	 * If we're setting all cpus we can unbind.  Otherwise make sure
458 	 * only one cpu is in the set.
459 	 */
460 	if (CPU_CMP(cpuset_root, mask)) {
461 		for (n = 0; n < CPU_SETSIZE; n++) {
462 			if (!CPU_ISSET(n, mask))
463 				continue;
464 			if (cpu != NOCPU)
465 				return (EINVAL);
466 			cpu = n;
467 		}
468 	}
469 	ie = intr_lookup(irq);
470 	if (ie == NULL)
471 		return (ESRCH);
472 	switch (mode) {
473 	case CPU_WHICH_IRQ:
474 		return (intr_event_bind(ie, cpu));
475 	case CPU_WHICH_INTRHANDLER:
476 		return (intr_event_bind_irqonly(ie, cpu));
477 	case CPU_WHICH_ITHREAD:
478 		return (intr_event_bind_ithread(ie, cpu));
479 	default:
480 		return (EINVAL);
481 	}
482 }
483 
484 int
485 intr_getaffinity(int irq, int mode, void *m)
486 {
487 	struct intr_event *ie;
488 	struct thread *td;
489 	struct proc *p;
490 	cpuset_t *mask;
491 	lwpid_t id;
492 	int error;
493 
494 	mask = m;
495 	ie = intr_lookup(irq);
496 	if (ie == NULL)
497 		return (ESRCH);
498 
499 	error = 0;
500 	CPU_ZERO(mask);
501 	switch (mode) {
502 	case CPU_WHICH_IRQ:
503 	case CPU_WHICH_INTRHANDLER:
504 		mtx_lock(&ie->ie_lock);
505 		if (ie->ie_cpu == NOCPU)
506 			CPU_COPY(cpuset_root, mask);
507 		else
508 			CPU_SET(ie->ie_cpu, mask);
509 		mtx_unlock(&ie->ie_lock);
510 		break;
511 	case CPU_WHICH_ITHREAD:
512 		mtx_lock(&ie->ie_lock);
513 		if (ie->ie_thread == NULL) {
514 			mtx_unlock(&ie->ie_lock);
515 			CPU_COPY(cpuset_root, mask);
516 		} else {
517 			id = ie->ie_thread->it_thread->td_tid;
518 			mtx_unlock(&ie->ie_lock);
519 			error = cpuset_which(CPU_WHICH_TID, id, &p, &td, NULL);
520 			if (error != 0)
521 				return (error);
522 			CPU_COPY(&td->td_cpuset->cs_mask, mask);
523 			PROC_UNLOCK(p);
524 		}
525 	default:
526 		return (EINVAL);
527 	}
528 	return (0);
529 }
530 
531 int
532 intr_event_destroy(struct intr_event *ie)
533 {
534 
535 	if (ie == NULL)
536 		return (EINVAL);
537 
538 	mtx_lock(&event_lock);
539 	mtx_lock(&ie->ie_lock);
540 	if (!CK_SLIST_EMPTY(&ie->ie_handlers)) {
541 		mtx_unlock(&ie->ie_lock);
542 		mtx_unlock(&event_lock);
543 		return (EBUSY);
544 	}
545 	TAILQ_REMOVE(&event_list, ie, ie_list);
546 #ifndef notyet
547 	if (ie->ie_thread != NULL) {
548 		ithread_destroy(ie->ie_thread);
549 		ie->ie_thread = NULL;
550 	}
551 #endif
552 	mtx_unlock(&ie->ie_lock);
553 	mtx_unlock(&event_lock);
554 	mtx_destroy(&ie->ie_lock);
555 	free(ie, M_ITHREAD);
556 	return (0);
557 }
558 
559 static struct intr_thread *
560 ithread_create(const char *name)
561 {
562 	struct intr_thread *ithd;
563 	struct thread *td;
564 	int error;
565 
566 	ithd = malloc(sizeof(struct intr_thread), M_ITHREAD, M_WAITOK | M_ZERO);
567 
568 	error = kproc_kthread_add(ithread_loop, ithd, &intrproc,
569 		    &td, RFSTOPPED | RFHIGHPID,
570 		    0, "intr", "%s", name);
571 	if (error)
572 		panic("kproc_create() failed with %d", error);
573 	thread_lock(td);
574 	sched_class(td, PRI_ITHD);
575 	TD_SET_IWAIT(td);
576 	thread_unlock(td);
577 	td->td_pflags |= TDP_ITHREAD;
578 	ithd->it_thread = td;
579 	CTR2(KTR_INTR, "%s: created %s", __func__, name);
580 	return (ithd);
581 }
582 
583 static void
584 ithread_destroy(struct intr_thread *ithread)
585 {
586 	struct thread *td;
587 
588 	CTR2(KTR_INTR, "%s: killing %s", __func__, ithread->it_event->ie_name);
589 	td = ithread->it_thread;
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 }
598 
599 int
600 intr_event_add_handler(struct intr_event *ie, const char *name,
601     driver_filter_t filter, driver_intr_t handler, void *arg, u_char pri,
602     enum intr_type flags, void **cookiep)
603 {
604 	struct intr_handler *ih, *temp_ih;
605 	struct intr_handler **prevptr;
606 	struct intr_thread *it;
607 
608 	if (ie == NULL || name == NULL || (handler == NULL && filter == NULL))
609 		return (EINVAL);
610 
611 	/* Allocate and populate an interrupt handler structure. */
612 	ih = malloc(sizeof(struct intr_handler), M_ITHREAD, M_WAITOK | M_ZERO);
613 	ih->ih_filter = filter;
614 	ih->ih_handler = handler;
615 	ih->ih_argument = arg;
616 	strlcpy(ih->ih_name, name, sizeof(ih->ih_name));
617 	ih->ih_event = ie;
618 	ih->ih_pri = pri;
619 	if (flags & INTR_EXCL)
620 		ih->ih_flags = IH_EXCLUSIVE;
621 	if (flags & INTR_MPSAFE)
622 		ih->ih_flags |= IH_MPSAFE;
623 	if (flags & INTR_ENTROPY)
624 		ih->ih_flags |= IH_ENTROPY;
625 	if (flags & INTR_TYPE_NET)
626 		ih->ih_flags |= IH_NET;
627 
628 	/* We can only have one exclusive handler in a event. */
629 	mtx_lock(&ie->ie_lock);
