xref: /netbsd-src/sys/kern/kern_softint.c (revision b1c86f5f087524e68db12794ee9c3e3da1ab17a0)
1 /*	$NetBSD: kern_softint.c,v 1.31 2010/01/09 19:02:17 rmind Exp $	*/
2 
3 /*-
4  * Copyright (c) 2007, 2008 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  *
19  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
20  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
21  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
23  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
24  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
25  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
26  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
27  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
28  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
29  * POSSIBILITY OF SUCH DAMAGE.
30  */
31 
32 /*
33  * Generic software interrupt framework.
34  *
35  * Overview
36  *
37  *	The soft interrupt framework provides a mechanism to schedule a
38  *	low priority callback that runs with thread context.  It allows
39  *	for dynamic registration of software interrupts, and for fair
40  *	queueing and prioritization of those interrupts.  The callbacks
41  *	can be scheduled to run from nearly any point in the kernel: by
42  *	code running with thread context, by code running from a
43  *	hardware interrupt handler, and at any interrupt priority
44  *	level.
45  *
46  * Priority levels
47  *
48  *	Since soft interrupt dispatch can be tied to the underlying
49  *	architecture's interrupt dispatch code, it can be limited
50  *	both by the capabilities of the hardware and the capabilities
51  *	of the interrupt dispatch code itself.  The number of priority
52  *	levels is restricted to four.  In order of priority (lowest to
53  *	highest) the levels are: clock, bio, net, serial.
54  *
55  *	The names are symbolic and in isolation do not have any direct
56  *	connection with a particular kind of device activity: they are
57  *	only meant as a guide.
58  *
59  *	The four priority levels map directly to scheduler priority
60  *	levels, and where the architecture implements 'fast' software
61  *	interrupts, they also map onto interrupt priorities.  The
62  *	interrupt priorities are intended to be hidden from machine
63  *	independent code, which should use thread-safe mechanisms to
64  *	synchronize with software interrupts (for example: mutexes).
65  *
66  * Capabilities
67  *
68  *	Software interrupts run with limited machine context.  In
69  *	particular, they do not posess any address space context.  They
70  *	should not try to operate on user space addresses, or to use
71  *	virtual memory facilities other than those noted as interrupt
72  *	safe.
73  *
74  *	Unlike hardware interrupts, software interrupts do have thread
75  *	context.  They may block on synchronization objects, sleep, and
76  *	resume execution at a later time.
77  *
78  *	Since software interrupts are a limited resource and run with
79  *	higher priority than most other LWPs in the system, all
80  *	block-and-resume activity by a software interrupt must be kept
81  *	short to allow futher processing at that level to continue.  By
82  *	extension, code running with process context must take care to
83  *	ensure that any lock that may be taken from a software interrupt
84  *	can not be held for more than a short period of time.
85  *
86  *	The kernel does not allow software interrupts to use facilities
87  *	or perform actions that may block for a significant amount of
88  *	time.  This means that it's not valid for a software interrupt
89  *	to sleep on condition variables	or wait for resources to become
90  *	available (for example,	memory).
91  *
92  * Per-CPU operation
93  *
94  *	If a soft interrupt is triggered on a CPU, it can only be
95  *	dispatched on the same CPU.  Each LWP dedicated to handling a
96  *	soft interrupt is bound to its home CPU, so if the LWP blocks
97  *	and needs to run again, it can only run there.  Nearly all data
98  *	structures used to manage software interrupts are per-CPU.
99  *
100  *	The per-CPU requirement is intended to reduce "ping-pong" of
101  *	cache lines between CPUs: lines occupied by data structures
102  *	used to manage the soft interrupts, and lines occupied by data
103  *	items being passed down to the soft interrupt.  As a positive
104  *	side effect, this also means that the soft interrupt dispatch
105  *	code does not need to to use spinlocks to synchronize.
