Home | History | Annotate | Line # | Download | only in kern
kern_softint.c revision 1.32
      1  1.32    matt /*	$NetBSD: kern_softint.c,v 1.32 2010/12/11 22:32:13 matt 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  *
     19   1.2      ad  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
     20   1.2      ad  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
     21   1.2      ad  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
     22   1.2      ad  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
     23   1.2      ad  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
     24   1.2      ad  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
     25   1.2      ad  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
     26   1.2      ad  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
     27   1.2      ad  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
     28   1.2      ad  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
     29   1.2      ad  * POSSIBILITY OF SUCH DAMAGE.
     30   1.2      ad  */
     31   1.2      ad 
     32   1.2      ad /*
     33   1.5      ad  * Generic software interrupt framework.
     34   1.5      ad  *
     35   1.5      ad  * Overview
     36   1.5      ad  *
     37   1.5      ad  *	The soft interrupt framework provides a mechanism to schedule a
     38   1.5      ad  *	low priority callback that runs with thread context.  It allows
     39   1.5      ad  *	for dynamic registration of software interrupts, and for fair
     40   1.5      ad  *	queueing and prioritization of those interrupts.  The callbacks
     41   1.5      ad  *	can be scheduled to run from nearly any point in the kernel: by
     42   1.5      ad  *	code running with thread context, by code running from a
     43   1.5      ad  *	hardware interrupt handler, and at any interrupt priority
     44   1.5      ad  *	level.
     45   1.5      ad  *
     46   1.5      ad  * Priority levels
     47   1.5      ad  *
     48   1.5      ad  *	Since soft interrupt dispatch can be tied to the underlying
     49   1.5      ad  *	architecture's interrupt dispatch code, it can be limited
     50   1.5      ad  *	both by the capabilities of the hardware and the capabilities
     51   1.5      ad  *	of the interrupt dispatch code itself.  The number of priority
     52   1.5      ad  *	levels is restricted to four.  In order of priority (lowest to
     53   1.5      ad  *	highest) the levels are: clock, bio, net, serial.
     54   1.5      ad  *
     55   1.5      ad  *	The names are symbolic and in isolation do not have any direct
     56   1.5      ad  *	connection with a particular kind of device activity: they are
     57   1.5      ad  *	only meant as a guide.
     58   1.5      ad  *
     59   1.5      ad  *	The four priority levels map directly to scheduler priority
     60   1.5      ad  *	levels, and where the architecture implements 'fast' software
     61   1.5      ad  *	interrupts, they also map onto interrupt priorities.  The
     62   1.5      ad  *	interrupt priorities are intended to be hidden from machine
     63   1.5      ad  *	independent code, which should use thread-safe mechanisms to
     64   1.5      ad  *	synchronize with software interrupts (for example: mutexes).
     65   1.5      ad  *
     66   1.5      ad  * Capabilities
     67   1.5      ad  *
     68   1.5      ad  *	Software interrupts run with limited machine context.  In
     69   1.5      ad  *	particular, they do not posess any address space context.  They
     70   1.5      ad  *	should not try to operate on user space addresses, or to use
     71   1.5      ad  *	virtual memory facilities other than those noted as interrupt
     72   1.5      ad  *	safe.
     73   1.5      ad  *
     74   1.5      ad  *	Unlike hardware interrupts, software interrupts do have thread
     75   1.5      ad  *	context.  They may block on synchronization objects, sleep, and
     76   1.5      ad  *	resume execution at a later time.
     77   1.5      ad  *
     78   1.5      ad  *	Since software interrupts are a limited resource and run with
     79   1.5      ad  *	higher priority than most other LWPs in the system, all
     80   1.5      ad  *	block-and-resume activity by a software interrupt must be kept
     81   1.5      ad  *	short to allow futher processing at that level to continue.  By
     82   1.5      ad  *	extension, code running with process context must take care to
     83   1.5      ad  *	ensure that any lock that may be taken from a software interrupt
     84   1.5      ad  *	can not be held for more than a short period of time.
     85   1.5      ad  *
     86   1.5      ad  *	The kernel does not allow software interrupts to use facilities
     87   1.5      ad  *	or perform actions that may block for a significant amount of
     88   1.5      ad  *	time.  This means that it's not valid for a software interrupt
     89  1.10      ad  *	to sleep on condition variables	or wait for resources to become
     90  1.10      ad  *	available (for example,	memory).
     91   1.5      ad  *
     92   1.5      ad  * Per-CPU operation
     93   1.5      ad  *
     94   1.5      ad  *	If a soft interrupt is triggered on a CPU, it can only be
     95   1.5      ad  *	dispatched on the same CPU.  Each LWP dedicated to handling a
     96   1.5      ad  *	soft interrupt is bound to its home CPU, so if the LWP blocks
     97   1.5      ad  *	and needs to run again, it can only run there.  Nearly all data
     98   1.5      ad  *	structures used to manage software interrupts are per-CPU.
     99   1.5      ad  *
    100   1.5      ad  *	The per-CPU requirement is intended to reduce "ping-pong" of
    101   1.5      ad  *	cache lines between CPUs: lines occupied by data structures
    102   1.5      ad  *	used to manage the soft interrupts, and lines occupied by data
    103   1.5      ad  *	items being passed down to the soft interrupt.  As a positive
    104   1.5      ad  *	side effect, this also means that the soft interrupt dispatch
    105   1.5      ad  *	code does not need to to use spinlocks to synchronize.
    106   1.5      ad  *
    107   1.5      ad  * Generic implementation
    108   1.5      ad  *
    109   1.5      ad  *	A generic, low performance implementation is provided that
    110   1.5      ad  *	works across all architectures, with no machine-dependent
    111   1.5      ad  *	modifications needed.  This implementation uses the scheduler,
    112   1.5      ad  *	and so has a number of restrictions:
    113   1.5      ad  *
    114   1.5      ad  *	1) The software interrupts are not currently preemptive, so
    115   1.5      ad  *	must wait for the currently executing LWP to yield the CPU.
