Home | History | Annotate | Line # | Download | only in kern
kern_softint.c revision 1.10
      1  1.10    ad /*	$NetBSD: kern_softint.c,v 1.10 2008/01/29 18:06:14 ad Exp $	*/
      2   1.2    ad 
      3   1.2    ad /*-
      4  1.10    ad  * Copyright (c) 2007, 2008 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  1.10    ad  *	to sleep on condition variables	or wait for resources to become
     97  1.10    ad  *	available (for example,	memory).
     98   1.5    ad  *
     99   1.5    ad  * Per-CPU operation
    100   1.5    ad  *
    101   1.5    ad  *	If a soft interrupt is triggered on a CPU, it can only be
    102   1.5    ad  *	dispatched on the same CPU.  Each LWP dedicated to handling a
    103   1.5    ad  *	soft interrupt is bound to its home CPU, so if the LWP blocks
    104   1.5    ad  *	and needs to run again, it can only run there.  Nearly all data
    105   1.5    ad  *	structures used to manage software interrupts are per-CPU.
    106   1.5    ad  *
    107   1.5    ad  *	The per-CPU requirement is intended to reduce "ping-pong" of
    108   1.5    ad  *	cache lines between CPUs: lines occupied by data structures
    109   1.5    ad  *	used to manage the soft interrupts, and lines occupied by data
    110   1.5    ad  *	items being passed down to the soft interrupt.  As a positive
    111   1.5    ad  *	side effect, this also means that the soft interrupt dispatch
    112   1.5    ad  *	code does not need to to use spinlocks to synchronize.
    113   1.5    ad  *
    114   1.5    ad  * Generic implementation
    115   1.5    ad  *
    116   1.5    ad  *	A generic, low performance implementation is provided that
    117   1.5    ad  *	works across all architectures, with no machine-dependent
    118   1.5    ad  *	modifications needed.  This implementation uses the scheduler,
    119   1.5    ad  *	and so has a number of restrictions:
    120   1.5    ad  *
    121   1.5    ad  *	1) The software interrupts are not currently preemptive, so
    122   1.5    ad  *	must wait for the currently executing LWP to yield the CPU.
    123   1.5    ad  *	This can introduce latency.
    124   1.5    ad  *
    125   1.5    ad  *	2) An expensive context switch is required for a software
    126   1.5    ad  *	interrupt to be handled.
    127   1.5    ad  *
    128   1.5    ad  * 'Fast' software interrupts
    129   1.5    ad  *
    130   1.5    ad  *	If an architectures defines __HAVE_FAST_SOFTINTS, it implements
    131   1.5    ad  *	the fast mechanism.  Threads running either in the kernel or in
    132   1.5    ad  *	userspace will be interrupted, but will not be preempted.  When
    133   1.5    ad  *	the soft interrupt completes execution, the interrupted LWP
    134   1.5    ad  *	is resumed.  Interrupt dispatch code must provide the minimum
    135   1.5    ad  *	level of context necessary for the soft interrupt to block and
    136   1.5    ad  *	be resumed at a later time.  The machine-dependent dispatch
    137   1.5    ad  *	path looks something like the following:
    138   1.5    ad  *
    139   1.5    ad  *	softintr()
    140   1.5    ad  *	{
    141   1.5    ad  *		go to IPL_HIGH if necessary for switch;
    142   1.5    ad  *		save any necessary registers in a format that can be
    143   1.5    ad  *		    restored by cpu_switchto if the softint blocks;
    144   1.5    ad  *		arrange for cpu_switchto() to restore into the
    145   1.5    ad  *		    trampoline function;
    146   1.5    ad  *		identify LWP to handle this interrupt;
    147   1.5    ad  *		switch to the LWP's stack;
    148   1.5    ad  *		switch register stacks, if necessary;
    149   1.5    ad  *		assign new value of curlwp;
    150   1.5    ad  *		call MI softint_dispatch, passing old curlwp and IPL
    151   1.5    ad  *		    to execute interrupt at;
    152   1.5    ad  *		switch back to old stack;
    153   1.5    ad  *		switch back to old register stack, if necessary;
    154   1.5    ad  *		restore curlwp;
    155   1.5    ad  *		return to interrupted LWP;
    156   1.5    ad  *	}
    157   1.5    ad  *
    158   1.5    ad  *	If the soft interrupt blocks, a trampoline function is returned
    159   1.5    ad  *	to in the context of the interrupted LWP, as arranged for by
    160   1.5    ad  *	softint():
    161   1.5    ad  *
    162   1.5    ad  *	softint_ret()
    163   1.5    ad  *	{
    164   1.5    ad  *		unlock soft interrupt LWP;
    165   1.5    ad  *		resume interrupt processing, likely returning to
    166   1.5    ad  *		    interrupted LWP or dispatching another, different
    167   1.5    ad  *		    interrupt;
    168   1.5    ad  *	}
    169   1.5    ad  *
    170   1.5    ad  *	Once the soft interrupt has fired (and even if it has blocked),
    171   1.5    ad  *	no further soft interrupts at that level will be triggered by
    172   1.5    ad  *	MI code until the soft interrupt handler has ceased execution.
