Home | History | Annotate | Line # | Download | only in kern
kern_softint.c revision 1.6.2.1
      1  1.6.2.1  ad /*	$NetBSD: kern_softint.c,v 1.6.2.1 2007/12/26 19:57:09 ad Exp $	*/
      2      1.2  ad 
      3      1.2  ad /*-
      4      1.2  ad  * Copyright (c) 2007 The NetBSD Foundation, Inc.
      5      1.2  ad  * All rights reserved.
      6      1.2  ad  *
      7      1.2  ad  * This code is derived from software contributed to The NetBSD Foundation
      8      1.2  ad  * by Andrew Doran.
      9      1.2  ad  *
     10      1.2  ad  * Redistribution and use in source and binary forms, with or without
     11      1.2  ad  * modification, are permitted provided that the following conditions
     12      1.2  ad  * are met:
     13      1.2  ad  * 1. Redistributions of source code must retain the above copyright
     14      1.2  ad  *    notice, this list of conditions and the following disclaimer.
     15      1.2  ad  * 2. Redistributions in binary form must reproduce the above copyright
     16      1.2  ad  *    notice, this list of conditions and the following disclaimer in the
     17      1.2  ad  *    documentation and/or other materials provided with the distribution.
     18      1.2  ad  * 3. All advertising materials mentioning features or use of this software
     19      1.2  ad  *    must display the following acknowledgement:
     20      1.2  ad  *	This product includes software developed by the NetBSD
     21      1.2  ad  *	Foundation, Inc. and its contributors.
     22      1.2  ad  * 4. Neither the name of The NetBSD Foundation nor the names of its
     23      1.2  ad  *    contributors may be used to endorse or promote products derived
     24      1.2  ad  *    from this software without specific prior written permission.
     25      1.2  ad  *
     26      1.2  ad  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
     27      1.2  ad  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
     28      1.2  ad  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
     29      1.2  ad  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
     30      1.2  ad  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
     31      1.2  ad  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
     32      1.2  ad  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
     33      1.2  ad  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
     34      1.2  ad  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
     35      1.2  ad  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
     36      1.2  ad  * POSSIBILITY OF SUCH DAMAGE.
     37      1.2  ad  */
     38      1.2  ad 
     39      1.2  ad /*
     40      1.5  ad  * Generic software interrupt framework.
     41      1.5  ad  *
     42      1.5  ad  * Overview
     43      1.5  ad  *
     44      1.5  ad  *	The soft interrupt framework provides a mechanism to schedule a
     45      1.5  ad  *	low priority callback that runs with thread context.  It allows
     46      1.5  ad  *	for dynamic registration of software interrupts, and for fair
     47      1.5  ad  *	queueing and prioritization of those interrupts.  The callbacks
     48      1.5  ad  *	can be scheduled to run from nearly any point in the kernel: by
     49      1.5  ad  *	code running with thread context, by code running from a
     50      1.5  ad  *	hardware interrupt handler, and at any interrupt priority
     51      1.5  ad  *	level.
     52      1.5  ad  *
     53      1.5  ad  * Priority levels
     54      1.5  ad  *
     55      1.5  ad  *	Since soft interrupt dispatch can be tied to the underlying
     56      1.5  ad  *	architecture's interrupt dispatch code, it can be limited
     57      1.5  ad  *	both by the capabilities of the hardware and the capabilities
     58      1.5  ad  *	of the interrupt dispatch code itself.  The number of priority
     59      1.5  ad  *	levels is restricted to four.  In order of priority (lowest to
     60      1.5  ad  *	highest) the levels are: clock, bio, net, serial.
     61      1.5  ad  *
     62      1.5  ad  *	The names are symbolic and in isolation do not have any direct
     63      1.5  ad  *	connection with a particular kind of device activity: they are
     64      1.5  ad  *	only meant as a guide.
     65      1.5  ad  *
     66      1.5  ad  *	The four priority levels map directly to scheduler priority
     67      1.5  ad  *	levels, and where the architecture implements 'fast' software
     68      1.5  ad  *	interrupts, they also map onto interrupt priorities.  The
     69      1.5  ad  *	interrupt priorities are intended to be hidden from machine
     70      1.5  ad  *	independent code, which should use thread-safe mechanisms to
     71      1.5  ad  *	synchronize with software interrupts (for example: mutexes).
     72      1.5  ad  *
     73      1.5  ad  * Capabilities
     74      1.5  ad  *
     75      1.5  ad  *	Software interrupts run with limited machine context.  In
     76      1.5  ad  *	particular, they do not posess any address space context.  They
     77      1.5  ad  *	should not try to operate on user space addresses, or to use
     78      1.5  ad  *	virtual memory facilities other than those noted as interrupt
     79      1.5  ad  *	safe.
     80      1.5  ad  *
     81      1.5  ad  *	Unlike hardware interrupts, software interrupts do have thread
     82      1.5  ad  *	context.  They may block on synchronization objects, sleep, and
     83      1.5  ad  *	resume execution at a later time.
     84      1.5  ad  *
     85      1.5  ad  *	Since software interrupts are a limited resource and run with
     86      1.5  ad  *	higher priority than most other LWPs in the system, all
     87      1.5  ad  *	block-and-resume activity by a software interrupt must be kept
     88      1.5  ad  *	short to allow futher processing at that level to continue.  By
     89      1.5  ad  *	extension, code running with process context must take care to
     90      1.5  ad  *	ensure that any lock that may be taken from a software interrupt
     91      1.5  ad  *	can not be held for more than a short period of time.
