xref: /netbsd-src/sys/kern/kern_softint.c (revision 8b0f9554ff8762542c4defc4f70e1eb76fb508fa)
1 /*	$NetBSD: kern_softint.c,v 1.8 2007/12/11 19:07:28 ad Exp $	*/
2 
3 /*-
4  * Copyright (c) 2007 The NetBSD Foundation, Inc.
5  * All rights reserved.
6  *
7  * This code is derived from software contributed to The NetBSD Foundation
8  * by Andrew Doran.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  * 3. All advertising materials mentioning features or use of this software
19  *    must display the following acknowledgement:
20  *	This product includes software developed by the NetBSD
21  *	Foundation, Inc. and its contributors.
22  * 4. Neither the name of The NetBSD Foundation nor the names of its
23  *    contributors may be used to endorse or promote products derived
24  *    from this software without specific prior written permission.
25  *
26  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
27  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
28  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
29  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
30  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
31  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
32  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
33  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
34  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
35  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
36  * POSSIBILITY OF SUCH DAMAGE.
37  */
38 
39 /*
40  * Generic software interrupt framework.
41  *
42  * Overview
43  *
44  *	The soft interrupt framework provides a mechanism to schedule a
45  *	low priority callback that runs with thread context.  It allows
46  *	for dynamic registration of software interrupts, and for fair
47  *	queueing and prioritization of those interrupts.  The callbacks
48  *	can be scheduled to run from nearly any point in the kernel: by
49  *	code running with thread context, by code running from a
50  *	hardware interrupt handler, and at any interrupt priority
51  *	level.
52  *
53  * Priority levels
54  *
55  *	Since soft interrupt dispatch can be tied to the underlying
56  *	architecture's interrupt dispatch code, it can be limited
57  *	both by the capabilities of the hardware and the capabilities
58  *	of the interrupt dispatch code itself.  The number of priority
59  *	levels is restricted to four.  In order of priority (lowest to
60  *	highest) the levels are: clock, bio, net, serial.
61  *
62  *	The names are symbolic and in isolation do not have any direct
63  *	connection with a particular kind of device activity: they are
64  *	only meant as a guide.
65  *
66  *	The four priority levels map directly to scheduler priority
67  *	levels, and where the architecture implements 'fast' software
68  *	interrupts, they also map onto interrupt priorities.  The
69  *	interrupt priorities are intended to be hidden from machine
70  *	independent code, which should use thread-safe mechanisms to
71  *	synchronize with software interrupts (for example: mutexes).
72  *
73  * Capabilities
74  *
75  *	Software interrupts run with limited machine context.  In
76  *	particular, they do not posess any address space context.  They
77  *	should not try to operate on user space addresses, or to use
78  *	virtual memory facilities other than those noted as interrupt
79  *	safe.
80  *
81  *	Unlike hardware interrupts, software interrupts do have thread
82  *	context.  They may block on synchronization objects, sleep, and
83  *	resume execution at a later time.
84  *
85  *	Since software interrupts are a limited resource and run with
86  *	higher priority than most other LWPs in the system, all
87  *	block-and-resume activity by a software interrupt must be kept
88  *	short to allow futher processing at that level to continue.  By
89  *	extension, code running with process context must take care to
90  *	ensure that any lock that may be taken from a software interrupt
91  *	can not be held for more than a short period of time.
92  *
93  *	The kernel does not allow software interrupts to use facilities
94  *	or perform actions that may block for a significant amount of
95  *	time.  This means that it's not valid for a software interrupt
96  *	to: sleep on condition variables, use the lockmgr() facility,
97  *	or wait for resources to become available (for example,
98  *	memory).
99  *
100  * Per-CPU operation
101  *
102  *	If a soft interrupt is triggered on a CPU, it can only be
103  *	dispatched on the same CPU.  Each LWP dedicated to handling a
104  *	soft interrupt is bound to its home CPU, so if the LWP blocks
105  *	and needs to run again, it can only run there.  Nearly all data
106  *	structures used to manage software interrupts are per-CPU.
107  *
108  *	The per-CPU requirement is intended to reduce "ping-pong" of
109  *	cache lines between CPUs: lines occupied by data structures
110  *	used to manage the soft interrupts, and lines occupied by data
111  *	items being passed down to the soft interrupt.  As a positive
112  *	side effect, this also means that the soft interrupt dispatch
113  *	code does not need to to use spinlocks to synchronize.