    173   1.5    ad  *	If a soft interrupt handler blocks and is resumed, it resumes
    174   1.5    ad  *	execution as a normal LWP (kthread) and gains VM context.  Only
    175   1.5    ad  *	when it has completed and is ready to fire again will it
    176   1.5    ad  *	interrupt other threads.
    177   1.5    ad  *
    178   1.5    ad  * Future directions
    179   1.5    ad  *
    180   1.5    ad  *	Provide a cheap way to direct software interrupts to remote
    181   1.5    ad  *	CPUs.  Provide a way to enqueue work items into the handler
    182   1.5    ad  *	record,	removing additional spl calls (see subr_workqueue.c).
    183   1.2    ad  */
    184   1.2    ad 
    185   1.2    ad #include <sys/cdefs.h>
    186  1.10    ad __KERNEL_RCSID(0, "$NetBSD: kern_softint.c,v 1.10 2008/01/29 18:06:14 ad Exp $");
    187   1.2    ad 
    188   1.2    ad #include <sys/param.h>
    189   1.5    ad #include <sys/malloc.h>
    190   1.5    ad #include <sys/proc.h>
    191   1.2    ad #include <sys/intr.h>
    192   1.5    ad #include <sys/mutex.h>
    193   1.5    ad #include <sys/kthread.h>
    194   1.5    ad #include <sys/evcnt.h>
    195   1.5    ad #include <sys/cpu.h>
    196   1.5    ad 
    197   1.5    ad #include <net/netisr.h>
    198   1.5    ad 
    199   1.5    ad #include <uvm/uvm_extern.h>
    200   1.5    ad 
    201   1.5    ad /* This could overlap with signal info in struct lwp. */
    202   1.5    ad typedef struct softint {
    203   1.5    ad 	SIMPLEQ_HEAD(, softhand) si_q;
    204   1.5    ad 	struct lwp		*si_lwp;
    205   1.5    ad 	struct cpu_info		*si_cpu;
    206   1.5    ad 	uintptr_t		si_machdep;
    207   1.5    ad 	struct evcnt		si_evcnt;
    208   1.5    ad 	struct evcnt		si_evcnt_block;
    209   1.5    ad 	int			si_active;
    210   1.5    ad 	char			si_name[8];
    211   1.5    ad 	char			si_name_block[8+6];
    212   1.5    ad } softint_t;
    213   1.5    ad 
    214   1.5    ad typedef struct softhand {
    215   1.5    ad 	SIMPLEQ_ENTRY(softhand)	sh_q;
    216   1.5    ad 	void			(*sh_func)(void *);
    217   1.5    ad 	void			*sh_arg;
    218   1.5    ad 	softint_t		*sh_isr;
    219   1.5    ad 	u_int			sh_pending;
    220   1.5    ad 	u_int			sh_flags;
    221   1.5    ad } softhand_t;
    222   1.5    ad 
    223   1.5    ad typedef struct softcpu {
    224   1.5    ad 	struct cpu_info		*sc_cpu;
    225   1.5    ad 	softint_t		sc_int[SOFTINT_COUNT];
    226   1.5    ad 	softhand_t		sc_hand[1];
    227   1.5    ad } softcpu_t;
    228   1.5    ad 
    229   1.5    ad static void	softint_thread(void *);
    230   1.5    ad 
    231   1.5    ad u_int		softint_bytes = 8192;
    232   1.5    ad u_int		softint_timing;
    233   1.5    ad static u_int	softint_max;
    234   1.5    ad static kmutex_t	softint_lock;
    235   1.5    ad static void	*softint_netisrs[32];
    236   1.2    ad 
    237   1.5    ad /*
    238   1.5    ad  * softint_init_isr:
    239   1.5    ad  *
    240   1.5    ad  *	Initialize a single interrupt level for a single CPU.