     92      1.5  ad  *
     93      1.5  ad  *	The kernel does not allow software interrupts to use facilities
     94      1.5  ad  *	or perform actions that may block for a significant amount of
     95      1.5  ad  *	time.  This means that it's not valid for a software interrupt
     96      1.5  ad  *	to: sleep on condition variables, use the lockmgr() facility,
     97      1.5  ad  *	or wait for resources to become available (for example,
     98      1.5  ad  *	memory).
     99      1.5  ad  *
    100      1.5  ad  * Per-CPU operation
    101      1.5  ad  *
    102      1.5  ad  *	If a soft interrupt is triggered on a CPU, it can only be
    103      1.5  ad  *	dispatched on the same CPU.  Each LWP dedicated to handling a
    104      1.5  ad  *	soft interrupt is bound to its home CPU, so if the LWP blocks
    105      1.5  ad  *	and needs to run again, it can only run there.  Nearly all data
    106      1.5  ad  *	structures used to manage software interrupts are per-CPU.
    107      1.5  ad  *
    108      1.5  ad  *	The per-CPU requirement is intended to reduce "ping-pong" of
    109      1.5  ad  *	cache lines between CPUs: lines occupied by data structures
    110      1.5  ad  *	used to manage the soft interrupts, and lines occupied by data
    111      1.5  ad  *	items being passed down to the soft interrupt.  As a positive
    112      1.5  ad  *	side effect, this also means that the soft interrupt dispatch
    113      1.5  ad  *	code does not need to to use spinlocks to synchronize.
    114      1.5  ad  *
    115      1.5  ad  * Generic implementation
    116      1.5  ad  *
    117      1.5  ad  *	A generic, low performance implementation is provided that
    118      1.5  ad  *	works across all architectures, with no machine-dependent
    119      1.5  ad  *	modifications needed.  This implementation uses the scheduler,
    120      1.5  ad  *	and so has a number of restrictions:
    121      1.5  ad  *
    122      1.5  ad  *	1) The software interrupts are not currently preemptive, so
    123      1.5  ad  *	must wait for the currently executing LWP to yield the CPU.
    124      1.5  ad  *	This can introduce latency.
    125      1.5  ad  *
    126      1.5  ad  *	2) An expensive context switch is required for a software
    127      1.5  ad  *	interrupt to be handled.
    128      1.5  ad  *
    129      1.5  ad  * 'Fast' software interrupts
    130      1.5  ad  *
    131      1.5  ad  *	If an architectures defines __HAVE_FAST_SOFTINTS, it implements
    132      1.5  ad  *	the fast mechanism.  Threads running either in the kernel or in
    133      1.5  ad  *	userspace will be interrupted, but will not be preempted.  When
    134      1.5  ad  *	the soft interrupt completes execution, the interrupted LWP
    135      1.5  ad  *	is resumed.  Interrupt dispatch code must provide the minimum
    136      1.5  ad  *	level of context necessary for the soft interrupt to block and
    137      1.5  ad  *	be resumed at a later time.  The machine-dependent dispatch
    138      1.5  ad  *	path looks something like the following:
    139      1.5  ad  *
    140      1.5  ad  *	softintr()
    141      1.5  ad  *	{
    142      1.5  ad  *		go to IPL_HIGH if necessary for switch;
    143      1.5  ad  *		save any necessary registers in a format that can be
    144      1.5  ad  *		    restored by cpu_switchto if the softint blocks;
    145      1.5  ad  *		arrange for cpu_switchto() to restore into the
    146      1.5  ad  *		    trampoline function;
    147      1.5  ad  *		identify LWP to handle this interrupt;
    148      1.5  ad  *		switch to the LWP's stack;
    149      1.5  ad  *		switch register stacks, if necessary;
    150      1.5  ad  *		assign new value of curlwp;
    151      1.5  ad  *		call MI softint_dispatch, passing old curlwp and IPL
    152      1.5  ad  *		    to execute interrupt at;
    153      1.5  ad  *		switch back to old stack;
    154      1.5  ad  *		switch back to old register stack, if necessary;
    155      1.5  ad  *		restore curlwp;
    156      1.5  ad  *		return to interrupted LWP;
    157      1.5  ad  *	}
    158      1.5  ad  *
    159      1.5  ad  *	If the soft interrupt blocks, a trampoline function is returned
    160      1.5  ad  *	to in the context of the interrupted LWP, as arranged for by
    161      1.5  ad  *	softint():
    162      1.5  ad  *
    163      1.5  ad  *	softint_ret()
    164      1.5  ad  *	{
    165      1.5  ad  *		unlock soft interrupt LWP;
    166      1.5  ad  *		resume interrupt processing, likely returning to
    167      1.5  ad  *		    interrupted LWP or dispatching another, different
    168      1.5  ad  *		    interrupt;
    169      1.5  ad  *	}
    170      1.5  ad  *
    171      1.5  ad  *	Once the soft interrupt has fired (and even if it has blocked),
    172      1.5  ad  *	no further soft interrupts at that level will be triggered by
    173      1.5  ad  *	MI code until the soft interrupt handler has ceased execution.
    174      1.5  ad  *	If a soft interrupt handler blocks and is resumed, it resumes
    175      1.5  ad  *	execution as a normal LWP (kthread) and gains VM context.  Only
    176      1.5  ad  *	when it has completed and is ready to fire again will it
    177      1.5  ad  *	interrupt other threads.
    178      1.5  ad  *
    179      1.5  ad  * Future directions
    180      1.5  ad  *
    181      1.5  ad  *	Provide a cheap way to direct software interrupts to remote
    182      1.5  ad  *	CPUs.  Provide a way to enqueue work items into the handler
    183      1.5  ad  *	record,	removing additional spl calls (see subr_workqueue.c).