    116   1.5      ad  *	This can introduce latency.
    117   1.5      ad  *
    118   1.5      ad  *	2) An expensive context switch is required for a software
    119   1.5      ad  *	interrupt to be handled.
    120   1.5      ad  *
    121   1.5      ad  * 'Fast' software interrupts
    122   1.5      ad  *
    123   1.5      ad  *	If an architectures defines __HAVE_FAST_SOFTINTS, it implements
    124   1.5      ad  *	the fast mechanism.  Threads running either in the kernel or in
    125   1.5      ad  *	userspace will be interrupted, but will not be preempted.  When
    126   1.5      ad  *	the soft interrupt completes execution, the interrupted LWP
    127   1.5      ad  *	is resumed.  Interrupt dispatch code must provide the minimum
    128   1.5      ad  *	level of context necessary for the soft interrupt to block and
    129   1.5      ad  *	be resumed at a later time.  The machine-dependent dispatch
    130   1.5      ad  *	path looks something like the following:
    131   1.5      ad  *
    132   1.5      ad  *	softintr()
    133   1.5      ad  *	{
    134   1.5      ad  *		go to IPL_HIGH if necessary for switch;
    135   1.5      ad  *		save any necessary registers in a format that can be
    136   1.5      ad  *		    restored by cpu_switchto if the softint blocks;
    137   1.5      ad  *		arrange for cpu_switchto() to restore into the
    138   1.5      ad  *		    trampoline function;
    139   1.5      ad  *		identify LWP to handle this interrupt;
    140   1.5      ad  *		switch to the LWP's stack;
    141   1.5      ad  *		switch register stacks, if necessary;
    142   1.5      ad  *		assign new value of curlwp;
    143   1.5      ad  *		call MI softint_dispatch, passing old curlwp and IPL
    144   1.5      ad  *		    to execute interrupt at;
    145   1.5      ad  *		switch back to old stack;
    146   1.5      ad  *		switch back to old register stack, if necessary;
    147   1.5      ad  *		restore curlwp;
    148   1.5      ad  *		return to interrupted LWP;
    149   1.5      ad  *	}
    150   1.5      ad  *
    151   1.5      ad  *	If the soft interrupt blocks, a trampoline function is returned
    152   1.5      ad  *	to in the context of the interrupted LWP, as arranged for by
    153   1.5      ad  *	softint():
    154   1.5      ad  *
    155   1.5      ad  *	softint_ret()
    156   1.5      ad  *	{
    157   1.5      ad  *		unlock soft interrupt LWP;
    158   1.5      ad  *		resume interrupt processing, likely returning to
    159   1.5      ad  *		    interrupted LWP or dispatching another, different
    160   1.5      ad  *		    interrupt;
    161   1.5      ad  *	}
    162   1.5      ad  *
    163   1.5      ad  *	Once the soft interrupt has fired (and even if it has blocked),
    164   1.5      ad  *	no further soft interrupts at that level will be triggered by
    165   1.5      ad  *	MI code until the soft interrupt handler has ceased execution.
    166   1.5      ad  *	If a soft interrupt handler blocks and is resumed, it resumes
    167   1.5      ad  *	execution as a normal LWP (kthread) and gains VM context.  Only
    168   1.5      ad  *	when it has completed and is ready to fire again will it
    169   1.5      ad  *	interrupt other threads.
    170   1.5      ad  *
    171   1.5      ad  * Future directions
    172   1.5      ad  *
    173   1.5      ad  *	Provide a cheap way to direct software interrupts to remote
    174   1.5      ad  *	CPUs.  Provide a way to enqueue work items into the handler
    175   1.5      ad  *	record,	removing additional spl calls (see subr_workqueue.c).
    176   1.2      ad  */
    177   1.2      ad 
    178   1.2      ad #include <sys/cdefs.h>
    179  1.32    matt __KERNEL_RCSID(0, "$NetBSD: kern_softint.c,v 1.32 2010/12/11 22:32:13 matt Exp $");
    180   1.2      ad 
    181   1.2      ad #include <sys/param.h>
    182   1.5      ad #include <sys/malloc.h>
    183   1.5      ad #include <sys/proc.h>
    184   1.2      ad #include <sys/intr.h>
    185   1.5      ad #include <sys/mutex.h>
    186   1.5      ad #include <sys/kthread.h>
    187   1.5      ad #include <sys/evcnt.h>
    188   1.5      ad #include <sys/cpu.h>
    189  1.24      ad #include <sys/xcall.h>
    190   1.5      ad 
    191   1.5      ad #include <net/netisr.h>
    192   1.5      ad 
    193   1.5      ad #include <uvm/uvm_extern.h>
    194   1.5      ad 
    195   1.5      ad /* This could overlap with signal info in struct lwp. */
    196   1.5      ad typedef struct softint {
    197   1.5      ad 	SIMPLEQ_HEAD(, softhand) si_q;
    198   1.5      ad 	struct lwp		*si_lwp;
    199   1.5      ad 	struct cpu_info		*si_cpu;
    200   1.5      ad 	uintptr_t		si_machdep;
    201   1.5      ad 	struct evcnt		si_evcnt;
    202   1.5      ad 	struct evcnt		si_evcnt_block;
    203   1.5      ad 	int			si_active;
    204   1.5      ad 	char			si_name[8];
    205   1.5      ad 	char			si_name_block[8+6];
    206   1.5      ad } softint_t;
    207   1.5      ad 
    208   1.5      ad typedef struct softhand {
    209   1.5      ad 	SIMPLEQ_ENTRY(softhand)	sh_q;
    210   1.5      ad 	void			(*sh_func)(void *);
    211   1.5      ad 	void			*sh_arg;
    212   1.5      ad 	softint_t		*sh_isr;
    213  1.28  bouyer 	u_int			sh_flags;
    214   1.5      ad } softhand_t;
    215   1.5      ad 
    216   1.5      ad typedef struct softcpu {
    217   1.5      ad 	struct cpu_info		*sc_cpu;
    218   1.5      ad 	softint_t		sc_int[SOFTINT_COUNT];
    219   1.5      ad 	softhand_t		sc_hand[1];
    220   1.5      ad } softcpu_t;
    221   1.5      ad 
    222   1.5      ad static void	softint_thread(void *);
    223   1.5      ad 
    224   1.5      ad u_int		softint_bytes = 8192;
    225   1.5      ad u_int		softint_timing;
    226   1.5      ad static u_int	softint_max;
    227   1.5      ad static kmutex_t	softint_lock;
    228  1.23   pooka static void	*softint_netisrs[NETISR_MAX];
    229   1.2      ad 
    230   1.5      ad /*
    231   1.5      ad  * softint_init_isr:
    232   1.5      ad  *
    233   1.5      ad  *	Initialize a single interrupt level for a single CPU.