    241   1.5    ad  */
    242   1.5    ad static void
    243   1.5    ad softint_init_isr(softcpu_t *sc, const char *desc, pri_t pri, u_int level)
    244   1.5    ad {
    245   1.5    ad 	struct cpu_info *ci;
    246   1.5    ad 	softint_t *si;
    247   1.5    ad 	int error;
    248   1.5    ad 
    249   1.5    ad 	si = &sc->sc_int[level];
    250   1.5    ad 	ci = sc->sc_cpu;
    251   1.5    ad 	si->si_cpu = ci;
    252   1.5    ad 
    253   1.5    ad 	SIMPLEQ_INIT(&si->si_q);
    254   1.5    ad 
    255   1.5    ad 	error = kthread_create(pri, KTHREAD_MPSAFE | KTHREAD_INTR |
    256   1.5    ad 	    KTHREAD_IDLE, ci, softint_thread, si, &si->si_lwp,
    257   1.5    ad 	    "soft%s/%d", desc, (int)ci->ci_cpuid);
    258   1.5    ad 	if (error != 0)
    259   1.5    ad 		panic("softint_init_isr: error %d", error);
    260   1.5    ad 
    261   1.5    ad 	snprintf(si->si_name, sizeof(si->si_name), "%s/%d", desc,
    262   1.5    ad 	    (int)ci->ci_cpuid);
    263   1.5    ad 	evcnt_attach_dynamic(&si->si_evcnt, EVCNT_TYPE_INTR, NULL,
    264   1.5    ad 	   "softint", si->si_name);
    265   1.5    ad 	snprintf(si->si_name_block, sizeof(si->si_name_block), "%s block/%d",
    266   1.5    ad 	    desc, (int)ci->ci_cpuid);
    267   1.5    ad 	evcnt_attach_dynamic(&si->si_evcnt_block, EVCNT_TYPE_INTR, NULL,
    268   1.5    ad 	   "softint", si->si_name_block);
    269   1.3    ad 
    270   1.5    ad 	si->si_lwp->l_private = si;
    271   1.5    ad 	softint_init_md(si->si_lwp, level, &si->si_machdep);
    272   1.5    ad }
    273   1.2    ad /*
    274   1.2    ad  * softint_init:
    275   1.2    ad  *
    276   1.2    ad  *	Initialize per-CPU data structures.  Called from mi_cpu_attach().
    277   1.2    ad  */
    278   1.2    ad void
    279   1.2    ad softint_init(struct cpu_info *ci)
    280   1.2    ad {
    281   1.5    ad 	static struct cpu_info *first;
    282   1.5    ad 	softcpu_t *sc, *scfirst;
    283   1.5    ad 	softhand_t *sh, *shmax;
    284   1.5    ad 
    285   1.5    ad 	if (first == NULL) {
    286   1.5    ad 		/* Boot CPU. */
    287   1.5    ad 		first = ci;
    288   1.5    ad 		mutex_init(&softint_lock, MUTEX_DEFAULT, IPL_NONE);
    289   1.5    ad 		softint_bytes = round_page(softint_bytes);
    290   1.5    ad 		softint_max = (softint_bytes - sizeof(softcpu_t)) /
    291   1.5    ad 		    sizeof(softhand_t);
    292   1.5    ad 	}
    293   1.2    ad 
    294   1.5    ad 	sc = (softcpu_t *)uvm_km_alloc(kernel_map, softint_bytes, 0,
    295   1.5    ad 	    UVM_KMF_WIRED | UVM_KMF_ZERO);
    296   1.5    ad 	if (sc == NULL)
    297   1.5    ad 		panic("softint_init_cpu: cannot allocate memory");
    298   1.5    ad 
    299   1.5    ad 	ci->ci_data.cpu_softcpu = sc;
    300   1.5    ad 	ci->ci_data.cpu_softints = 0;
    301   1.5    ad 	sc->sc_cpu = ci;
    302   1.5    ad 
    303   1.5    ad 	softint_init_isr(sc, "net", PRI_SOFTNET, SOFTINT_NET);
    304   1.5    ad 	softint_init_isr(sc, "bio", PRI_SOFTBIO, SOFTINT_BIO);
    305   1.5    ad 	softint_init_isr(sc, "clk", PRI_SOFTCLOCK, SOFTINT_CLOCK);
    306   1.5    ad 	softint_init_isr(sc, "ser", PRI_SOFTSERIAL, SOFTINT_SERIAL);
    307   1.5    ad 
    308   1.5    ad 	if (first != ci) {
    309   1.5    ad 		mutex_enter(&softint_lock);
    310   1.5    ad 		scfirst = first->ci_data.cpu_softcpu;
    311   1.5    ad 		sh = sc->sc_hand;
    312   1.5    ad 		memcpy(sh, scfirst->sc_hand, sizeof(*sh) * softint_max);
    313   1.5    ad 		/* Update pointers for this CPU. */
    314   1.5    ad 		for (shmax = sh + softint_max; sh < shmax; sh++) {
    315   1.5    ad 			if (sh->sh_func == NULL)
    316   1.5    ad 				continue;
    317   1.5    ad 			sh->sh_isr =
    318   1.5    ad 			    &sc->sc_int[sh->sh_flags & SOFTINT_LVLMASK];
    319   1.5    ad 		}
    320   1.5    ad 		mutex_exit(&softint_lock);
    321   1.5    ad 	} else {
    322   1.5    ad 		/*
    323   1.5    ad 		 * Establish handlers for legacy net interrupts.
    324   1.5    ad 		 * XXX Needs to go away.
    325   1.5    ad 		 */
    326   1.5    ad #define DONETISR(n, f)							\
    327   1.5    ad     softint_netisrs[(n)] = 						\
    328   1.5    ad         softint_establish(SOFTINT_NET, (void (*)(void *))(f), NULL)
    329   1.5    ad #include <net/netisr_dispatch.h>
    330   1.5    ad 	}
    331   1.2    ad }
    332   1.2    ad 
    333   1.2    ad /*
    334   1.2    ad  * softint_establish:
    335   1.2    ad  *
    336   1.2    ad  *	Register a software interrupt handler.