114  *
115  * Generic implementation
116  *
117  *	A generic, low performance implementation is provided that
118  *	works across all architectures, with no machine-dependent
119  *	modifications needed.  This implementation uses the scheduler,
120  *	and so has a number of restrictions:
121  *
122  *	1) The software interrupts are not currently preemptive, so
123  *	must wait for the currently executing LWP to yield the CPU.
124  *	This can introduce latency.
125  *
126  *	2) An expensive context switch is required for a software
127  *	interrupt to be handled.
128  *
129  * 'Fast' software interrupts
130  *
131  *	If an architectures defines __HAVE_FAST_SOFTINTS, it implements
132  *	the fast mechanism.  Threads running either in the kernel or in
133  *	userspace will be interrupted, but will not be preempted.  When
134  *	the soft interrupt completes execution, the interrupted LWP
135  *	is resumed.  Interrupt dispatch code must provide the minimum
136  *	level of context necessary for the soft interrupt to block and
137  *	be resumed at a later time.  The machine-dependent dispatch
138  *	path looks something like the following:
139  *
140  *	softintr()
141  *	{
142  *		go to IPL_HIGH if necessary for switch;
143  *		save any necessary registers in a format that can be
144  *		    restored by cpu_switchto if the softint blocks;
145  *		arrange for cpu_switchto() to restore into the
146  *		    trampoline function;
147  *		identify LWP to handle this interrupt;
148  *		switch to the LWP's stack;
149  *		switch register stacks, if necessary;
150  *		assign new value of curlwp;
151  *		call MI softint_dispatch, passing old curlwp and IPL
152  *		    to execute interrupt at;
153  *		switch back to old stack;
154  *		switch back to old register stack, if necessary;
155  *		restore curlwp;
156  *		return to interrupted LWP;
157  *	}
158  *
159  *	If the soft interrupt blocks, a trampoline function is returned
160  *	to in the context of the interrupted LWP, as arranged for by
161  *	softint():
162  *
163  *	softint_ret()
164  *	{
165  *		unlock soft interrupt LWP;
166  *		resume interrupt processing, likely returning to
167  *		    interrupted LWP or dispatching another, different
168  *		    interrupt;
169  *	}
170  *
171  *	Once the soft interrupt has fired (and even if it has blocked),
172  *	no further soft interrupts at that level will be triggered by
173  *	MI code until the soft interrupt handler has ceased execution.
174  *	If a soft interrupt handler blocks and is resumed, it resumes
175  *	execution as a normal LWP (kthread) and gains VM context.  Only
176  *	when it has completed and is ready to fire again will it
177  *	interrupt other threads.
178  *
179  * Future directions
180  *
181  *	Provide a cheap way to direct software interrupts to remote
182  *	CPUs.  Provide a way to enqueue work items into the handler
183  *	record,	removing additional spl calls (see subr_workqueue.c).
184  */
185 
186 #include <sys/cdefs.h>
187 __KERNEL_RCSID(0, "$NetBSD: kern_softint.c,v 1.8 2007/12/11 19:07:28 ad Exp $");
188 
189 #include <sys/param.h>
190 #include <sys/malloc.h>
191 #include <sys/proc.h>
192 #include <sys/intr.h>
193 #include <sys/mutex.h>
194 #include <sys/kthread.h>
195 #include <sys/evcnt.h>
196 #include <sys/cpu.h>
197 
198 #include <net/netisr.h>
199 
200 #include <uvm/uvm_extern.h>
201 
202 /* This could overlap with signal info in struct lwp. */
203 typedef struct softint {
204 	SIMPLEQ_HEAD(, softhand) si_q;
205 	struct lwp		*si_lwp;
206 	struct cpu_info		*si_cpu;
207 	uintptr_t		si_machdep;
208 	struct evcnt		si_evcnt;
209 	struct evcnt		si_evcnt_block;
210 	int			si_active;
211 	char			si_name[8];
212 	char			si_name_block[8+6];
213 } softint_t;
214 
215 typedef struct softhand {
216 	SIMPLEQ_ENTRY(softhand)	sh_q;
217 	void			(*sh_func)(void *);
218 	void			*sh_arg;
219 	softint_t		*sh_isr;
220 	u_int			sh_pending;
221 	u_int			sh_flags;
222 } softhand_t;
223 
224 typedef struct softcpu {
225 	struct cpu_info		*sc_cpu;
226 	softint_t		sc_int[SOFTINT_COUNT];
227 	softhand_t		sc_hand[1];
228 } softcpu_t;
229 
230 static void	softint_thread(void *);
231 
232 u_int		softint_bytes = 8192;
233 u_int		softint_timing;
234 static u_int	softint_max;
235 static kmutex_t	softint_lock;
236 static void	*softint_netisrs[32];
237 
238 /*
239  * softint_init_isr:
240  *
241  *	Initialize a single interrupt level for a single CPU.