    184      1.2  ad  */
    185      1.2  ad 
    186      1.2  ad #include <sys/cdefs.h>
    187  1.6.2.1  ad __KERNEL_RCSID(0, "$NetBSD: kern_softint.c,v 1.6.2.1 2007/12/26 19:57:09 ad Exp $");
    188      1.2  ad 
    189      1.2  ad #include <sys/param.h>
    190      1.5  ad #include <sys/malloc.h>
    191      1.5  ad #include <sys/proc.h>
    192      1.2  ad #include <sys/intr.h>
    193      1.5  ad #include <sys/mutex.h>
    194      1.5  ad #include <sys/kthread.h>
    195      1.5  ad #include <sys/evcnt.h>
    196      1.5  ad #include <sys/cpu.h>
    197      1.5  ad 
    198      1.5  ad #include <net/netisr.h>
    199      1.5  ad 
    200      1.5  ad #include <uvm/uvm_extern.h>
    201      1.5  ad 
    202      1.5  ad /* This could overlap with signal info in struct lwp. */
    203      1.5  ad typedef struct softint {
    204      1.5  ad 	SIMPLEQ_HEAD(, softhand) si_q;
    205      1.5  ad 	struct lwp		*si_lwp;
    206      1.5  ad 	struct cpu_info		*si_cpu;
    207      1.5  ad 	uintptr_t		si_machdep;
    208      1.5  ad 	struct evcnt		si_evcnt;
    209      1.5  ad 	struct evcnt		si_evcnt_block;
    210      1.5  ad 	int			si_active;
    211      1.5  ad 	char			si_name[8];
    212      1.5  ad 	char			si_name_block[8+6];
    213      1.5  ad } softint_t;
    214      1.5  ad 
    215      1.5  ad typedef struct softhand {
    216      1.5  ad 	SIMPLEQ_ENTRY(softhand)	sh_q;
    217      1.5  ad 	void			(*sh_func)(void *);
    218      1.5  ad 	void			*sh_arg;
    219      1.5  ad 	softint_t		*sh_isr;
    220      1.5  ad 	u_int			sh_pending;
    221      1.5  ad 	u_int			sh_flags;
    222      1.5  ad } softhand_t;
    223      1.5  ad 
    224      1.5  ad typedef struct softcpu {
    225      1.5  ad 	struct cpu_info		*sc_cpu;
    226      1.5  ad 	softint_t		sc_int[SOFTINT_COUNT];
    227      1.5  ad 	softhand_t		sc_hand[1];
    228      1.5  ad } softcpu_t;
    229      1.5  ad 
    230      1.5  ad static void	softint_thread(void *);
    231      1.5  ad 
    232      1.5  ad u_int		softint_bytes = 8192;
    233      1.5  ad u_int		softint_timing;
    234      1.5  ad static u_int	softint_max;
    235      1.5  ad static kmutex_t	softint_lock;
    236      1.5  ad static void	*softint_netisrs[32];
    237      1.2  ad 
    238      1.5  ad /*
    239      1.5  ad  * softint_init_isr:
    240      1.5  ad  *
    241      1.5  ad  *	Initialize a single interrupt level for a single CPU.
    242      1.5  ad  */
    243      1.5  ad static void
    244      1.5  ad softint_init_isr(softcpu_t *sc, const char *desc, pri_t pri, u_int level)
    245      1.5  ad {
    246      1.5  ad 	struct cpu_info *ci;
    247      1.5  ad 	softint_t *si;
    248      1.5  ad 	int error;
    249      1.5  ad 
    250      1.5  ad 	si = &sc->sc_int[level];
    251      1.5  ad 	ci = sc->sc_cpu;
    252      1.5  ad 	si->si_cpu = ci;
    253      1.5  ad 
    254      1.5  ad 	SIMPLEQ_INIT(&si->si_q);
    255      1.5  ad 
    256      1.5  ad 	error = kthread_create(pri, KTHREAD_MPSAFE | KTHREAD_INTR |
    257      1.5  ad 	    KTHREAD_IDLE, ci, softint_thread, si, &si->si_lwp,
    258      1.5  ad 	    "soft%s/%d", desc, (int)ci->ci_cpuid);
    259      1.5  ad 	if (error != 0)
    260      1.5  ad 		panic("softint_init_isr: error %d", error);
    261      1.5  ad 
    262      1.5  ad 	snprintf(si->si_name, sizeof(si->si_name), "%s/%d", desc,
    263      1.5  ad 	    (int)ci->ci_cpuid);
    264      1.5  ad 	evcnt_attach_dynamic(&si->si_evcnt, EVCNT_TYPE_INTR, NULL,
    265      1.5  ad 	   "softint", si->si_name);
    266      1.5  ad 	snprintf(si->si_name_block, sizeof(si->si_name_block), "%s block/%d",
    267      1.5  ad 	    desc, (int)ci->ci_cpuid);
    268      1.5  ad 	evcnt_attach_dynamic(&si->si_evcnt_block, EVCNT_TYPE_INTR, NULL,
    269      1.5  ad 	   "softint", si->si_name_block);
    270      1.3  ad 
    271      1.5  ad 	si->si_lwp->l_private = si;
    272      1.5  ad 	softint_init_md(si->si_lwp, level, &si->si_machdep);
    273      1.5  ad }
    274      1.2  ad /*
    275      1.2  ad  * softint_init:
    276      1.2  ad  *
    277      1.2  ad  *	Initialize per-CPU data structures.  Called from mi_cpu_attach().
