Home | History | Annotate | Line # | Download | only in kern
kern_sleepq.c revision 1.28.2.1
      1 /*	$NetBSD: kern_sleepq.c,v 1.28.2.1 2008/05/23 05:24:16 wrstuden Exp $	*/
      2 
      3 /*-
      4  * Copyright (c) 2006, 2007, 2008 The NetBSD Foundation, Inc.
      5  * All rights reserved.
      6  *
      7  * This code is derived from software contributed to The NetBSD Foundation
      8  * by Andrew Doran.
      9  *
     10  * Redistribution and use in source and binary forms, with or without
     11  * modification, are permitted provided that the following conditions
     12  * are met:
     13  * 1. Redistributions of source code must retain the above copyright
     14  *    notice, this list of conditions and the following disclaimer.
     15  * 2. Redistributions in binary form must reproduce the above copyright
     16  *    notice, this list of conditions and the following disclaimer in the
     17  *    documentation and/or other materials provided with the distribution.
     18  *
     19  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
     20  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
     21  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
     22  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
     23  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
     24  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
     25  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
     26  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
     27  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
     28  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
     29  * POSSIBILITY OF SUCH DAMAGE.
     30  */
     31 
     32 /*
     33  * Sleep queue implementation, used by turnstiles and general sleep/wakeup
     34  * interfaces.
     35  */
     36 
     37 #include <sys/cdefs.h>
     38 __KERNEL_RCSID(0, "$NetBSD: kern_sleepq.c,v 1.28.2.1 2008/05/23 05:24:16 wrstuden Exp $");
     39 
     40 #include <sys/param.h>
     41 #include <sys/kernel.h>
     42 #include <sys/cpu.h>
     43 #include <sys/pool.h>
     44 #include <sys/proc.h>
     45 #include <sys/resourcevar.h>
     46 #include <sys/sa.h>
     47 #include <sys/savar.h>
     48 #include <sys/sched.h>
     49 #include <sys/systm.h>
     50 #include <sys/sleepq.h>
     51 #include <sys/ktrace.h>
     52 
     53 #include <uvm/uvm_extern.h>
     54 
     55 int	sleepq_sigtoerror(lwp_t *, int);
     56 
     57 /* General purpose sleep table, used by ltsleep() and condition variables. */
     58 sleeptab_t	sleeptab;
     59 
     60 /*
     61  * sleeptab_init:
     62  *
     63  *	Initialize a sleep table.
     64  */
     65 void
     66 sleeptab_init(sleeptab_t *st)
     67 {
     68 	sleepq_t *sq;
     69 	int i;
     70 
     71 	for (i = 0; i < SLEEPTAB_HASH_SIZE; i++) {
     72 		sq = &st->st_queues[i].st_queue;
     73 		mutex_init(&st->st_queues[i].st_mutex, MUTEX_DEFAULT,
     74 		    IPL_SCHED);
     75 		sleepq_init(sq, &st->st_queues[i].st_mutex);
     76 	}
     77 }
     78 
     79 /*
     80  * sleepq_init:
     81  *
     82  *	Prepare a sleep queue for use.
     83  */
     84 void
     85 sleepq_init(sleepq_t *sq, kmutex_t *mtx)
     86 {
     87 
     88 	sq->sq_waiters = 0;
     89 	sq->sq_mutex = mtx;
     90 	TAILQ_INIT(&sq->sq_queue);
     91 }
     92 
     93 /*
     94  * sleepq_remove:
     95  *
     96  *	Remove an LWP from a sleep queue and wake it up.  Return non-zero if
     97  *	the LWP is swapped out; if so the caller needs to awaken the swapper
     98  *	to bring the LWP into memory.
     99  */
    100 int
    101 sleepq_remove(sleepq_t *sq, lwp_t *l)
    102 {
    103 	struct schedstate_percpu *spc;
    104 	struct cpu_info *ci;
    105 
    106 	KASSERT(lwp_locked(l, sq->sq_mutex));
    107 	KASSERT(sq->sq_waiters > 0);
    108 
    109 	sq->sq_waiters--;
    110 	TAILQ_REMOVE(&sq->sq_queue, l, l_sleepchain);
    111 
    112 #ifdef DIAGNOSTIC
    113 	if (sq->sq_waiters == 0)
    114 		KASSERT(TAILQ_FIRST(&sq->sq_queue) == NULL);
    115 	else
    116 		KASSERT(TAILQ_FIRST(&sq->sq_queue) != NULL);
    117 #endif
    118 
    119 	l->l_syncobj = &sched_syncobj;
    120 	l->l_wchan = NULL;
    121 	l->l_sleepq = NULL;
    122 	l->l_flag &= ~LW_SINTR;
    123 
    124 	ci = l->l_cpu;
    125 	spc = &ci->ci_schedstate;
    126 
    127 	/*
    128 	 * If not sleeping, the LWP must have been suspended.  Let whoever
    129 	 * holds it stopped set it running again.