242  */
243 static void
244 softint_init_isr(softcpu_t *sc, const char *desc, pri_t pri, u_int level)
245 {
246 	struct cpu_info *ci;
247 	softint_t *si;
248 	int error;
249 
250 	si = &sc->sc_int[level];
251 	ci = sc->sc_cpu;
252 	si->si_cpu = ci;
253 
254 	SIMPLEQ_INIT(&si->si_q);
255 
256 	error = kthread_create(pri, KTHREAD_MPSAFE | KTHREAD_INTR |
257 	    KTHREAD_IDLE, ci, softint_thread, si, &si->si_lwp,
258 	    "soft%s/%d", desc, (int)ci->ci_cpuid);
259 	if (error != 0)
260 		panic("softint_init_isr: error %d", error);
261 
262 	snprintf(si->si_name, sizeof(si->si_name), "%s/%d", desc,
263 	    (int)ci->ci_cpuid);
264 	evcnt_attach_dynamic(&si->si_evcnt, EVCNT_TYPE_INTR, NULL,
265 	   "softint", si->si_name);
266 	snprintf(si->si_name_block, sizeof(si->si_name_block), "%s block/%d",
267 	    desc, (int)ci->ci_cpuid);
268 	evcnt_attach_dynamic(&si->si_evcnt_block, EVCNT_TYPE_INTR, NULL,
269 	   "softint", si->si_name_block);
270 
271 	si->si_lwp->l_private = si;
272 	softint_init_md(si->si_lwp, level, &si->si_machdep);
273 }
274 /*
275  * softint_init:
276  *
277  *	Initialize per-CPU data structures.  Called from mi_cpu_attach().
278  */
279 void
280 softint_init(struct cpu_info *ci)
281 {
282 	static struct cpu_info *first;
283 	softcpu_t *sc, *scfirst;
284 	softhand_t *sh, *shmax;
285 
286 	if (first == NULL) {
287 		/* Boot CPU. */
288 		first = ci;
289 		mutex_init(&softint_lock, MUTEX_DEFAULT, IPL_NONE);
290 		softint_bytes = round_page(softint_bytes);
291 		softint_max = (softint_bytes - sizeof(softcpu_t)) /
292 		    sizeof(softhand_t);
293 	}
294 
295 	sc = (softcpu_t *)uvm_km_alloc(kernel_map, softint_bytes, 0,
296 	    UVM_KMF_WIRED | UVM_KMF_ZERO);
297 	if (sc == NULL)
298 		panic("softint_init_cpu: cannot allocate memory");
299 
300 	ci->ci_data.cpu_softcpu = sc;
301 	ci->ci_data.cpu_softints = 0;
302 	sc->sc_cpu = ci;
303 
304 	softint_init_isr(sc, "net", PRI_SOFTNET, SOFTINT_NET);
305 	softint_init_isr(sc, "bio", PRI_SOFTBIO, SOFTINT_BIO);
306 	softint_init_isr(sc, "clk", PRI_SOFTCLOCK, SOFTINT_CLOCK);
307 	softint_init_isr(sc, "ser", PRI_SOFTSERIAL, SOFTINT_SERIAL);
308 
309 	if (first != ci) {
310 		mutex_enter(&softint_lock);
311 		scfirst = first->ci_data.cpu_softcpu;
312 		sh = sc->sc_hand;
313 		memcpy(sh, scfirst->sc_hand, sizeof(*sh) * softint_max);
314 		/* Update pointers for this CPU. */
315 		for (shmax = sh + softint_max; sh < shmax; sh++) {
316 			if (sh->sh_func == NULL)
317 				continue;
318 			sh->sh_isr =
319 			    &sc->sc_int[sh->sh_flags & SOFTINT_LVLMASK];
320 		}
321 		mutex_exit(&softint_lock);
322 	} else {
323 		/*
324 		 * Establish handlers for legacy net interrupts.
