1 /*	$NetBSD: kern_softint.c,v 1.43 2016/07/04 04:20:14 knakahara 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 
172 #include <sys/cdefs.h>
173 __KERNEL_RCSID(0, "$NetBSD: kern_softint.c,v 1.43 2016/07/04 04:20:14 knakahara Exp $");
174 
175 #include <sys/param.h>
176 #include <sys/proc.h>
177 #include <sys/intr.h>
178 #include <sys/ipi.h>
179 #include <sys/mutex.h>
180 #include <sys/kthread.h>
181 #include <sys/evcnt.h>
182 #include <sys/cpu.h>
183 #include <sys/xcall.h>
184 #include <sys/pserialize.h>
185 
186 #include <net/netisr.h>
187 
188 #include <uvm/uvm_extern.h>
189 
190 /* This could overlap with signal info in struct lwp. */
191 typedef struct softint {
192 	SIMPLEQ_HEAD(, softhand) si_q;
193 	struct lwp		*si_lwp;
194 	struct cpu_info		*si_cpu;
195 	uintptr_t		si_machdep;
196 	struct evcnt		si_evcnt;
197 	struct evcnt		si_evcnt_block;
198 	int			si_active;
199 	char			si_name[8];
200 	char			si_name_block[8+6];
201 } softint_t;
202 
203 typedef struct softhand {
204 	SIMPLEQ_ENTRY(softhand)	sh_q;
205 	void			(*sh_func)(void *);
206 	void			*sh_arg;
207 	softint_t		*sh_isr;
208 	u_int			sh_flags;
209 	u_int			sh_ipi_id;
210 } softhand_t;
211 
212 typedef struct softcpu {
213 	struct cpu_info		*sc_cpu;
214 	softint_t		sc_int[SOFTINT_COUNT];
215 	softhand_t		sc_hand[1];
216 } softcpu_t;
217 
218 static void	softint_thread(void *);
219 
220 u_int		softint_bytes = 8192;
221 u_int		softint_timing;
222 static u_int	softint_max;
223 static kmutex_t	softint_lock;
224 static void	*softint_netisrs[NETISR_MAX];
225 
226 /*
227  * softint_init_isr:
228  *
229  *	Initialize a single interrupt level for a single CPU.
230  */
231 static void
softint_init_isr(softcpu_t * sc,const char * desc,pri_t pri,u_int level)232 softint_init_isr(softcpu_t *sc, const char *desc, pri_t pri, u_int level)
233 {
234 	struct cpu_info *ci;
235 	softint_t *si;
236 	int error;
237 
238 	si = &sc->sc_int[level];
239 	ci = sc->sc_cpu;
240 	si->si_cpu = ci;
241 
242 	SIMPLEQ_INIT(&si->si_q);
243 
244 	error = kthread_create(pri, KTHREAD_MPSAFE | KTHREAD_INTR |
245 	    KTHREAD_IDLE, ci, softint_thread, si, &si->si_lwp,
246 	    "soft%s/%u", desc, ci->ci_index);
247 	if (error != 0)
248 		panic("softint_init_isr: error %d", error);
249 
250 	snprintf(si->si_name, sizeof(si->si_name), "%s/%u", desc,
251 	    ci->ci_index);
252 	evcnt_attach_dynamic(&si->si_evcnt, EVCNT_TYPE_MISC, NULL,
253 	   "softint", si->si_name);
254 	snprintf(si->si_name_block, sizeof(si->si_name_block), "%s block/%u",
255 	    desc, ci->ci_index);
256 	evcnt_attach_dynamic(&si->si_evcnt_block, EVCNT_TYPE_MISC, NULL,
257 	   "softint", si->si_name_block);
258 
259 	si->si_lwp->l_private = si;
260 	softint_init_md(si->si_lwp, level, &si->si_machdep);
261 }
262 
263 /*
264  * softint_init:
265  *
266  *	Initialize per-CPU data structures.  Called from mi_cpu_attach().