630 	if (!CK_SLIST_EMPTY(&ie->ie_handlers)) {
631 		if ((flags & INTR_EXCL) ||
632 		    (CK_SLIST_FIRST(&ie->ie_handlers)->ih_flags & IH_EXCLUSIVE)) {
633 			mtx_unlock(&ie->ie_lock);
634 			free(ih, M_ITHREAD);
635 			return (EINVAL);
636 		}
637 	}
638 
639 	/* Create a thread if we need one. */
640 	while (ie->ie_thread == NULL && handler != NULL) {
641 		if (ie->ie_flags & IE_ADDING_THREAD)
642 			msleep(ie, &ie->ie_lock, 0, "ithread", 0);
643 		else {
644 			ie->ie_flags |= IE_ADDING_THREAD;
645 			mtx_unlock(&ie->ie_lock);
646 			it = ithread_create("intr: newborn");
647 			mtx_lock(&ie->ie_lock);
648 			ie->ie_flags &= ~IE_ADDING_THREAD;
649 			ie->ie_thread = it;
650 			it->it_event = ie;
651 			ithread_update(it);
652 			wakeup(ie);
653 		}
654 	}
655 
656 	/* Add the new handler to the event in priority order. */
657 	CK_SLIST_FOREACH_PREVPTR(temp_ih, prevptr, &ie->ie_handlers, ih_next) {
658 		if (temp_ih->ih_pri > ih->ih_pri)
659 			break;
660 	}
661 	CK_SLIST_INSERT_PREVPTR(prevptr, temp_ih, ih, ih_next);
662 
663 	intr_event_update(ie);
664 
665 	CTR3(KTR_INTR, "%s: added %s to %s", __func__, ih->ih_name,
666 	    ie->ie_name);
667 	mtx_unlock(&ie->ie_lock);
668 
669 	if (cookiep != NULL)
670 		*cookiep = ih;
671 	return (0);
672 }
673 
674 /*
675  * Append a description preceded by a ':' to the name of the specified
676  * interrupt handler.
677  */
678 int
679 intr_event_describe_handler(struct intr_event *ie, void *cookie,
680     const char *descr)
681 {
682 	struct intr_handler *ih;
683 	size_t space;
684 	char *start;
685 
686 	mtx_lock(&ie->ie_lock);
687 #ifdef INVARIANTS
688 	CK_SLIST_FOREACH(ih, &ie->ie_handlers, ih_next) {
689 		if (ih == cookie)
690 			break;
691 	}
692 	if (ih == NULL) {
693 		mtx_unlock(&ie->ie_lock);
694 		panic("handler %p not found in interrupt event %p", cookie, ie);
695 	}
696 #endif
697 	ih = cookie;
698 
699 	/*
700 	 * Look for an existing description by checking for an
701 	 * existing ":".  This assumes device names do not include
702 	 * colons.  If one is found, prepare to insert the new
703 	 * description at that point.  If one is not found, find the
704 	 * end of the name to use as the insertion point.
705 	 */
706 	start = strchr(ih->ih_name, ':');
707 	if (start == NULL)
708 		start = strchr(ih->ih_name, 0);
709 
710 	/*
711 	 * See if there is enough remaining room in the string for the
712 	 * description + ":".  The "- 1" leaves room for the trailing
713 	 * '\0'.  The "+ 1" accounts for the colon.
714 	 */
715 	space = sizeof(ih->ih_name) - (start - ih->ih_name) - 1;
716 	if (strlen(descr) + 1 > space) {
717 		mtx_unlock(&ie->ie_lock);
718 		return (ENOSPC);
719 	}
720 
721 	/* Append a colon followed by the description. */
722 	*start = ':';
723 	strcpy(start + 1, descr);
724 	intr_event_update(ie);
725 	mtx_unlock(&ie->ie_lock);
726 	return (0);
727 }
728 
729 /*
730  * Return the ie_source field from the intr_event an intr_handler is
731  * associated with.
732  */
733 void *
734 intr_handler_source(void *cookie)
735 {
736 	struct intr_handler *ih;
737 	struct intr_event *ie;
738 
739 	ih = (struct intr_handler *)cookie;
740 	if (ih == NULL)
741 		return (NULL);
742 	ie = ih->ih_event;
743 	KASSERT(ie != NULL,
744 	    ("interrupt handler \"%s\" has a NULL interrupt event",
745 	    ih->ih_name));
746 	return (ie->ie_source);
747 }
748 
749 /*
750  * If intr_event_handle() is running in the ISR context at the time of the call,
751  * then wait for it to complete.
752  */
753 static void
754 intr_event_barrier(struct intr_event *ie)
755 {
756 	int phase;
757 
758 	mtx_assert(&ie->ie_lock, MA_OWNED);
759 	phase = ie->ie_phase;
760 
761 	/*
762 	 * Switch phase to direct future interrupts to the other active counter.
763 	 * Make sure that any preceding stores are visible before the switch.
764 	 */
765 	KASSERT(ie->ie_active[!phase] == 0, ("idle phase has activity"));
766 	atomic_store_rel_int(&ie->ie_phase, !phase);
767 
768 	/*
769 	 * This code cooperates with wait-free iteration of ie_handlers
770 	 * in intr_event_handle.
771 	 * Make sure that the removal and the phase update are not reordered
772 	 * with the active count check.
773 	 * Note that no combination of acquire and release fences can provide
774 	 * that guarantee as Store->Load sequences can always be reordered.
775 	 */
776 	atomic_thread_fence_seq_cst();
777 
778 	/*
779 	 * Now wait on the inactive phase.
780 	 * The acquire fence is needed so that all post-barrier accesses
781 	 * are after the check.
