xref: /netbsd-src/sys/kern/kern_softint.c (revision a536ee5124e62c9a0051a252f7833dc8f50f44c9)
1 /*	$NetBSD: kern_softint.c,v 1.38 2011/09/27 01:02:38 jym 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 1.38 2011/09/27 01:02:38 jym Exp $");
180 
181 #include <sys/param.h>
182 #include <sys/proc.h>
183 #include <sys/intr.h>
184 #include <sys/mutex.h>
185 #include <sys/kthread.h>
186 #include <sys/evcnt.h>
187 #include <sys/cpu.h>
188 #include <sys/xcall.h>
189 
190 #include <net/netisr.h>
191 
192 #include <uvm/uvm_extern.h>
193 
194 /* This could overlap with signal info in struct lwp. */
195 typedef struct softint {
196 	SIMPLEQ_HEAD(, softhand) si_q;
197 	struct lwp		*si_lwp;
198 	struct cpu_info		*si_cpu;
199 	uintptr_t		si_machdep;
200 	struct evcnt		si_evcnt;
201 	struct evcnt		si_evcnt_block;
202 	int			si_active;
203 	char			si_name[8];
204 	char			si_name_block[8+6];
205 } softint_t;
206 
207 typedef struct softhand {
208 	SIMPLEQ_ENTRY(softhand)	sh_q;
209 	void			(*sh_func)(void *);
210 	void			*sh_arg;
211 	softint_t		*sh_isr;
212 	u_int			sh_flags;
213 } softhand_t;
214 
215 typedef struct softcpu {
216 	struct cpu_info		*sc_cpu;
217 	softint_t		sc_int[SOFTINT_COUNT];
218 	softhand_t		sc_hand[1];
219 } softcpu_t;
220 
221 static void	softint_thread(void *);
222 
223 u_int		softint_bytes = 8192;
224 u_int		softint_timing;
225 static u_int	softint_max;
226 static kmutex_t	softint_lock;
227 static void	*softint_netisrs[NETISR_MAX];
228 
229 /*
230  * softint_init_isr:
231  *
232  *	Initialize a single interrupt level for a single CPU.
233  */
234 static void
235 softint_init_isr(softcpu_t *sc, const char *desc, pri_t pri, u_int level)
236 {
237 	struct cpu_info *ci;
238 	softint_t *si;
239 	int error;
240 
241 	si = &sc->sc_int[level];
242 	ci = sc->sc_cpu;
243 	si->si_cpu = ci;
244 
245 	SIMPLEQ_INIT(&si->si_q);
246 
247 	error = kthread_create(pri, KTHREAD_MPSAFE | KTHREAD_INTR |
248 	    KTHREAD_IDLE, ci, softint_thread, si, &si->si_lwp,
249 	    "soft%s/%u", desc, ci->ci_index);
250 	if (error != 0)
251 		panic("softint_init_isr: error %d", error);
252 
253 	snprintf(si->si_name, sizeof(si->si_name), "%s/%u", desc,
254 	    ci->ci_index);
255 	evcnt_attach_dynamic(&si->si_evcnt, EVCNT_TYPE_MISC, NULL,
256 	   "softint", si->si_name);
257 	snprintf(si->si_name_block, sizeof(si->si_name_block), "%s block/%u",
258 	    desc, ci->ci_index);
259 	evcnt_attach_dynamic(&si->si_evcnt_block, EVCNT_TYPE_MISC, NULL,
260 	   "softint", si->si_name_block);
261 
262 	si->si_lwp->l_private = si;
263 	softint_init_md(si->si_lwp, level, &si->si_machdep);
264 }
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 	/* Use uvm_km(9) for persistent, page-aligned allocation. */
288 	sc = (softcpu_t *)uvm_km_alloc(kernel_map, softint_bytes, 0,
289 	    UVM_KMF_WIRED | UVM_KMF_ZERO);
290 	if (sc == NULL)
291 		panic("softint_init_cpu: cannot allocate memory");
292 
293 	ci->ci_data.cpu_softcpu = sc;
294 	ci->ci_data.cpu_softints = 0;
295 	sc->sc_cpu = ci;
296 
297 	softint_init_isr(sc, "net", PRI_SOFTNET, SOFTINT_NET);
298 	softint_init_isr(sc, "bio", PRI_SOFTBIO, SOFTINT_BIO);
299 	softint_init_isr(sc, "clk", PRI_SOFTCLOCK, SOFTINT_CLOCK);
300 	softint_init_isr(sc, "ser", PRI_SOFTSERIAL, SOFTINT_SERIAL);
301 
302 	if (first != ci) {
303 		mutex_enter(&softint_lock);
304 		scfirst = first->ci_data.cpu_softcpu;
305 		sh = sc->sc_hand;
306 		memcpy(sh, scfirst->sc_hand, sizeof(*sh) * softint_max);
307 		/* Update pointers for this CPU. */
308 		for (shmax = sh + softint_max; sh < shmax; sh++) {
309 			if (sh->sh_func == NULL)
310 				continue;
311 			sh->sh_isr =
312 			    &sc->sc_int[sh->sh_flags & SOFTINT_LVLMASK];
313 		}
314 		mutex_exit(&softint_lock);
315 	} else {
316 		/*
317 		 * Establish handlers for legacy net interrupts.