267  */
268 void
softint_init(struct cpu_info * ci)269 softint_init(struct cpu_info *ci)
270 {
271 	static struct cpu_info *first;
272 	softcpu_t *sc, *scfirst;
273 	softhand_t *sh, *shmax;
274 
275 	if (first == NULL) {
276 		/* Boot CPU. */
277 		first = ci;
278 		mutex_init(&softint_lock, MUTEX_DEFAULT, IPL_NONE);
279 		softint_bytes = round_page(softint_bytes);
280 		softint_max = (softint_bytes - sizeof(softcpu_t)) /
281 		    sizeof(softhand_t);
282 	}
283 
284 	/* Use uvm_km(9) for persistent, page-aligned allocation. */
285 	sc = (softcpu_t *)uvm_km_alloc(kernel_map, softint_bytes, 0,
286 	    UVM_KMF_WIRED | UVM_KMF_ZERO);
287 	if (sc == NULL)
288 		panic("softint_init_cpu: cannot allocate memory");
289 
290 	ci->ci_data.cpu_softcpu = sc;
291 	ci->ci_data.cpu_softints = 0;
292 	sc->sc_cpu = ci;
293 
294 	softint_init_isr(sc, "net", PRI_SOFTNET, SOFTINT_NET);
295 	softint_init_isr(sc, "bio", PRI_SOFTBIO, SOFTINT_BIO);
296 	softint_init_isr(sc, "clk", PRI_SOFTCLOCK, SOFTINT_CLOCK);
297 	softint_init_isr(sc, "ser", PRI_SOFTSERIAL, SOFTINT_SERIAL);
298 
299 	if (first != ci) {
300 		mutex_enter(&softint_lock);
301 		scfirst = first->ci_data.cpu_softcpu;
302 		sh = sc->sc_hand;
303 		memcpy(sh, scfirst->sc_hand, sizeof(*sh) * softint_max);
304 		/* Update pointers for this CPU. */
305 		for (shmax = sh + softint_max; sh < shmax; sh++) {
306 			if (sh->sh_func == NULL)
307 				continue;
308 			sh->sh_isr =
309 			    &sc->sc_int[sh->sh_flags & SOFTINT_LVLMASK];
310 		}
311 		mutex_exit(&softint_lock);
312 	} else {
313 		/*
314 		 * Establish handlers for legacy net interrupts.
315 		 * XXX Needs to go away.
316 		 */
317 #define DONETISR(n, f)							\
318     softint_netisrs[(n)] = softint_establish(SOFTINT_NET|SOFTINT_MPSAFE,\
319         (void (*)(void *))(f), NULL)
320 #include <net/netisr_dispatch.h>
321 	}
322 }
323 
324 /*
325  * softint_establish:
326  *
327  *	Register a software interrupt handler.
328  */
329 void *
softint_establish(u_int flags,void (* func)(void *),void * arg)330 softint_establish(u_int flags, void (*func)(void *), void *arg)
331 {
332 	CPU_INFO_ITERATOR cii;
333 	struct cpu_info *ci;
334 	softcpu_t *sc;
335 	softhand_t *sh;
336 	u_int level, index;
337 	u_int ipi_id = 0;
338 	void *sih;
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 	}
352 	if (index == softint_max) {
353 		mutex_exit(&softint_lock);
354 		printf("WARNING: softint_establish: table full, "
355 		    "increase softint_bytes\n");
356 		return NULL;
357 	}
358 	sih = (void *)((uint8_t *)&sc->sc_hand[index] - (uint8_t *)sc);
359 
360 	if (flags & SOFTINT_RCPU) {
361 		if ((ipi_id = ipi_register(softint_schedule, sih)) == 0) {
362 			mutex_exit(&softint_lock);
363 			return NULL;
364 		}
365 	}
366 
367 	/* Set up the handler on each CPU. */
368 	if (ncpu < 2) {
369 		/* XXX hack for machines with no CPU_INFO_FOREACH() early on */
370 		sc = curcpu()->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 		sh->sh_ipi_id = ipi_id;
377 	} else for (CPU_INFO_FOREACH(cii, ci)) {
378 		sc = ci->ci_data.cpu_softcpu;
379 		sh = &sc->sc_hand[index];
380 		sh->sh_isr = &sc->sc_int[level];
381 		sh->sh_func = func;
382 		sh->sh_arg = arg;
383 		sh->sh_flags = flags;
384 		sh->sh_ipi_id = ipi_id;
385 	}
386 	mutex_exit(&softint_lock);
387 
388 	return sih;
389 }
390 
391 /*
392  * softint_disestablish:
393  *
394  *	Unregister a software interrupt handler.  The soft interrupt could
395  *	still be active at this point, but the caller commits not to try
396  *	and trigger it again once this call is made.  The caller must not
397  *	hold any locks that could be taken from soft interrupt context,
398  *	because we will wait for the softint to complete if it's still
399  *	running.