    337   1.2    ad  */
    338   1.2    ad void *
    339   1.2    ad softint_establish(u_int flags, void (*func)(void *), void *arg)
    340   1.2    ad {
    341   1.5    ad 	CPU_INFO_ITERATOR cii;
    342   1.5    ad 	struct cpu_info *ci;
    343   1.5    ad 	softcpu_t *sc;
    344   1.5    ad 	softhand_t *sh;
    345   1.5    ad 	u_int level, index;
    346   1.2    ad 
    347   1.2    ad 	level = (flags & SOFTINT_LVLMASK);
    348   1.2    ad 	KASSERT(level < SOFTINT_COUNT);
    349   1.2    ad 
    350   1.5    ad 	mutex_enter(&softint_lock);
    351   1.5    ad 
    352   1.5    ad 	/* Find a free slot. */
    353   1.5    ad 	sc = curcpu()->ci_data.cpu_softcpu;
    354   1.5    ad 	for (index = 1; index < softint_max; index++)
    355   1.5    ad 		if (sc->sc_hand[index].sh_func == NULL)
    356   1.5    ad 			break;
    357   1.5    ad 	if (index == softint_max) {
    358   1.5    ad 		mutex_exit(&softint_lock);
    359   1.5    ad 		printf("WARNING: softint_establish: table full, "
    360   1.5    ad 		    "increase softint_bytes\n");
    361   1.5    ad 		return NULL;
    362   1.5    ad 	}
    363   1.5    ad 
    364   1.5    ad 	/* Set up the handler on each CPU. */
    365   1.8    ad 	if (ncpu < 2) {
    366   1.7    ad 		/* XXX hack for machines with no CPU_INFO_FOREACH() early on */
    367   1.7    ad 		sc = curcpu()->ci_data.cpu_softcpu;
    368   1.7    ad 		sh = &sc->sc_hand[index];
    369   1.7    ad 		sh->sh_isr = &sc->sc_int[level];
    370   1.7    ad 		sh->sh_func = func;
    371   1.7    ad 		sh->sh_arg = arg;
    372   1.7    ad 		sh->sh_flags = flags;
    373   1.7    ad 		sh->sh_pending = 0;
    374   1.7    ad 	} else for (CPU_INFO_FOREACH(cii, ci)) {
    375   1.5    ad 		sc = ci->ci_data.cpu_softcpu;
    376   1.5    ad 		sh = &sc->sc_hand[index];
    377   1.5    ad 		sh->sh_isr = &sc->sc_int[level];
    378   1.5    ad 		sh->sh_func = func;
    379   1.5    ad 		sh->sh_arg = arg;
    380   1.5    ad 		sh->sh_flags = flags;
    381   1.5    ad 		sh->sh_pending = 0;
    382   1.2    ad 	}
    383   1.2    ad 
    384   1.5    ad 	mutex_exit(&softint_lock);
    385   1.5    ad 
    386   1.5    ad 	return (void *)((uint8_t *)&sc->sc_hand[index] - (uint8_t *)sc);
    387   1.2    ad }
    388   1.2    ad 
    389   1.2    ad /*
    390   1.2    ad  * softint_disestablish:
    391   1.2    ad  *
    392   1.2    ad  *	Unregister a software interrupt handler.
    393   1.2    ad  */
    394   1.2    ad void
    395   1.2    ad softint_disestablish(void *arg)
    396   1.2    ad {
    397   1.5    ad 	CPU_INFO_ITERATOR cii;
    398   1.5    ad 	struct cpu_info *ci;
    399   1.5    ad 	softcpu_t *sc;
    400   1.5    ad 	softhand_t *sh;
    401   1.5    ad 	uintptr_t offset;
    402   1.5    ad 
    403   1.5    ad 	offset = (uintptr_t)arg;
    404   1.5    ad 	KASSERT(offset != 0 && offset < softint_bytes);
    405   1.5    ad 
    406   1.5    ad 	mutex_enter(&softint_lock);
    407   1.5    ad 
    408   1.5    ad 	/* Clear the handler on each CPU. */
    409   1.5    ad 	for (CPU_INFO_FOREACH(cii, ci)) {
    410   1.5    ad 		sc = ci->ci_data.cpu_softcpu;
    411   1.5    ad 		sh = (softhand_t *)((uint8_t *)sc + offset);
    412   1.5    ad 		KASSERT(sh->sh_func != NULL);
    413   1.5    ad 		KASSERT(sh->sh_pending == 0);
    414   1.5    ad 		sh->sh_func = NULL;
    415   1.5    ad 	}
    416   1.2    ad 
    417   1.5    ad 	mutex_exit(&softint_lock);
    418   1.2    ad }
    419   1.2    ad 
    420   1.2    ad /*
    421   1.2    ad  * softint_schedule:
    422   1.2    ad  *
    423   1.2    ad  *	Trigger a software interrupt.  Must be called from a hardware
    424   1.2    ad  *	interrupt handler, or with preemption disabled (since we are
    425   1.2    ad  *	using the value of curcpu()).