    234   1.5      ad  */
    235   1.5      ad static void
    236   1.5      ad softint_init_isr(softcpu_t *sc, const char *desc, pri_t pri, u_int level)
    237   1.5      ad {
    238   1.5      ad 	struct cpu_info *ci;
    239   1.5      ad 	softint_t *si;
    240   1.5      ad 	int error;
    241   1.5      ad 
    242   1.5      ad 	si = &sc->sc_int[level];
    243   1.5      ad 	ci = sc->sc_cpu;
    244   1.5      ad 	si->si_cpu = ci;
    245   1.5      ad 
    246   1.5      ad 	SIMPLEQ_INIT(&si->si_q);
    247   1.5      ad 
    248   1.5      ad 	error = kthread_create(pri, KTHREAD_MPSAFE | KTHREAD_INTR |
    249   1.5      ad 	    KTHREAD_IDLE, ci, softint_thread, si, &si->si_lwp,
    250  1.12  martin 	    "soft%s/%u", desc, ci->ci_index);
    251   1.5      ad 	if (error != 0)
    252   1.5      ad 		panic("softint_init_isr: error %d", error);
    253   1.5      ad 
    254  1.12  martin 	snprintf(si->si_name, sizeof(si->si_name), "%s/%u", desc,
    255  1.12  martin 	    ci->ci_index);
    256  1.20      ad 	evcnt_attach_dynamic(&si->si_evcnt, EVCNT_TYPE_MISC, NULL,
    257   1.5      ad 	   "softint", si->si_name);
    258  1.12  martin 	snprintf(si->si_name_block, sizeof(si->si_name_block), "%s block/%u",
    259  1.12  martin 	    desc, ci->ci_index);
    260  1.20      ad 	evcnt_attach_dynamic(&si->si_evcnt_block, EVCNT_TYPE_MISC, NULL,
    261   1.5      ad 	   "softint", si->si_name_block);
    262   1.3      ad 
    263   1.5      ad 	si->si_lwp->l_private = si;
    264   1.5      ad 	softint_init_md(si->si_lwp, level, &si->si_machdep);
    265   1.5      ad }
    266   1.2      ad /*
    267   1.2      ad  * softint_init:
    268   1.2      ad  *
    269   1.2      ad  *	Initialize per-CPU data structures.  Called from mi_cpu_attach().
    270   1.2      ad  */
    271   1.2      ad void
    272   1.2      ad softint_init(struct cpu_info *ci)
    273   1.2      ad {
    274   1.5      ad 	static struct cpu_info *first;
    275   1.5      ad 	softcpu_t *sc, *scfirst;
    276   1.5      ad 	softhand_t *sh, *shmax;
    277   1.5      ad 
    278   1.5      ad 	if (first == NULL) {
    279   1.5      ad 		/* Boot CPU. */
    280   1.5      ad 		first = ci;
    281   1.5      ad 		mutex_init(&softint_lock, MUTEX_DEFAULT, IPL_NONE);
    282   1.5      ad 		softint_bytes = round_page(softint_bytes);
    283   1.5      ad 		softint_max = (softint_bytes - sizeof(softcpu_t)) /
    284   1.5      ad 		    sizeof(softhand_t);
    285   1.5      ad 	}
    286   1.2      ad 
    287   1.5      ad 	sc = (softcpu_t *)uvm_km_alloc(kernel_map, softint_bytes, 0,
    288   1.5      ad 	    UVM_KMF_WIRED | UVM_KMF_ZERO);
    289   1.5      ad 	if (sc == NULL)
    290   1.5      ad 		panic("softint_init_cpu: cannot allocate memory");
    291   1.5      ad 
    292   1.5      ad 	ci->ci_data.cpu_softcpu = sc;
    293   1.5      ad 	ci->ci_data.cpu_softints = 0;
    294   1.5      ad 	sc->sc_cpu = ci;
    295   1.5      ad 
    296   1.5      ad 	softint_init_isr(sc, "net", PRI_SOFTNET, SOFTINT_NET);
    297   1.5      ad 	softint_init_isr(sc, "bio", PRI_SOFTBIO, SOFTINT_BIO);
    298   1.5      ad 	softint_init_isr(sc, "clk", PRI_SOFTCLOCK, SOFTINT_CLOCK);
    299   1.5      ad 	softint_init_isr(sc, "ser", PRI_SOFTSERIAL, SOFTINT_SERIAL);
    300   1.5      ad 
    301   1.5      ad 	if (first != ci) {
    302   1.5      ad 		mutex_enter(&softint_lock);
    303   1.5      ad 		scfirst = first->ci_data.cpu_softcpu;
    304   1.5      ad 		sh = sc->sc_hand;
    305   1.5      ad 		memcpy(sh, scfirst->sc_hand, sizeof(*sh) * softint_max);
    306   1.5      ad 		/* Update pointers for this CPU. */
    307   1.5      ad 		for (shmax = sh + softint_max; sh < shmax; sh++) {
    308   1.5      ad 			if (sh->sh_func == NULL)
    309   1.5      ad 				continue;
    310   1.5      ad 			sh->sh_isr =
    311   1.5      ad 			    &sc->sc_int[sh->sh_flags & SOFTINT_LVLMASK];
    312   1.5      ad 		}
    313   1.5      ad 		mutex_exit(&softint_lock);
    314   1.5      ad 	} else {
    315   1.5      ad 		/*
    316   1.5      ad 		 * Establish handlers for legacy net interrupts.
