Home | History | Annotate | Line # | Download | only in kern
kern_softint.c revision 1.1.2.19
      1 /*	$NetBSD: kern_softint.c,v 1.1.2.19 2007/10/30 15:32:52 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 1.1.2.19 2007/10/30 15:32:52 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 #define	PRI_SOFTSERIAL	(PRI_COUNT - 1)
    203 #define	PRI_SOFTNET	(PRI_SOFTSERIAL - schedppq * 1)
    204 #define	PRI_SOFTBIO	(PRI_SOFTSERIAL - schedppq * 2)
    205 #define	PRI_SOFTCLOCK	(PRI_SOFTSERIAL - schedppq * 3)
    206 
    207 /* This could overlap with signal info in struct lwp. */
    208 typedef struct softint {
    209 	SIMPLEQ_HEAD(, softhand) si_q;
    210 	struct lwp		*si_lwp;
    211 	struct cpu_info		*si_cpu;
    212 	uintptr_t		si_machdep;
    213 	struct evcnt		si_evcnt;
    214 	struct evcnt		si_evcnt_block;
    215 	int			si_active;
    216 	char			si_name[8];
    217 	char			si_name_block[8+6];
    218 } softint_t;
    219 
    220 typedef struct softhand {
    221 	SIMPLEQ_ENTRY(softhand)	sh_q;
    222 	void			(*sh_func)(void *);
    223 	void			*sh_arg;
    224 	softint_t		*sh_isr;
    225 	u_int			sh_pending;
    226 	u_int			sh_flags;
    227 } softhand_t;
    228 
    229 typedef struct softcpu {
    230 	struct cpu_info		*sc_cpu;
    231 	softint_t		sc_int[SOFTINT_COUNT];
    232 	softhand_t		sc_hand[1];
    233 } softcpu_t;
    234 
    235 static void	softint_thread(void *);
    236 
    237 u_int		softint_bytes = 8192;
    238 u_int		softint_timing;
    239 static u_int	softint_max;
    240 static kmutex_t	softint_lock;
    241 static void	*softint_netisrs[32];
    242 
    243 /*
    244  * softint_init_isr:
    245  *
    246  *	Initialize a single interrupt level for a single CPU.
    247  */
    248 static void
    249 softint_init_isr(softcpu_t *sc, const char *desc, pri_t pri, u_int level)
    250 {
    251 	struct cpu_info *ci;
    252 	softint_t *si;
    253 	int error;
    254 
    255 	si = &sc->sc_int[level];
    256 	ci = sc->sc_cpu;
    257 	si->si_cpu = ci;
    258 
    259 	SIMPLEQ_INIT(&si->si_q);
    260 
    261 	error = kthread_create(pri, KTHREAD_MPSAFE | KTHREAD_INTR |
    262 	    KTHREAD_IDLE, ci, softint_thread, si, &si->si_lwp,
    263 	    "soft%s/%d", desc, (int)ci->ci_cpuid);
    264 	if (error != 0)
    265 		panic("softint_init_isr: error %d", error);
    266 
    267 	snprintf(si->si_name, sizeof(si->si_name), "%s/%d", desc,
    268 	    (int)ci->ci_cpuid);
    269 	evcnt_attach_dynamic(&si->si_evcnt, EVCNT_TYPE_INTR, NULL,
    270 	   "softint", si->si_name);
    271 	snprintf(si->si_name_block, sizeof(si->si_name_block), "%s block/%d",
    272 	    desc, (int)ci->ci_cpuid);
    273 	evcnt_attach_dynamic(&si->si_evcnt_block, EVCNT_TYPE_INTR, NULL,
    274 	   "softint", si->si_name_block);
    275 
    276 	si->si_lwp->l_private = si;
    277 	softint_init_md(si->si_lwp, level, &si->si_machdep);
    278 #ifdef __HAVE_FAST_SOFTINTS
    279 	si->si_lwp->l_mutex = &ci->ci_schedstate.spc_lwplock;
    280 #endif
    281 }
    282 /*
    283  * softint_init:
    284  *
    285  *	Initialize per-CPU data structures.  Called from mi_cpu_attach().