782 	 */
783 	while (ie->ie_active[phase] > 0)
784 		cpu_spinwait();
785 	atomic_thread_fence_acq();
786 }
787 
788 static void
789 intr_handler_barrier(struct intr_handler *handler)
790 {
791 	struct intr_event *ie;
792 
793 	ie = handler->ih_event;
794 	mtx_assert(&ie->ie_lock, MA_OWNED);
795 	KASSERT((handler->ih_flags & IH_DEAD) == 0,
796 	    ("update for a removed handler"));
797 
798 	if (ie->ie_thread == NULL) {
799 		intr_event_barrier(ie);
800 		return;
801 	}
802 	if ((handler->ih_flags & IH_CHANGED) == 0) {
803 		handler->ih_flags |= IH_CHANGED;
804 		intr_event_schedule_thread(ie, NULL);
805 	}
806 	while ((handler->ih_flags & IH_CHANGED) != 0)
807 		msleep(handler, &ie->ie_lock, 0, "ih_barr", 0);
808 }
809 
810 /*
811  * Sleep until an ithread finishes executing an interrupt handler.
812  *
813  * XXX Doesn't currently handle interrupt filters or fast interrupt
814  * handlers. This is intended for LinuxKPI drivers only.
815  * Do not use in BSD code.
816  */
817 void
818 _intr_drain(int irq)
819 {
820 	struct intr_event *ie;
821 	struct intr_thread *ithd;
822 	struct thread *td;
823 
824 	ie = intr_lookup(irq);
825 	if (ie == NULL)
826 		return;
827 	if (ie->ie_thread == NULL)
828 		return;
829 	ithd = ie->ie_thread;
830 	td = ithd->it_thread;
831 	/*
832 	 * We set the flag and wait for it to be cleared to avoid
833 	 * long delays with potentially busy interrupt handlers
834 	 * were we to only sample TD_AWAITING_INTR() every tick.
835 	 */
836 	thread_lock(td);
837 	if (!TD_AWAITING_INTR(td)) {
838 		ithd->it_flags |= IT_WAIT;
839 		while (ithd->it_flags & IT_WAIT) {
840 			thread_unlock(td);
841 			pause("idrain", 1);
842 			thread_lock(td);
843 		}
844 	}
845 	thread_unlock(td);
846 	return;
847 }
848 
849 int
850 intr_event_remove_handler(void *cookie)
851 {
852 	struct intr_handler *handler = (struct intr_handler *)cookie;
853 	struct intr_event *ie;
854 	struct intr_handler *ih;
855 	struct intr_handler **prevptr;
856 #ifdef notyet
857 	int dead;
858 #endif
859 
860 	if (handler == NULL)
861 		return (EINVAL);
862 	ie = handler->ih_event;
863 	KASSERT(ie != NULL,
864 	    ("interrupt handler \"%s\" has a NULL interrupt event",
865 	    handler->ih_name));
866 
867 	mtx_lock(&ie->ie_lock);
868 	CTR3(KTR_INTR, "%s: removing %s from %s", __func__, handler->ih_name,
869 	    ie->ie_name);
870 	CK_SLIST_FOREACH_PREVPTR(ih, prevptr, &ie->ie_handlers, ih_next) {
871 		if (ih == handler)
872 			break;
873 	}
874 	if (ih == NULL) {
875 		panic("interrupt handler \"%s\" not found in "
876 		    "interrupt event \"%s\"", handler->ih_name, ie->ie_name);
877 	}
878 
879 	/*
880 	 * If there is no ithread, then directly remove the handler.  Note that
881 	 * intr_event_handle() iterates ie_handlers in a lock-less fashion, so
882 	 * care needs to be taken to keep ie_handlers consistent and to free
883 	 * the removed handler only when ie_handlers is quiescent.
884 	 */
885 	if (ie->ie_thread == NULL) {
886 		CK_SLIST_REMOVE_PREVPTR(prevptr, ih, ih_next);
887 		intr_event_barrier(ie);
888 		intr_event_update(ie);
889 		mtx_unlock(&ie->ie_lock);
890 		free(handler, M_ITHREAD);
891 		return (0);
892 	}
893 
894 	/*
895 	 * Let the interrupt thread do the job.
896 	 * The interrupt source is disabled when the interrupt thread is
897 	 * running, so it does not have to worry about interaction with
898 	 * intr_event_handle().
899 	 */
900 	KASSERT((handler->ih_flags & IH_DEAD) == 0,
901 	    ("duplicate handle remove"));
902 	handler->ih_flags |= IH_DEAD;
903 	intr_event_schedule_thread(ie, NULL);
904 	while (handler->ih_flags & IH_DEAD)
905 		msleep(handler, &ie->ie_lock, 0, "iev_rmh", 0);
906 	intr_event_update(ie);
907 
908 #ifdef notyet
909 	/*
910 	 * XXX: This could be bad in the case of ppbus(8).  Also, I think
911 	 * this could lead to races of stale data when servicing an
912 	 * interrupt.
913 	 */
914 	dead = 1;
915 	CK_SLIST_FOREACH(ih, &ie->ie_handlers, ih_next) {
916 		if (ih->ih_handler != NULL) {
917 			dead = 0;
918 			break;
919 		}
920 	}
921 	if (dead) {
922 		ithread_destroy(ie->ie_thread);
923 		ie->ie_thread = NULL;
924 	}
925 #endif
926 	mtx_unlock(&ie->ie_lock);
927 	free(handler, M_ITHREAD);
928 	return (0);
929 }
930 
931 int
932 intr_event_suspend_handler(void *cookie)
933 {
934 	struct intr_handler *handler = (struct intr_handler *)cookie;
935 	struct intr_event *ie;
936 
937 	if (handler == NULL)
938 		return (EINVAL);
939 	ie = handler->ih_event;
940 	KASSERT(ie != NULL,
941 	    ("interrupt handler \"%s\" has a NULL interrupt event",
942 	    handler->ih_name));
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 int
951 intr_event_resume_handler(void *cookie)
952 {
953 	struct intr_handler *handler = (struct intr_handler *)cookie;
954 	struct intr_event *ie;
955 
956 	if (handler == NULL)
957 		return (EINVAL);
958 	ie = handler->ih_event;
959 	KASSERT(ie != NULL,
960 	    ("interrupt handler \"%s\" has a NULL interrupt event",
961 	    handler->ih_name));
962 
963 	/*
964 	 * intr_handler_barrier() acts not only as a barrier,
965 	 * it also allows to check for any pending interrupts.