    278      1.2  ad  */
    279      1.2  ad void
    280      1.2  ad softint_init(struct cpu_info *ci)
    281      1.2  ad {
    282      1.5  ad 	static struct cpu_info *first;
    283      1.5  ad 	softcpu_t *sc, *scfirst;
    284      1.5  ad 	softhand_t *sh, *shmax;
    285      1.5  ad 
    286      1.5  ad 	if (first == NULL) {
    287      1.5  ad 		/* Boot CPU. */
    288      1.5  ad 		first = ci;
    289      1.5  ad 		mutex_init(&softint_lock, MUTEX_DEFAULT, IPL_NONE);
    290      1.5  ad 		softint_bytes = round_page(softint_bytes);
    291      1.5  ad 		softint_max = (softint_bytes - sizeof(softcpu_t)) /
    292      1.5  ad 		    sizeof(softhand_t);
    293      1.5  ad 	}
    294      1.2  ad 
    295      1.5  ad 	sc = (softcpu_t *)uvm_km_alloc(kernel_map, softint_bytes, 0,
    296      1.5  ad 	    UVM_KMF_WIRED | UVM_KMF_ZERO);
    297      1.5  ad 	if (sc == NULL)
    298      1.5  ad 		panic("softint_init_cpu: cannot allocate memory");
    299      1.5  ad 
    300      1.5  ad 	ci->ci_data.cpu_softcpu = sc;
    301      1.5  ad 	ci->ci_data.cpu_softints = 0;
    302      1.5  ad 	sc->sc_cpu = ci;
    303      1.5  ad 
    304      1.5  ad 	softint_init_isr(sc, "net", PRI_SOFTNET, SOFTINT_NET);
    305      1.5  ad 	softint_init_isr(sc, "bio", PRI_SOFTBIO, SOFTINT_BIO);
    306      1.5  ad 	softint_init_isr(sc, "clk", PRI_SOFTCLOCK, SOFTINT_CLOCK);
    307      1.5  ad 	softint_init_isr(sc, "ser", PRI_SOFTSERIAL, SOFTINT_SERIAL);
    308      1.5  ad 
    309      1.5  ad 	if (first != ci) {
    310      1.5  ad 		mutex_enter(&softint_lock);
    311      1.5  ad 		scfirst = first->ci_data.cpu_softcpu;
    312      1.5  ad 		sh = sc->sc_hand;
    313      1.5  ad 		memcpy(sh, scfirst->sc_hand, sizeof(*sh) * softint_max);
    314      1.5  ad 		/* Update pointers for this CPU. */
    315      1.5  ad 		for (shmax = sh + softint_max; sh < shmax; sh++) {
    316      1.5  ad 			if (sh->sh_func == NULL)
    317      1.5  ad 				continue;
    318      1.5  ad 			sh->sh_isr =
    319      1.5  ad 			    &sc->sc_int[sh->sh_flags & SOFTINT_LVLMASK];
    320      1.5  ad 		}
    321      1.5  ad 		mutex_exit(&softint_lock);
    322      1.5  ad 	} else {
    323      1.5  ad 		/*
    324      1.5  ad 		 * Establish handlers for legacy net interrupts.
    325      1.5  ad 		 * XXX Needs to go away.
    326      1.5  ad 		 */
    327      1.5  ad #define DONETISR(n, f)							\
    328      1.5  ad     softint_netisrs[(n)] = 						\
    329      1.5  ad         softint_establish(SOFTINT_NET, (void (*)(void *))(f), NULL)
    330      1.5  ad #include <net/netisr_dispatch.h>
    331      1.5  ad 	}
    332      1.2  ad }
    333      1.2  ad 
    334      1.2  ad /*
    335      1.2  ad  * softint_establish:
    336      1.2  ad  *
    337      1.2  ad  *	Register a software interrupt handler.
    338      1.2  ad  */
    339      1.2  ad void *
    340      1.2  ad softint_establish(u_int flags, void (*func)(void *), void *arg)
    341      1.2  ad {
    342      1.5  ad 	CPU_INFO_ITERATOR cii;
    343      1.5  ad 	struct cpu_info *ci;
    344      1.5  ad 	softcpu_t *sc;
    345      1.5  ad 	softhand_t *sh;
    346      1.5  ad 	u_int level, index;
    347      1.2  ad 
    348      1.2  ad 	level = (flags & SOFTINT_LVLMASK);
    349      1.2  ad 	KASSERT(level < SOFTINT_COUNT);
    350      1.2  ad 
    351      1.5  ad 	mutex_enter(&softint_lock);
    352      1.5  ad 
    353      1.5  ad 	/* Find a free slot. */
    354      1.5  ad 	sc = curcpu()->ci_data.cpu_softcpu;
    355      1.5  ad 	for (index = 1; index < softint_max; index++)
    356      1.5  ad 		if (sc->sc_hand[index].sh_func == NULL)
    357      1.5  ad 			break;
    358      1.5  ad 	if (index == softint_max) {
    359      1.5  ad 		mutex_exit(&softint_lock);
    360      1.5  ad 		printf("WARNING: softint_establish: table full, "
    361      1.5  ad 		    "increase softint_bytes\n");
    362      1.5  ad 		return NULL;
    363      1.5  ad 	}
    364      1.5  ad 
    365      1.5  ad 	/* Set up the handler on each CPU. */
    366  1.6.2.1  ad 	if (ncpu < 2) {
    367  1.6.2.1  ad 		/* XXX hack for machines with no CPU_INFO_FOREACH() early on */
    368  1.6.2.1  ad 		sc = curcpu()->ci_data.cpu_softcpu;
    369  1.6.2.1  ad 		sh = &sc->sc_hand[index];
    370  1.6.2.1  ad 		sh->sh_isr = &sc->sc_int[level];
    371  1.6.2.1  ad 		sh->sh_func = func;
    372  1.6.2.1  ad 		sh->sh_arg = arg;
    373  1.6.2.1  ad 		sh->sh_flags = flags;
    374  1.6.2.1  ad 		sh->sh_pending = 0;
    375  1.6.2.1  ad 	} else for (CPU_INFO_FOREACH(cii, ci)) {
    376      1.5  ad 		sc = ci->ci_data.cpu_softcpu;
    377      1.5  ad 		sh = &sc->sc_hand[index];
    378      1.5  ad 		sh->sh_isr = &sc->sc_int[level];
    379      1.5  ad 		sh->sh_func = func;
    380      1.5  ad 		sh->sh_arg = arg;
    381      1.5  ad 		sh->sh_flags = flags;
    382      1.5  ad 		sh->sh_pending = 0;
    383      1.2  ad 	}
    384      1.2  ad 
    385      1.5  ad 	mutex_exit(&softint_lock);
    386      1.5  ad 
    387      1.5  ad 	return (void *)((uint8_t *)&sc->sc_hand[index] - (uint8_t *)sc);
    388      1.2  ad }
    389      1.2  ad 
    390      1.2  ad /*
    391      1.2  ad  * softint_disestablish:
    392      1.2  ad  *
    393      1.2  ad  *	Unregister a software interrupt handler.