106  *
107  * Generic implementation
108  *
109  *	A generic, low performance implementation is provided that
110  *	works across all architectures, with no machine-dependent
111  *	modifications needed.  This implementation uses the scheduler,
112  *	and so has a number of restrictions:
113  *
114  *	1) The software interrupts are not currently preemptive, so
115  *	must wait for the currently executing LWP to yield the CPU.
116  *	This can introduce latency.
117  *
118  *	2) An expensive context switch is required for a software
119  *	interrupt to be handled.
120  *
121  * 'Fast' software interrupts
122  *
123  *	If an architectures defines __HAVE_FAST_SOFTINTS, it implements
124  *	the fast mechanism.  Threads running either in the kernel or in
125  *	userspace will be interrupted, but will not be preempted.  When
126  *	the soft interrupt completes execution, the interrupted LWP
127  *	is resumed.  Interrupt dispatch code must provide the minimum
128  *	level of context necessary for the soft interrupt to block and
129  *	be resumed at a later time.  The machine-dependent dispatch
130  *	path looks something like the following:
131  *
132  *	softintr()
133  *	{
134  *		go to IPL_HIGH if necessary for switch;
135  *		save any necessary registers in a format that can be
136  *		    restored by cpu_switchto if the softint blocks;
137  *		arrange for cpu_switchto() to restore into the
138  *		    trampoline function;
139  *		identify LWP to handle this interrupt;
140  *		switch to the LWP's stack;
141  *		switch register stacks, if necessary;
142  *		assign new value of curlwp;
143  *		call MI softint_dispatch, passing old curlwp and IPL
144  *		    to execute interrupt at;
145  *		switch back to old stack;
146  *		switch back to old register stack, if necessary;
147  *		restore curlwp;
148  *		return to interrupted LWP;
149  *	}
150  *
151  *	If the soft interrupt blocks, a trampoline function is returned
152  *	to in the context of the interrupted LWP, as arranged for by
153  *	softint():
154  *
155  *	softint_ret()
156  *	{
157  *		unlock soft interrupt LWP;
158  *		resume interrupt processing, likely returning to
159  *		    interrupted LWP or dispatching another, different
160  *		    interrupt;
161  *	}
162  *
163  *	Once the soft interrupt has fired (and even if it has blocked),
164  *	no further soft interrupts at that level will be triggered by
165  *	MI code until the soft interrupt handler has ceased execution.
166  *	If a soft interrupt handler blocks and is resumed, it resumes
167  *	execution as a normal LWP (kthread) and gains VM context.  Only
168  *	when it has completed and is ready to fire again will it
169  *	interrupt other threads.
170  *
171  * Future directions
172  *
173  *	Provide a cheap way to direct software interrupts to remote
174  *	CPUs.  Provide a way to enqueue work items into the handler
175  *	record,	removing additional spl calls (see subr_workqueue.c).
176  */
177 
178 #include <sys/cdefs.h>
179 __KERNEL_RCSID(0, "$NetBSD: kern_softint.c,v 1.31 2010/01/09 19:02:17 rmind Exp $");
180 
181 #include <sys/param.h>
182 #include <sys/malloc.h>
183 #include <sys/proc.h>
184 #include <sys/intr.h>
185 #include <sys/mutex.h>
186 #include <sys/kthread.h>
187 #include <sys/evcnt.h>
188 #include <sys/cpu.h>
189 #include <sys/xcall.h>
190 
191 #include <net/netisr.h>
192 
193 #include <uvm/uvm_extern.h>
194 
195 /* This could overlap with signal info in struct lwp. */
196 typedef struct softint {
197 	SIMPLEQ_HEAD(, softhand) si_q;
198 	struct lwp		*si_lwp;
199 	struct cpu_info		*si_cpu;
200 	uintptr_t		si_machdep;
201 	struct evcnt		si_evcnt;
202 	struct evcnt		si_evcnt_block;
203 	int			si_active;
204 	char			si_name[8];
205 	char			si_name_block[8+6];
206 } softint_t;
207 
208 typedef struct softhand {
209 	SIMPLEQ_ENTRY(softhand)	sh_q;
210 	void			(*sh_func)(void *);
211 	void			*sh_arg;
212 	softint_t		*sh_isr;
213 	u_int			sh_flags;
214 } softhand_t;
215 
216 typedef struct softcpu {
217 	struct cpu_info		*sc_cpu;
218 	softint_t		sc_int[SOFTINT_COUNT];
219 	softhand_t		sc_hand[1];
220 } softcpu_t;
221 
222 static void	softint_thread(void *);
223 
224 u_int		softint_bytes = 8192;
225 u_int		softint_timing;
226 static u_int	softint_max;
227 static kmutex_t	softint_lock;
228 static void	*softint_netisrs[NETISR_MAX];
229 
230 /*
231  * softint_init_isr:
232  *
233  *	Initialize a single interrupt level for a single CPU.