325 		 * XXX Needs to go away.
326 		 */
327 #define DONETISR(n, f)							\
328     softint_netisrs[(n)] = 						\
329         softint_establish(SOFTINT_NET, (void (*)(void *))(f), NULL)
330 #include <net/netisr_dispatch.h>
331 	}
332 }
333 
334 /*
335  * softint_establish:
336  *
337  *	Register a software interrupt handler.
338  */
339 void *
340 softint_establish(u_int flags, void (*func)(void *), void *arg)
341 {
342 	CPU_INFO_ITERATOR cii;
343 	struct cpu_info *ci;
344 	softcpu_t *sc;
345 	softhand_t *sh;
346 	u_int level, index;
347 
348 	level = (flags & SOFTINT_LVLMASK);
349 	KASSERT(level < SOFTINT_COUNT);
350 
351 	mutex_enter(&softint_lock);
352 
353 	/* Find a free slot. */
354 	sc = curcpu()->ci_data.cpu_softcpu;
355 	for (index = 1; index < softint_max; index++)
356 		if (sc->sc_hand[index].sh_func == NULL)
357 			break;
358 	if (index == softint_max) {
359 		mutex_exit(&softint_lock);
360 		printf("WARNING: softint_establish: table full, "
361 		    "increase softint_bytes\n");
362 		return NULL;
363 	}
364 
365 	/* Set up the handler on each CPU. */
366 	if (ncpu < 2) {
367 		/* XXX hack for machines with no CPU_INFO_FOREACH() early on */
368 		sc = curcpu()->ci_data.cpu_softcpu;
369 		sh = &sc->sc_hand[index];
370 		sh->sh_isr = &sc->sc_int[level];
371 		sh->sh_func = func;
372 		sh->sh_arg = arg;
373 		sh->sh_flags = flags;
374 		sh->sh_pending = 0;
375 	} else for (CPU_INFO_FOREACH(cii, ci)) {
376 		sc = ci->ci_data.cpu_softcpu;
377 		sh = &sc->sc_hand[index];
378 		sh->sh_isr = &sc->sc_int[level];
379 		sh->sh_func = func;
380 		sh->sh_arg = arg;
381 		sh->sh_flags = flags;
382 		sh->sh_pending = 0;
383 	}
384 
385 	mutex_exit(&softint_lock);
386 
387 	return (void *)((uint8_t *)&sc->sc_hand[index] - (uint8_t *)sc);
388 }
389 
390 /*
391  * softint_disestablish:
392  *
393  *	Unregister a software interrupt handler.
394  */
395 void
396 softint_disestablish(void *arg)
397 {
398 	CPU_INFO_ITERATOR cii;
399 	struct cpu_info *ci;
400 	softcpu_t *sc;
401 	softhand_t *sh;
402 	uintptr_t offset;
403 
404 	offset = (uintptr_t)arg;
405 	KASSERT(offset != 0 && offset < softint_bytes);
406 
407 	mutex_enter(&softint_lock);
408 
409 	/* Clear the handler on each CPU. */
410 	for (CPU_INFO_FOREACH(cii, ci)) {
411 		sc = ci->ci_data.cpu_softcpu;
412 		sh = (softhand_t *)((uint8_t *)sc + offset);
413 		KASSERT(sh->sh_func != NULL);
414 		KASSERT(sh->sh_pending == 0);
415 		sh->sh_func = NULL;
416 	}
417 
418 	mutex_exit(&softint_lock);
419 }
420 
421 /*
422  * softint_schedule:
423  *
424  *	Trigger a software interrupt.  Must be called from a hardware
425  *	interrupt handler, or with preemption disabled (since we are
426  *	using the value of curcpu()).