    394      1.2  ad  */
    395      1.2  ad void
    396      1.2  ad softint_disestablish(void *arg)
    397      1.2  ad {
    398      1.5  ad 	CPU_INFO_ITERATOR cii;
    399      1.5  ad 	struct cpu_info *ci;
    400      1.5  ad 	softcpu_t *sc;
    401      1.5  ad 	softhand_t *sh;
    402      1.5  ad 	uintptr_t offset;
    403      1.5  ad 
    404      1.5  ad 	offset = (uintptr_t)arg;
    405      1.5  ad 	KASSERT(offset != 0 && offset < softint_bytes);
    406      1.5  ad 
    407      1.5  ad 	mutex_enter(&softint_lock);
    408      1.5  ad 
    409      1.5  ad 	/* Clear the handler on each CPU. */
    410      1.5  ad 	for (CPU_INFO_FOREACH(cii, ci)) {
    411      1.5  ad 		sc = ci->ci_data.cpu_softcpu;
    412      1.5  ad 		sh = (softhand_t *)((uint8_t *)sc + offset);
    413      1.5  ad 		KASSERT(sh->sh_func != NULL);
    414      1.5  ad 		KASSERT(sh->sh_pending == 0);
    415      1.5  ad 		sh->sh_func = NULL;
    416      1.5  ad 	}
    417      1.2  ad 
    418      1.5  ad 	mutex_exit(&softint_lock);
    419      1.2  ad }
    420      1.2  ad 
    421      1.2  ad /*
    422      1.2  ad  * softint_schedule:
    423      1.2  ad  *
    424      1.2  ad  *	Trigger a software interrupt.  Must be called from a hardware
    425      1.2  ad  *	interrupt handler, or with preemption disabled (since we are
    426      1.2  ad  *	using the value of curcpu()).
    427      1.2  ad  */
    428      1.2  ad void
    429      1.2  ad softint_schedule(void *arg)
    430      1.2  ad {
    431      1.5  ad 	softhand_t *sh;
    432      1.5  ad 	softint_t *si;
    433      1.5  ad 	uintptr_t offset;
    434      1.5  ad 	int s;
    435      1.5  ad 
    436      1.5  ad 	/* Find the handler record for this CPU. */
    437      1.5  ad 	offset = (uintptr_t)arg;
    438      1.5  ad 	KASSERT(offset != 0 && offset < softint_bytes);
    439      1.5  ad 	sh = (softhand_t *)((uint8_t *)curcpu()->ci_data.cpu_softcpu + offset);
    440      1.5  ad 
    441      1.5  ad 	/* If it's already pending there's nothing to do. */
    442      1.5  ad 	if (sh->sh_pending)
    443      1.5  ad 		return;
    444      1.5  ad 
    445      1.5  ad 	/*
    446      1.5  ad 	 * Enqueue the handler into the LWP's pending list.
    447      1.5  ad 	 * If the LWP is completely idle, then make it run.
    448      1.5  ad 	 */
    449      1.5  ad 	s = splhigh();
    450      1.5  ad 	if (!sh->sh_pending) {
    451      1.5  ad 		si = sh->sh_isr;
    452      1.5  ad 		sh->sh_pending = 1;
    453      1.5  ad 		SIMPLEQ_INSERT_TAIL(&si->si_q, sh, sh_q);
    454      1.5  ad 		if (si->si_active == 0) {
    455      1.5  ad 			si->si_active = 1;
    456      1.5  ad 			softint_trigger(si->si_machdep);
    457      1.5  ad 		}
    458      1.5  ad 	}
    459      1.5  ad 	splx(s);
    460      1.5  ad }
    461      1.5  ad 
    462      1.5  ad /*
    463      1.5  ad  * softint_execute:
    464      1.5  ad  *
    465      1.5  ad  *	Invoke handlers for the specified soft interrupt.
    466      1.5  ad  *	Must be entered at splhigh.  Will drop the priority
    467      1.5  ad  *	to the level specified, but returns back at splhigh.
    468      1.5  ad  */
    469      1.5  ad static inline void
    470      1.5  ad softint_execute(softint_t *si, lwp_t *l, int s)
    471      1.5  ad {
    472      1.5  ad 	softhand_t *sh;
    473      1.5  ad 	bool havelock;
    474      1.5  ad 
    475      1.5  ad #ifdef __HAVE_FAST_SOFTINTS
    476      1.5  ad 	KASSERT(si->si_lwp == curlwp);
    477      1.5  ad #else
    478      1.5  ad 	/* May be running in user context. */
    479      1.5  ad #endif
    480      1.5  ad 	KASSERT(si->si_cpu == curcpu());
    481      1.5  ad 	KASSERT(si->si_lwp->l_wchan == NULL);
    482      1.5  ad 	KASSERT(si->si_active);
    483      1.5  ad 
    484      1.5  ad 	havelock = false;
    485      1.5  ad 
    486      1.5  ad 	/*
    487      1.5  ad 	 * Note: due to priority inheritance we may have interrupted a
    488      1.5  ad 	 * higher priority LWP.  Since the soft interrupt must be quick
    489      1.5  ad 	 * and is non-preemptable, we don't bother yielding.