234  */
235 static void
236 softint_init_isr(softcpu_t *sc, const char *desc, pri_t pri, u_int level)
237 {
238 	struct cpu_info *ci;
239 	softint_t *si;
240 	int error;
241 
242 	si = &sc->sc_int[level];
243 	ci = sc->sc_cpu;
244 	si->si_cpu = ci;
245 
246 	SIMPLEQ_INIT(&si->si_q);
247 
248 	error = kthread_create(pri, KTHREAD_MPSAFE | KTHREAD_INTR |
249 	    KTHREAD_IDLE, ci, softint_thread, si, &si->si_lwp,
250 	    "soft%s/%u", desc, ci->ci_index);
251 	if (error != 0)
252 		panic("softint_init_isr: error %d", error);
253 
254 	snprintf(si->si_name, sizeof(si->si_name), "%s/%u", desc,
255 	    ci->ci_index);
256 	evcnt_attach_dynamic(&si->si_evcnt, EVCNT_TYPE_MISC, NULL,
257 	   "softint", si->si_name);
258 	snprintf(si->si_name_block, sizeof(si->si_name_block), "%s block/%u",
259 	    desc, ci->ci_index);
260 	evcnt_attach_dynamic(&si->si_evcnt_block, EVCNT_TYPE_MISC, NULL,
261 	   "softint", si->si_name_block);
262 
263 	si->si_lwp->l_private = si;
264 	softint_init_md(si->si_lwp, level, &si->si_machdep);
265 }
266 /*
267  * softint_init:
268  *
269  *	Initialize per-CPU data structures.  Called from mi_cpu_attach().
270  */
271 void
272 softint_init(struct cpu_info *ci)
273 {
274 	static struct cpu_info *first;
275 	softcpu_t *sc, *scfirst;
276 	softhand_t *sh, *shmax;
277 
278 	if (first == NULL) {
279 		/* Boot CPU. */
280 		first = ci;
281 		mutex_init(&softint_lock, MUTEX_DEFAULT, IPL_NONE);
282 		softint_bytes = round_page(softint_bytes);
283 		softint_max = (softint_bytes - sizeof(softcpu_t)) /
284 		    sizeof(softhand_t);
285 	}
286 
287 	sc = (softcpu_t *)uvm_km_alloc(kernel_map, softint_bytes, 0,
288 	    UVM_KMF_WIRED | UVM_KMF_ZERO);
289 	if (sc == NULL)
290 		panic("softint_init_cpu: cannot allocate memory");
291 
292 	ci->ci_data.cpu_softcpu = sc;
293 	ci->ci_data.cpu_softints = 0;
294 	sc->sc_cpu = ci;
295 
296 	softint_init_isr(sc, "net", PRI_SOFTNET, SOFTINT_NET);
297 	softint_init_isr(sc, "bio", PRI_SOFTBIO, SOFTINT_BIO);
298 	softint_init_isr(sc, "clk", PRI_SOFTCLOCK, SOFTINT_CLOCK);
299 	softint_init_isr(sc, "ser", PRI_SOFTSERIAL, SOFTINT_SERIAL);
300 
301 	if (first != ci) {
302 		mutex_enter(&softint_lock);
303 		scfirst = first->ci_data.cpu_softcpu;
304 		sh = sc->sc_hand;
305 		memcpy(sh, scfirst->sc_hand, sizeof(*sh) * softint_max);
306 		/* Update pointers for this CPU. */
307 		for (shmax = sh + softint_max; sh < shmax; sh++) {
308 			if (sh->sh_func == NULL)
309 				continue;
310 			sh->sh_isr =
311 			    &sc->sc_int[sh->sh_flags & SOFTINT_LVLMASK];
312 		}
313 		mutex_exit(&softint_lock);
314 	} else {
315 		/*
316 		 * Establish handlers for legacy net interrupts.