    286  */
    287 void
    288 softint_init(struct cpu_info *ci)
    289 {
    290 	static struct cpu_info *first;
    291 	softcpu_t *sc, *scfirst;
    292 	softhand_t *sh, *shmax;
    293 
    294 	if (first == NULL) {
    295 		/* Boot CPU. */
    296 		first = ci;
    297 		mutex_init(&softint_lock, MUTEX_DEFAULT, IPL_NONE);
    298 		softint_bytes = round_page(softint_bytes);
    299 		softint_max = (softint_bytes - sizeof(softcpu_t)) /
    300 		    sizeof(softhand_t);
    301 	}
    302 
    303 	sc = (softcpu_t *)uvm_km_alloc(kernel_map, softint_bytes, 0,
    304 	    UVM_KMF_WIRED | UVM_KMF_ZERO);
    305 	if (sc == NULL)
    306 		panic("softint_init_cpu: cannot allocate memory");
    307 
    308 	ci->ci_data.cpu_softcpu = sc;
    309 	ci->ci_data.cpu_softints = 0;
    310 	sc->sc_cpu = ci;
    311 
    312 	softint_init_isr(sc, "net", PRI_SOFTNET, SOFTINT_NET);
    313 	softint_init_isr(sc, "bio", PRI_SOFTBIO, SOFTINT_BIO);
    314 	softint_init_isr(sc, "clk", PRI_SOFTCLOCK, SOFTINT_CLOCK);
    315 	softint_init_isr(sc, "ser", PRI_SOFTSERIAL, SOFTINT_SERIAL);
    316 
    317 	if (first != ci) {
    318 		mutex_enter(&softint_lock);
    319 		scfirst = first->ci_data.cpu_softcpu;
    320 		sh = sc->sc_hand;
    321 		memcpy(sh, scfirst->sc_hand, sizeof(*sh) * softint_max);
    322 		/* Update pointers for this CPU. */
    323 		for (shmax = sh + softint_max; sh < shmax; sh++) {
    324 			if (sh->sh_func == NULL)
    325 				continue;
    326 			sh->sh_isr =
    327 			    &sc->sc_int[sh->sh_flags & SOFTINT_LVLMASK];
    328 		}
    329 		mutex_exit(&softint_lock);
    330 	} else {
    331 		/*
    332 		 * Establish handlers for legacy net interrupts.
    333 		 * XXX Needs to go away.
    334 		 */
    335 #define DONETISR(n, f)							\
    336     softint_netisrs[(n)] = 						\
    337         softint_establish(SOFTINT_NET, (void (*)(void *))(f), NULL)
    338 #include <net/netisr_dispatch.h>
    339 	}
    340 }
    341 
    342 /*
    343  * softint_establish:
    344  *
    345  *	Register a software interrupt handler.
    346  */
    347 void *
    348 softint_establish(u_int flags, void (*func)(void *), void *arg)
    349 {
    350 	CPU_INFO_ITERATOR cii;
    351 	struct cpu_info *ci;
    352 	softcpu_t *sc;
    353 	softhand_t *sh;
    354 	u_int level, index;
    355 
    356 	level = (flags & SOFTINT_LVLMASK);
    357 	KASSERT(level < SOFTINT_COUNT);
    358 
    359 	mutex_enter(&softint_lock);
    360 
    361 	/* Find a free slot. */
    362 	sc = curcpu()->ci_data.cpu_softcpu;
    363 	for (index = 1; index < softint_max; index++)
    364 		if (sc->sc_hand[index].sh_func == NULL)
    365 			break;
    366 	if (index == softint_max) {
    367 		mutex_exit(&softint_lock);
    368 		printf("WARNING: softint_establish: table full, "
    369 		    "increase softint_bytes\n");
    370 		return NULL;
    371 	}
    372 
    373 	/* Set up the handler on each CPU. */
    374 	for (CPU_INFO_FOREACH(cii, ci)) {
    375 		sc = ci->ci_data.cpu_softcpu;
    376 		sh = &sc->sc_hand[index];
    377 
    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 	/* Cheat and make the KASSERT in softint_thread() happy. */
    570 	si->si_active = 1;
    571 	l->l_stat = LSRUN;
    572 	sched_enqueue(l, false);
    573 	lwp_unlock(l);
    574 }
    575 
    576 /*
    577  * softint_trigger:
    578  *
    579  *	Slow path: cause a soft interrupt handler to begin executing.
