Home | History | Annotate | Line # | Download | only in kern
sched_4bsd.c revision 1.1.2.23
      1 /*	$NetBSD: sched_4bsd.c,v 1.1.2.23 2007/04/02 00:28:09 rmind Exp $	*/
      2 
      3 /*-
      4  * Copyright (c) 1999, 2000, 2004, 2006, 2007 The NetBSD Foundation, Inc.
      5  * All rights reserved.
      6  *
      7  * This code is derived from software contributed to The NetBSD Foundation
      8  * by Jason R. Thorpe of the Numerical Aerospace Simulation Facility,
      9  * NASA Ames Research Center, by Charles M. Hannum, Andrew Doran, and
     10  * Daniel Sieger.
     11  *
     12  * Redistribution and use in source and binary forms, with or without
     13  * modification, are permitted provided that the following conditions
     14  * are met:
     15  * 1. Redistributions of source code must retain the above copyright
     16  *    notice, this list of conditions and the following disclaimer.
     17  * 2. Redistributions in binary form must reproduce the above copyright
     18  *    notice, this list of conditions and the following disclaimer in the
     19  *    documentation and/or other materials provided with the distribution.
     20  * 3. All advertising materials mentioning features or use of this software
     21  *    must display the following acknowledgement:
     22  *	This product includes software developed by the NetBSD
     23  *	Foundation, Inc. and its contributors.
     24  * 4. Neither the name of The NetBSD Foundation nor the names of its
     25  *    contributors may be used to endorse or promote products derived
     26  *    from this software without specific prior written permission.
     27  *
     28  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
     29  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
     30  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
     31  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
     32  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
     33  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
     34  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
     35  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
     36  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
     37  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
     38  * POSSIBILITY OF SUCH DAMAGE.
     39  */
     40 
     41 /*-
     42  * Copyright (c) 1982, 1986, 1990, 1991, 1993
     43  *	The Regents of the University of California.  All rights reserved.
     44  * (c) UNIX System Laboratories, Inc.
     45  * All or some portions of this file are derived from material licensed
     46  * to the University of California by American Telephone and Telegraph
     47  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
     48  * the permission of UNIX System Laboratories, Inc.
     49  *
     50  * Redistribution and use in source and binary forms, with or without
     51  * modification, are permitted provided that the following conditions
     52  * are met:
     53  * 1. Redistributions of source code must retain the above copyright
     54  *    notice, this list of conditions and the following disclaimer.
     55  * 2. Redistributions in binary form must reproduce the above copyright
     56  *    notice, this list of conditions and the following disclaimer in the
     57  *    documentation and/or other materials provided with the distribution.
     58  * 3. Neither the name of the University nor the names of its contributors
     59  *    may be used to endorse or promote products derived from this software
     60  *    without specific prior written permission.
     61  *
     62  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
     63  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
     64  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
     65  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
     66  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
     67  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
     68  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
     69  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
     70  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
     71  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
     72  * SUCH DAMAGE.
     73  *
     74  *	@(#)kern_synch.c	8.9 (Berkeley) 5/19/95
     75  */
     76 
     77 #include <sys/cdefs.h>
     78 __KERNEL_RCSID(0, "$NetBSD: sched_4bsd.c,v 1.1.2.23 2007/04/02 00:28:09 rmind Exp $");
     79 
     80 #include "opt_ddb.h"
     81 #include "opt_lockdebug.h"
     82 #include "opt_perfctrs.h"
     83 
     84 #define	__MUTEX_PRIVATE
     85 
     86 #include <sys/param.h>
     87 #include <sys/systm.h>
     88 #include <sys/callout.h>
     89 #include <sys/cpu.h>
     90 #include <sys/proc.h>
     91 #include <sys/kernel.h>
     92 #include <sys/signalvar.h>
     93 #include <sys/resourcevar.h>
     94 #include <sys/sched.h>
     95 #include <sys/sysctl.h>
     96 #include <sys/kauth.h>
     97 #include <sys/lockdebug.h>
     98 
     99 #include <uvm/uvm_extern.h>
    100 
    101 /*
    102  * Run queues.