317 		 * XXX Needs to go away.
318 		 */
319 #define DONETISR(n, f)							\
320     softint_netisrs[(n)] = softint_establish(SOFTINT_NET|SOFTINT_MPSAFE,\
321         (void (*)(void *))(f), NULL)
322 #include <net/netisr_dispatch.h>
323 	}
324 }
325 
326 /*
327  * softint_establish:
328  *
329  *	Register a software interrupt handler.
330  */
331 void *
332 softint_establish(u_int flags, void (*func)(void *), void *arg)
333 {
334 	CPU_INFO_ITERATOR cii;
335 	struct cpu_info *ci;
336 	softcpu_t *sc;
337 	softhand_t *sh;
338 	u_int level, index;
339 
340 	level = (flags & SOFTINT_LVLMASK);
341 	KASSERT(level < SOFTINT_COUNT);
342 	KASSERT((flags & SOFTINT_IMPMASK) == 0);
343 
344 	mutex_enter(&softint_lock);
345 
346 	/* Find a free slot. */
347 	sc = curcpu()->ci_data.cpu_softcpu;
348 	for (index = 1; index < softint_max; index++)
349 		if (sc->sc_hand[index].sh_func == NULL)
350 			break;
351 	if (index == softint_max) {
352 		mutex_exit(&softint_lock);
353 		printf("WARNING: softint_establish: table full, "
354 		    "increase softint_bytes\n");
355 		return NULL;
356 	}
357 
358 	/* Set up the handler on each CPU. */
359 	if (ncpu < 2) {
360 		/* XXX hack for machines with no CPU_INFO_FOREACH() early on */
361 		sc = curcpu()->ci_data.cpu_softcpu;
362 		sh = &sc->sc_hand[index];
363 		sh->sh_isr = &sc->sc_int[level];
364 		sh->sh_func = func;
365 		sh->sh_arg = arg;
366 		sh->sh_flags = flags;
367 	} else for (CPU_INFO_FOREACH(cii, ci)) {
368 		sc = ci->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 	}
375 
376 	mutex_exit(&softint_lock);
377 
378 	return (void *)((uint8_t *)&sc->sc_hand[index] - (uint8_t *)sc);
379 }
380 
381 /*
382  * softint_disestablish:
383  *
384  *	Unregister a software interrupt handler.  The soft interrupt could
385  *	still be active at this point, but the caller commits not to try
386  *	and trigger it again once this call is made.  The caller must not
387  *	hold any locks that could be taken from soft interrupt context,
388  *	because we will wait for the softint to complete if it's still
389  *	running.
390  */
391 void
392 softint_disestablish(void *arg)
393 {
394 	CPU_INFO_ITERATOR cii;
395 	struct cpu_info *ci;
396 	softcpu_t *sc;
397 	softhand_t *sh;
398 	uintptr_t offset;
399 	uint64_t where;
400 	u_int flags;
401 
402 	offset = (uintptr_t)arg;
403 	KASSERT(offset != 0 && offset < softint_bytes);
404 
405 	/*
406 	 * Run a cross call so we see up to date values of sh_flags from
407 	 * all CPUs.  Once softint_disestablish() is called, the caller
408 	 * commits to not trigger the interrupt and set SOFTINT_ACTIVE on
409 	 * it again.  So, we are only looking for handler records with
410 	 * SOFTINT_ACTIVE already set.