    317   1.5      ad 		 * XXX Needs to go away.
    318   1.5      ad 		 */
    319   1.5      ad #define DONETISR(n, f)							\
    320  1.16      ad     softint_netisrs[(n)] = softint_establish(SOFTINT_NET|SOFTINT_MPSAFE,\
    321  1.16      ad         (void (*)(void *))(f), NULL)
    322   1.5      ad #include <net/netisr_dispatch.h>
    323   1.5      ad 	}
    324   1.2      ad }
    325   1.2      ad 
    326   1.2      ad /*
    327   1.2      ad  * softint_establish:
    328   1.2      ad  *
    329   1.2      ad  *	Register a software interrupt handler.
    330   1.2      ad  */
    331   1.2      ad void *
    332   1.2      ad softint_establish(u_int flags, void (*func)(void *), void *arg)
    333   1.2      ad {
    334   1.5      ad 	CPU_INFO_ITERATOR cii;
    335   1.5      ad 	struct cpu_info *ci;
    336   1.5      ad 	softcpu_t *sc;
    337   1.5      ad 	softhand_t *sh;
    338   1.5      ad 	u_int level, index;
    339   1.2      ad 
    340   1.2      ad 	level = (flags & SOFTINT_LVLMASK);
    341   1.2      ad 	KASSERT(level < SOFTINT_COUNT);
    342  1.24      ad 	KASSERT((flags & SOFTINT_IMPMASK) == 0);
    343   1.2      ad 
    344   1.5      ad 	mutex_enter(&softint_lock);
    345   1.5      ad 
    346   1.5      ad 	/* Find a free slot. */
    347   1.5      ad 	sc = curcpu()->ci_data.cpu_softcpu;
    348  1.32    matt 	for (index = 1; index < softint_max; index++) {
    349   1.5      ad 		if (sc->sc_hand[index].sh_func == NULL)
    350   1.5      ad 			break;
    351  1.32    matt 	}
    352   1.5      ad 	if (index == softint_max) {
    353   1.5      ad 		mutex_exit(&softint_lock);
    354   1.5      ad 		printf("WARNING: softint_establish: table full, "
    355   1.5      ad 		    "increase softint_bytes\n");
    356   1.5      ad 		return NULL;
    357   1.5      ad 	}
    358   1.5      ad 
    359   1.5      ad 	/* Set up the handler on each CPU. */
    360   1.8      ad 	if (ncpu < 2) {
    361   1.7      ad 		/* XXX hack for machines with no CPU_INFO_FOREACH() early on */
    362   1.7      ad 		sc = curcpu()->ci_data.cpu_softcpu;
    363   1.7      ad 		sh = &sc->sc_hand[index];
    364   1.7      ad 		sh->sh_isr = &sc->sc_int[level];
    365   1.7      ad 		sh->sh_func = func;
    366   1.7      ad 		sh->sh_arg = arg;
    367   1.7      ad 		sh->sh_flags = flags;
    368   1.7      ad 	} else for (CPU_INFO_FOREACH(cii, ci)) {
    369   1.5      ad 		sc = ci->ci_data.cpu_softcpu;
    370   1.5      ad 		sh = &sc->sc_hand[index];
    371   1.5      ad 		sh->sh_isr = &sc->sc_int[level];
    372   1.5      ad 		sh->sh_func = func;
    373   1.5      ad 		sh->sh_arg = arg;
    374   1.5      ad 		sh->sh_flags = flags;
    375   1.2      ad 	}
    376   1.2      ad 
    377   1.5      ad 	mutex_exit(&softint_lock);
    378   1.5      ad 
    379   1.5      ad 	return (void *)((uint8_t *)&sc->sc_hand[index] - (uint8_t *)sc);
    380   1.2      ad }
    381   1.2      ad 
    382   1.2      ad /*
    383   1.2      ad  * softint_disestablish:
    384   1.2      ad  *
    385  1.24      ad  *	Unregister a software interrupt handler.  The soft interrupt could
    386  1.24      ad  *	still be active at this point, but the caller commits not to try
    387  1.24      ad  *	and trigger it again once this call is made.  The caller must not
    388  1.24      ad  *	hold any locks that could be taken from soft interrupt context,
    389  1.24      ad  *	because we will wait for the softint to complete if it's still
    390  1.24      ad  *	running.
    391   1.2      ad  */
    392   1.2      ad void
    393   1.2      ad softint_disestablish(void *arg)
    394   1.2      ad {
    395   1.5      ad 	CPU_INFO_ITERATOR cii;
    396   1.5      ad 	struct cpu_info *ci;
    397   1.5      ad 	softcpu_t *sc;
    398   1.5      ad 	softhand_t *sh;
    399   1.5      ad 	uintptr_t offset;
    400  1.24      ad 	uint64_t where;
    401  1.24      ad 	u_int flags;
    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  1.24      ad 	/*
    407  1.24      ad 	 * Run a cross call so we see up to date values of sh_flags from
    408  1.24      ad 	 * all CPUs.  Once softint_disestablish() is called, the caller
    409  1.24      ad 	 * commits to not trigger the interrupt and set SOFTINT_ACTIVE on
    410  1.24      ad 	 * it again.  So, we are only looking for handler records with
    411  1.26  dyoung 	 * SOFTINT_ACTIVE already set.