    426   1.2    ad  */
    427   1.2    ad void
    428   1.2    ad softint_schedule(void *arg)
    429   1.2    ad {
    430   1.5    ad 	softhand_t *sh;
    431   1.5    ad 	softint_t *si;
    432   1.5    ad 	uintptr_t offset;
    433   1.5    ad 	int s;
    434   1.5    ad 
    435   1.5    ad 	/* Find the handler record for this CPU. */
    436   1.5    ad 	offset = (uintptr_t)arg;
    437   1.5    ad 	KASSERT(offset != 0 && offset < softint_bytes);
    438   1.5    ad 	sh = (softhand_t *)((uint8_t *)curcpu()->ci_data.cpu_softcpu + offset);
    439   1.5    ad 
    440   1.5    ad 	/* If it's already pending there's nothing to do. */
    441   1.5    ad 	if (sh->sh_pending)
    442   1.5    ad 		return;
    443   1.5    ad 
    444   1.5    ad 	/*
    445   1.5    ad 	 * Enqueue the handler into the LWP's pending list.
    446   1.5    ad 	 * If the LWP is completely idle, then make it run.
    447   1.5    ad 	 */
    448   1.5    ad 	s = splhigh();
    449   1.5    ad 	if (!sh->sh_pending) {
    450   1.5    ad 		si = sh->sh_isr;
    451   1.5    ad 		sh->sh_pending = 1;
    452   1.5    ad 		SIMPLEQ_INSERT_TAIL(&si->si_q, sh, sh_q);
    453   1.5    ad 		if (si->si_active == 0) {
    454   1.5    ad 			si->si_active = 1;
    455   1.5    ad 			softint_trigger(si->si_machdep);
    456   1.5    ad 		}
    457   1.5    ad 	}
    458   1.5    ad 	splx(s);
    459   1.5    ad }
    460   1.5    ad 
    461   1.5    ad /*
    462   1.5    ad  * softint_execute:
    463   1.5    ad  *
    464   1.5    ad  *	Invoke handlers for the specified soft interrupt.
    465   1.5    ad  *	Must be entered at splhigh.  Will drop the priority
    466   1.5    ad  *	to the level specified, but returns back at splhigh.
    467   1.5    ad  */
    468   1.5    ad static inline void
    469   1.5    ad softint_execute(softint_t *si, lwp_t *l, int s)
    470   1.5    ad {
    471   1.5    ad 	softhand_t *sh;
    472   1.5    ad 	bool havelock;
    473   1.5    ad 
    474   1.5    ad #ifdef __HAVE_FAST_SOFTINTS
    475   1.5    ad 	KASSERT(si->si_lwp == curlwp);
    476   1.5    ad #else
    477   1.5    ad 	/* May be running in user context. */
    478   1.5    ad #endif
    479   1.5    ad 	KASSERT(si->si_cpu == curcpu());
    480   1.5    ad 	KASSERT(si->si_lwp->l_wchan == NULL);
    481   1.5    ad 	KASSERT(si->si_active);
    482   1.5    ad 
    483   1.5    ad 	havelock = false;
    484   1.5    ad 
    485   1.5    ad 	/*
    486   1.5    ad 	 * Note: due to priority inheritance we may have interrupted a
    487   1.5    ad 	 * higher priority LWP.  Since the soft interrupt must be quick
    488   1.5    ad 	 * and is non-preemptable, we don't bother yielding.
    489   1.5    ad 	 */
    490   1.5    ad 
    491   1.5    ad 	while (!SIMPLEQ_EMPTY(&si->si_q)) {
    492   1.5    ad 		/*
    493   1.5    ad 		 * Pick the longest waiting handler to run.  We block
    494   1.5    ad 		 * interrupts but do not lock in order to do this, as
    495   1.5    ad 		 * we are protecting against the local CPU only.
    496   1.5    ad 		 */
    497   1.5    ad 		sh = SIMPLEQ_FIRST(&si->si_q);
    498   1.5    ad 		SIMPLEQ_REMOVE_HEAD(&si->si_q, sh_q);
    499   1.5    ad 		sh->sh_pending = 0;
    500   1.5    ad 		splx(s);
    501   1.5    ad 
    502   1.5    ad 		/* Run the handler. */
    503   1.5    ad 		if ((sh->sh_flags & SOFTINT_MPSAFE) == 0 && !havelock) {
    504   1.5    ad 			KERNEL_LOCK(1, l);
    505   1.5    ad 			havelock = true;
    506   1.5    ad 		}
    507   1.5    ad 		(*sh->sh_func)(sh->sh_arg);
    508   1.5    ad 
    509   1.5    ad 		(void)splhigh();
    510   1.5    ad 	}
    511   1.2    ad 
    512   1.5    ad 	if (havelock) {
    513   1.5    ad 		KERNEL_UNLOCK_ONE(l);
    514   1.5    ad 	}
    515   1.5    ad 
    516   1.5    ad 	/*
    517   1.5    ad 	 * Unlocked, but only for statistics.