411 	 */
412 	where = xc_broadcast(0, (xcfunc_t)nullop, NULL, NULL);
413 	xc_wait(where);
414 
415 	for (;;) {
416 		/* Collect flag values from each CPU. */
417 		flags = 0;
418 		for (CPU_INFO_FOREACH(cii, ci)) {
419 			sc = ci->ci_data.cpu_softcpu;
420 			sh = (softhand_t *)((uint8_t *)sc + offset);
421 			KASSERT(sh->sh_func != NULL);
422 			flags |= sh->sh_flags;
423 		}
424 		/* Inactive on all CPUs? */
425 		if ((flags & SOFTINT_ACTIVE) == 0) {
426 			break;
427 		}
428 		/* Oops, still active.  Wait for it to clear. */
429 		(void)kpause("softdis", false, 1, NULL);
430 	}
431 
432 	/* Clear the handler on each CPU. */
433 	mutex_enter(&softint_lock);
434 	for (CPU_INFO_FOREACH(cii, ci)) {
435 		sc = ci->ci_data.cpu_softcpu;
436 		sh = (softhand_t *)((uint8_t *)sc + offset);
437 		KASSERT(sh->sh_func != NULL);
438 		sh->sh_func = NULL;
439 	}
440 	mutex_exit(&softint_lock);
441 }
442 
443 /*
444  * softint_schedule:
445  *
446  *	Trigger a software interrupt.  Must be called from a hardware
447  *	interrupt handler, or with preemption disabled (since we are
448  *	using the value of curcpu()).
449  */
450 void
451 softint_schedule(void *arg)
452 {
453 	softhand_t *sh;
454 	softint_t *si;
455 	uintptr_t offset;
456 	int s;
457 
458 	KASSERT(kpreempt_disabled());
459 
460 	/* Find the handler record for this CPU. */
461 	offset = (uintptr_t)arg;
462 	KASSERT(offset != 0 && offset < softint_bytes);
463 	sh = (softhand_t *)((uint8_t *)curcpu()->ci_data.cpu_softcpu + offset);
464 
465 	/* If it's already pending there's nothing to do. */
466 	if ((sh->sh_flags & SOFTINT_PENDING) != 0)
467 		return;
468 
469 	/*
470 	 * Enqueue the handler into the LWP's pending list.
471 	 * If the LWP is completely idle, then make it run.
472 	 */
473 	s = splhigh();
474 	if ((sh->sh_flags & SOFTINT_PENDING) == 0) {
475 		si = sh->sh_isr;
476 		sh->sh_flags |= SOFTINT_PENDING;
477 		SIMPLEQ_INSERT_TAIL(&si->si_q, sh, sh_q);
478 		if (si->si_active == 0) {
479 			si->si_active = 1;
480 			softint_trigger(si->si_machdep);
481 		}
482 	}
483 	splx(s);
484 }
485 
486 /*
487  * softint_execute:
488  *
489  *	Invoke handlers for the specified soft interrupt.
490  *	Must be entered at splhigh.  Will drop the priority
491  *	to the level specified, but returns back at splhigh.
492  */
493 static inline void
494 softint_execute(softint_t *si, lwp_t *l, int s)
495 {
496 	softhand_t *sh;
497 	bool havelock;
498 
499 #ifdef __HAVE_FAST_SOFTINTS
500 	KASSERT(si->si_lwp == curlwp);
501 #else
502 	/* May be running in user context. */
503 #endif
504 	KASSERT(si->si_cpu == curcpu());
505 	KASSERT(si->si_lwp->l_wchan == NULL);
506 	KASSERT(si->si_active);
507 
508 	havelock = false;
509 
510 	/*
511 	 * Note: due to priority inheritance we may have interrupted a
512 	 * higher priority LWP.  Since the soft interrupt must be quick
513 	 * and is non-preemptable, we don't bother yielding.