    103  *
    104  * We have 32 run queues in descending priority of 0..31.  We maintain
    105  * a bitmask of non-empty queues in order speed up finding the first
    106  * runnable process.  The bitmask is maintained only by machine-dependent
    107  * code, allowing the most efficient instructions to be used to find the
    108  * first non-empty queue.
    109  */
    110 
    111 #define	RUNQUE_NQS		32      /* number of runqueues */
    112 #define	PPQ	(128 / RUNQUE_NQS)	/* priorities per queue */
    113 
    114 typedef struct subqueue {
    115 	TAILQ_HEAD(, lwp) sq_queue;
    116 } subqueue_t;
    117 typedef struct runqueue {
    118 	subqueue_t rq_subqueues[RUNQUE_NQS];	/* run queues */
    119 	uint32_t rq_bitmap;	/* bitmap of non-empty queues */
    120 } runqueue_t;
    121 static runqueue_t global_queue;
    122 
    123 static void updatepri(struct lwp *);
    124 static void resetpriority(struct lwp *);
    125 static void resetprocpriority(struct proc *);
    126 
    127 extern unsigned int sched_pstats_ticks; /* defined in kern_synch.c */
    128 
    129 /* The global scheduler state */
    130 kmutex_t sched_mutex;
    131 
    132 /* Number of hardclock ticks per sched_tick() */
    133 int rrticks;
    134 
    135 /*
    136  * Force switch among equal priority processes every 100ms.
    137  * Called from hardclock every hz/10 == rrticks hardclock ticks.
    138  */
    139 /* ARGSUSED */
    140 void
    141 sched_tick(struct cpu_info *ci)
    142 {
    143 	struct schedstate_percpu *spc = &ci->ci_schedstate;
    144 
    145 	spc->spc_ticks = rrticks;
    146 
    147 	if (!CURCPU_IDLE_P()) {
    148 		if (spc->spc_flags & SPCF_SEENRR) {
    149 			/*
    150 			 * The process has already been through a roundrobin
    151 			 * without switching and may be hogging the CPU.
    152 			 * Indicate that the process should yield.
    153 			 */
    154 			spc->spc_flags |= SPCF_SHOULDYIELD;
    155 		} else
    156 			spc->spc_flags |= SPCF_SEENRR;
    157 	}
    158 	cpu_need_resched(curcpu(), 0);
    159 }
    160 
    161 #define	NICE_WEIGHT 2			/* priorities per nice level */
    162 
    163 #define	ESTCPU_SHIFT	11
    164 #define	ESTCPU_MAX	((NICE_WEIGHT * PRIO_MAX - PPQ) << ESTCPU_SHIFT)
    165 #define	ESTCPULIM(e)	min((e), ESTCPU_MAX)
    166 
    167 /*
    168  * Constants for digital decay and forget:
    169  *	90% of (p_estcpu) usage in 5 * loadav time
    170  *	95% of (p_pctcpu) usage in 60 seconds (load insensitive)
    171  *          Note that, as ps(1) mentions, this can let percentages
    172  *          total over 100% (I've seen 137.9% for 3 processes).
    173  *
    174  * Note that hardclock updates p_estcpu and p_cpticks independently.
    175  *
    176  * We wish to decay away 90% of p_estcpu in (5 * loadavg) seconds.
    177  * That is, the system wants to compute a value of decay such
    178  * that the following for loop:
    179  * 	for (i = 0; i < (5 * loadavg); i++)
    180  * 		p_estcpu *= decay;
    181  * will compute
    182  * 	p_estcpu *= 0.1;
    183  * for all values of loadavg:
    184  *
    185  * Mathematically this loop can be expressed by saying:
    186  * 	decay ** (5 * loadavg) ~= .1
    187  *
    188  * The system computes decay as:
    189  * 	decay = (2 * loadavg) / (2 * loadavg + 1)
    190  *
    191  * We wish to prove that the system's computation of decay
    192  * will always fulfill the equation:
    193  * 	decay ** (5 * loadavg) ~= .1
    194  *
    195  * If we compute b as:
    196  * 	b = 2 * loadavg
    197  * then
    198  * 	decay = b / (b + 1)
    199  *
    200  * We now need to prove two things:
    201  *	1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1)
    202  *	2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg)
    203  *
    204  * Facts:
    205  *         For x close to zero, exp(x) =~ 1 + x, since
    206  *              exp(x) = 0! + x**1/1! + x**2/2! + ... .