318 		 * XXX Needs to go away.
319 		 */
320 #define DONETISR(n, f)							\
321     softint_netisrs[(n)] = softint_establish(SOFTINT_NET|SOFTINT_MPSAFE,\
322         (void (*)(void *))(f), NULL)
323 #include <net/netisr_dispatch.h>
324 	}
325 }
326 
327 /*
328  * softint_establish:
329  *
330  *	Register a software interrupt handler.
331  */
332 void *
333 softint_establish(u_int flags, void (*func)(void *), void *arg)
334 {
335 	CPU_INFO_ITERATOR cii;
336 	struct cpu_info *ci;
337 	softcpu_t *sc;
338 	softhand_t *sh;
339 	u_int level, index;
340 
341 	level = (flags & SOFTINT_LVLMASK);
342 	KASSERT(level < SOFTINT_COUNT);
343 	KASSERT((flags & SOFTINT_IMPMASK) == 0);
344 
345 	mutex_enter(&softint_lock);
346 
347 	/* Find a free slot. */
348 	sc = curcpu()->ci_data.cpu_softcpu;
349 	for (index = 1; index < softint_max; index++) {
350 		if (sc->sc_hand[index].sh_func == NULL)
351 			break;
352 	}
353 	if (index == softint_max) {
354 		mutex_exit(&softint_lock);
355 		printf("WARNING: softint_establish: table full, "
356 		    "increase softint_bytes\n");
357 		return NULL;
358 	}
359 
360 	/* Set up the handler on each CPU. */
361 	if (ncpu < 2) {
362 		/* XXX hack for machines with no CPU_INFO_FOREACH() early on */
363 		sc = curcpu()->ci_data.cpu_softcpu;
364 		sh = &sc->sc_hand[index];
365 		sh->sh_isr = &sc->sc_int[level];
366 		sh->sh_func = func;
367 		sh->sh_arg = arg;
368 		sh->sh_flags = flags;
369 	} else for (CPU_INFO_FOREACH(cii, ci)) {
370 		sc = ci->ci_data.cpu_softcpu;
371 		sh = &sc->sc_hand[index];
372 		sh->sh_isr = &sc->sc_int[level];
373 		sh->sh_func = func;
374 		sh->sh_arg = arg;
375 		sh->sh_flags = flags;
376 	}
377 
378 	mutex_exit(&softint_lock);
379 
380 	return (void *)((uint8_t *)&sc->sc_hand[index] - (uint8_t *)sc);
381 }
382 
383 /*
384  * softint_disestablish:
385  *
386  *	Unregister a software interrupt handler.  The soft interrupt could
387  *	still be active at this point, but the caller commits not to try
388  *	and trigger it again once this call is made.  The caller must not
389  *	hold any locks that could be taken from soft interrupt context,
390  *	because we will wait for the softint to complete if it's still
391  *	running.