400  */
401 void
softint_disestablish(void * arg)402 softint_disestablish(void *arg)
403 {
404 	CPU_INFO_ITERATOR cii;
405 	struct cpu_info *ci;
406 	softcpu_t *sc;
407 	softhand_t *sh;
408 	uintptr_t offset;
409 	uint64_t where;
410 	u_int flags;
411 
412 	offset = (uintptr_t)arg;
413 	KASSERTMSG(offset != 0 && offset < softint_bytes, "%"PRIuPTR" %u",
414 	    offset, softint_bytes);
415 
416 	/*
417 	 * Unregister an IPI handler if there is any.  Note: there is
418 	 * no need to disable preemption here - ID is stable.
419 	 */
420 	sc = curcpu()->ci_data.cpu_softcpu;
421 	sh = (softhand_t *)((uint8_t *)sc + offset);
422 	if (sh->sh_ipi_id) {
423 		ipi_unregister(sh->sh_ipi_id);
424 	}
425 
426 	/*
427 	 * Run a cross call so we see up to date values of sh_flags from
428 	 * all CPUs.  Once softint_disestablish() is called, the caller
429 	 * commits to not trigger the interrupt and set SOFTINT_ACTIVE on
430 	 * it again.  So, we are only looking for handler records with
431 	 * SOFTINT_ACTIVE already set.
432 	 */
433 	where = xc_broadcast(0, (xcfunc_t)nullop, NULL, NULL);
434 	xc_wait(where);
435 
436 	for (;;) {
437 		/* Collect flag values from each CPU. */
438 		flags = 0;
439 		for (CPU_INFO_FOREACH(cii, ci)) {
440 			sc = ci->ci_data.cpu_softcpu;
441 			sh = (softhand_t *)((uint8_t *)sc + offset);
442 			KASSERT(sh->sh_func != NULL);
443 			flags |= sh->sh_flags;
444 		}
445 		/* Inactive on all CPUs? */
446 		if ((flags & SOFTINT_ACTIVE) == 0) {
447 			break;
448 		}
449 		/* Oops, still active.  Wait for it to clear. */
450 		(void)kpause("softdis", false, 1, NULL);
451 	}
452 
453 	/* Clear the handler on each CPU. */
454 	mutex_enter(&softint_lock);
455 	for (CPU_INFO_FOREACH(cii, ci)) {
456 		sc = ci->ci_data.cpu_softcpu;
457 		sh = (softhand_t *)((uint8_t *)sc + offset);
458 		KASSERT(sh->sh_func != NULL);
459 		sh->sh_func = NULL;
460 	}
461 	mutex_exit(&softint_lock);
462 }
463 
464 /*
465  * softint_schedule:
466  *
467  *	Trigger a software interrupt.  Must be called from a hardware
468  *	interrupt handler, or with preemption disabled (since we are
469  *	using the value of curcpu()).