    412  1.24      ad 	 */
    413  1.24      ad 	where = xc_broadcast(0, (xcfunc_t)nullop, NULL, NULL);
    414  1.24      ad 	xc_wait(where);
    415  1.24      ad 
    416  1.24      ad 	for (;;) {
    417  1.24      ad 		/* Collect flag values from each CPU. */
    418  1.24      ad 		flags = 0;
    419  1.24      ad 		for (CPU_INFO_FOREACH(cii, ci)) {
    420  1.24      ad 			sc = ci->ci_data.cpu_softcpu;
    421  1.24      ad 			sh = (softhand_t *)((uint8_t *)sc + offset);
    422  1.24      ad 			KASSERT(sh->sh_func != NULL);
    423  1.24      ad 			flags |= sh->sh_flags;
    424  1.24      ad 		}
    425  1.24      ad 		/* Inactive on all CPUs? */
    426  1.24      ad 		if ((flags & SOFTINT_ACTIVE) == 0) {
    427  1.24      ad 			break;
    428  1.24      ad 		}
    429  1.24      ad 		/* Oops, still active.  Wait for it to clear. */
    430  1.25      ad 		(void)kpause("softdis", false, 1, NULL);
    431  1.24      ad 	}
    432   1.5      ad 
    433   1.5      ad 	/* Clear the handler on each CPU. */
    434  1.24      ad 	mutex_enter(&softint_lock);
    435   1.5      ad 	for (CPU_INFO_FOREACH(cii, ci)) {
    436   1.5      ad 		sc = ci->ci_data.cpu_softcpu;
    437   1.5      ad 		sh = (softhand_t *)((uint8_t *)sc + offset);
    438   1.5      ad 		KASSERT(sh->sh_func != NULL);
    439   1.5      ad 		sh->sh_func = NULL;
    440   1.5      ad 	}
    441   1.5      ad 	mutex_exit(&softint_lock);
    442   1.2      ad }
    443   1.2      ad 
    444   1.2      ad /*
    445   1.2      ad  * softint_schedule:
    446   1.2      ad  *
    447   1.2      ad  *	Trigger a software interrupt.  Must be called from a hardware
    448   1.2      ad  *	interrupt handler, or with preemption disabled (since we are
    449   1.2      ad  *	using the value of curcpu()).
    450   1.2      ad  */
    451   1.2      ad void
    452   1.2      ad softint_schedule(void *arg)
    453   1.2      ad {
    454   1.5      ad 	softhand_t *sh;
    455   1.5      ad 	softint_t *si;
    456   1.5      ad 	uintptr_t offset;
    457   1.5      ad 	int s;
    458   1.5      ad 
    459  1.17      ad 	KASSERT(kpreempt_disabled());
    460  1.17      ad 
    461   1.5      ad 	/* Find the handler record for this CPU. */
    462   1.5      ad 	offset = (uintptr_t)arg;
    463   1.5      ad 	KASSERT(offset != 0 && offset < softint_bytes);
    464   1.5      ad 	sh = (softhand_t *)((uint8_t *)curcpu()->ci_data.cpu_softcpu + offset);
    465   1.5      ad 
    466   1.5      ad 	/* If it's already pending there's nothing to do. */
    467  1.32    matt 	if ((sh->sh_flags & SOFTINT_PENDING) != 0) {
    468   1.5      ad 		return;
    469  1.32    matt 	}
    470   1.5      ad 
    471   1.5      ad 	/*
    472   1.5      ad 	 * Enqueue the handler into the LWP's pending list.
    473   1.5      ad 	 * If the LWP is completely idle, then make it run.
    474   1.5      ad 	 */
    475   1.5      ad 	s = splhigh();
    476  1.24      ad 	if ((sh->sh_flags & SOFTINT_PENDING) == 0) {
    477   1.5      ad 		si = sh->sh_isr;
    478  1.24      ad 		sh->sh_flags |= SOFTINT_PENDING;
    479   1.5      ad 		SIMPLEQ_INSERT_TAIL(&si->si_q, sh, sh_q);
    480   1.5      ad 		if (si->si_active == 0) {
    481   1.5      ad 			si->si_active = 1;
    482   1.5      ad 			softint_trigger(si->si_machdep);
    483   1.5      ad 		}
    484   1.5      ad 	}
    485   1.5      ad 	splx(s);
    486   1.5      ad }
    487   1.5      ad 
    488   1.5      ad /*
    489   1.5      ad  * softint_execute:
    490   1.5      ad  *
    491   1.5      ad  *	Invoke handlers for the specified soft interrupt.
    492   1.5      ad  *	Must be entered at splhigh.  Will drop the priority
    493   1.5      ad  *	to the level specified, but returns back at splhigh.
    494   1.5      ad  */
    495   1.5      ad static inline void
    496   1.5      ad softint_execute(softint_t *si, lwp_t *l, int s)
    497   1.5      ad {
    498   1.5      ad 	softhand_t *sh;
    499   1.5      ad 	bool havelock;
    500   1.5      ad 
    501   1.5      ad #ifdef __HAVE_FAST_SOFTINTS
    502   1.5      ad 	KASSERT(si->si_lwp == curlwp);
    503   1.5      ad #else
    504   1.5      ad 	/* May be running in user context. */
    505   1.5      ad #endif
    506   1.5      ad 	KASSERT(si->si_cpu == curcpu());
    507   1.5      ad 	KASSERT(si->si_lwp->l_wchan == NULL);
    508   1.5      ad 	KASSERT(si->si_active);
    509   1.5      ad 
    510   1.5      ad 	havelock = false;
    511   1.5      ad 
    512   1.5      ad 	/*
    513   1.5      ad 	 * Note: due to priority inheritance we may have interrupted a
    514   1.5      ad 	 * higher priority LWP.  Since the soft interrupt must be quick
    515   1.5      ad 	 * and is non-preemptable, we don't bother yielding.
    516   1.5      ad 	 */
    517   1.5      ad 
    518   1.5      ad 	while (!SIMPLEQ_EMPTY(&si->si_q)) {
    519   1.5      ad 		/*
    520   1.5      ad 		 * Pick the longest waiting handler to run.  We block
    521   1.5      ad 		 * interrupts but do not lock in order to do this, as
    522   1.5      ad 		 * we are protecting against the local CPU only.