966 	 */
967 	mtx_lock(&ie->ie_lock);
968 	handler->ih_flags &= ~IH_SUSP;
969 	intr_handler_barrier(handler);
970 	mtx_unlock(&ie->ie_lock);
971 	return (0);
972 }
973 
974 static int
975 intr_event_schedule_thread(struct intr_event *ie, struct trapframe *frame)
976 {
977 	struct intr_entropy entropy;
978 	struct intr_thread *it;
979 	struct thread *td;
980 	struct thread *ctd;
981 
982 	/*
983 	 * If no ithread or no handlers, then we have a stray interrupt.
984 	 */
985 	if (ie == NULL || CK_SLIST_EMPTY(&ie->ie_handlers) ||
986 	    ie->ie_thread == NULL)
987 		return (EINVAL);
988 
989 	ctd = curthread;
990 	it = ie->ie_thread;
991 	td = it->it_thread;
992 
993 	/*
994 	 * If any of the handlers for this ithread claim to be good
995 	 * sources of entropy, then gather some.
996 	 */
997 	if (ie->ie_hflags & IH_ENTROPY) {
998 		entropy.event = (uintptr_t)ie;
999 		entropy.td = ctd;
1000 		random_harvest_queue(&entropy, sizeof(entropy), RANDOM_INTERRUPT);
1001 	}
1002 
1003 	KASSERT(td->td_proc != NULL, ("ithread %s has no process", ie->ie_name));
1004 
1005 	/*
1006 	 * Set it_need to tell the thread to keep running if it is already
1007 	 * running.  Then, lock the thread and see if we actually need to
1008 	 * put it on the runqueue.
1009 	 *
1010 	 * Use store_rel to arrange that the store to ih_need in
1011 	 * swi_sched() is before the store to it_need and prepare for
1012 	 * transfer of this order to loads in the ithread.
1013 	 */
1014 	atomic_store_rel_int(&it->it_need, 1);
1015 	thread_lock(td);
1016 	if (TD_AWAITING_INTR(td)) {
1017 #ifdef HWPMC_HOOKS
1018 		it->it_waiting = 0;
1019 		if (PMC_HOOK_INSTALLED_ANY())
1020 			PMC_SOFT_CALL_INTR_HLPR(schedule, frame);
1021 #endif
1022 		CTR3(KTR_INTR, "%s: schedule pid %d (%s)", __func__, td->td_proc->p_pid,
1023 		    td->td_name);
1024 		TD_CLR_IWAIT(td);
1025 		sched_wakeup(td, SRQ_INTR);
1026 	} else {
1027 #ifdef HWPMC_HOOKS
1028 		it->it_waiting++;
1029 		if (PMC_HOOK_INSTALLED_ANY() &&
1030 		    (it->it_waiting >= intr_hwpmc_waiting_report_threshold))
1031 			PMC_SOFT_CALL_INTR_HLPR(waiting, frame);
1032 #endif
1033 		CTR5(KTR_INTR, "%s: pid %d (%s): it_need %d, state %d",
1034 		    __func__, td->td_proc->p_pid, td->td_name, it->it_need, TD_GET_STATE(td));
1035 		thread_unlock(td);
1036 	}
1037 
1038 	return (0);
1039 }
1040 
1041 /*
1042  * Allow interrupt event binding for software interrupt handlers -- a no-op,
1043  * since interrupts are generated in software rather than being directed by
1044  * a PIC.
1045  */
1046 static int
1047 swi_assign_cpu(void *arg, int cpu)
1048 {
1049 
1050 	return (0);
1051 }
1052 
1053 /*
1054  * Add a software interrupt handler to a specified event.  If a given event
1055  * is not specified, then a new event is created.
1056  */
1057 int
1058 swi_add(struct intr_event **eventp, const char *name, driver_intr_t handler,
1059 	    void *arg, int pri, enum intr_type flags, void **cookiep)
1060 {
1061 	struct intr_event *ie;
1062 	int error = 0;
1063 
1064 	if (flags & INTR_ENTROPY)
1065 		return (EINVAL);
1066 
1067 	ie = (eventp != NULL) ? *eventp : NULL;
1068 
1069 	if (ie != NULL) {
1070 		if (!(ie->ie_flags & IE_SOFT))
1071 			return (EINVAL);
1072 	} else {
1073 		error = intr_event_create(&ie, NULL, IE_SOFT, 0,
1074 		    NULL, NULL, NULL, swi_assign_cpu, "swi%d:", pri);
1075 		if (error)
1076 			return (error);
1077 		if (eventp != NULL)
1078 			*eventp = ie;
1079 	}
1080 	if (handler != NULL) {
1081 		error = intr_event_add_handler(ie, name, NULL, handler, arg,
1082 		    PI_SWI(pri), flags, cookiep);
1083 	}
1084 	return (error);
1085 }
1086 
1087 /*
1088  * Schedule a software interrupt thread.
1089  */
1090 void
1091 swi_sched(void *cookie, int flags)
1092 {
1093 	struct intr_handler *ih = (struct intr_handler *)cookie;
1094 	struct intr_event *ie = ih->ih_event;
1095 	struct intr_entropy entropy;
1096 	int error __unused;
1097 
1098 	CTR3(KTR_INTR, "swi_sched: %s %s need=%d", ie->ie_name, ih->ih_name,
1099 	    ih->ih_need);
1100 
1101 	if ((flags & SWI_FROMNMI) == 0) {
1102 		entropy.event = (uintptr_t)ih;
1103 		entropy.td = curthread;
1104 		random_harvest_queue(&entropy, sizeof(entropy), RANDOM_SWI);
1105 	}
1106 
1107 	/*
1108 	 * Set ih_need for this handler so that if the ithread is already
1109 	 * running it will execute this handler on the next pass.  Otherwise,
1110 	 * it will execute it the next time it runs.