    580  *	Called at IPL_HIGH.
    581  */
    582 void
    583 softint_trigger(uintptr_t machdep)
    584 {
    585 	struct cpu_info *ci;
    586 	lwp_t *l;
    587 
    588 	l = curlwp;
    589 	ci = l->l_cpu;
    590 	ci->ci_data.cpu_softints |= machdep;
    591 	if (l == ci->ci_data.cpu_idlelwp) {
    592 		cpu_need_resched(ci, 0);
    593 	} else {
    594 		/* MI equivalent of aston() */
    595 		cpu_signotify(l);
    596 	}
    597 }
    598 
    599 /*
    600  * softint_thread:
    601  *
    602  *	Slow path: MI software interrupt dispatch.
    603  */
    604 void
    605 softint_thread(void *cookie)
    606 {
    607 	softint_t *si;
    608 	lwp_t *l;
    609 	int s;
    610 
    611 	l = curlwp;
    612 	si = l->l_private;
    613 
    614 	for (;;) {
    615 		/*
    616 		 * Clear pending status and run it.  We must drop the
    617 		 * spl before mi_switch(), since IPL_HIGH may be higher
    618 		 * than IPL_SCHED (and it is not safe to switch at a
    619 		 * higher level).
    620 		 */
    621 		s = splhigh();
    622 		l->l_cpu->ci_data.cpu_softints &= ~si->si_machdep;
    623 		softint_execute(si, l, s);
    624 		splx(s);
    625 
    626 		lwp_lock(l);
    627 		l->l_stat = LSIDL;
    628 		mi_switch(l);
    629 	}
    630 }
    631 
    632 /*
    633  * softint_picklwp:
    634  *
    635  *	Slow path: called from mi_switch() to pick the highest priority
    636  *	soft interrupt LWP that needs to run.
    637  */
    638 lwp_t *
    639 softint_picklwp(void)
    640 {
    641 	struct cpu_info *ci;
    642 	u_int mask;
    643 	softint_t *si;
    644 	lwp_t *l;
    645 
    646 	ci = curcpu();
    647 	si = ((softcpu_t *)ci->ci_data.cpu_softcpu)->sc_int;
    648 	mask = ci->ci_data.cpu_softints;
    649 
    650 	if ((mask & (1 << SOFTINT_SERIAL)) != 0) {
    651 		l = si[SOFTINT_SERIAL].si_lwp;
    652 	} else if ((mask & (1 << SOFTINT_NET)) != 0) {
    653 		l = si[SOFTINT_NET].si_lwp;
    654 	} else if ((mask & (1 << SOFTINT_BIO)) != 0) {
    655 		l = si[SOFTINT_BIO].si_lwp;
    656 	} else if ((mask & (1 << SOFTINT_CLOCK)) != 0) {
    657 		l = si[SOFTINT_CLOCK].si_lwp;
    658 	} else {
    659 		panic("softint_picklwp");
    660 	}
    661 
    662 	return l;
    663 }
    664 
    665 /*
    666  * softint_overlay:
    667  *
    668  *	Slow path: called from lwp_userret() to run a soft interrupt
    669  *	within the context of a user thread.  If the LWP blocks,
    670  *	priority will be elevated in sched_kpri().