    130 	 */
    131 	if (l->l_stat != LSSLEEP) {
    132 		KASSERT(l->l_stat == LSSTOP || l->l_stat == LSSUSPENDED);
    133 		lwp_setlock(l, spc->spc_lwplock);
    134 		return 0;
    135 	}
    136 
    137 	/*
    138 	 * If the LWP is still on the CPU, mark it as LSONPROC.  It may be
    139 	 * about to call mi_switch(), in which case it will yield.
    140 	 */
    141 	if ((l->l_flag & LW_RUNNING) != 0) {
    142 		l->l_stat = LSONPROC;
    143 		l->l_slptime = 0;
    144 		lwp_setlock(l, spc->spc_lwplock);
    145 		return 0;
    146 	}
    147 
    148 	/*
    149 	 * Call the wake-up handler of scheduler.
    150 	 * It might change the CPU for this thread.
    151 	 */
    152 	sched_wakeup(l);
    153 	ci = l->l_cpu;
    154 	spc = &ci->ci_schedstate;
    155 
    156 	/*
    157 	 * Set it running.
    158 	 */
    159 	spc_lock(ci);
    160 	lwp_setlock(l, spc->spc_mutex);
    161 	if (l->l_proc->p_sa != NULL)
    162 		sa_awaken(l);
    163 	sched_setrunnable(l);
    164 	l->l_stat = LSRUN;
    165 	l->l_slptime = 0;
    166 	if ((l->l_flag & LW_INMEM) != 0) {
    167 		sched_enqueue(l, false);
    168 		spc_unlock(ci);
    169 		return 0;
    170 	}
    171 	spc_unlock(ci);
    172 	return 1;
    173 }
    174 
    175 /*
    176  * sleepq_insert:
    177  *
    178  *	Insert an LWP into the sleep queue, optionally sorting by priority.
    179  */
    180 inline void
    181 sleepq_insert(sleepq_t *sq, lwp_t *l, syncobj_t *sobj)
    182 {
    183 	lwp_t *l2;
    184 	const int pri = lwp_eprio(l);
    185 
    186 	if ((sobj->sobj_flag & SOBJ_SLEEPQ_SORTED) != 0) {
    187 		TAILQ_FOREACH(l2, &sq->sq_queue, l_sleepchain) {
    188 			if (lwp_eprio(l2) < pri) {
    189 				TAILQ_INSERT_BEFORE(l2, l, l_sleepchain);
    190 				return;
    191 			}
    192 		}
    193 	}
    194 
    195 	if ((sobj->sobj_flag & SOBJ_SLEEPQ_LIFO) != 0)
    196 		TAILQ_INSERT_HEAD(&sq->sq_queue, l, l_sleepchain);
    197 	else
    198 		TAILQ_INSERT_TAIL(&sq->sq_queue, l, l_sleepchain);
    199 }
    200 
    201 /*
    202  * sleepq_enqueue:
    203  *
    204  *	Enter an LWP into the sleep queue and prepare for sleep.  The sleep
    205  *	queue must already be locked, and any interlock (such as the kernel
    206  *	lock) must have be released (see sleeptab_lookup(), sleepq_enter()).
    207  */
    208 void
    209 sleepq_enqueue(sleepq_t *sq, wchan_t wchan, const char *wmesg, syncobj_t *sobj)
    210 {
    211 	lwp_t *l = curlwp;
    212 
    213 	KASSERT(lwp_locked(l, sq->sq_mutex));
    214 	KASSERT(l->l_stat == LSONPROC);
    215 	KASSERT(l->l_wchan == NULL && l->l_sleepq == NULL);
    216 
    217 	l->l_syncobj = sobj;
    218 	l->l_wchan = wchan;
    219 	l->l_sleepq = sq;
    220 	l->l_wmesg = wmesg;
    221 	l->l_slptime = 0;
    222 	l->l_stat = LSSLEEP;
    223 	l->l_sleeperr = 0;
    224 
    225 	sq->sq_waiters++;
    226 	sleepq_insert(sq, l, sobj);
    227 	sched_slept(l);
    228 }
    229 
    230 /*
    231  * sleepq_block:
    232  *
    233  *	After any intermediate step such as releasing an interlock, switch.