392  */
393 void
394 softint_disestablish(void *arg)
395 {
396 	CPU_INFO_ITERATOR cii;
397 	struct cpu_info *ci;
398 	softcpu_t *sc;
399 	softhand_t *sh;
400 	uintptr_t offset;
401 	uint64_t where;
402 	u_int flags;
403 
404 	offset = (uintptr_t)arg;
405 	KASSERT(offset != 0 && offset < softint_bytes);
406 
407 	/*
408 	 * Run a cross call so we see up to date values of sh_flags from
409 	 * all CPUs.  Once softint_disestablish() is called, the caller
410 	 * commits to not trigger the interrupt and set SOFTINT_ACTIVE on
411 	 * it again.  So, we are only looking for handler records with
412 	 * SOFTINT_ACTIVE already set.
413 	 */
414 	where = xc_broadcast(0, (xcfunc_t)nullop, NULL, NULL);
415 	xc_wait(where);
416 
417 	for (;;) {
418 		/* Collect flag values from each CPU. */
419 		flags = 0;
420 		for (CPU_INFO_FOREACH(cii, ci)) {
421 			sc = ci->ci_data.cpu_softcpu;
422 			sh = (softhand_t *)((uint8_t *)sc + offset);
423 			KASSERT(sh->sh_func != NULL);
424 			flags |= sh->sh_flags;
425 		}
426 		/* Inactive on all CPUs? */
427 		if ((flags & SOFTINT_ACTIVE) == 0) {
428 			break;
429 		}
430 		/* Oops, still active.  Wait for it to clear. */
431 		(void)kpause("softdis", false, 1, NULL);
432 	}
433 
434 	/* Clear the handler on each CPU. */
435 	mutex_enter(&softint_lock);
436 	for (CPU_INFO_FOREACH(cii, ci)) {
437 		sc = ci->ci_data.cpu_softcpu;
438 		sh = (softhand_t *)((uint8_t *)sc + offset);
439 		KASSERT(sh->sh_func != NULL);
440 		sh->sh_func = NULL;
441 	}
442 	mutex_exit(&softint_lock);
443 }
444 
445 /*
446  * softint_schedule:
447  *
448  *	Trigger a software interrupt.  Must be called from a hardware
449  *	interrupt handler, or with preemption disabled (since we are
450  *	using the value of curcpu()).
451  */
452 void
453 softint_schedule(void *arg)
454 {
455 	softhand_t *sh;
456 	softint_t *si;
457 	uintptr_t offset;
458 	int s;
459 
460 	KASSERT(kpreempt_disabled());
461 
462 	/* Find the handler record for this CPU. */
463 	offset = (uintptr_t)arg;
464 	KASSERT(offset != 0 && offset < softint_bytes);
465 	sh = (softhand_t *)((uint8_t *)curcpu()->ci_data.cpu_softcpu + offset);
466 
467 	/* If it's already pending there's nothing to do. */
468 	if ((sh->sh_flags & SOFTINT_PENDING) != 0) {
469 		return;
470 	}
471 
472 	/*
473 	 * Enqueue the handler into the LWP's pending list.
474 	 * If the LWP is completely idle, then make it run.
475 	 */
476 	s = splhigh();
477 	if ((sh->sh_flags & SOFTINT_PENDING) == 0) {
478 		si = sh->sh_isr;
479 		sh->sh_flags |= SOFTINT_PENDING;
480 		SIMPLEQ_INSERT_TAIL(&si->si_q, sh, sh_q);
481 		if (si->si_active == 0) {
482 			si->si_active = 1;
483 			softint_trigger(si->si_machdep);
484 		}
485 	}
486 	splx(s);
487 }
488 
489 /*
490  * softint_execute:
491  *
492  *	Invoke handlers for the specified soft interrupt.
493  *	Must be entered at splhigh.  Will drop the priority
494  *	to the level specified, but returns back at splhigh.
495  */
496 static inline void
497 softint_execute(softint_t *si, lwp_t *l, int s)
498 {
499 	softhand_t *sh;
500 	bool havelock;
501 
502 #ifdef __HAVE_FAST_SOFTINTS
503 	KASSERT(si->si_lwp == curlwp);
504 #else
505 	/* May be running in user context. */
506 #endif
507 	KASSERT(si->si_cpu == curcpu());
508 	KASSERT(si->si_lwp->l_wchan == NULL);
509 	KASSERT(si->si_active);
510 
511 	havelock = false;
512 
513 	/*
514 	 * Note: due to priority inheritance we may have interrupted a
515 	 * higher priority LWP.  Since the soft interrupt must be quick
516 	 * and is non-preemptable, we don't bother yielding.