1111 	 */
1112 	ih->ih_need = 1;
1113 
1114 	if (flags & SWI_DELAY)
1115 		return;
1116 
1117 	if (flags & SWI_FROMNMI) {
1118 #if defined(SMP) && (defined(__i386__) || defined(__amd64__))
1119 		KASSERT(ie == clk_intr_event,
1120 		    ("SWI_FROMNMI used not with clk_intr_event"));
1121 		ipi_self_from_nmi(IPI_SWI);
1122 #endif
1123 	} else {
1124 		VM_CNT_INC(v_soft);
1125 		error = intr_event_schedule_thread(ie, NULL);
1126 		KASSERT(error == 0, ("stray software interrupt"));
1127 	}
1128 }
1129 
1130 /*
1131  * Remove a software interrupt handler.  Currently this code does not
1132  * remove the associated interrupt event if it becomes empty.  Calling code
1133  * may do so manually via intr_event_destroy(), but that's not really
1134  * an optimal interface.
1135  */
1136 int
1137 swi_remove(void *cookie)
1138 {
1139 
1140 	return (intr_event_remove_handler(cookie));
1141 }
1142 
1143 static void
1144 intr_event_execute_handlers(struct proc *p, struct intr_event *ie)
1145 {
1146 	struct intr_handler *ih, *ihn, *ihp;
1147 
1148 	ihp = NULL;
1149 	CK_SLIST_FOREACH_SAFE(ih, &ie->ie_handlers, ih_next, ihn) {
1150 		/*
1151 		 * If this handler is marked for death, remove it from
1152 		 * the list of handlers and wake up the sleeper.
1153 		 */
1154 		if (ih->ih_flags & IH_DEAD) {
1155 			mtx_lock(&ie->ie_lock);
1156 			if (ihp == NULL)
1157 				CK_SLIST_REMOVE_HEAD(&ie->ie_handlers, ih_next);
1158 			else
1159 				CK_SLIST_REMOVE_AFTER(ihp, ih_next);
1160 			ih->ih_flags &= ~IH_DEAD;
1161 			wakeup(ih);
1162 			mtx_unlock(&ie->ie_lock);
1163 			continue;
1164 		}
1165 
1166 		/*
1167 		 * Now that we know that the current element won't be removed
1168 		 * update the previous element.
1169 		 */
1170 		ihp = ih;
1171 
1172 		if ((ih->ih_flags & IH_CHANGED) != 0) {
1173 			mtx_lock(&ie->ie_lock);
1174 			ih->ih_flags &= ~IH_CHANGED;
1175 			wakeup(ih);
1176 			mtx_unlock(&ie->ie_lock);
1177 		}
1178 
1179 		/* Skip filter only handlers */
1180 		if (ih->ih_handler == NULL)
1181 			continue;
1182 
1183 		/* Skip suspended handlers */
1184 		if ((ih->ih_flags & IH_SUSP) != 0)
1185 			continue;
1186 
1187 		/*
1188 		 * For software interrupt threads, we only execute
1189 		 * handlers that have their need flag set.  Hardware
1190 		 * interrupt threads always invoke all of their handlers.
1191 		 *
1192 		 * ih_need can only be 0 or 1.  Failed cmpset below
1193 		 * means that there is no request to execute handlers,
1194 		 * so a retry of the cmpset is not needed.
1195 		 */
1196 		if ((ie->ie_flags & IE_SOFT) != 0 &&
1197 		    atomic_cmpset_int(&ih->ih_need, 1, 0) == 0)
1198 			continue;
1199 
1200 		/* Execute this handler. */
1201 		CTR6(KTR_INTR, "%s: pid %d exec %p(%p) for %s flg=%x",
1202 		    __func__, p->p_pid, (void *)ih->ih_handler,
1203 		    ih->ih_argument, ih->ih_name, ih->ih_flags);
1204 
1205 		if (!(ih->ih_flags & IH_MPSAFE))
1206 			mtx_lock(&Giant);
1207 		ih->ih_handler(ih->ih_argument);
1208 		if (!(ih->ih_flags & IH_MPSAFE))
1209 			mtx_unlock(&Giant);
1210 	}
1211 }
1212 
1213 static void
1214 ithread_execute_handlers(struct proc *p, struct intr_event *ie)
1215 {
1216 
1217 	/* Interrupt handlers should not sleep. */
1218 	if (!(ie->ie_flags & IE_SOFT))
1219 		THREAD_NO_SLEEPING();
1220 	intr_event_execute_handlers(p, ie);
1221 	if (!(ie->ie_flags & IE_SOFT))
1222 		THREAD_SLEEPING_OK();
1223 
1224 	/*
1225 	 * Interrupt storm handling:
1226 	 *
1227 	 * If this interrupt source is currently storming, then throttle
1228 	 * it to only fire the handler once  per clock tick.
1229 	 *
1230 	 * If this interrupt source is not currently storming, but the
1231 	 * number of back to back interrupts exceeds the storm threshold,
1232 	 * then enter storming mode.
1233 	 */
1234 	if (intr_storm_threshold != 0 && ie->ie_count >= intr_storm_threshold &&
1235 	    !(ie->ie_flags & IE_SOFT)) {
1236 		/* Report the message only once every second. */
1237 		if (ppsratecheck(&ie->ie_warntm, &ie->ie_warncnt, 1)) {
1238 			printf(
1239 	"interrupt storm detected on \"%s\"; throttling interrupt source\n",
1240 			    ie->ie_name);
1241 		}
1242 		pause("istorm", 1);
1243 	} else
1244 		ie->ie_count++;
1245 
1246 	/*
1247 	 * Now that all the handlers have had a chance to run, reenable
1248 	 * the interrupt source.