    207  *              therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b.
    208  *         For x close to zero, ln(1+x) =~ x, since
    209  *              ln(1+x) = x - x**2/2 + x**3/3 - ...     -1 < x < 1
    210  *              therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1).
    211  *         ln(.1) =~ -2.30
    212  *
    213  * Proof of (1):
    214  *    Solve (factor)**(power) =~ .1 given power (5*loadav):
    215  *	solving for factor,
    216  *      ln(factor) =~ (-2.30/5*loadav), or
    217  *      factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) =
    218  *          exp(-1/b) =~ (b-1)/b =~ b/(b+1).                    QED
    219  *
    220  * Proof of (2):
    221  *    Solve (factor)**(power) =~ .1 given factor == (b/(b+1)):
    222  *	solving for power,
    223  *      power*ln(b/(b+1)) =~ -2.30, or
    224  *      power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav.  QED
    225  *
    226  * Actual power values for the implemented algorithm are as follows:
    227  *      loadav: 1       2       3       4
    228  *      power:  5.68    10.32   14.94   19.55
    229  */
    230 
    231 /* calculations for digital decay to forget 90% of usage in 5*loadav sec */
    232 #define	loadfactor(loadav)	(2 * (loadav))
    233 
    234 static fixpt_t
    235 decay_cpu(fixpt_t loadfac, fixpt_t estcpu)
    236 {
    237 
    238 	if (estcpu == 0) {
    239 		return 0;
    240 	}
    241 
    242 #if !defined(_LP64)
    243 	/* avoid 64bit arithmetics. */
    244 #define	FIXPT_MAX ((fixpt_t)((UINTMAX_C(1) << sizeof(fixpt_t) * CHAR_BIT) - 1))
    245 	if (__predict_true(loadfac <= FIXPT_MAX / ESTCPU_MAX)) {
    246 		return estcpu * loadfac / (loadfac + FSCALE);
    247 	}
    248 #endif /* !defined(_LP64) */
    249 
    250 	return (uint64_t)estcpu * loadfac / (loadfac + FSCALE);
    251 }
    252 
    253 /*
    254  * For all load averages >= 1 and max p_estcpu of (255 << ESTCPU_SHIFT),
    255  * sleeping for at least seven times the loadfactor will decay p_estcpu to
    256  * less than (1 << ESTCPU_SHIFT).
    257  *
    258  * note that our ESTCPU_MAX is actually much smaller than (255 << ESTCPU_SHIFT).
    259  */
    260 static fixpt_t
    261 decay_cpu_batch(fixpt_t loadfac, fixpt_t estcpu, unsigned int n)
    262 {
    263 
    264 	if ((n << FSHIFT) >= 7 * loadfac) {
    265 		return 0;
    266 	}
    267 
    268 	while (estcpu != 0 && n > 1) {
    269 		estcpu = decay_cpu(loadfac, estcpu);
    270 		n--;
    271 	}
    272 
    273 	return estcpu;
    274 }
    275 
    276 /*
    277  * sched_pstats_hook:
    278  *
    279  * Periodically called from sched_pstats(); used to recalculate priorities.
    280  */
    281 inline void
    282 sched_pstats_hook(struct proc *p, int minslp)
    283 {
    284 	struct lwp *l;
    285 	fixpt_t loadfac = loadfactor(averunnable.ldavg[0]);
    286 
    287 	/*
    288 	 * If the process has slept the entire second,
    289 	 * stop recalculating its priority until it wakes up.