517 	 */
518 
519 	while (!SIMPLEQ_EMPTY(&si->si_q)) {
520 		/*
521 		 * Pick the longest waiting handler to run.  We block
522 		 * interrupts but do not lock in order to do this, as
523 		 * we are protecting against the local CPU only.
524 		 */
525 		sh = SIMPLEQ_FIRST(&si->si_q);
526 		SIMPLEQ_REMOVE_HEAD(&si->si_q, sh_q);
527 		KASSERT((sh->sh_flags & SOFTINT_PENDING) != 0);
528 		KASSERT((sh->sh_flags & SOFTINT_ACTIVE) == 0);
529 		sh->sh_flags ^= (SOFTINT_PENDING | SOFTINT_ACTIVE);
530 		splx(s);
531 
532 		/* Run the handler. */
533 		if (sh->sh_flags & SOFTINT_MPSAFE) {
534 			if (havelock) {
535 				KERNEL_UNLOCK_ONE(l);
536 				havelock = false;
537 			}
538 		} else if (!havelock) {
539 			KERNEL_LOCK(1, l);
540 			havelock = true;
541 		}
542 		(*sh->sh_func)(sh->sh_arg);
543 
544 		/* Diagnostic: check that spin-locks have not leaked. */
545 		KASSERTMSG(curcpu()->ci_mtx_count == 0,
546 		    "%s: ci_mtx_count (%d) != 0, sh_func %p\n",
547 		    __func__, curcpu()->ci_mtx_count, sh->sh_func);
548 
549 		(void)splhigh();
550 		KASSERT((sh->sh_flags & SOFTINT_ACTIVE) != 0);
551 		sh->sh_flags ^= SOFTINT_ACTIVE;
552 	}
553 
554 	if (havelock) {
555 		KERNEL_UNLOCK_ONE(l);
556 	}
557 
558 	/*
559 	 * Unlocked, but only for statistics.
560 	 * Should be per-CPU to prevent cache ping-pong.
561 	 */
562 	curcpu()->ci_data.cpu_nsoft++;
563 
564 	KASSERT(si->si_cpu == curcpu());
565 	KASSERT(si->si_lwp->l_wchan == NULL);
566 	KASSERT(si->si_active);
567 	si->si_evcnt.ev_count++;
568 	si->si_active = 0;
569 }
570 
571 /*
572  * softint_block:
573  *
574  *	Update statistics when the soft interrupt blocks.
575  */
576 void
577 softint_block(lwp_t *l)
578 {
579 	softint_t *si = l->l_private;
580 
581 	KASSERT((l->l_pflag & LP_INTR) != 0);
582 	si->si_evcnt_block.ev_count++;
583 }
584 
585 /*
586  * schednetisr:
587  *
588  *	Trigger a legacy network interrupt.  XXX Needs to go away.
589  */
590 void
591 schednetisr(int isr)
592 {
593 
594 	softint_schedule(softint_netisrs[isr]);
595 }
596 
597 #ifndef __HAVE_FAST_SOFTINTS
598 
599 #ifdef __HAVE_PREEMPTION
600 #error __HAVE_PREEMPTION requires __HAVE_FAST_SOFTINTS
601 #endif
602 
603 /*
604  * softint_init_md:
605  *
606  *	Slow path: perform machine-dependent initialization.
607  */
608 void
609 softint_init_md(lwp_t *l, u_int level, uintptr_t *machdep)
610 {
611 	softint_t *si;
612 
613 	*machdep = (1 << level);
614 	si = l->l_private;
615 
616 	lwp_lock(l);
617 	lwp_unlock_to(l, l->l_cpu->ci_schedstate.spc_mutex);
618 	lwp_lock(l);
619 	/* Cheat and make the KASSERT in softint_thread() happy. */
620 	si->si_active = 1;
621 	l->l_stat = LSRUN;
622 	sched_enqueue(l, false);
623 	lwp_unlock(l);
624 }
625 
626 /*
627  * softint_trigger:
628  *
629  *	Slow path: cause a soft interrupt handler to begin executing.