    234  * 	sleepq_block() may return early under exceptional conditions, for
    235  * 	example if the LWP's containing process is exiting.
    236  */
    237 int
    238 sleepq_block(int timo, bool catch)
    239 {
    240 	int error = 0, sig;
    241 	struct proc *p;
    242 	lwp_t *l = curlwp;
    243 	bool early = false;
    244 
    245 	ktrcsw(1, 0);
    246 
    247 	/*
    248 	 * If sleeping interruptably, check for pending signals, exits or
    249 	 * core dump events.
    250 	 */
    251 	if (catch) {
    252 		l->l_flag |= LW_SINTR;
    253 		if ((l->l_flag & (LW_CANCELLED|LW_WEXIT|LW_WCORE)) != 0) {
    254 			l->l_flag &= ~LW_CANCELLED;
    255 			error = EINTR;
    256 			early = true;
    257 		} else if ((l->l_flag & LW_PENDSIG) != 0 && sigispending(l, 0))
    258 			early = true;
    259 	}
    260 
    261 	if (early) {
    262 		/* lwp_unsleep() will release the lock */
    263 		lwp_unsleep(l, true);
    264 	} else {
    265 		if (timo)
    266 			callout_schedule(&l->l_timeout_ch, timo);
    267 
    268 		if ((l->l_flag & LW_SA) != 0)
    269 			sa_switch(l);
    270 		else
    271 			mi_switch(l);
    272 
    273 		/* The LWP and sleep queue are now unlocked. */
    274 		if (timo) {
    275 			/*
    276 			 * Even if the callout appears to have fired, we need to
    277 			 * stop it in order to synchronise with other CPUs.
    278 			 */
    279 			if (callout_halt(&l->l_timeout_ch, NULL))
    280 				error = EWOULDBLOCK;
    281 		}
    282 	}
    283 
    284 	if (catch && error == 0) {
    285 		p = l->l_proc;
    286 		if ((l->l_flag & (LW_CANCELLED | LW_WEXIT | LW_WCORE)) != 0)
    287 			error = EINTR;
    288 		else if ((l->l_flag & LW_PENDSIG) != 0) {
    289 			mutex_enter(p->p_lock);
    290 			if ((sig = issignal(l)) != 0)
    291 				error = sleepq_sigtoerror(l, sig);
    292 			mutex_exit(p->p_lock);
    293 		}
    294 	}
    295 
    296 	ktrcsw(0, 0);
    297 
    298 	KERNEL_LOCK(l->l_biglocks, l);
    299 	return error;
    300 }
    301 
    302 /*
    303  * sleepq_wake:
    304  *
    305  *	Wake zero or more LWPs blocked on a single wait channel.
    306  */
    307 lwp_t *
    308 sleepq_wake(sleepq_t *sq, wchan_t wchan, u_int expected)
    309 {
    310 	lwp_t *l, *next;
    311 	int swapin = 0;
    312 
    313 	KASSERT(mutex_owned(sq->sq_mutex));
    314 
    315 	for (l = TAILQ_FIRST(&sq->sq_queue); l != NULL; l = next) {
    316 		KASSERT(l->l_sleepq == sq);
    317 		KASSERT(l->l_mutex == sq->sq_mutex);
    318 		next = TAILQ_NEXT(l, l_sleepchain);
    319 		if (l->l_wchan != wchan)
    320 			continue;
    321 		swapin |= sleepq_remove(sq, l);
    322 		if (--expected == 0)
    323 			break;
    324 	}
    325 
    326 	sleepq_unlock(sq);
    327 
    328 	/*
    329 	 * If there are newly awakend threads that need to be swapped in,
    330 	 * then kick the swapper into action.
    331 	 */
    332 	if (swapin)
    333 		uvm_kick_scheduler();
    334 
    335 	return l;
    336 }
    337 
    338 /*
    339  * sleepq_unsleep:
    340  *
    341  *	Remove an LWP from its sleep queue and set it runnable again.
    342  *	sleepq_unsleep() is called with the LWP's mutex held, and will
    343  *	always release it.