470  */
471 void
softint_schedule(void * arg)472 softint_schedule(void *arg)
473 {
474 	softhand_t *sh;
475 	softint_t *si;
476 	uintptr_t offset;
477 	int s;
478 
479 	KASSERT(kpreempt_disabled());
480 
481 	/* Find the handler record for this CPU. */
482 	offset = (uintptr_t)arg;
483 	KASSERTMSG(offset != 0 && offset < softint_bytes, "%"PRIuPTR" %u",
484 	    offset, softint_bytes);
485 	sh = (softhand_t *)((uint8_t *)curcpu()->ci_data.cpu_softcpu + offset);
486 
487 	/* If it's already pending there's nothing to do. */
488 	if ((sh->sh_flags & SOFTINT_PENDING) != 0) {
489 		return;
490 	}
491 
492 	/*
493 	 * Enqueue the handler into the LWP's pending list.
494 	 * If the LWP is completely idle, then make it run.
495 	 */
496 	s = splhigh();
497 	if ((sh->sh_flags & SOFTINT_PENDING) == 0) {
498 		si = sh->sh_isr;
499 		sh->sh_flags |= SOFTINT_PENDING;
500 		SIMPLEQ_INSERT_TAIL(&si->si_q, sh, sh_q);
501 		if (si->si_active == 0) {
502 			si->si_active = 1;
503 			softint_trigger(si->si_machdep);
504 		}
505 	}
506 	splx(s);
507 }
508 
509 /*
510  * softint_schedule_cpu:
511  *
512  *	Trigger a software interrupt on a target CPU.  This invokes
513  *	softint_schedule() for the local CPU or send an IPI to invoke
514  *	this routine on the remote CPU.  Preemption must be disabled.
515  */
516 void
softint_schedule_cpu(void * arg,struct cpu_info * ci)517 softint_schedule_cpu(void *arg, struct cpu_info *ci)
518 {
519 	KASSERT(kpreempt_disabled());
520 
521 	if (curcpu() != ci) {
522 		const softcpu_t *sc = ci->ci_data.cpu_softcpu;
523 		const uintptr_t offset = (uintptr_t)arg;
524 		const softhand_t *sh;
525 
526 		sh = (const softhand_t *)((const uint8_t *)sc + offset);
527 		KASSERT((sh->sh_flags & SOFTINT_RCPU) != 0);
528 		ipi_trigger(sh->sh_ipi_id, ci);
529 		return;
530 	}
531 
532 	/* Just a local CPU. */
533 	softint_schedule(arg);
534 }
535 
536 /*
537  * softint_execute:
538  *
539  *	Invoke handlers for the specified soft interrupt.
540  *	Must be entered at splhigh.  Will drop the priority
541  *	to the level specified, but returns back at splhigh.
542  */
543 static inline void
softint_execute(softint_t * si,lwp_t * l,int s)544 softint_execute(softint_t *si, lwp_t *l, int s)
545 {
546 	softhand_t *sh;
547 	bool havelock;
548 
549 #ifdef __HAVE_FAST_SOFTINTS
550 	KASSERT(si->si_lwp == curlwp);
551 #else
552 	/* May be running in user context. */
553 #endif
554 	KASSERT(si->si_cpu == curcpu());
555 	KASSERT(si->si_lwp->l_wchan == NULL);
556 	KASSERT(si->si_active);
557 
558 	havelock = false;
559 
560 	/*
561 	 * Note: due to priority inheritance we may have interrupted a
562 	 * higher priority LWP.  Since the soft interrupt must be quick
563 	 * and is non-preemptable, we don't bother yielding.
564 	 */
565 
566 	while (!SIMPLEQ_EMPTY(&si->si_q)) {
567 		/*
568 		 * Pick the longest waiting handler to run.  We block
569 		 * interrupts but do not lock in order to do this, as
570 		 * we are protecting against the local CPU only.