630  *	Called at IPL_HIGH.
631  */
632 void
633 softint_trigger(uintptr_t machdep)
634 {
635 	struct cpu_info *ci;
636 	lwp_t *l;
637 
638 	l = curlwp;
639 	ci = l->l_cpu;
640 	ci->ci_data.cpu_softints |= machdep;
641 	if (l == ci->ci_data.cpu_idlelwp) {
642 		cpu_need_resched(ci, 0);
643 	} else {
644 		/* MI equivalent of aston() */
645 		cpu_signotify(l);
646 	}
647 }
648 
649 /*
650  * softint_thread:
651  *
652  *	Slow path: MI software interrupt dispatch.
653  */
654 void
655 softint_thread(void *cookie)
656 {
657 	softint_t *si;
658 	lwp_t *l;
659 	int s;
660 
661 	l = curlwp;
662 	si = l->l_private;
663 
664 	for (;;) {
665 		/*
666 		 * Clear pending status and run it.  We must drop the
667 		 * spl before mi_switch(), since IPL_HIGH may be higher
668 		 * than IPL_SCHED (and it is not safe to switch at a
669 		 * higher level).
670 		 */
671 		s = splhigh();
672 		l->l_cpu->ci_data.cpu_softints &= ~si->si_machdep;
673 		softint_execute(si, l, s);
674 		splx(s);
675 
676 		lwp_lock(l);
677 		l->l_stat = LSIDL;
678 		mi_switch(l);
679 	}
680 }
681 
682 /*
683  * softint_picklwp:
684  *
685  *	Slow path: called from mi_switch() to pick the highest priority
686  *	soft interrupt LWP that needs to run.
687  */
688 lwp_t *
689 softint_picklwp(void)
690 {
691 	struct cpu_info *ci;
692 	u_int mask;
693 	softint_t *si;
694 	lwp_t *l;
695 
696 	ci = curcpu();
697 	si = ((softcpu_t *)ci->ci_data.cpu_softcpu)->sc_int;
698 	mask = ci->ci_data.cpu_softints;
699 
700 	if ((mask & (1 << SOFTINT_SERIAL)) != 0) {
701 		l = si[SOFTINT_SERIAL].si_lwp;
702 	} else if ((mask & (1 << SOFTINT_NET)) != 0) {
703 		l = si[SOFTINT_NET].si_lwp;
704 	} else if ((mask & (1 << SOFTINT_BIO)) != 0) {
705 		l = si[SOFTINT_BIO].si_lwp;
706 	} else if ((mask & (1 << SOFTINT_CLOCK)) != 0) {
707 		l = si[SOFTINT_CLOCK].si_lwp;
708 	} else {
709 		panic("softint_picklwp");
710 	}
711 
712 	return l;
713 }
714 
715 /*
716  * softint_overlay:
717  *
718  *	Slow path: called from lwp_userret() to run a soft interrupt
719  *	within the context of a user thread.
720  */
721 void
722 softint_overlay(void)
723 {
724 	struct cpu_info *ci;
725 	u_int softints, oflag;
726 	softint_t *si;
727 	pri_t obase;
728 	lwp_t *l;
729 	int s;
730 
731 	l = curlwp;
732 	KASSERT((l->l_pflag & LP_INTR) == 0);
733 
734 	/*
735 	 * Arrange to elevate priority if the LWP blocks.  Also, bind LWP
736 	 * to the CPU.  Note: disable kernel preemption before doing that.