    490      1.5  ad 	 */
    491      1.5  ad 
    492      1.5  ad 	while (!SIMPLEQ_EMPTY(&si->si_q)) {
    493      1.5  ad 		/*
    494      1.5  ad 		 * Pick the longest waiting handler to run.  We block
    495      1.5  ad 		 * interrupts but do not lock in order to do this, as
    496      1.5  ad 		 * we are protecting against the local CPU only.
    497      1.5  ad 		 */
    498      1.5  ad 		sh = SIMPLEQ_FIRST(&si->si_q);
    499      1.5  ad 		SIMPLEQ_REMOVE_HEAD(&si->si_q, sh_q);
    500      1.5  ad 		sh->sh_pending = 0;
    501      1.5  ad 		splx(s);
    502      1.5  ad 
    503      1.5  ad 		/* Run the handler. */
    504      1.5  ad 		if ((sh->sh_flags & SOFTINT_MPSAFE) == 0 && !havelock) {
    505      1.5  ad 			KERNEL_LOCK(1, l);
    506      1.5  ad 			havelock = true;
    507      1.5  ad 		}
    508      1.5  ad 		(*sh->sh_func)(sh->sh_arg);
    509      1.5  ad 
    510      1.5  ad 		(void)splhigh();
    511      1.5  ad 	}
    512      1.2  ad 
    513      1.5  ad 	if (havelock) {
    514      1.5  ad 		KERNEL_UNLOCK_ONE(l);
    515      1.5  ad 	}
    516      1.5  ad 
    517      1.5  ad 	/*
    518      1.5  ad 	 * Unlocked, but only for statistics.
    519      1.5  ad 	 * Should be per-CPU to prevent cache ping-pong.
    520      1.5  ad 	 */
    521      1.5  ad 	uvmexp.softs++;
    522      1.5  ad 
    523      1.5  ad 	si->si_evcnt.ev_count++;
    524      1.5  ad 	si->si_active = 0;
    525      1.2  ad }
    526      1.2  ad 
    527      1.2  ad /*
    528      1.2  ad  * softint_block:
    529      1.2  ad  *
    530      1.2  ad  *	Update statistics when the soft interrupt blocks.
    531      1.2  ad  */
    532      1.2  ad void
    533      1.2  ad softint_block(lwp_t *l)
    534      1.2  ad {
    535      1.5  ad 	softint_t *si = l->l_private;
    536      1.5  ad 
    537      1.5  ad 	KASSERT((l->l_pflag & LP_INTR) != 0);
    538      1.5  ad 	si->si_evcnt_block.ev_count++;
    539      1.5  ad }
    540      1.5  ad 
    541      1.5  ad /*
    542      1.5  ad  * schednetisr:
    543      1.5  ad  *
    544      1.5  ad  *	Trigger a legacy network interrupt.  XXX Needs to go away.
    545      1.5  ad  */
    546      1.5  ad void
    547      1.5  ad schednetisr(int isr)
    548      1.5  ad {
    549      1.5  ad 
    550      1.5  ad 	softint_schedule(softint_netisrs[isr]);
    551      1.5  ad }
    552      1.5  ad 
    553      1.5  ad #ifndef __HAVE_FAST_SOFTINTS
    554      1.5  ad 
    555      1.5  ad /*
    556      1.5  ad  * softint_init_md:
    557      1.5  ad  *
    558      1.5  ad  *	Slow path: perform machine-dependent initialization.
    559      1.5  ad  */
    560      1.5  ad void
    561      1.5  ad softint_init_md(lwp_t *l, u_int level, uintptr_t *machdep)
    562      1.5  ad {
    563      1.5  ad 	softint_t *si;
    564      1.5  ad 
    565      1.5  ad 	*machdep = (1 << level);
    566      1.5  ad 	si = l->l_private;
    567      1.5  ad 
    568      1.5  ad 	lwp_lock(l);
    569      1.5  ad 	lwp_unlock_to(l, l->l_cpu->ci_schedstate.spc_mutex);
    570      1.5  ad 	lwp_lock(l);
    571      1.5  ad 	/* Cheat and make the KASSERT in softint_thread() happy. */
    572      1.5  ad 	si->si_active = 1;
    573      1.5  ad 	l->l_stat = LSRUN;
    574      1.5  ad 	sched_enqueue(l, false);
    575      1.5  ad 	lwp_unlock(l);
    576      1.5  ad }
    577      1.5  ad 
    578      1.5  ad /*
    579      1.5  ad  * softint_trigger:
    580      1.5  ad  *
    581      1.5  ad  *	Slow path: cause a soft interrupt handler to begin executing.
    582      1.5  ad  *	Called at IPL_HIGH.
    583      1.5  ad  */
    584      1.5  ad void
    585      1.5  ad softint_trigger(uintptr_t machdep)
    586      1.5  ad {
    587      1.5  ad 	struct cpu_info *ci;
    588      1.5  ad 	lwp_t *l;
    589      1.2  ad 
    590      1.5  ad 	l = curlwp;
    591      1.5  ad 	ci = l->l_cpu;
    592      1.5  ad 	ci->ci_data.cpu_softints |= machdep;
    593      1.5  ad 	if (l == ci->ci_data.cpu_idlelwp) {
    594      1.5  ad 		cpu_need_resched(ci, 0);
    595      1.5  ad 	} else {
    596      1.5  ad 		/* MI equivalent of aston() */
    597      1.5  ad 		cpu_signotify(l);
    598      1.5  ad 	}
    599      1.5  ad }
    600      1.5  ad 
    601      1.5  ad /*
    602      1.5  ad  * softint_thread:
    603      1.5  ad  *
    604      1.5  ad  *	Slow path: MI software interrupt dispatch.