571 		 */
572 		sh = SIMPLEQ_FIRST(&si->si_q);
573 		SIMPLEQ_REMOVE_HEAD(&si->si_q, sh_q);
574 		KASSERT((sh->sh_flags & SOFTINT_PENDING) != 0);
575 		KASSERT((sh->sh_flags & SOFTINT_ACTIVE) == 0);
576 		sh->sh_flags ^= (SOFTINT_PENDING | SOFTINT_ACTIVE);
577 		splx(s);
578 
579 		/* Run the handler. */
580 		if (sh->sh_flags & SOFTINT_MPSAFE) {
581 			if (havelock) {
582 				KERNEL_UNLOCK_ONE(l);
583 				havelock = false;
584 			}
585 		} else if (!havelock) {
586 			KERNEL_LOCK(1, l);
587 			havelock = true;
588 		}
589 		(*sh->sh_func)(sh->sh_arg);
590 
591 		/* Diagnostic: check that spin-locks have not leaked. */
592 		KASSERTMSG(curcpu()->ci_mtx_count == 0,
593 		    "%s: ci_mtx_count (%d) != 0, sh_func %p\n",
594 		    __func__, curcpu()->ci_mtx_count, sh->sh_func);
595 
596 		(void)splhigh();
597 		KASSERT((sh->sh_flags & SOFTINT_ACTIVE) != 0);
598 		sh->sh_flags ^= SOFTINT_ACTIVE;
599 	}
600 
601 	if (havelock) {
602 		KERNEL_UNLOCK_ONE(l);
603 	}
604 
605 	/*
606 	 * Unlocked, but only for statistics.
607 	 * Should be per-CPU to prevent cache ping-pong.
608 	 */
609 	curcpu()->ci_data.cpu_nsoft++;
610 
611 	KASSERT(si->si_cpu == curcpu());
612 	KASSERT(si->si_lwp->l_wchan == NULL);
613 	KASSERT(si->si_active);
614 	si->si_evcnt.ev_count++;
615 	si->si_active = 0;
616 }
617 
618 /*
619  * softint_block:
620  *
621  *	Update statistics when the soft interrupt blocks.
622  */
623 void
softint_block(lwp_t * l)624 softint_block(lwp_t *l)
625 {
626 	softint_t *si = l->l_private;
627 
628 	KASSERT((l->l_pflag & LP_INTR) != 0);
629 	si->si_evcnt_block.ev_count++;
630 }
631 
632 /*
633  * schednetisr:
634  *
635  *	Trigger a legacy network interrupt.  XXX Needs to go away.
636  */
637 void
schednetisr(int isr)638 schednetisr(int isr)
639 {
640 
641 	softint_schedule(softint_netisrs[isr]);
642 }
643 
644 #ifndef __HAVE_FAST_SOFTINTS
645 
646 #ifdef __HAVE_PREEMPTION
647 #error __HAVE_PREEMPTION requires __HAVE_FAST_SOFTINTS
648 #endif
649 
650 /*
651  * softint_init_md:
652  *
653  *	Slow path: perform machine-dependent initialization.
654  */
655 void
softint_init_md(lwp_t * l,u_int level,uintptr_t * machdep)656 softint_init_md(lwp_t *l, u_int level, uintptr_t *machdep)
657 {
658 	softint_t *si;
659 
660 	*machdep = (1 << level);
661 	si = l->l_private;
662 
663 	lwp_lock(l);
664 	lwp_unlock_to(l, l->l_cpu->ci_schedstate.spc_mutex);
665 	lwp_lock(l);
666 	/* Cheat and make the KASSERT in softint_thread() happy. */
667 	si->si_active = 1;
668 	l->l_stat = LSRUN;
669 	sched_enqueue(l, false);
670 	lwp_unlock(l);
671 }
672 
673 /*
674  * softint_trigger:
675  *
676  *	Slow path: cause a soft interrupt handler to begin executing.
677  *	Called at IPL_HIGH.
678  */
679 void
softint_trigger(uintptr_t machdep)680 softint_trigger(uintptr_t machdep)
681 {
682 	struct cpu_info *ci;
683 	lwp_t *l;
684 
685 	l = curlwp;
686 	ci = l->l_cpu;
687 	ci->ci_data.cpu_softints |= machdep;
688 	if (l == ci->ci_data.cpu_idlelwp) {
689 		cpu_need_resched(ci, 0);
690 	} else {
691 		/* MI equivalent of aston() */
692 		cpu_signotify(l);
693 	}
694 }
695 
696 /*
697  * softint_thread:
698  *
699  *	Slow path: MI software interrupt dispatch.