427  */
428 void
429 softint_schedule(void *arg)
430 {
431 	softhand_t *sh;
432 	softint_t *si;
433 	uintptr_t offset;
434 	int s;
435 
436 	/* Find the handler record for this CPU. */
437 	offset = (uintptr_t)arg;
438 	KASSERT(offset != 0 && offset < softint_bytes);
439 	sh = (softhand_t *)((uint8_t *)curcpu()->ci_data.cpu_softcpu + offset);
440 
441 	/* If it's already pending there's nothing to do. */
442 	if (sh->sh_pending)
443 		return;
444 
445 	/*
446 	 * Enqueue the handler into the LWP's pending list.
447 	 * If the LWP is completely idle, then make it run.
448 	 */
449 	s = splhigh();
450 	if (!sh->sh_pending) {
451 		si = sh->sh_isr;
452 		sh->sh_pending = 1;
453 		SIMPLEQ_INSERT_TAIL(&si->si_q, sh, sh_q);
454 		if (si->si_active == 0) {
455 			si->si_active = 1;
456 			softint_trigger(si->si_machdep);
457 		}
458 	}
459 	splx(s);
460 }
461 
462 /*
463  * softint_execute:
464  *
465  *	Invoke handlers for the specified soft interrupt.
466  *	Must be entered at splhigh.  Will drop the priority
467  *	to the level specified, but returns back at splhigh.
468  */
469 static inline void
470 softint_execute(softint_t *si, lwp_t *l, int s)
471 {
472 	softhand_t *sh;
473 	bool havelock;
474 
475 #ifdef __HAVE_FAST_SOFTINTS
476 	KASSERT(si->si_lwp == curlwp);
477 #else
478 	/* May be running in user context. */
479 #endif
480 	KASSERT(si->si_cpu == curcpu());
481 	KASSERT(si->si_lwp->l_wchan == NULL);
482 	KASSERT(si->si_active);
483 
484 	havelock = false;
485 
486 	/*
487 	 * Note: due to priority inheritance we may have interrupted a
488 	 * higher priority LWP.  Since the soft interrupt must be quick
489 	 * and is non-preemptable, we don't bother yielding.
490 	 */
491 
492 	while (!SIMPLEQ_EMPTY(&si->si_q)) {
493 		/*
494 		 * Pick the longest waiting handler to run.  We block
495 		 * interrupts but do not lock in order to do this, as
496 		 * we are protecting against the local CPU only.
497 		 */
498 		sh = SIMPLEQ_FIRST(&si->si_q);
499 		SIMPLEQ_REMOVE_HEAD(&si->si_q, sh_q);
500 		sh->sh_pending = 0;
501 		splx(s);
502 
503 		/* Run the handler. */
504 		if ((sh->sh_flags & SOFTINT_MPSAFE) == 0 && !havelock) {
505 			KERNEL_LOCK(1, l);
506 			havelock = true;
507 		}
508 		(*sh->sh_func)(sh->sh_arg);
509 
510 		(void)splhigh();
511 	}
512 
513 	if (havelock) {
514 		KERNEL_UNLOCK_ONE(l);
515 	}
516 
517 	/*
518 	 * Unlocked, but only for statistics.
519 	 * Should be per-CPU to prevent cache ping-pong.
520 	 */
521 	uvmexp.softs++;
522 
523 	si->si_evcnt.ev_count++;
524 	si->si_active = 0;
525 }
526 
527 /*
528  * softint_block:
529  *
530  *	Update statistics when the soft interrupt blocks.
531  */
532 void
533 softint_block(lwp_t *l)
534 {
535 	softint_t *si = l->l_private;
536 
537 	KASSERT((l->l_pflag & LP_INTR) != 0);
538 	si->si_evcnt_block.ev_count++;
539 }
540 
541 /*
542  * schednetisr:
543  *
544  *	Trigger a legacy network interrupt.  XXX Needs to go away.
545  */
546 void
547 schednetisr(int isr)
548 {
549 
550 	softint_schedule(softint_netisrs[isr]);
551 }
552 
553 #ifndef __HAVE_FAST_SOFTINTS
554 
555 /*
556  * softint_init_md:
557  *
558  *	Slow path: perform machine-dependent initialization.