    344  */
    345 u_int
    346 sleepq_unsleep(lwp_t *l, bool cleanup)
    347 {
    348 	sleepq_t *sq = l->l_sleepq;
    349 	int swapin;
    350 
    351 	KASSERT(lwp_locked(l, sq->sq_mutex));
    352 	KASSERT(l->l_wchan != NULL);
    353 
    354 	swapin = sleepq_remove(sq, l);
    355 
    356 	if (cleanup) {
    357 		sleepq_unlock(sq);
    358 		if (swapin)
    359 			uvm_kick_scheduler();
    360 	}
    361 
    362 	return swapin;
    363 }
    364 
    365 /*
    366  * sleepq_timeout:
    367  *
    368  *	Entered via the callout(9) subsystem to time out an LWP that is on a
    369  *	sleep queue.
    370  */
    371 void
    372 sleepq_timeout(void *arg)
    373 {
    374 	lwp_t *l = arg;
    375 
    376 	/*
    377 	 * Lock the LWP.  Assuming it's still on the sleep queue, its
    378 	 * current mutex will also be the sleep queue mutex.
    379 	 */
    380 	lwp_lock(l);
    381 
    382 	if (l->l_wchan == NULL) {
    383 		/* Somebody beat us to it. */
    384 		lwp_unlock(l);
    385 		return;
    386 	}
    387 
    388 	lwp_unsleep(l, true);
    389 }
    390 
    391 /*
    392  * sleepq_sigtoerror:
    393  *
    394  *	Given a signal number, interpret and return an error code.
    395  */
    396 int
    397 sleepq_sigtoerror(lwp_t *l, int sig)
    398 {
    399 	struct proc *p = l->l_proc;
    400 	int error;
    401 
    402 	KASSERT(mutex_owned(p->p_lock));
    403 
    404 	/*
    405 	 * If this sleep was canceled, don't let the syscall restart.
    406 	 */
    407 	if ((SIGACTION(p, sig).sa_flags & SA_RESTART) == 0)
    408 		error = EINTR;
    409 	else
    410 		error = ERESTART;
    411 
    412 	return error;
    413 }
    414 
    415 /*
    416  * sleepq_abort:
    417  *
    418  *	After a panic or during autoconfiguration, lower the interrupt
    419  *	priority level to give pending interrupts a chance to run, and
    420  *	then return.  Called if sleepq_dontsleep() returns non-zero, and
    421  *	always returns zero.
    422  */
    423 int
    424 sleepq_abort(kmutex_t *mtx, int unlock)
    425 {
    426 	extern int safepri;
    427 	int s;
    428 
    429 	s = splhigh();
    430 	splx(safepri);
    431 	splx(s);
    432 	if (mtx != NULL && unlock != 0)
    433 		mutex_exit(mtx);
    434 
    435 	return 0;
    436 }
    437 
    438 /*
    439  * sleepq_changepri:
    440  *
    441  *	Adjust the priority of an LWP residing on a sleepq.  This method
    442  *	will only alter the user priority; the effective priority is
    443  *	assumed to have been fixed at the time of insertion into the queue.
    444  */
    445 void
    446 sleepq_changepri(lwp_t *l, pri_t pri)
    447 {
    448 	sleepq_t *sq = l->l_sleepq;
    449 	pri_t opri;
    450 
    451 	KASSERT(lwp_locked(l, sq->sq_mutex));
    452 
    453 	opri = lwp_eprio(l);
    454 	l->l_priority = pri;
    455 	if (lwp_eprio(l) != opri) {
    456 		TAILQ_REMOVE(&sq->sq_queue, l, l_sleepchain);
    457 		sleepq_insert(sq, l, l->l_syncobj);
    458 	}
    459 }
    460 
    461 void
    462 sleepq_lendpri(lwp_t *l, pri_t pri)
    463 {
    464 	sleepq_t *sq = l->l_sleepq;
    465 	pri_t opri;
    466 
    467 	KASSERT(lwp_locked(l, sq->sq_mutex));
    468 
    469 	opri = lwp_eprio(l);
    470 	l->l_inheritedprio = pri;
    471 
    472 	if (lwp_eprio(l) != opri &&
    473 	    (l->l_syncobj->sobj_flag & SOBJ_SLEEPQ_SORTED) != 0) {
    474 		TAILQ_REMOVE(&sq->sq_queue, l, l_sleepchain);
    475 		sleepq_insert(sq, l, l->l_syncobj);
    476 	}
    477 }
    478