    290 	 */
    291 	if (minslp <= 1) {
    292 		p->p_estcpu = decay_cpu(loadfac, p->p_estcpu);
    293 
    294 		LIST_FOREACH(l, &p->p_lwps, l_sibling) {
    295 			if ((l->l_flag & LW_IDLE) != 0)
    296 				continue;
    297 			lwp_lock(l);
    298 			if (l->l_slptime <= 1 &&
    299 			    l->l_priority >= PUSER)
    300 				resetpriority(l);
    301 			lwp_unlock(l);
    302 		}
    303 	}
    304 }
    305 
    306 /*
    307  * Recalculate the priority of a process after it has slept for a while.
    308  */
    309 static void
    310 updatepri(struct lwp *l)
    311 {
    312 	struct proc *p = l->l_proc;
    313 	fixpt_t loadfac;
    314 
    315 	LOCK_ASSERT(lwp_locked(l, NULL));
    316 	KASSERT(l->l_slptime > 1);
    317 
    318 	loadfac = loadfactor(averunnable.ldavg[0]);
    319 
    320 	l->l_slptime--; /* the first time was done in sched_pstats */
    321 	/* XXX NJWLWP */
    322 	/* XXXSMP occasionally unlocked, should be per-LWP */
    323 	p->p_estcpu = decay_cpu_batch(loadfac, p->p_estcpu, l->l_slptime);
    324 	resetpriority(l);
    325 }
    326 
    327 /*
    328  * On some architectures, it's faster to use a MSB ordering for the priorites
    329  * than the traditional LSB ordering.
    330  */
    331 #ifdef __HAVE_BIGENDIAN_BITOPS
    332 #define	RQMASK(n) (0x80000000 >> (n))
    333 #else
    334 #define	RQMASK(n) (0x00000001 << (n))
    335 #endif
    336 
    337 /*
    338  * The primitives that manipulate the run queues.  whichqs tells which
    339  * of the 32 queues qs have processes in them.  sched_enqueue() puts processes
    340  * into queues, sched_dequeue removes them from queues.  The running process is
    341  * on no queue, other processes are on a queue related to p->p_priority,
    342  * divided by 4 actually to shrink the 0-127 range of priorities into the 32
    343  * available queues.
    344  */
    345 #ifdef RQDEBUG
    346 static void
    347 runqueue_check(const runqueue_t *rq, int whichq, struct lwp *l)
    348 {
    349 	const subqueue_t * const sq = &rq->rq_subqueues[whichq];
    350 	const uint32_t bitmap = rq->rq_bitmap;
    351 	struct lwp *l2;
    352 	int found = 0;
    353 	int die = 0;
    354 	int empty = 1;
    355 
    356 	TAILQ_FOREACH(l2, &sq->sq_queue, l_runq) {
    357 		if (l2->l_stat != LSRUN) {
    358 			printf("runqueue_check[%d]: lwp %p state (%d) "
    359 			    " != LSRUN\n", whichq, l2, l2->l_stat);
    360 		}
    361 		if (l2 == l)
    362 			found = 1;
    363 		empty = 0;
    364 	}
    365 	if (empty && (bitmap & RQMASK(whichq)) != 0) {
    366 		printf("runqueue_check[%d]: bit set for empty run-queue %p\n",
    367 		    whichq, rq);
    368 		die = 1;
    369 	} else if (!empty && (bitmap & RQMASK(whichq)) == 0) {
    370 		printf("runqueue_check[%d]: bit clear for non-empty "
    371 		    "run-queue %p\n", whichq, rq);
    372 		die = 1;
    373 	}
    374 	if (l != NULL && (bitmap & RQMASK(whichq)) == 0) {
    375 		printf("runqueue_check[%d]: bit clear for active lwp %p\n",
    376 		    whichq, l);
    377 		die = 1;
    378 	}
    379 	if (l != NULL && empty) {
    380 		printf("runqueue_check[%d]: empty run-queue %p with "
    381 		    "active lwp %p\n", whichq, rq, l);
