sched_4bsd.c revision 1.1.6.1 1 1.1.6.1 ad /* $NetBSD: sched_4bsd.c,v 1.1.6.1 2007/06/08 14:17:24 ad Exp $ */
2 1.1.6.1 ad
3 1.1.6.1 ad /*-
4 1.1.6.1 ad * Copyright (c) 1999, 2000, 2004, 2006, 2007 The NetBSD Foundation, Inc.
5 1.1.6.1 ad * All rights reserved.
6 1.1.6.1 ad *
7 1.1.6.1 ad * This code is derived from software contributed to The NetBSD Foundation
8 1.1.6.1 ad * by Jason R. Thorpe of the Numerical Aerospace Simulation Facility,
9 1.1.6.1 ad * NASA Ames Research Center, by Charles M. Hannum, Andrew Doran, and
10 1.1.6.1 ad * Daniel Sieger.
11 1.1.6.1 ad *
12 1.1.6.1 ad * Redistribution and use in source and binary forms, with or without
13 1.1.6.1 ad * modification, are permitted provided that the following conditions
14 1.1.6.1 ad * are met:
15 1.1.6.1 ad * 1. Redistributions of source code must retain the above copyright
16 1.1.6.1 ad * notice, this list of conditions and the following disclaimer.
17 1.1.6.1 ad * 2. Redistributions in binary form must reproduce the above copyright
18 1.1.6.1 ad * notice, this list of conditions and the following disclaimer in the
19 1.1.6.1 ad * documentation and/or other materials provided with the distribution.
20 1.1.6.1 ad * 3. All advertising materials mentioning features or use of this software
21 1.1.6.1 ad * must display the following acknowledgement:
22 1.1.6.1 ad * This product includes software developed by the NetBSD
23 1.1.6.1 ad * Foundation, Inc. and its contributors.
24 1.1.6.1 ad * 4. Neither the name of The NetBSD Foundation nor the names of its
25 1.1.6.1 ad * contributors may be used to endorse or promote products derived
26 1.1.6.1 ad * from this software without specific prior written permission.
27 1.1.6.1 ad *
28 1.1.6.1 ad * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
29 1.1.6.1 ad * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
30 1.1.6.1 ad * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
31 1.1.6.1 ad * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
32 1.1.6.1 ad * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
33 1.1.6.1 ad * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
34 1.1.6.1 ad * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
35 1.1.6.1 ad * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
36 1.1.6.1 ad * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
37 1.1.6.1 ad * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
38 1.1.6.1 ad * POSSIBILITY OF SUCH DAMAGE.
39 1.1.6.1 ad */
40 1.1.6.1 ad
41 1.1.6.1 ad /*-
42 1.1.6.1 ad * Copyright (c) 1982, 1986, 1990, 1991, 1993
43 1.1.6.1 ad * The Regents of the University of California. All rights reserved.
44 1.1.6.1 ad * (c) UNIX System Laboratories, Inc.
45 1.1.6.1 ad * All or some portions of this file are derived from material licensed
46 1.1.6.1 ad * to the University of California by American Telephone and Telegraph
47 1.1.6.1 ad * Co. or Unix System Laboratories, Inc. and are reproduced herein with
48 1.1.6.1 ad * the permission of UNIX System Laboratories, Inc.
49 1.1.6.1 ad *
50 1.1.6.1 ad * Redistribution and use in source and binary forms, with or without
51 1.1.6.1 ad * modification, are permitted provided that the following conditions
52 1.1.6.1 ad * are met:
53 1.1.6.1 ad * 1. Redistributions of source code must retain the above copyright
54 1.1.6.1 ad * notice, this list of conditions and the following disclaimer.
55 1.1.6.1 ad * 2. Redistributions in binary form must reproduce the above copyright
56 1.1.6.1 ad * notice, this list of conditions and the following disclaimer in the
57 1.1.6.1 ad * documentation and/or other materials provided with the distribution.
58 1.1.6.1 ad * 3. Neither the name of the University nor the names of its contributors
59 1.1.6.1 ad * may be used to endorse or promote products derived from this software
60 1.1.6.1 ad * without specific prior written permission.