700  */
701 void
softint_thread(void * cookie)702 softint_thread(void *cookie)
703 {
704 	softint_t *si;
705 	lwp_t *l;
706 	int s;
707 
708 	l = curlwp;
709 	si = l->l_private;
710 
711 	for (;;) {
712 		/*
713 		 * Clear pending status and run it.  We must drop the
714 		 * spl before mi_switch(), since IPL_HIGH may be higher
715 		 * than IPL_SCHED (and it is not safe to switch at a
716 		 * higher level).
717 		 */
718 		s = splhigh();
719 		l->l_cpu->ci_data.cpu_softints &= ~si->si_machdep;
720 		softint_execute(si, l, s);
721 		splx(s);
722 
723 		lwp_lock(l);
724 		l->l_stat = LSIDL;
725 		mi_switch(l);
726 	}
727 }
728 
729 /*
730  * softint_picklwp:
731  *
732  *	Slow path: called from mi_switch() to pick the highest priority
733  *	soft interrupt LWP that needs to run.
734  */
735 lwp_t *
softint_picklwp(void)736 softint_picklwp(void)
737 {
738 	struct cpu_info *ci;
739 	u_int mask;
740 	softint_t *si;
741 	lwp_t *l;
742 
743 	ci = curcpu();
744 	si = ((softcpu_t *)ci->ci_data.cpu_softcpu)->sc_int;
745 	mask = ci->ci_data.cpu_softints;
746 
747 	if ((mask & (1 << SOFTINT_SERIAL)) != 0) {
748 		l = si[SOFTINT_SERIAL].si_lwp;
749 	} else if ((mask & (1 << SOFTINT_NET)) != 0) {
750 		l = si[SOFTINT_NET].si_lwp;
751 	} else if ((mask & (1 << SOFTINT_BIO)) != 0) {
752 		l = si[SOFTINT_BIO].si_lwp;
753 	} else if ((mask & (1 << SOFTINT_CLOCK)) != 0) {
754 		l = si[SOFTINT_CLOCK].si_lwp;
755 	} else {
756 		panic("softint_picklwp");
757 	}
758 
759 	return l;
760 }
761 
762 /*
763  * softint_overlay:
764  *
765  *	Slow path: called from lwp_userret() to run a soft interrupt
766  *	within the context of a user thread.
767  */
768 void
softint_overlay(void)769 softint_overlay(void)
770 {
771 	struct cpu_info *ci;
772 	u_int softints, oflag;
773 	softint_t *si;
774 	pri_t obase;
775 	lwp_t *l;
776 	int s;
777 
778 	l = curlwp;
779 	KASSERT((l->l_pflag & LP_INTR) == 0);
780 
781 	/*
782 	 * Arrange to elevate priority if the LWP blocks.  Also, bind LWP
783 	 * to the CPU.  Note: disable kernel preemption before doing that.