    382 		die = 1;
    383 	}
    384 	if (l != NULL && !found) {
    385 		printf("runqueue_check[%d]: lwp %p not in runqueue %p!",
    386 		    whichq, l, rq);
    387 		die = 1;
    388 	}
    389 	if (die)
    390 		panic("runqueue_check: inconsistency found");
    391 }
    392 #else /* RQDEBUG */
    393 #define	runqueue_check(a, b, c)	/* nothing */
    394 #endif /* RQDEBUG */
    395 
    396 static void
    397 runqueue_init(runqueue_t *rq)
    398 {
    399 	int i;
    400 
    401 	for (i = 0; i < RUNQUE_NQS; i++)
    402 		TAILQ_INIT(&rq->rq_subqueues[i].sq_queue);
    403 }
    404 
    405 static void
    406 runqueue_enqueue(runqueue_t *rq, struct lwp *l)
    407 {
    408 	subqueue_t *sq;
    409 	const int whichq = lwp_eprio(l) / PPQ;
    410 
    411 	LOCK_ASSERT(lwp_locked(l, l->l_cpu->ci_schedstate.spc_mutex));
    412 
    413 	runqueue_check(rq, whichq, NULL);
    414 	rq->rq_bitmap |= RQMASK(whichq);
    415 	sq = &rq->rq_subqueues[whichq];
    416 	TAILQ_INSERT_TAIL(&sq->sq_queue, l, l_runq);
    417 	runqueue_check(rq, whichq, l);
    418 }
    419 
    420 static void
    421 runqueue_dequeue(runqueue_t *rq, struct lwp *l)
    422 {
    423 	subqueue_t *sq;
    424 	const int whichq = lwp_eprio(l) / PPQ;
    425 
    426 	LOCK_ASSERT(lwp_locked(l, l->l_cpu->ci_schedstate.spc_mutex));
    427 
    428 	runqueue_check(rq, whichq, l);
    429 	KASSERT((rq->rq_bitmap & RQMASK(whichq)) != 0);
    430 	sq = &rq->rq_subqueues[whichq];
    431 	TAILQ_REMOVE(&sq->sq_queue, l, l_runq);
    432 	if (TAILQ_EMPTY(&sq->sq_queue))
    433 		rq->rq_bitmap &= ~RQMASK(whichq);
    434 	runqueue_check(rq, whichq, NULL);
    435 }
    436 
    437 static struct lwp *
    438 runqueue_nextlwp(runqueue_t *rq)
    439 {
    440 	const uint32_t bitmap = rq->rq_bitmap;
    441 	int whichq;
    442 
    443 	if (bitmap == 0) {
    444 		return NULL;
    445 	}
    446 #ifdef __HAVE_BIGENDIAN_BITOPS
    447 	/* XXX should introduce a fast "fls" function. */
    448 	for (whichq = 0; ; whichq++) {
    449 		if ((bitmap & RQMASK(whichq)) != 0) {
    450 			break;
    451 		}
    452 	}
    453 #else
    454 	whichq = ffs(bitmap) - 1;
    455 #endif
    456 	return TAILQ_FIRST(&rq->rq_subqueues[whichq].sq_queue);
    457 }
    458 
    459 #if defined(DDB)
    460 static void
    461 runqueue_print(const runqueue_t *rq, void (*pr)(const char *, ...))
    462 {
    463 	const uint32_t bitmap = rq->rq_bitmap;
    464 	struct lwp *l;
    465 	int i, first;
    466 
    467 	for (i = 0; i < RUNQUE_NQS; i++) {
    468 		const subqueue_t *sq;
    469 		first = 1;
    470 		sq = &rq->rq_subqueues[i];
    471 		TAILQ_FOREACH(l, &sq->sq_queue, l_runq) {
    472 			if (first) {
    473 				(*pr)("%c%d",
    474 				    (bitmap & RQMASK(i)) ? ' ' : '!', i);
    475 				first = 0;
    476 			}
    477 			(*pr)("\t%d.%d (%s) pri=%d usrpri=%d\n",
    478 			    l->l_proc->p_pid,
    479 			    l->l_lid, l->l_proc->p_comm,
    480 			    (int)l->l_priority, (int)l->l_usrpri);
    481 		}
    482 	}
    483 }
    484 #endif /* defined(DDB) */
    485 #undef RQMASK
    486 
    487 /*
    488  * Initialize the (doubly-linked) run queues
    489  * to be empty.