    605      1.5  ad  */
    606      1.5  ad void
    607      1.5  ad softint_thread(void *cookie)
    608      1.5  ad {
    609      1.5  ad 	softint_t *si;
    610      1.5  ad 	lwp_t *l;
    611      1.5  ad 	int s;
    612      1.5  ad 
    613      1.5  ad 	l = curlwp;
    614      1.5  ad 	si = l->l_private;
    615      1.5  ad 
    616      1.5  ad 	for (;;) {
    617      1.5  ad 		/*
    618      1.5  ad 		 * Clear pending status and run it.  We must drop the
    619      1.5  ad 		 * spl before mi_switch(), since IPL_HIGH may be higher
    620      1.5  ad 		 * than IPL_SCHED (and it is not safe to switch at a
    621      1.5  ad 		 * higher level).
    622      1.5  ad 		 */
    623      1.5  ad 		s = splhigh();
    624      1.5  ad 		l->l_cpu->ci_data.cpu_softints &= ~si->si_machdep;
    625      1.5  ad 		softint_execute(si, l, s);
    626      1.5  ad 		splx(s);
    627      1.5  ad 
    628      1.5  ad 		lwp_lock(l);
    629      1.5  ad 		l->l_stat = LSIDL;
    630      1.5  ad 		mi_switch(l);
    631      1.5  ad 	}
    632      1.2  ad }
    633      1.4  ad 
    634      1.4  ad /*
    635      1.4  ad  * softint_picklwp:
    636      1.4  ad  *
    637      1.4  ad  *	Slow path: called from mi_switch() to pick the highest priority
    638      1.4  ad  *	soft interrupt LWP that needs to run.
    639      1.4  ad  */
    640      1.4  ad lwp_t *
    641      1.4  ad softint_picklwp(void)
    642      1.4  ad {
    643      1.5  ad 	struct cpu_info *ci;
    644      1.5  ad 	u_int mask;
    645      1.5  ad 	softint_t *si;
    646      1.5  ad 	lwp_t *l;
    647      1.5  ad 
    648      1.5  ad 	ci = curcpu();
    649      1.5  ad 	si = ((softcpu_t *)ci->ci_data.cpu_softcpu)->sc_int;
    650      1.5  ad 	mask = ci->ci_data.cpu_softints;
    651      1.5  ad 
    652      1.5  ad 	if ((mask & (1 << SOFTINT_SERIAL)) != 0) {
    653      1.5  ad 		l = si[SOFTINT_SERIAL].si_lwp;
    654      1.5  ad 	} else if ((mask & (1 << SOFTINT_NET)) != 0) {
    655      1.5  ad 		l = si[SOFTINT_NET].si_lwp;
    656      1.5  ad 	} else if ((mask & (1 << SOFTINT_BIO)) != 0) {
    657      1.5  ad 		l = si[SOFTINT_BIO].si_lwp;
    658      1.5  ad 	} else if ((mask & (1 << SOFTINT_CLOCK)) != 0) {
    659      1.5  ad 		l = si[SOFTINT_CLOCK].si_lwp;
    660      1.5  ad 	} else {
    661      1.5  ad 		panic("softint_picklwp");
    662      1.5  ad 	}
    663      1.4  ad 
    664      1.5  ad 	return l;
    665      1.4  ad }
    666      1.4  ad 
    667      1.4  ad /*
    668      1.4  ad  * softint_overlay:
    669      1.4  ad  *
    670      1.4  ad  *	Slow path: called from lwp_userret() to run a soft interrupt
    671      1.6  ad  *	within the context of a user thread.