    523   1.5      ad 		 */
    524   1.5      ad 		sh = SIMPLEQ_FIRST(&si->si_q);
    525   1.5      ad 		SIMPLEQ_REMOVE_HEAD(&si->si_q, sh_q);
    526  1.24      ad 		KASSERT((sh->sh_flags & SOFTINT_PENDING) != 0);
    527  1.24      ad 		KASSERT((sh->sh_flags & SOFTINT_ACTIVE) == 0);
    528  1.24      ad 		sh->sh_flags ^= (SOFTINT_PENDING | SOFTINT_ACTIVE);
    529   1.5      ad 		splx(s);
    530   1.5      ad 
    531   1.5      ad 		/* Run the handler. */
    532  1.30   rmind 		if (sh->sh_flags & SOFTINT_MPSAFE) {
    533  1.30   rmind 			if (havelock) {
    534  1.30   rmind 				KERNEL_UNLOCK_ONE(l);
    535  1.30   rmind 				havelock = false;
    536  1.30   rmind 			}
    537  1.30   rmind 		} else if (!havelock) {
    538   1.5      ad 			KERNEL_LOCK(1, l);
    539   1.5      ad 			havelock = true;
    540   1.5      ad 		}
    541   1.5      ad 		(*sh->sh_func)(sh->sh_arg);
    542   1.5      ad 
    543   1.5      ad 		(void)splhigh();
    544  1.24      ad 		KASSERT((sh->sh_flags & SOFTINT_ACTIVE) != 0);
    545  1.24      ad 		sh->sh_flags ^= SOFTINT_ACTIVE;
    546   1.5      ad 	}
    547   1.2      ad 
    548   1.5      ad 	if (havelock) {
    549   1.5      ad 		KERNEL_UNLOCK_ONE(l);
    550   1.5      ad 	}
    551   1.5      ad 
    552   1.5      ad 	/*
    553   1.5      ad 	 * Unlocked, but only for statistics.
    554   1.5      ad 	 * Should be per-CPU to prevent cache ping-pong.
    555   1.5      ad 	 */
    556   1.5      ad 	uvmexp.softs++;
    557   1.5      ad 
    558  1.13      ad 	KASSERT(si->si_cpu == curcpu());
    559  1.13      ad 	KASSERT(si->si_lwp->l_wchan == NULL);
    560  1.13      ad 	KASSERT(si->si_active);
    561   1.5      ad 	si->si_evcnt.ev_count++;
    562   1.5      ad 	si->si_active = 0;
    563   1.2      ad }
    564   1.2      ad 
    565   1.2      ad /*
    566   1.2      ad  * softint_block:
    567   1.2      ad  *
    568   1.2      ad  *	Update statistics when the soft interrupt blocks.
    569   1.2      ad  */
    570   1.2      ad void
    571   1.2      ad softint_block(lwp_t *l)
    572   1.2      ad {
    573   1.5      ad 	softint_t *si = l->l_private;
    574   1.5      ad 
    575   1.5      ad 	KASSERT((l->l_pflag & LP_INTR) != 0);
    576   1.5      ad 	si->si_evcnt_block.ev_count++;
    577   1.5      ad }
    578   1.5      ad 
    579   1.5      ad /*
    580   1.5      ad  * schednetisr:
    581   1.5      ad  *
    582   1.5      ad  *	Trigger a legacy network interrupt.  XXX Needs to go away.
    583   1.5      ad  */
    584   1.5      ad void
    585   1.5      ad schednetisr(int isr)
    586   1.5      ad {
    587   1.5      ad 
    588   1.5      ad 	softint_schedule(softint_netisrs[isr]);
    589   1.5      ad }
    590   1.5      ad 
    591   1.5      ad #ifndef __HAVE_FAST_SOFTINTS
    592   1.5      ad 
    593  1.19      ad #ifdef __HAVE_PREEMPTION
    594  1.19      ad #error __HAVE_PREEMPTION requires __HAVE_FAST_SOFTINTS
    595  1.17      ad #endif
    596  1.17      ad 
    597   1.5      ad /*
    598   1.5      ad  * softint_init_md:
    599   1.5      ad  *
    600   1.5      ad  *	Slow path: perform machine-dependent initialization.
    601   1.5      ad  */
    602   1.5      ad void
    603   1.5      ad softint_init_md(lwp_t *l, u_int level, uintptr_t *machdep)
    604   1.5      ad {
    605   1.5      ad 	softint_t *si;
    606   1.5      ad 
    607   1.5      ad 	*machdep = (1 << level);
    608   1.5      ad 	si = l->l_private;
    609   1.5      ad 
    610   1.5      ad 	lwp_lock(l);
    611   1.5      ad 	lwp_unlock_to(l, l->l_cpu->ci_schedstate.spc_mutex);
    612   1.5      ad 	lwp_lock(l);
    613   1.5      ad 	/* Cheat and make the KASSERT in softint_thread() happy. */
    614   1.5      ad 	si->si_active = 1;
    615   1.5      ad 	l->l_stat = LSRUN;
    616   1.5      ad 	sched_enqueue(l, false);
    617   1.5      ad 	lwp_unlock(l);
    618   1.5      ad }
    619   1.5      ad 
    620   1.5      ad /*
    621   1.5      ad  * softint_trigger:
    622   1.5      ad  *
    623   1.5      ad  *	Slow path: cause a soft interrupt handler to begin executing.
    624   1.5      ad  *	Called at IPL_HIGH.