    490  */
    491 void
    492 sched_rqinit()
    493 {
    494 
    495 	runqueue_init(&global_queue);
    496 	mutex_init(&sched_mutex, MUTEX_SPIN, IPL_SCHED);
    497 	/* Initialize the lock pointer for lwp0 */
    498 	lwp0.l_mutex = &sched_mutex;
    499 }
    500 
    501 void
    502 sched_cpuattach(struct cpu_info *ci)
    503 {
    504 
    505 	ci->ci_schedstate.spc_mutex = &sched_mutex;
    506 }
    507 
    508 void
    509 sched_setup()
    510 {
    511 
    512 	rrticks = hz / 10;
    513 	sched_pstats(NULL);
    514 }
    515 
    516 void
    517 sched_setrunnable(struct lwp *l)
    518 {
    519 
    520  	if (l->l_slptime > 1)
    521  		updatepri(l);
    522 }
    523 
    524 bool
    525 sched_curcpu_runnable_p(void)
    526 {
    527 
    528 	return global_queue.rq_bitmap != 0;
    529 }
    530 
    531 void
    532 sched_nice(struct proc *chgp, int n)
    533 {
    534 
    535 	chgp->p_nice = n;
    536 	(void)resetprocpriority(chgp);
    537 }
    538 
    539 /*
    540  * Compute the priority of a process when running in user mode.
    541  * Arrange to reschedule if the resulting priority is better
    542  * than that of the current process.
    543  */
    544 static void
    545 resetpriority(struct lwp *l)
    546 {
    547 	unsigned int newpriority;
    548 	struct proc *p = l->l_proc;
    549 
    550 	/* XXXSMP LOCK_ASSERT(mutex_owned(&p->p_stmutex)); */
    551 	LOCK_ASSERT(lwp_locked(l, NULL));
    552 
    553 	if ((l->l_flag & LW_SYSTEM) != 0)
    554 		return;
    555 
    556 	newpriority = PUSER + (p->p_estcpu >> ESTCPU_SHIFT) +
    557 	    NICE_WEIGHT * (p->p_nice - NZERO);
    558 	newpriority = min(newpriority, MAXPRI);
    559 	lwp_changepri(l, newpriority);
    560 }
    561 
    562 /*
    563  * Recompute priority for all LWPs in a process.
    564  */
    565 static void
    566 resetprocpriority(struct proc *p)
    567 {
    568 	struct lwp *l;
    569 
    570 	LOCK_ASSERT(mutex_owned(&p->p_stmutex));
    571 
    572 	LIST_FOREACH(l, &p->p_lwps, l_sibling) {
    573 		lwp_lock(l);
    574 		resetpriority(l);
    575 		lwp_unlock(l);
    576 	}
    577 }
    578 
    579 /*
    580  * We adjust the priority of the current process.  The priority of a process
    581  * gets worse as it accumulates CPU time.  The CPU usage estimator (p_estcpu)
    582  * is increased here.  The formula for computing priorities (in kern_synch.c)
    583  * will compute a different value each time p_estcpu increases. This can
    584  * cause a switch, but unless the priority crosses a PPQ boundary the actual
    585  * queue will not change.  The CPU usage estimator ramps up quite quickly
    586  * when the process is running (linearly), and decays away exponentially, at
    587  * a rate which is proportionally slower when the system is busy.  The basic
    588  * principle is that the system will 90% forget that the process used a lot
    589  * of CPU time in 5 * loadav seconds.  This causes the system to favor
    590  * processes which haven't run much recently, and to round-robin among other
    591  * processes.
    592  */
    593 
    594 void
    595 sched_schedclock(struct lwp *l)
    596 {
    597 	struct proc *p = l->l_proc;
    598 
    599 	KASSERT(!CURCPU_IDLE_P());
    600 	mutex_spin_enter(&p->p_stmutex);
    601 	p->p_estcpu = ESTCPULIM(p->p_estcpu + (1 << ESTCPU_SHIFT));
    602 	lwp_lock(l);
    603 	resetpriority(l);
    604 	mutex_spin_exit(&p->p_stmutex);
    605 	if ((l->l_flag & LW_SYSTEM) == 0 && l->l_priority >= PUSER)
    606 		l->l_priority = l->l_usrpri;
    607 	lwp_unlock(l);
    608 }
    609 
    610 /*
    611  * scheduler_fork_hook:
    612  *
    613  *	Inherit the parent's scheduler history.