    672      1.4  ad  */
    673      1.4  ad void
    674      1.4  ad softint_overlay(void)
    675      1.4  ad {
    676      1.5  ad 	struct cpu_info *ci;
    677      1.5  ad 	u_int softints;
    678      1.5  ad 	softint_t *si;
    679      1.6  ad 	pri_t obase;
    680      1.5  ad 	lwp_t *l;
    681      1.5  ad 	int s;
    682      1.5  ad 
    683      1.5  ad 	l = curlwp;
    684      1.5  ad 	ci = l->l_cpu;
    685      1.5  ad 	si = ((softcpu_t *)ci->ci_data.cpu_softcpu)->sc_int;
    686      1.5  ad 
    687      1.5  ad 	KASSERT((l->l_pflag & LP_INTR) == 0);
    688      1.5  ad 
    689      1.6  ad 	/* Arrange to elevate priority if the LWP blocks. */
    690      1.6  ad 	obase = l->l_kpribase;
    691      1.6  ad 	l->l_kpribase = PRI_KERNEL_RT;
    692      1.5  ad 	l->l_pflag |= LP_INTR;
    693      1.5  ad 	s = splhigh();
    694      1.5  ad 	while ((softints = ci->ci_data.cpu_softints) != 0) {
    695      1.5  ad 		if ((softints & (1 << SOFTINT_SERIAL)) != 0) {
    696      1.5  ad 			ci->ci_data.cpu_softints &= ~(1 << SOFTINT_SERIAL);
    697      1.5  ad 			softint_execute(&si[SOFTINT_SERIAL], l, s);
    698      1.5  ad 			continue;
    699      1.5  ad 		}
    700      1.5  ad 		if ((softints & (1 << SOFTINT_NET)) != 0) {
    701      1.5  ad 			ci->ci_data.cpu_softints &= ~(1 << SOFTINT_NET);
    702      1.5  ad 			softint_execute(&si[SOFTINT_NET], l, s);
    703      1.5  ad 			continue;
    704      1.5  ad 		}
    705      1.5  ad 		if ((softints & (1 << SOFTINT_BIO)) != 0) {
    706      1.5  ad 			ci->ci_data.cpu_softints &= ~(1 << SOFTINT_BIO);
    707      1.5  ad 			softint_execute(&si[SOFTINT_BIO], l, s);
    708      1.5  ad 			continue;
    709      1.5  ad 		}
    710      1.5  ad 		if ((softints & (1 << SOFTINT_CLOCK)) != 0) {
    711      1.5  ad 			ci->ci_data.cpu_softints &= ~(1 << SOFTINT_CLOCK);
    712      1.5  ad 			softint_execute(&si[SOFTINT_CLOCK], l, s);
    713      1.5  ad 			continue;
    714      1.5  ad 		}
    715      1.5  ad 	}
    716      1.5  ad 	splx(s);
    717      1.5  ad 	l->l_pflag &= ~LP_INTR;
    718      1.6  ad 	l->l_kpribase = obase;
    719      1.4  ad }
    720      1.5  ad 
    721      1.5  ad #else	/*  !__HAVE_FAST_SOFTINTS */
    722      1.5  ad 
    723      1.5  ad /*
    724      1.5  ad  * softint_thread:
    725      1.5  ad  *
    726      1.5  ad  *	Fast path: the LWP is switched to without restoring any state,
    727      1.5  ad  *	so we should not arrive here - there is a direct handoff between
    728      1.5  ad  *	the interrupt stub and softint_dispatch().
    729      1.5  ad  */
    730      1.5  ad void
    731      1.5  ad softint_thread(void *cookie)
    732      1.5  ad {
    733      1.5  ad 
    734      1.5  ad 	panic("softint_thread");
    735      1.5  ad }
    736      1.5  ad 
    737      1.5  ad /*
    738      1.5  ad  * softint_dispatch:
    739      1.5  ad  *
    740      1.5  ad  *	Fast path: entry point from machine-dependent code.
    741      1.5  ad  */
    742      1.5  ad void
    743      1.5  ad softint_dispatch(lwp_t *pinned, int s)
    744      1.5  ad {
    745  1.6.2.1  ad 	struct bintime now;
    746      1.5  ad 	softint_t *si;
    747      1.5  ad 	u_int timing;
    748      1.5  ad 	lwp_t *l;
    749      1.5  ad 
    750      1.5  ad 	l = curlwp;
    751      1.5  ad 	si = l->l_private;
    752      1.5  ad 
    753      1.5  ad 	/*
    754      1.5  ad 	 * Note the interrupted LWP, and mark the current LWP as running
    755      1.5  ad 	 * before proceeding.  Although this must as a rule be done with
    756      1.5  ad 	 * the LWP locked, at this point no external agents will want to
    757      1.5  ad 	 * modify the interrupt LWP's state.
    758      1.5  ad 	 */
    759      1.5  ad 	timing = (softint_timing ? LW_TIMEINTR : 0);
    760      1.5  ad 	l->l_switchto = pinned;
    761      1.5  ad 	l->l_stat = LSONPROC;
    762      1.5  ad 	l->l_flag |= (LW_RUNNING | timing);
    763      1.5  ad 
    764      1.5  ad 	/*
    765      1.5  ad 	 * Dispatch the interrupt.  If softints are being timed, charge
    766      1.5  ad 	 * for it.
    767      1.5  ad 	 */
    768      1.5  ad 	if (timing)
    769  1.6.2.1  ad 		bintime(&l->l_stime);
    770      1.5  ad 	softint_execute(si, l, s);
    771      1.5  ad 	if (timing) {
    772  1.6.2.1  ad 		bintime(&now);
    773      1.5  ad 		updatertime(l, &now);
    774      1.5  ad 		l->l_flag &= ~LW_TIMEINTR;
    775      1.5  ad 	}
    776      1.5  ad 
    777      1.5  ad 	/*
    778      1.5  ad 	 * If we blocked while handling the interrupt, the pinned LWP is
    779      1.5  ad 	 * gone so switch to the idle LWP.  It will select a new LWP to
    780      1.5  ad 	 * run.
    781      1.5  ad 	 *
    782      1.5  ad 	 * We must drop the priority level as switching at IPL_HIGH could
    783      1.5  ad 	 * deadlock the system.  We have already set si->si_active = 0,
    784      1.5  ad 	 * which means another interrupt at this level can be triggered.
    785      1.5  ad 	 * That's not be a problem: we are lowering to level 's' which will
    786      1.5  ad 	 * prevent softint_dispatch() from being reentered at level 's',
    787      1.5  ad 	 * until the priority is finally dropped to IPL_NONE on entry to
    788      1.5  ad 	 * the idle loop.
    789      1.5  ad 	 */
    790      1.5  ad 	l->l_stat = LSIDL;
    791      1.5  ad 	if (l->l_switchto == NULL) {
    792      1.5  ad 		splx(s);
    793      1.5  ad 		pmap_deactivate(l);
    794      1.5  ad 		lwp_exit_switchaway(l);
    795      1.5  ad 		/* NOTREACHED */
    796      1.5  ad 	}
    797      1.5  ad 	l->l_switchto = NULL;
    798      1.5  ad 	l->l_flag &= ~LW_RUNNING;
    799      1.5  ad }
    800      1.5  ad 
    801      1.5  ad #endif	/* !__HAVE_FAST_SOFTINTS */
    802