    625   1.5      ad  */
    626   1.5      ad void
    627   1.5      ad softint_trigger(uintptr_t machdep)
    628   1.5      ad {
    629   1.5      ad 	struct cpu_info *ci;
    630   1.5      ad 	lwp_t *l;
    631   1.2      ad 
    632   1.5      ad 	l = curlwp;
    633   1.5      ad 	ci = l->l_cpu;
    634   1.5      ad 	ci->ci_data.cpu_softints |= machdep;
    635   1.5      ad 	if (l == ci->ci_data.cpu_idlelwp) {
    636   1.5      ad 		cpu_need_resched(ci, 0);
    637   1.5      ad 	} else {
    638   1.5      ad 		/* MI equivalent of aston() */
    639   1.5      ad 		cpu_signotify(l);
    640   1.5      ad 	}
    641   1.5      ad }
    642   1.5      ad 
    643   1.5      ad /*
    644   1.5      ad  * softint_thread:
    645   1.5      ad  *
    646   1.5      ad  *	Slow path: MI software interrupt dispatch.
    647   1.5      ad  */
    648   1.5      ad void
    649   1.5      ad softint_thread(void *cookie)
    650   1.5      ad {
    651   1.5      ad 	softint_t *si;
    652   1.5      ad 	lwp_t *l;
    653   1.5      ad 	int s;
    654   1.5      ad 
    655   1.5      ad 	l = curlwp;
    656   1.5      ad 	si = l->l_private;
    657   1.5      ad 
    658   1.5      ad 	for (;;) {
    659   1.5      ad 		/*
    660   1.5      ad 		 * Clear pending status and run it.  We must drop the
    661   1.5      ad 		 * spl before mi_switch(), since IPL_HIGH may be higher
    662   1.5      ad 		 * than IPL_SCHED (and it is not safe to switch at a
    663   1.5      ad 		 * higher level).
    664   1.5      ad 		 */
    665   1.5      ad 		s = splhigh();
    666   1.5      ad 		l->l_cpu->ci_data.cpu_softints &= ~si->si_machdep;
    667   1.5      ad 		softint_execute(si, l, s);
    668   1.5      ad 		splx(s);
    669   1.5      ad 
    670   1.5      ad 		lwp_lock(l);
    671   1.5      ad 		l->l_stat = LSIDL;
    672   1.5      ad 		mi_switch(l);
    673   1.5      ad 	}
    674   1.2      ad }
    675   1.4      ad 
    676   1.4      ad /*
    677   1.4      ad  * softint_picklwp:
    678   1.4      ad  *
    679   1.4      ad  *	Slow path: called from mi_switch() to pick the highest priority
    680   1.4      ad  *	soft interrupt LWP that needs to run.
    681   1.4      ad  */
    682   1.4      ad lwp_t *
    683   1.4      ad softint_picklwp(void)
    684   1.4      ad {
    685   1.5      ad 	struct cpu_info *ci;
    686   1.5      ad 	u_int mask;
    687   1.5      ad 	softint_t *si;
    688   1.5      ad 	lwp_t *l;
    689   1.5      ad 
    690   1.5      ad 	ci = curcpu();
    691   1.5      ad 	si = ((softcpu_t *)ci->ci_data.cpu_softcpu)->sc_int;
    692   1.5      ad 	mask = ci->ci_data.cpu_softints;
    693   1.5      ad 
    694   1.5      ad 	if ((mask & (1 << SOFTINT_SERIAL)) != 0) {
    695   1.5      ad 		l = si[SOFTINT_SERIAL].si_lwp;
    696   1.5      ad 	} else if ((mask & (1 << SOFTINT_NET)) != 0) {
    697   1.5      ad 		l = si[SOFTINT_NET].si_lwp;
    698   1.5      ad 	} else if ((mask & (1 << SOFTINT_BIO)) != 0) {
    699   1.5      ad 		l = si[SOFTINT_BIO].si_lwp;
    700   1.5      ad 	} else if ((mask & (1 << SOFTINT_CLOCK)) != 0) {
    701   1.5      ad 		l = si[SOFTINT_CLOCK].si_lwp;
    702   1.5      ad 	} else {
    703   1.5      ad 		panic("softint_picklwp");
    704   1.5      ad 	}
    705   1.4      ad 
    706   1.5      ad 	return l;
    707   1.4      ad }
    708   1.4      ad 
    709   1.4      ad /*
    710   1.4      ad  * softint_overlay:
    711   1.4      ad  *
    712   1.4      ad  *	Slow path: called from lwp_userret() to run a soft interrupt
    713   1.6      ad  *	within the context of a user thread.
    714   1.4      ad  */
    715   1.4      ad void
    716   1.4      ad softint_overlay(void)
    717   1.4      ad {
    718   1.5      ad 	struct cpu_info *ci;
    719  1.14      ad 	u_int softints, oflag;
    720   1.5      ad 	softint_t *si;
    721   1.6      ad 	pri_t obase;
    722   1.5      ad 	lwp_t *l;
    723   1.5      ad 	int s;
    724   1.5      ad 
    725   1.5      ad 	l = curlwp;
    726  1.31   rmind 	KASSERT((l->l_pflag & LP_INTR) == 0);
    727  1.31   rmind 
    728  1.31   rmind 	/*
    729  1.31   rmind 	 * Arrange to elevate priority if the LWP blocks.  Also, bind LWP
    730  1.31   rmind 	 * to the CPU.  Note: disable kernel preemption before doing that.