    671  */
    672 void
    673 softint_overlay(void)
    674 {
    675 	struct cpu_info *ci;
    676 	u_int softints;
    677 	softint_t *si;
    678 	lwp_t *l;
    679 	int s;
    680 
    681 	l = curlwp;
    682 	ci = l->l_cpu;
    683 	si = ((softcpu_t *)ci->ci_data.cpu_softcpu)->sc_int;
    684 
    685 	KASSERT((l->l_pflag & LP_INTR) == 0);
    686 
    687 	l->l_pflag |= LP_INTR;
    688 	s = splhigh();
    689 	while ((softints = ci->ci_data.cpu_softints) != 0) {
    690 		if ((softints & (1 << SOFTINT_SERIAL)) != 0) {
    691 			ci->ci_data.cpu_softints &= ~(1 << SOFTINT_SERIAL);
    692 			softint_execute(&si[SOFTINT_SERIAL], l, safepri);
    693 			continue;
    694 		}
    695 		if ((softints & (1 << SOFTINT_NET)) != 0) {
    696 			ci->ci_data.cpu_softints &= ~(1 << SOFTINT_NET);
    697 			softint_execute(&si[SOFTINT_NET], l, safepri);
    698 			continue;
    699 		}
    700 		if ((softints & (1 << SOFTINT_BIO)) != 0) {
    701 			ci->ci_data.cpu_softints &= ~(1 << SOFTINT_BIO);
    702 			softint_execute(&si[SOFTINT_BIO], l, safepri);
    703 			continue;
    704 		}
    705 		if ((softints & (1 << SOFTINT_CLOCK)) != 0) {
    706 			ci->ci_data.cpu_softints &= ~(1 << SOFTINT_CLOCK);
    707 			softint_execute(&si[SOFTINT_CLOCK], l, safepri);
    708 			continue;
    709 		}
    710 	}
    711 	splx(s);
    712 	l->l_pflag &= ~LP_INTR;
    713 }
    714 
    715 /*
    716  * softint_kpri:
    717  *
    718  *	Adjust priority for a blocking user LWP that is handling a
    719  *	soft interrupt.
    720  */
    721 pri_t
    722 softint_kpri(lwp_t *l)
    723 {
    724 
    725 	/* No point doing anything more fair / complicated. */
    726 	return PRI_SOFTSERIAL;
    727 }
    728 
    729 #else	/*  !__HAVE_FAST_SOFTINTS */
    730 
    731 /*
    732  * softint_thread:
    733  *
    734  *	Fast path: the LWP is switched to without restoring any state,
    735  *	so we should not arrive here - there is a direct handoff between
    736  *	the interrupt stub and softint_dispatch().
    737  */
    738 void
    739 softint_thread(void *cookie)
    740 {
    741 
    742 	panic("softint_thread");
    743 }
    744 
    745 /*
    746  * softint_dispatch:
    747  *
    748  *	Fast path: entry point from machine-dependent code.
    749  */
    750 void
    751 softint_dispatch(lwp_t *pinned, int s)
    752 {
    753 	struct timeval now;
    754 	softint_t *si;
    755 	u_int timing;
    756 	lwp_t *l;
    757 
    758 	l = curlwp;
    759 	si = l->l_private;
    760 
    761 	/*
    762 	 * Note the interrupted LWP, and mark the current LWP as running
    763 	 * before proceeding.  Although this must as a rule be done with
    764 	 * the LWP locked, at this point no external agents will want to
    765 	 * modify the interrupt LWP's state.
    766 	 */
    767 	timing = (softint_timing ? LW_TIMEINTR : 0);
    768 	l->l_switchto = pinned;
    769 	l->l_stat = LSONPROC;
    770 	l->l_flag |= (LW_RUNNING | timing);
    771 
    772 	/*
    773 	 * Dispatch the interrupt.  If softints are being timed, charge
    774 	 * for it.
    775 	 */
    776 	if (timing)
    777 		microtime(&l->l_stime);
    778 	softint_execute(si, l, s);
    779 	if (timing) {
    780 		microtime(&now);
    781 		updatertime(l, &now);
    782 		l->l_flag &= ~LW_TIMEINTR;
    783 	}
    784 
    785 	/*
    786 	 * If we blocked while handling the interrupt, the pinned LWP is
    787 	 * gone so switch to the idle LWP.  It will select a new LWP to
    788 	 * run.
    789 	 *
    790 	 * We must drop the priority level as switching at IPL_HIGH could
    791 	 * deadlock the system.  We have already set si->si_active = 0,
    792 	 * which means another interrupt at this level can be triggered.
    793 	 * That's not be a problem: we are lowering to level 's' which will
    794 	 * prevent softint_dispatch() from being reentered at level 's',
    795 	 * until the priority is finally dropped to IPL_NONE on entry to
    796 	 * the idle loop.
    797 	 */
    798 	l->l_stat = LSIDL;
    799 	if (l->l_switchto == NULL) {
    800 		splx(s);
    801 		pmap_deactivate(l);
    802 		lwp_exit_switchaway(l);
    803 		/* NOTREACHED */
    804 	}
    805 	l->l_switchto = NULL;
    806 	l->l_flag &= ~LW_RUNNING;
    807 }
    808 
    809 #endif	/* !__HAVE_FAST_SOFTINTS */
    810