559  */
560 void
561 softint_init_md(lwp_t *l, u_int level, uintptr_t *machdep)
562 {
563 	softint_t *si;
564 
565 	*machdep = (1 << level);
566 	si = l->l_private;
567 
568 	lwp_lock(l);
569 	lwp_unlock_to(l, l->l_cpu->ci_schedstate.spc_mutex);
570 	lwp_lock(l);
571 	/* Cheat and make the KASSERT in softint_thread() happy. */
572 	si->si_active = 1;
573 	l->l_stat = LSRUN;
574 	sched_enqueue(l, false);
575 	lwp_unlock(l);
576 }
577 
578 /*
579  * softint_trigger:
580  *
581  *	Slow path: cause a soft interrupt handler to begin executing.
582  *	Called at IPL_HIGH.
583  */
584 void
585 softint_trigger(uintptr_t machdep)
586 {
587 	struct cpu_info *ci;
588 	lwp_t *l;
589 
590 	l = curlwp;
591 	ci = l->l_cpu;
592 	ci->ci_data.cpu_softints |= machdep;
593 	if (l == ci->ci_data.cpu_idlelwp) {
594 		cpu_need_resched(ci, 0);
595 	} else {
596 		/* MI equivalent of aston() */
597 		cpu_signotify(l);
598 	}
599 }
600 
601 /*
602  * softint_thread:
603  *
604  *	Slow path: MI software interrupt dispatch.
605  */
606 void
607 softint_thread(void *cookie)
608 {
609 	softint_t *si;
610 	lwp_t *l;
611 	int s;
612 
613 	l = curlwp;
614 	si = l->l_private;
615 
616 	for (;;) {
617 		/*
618 		 * Clear pending status and run it.  We must drop the
619 		 * spl before mi_switch(), since IPL_HIGH may be higher
620 		 * than IPL_SCHED (and it is not safe to switch at a
621 		 * higher level).
622 		 */
623 		s = splhigh();
624 		l->l_cpu->ci_data.cpu_softints &= ~si->si_machdep;
625 		softint_execute(si, l, s);
626 		splx(s);
627 
628 		lwp_lock(l);
629 		l->l_stat = LSIDL;
630 		mi_switch(l);
631 	}
632 }
633 
634 /*
635  * softint_picklwp:
636  *
637  *	Slow path: called from mi_switch() to pick the highest priority
638  *	soft interrupt LWP that needs to run.
639  */
640 lwp_t *
641 softint_picklwp(void)
642 {
643 	struct cpu_info *ci;
644 	u_int mask;
645 	softint_t *si;
646 	lwp_t *l;
647 
648 	ci = curcpu();
649 	si = ((softcpu_t *)ci->ci_data.cpu_softcpu)->sc_int;
650 	mask = ci->ci_data.cpu_softints;
651 
652 	if ((mask & (1 << SOFTINT_SERIAL)) != 0) {
653 		l = si[SOFTINT_SERIAL].si_lwp;
654 	} else if ((mask & (1 << SOFTINT_NET)) != 0) {
655 		l = si[SOFTINT_NET].si_lwp;
656 	} else if ((mask & (1 << SOFTINT_BIO)) != 0) {
657 		l = si[SOFTINT_BIO].si_lwp;
658 	} else if ((mask & (1 << SOFTINT_CLOCK)) != 0) {
659 		l = si[SOFTINT_CLOCK].si_lwp;
660 	} else {
661 		panic("softint_picklwp");
662 	}
663 
664 	return l;
665 }
666 
667 /*
668  * softint_overlay:
669  *
670  *	Slow path: called from lwp_userret() to run a soft interrupt
671  *	within the context of a user thread.