    518   1.5    ad 	 * Should be per-CPU to prevent cache ping-pong.
    519   1.5    ad 	 */
    520   1.5    ad 	uvmexp.softs++;
    521   1.5    ad 
    522   1.5    ad 	si->si_evcnt.ev_count++;
    523   1.5    ad 	si->si_active = 0;
    524   1.2    ad }
    525   1.2    ad 
    526   1.2    ad /*
    527   1.2    ad  * softint_block:
    528   1.2    ad  *
    529   1.2    ad  *	Update statistics when the soft interrupt blocks.
    530   1.2    ad  */
    531   1.2    ad void
    532   1.2    ad softint_block(lwp_t *l)
    533   1.2    ad {
    534   1.5    ad 	softint_t *si = l->l_private;
    535   1.5    ad 
    536   1.5    ad 	KASSERT((l->l_pflag & LP_INTR) != 0);
    537   1.5    ad 	si->si_evcnt_block.ev_count++;
    538   1.5    ad }
    539   1.5    ad 
    540   1.5    ad /*
    541   1.5    ad  * schednetisr:
    542   1.5    ad  *
    543   1.5    ad  *	Trigger a legacy network interrupt.  XXX Needs to go away.
    544   1.5    ad  */
    545   1.5    ad void
    546   1.5    ad schednetisr(int isr)
    547   1.5    ad {
    548   1.5    ad 
    549   1.5    ad 	softint_schedule(softint_netisrs[isr]);
    550   1.5    ad }
    551   1.5    ad 
    552   1.5    ad #ifndef __HAVE_FAST_SOFTINTS
    553   1.5    ad 
    554   1.5    ad /*
    555   1.5    ad  * softint_init_md:
    556   1.5    ad  *
    557   1.5    ad  *	Slow path: perform machine-dependent initialization.
    558   1.5    ad  */
    559   1.5    ad void
    560   1.5    ad softint_init_md(lwp_t *l, u_int level, uintptr_t *machdep)
    561   1.5    ad {
    562   1.5    ad 	softint_t *si;
    563   1.5    ad 
    564   1.5    ad 	*machdep = (1 << level);
    565   1.5    ad 	si = l->l_private;
    566   1.5    ad 
    567   1.5    ad 	lwp_lock(l);
    568   1.5    ad 	lwp_unlock_to(l, l->l_cpu->ci_schedstate.spc_mutex);
    569   1.5    ad 	lwp_lock(l);
    570   1.5    ad 	/* Cheat and make the KASSERT in softint_thread() happy. */
    571   1.5    ad 	si->si_active = 1;
    572   1.5    ad 	l->l_stat = LSRUN;
    573   1.5    ad 	sched_enqueue(l, false);
    574   1.5    ad 	lwp_unlock(l);
    575   1.5    ad }
    576   1.5    ad 
    577   1.5    ad /*
    578   1.5    ad  * softint_trigger:
    579   1.5    ad  *
    580   1.5    ad  *	Slow path: cause a soft interrupt handler to begin executing.
    581   1.5    ad  *	Called at IPL_HIGH.
    582   1.5    ad  */
    583   1.5    ad void
    584   1.5    ad softint_trigger(uintptr_t machdep)
    585   1.5    ad {
    586   1.5    ad 	struct cpu_info *ci;
    587   1.5    ad 	lwp_t *l;
    588   1.2    ad 
    589   1.5    ad 	l = curlwp;
    590   1.5    ad 	ci = l->l_cpu;
    591   1.5    ad 	ci->ci_data.cpu_softints |= machdep;
    592   1.5    ad 	if (l == ci->ci_data.cpu_idlelwp) {
    593   1.5    ad 		cpu_need_resched(ci, 0);
    594   1.5    ad 	} else {
    595   1.5    ad 		/* MI equivalent of aston() */
    596   1.5    ad 		cpu_signotify(l);
    597   1.5    ad 	}
    598   1.5    ad }
    599   1.5    ad 
    600   1.5    ad /*
    601   1.5    ad  * softint_thread:
    602   1.5    ad  *
    603   1.5    ad  *	Slow path: MI software interrupt dispatch.
    604   1.5    ad  */
    605   1.5    ad void
    606   1.5    ad softint_thread(void *cookie)
    607   1.5    ad {
    608   1.5    ad 	softint_t *si;
    609   1.5    ad 	lwp_t *l;
    610   1.5    ad 	int s;
    611   1.5    ad 
    612   1.5    ad 	l = curlwp;
    613   1.5    ad 	si = l->l_private;
    614   1.5    ad 
    615   1.5    ad 	for (;;) {
    616   1.5    ad 		/*
    617   1.5    ad 		 * Clear pending status and run it.  We must drop the
    618   1.5    ad 		 * spl before mi_switch(), since IPL_HIGH may be higher
    619   1.5    ad 		 * than IPL_SCHED (and it is not safe to switch at a
    620   1.5    ad 		 * higher level).