737 	 */
738 	s = splhigh();
739 	ci = l->l_cpu;
740 	si = ((softcpu_t *)ci->ci_data.cpu_softcpu)->sc_int;
741 
742 	obase = l->l_kpribase;
743 	l->l_kpribase = PRI_KERNEL_RT;
744 	oflag = l->l_pflag;
745 	l->l_pflag = oflag | LP_INTR | LP_BOUND;
746 
747 	while ((softints = ci->ci_data.cpu_softints) != 0) {
748 		if ((softints & (1 << SOFTINT_SERIAL)) != 0) {
749 			ci->ci_data.cpu_softints &= ~(1 << SOFTINT_SERIAL);
750 			softint_execute(&si[SOFTINT_SERIAL], l, s);
751 			continue;
752 		}
753 		if ((softints & (1 << SOFTINT_NET)) != 0) {
754 			ci->ci_data.cpu_softints &= ~(1 << SOFTINT_NET);
755 			softint_execute(&si[SOFTINT_NET], l, s);
756 			continue;
757 		}
758 		if ((softints & (1 << SOFTINT_BIO)) != 0) {
759 			ci->ci_data.cpu_softints &= ~(1 << SOFTINT_BIO);
760 			softint_execute(&si[SOFTINT_BIO], l, s);
761 			continue;
762 		}
763 		if ((softints & (1 << SOFTINT_CLOCK)) != 0) {
764 			ci->ci_data.cpu_softints &= ~(1 << SOFTINT_CLOCK);
765 			softint_execute(&si[SOFTINT_CLOCK], l, s);
766 			continue;
767 		}
768 	}
769 	l->l_pflag = oflag;
770 	l->l_kpribase = obase;
771 	splx(s);
772 }
773 
774 #else	/*  !__HAVE_FAST_SOFTINTS */
775 
776 /*
777  * softint_thread:
778  *
779  *	Fast path: the LWP is switched to without restoring any state,
780  *	so we should not arrive here - there is a direct handoff between
781  *	the interrupt stub and softint_dispatch().
782  */
783 void
784 softint_thread(void *cookie)
785 {
786 
787 	panic("softint_thread");
788 }
789 
790 /*
791  * softint_dispatch:
792  *
793  *	Fast path: entry point from machine-dependent code.
794  */
795 void
796 softint_dispatch(lwp_t *pinned, int s)
797 {
798 	struct bintime now;
799 	softint_t *si;
800 	u_int timing;
801 	lwp_t *l;
802 
803 	KASSERT((pinned->l_pflag & LP_RUNNING) != 0);
804 	l = curlwp;
805 	si = l->l_private;
806 
807 	/*
808 	 * Note the interrupted LWP, and mark the current LWP as running
809 	 * before proceeding.  Although this must as a rule be done with
810 	 * the LWP locked, at this point no external agents will want to
811 	 * modify the interrupt LWP's state.
812 	 */
813 	timing = (softint_timing ? LP_TIMEINTR : 0);
814 	l->l_switchto = pinned;
815 	l->l_stat = LSONPROC;
816 	l->l_pflag |= (LP_RUNNING | timing);
817 
818 	/*
819 	 * Dispatch the interrupt.  If softints are being timed, charge
820 	 * for it.
821 	 */
822 	if (timing)
823 		binuptime(&l->l_stime);
824 	softint_execute(si, l, s);
825 	if (timing) {
826 		binuptime(&now);
827 		updatertime(l, &now);
828 		l->l_pflag &= ~LP_TIMEINTR;
829 	}
830 
831 	/*
832 	 * If we blocked while handling the interrupt, the pinned LWP is
833 	 * gone so switch to the idle LWP.  It will select a new LWP to
834 	 * run.
835 	 *
836 	 * We must drop the priority level as switching at IPL_HIGH could
837 	 * deadlock the system.  We have already set si->si_active = 0,
838 	 * which means another interrupt at this level can be triggered.
839 	 * That's not be a problem: we are lowering to level 's' which will
840 	 * prevent softint_dispatch() from being reentered at level 's',
841 	 * until the priority is finally dropped to IPL_NONE on entry to
842 	 * the LWP chosen by lwp_exit_switchaway().
843 	 */
844 	l->l_stat = LSIDL;
845 	if (l->l_switchto == NULL) {
846 		splx(s);
847 		pmap_deactivate(l);
848 		lwp_exit_switchaway(l);
849 		/* NOTREACHED */
850 	}
851 	l->l_switchto = NULL;
852 	l->l_pflag &= ~LP_RUNNING;
853 }
854 
855 #endif	/* !__HAVE_FAST_SOFTINTS */
856