1249 	 */
1250 	if (ie->ie_post_ithread != NULL)
1251 		ie->ie_post_ithread(ie->ie_source);
1252 }
1253 
1254 /*
1255  * This is the main code for interrupt threads.
1256  */
1257 static void
1258 ithread_loop(void *arg)
1259 {
1260 	struct epoch_tracker et;
1261 	struct intr_thread *ithd;
1262 	struct intr_event *ie;
1263 	struct thread *td;
1264 	struct proc *p;
1265 	int wake, epoch_count;
1266 	bool needs_epoch;
1267 
1268 	td = curthread;
1269 	p = td->td_proc;
1270 	ithd = (struct intr_thread *)arg;
1271 	KASSERT(ithd->it_thread == td,
1272 	    ("%s: ithread and proc linkage out of sync", __func__));
1273 	ie = ithd->it_event;
1274 	ie->ie_count = 0;
1275 	wake = 0;
1276 
1277 	/*
1278 	 * As long as we have interrupts outstanding, go through the
1279 	 * list of handlers, giving each one a go at it.
1280 	 */
1281 	for (;;) {
1282 		/*
1283 		 * If we are an orphaned thread, then just die.
1284 		 */
1285 		if (ithd->it_flags & IT_DEAD) {
1286 			CTR3(KTR_INTR, "%s: pid %d (%s) exiting", __func__,
1287 			    p->p_pid, td->td_name);
1288 			free(ithd, M_ITHREAD);
1289 			kthread_exit();
1290 		}
1291 
1292 		/*
1293 		 * Service interrupts.  If another interrupt arrives while
1294 		 * we are running, it will set it_need to note that we
1295 		 * should make another pass.
1296 		 *
1297 		 * The load_acq part of the following cmpset ensures
1298 		 * that the load of ih_need in ithread_execute_handlers()
1299 		 * is ordered after the load of it_need here.
1300 		 */
1301 		needs_epoch =
1302 		    (atomic_load_int(&ie->ie_hflags) & IH_NET) != 0;
1303 		if (needs_epoch) {
1304 			epoch_count = 0;
1305 			NET_EPOCH_ENTER(et);
1306 		}
1307 		while (atomic_cmpset_acq_int(&ithd->it_need, 1, 0) != 0) {
1308 			ithread_execute_handlers(p, ie);
1309 			if (needs_epoch &&
1310 			    ++epoch_count >= intr_epoch_batch) {
1311 				NET_EPOCH_EXIT(et);
1312 				epoch_count = 0;
1313 				NET_EPOCH_ENTER(et);
1314 			}
1315 		}
1316 		if (needs_epoch)
1317 			NET_EPOCH_EXIT(et);
1318 		WITNESS_WARN(WARN_PANIC, NULL, "suspending ithread");
1319 		mtx_assert(&Giant, MA_NOTOWNED);
1320 
1321 		/*
1322 		 * Processed all our interrupts.  Now get the sched
1323 		 * lock.  This may take a while and it_need may get
1324 		 * set again, so we have to check it again.
1325 		 */
1326 		thread_lock(td);
1327 		if (atomic_load_acq_int(&ithd->it_need) == 0 &&
1328 		    (ithd->it_flags & (IT_DEAD | IT_WAIT)) == 0) {
1329 			TD_SET_IWAIT(td);
1330 			ie->ie_count = 0;
1331 			mi_switch(SW_VOL | SWT_IWAIT);
1332 		} else {
1333 			if (ithd->it_flags & IT_WAIT) {
1334 				wake = 1;
1335 				ithd->it_flags &= ~IT_WAIT;
1336 			}
1337 			thread_unlock(td);
1338 		}
1339 		if (wake) {
1340 			wakeup(ithd);
1341 			wake = 0;
1342 		}
1343 	}
1344 }
1345 
1346 /*
1347  * Main interrupt handling body.
1348  *
1349  * Input:
1350  * o ie:                        the event connected to this interrupt.
1351  * o frame:                     some archs (i.e. i386) pass a frame to some.
1352  *                              handlers as their main argument.
1353  * Return value:
1354  * o 0:                         everything ok.
1355  * o EINVAL:                    stray interrupt.
1356  */
1357 int
1358 intr_event_handle(struct intr_event *ie, struct trapframe *frame)
1359 {
1360 	struct intr_handler *ih;
1361 	struct trapframe *oldframe;
1362 	struct thread *td;
1363 	int phase;
1364 	int ret;
1365 	bool filter, thread;
1366 
1367 	td = curthread;
1368 
1369 #ifdef KSTACK_USAGE_PROF
1370 	intr_prof_stack_use(td, frame);
1371 #endif
1372 
1373 	/* An interrupt with no event or handlers is a stray interrupt. */
1374 	if (ie == NULL || CK_SLIST_EMPTY(&ie->ie_handlers))
1375 		return (EINVAL);
1376 
1377 	/*
1378 	 * Execute fast interrupt handlers directly.
1379 	 * To support clock handlers, if a handler registers
1380 	 * with a NULL argument, then we pass it a pointer to
1381 	 * a trapframe as its argument.
1382 	 */
1383 	td->td_intr_nesting_level++;
1384 	filter = false;
1385 	thread = false;
1386 	ret = 0;
1387 	critical_enter();
1388 	oldframe = td->td_intr_frame;
1389 	td->td_intr_frame = frame;
1390 
1391 	phase = ie->ie_phase;
1392 	atomic_add_int(&ie->ie_active[phase], 1);
1393 
1394 	/*
1395 	 * This fence is required to ensure that no later loads are
1396 	 * re-ordered before the ie_active store.