784 	 */
785 	s = splhigh();
786 	ci = l->l_cpu;
787 	si = ((softcpu_t *)ci->ci_data.cpu_softcpu)->sc_int;
788 
789 	obase = l->l_kpribase;
790 	l->l_kpribase = PRI_KERNEL_RT;
791 	oflag = l->l_pflag;
792 	l->l_pflag = oflag | LP_INTR | LP_BOUND;
793 
794 	while ((softints = ci->ci_data.cpu_softints) != 0) {
795 		if ((softints & (1 << SOFTINT_SERIAL)) != 0) {
796 			ci->ci_data.cpu_softints &= ~(1 << SOFTINT_SERIAL);
797 			softint_execute(&si[SOFTINT_SERIAL], l, s);
798 			continue;
799 		}
800 		if ((softints & (1 << SOFTINT_NET)) != 0) {
801 			ci->ci_data.cpu_softints &= ~(1 << SOFTINT_NET);
802 			softint_execute(&si[SOFTINT_NET], l, s);
803 			continue;
804 		}
805 		if ((softints & (1 << SOFTINT_BIO)) != 0) {
806 			ci->ci_data.cpu_softints &= ~(1 << SOFTINT_BIO);
807 			softint_execute(&si[SOFTINT_BIO], l, s);
808 			continue;
809 		}
810 		if ((softints & (1 << SOFTINT_CLOCK)) != 0) {
811 			ci->ci_data.cpu_softints &= ~(1 << SOFTINT_CLOCK);
812 			softint_execute(&si[SOFTINT_CLOCK], l, s);
813 			continue;
814 		}
815 	}
816 	l->l_pflag = oflag;
817 	l->l_kpribase = obase;
818 	splx(s);
819 }
820 
821 #else	/*  !__HAVE_FAST_SOFTINTS */
822 
823 /*
824  * softint_thread:
825  *
826  *	Fast path: the LWP is switched to without restoring any state,
827  *	so we should not arrive here - there is a direct handoff between
828  *	the interrupt stub and softint_dispatch().
829  */
830 void
softint_thread(void * cookie)831 softint_thread(void *cookie)
832 {
833 
834 	panic("softint_thread");
835 }
836 
837 /*
838  * softint_dispatch:
839  *
840  *	Fast path: entry point from machine-dependent code.
841  */
842 void
softint_dispatch(lwp_t * pinned,int s)843 softint_dispatch(lwp_t *pinned, int s)
844 {
845 	struct bintime now;
846 	softint_t *si;
847 	u_int timing;
848 	lwp_t *l;
849 
850 	KASSERT((pinned->l_pflag & LP_RUNNING) != 0);
851 	l = curlwp;
852 	si = l->l_private;
853 
854 	/*
855 	 * Note the interrupted LWP, and mark the current LWP as running
856 	 * before proceeding.  Although this must as a rule be done with
857 	 * the LWP locked, at this point no external agents will want to
858 	 * modify the interrupt LWP's state.
859 	 */
860 	timing = (softint_timing ? LP_TIMEINTR : 0);
861 	l->l_switchto = pinned;
862 	l->l_stat = LSONPROC;
863 	l->l_pflag |= (LP_RUNNING | timing);
864 
865 	/*
866 	 * Dispatch the interrupt.  If softints are being timed, charge
867 	 * for it.
868 	 */
869 	if (timing)
870 		binuptime(&l->l_stime);
871 	softint_execute(si, l, s);
872 	if (timing) {
873 		binuptime(&now);
874 		updatertime(l, &now);
875 		l->l_pflag &= ~LP_TIMEINTR;
876 	}
877 
878 	/* Indicate a soft-interrupt switch. */
879 	pserialize_switchpoint();
880 
881 	/*
882 	 * If we blocked while handling the interrupt, the pinned LWP is
883 	 * gone so switch to the idle LWP.  It will select a new LWP to
884 	 * run.
885 	 *
886 	 * We must drop the priority level as switching at IPL_HIGH could
887 	 * deadlock the system.  We have already set si->si_active = 0,
888 	 * which means another interrupt at this level can be triggered.
889 	 * That's not be a problem: we are lowering to level 's' which will
890 	 * prevent softint_dispatch() from being reentered at level 's',
891 	 * until the priority is finally dropped to IPL_NONE on entry to
892 	 * the LWP chosen by lwp_exit_switchaway().
893 	 */
894 	l->l_stat = LSIDL;
895 	if (l->l_switchto == NULL) {
896 		splx(s);
897 		pmap_deactivate(l);
898 		lwp_exit_switchaway(l);
899 		/* NOTREACHED */
900 	}
901 	l->l_switchto = NULL;
902 	l->l_pflag &= ~LP_RUNNING;
903 }
904 
905 #endif	/* !__HAVE_FAST_SOFTINTS */
906