    621   1.5    ad 		 */
    622   1.5    ad 		s = splhigh();
    623   1.5    ad 		l->l_cpu->ci_data.cpu_softints &= ~si->si_machdep;
    624   1.5    ad 		softint_execute(si, l, s);
    625   1.5    ad 		splx(s);
    626   1.5    ad 
    627   1.5    ad 		lwp_lock(l);
    628   1.5    ad 		l->l_stat = LSIDL;
    629   1.5    ad 		mi_switch(l);
    630   1.5    ad 	}
    631   1.2    ad }
    632   1.4    ad 
    633   1.4    ad /*
    634   1.4    ad  * softint_picklwp:
    635   1.4    ad  *
    636   1.4    ad  *	Slow path: called from mi_switch() to pick the highest priority
    637   1.4    ad  *	soft interrupt LWP that needs to run.
    638   1.4    ad  */
    639   1.4    ad lwp_t *
    640   1.4    ad softint_picklwp(void)
    641   1.4    ad {
    642   1.5    ad 	struct cpu_info *ci;
    643   1.5    ad 	u_int mask;
    644   1.5    ad 	softint_t *si;
    645   1.5    ad 	lwp_t *l;
    646   1.5    ad 
    647   1.5    ad 	ci = curcpu();
    648   1.5    ad 	si = ((softcpu_t *)ci->ci_data.cpu_softcpu)->sc_int;
    649   1.5    ad 	mask = ci->ci_data.cpu_softints;
    650   1.5    ad 
    651   1.5    ad 	if ((mask & (1 << SOFTINT_SERIAL)) != 0) {
    652   1.5    ad 		l = si[SOFTINT_SERIAL].si_lwp;
    653   1.5    ad 	} else if ((mask & (1 << SOFTINT_NET)) != 0) {
    654   1.5    ad 		l = si[SOFTINT_NET].si_lwp;
    655   1.5    ad 	} else if ((mask & (1 << SOFTINT_BIO)) != 0) {
    656   1.5    ad 		l = si[SOFTINT_BIO].si_lwp;
    657   1.5    ad 	} else if ((mask & (1 << SOFTINT_CLOCK)) != 0) {
    658   1.5    ad 		l = si[SOFTINT_CLOCK].si_lwp;
    659   1.5    ad 	} else {
    660   1.5    ad 		panic("softint_picklwp");
    661   1.5    ad 	}
    662   1.4    ad 
    663   1.5    ad 	return l;
    664   1.4    ad }
    665   1.4    ad 
    666   1.4    ad /*
    667   1.4    ad  * softint_overlay:
    668   1.4    ad  *
    669   1.4    ad  *	Slow path: called from lwp_userret() to run a soft interrupt
    670   1.6    ad  *	within the context of a user thread.
    671   1.4    ad  */
    672   1.4    ad void
    673   1.4    ad softint_overlay(void)
    674   1.4    ad {
    675   1.5    ad 	struct cpu_info *ci;
    676   1.5    ad 	u_int softints;
    677   1.5    ad 	softint_t *si;
    678   1.6    ad 	pri_t obase;
    679   1.5    ad 	lwp_t *l;
    680   1.5    ad 	int s;
    681   1.5    ad 
    682   1.5    ad 	l = curlwp;
    683   1.5    ad 	ci = l->l_cpu;
    684   1.5    ad 	si = ((softcpu_t *)ci->ci_data.cpu_softcpu)->sc_int;
    685   1.5    ad 
    686   1.5    ad 	KASSERT((l->l_pflag & LP_INTR) == 0);
    687   1.5    ad 
    688   1.6    ad 	/* Arrange to elevate priority if the LWP blocks. */
    689   1.6    ad 	obase = l->l_kpribase;
    690   1.6    ad 	l->l_kpribase = PRI_KERNEL_RT;
    691   1.5    ad 	l->l_pflag |= LP_INTR;
    692   1.5    ad 	s = splhigh();
    693   1.5    ad 	while ((softints = ci->ci_data.cpu_softints) != 0) {
    694   1.5    ad 		if ((softints & (1 << SOFTINT_SERIAL)) != 0) {
    695   1.5    ad 			ci->ci_data.cpu_softints &= ~(1 << SOFTINT_SERIAL);
    696   1.5    ad 			softint_execute(&si[SOFTINT_SERIAL], l, s);
    697   1.5    ad 			continue;
    698   1.5    ad 		}
    699   1.5    ad 		if ((softints & (1 << SOFTINT_NET)) != 0) {
    700   1.5    ad 			ci->ci_data.cpu_softints &= ~(1 << SOFTINT_NET);
    701   1.5    ad 			softint_execute(&si[SOFTINT_NET], l, s);
    702   1.5    ad 			continue;
    703   1.5    ad 		}
    704   1.5    ad 		if ((softints & (1 << SOFTINT_BIO)) != 0) {
    705   1.5    ad 			ci->ci_data.cpu_softints &= ~(1 << SOFTINT_BIO);
    706   1.5    ad 			softint_execute(&si[SOFTINT_BIO], l, s);
    707   1.5    ad 			continue;
    708   1.5    ad 		}
    709   1.5    ad 		if ((softints & (1 << SOFTINT_CLOCK)) != 0) {
    710   1.5    ad 			ci->ci_data.cpu_softints &= ~(1 << SOFTINT_CLOCK);
    711   1.5    ad 			softint_execute(&si[SOFTINT_CLOCK], l, s);
    712   1.5    ad 			continue;
    713   1.5    ad 		}
    714   1.5    ad 	}
    715   1.5    ad 	splx(s);
    716   1.5    ad 	l->l_pflag &= ~LP_INTR;
    717   1.6    ad 	l->l_kpribase = obase;
    718   1.4    ad }
    719   1.5    ad 
    720   1.5    ad #else	/*  !__HAVE_FAST_SOFTINTS */
    721   1.5    ad 
    722   1.5    ad /*
    723   1.5    ad  * softint_thread:
    724   1.5    ad  *
    725   1.5    ad  *	Fast path: the LWP is switched to without restoring any state,
    726   1.5    ad  *	so we should not arrive here - there is a direct handoff between
    727   1.5    ad  *	the interrupt stub and softint_dispatch().