1397 	 */
1398 	atomic_thread_fence_seq_cst();
1399 
1400 	CK_SLIST_FOREACH(ih, &ie->ie_handlers, ih_next) {
1401 		if ((ih->ih_flags & IH_SUSP) != 0)
1402 			continue;
1403 		if ((ie->ie_flags & IE_SOFT) != 0 && ih->ih_need == 0)
1404 			continue;
1405 		if (ih->ih_filter == NULL) {
1406 			thread = true;
1407 			continue;
1408 		}
1409 		CTR4(KTR_INTR, "%s: exec %p(%p) for %s", __func__,
1410 		    ih->ih_filter, ih->ih_argument == NULL ? frame :
1411 		    ih->ih_argument, ih->ih_name);
1412 		if (ih->ih_argument == NULL)
1413 			ret = ih->ih_filter(frame);
1414 		else
1415 			ret = ih->ih_filter(ih->ih_argument);
1416 #ifdef HWPMC_HOOKS
1417 		PMC_SOFT_CALL_TF( , , intr, all, frame);
1418 #endif
1419 		KASSERT(ret == FILTER_STRAY ||
1420 		    ((ret & (FILTER_SCHEDULE_THREAD | FILTER_HANDLED)) != 0 &&
1421 		    (ret & ~(FILTER_SCHEDULE_THREAD | FILTER_HANDLED)) == 0),
1422 		    ("%s: incorrect return value %#x from %s", __func__, ret,
1423 		    ih->ih_name));
1424 		filter = filter || ret == FILTER_HANDLED;
1425 #ifdef HWPMC_HOOKS
1426 		if (ret & FILTER_SCHEDULE_THREAD)
1427 			PMC_SOFT_CALL_TF( , , intr, ithread, frame);
1428 		else if (ret & FILTER_HANDLED)
1429 			PMC_SOFT_CALL_TF( , , intr, filter, frame);
1430 		else if (ret == FILTER_STRAY)
1431 			PMC_SOFT_CALL_TF( , , intr, stray, frame);
1432 #endif
1433 
1434 		/*
1435 		 * Wrapper handler special handling:
1436 		 *
1437 		 * in some particular cases (like pccard and pccbb),
1438 		 * the _real_ device handler is wrapped in a couple of
1439 		 * functions - a filter wrapper and an ithread wrapper.
1440 		 * In this case (and just in this case), the filter wrapper
1441 		 * could ask the system to schedule the ithread and mask
1442 		 * the interrupt source if the wrapped handler is composed
1443 		 * of just an ithread handler.
1444 		 *
1445 		 * TODO: write a generic wrapper to avoid people rolling
1446 		 * their own.
1447 		 */
1448 		if (!thread) {
1449 			if (ret == FILTER_SCHEDULE_THREAD)
1450 				thread = true;
1451 		}
1452 	}
1453 	atomic_add_rel_int(&ie->ie_active[phase], -1);
1454 
1455 	td->td_intr_frame = oldframe;
1456 
1457 	if (thread) {
1458 		if (ie->ie_pre_ithread != NULL)
1459 			ie->ie_pre_ithread(ie->ie_source);
1460 	} else {
1461 		if (ie->ie_post_filter != NULL)
1462 			ie->ie_post_filter(ie->ie_source);
1463 	}
1464 
1465 	/* Schedule the ithread if needed. */
1466 	if (thread) {
1467 		int error __unused;
1468 
1469 		error =  intr_event_schedule_thread(ie, frame);
1470 		KASSERT(error == 0, ("bad stray interrupt"));
1471 	}
1472 	critical_exit();
1473 	td->td_intr_nesting_level--;
1474 #ifdef notyet
1475 	/* The interrupt is not aknowledged by any filter and has no ithread. */
1476 	if (!thread && !filter)
1477 		return (EINVAL);
1478 #endif
1479 	return (0);
1480 }
1481 
1482 #ifdef DDB
1483 /*
1484  * Dump details about an interrupt handler
1485  */
1486 static void
1487 db_dump_intrhand(struct intr_handler *ih)
1488 {
1489 	int comma;
1490 
1491 	db_printf("\t%-10s ", ih->ih_name);
1492 	switch (ih->ih_pri) {
1493 	case PI_REALTIME:
1494 		db_printf("CLK ");
1495 		break;
1496 	case PI_INTR:
1497 		db_printf("INTR");
1498 		break;
1499 	default:
1500 		if (ih->ih_pri >= PI_SOFT)
1501 			db_printf("SWI ");
1502 		else
1503 			db_printf("%4u", ih->ih_pri);
1504 		break;
1505 	}
1506 	db_printf(" ");
1507 	if (ih->ih_filter != NULL) {
1508 		db_printf("[F]");
1509 		db_printsym((uintptr_t)ih->ih_filter, DB_STGY_PROC);
1510 	}
1511 	if (ih->ih_handler != NULL) {
1512 		if (ih->ih_filter != NULL)
1513 			db_printf(",");
1514 		db_printf("[H]");
1515 		db_printsym((uintptr_t)ih->ih_handler, DB_STGY_PROC);
1516 	}
1517 	db_printf("(%p)", ih->ih_argument);
1518 	if (ih->ih_need ||
1519 	    (ih->ih_flags & (IH_EXCLUSIVE | IH_ENTROPY | IH_DEAD |
1520 	    IH_MPSAFE)) != 0) {
1521 		db_printf(" {");
1522 		comma = 0;
1523 		if (ih->ih_flags & IH_EXCLUSIVE) {
1524 			if (comma)
1525 				db_printf(", ");
1526 			db_printf("EXCL");
1527 			comma = 1;
1528 		}
1529 		if (ih->ih_flags & IH_ENTROPY) {
1530 			if (comma)
1531 				db_printf(", ");
1532 			db_printf("ENTROPY");
1533 			comma = 1;
1534 		}
1535 		if (ih->ih_flags & IH_DEAD) {
1536 			if (comma)
1537 				db_printf(", ");
1538 			db_printf("DEAD");
1539 			comma = 1;
1540 		}
1541 		if (ih->ih_flags & IH_MPSAFE) {
1542 			if (comma)
1543 				db_printf(", ");
1544 			db_printf("MPSAFE");
1545 			comma = 1;
1546 		}
1547 		if (ih->ih_need) {
1548 			if (comma)
1549 				db_printf(", ");
1550 			db_printf("NEED");
1551 		}
1552 		db_printf("}");
1553 	}
1554 	db_printf("\n");
1555 }
1556 
1557 /*
1558  * Dump details about a event.