672  */
673 void
674 softint_overlay(void)
675 {
676 	struct cpu_info *ci;
677 	u_int softints;
678 	softint_t *si;
679 	pri_t obase;
680 	lwp_t *l;
681 	int s;
682 
683 	l = curlwp;
684 	ci = l->l_cpu;
685 	si = ((softcpu_t *)ci->ci_data.cpu_softcpu)->sc_int;
686 
687 	KASSERT((l->l_pflag & LP_INTR) == 0);
688 
689 	/* Arrange to elevate priority if the LWP blocks. */
690 	obase = l->l_kpribase;
691 	l->l_kpribase = PRI_KERNEL_RT;
692 	l->l_pflag |= LP_INTR;
693 	s = splhigh();
694 	while ((softints = ci->ci_data.cpu_softints) != 0) {
695 		if ((softints & (1 << SOFTINT_SERIAL)) != 0) {
696 			ci->ci_data.cpu_softints &= ~(1 << SOFTINT_SERIAL);
697 			softint_execute(&si[SOFTINT_SERIAL], l, s);
698 			continue;
699 		}
700 		if ((softints & (1 << SOFTINT_NET)) != 0) {
701 			ci->ci_data.cpu_softints &= ~(1 << SOFTINT_NET);
702 			softint_execute(&si[SOFTINT_NET], l, s);
703 			continue;
704 		}
705 		if ((softints & (1 << SOFTINT_BIO)) != 0) {
706 			ci->ci_data.cpu_softints &= ~(1 << SOFTINT_BIO);
707 			softint_execute(&si[SOFTINT_BIO], l, s);
708 			continue;
709 		}
710 		if ((softints & (1 << SOFTINT_CLOCK)) != 0) {
711 			ci->ci_data.cpu_softints &= ~(1 << SOFTINT_CLOCK);
712 			softint_execute(&si[SOFTINT_CLOCK], l, s);
713 			continue;
714 		}
715 	}
716 	splx(s);
717 	l->l_pflag &= ~LP_INTR;
718 	l->l_kpribase = obase;
719 }
720 
721 #else	/*  !__HAVE_FAST_SOFTINTS */
722 
723 /*
724  * softint_thread:
725  *
726  *	Fast path: the LWP is switched to without restoring any state,
727  *	so we should not arrive here - there is a direct handoff between
728  *	the interrupt stub and softint_dispatch().
729  */
730 void
731 softint_thread(void *cookie)
732 {
733 
734 	panic("softint_thread");
735 }
736 
737 /*
738  * softint_dispatch:
739  *
740  *	Fast path: entry point from machine-dependent code.
741  */
742 void
743 softint_dispatch(lwp_t *pinned, int s)
744 {
745 	struct timeval now;
746 	softint_t *si;
747 	u_int timing;
748 	lwp_t *l;
749 
750 	l = curlwp;
751 	si = l->l_private;
752 
753 	/*
754 	 * Note the interrupted LWP, and mark the current LWP as running
755 	 * before proceeding.  Although this must as a rule be done with
756 	 * the LWP locked, at this point no external agents will want to
757 	 * modify the interrupt LWP's state.
758 	 */
759 	timing = (softint_timing ? LW_TIMEINTR : 0);
760 	l->l_switchto = pinned;
761 	l->l_stat = LSONPROC;
762 	l->l_flag |= (LW_RUNNING | timing);
763 
764 	/*
765 	 * Dispatch the interrupt.  If softints are being timed, charge
766 	 * for it.
767 	 */
768 	if (timing)
769 		microtime(&l->l_stime);
770 	softint_execute(si, l, s);
771 	if (timing) {
772 		microtime(&now);
773 		updatertime(l, &now);
774 		l->l_flag &= ~LW_TIMEINTR;
775 	}
776 
777 	/*
778 	 * If we blocked while handling the interrupt, the pinned LWP is
779 	 * gone so switch to the idle LWP.  It will select a new LWP to
780 	 * run.
781 	 *
782 	 * We must drop the priority level as switching at IPL_HIGH could
783 	 * deadlock the system.  We have already set si->si_active = 0,
784 	 * which means another interrupt at this level can be triggered.
785 	 * That's not be a problem: we are lowering to level 's' which will
786 	 * prevent softint_dispatch() from being reentered at level 's',
787 	 * until the priority is finally dropped to IPL_NONE on entry to
788 	 * the idle loop.
789 	 */
790 	l->l_stat = LSIDL;
791 	if (l->l_switchto == NULL) {
792 		splx(s);
793 		pmap_deactivate(l);
794 		lwp_exit_switchaway(l);
795 		/* NOTREACHED */
796 	}
797 	l->l_switchto = NULL;
798 	l->l_flag &= ~LW_RUNNING;
799 }
800 
801 #endif	/* !__HAVE_FAST_SOFTINTS */
802