    731  1.31   rmind 	 */
    732  1.31   rmind 	s = splhigh();
    733   1.5      ad 	ci = l->l_cpu;
    734   1.5      ad 	si = ((softcpu_t *)ci->ci_data.cpu_softcpu)->sc_int;
    735   1.5      ad 
    736   1.6      ad 	obase = l->l_kpribase;
    737   1.6      ad 	l->l_kpribase = PRI_KERNEL_RT;
    738  1.14      ad 	oflag = l->l_pflag;
    739  1.14      ad 	l->l_pflag = oflag | LP_INTR | LP_BOUND;
    740  1.31   rmind 
    741   1.5      ad 	while ((softints = ci->ci_data.cpu_softints) != 0) {
    742   1.5      ad 		if ((softints & (1 << SOFTINT_SERIAL)) != 0) {
    743   1.5      ad 			ci->ci_data.cpu_softints &= ~(1 << SOFTINT_SERIAL);
    744   1.5      ad 			softint_execute(&si[SOFTINT_SERIAL], l, s);
    745   1.5      ad 			continue;
    746   1.5      ad 		}
    747   1.5      ad 		if ((softints & (1 << SOFTINT_NET)) != 0) {
    748   1.5      ad 			ci->ci_data.cpu_softints &= ~(1 << SOFTINT_NET);
    749   1.5      ad 			softint_execute(&si[SOFTINT_NET], l, s);
    750   1.5      ad 			continue;
    751   1.5      ad 		}
    752   1.5      ad 		if ((softints & (1 << SOFTINT_BIO)) != 0) {
    753   1.5      ad 			ci->ci_data.cpu_softints &= ~(1 << SOFTINT_BIO);
    754   1.5      ad 			softint_execute(&si[SOFTINT_BIO], l, s);
    755   1.5      ad 			continue;
    756   1.5      ad 		}
    757   1.5      ad 		if ((softints & (1 << SOFTINT_CLOCK)) != 0) {
    758   1.5      ad 			ci->ci_data.cpu_softints &= ~(1 << SOFTINT_CLOCK);
    759   1.5      ad 			softint_execute(&si[SOFTINT_CLOCK], l, s);
    760   1.5      ad 			continue;
    761   1.5      ad 		}
    762   1.5      ad 	}
    763  1.15      ad 	l->l_pflag = oflag;
    764  1.14      ad 	l->l_kpribase = obase;
    765   1.5      ad 	splx(s);
    766   1.4      ad }
    767   1.5      ad 
    768   1.5      ad #else	/*  !__HAVE_FAST_SOFTINTS */
    769   1.5      ad 
    770   1.5      ad /*
    771   1.5      ad  * softint_thread:
    772   1.5      ad  *
    773   1.5      ad  *	Fast path: the LWP is switched to without restoring any state,
    774   1.5      ad  *	so we should not arrive here - there is a direct handoff between
    775   1.5      ad  *	the interrupt stub and softint_dispatch().
    776   1.5      ad  */
    777   1.5      ad void
    778   1.5      ad softint_thread(void *cookie)
    779   1.5      ad {
    780   1.5      ad 
    781   1.5      ad 	panic("softint_thread");
    782   1.5      ad }
    783   1.5      ad 
    784   1.5      ad /*
    785   1.5      ad  * softint_dispatch:
    786   1.5      ad  *
    787   1.5      ad  *	Fast path: entry point from machine-dependent code.
    788   1.5      ad  */
    789   1.5      ad void
    790   1.5      ad softint_dispatch(lwp_t *pinned, int s)
    791   1.5      ad {
    792   1.9    yamt 	struct bintime now;
    793   1.5      ad 	softint_t *si;
    794   1.5      ad 	u_int timing;
    795   1.5      ad 	lwp_t *l;
    796   1.5      ad 
    797  1.29    yamt 	KASSERT((pinned->l_pflag & LP_RUNNING) != 0);
    798   1.5      ad 	l = curlwp;
    799   1.5      ad 	si = l->l_private;
    800   1.5      ad 
    801   1.5      ad 	/*
    802   1.5      ad 	 * Note the interrupted LWP, and mark the current LWP as running
    803   1.5      ad 	 * before proceeding.  Although this must as a rule be done with
    804   1.5      ad 	 * the LWP locked, at this point no external agents will want to
    805   1.5      ad 	 * modify the interrupt LWP's state.
    806   1.5      ad 	 */
    807  1.22      ad 	timing = (softint_timing ? LP_TIMEINTR : 0);
    808   1.5      ad 	l->l_switchto = pinned;
    809   1.5      ad 	l->l_stat = LSONPROC;
    810  1.22      ad 	l->l_pflag |= (LP_RUNNING | timing);
    811   1.5      ad 
    812   1.5      ad 	/*
    813   1.5      ad 	 * Dispatch the interrupt.  If softints are being timed, charge
    814   1.5      ad 	 * for it.
    815   1.5      ad 	 */
    816   1.5      ad 	if (timing)
    817  1.11    yamt 		binuptime(&l->l_stime);
    818   1.5      ad 	softint_execute(si, l, s);
    819   1.5      ad 	if (timing) {
    820  1.11    yamt 		binuptime(&now);
    821   1.5      ad 		updatertime(l, &now);
    822  1.22      ad 		l->l_pflag &= ~LP_TIMEINTR;
    823   1.5      ad 	}
    824   1.5      ad 
    825   1.5      ad 	/*
    826   1.5      ad 	 * If we blocked while handling the interrupt, the pinned LWP is
    827   1.5      ad 	 * gone so switch to the idle LWP.  It will select a new LWP to
    828   1.5      ad 	 * run.
    829   1.5      ad 	 *
    830   1.5      ad 	 * We must drop the priority level as switching at IPL_HIGH could
    831   1.5      ad 	 * deadlock the system.  We have already set si->si_active = 0,
    832   1.5      ad 	 * which means another interrupt at this level can be triggered.
    833   1.5      ad 	 * That's not be a problem: we are lowering to level 's' which will
    834   1.5      ad 	 * prevent softint_dispatch() from being reentered at level 's',
    835   1.5      ad 	 * until the priority is finally dropped to IPL_NONE on entry to
    836  1.21      ad 	 * the LWP chosen by lwp_exit_switchaway().
    837   1.5      ad 	 */
    838   1.5      ad 	l->l_stat = LSIDL;
    839   1.5      ad 	if (l->l_switchto == NULL) {
    840   1.5      ad 		splx(s);
    841   1.5      ad 		pmap_deactivate(l);
    842   1.5      ad 		lwp_exit_switchaway(l);
    843   1.5      ad 		/* NOTREACHED */
    844   1.5      ad 	}
    845   1.5      ad 	l->l_switchto = NULL;
    846  1.22      ad 	l->l_pflag &= ~LP_RUNNING;
    847   1.5      ad }
    848   1.5      ad 
    849   1.5      ad #endif	/* !__HAVE_FAST_SOFTINTS */
    850