1559  */
1560 void
1561 db_dump_intr_event(struct intr_event *ie, int handlers)
1562 {
1563 	struct intr_handler *ih;
1564 	struct intr_thread *it;
1565 	int comma;
1566 
1567 	db_printf("%s ", ie->ie_fullname);
1568 	it = ie->ie_thread;
1569 	if (it != NULL)
1570 		db_printf("(pid %d)", it->it_thread->td_proc->p_pid);
1571 	else
1572 		db_printf("(no thread)");
1573 	if ((ie->ie_flags & (IE_SOFT | IE_ADDING_THREAD)) != 0 ||
1574 	    (it != NULL && it->it_need)) {
1575 		db_printf(" {");
1576 		comma = 0;
1577 		if (ie->ie_flags & IE_SOFT) {
1578 			db_printf("SOFT");
1579 			comma = 1;
1580 		}
1581 		if (ie->ie_flags & IE_ADDING_THREAD) {
1582 			if (comma)
1583 				db_printf(", ");
1584 			db_printf("ADDING_THREAD");
1585 			comma = 1;
1586 		}
1587 		if (it != NULL && it->it_need) {
1588 			if (comma)
1589 				db_printf(", ");
1590 			db_printf("NEED");
1591 		}
1592 		db_printf("}");
1593 	}
1594 	db_printf("\n");
1595 
1596 	if (handlers)
1597 		CK_SLIST_FOREACH(ih, &ie->ie_handlers, ih_next)
1598 		    db_dump_intrhand(ih);
1599 }
1600 
1601 /*
1602  * Dump data about interrupt handlers
1603  */
1604 DB_SHOW_COMMAND_FLAGS(intr, db_show_intr, DB_CMD_MEMSAFE)
1605 {
1606 	struct intr_event *ie;
1607 	int all, verbose;
1608 
1609 	verbose = strchr(modif, 'v') != NULL;
1610 	all = strchr(modif, 'a') != NULL;
1611 	TAILQ_FOREACH(ie, &event_list, ie_list) {
1612 		if (!all && CK_SLIST_EMPTY(&ie->ie_handlers))
1613 			continue;
1614 		db_dump_intr_event(ie, verbose);
1615 		if (db_pager_quit)
1616 			break;
1617 	}
1618 }
1619 #endif /* DDB */
1620 
1621 /*
1622  * Start standard software interrupt threads
1623  */
1624 static void
1625 start_softintr(void *dummy)
1626 {
1627 
1628 	if (swi_add(&clk_intr_event, "clk", NULL, NULL, SWI_CLOCK,
1629 	    INTR_MPSAFE, NULL))
1630 		panic("died while creating clk swi ithread");
1631 }
1632 SYSINIT(start_softintr, SI_SUB_SOFTINTR, SI_ORDER_FIRST, start_softintr,
1633     NULL);
1634 
1635 /*
1636  * Sysctls used by systat and others: hw.intrnames and hw.intrcnt.
1637  * The data for this machine dependent, and the declarations are in machine
1638  * dependent code.  The layout of intrnames and intrcnt however is machine
1639  * independent.
1640  *
1641  * We do not know the length of intrcnt and intrnames at compile time, so
1642  * calculate things at run time.
1643  */
1644 static int
1645 sysctl_intrnames(SYSCTL_HANDLER_ARGS)
1646 {
1647 	return (sysctl_handle_opaque(oidp, intrnames, sintrnames, req));
1648 }
1649 
1650 SYSCTL_PROC(_hw, OID_AUTO, intrnames,
1651     CTLTYPE_OPAQUE | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
1652     sysctl_intrnames, "",
1653     "Interrupt Names");
1654 
1655 static int
1656 sysctl_intrcnt(SYSCTL_HANDLER_ARGS)
1657 {
1658 #ifdef SCTL_MASK32
1659 	uint32_t *intrcnt32;
1660 	unsigned i;
1661 	int error;
1662 
1663 	if (req->flags & SCTL_MASK32) {
1664 		if (!req->oldptr)
1665 			return (sysctl_handle_opaque(oidp, NULL, sintrcnt / 2, req));
1666 		intrcnt32 = malloc(sintrcnt / 2, M_TEMP, M_NOWAIT);
1667 		if (intrcnt32 == NULL)
1668 			return (ENOMEM);
1669 		for (i = 0; i < sintrcnt / sizeof (u_long); i++)
1670 			intrcnt32[i] = intrcnt[i];
1671 		error = sysctl_handle_opaque(oidp, intrcnt32, sintrcnt / 2, req);
1672 		free(intrcnt32, M_TEMP);
1673 		return (error);
1674 	}
1675 #endif
1676 	return (sysctl_handle_opaque(oidp, intrcnt, sintrcnt, req));
1677 }
1678 
1679 SYSCTL_PROC(_hw, OID_AUTO, intrcnt,
1680     CTLTYPE_OPAQUE | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
1681     sysctl_intrcnt, "",
1682     "Interrupt Counts");
1683 
1684 #ifdef DDB
1685 /*
1686  * DDB command to dump the interrupt statistics.
1687  */
1688 DB_SHOW_COMMAND_FLAGS(intrcnt, db_show_intrcnt, DB_CMD_MEMSAFE)
1689 {
1690 	u_long *i;
1691 	char *cp;
1692 	u_int j;
1693 
1694 	cp = intrnames;
1695 	j = 0;
1696 	for (i = intrcnt; j < (sintrcnt / sizeof(u_long)) && !db_pager_quit;
1697 	    i++, j++) {
1698 		if (*cp == '\0')
1699 			break;
1700 		if (*i != 0)
1701 			db_printf("%s\t%lu\n", cp, *i);
1702 		cp += strlen(cp) + 1;
1703 	}
1704 }
1705 #endif
1706