514 	 */
515 
516 	while (!SIMPLEQ_EMPTY(&si->si_q)) {
517 		/*
518 		 * Pick the longest waiting handler to run.  We block
519 		 * interrupts but do not lock in order to do this, as
520 		 * we are protecting against the local CPU only.
521 		 */
522 		sh = SIMPLEQ_FIRST(&si->si_q);
523 		SIMPLEQ_REMOVE_HEAD(&si->si_q, sh_q);
524 		KASSERT((sh->sh_flags & SOFTINT_PENDING) != 0);
525 		KASSERT((sh->sh_flags & SOFTINT_ACTIVE) == 0);
526 		sh->sh_flags ^= (SOFTINT_PENDING | SOFTINT_ACTIVE);
527 		splx(s);
528 
529 		/* Run the handler. */
530 		if (sh->sh_flags & SOFTINT_MPSAFE) {
531 			if (havelock) {
532 				KERNEL_UNLOCK_ONE(l);
533 				havelock = false;
534 			}
535 		} else if (!havelock) {
536 			KERNEL_LOCK(1, l);
537 			havelock = true;
538 		}
539 		(*sh->sh_func)(sh->sh_arg);
540 
541 		(void)splhigh();
542 		KASSERT((sh->sh_flags & SOFTINT_ACTIVE) != 0);
543 		sh->sh_flags ^= SOFTINT_ACTIVE;
544 	}
545 
546 	if (havelock) {
547 		KERNEL_UNLOCK_ONE(l);
548 	}
549 
550 	/*
551 	 * Unlocked, but only for statistics.
552 	 * Should be per-CPU to prevent cache ping-pong.
553 	 */
554 	uvmexp.softs++;
555 
556 	KASSERT(si->si_cpu == curcpu());
557 	KASSERT(si->si_lwp->l_wchan == NULL);
558 	KASSERT(si->si_active);
559 	si->si_evcnt.ev_count++;
560 	si->si_active = 0;
561 }
562 
563 /*
564  * softint_block:
565  *
566  *	Update statistics when the soft interrupt blocks.
567  */
568 void
569 softint_block(lwp_t *l)
570 {
571 	softint_t *si = l->l_private;
572 
573 	KASSERT((l->l_pflag & LP_INTR) != 0);
574 	si->si_evcnt_block.ev_count++;
575 }
576 
577 /*
578  * schednetisr:
579  *
580  *	Trigger a legacy network interrupt.  XXX Needs to go away.
581  */
582 void
583 schednetisr(int isr)
584 {
585 
586 	softint_schedule(softint_netisrs[isr]);
587 }
588 
589 #ifndef __HAVE_FAST_SOFTINTS
590 
591 #ifdef __HAVE_PREEMPTION
592 #error __HAVE_PREEMPTION requires __HAVE_FAST_SOFTINTS
593 #endif
594 
595 /*
596  * softint_init_md:
597  *
598  *	Slow path: perform machine-dependent initialization.
599  */
600 void
601 softint_init_md(lwp_t *l, u_int level, uintptr_t *machdep)
602 {
603 	softint_t *si;
604 
605 	*machdep = (1 << level);
606 	si = l->l_private;
607 
608 	lwp_lock(l);
609 	lwp_unlock_to(l, l->l_cpu->ci_schedstate.spc_mutex);
610 	lwp_lock(l);
611 	/* Cheat and make the KASSERT in softint_thread() happy. */
612 	si->si_active = 1;
613 	l->l_stat = LSRUN;
614 	sched_enqueue(l, false);
615 	lwp_unlock(l);
616 }
617 
618 /*
619  * softint_trigger:
620  *
621  *	Slow path: cause a soft interrupt handler to begin executing.
622  *	Called at IPL_HIGH.