    614  */
    615 void
    616 sched_proc_fork(struct proc *parent, struct proc *child)
    617 {
    618 
    619 	LOCK_ASSERT(mutex_owned(&parent->p_smutex));
    620 
    621 	child->p_estcpu = child->p_estcpu_inherited = parent->p_estcpu;
    622 	child->p_forktime = sched_pstats_ticks;
    623 }
    624 
    625 /*
    626  * scheduler_wait_hook:
    627  *
    628  *	Chargeback parents for the sins of their children.
    629  */
    630 void
    631 sched_proc_exit(struct proc *parent, struct proc *child)
    632 {
    633 	fixpt_t loadfac = loadfactor(averunnable.ldavg[0]);
    634 	fixpt_t estcpu;
    635 
    636 	/* XXX Only if parent != init?? */
    637 
    638 	mutex_spin_enter(&parent->p_stmutex);
    639 	estcpu = decay_cpu_batch(loadfac, child->p_estcpu_inherited,
    640 	    sched_pstats_ticks - child->p_forktime);
    641 	if (child->p_estcpu > estcpu)
    642 		parent->p_estcpu =
    643 		    ESTCPULIM(parent->p_estcpu + child->p_estcpu - estcpu);
    644 	mutex_spin_exit(&parent->p_stmutex);
    645 }
    646 
    647 void
    648 sched_enqueue(struct lwp *l, bool ctxswitch)
    649 {
    650 
    651 	runqueue_enqueue(&global_queue, l);
    652 }
    653 
    654 /*
    655  * XXXSMP When LWP dispatch (cpu_switch()) is changed to use sched_dequeue(),
    656  * drop of the effective priority level from kernel to user needs to be
    657  * moved here from userret().  The assignment in userret() is currently
    658  * done unlocked.
    659  */
    660 void
    661 sched_dequeue(struct lwp *l)
    662 {
    663 
    664 	runqueue_dequeue(&global_queue, l);
    665 }
    666 
    667 struct lwp *
    668 sched_nextlwp()
    669 {
    670 
    671 	return runqueue_nextlwp(&global_queue);
    672 }
    673 
    674 /* Dummy */
    675 void
    676 sched_lwp_fork(struct lwp *l)
    677 {
    678 
    679 }
    680 
    681 void
    682 sched_lwp_exit(struct lwp *l)
    683 {
    684 
    685 }
    686 
    687 void
    688 sched_slept(struct lwp *l)
    689 {
    690 
    691 }
    692 
    693 /* SysCtl */
    694 
    695 SYSCTL_SETUP(sysctl_sched_setup, "sysctl kern.sched subtree setup")
    696 {
    697 	sysctl_createv(clog, 0, NULL, NULL,
    698 		CTLFLAG_PERMANENT,
    699 		CTLTYPE_NODE, "kern", NULL,
    700 		NULL, 0, NULL, 0,
    701 		CTL_KERN, CTL_EOL);
    702 	sysctl_createv(clog, 0, NULL, NULL,
    703 		CTLFLAG_PERMANENT,
    704 		CTLTYPE_NODE, "sched",
    705 		SYSCTL_DESCR("Scheduler options"),
    706 		NULL, 0, NULL, 0,
    707 		CTL_KERN, KERN_SCHED, CTL_EOL);
    708 	sysctl_createv(clog, 0, NULL, NULL,
    709 		CTLFLAG_PERMANENT,
    710 		CTLTYPE_STRING, "name", NULL,
    711 		NULL, 0, __UNCONST("4.4BSD"), 0,
    712 		CTL_KERN, KERN_SCHED, CTL_CREATE, CTL_EOL);
    713 }
    714 
    715 #if defined(DDB)
    716 void
    717 sched_print_runqueue(void (*pr)(const char *, ...))
    718 {
    719 
    720 	runqueue_print(&global_queue, pr);
    721 }
    722 #endif /* defined(DDB) */
    723