    728   1.5    ad  */
    729   1.5    ad void
    730   1.5    ad softint_thread(void *cookie)
    731   1.5    ad {
    732   1.5    ad 
    733   1.5    ad 	panic("softint_thread");
    734   1.5    ad }
    735   1.5    ad 
    736   1.5    ad /*
    737   1.5    ad  * softint_dispatch:
    738   1.5    ad  *
    739   1.5    ad  *	Fast path: entry point from machine-dependent code.
    740   1.5    ad  */
    741   1.5    ad void
    742   1.5    ad softint_dispatch(lwp_t *pinned, int s)
    743   1.5    ad {
    744   1.9  yamt 	struct bintime now;
    745   1.5    ad 	softint_t *si;
    746   1.5    ad 	u_int timing;
    747   1.5    ad 	lwp_t *l;
    748   1.5    ad 
    749   1.5    ad 	l = curlwp;
    750   1.5    ad 	si = l->l_private;
    751   1.5    ad 
    752   1.5    ad 	/*
    753   1.5    ad 	 * Note the interrupted LWP, and mark the current LWP as running
    754   1.5    ad 	 * before proceeding.  Although this must as a rule be done with
    755   1.5    ad 	 * the LWP locked, at this point no external agents will want to
    756   1.5    ad 	 * modify the interrupt LWP's state.
    757   1.5    ad 	 */
    758   1.5    ad 	timing = (softint_timing ? LW_TIMEINTR : 0);
    759   1.5    ad 	l->l_switchto = pinned;
    760   1.5    ad 	l->l_stat = LSONPROC;
    761   1.5    ad 	l->l_flag |= (LW_RUNNING | timing);
    762   1.5    ad 
    763   1.5    ad 	/*
    764   1.5    ad 	 * Dispatch the interrupt.  If softints are being timed, charge
    765   1.5    ad 	 * for it.
    766   1.5    ad 	 */
    767   1.5    ad 	if (timing)
    768   1.9  yamt 		bintime(&l->l_stime);
    769   1.5    ad 	softint_execute(si, l, s);
    770   1.5    ad 	if (timing) {
    771   1.9  yamt 		bintime(&now);
    772   1.5    ad 		updatertime(l, &now);
    773   1.5    ad 		l->l_flag &= ~LW_TIMEINTR;
    774   1.5    ad 	}
    775   1.5    ad 
    776   1.5    ad 	/*
    777   1.5    ad 	 * If we blocked while handling the interrupt, the pinned LWP is
    778   1.5    ad 	 * gone so switch to the idle LWP.  It will select a new LWP to
    779   1.5    ad 	 * run.
    780   1.5    ad 	 *
    781   1.5    ad 	 * We must drop the priority level as switching at IPL_HIGH could
    782   1.5    ad 	 * deadlock the system.  We have already set si->si_active = 0,
    783   1.5    ad 	 * which means another interrupt at this level can be triggered.
    784   1.5    ad 	 * That's not be a problem: we are lowering to level 's' which will
    785   1.5    ad 	 * prevent softint_dispatch() from being reentered at level 's',
    786   1.5    ad 	 * until the priority is finally dropped to IPL_NONE on entry to
    787   1.5    ad 	 * the idle loop.
    788   1.5    ad 	 */
    789   1.5    ad 	l->l_stat = LSIDL;
    790   1.5    ad 	if (l->l_switchto == NULL) {
    791   1.5    ad 		splx(s);
    792   1.5    ad 		pmap_deactivate(l);
    793   1.5    ad 		lwp_exit_switchaway(l);
    794   1.5    ad 		/* NOTREACHED */
    795   1.5    ad 	}
    796   1.5    ad 	l->l_switchto = NULL;
    797   1.5    ad 	l->l_flag &= ~LW_RUNNING;
    798   1.5    ad }
    799   1.5    ad 
    800   1.5    ad #endif	/* !__HAVE_FAST_SOFTINTS */
    801