623  */
624 void
625 softint_trigger(uintptr_t machdep)
626 {
627 	struct cpu_info *ci;
628 	lwp_t *l;
629 
630 	l = curlwp;
631 	ci = l->l_cpu;
632 	ci->ci_data.cpu_softints |= machdep;
633 	if (l == ci->ci_data.cpu_idlelwp) {
634 		cpu_need_resched(ci, 0);
635 	} else {
636 		/* MI equivalent of aston() */
637 		cpu_signotify(l);
638 	}
639 }
640 
641 /*
642  * softint_thread:
643  *
644  *	Slow path: MI software interrupt dispatch.
645  */
646 void
647 softint_thread(void *cookie)
648 {
649 	softint_t *si;
650 	lwp_t *l;
651 	int s;
652 
653 	l = curlwp;
654 	si = l->l_private;
655 
656 	for (;;) {
657 		/*
658 		 * Clear pending status and run it.  We must drop the
659 		 * spl before mi_switch(), since IPL_HIGH may be higher
660 		 * than IPL_SCHED (and it is not safe to switch at a
661 		 * higher level).
662 		 */
663 		s = splhigh();
664 		l->l_cpu->ci_data.cpu_softints &= ~si->si_machdep;
665 		softint_execute(si, l, s);
666 		splx(s);
667 
668 		lwp_lock(l);
669 		l->l_stat = LSIDL;
670 		mi_switch(l);
671 	}
672 }
673 
674 /*
675  * softint_picklwp:
676  *
677  *	Slow path: called from mi_switch() to pick the highest priority
678  *	soft interrupt LWP that needs to run.
679  */
680 lwp_t *
681 softint_picklwp(void)
682 {
683 	struct cpu_info *ci;
684 	u_int mask;
685 	softint_t *si;
686 	lwp_t *l;
687 
688 	ci = curcpu();
689 	si = ((softcpu_t *)ci->ci_data.cpu_softcpu)->sc_int;
690 	mask = ci->ci_data.cpu_softints;
691 
692 	if ((mask & (1 << SOFTINT_SERIAL)) != 0) {
693 		l = si[SOFTINT_SERIAL].si_lwp;
694 	} else if ((mask & (1 << SOFTINT_NET)) != 0) {
695 		l = si[SOFTINT_NET].si_lwp;
696 	} else if ((mask & (1 << SOFTINT_BIO)) != 0) {
697 		l = si[SOFTINT_BIO].si_lwp;
698 	} else if ((mask & (1 << SOFTINT_CLOCK)) != 0) {
699 		l = si[SOFTINT_CLOCK].si_lwp;
700 	} else {
701 		panic("softint_picklwp");
702 	}
703 
704 	return l;
705 }
706 
707 /*
708  * softint_overlay:
709  *
710  *	Slow path: called from lwp_userret() to run a soft interrupt
711  *	within the context of a user thread.
712  */
713 void
714 softint_overlay(void)
715 {
716 	struct cpu_info *ci;
717 	u_int softints, oflag;
718 	softint_t *si;
719 	pri_t obase;
720 	lwp_t *l;
721 	int s;
722 
723 	l = curlwp;
724 	KASSERT((l->l_pflag & LP_INTR) == 0);
725 
726 	/*
727 	 * Arrange to elevate priority if the LWP blocks.  Also, bind LWP
728 	 * to the CPU.  Note: disable kernel preemption before doing that.
729 	 */
730 	s = splhigh();
731 	ci = l->l_cpu;
732 	si = ((softcpu_t *)ci->ci_data.cpu_softcpu)->sc_int;
733 
734 	obase = l->l_kpribase;
735 	l->l_kpribase = PRI_KERNEL_RT;
736 	oflag = l->l_pflag;
737 	l->l_pflag = oflag | LP_INTR | LP_BOUND;
738 
739 	while ((softints = ci->ci_data.cpu_softints) != 0) {
740 		if ((softints & (1 << SOFTINT_SERIAL)) != 0) {
741 			ci->ci_data.cpu_softints &= ~(1 << SOFTINT_SERIAL);
742 			softint_execute(&si[SOFTINT_SERIAL], l, s);
743 			continue;
744 		}
745 		if ((softints & (1 << SOFTINT_NET)) != 0) {
746 			ci->ci_data.cpu_softints &= ~(1 << SOFTINT_NET);
747 			softint_execute(&si[SOFTINT_NET], l, s);
748 			continue;
749 		}
750 		if ((softints & (1 << SOFTINT_BIO)) != 0) {
751 			ci->ci_data.cpu_softints &= ~(1 << SOFTINT_BIO);
752 			softint_execute(&si[SOFTINT_BIO], l, s);
753 			continue;
754 		}
755 		if ((softints & (1 << SOFTINT_CLOCK)) != 0) {
756 			ci->ci_data.cpu_softints &= ~(1 << SOFTINT_CLOCK);
757 			softint_execute(&si[SOFTINT_CLOCK], l, s);
758 			continue;
759 		}
760 	}
761 	l->l_pflag = oflag;
762 	l->l_kpribase = obase;
763 	splx(s);
764 }
765 
766 #else	/*  !__HAVE_FAST_SOFTINTS */
767 
768 /*
769  * softint_thread:
770  *
771  *	Fast path: the LWP is switched to without restoring any state,
772  *	so we should not arrive here - there is a direct handoff between
773  *	the interrupt stub and softint_dispatch().
774  */
775 void
776 softint_thread(void *cookie)
777 {
778 
779 	panic("softint_thread");
780 }
781 
782 /*
783  * softint_dispatch:
784  *
785  *	Fast path: entry point from machine-dependent code.
786  */
787 void
788 softint_dispatch(lwp_t *pinned, int s)
789 {
790 	struct bintime now;
791 	softint_t *si;
792 	u_int timing;
793 	lwp_t *l;
794 
795 	KASSERT((pinned->l_pflag & LP_RUNNING) != 0);
796 	l = curlwp;
797 	si = l->l_private;
798 
799 	/*
800 	 * Note the interrupted LWP, and mark the current LWP as running
801 	 * before proceeding.  Although this must as a rule be done with
802 	 * the LWP locked, at this point no external agents will want to
803 	 * modify the interrupt LWP's state.
804 	 */
805 	timing = (softint_timing ? LP_TIMEINTR : 0);
806 	l->l_switchto = pinned;
807 	l->l_stat = LSONPROC;
808 	l->l_pflag |= (LP_RUNNING | timing);
809 
810 	/*
811 	 * Dispatch the interrupt.  If softints are being timed, charge
812 	 * for it.
813 	 */
814 	if (timing)
815 		binuptime(&l->l_stime);
816 	softint_execute(si, l, s);
817 	if (timing) {
818 		binuptime(&now);
819 		updatertime(l, &now);
820 		l->l_pflag &= ~LP_TIMEINTR;
821 	}
822 
823 	/*
824 	 * If we blocked while handling the interrupt, the pinned LWP is
825 	 * gone so switch to the idle LWP.  It will select a new LWP to
826 	 * run.
827 	 *
828 	 * We must drop the priority level as switching at IPL_HIGH could
829 	 * deadlock the system.  We have already set si->si_active = 0,
830 	 * which means another interrupt at this level can be triggered.
831 	 * That's not be a problem: we are lowering to level 's' which will
832 	 * prevent softint_dispatch() from being reentered at level 's',
833 	 * until the priority is finally dropped to IPL_NONE on entry to
834 	 * the LWP chosen by lwp_exit_switchaway().
835 	 */
836 	l->l_stat = LSIDL;
837 	if (l->l_switchto == NULL) {
838 		splx(s);
839 		pmap_deactivate(l);
840 		lwp_exit_switchaway(l);
841 		/* NOTREACHED */
842 	}
843 	l->l_switchto = NULL;
844 	l->l_pflag &= ~LP_RUNNING;
845 }
846 
847 #endif	/* !__HAVE_FAST_SOFTINTS */
848