61 1.1.6.1 ad *
62 1.1.6.1 ad * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
63 1.1.6.1 ad * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
64 1.1.6.1 ad * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
65 1.1.6.1 ad * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
66 1.1.6.1 ad * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
67 1.1.6.1 ad * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
68 1.1.6.1 ad * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
69 1.1.6.1 ad * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
70 1.1.6.1 ad * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
71 1.1.6.1 ad * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
72 1.1.6.1 ad * SUCH DAMAGE.
73 1.1.6.1 ad *
74 1.1.6.1 ad * @(#)kern_synch.c 8.9 (Berkeley) 5/19/95
75 1.1.6.1 ad */
76 1.1.6.1 ad
77 1.1.6.1 ad #include <sys/cdefs.h>
78 1.1.6.1 ad __KERNEL_RCSID(0, "$NetBSD: sched_4bsd.c,v 1.1.6.1 2007/06/08 14:17:24 ad Exp $");
79 1.1.6.1 ad
80 1.1.6.1 ad #include "opt_ddb.h"
81 1.1.6.1 ad #include "opt_lockdebug.h"
82 1.1.6.1 ad #include "opt_perfctrs.h"
83 1.1.6.1 ad
84 1.1.6.1 ad #define __MUTEX_PRIVATE
85 1.1.6.1 ad
86 1.1.6.1 ad #include <sys/param.h>
87 1.1.6.1 ad #include <sys/systm.h>
88 1.1.6.1 ad #include <sys/callout.h>
89 1.1.6.1 ad #include <sys/cpu.h>
90 1.1.6.1 ad #include <sys/proc.h>
91 1.1.6.1 ad #include <sys/kernel.h>
92 1.1.6.1 ad #include <sys/signalvar.h>
93 1.1.6.1 ad #include <sys/resourcevar.h>
94 1.1.6.1 ad #include <sys/sched.h>
95 1.1.6.1 ad #include <sys/sysctl.h>
96 1.1.6.1 ad #include <sys/kauth.h>
97 1.1.6.1 ad #include <sys/lockdebug.h>
98 1.1.6.1 ad #include <sys/kmem.h>
99 1.1.6.1 ad
100 1.1.6.1 ad #include <uvm/uvm_extern.h>
101 1.1.6.1 ad
102 1.1.6.1 ad /*
103 1.1.6.1 ad * Run queues.
104 1.1.6.1 ad *
105 1.1.6.1 ad * We have 32 run queues in descending priority of 0..31. We maintain
106 1.1.6.1 ad * a bitmask of non-empty queues in order speed up finding the first
107 1.1.6.1 ad * runnable process. The bitmask is maintained only by machine-dependent
108 1.1.6.1 ad * code, allowing the most efficient instructions to be used to find the
109 1.1.6.1 ad * first non-empty queue.
110 1.1.6.1 ad */
111 1.1.6.1 ad
112 1.1.6.1 ad #define RUNQUE_NQS 32 /* number of runqueues */
113 1.1.6.1 ad #define PPQ (128 / RUNQUE_NQS) /* priorities per queue */
114 1.1.6.1 ad
115 1.1.6.1 ad typedef struct subqueue {
116 1.1.6.1 ad TAILQ_HEAD(, lwp) sq_queue;
117 1.1.6.1 ad } subqueue_t;
118 1.1.6.1 ad typedef struct runqueue {
119 1.1.6.1 ad subqueue_t rq_subqueues[RUNQUE_NQS]; /* run queues */
120 1.1.6.1 ad uint32_t rq_bitmap; /* bitmap of non-empty queues */
121 1.1.6.1 ad } runqueue_t;
122 1.1.6.1 ad static runqueue_t global_queue;
123 1.1.6.1 ad
124 1.1.6.1 ad static void updatepri(struct lwp *);
125 1.1.6.1 ad static void resetpriority(struct lwp *);
126 1.1.6.1 ad static void resetprocpriority(struct proc *);
127 1.1.6.1 ad
128 1.1.6.1 ad extern unsigned int sched_pstats_ticks; /* defined in kern_synch.c */
129 1.1.6.1 ad
130 1.1.6.1 ad /* The global scheduler state */
131 1.1.6.1 ad kmutex_t sched_mutex;
132 1.1.6.1 ad
133 1.1.6.1 ad /* Number of hardclock ticks per sched_tick() */
134 1.1.6.1 ad int rrticks;
135 1.1.6.1 ad
136 1.1.6.1 ad /*
137 1.1.6.1 ad * Force switch among equal priority processes every 100ms.
138 1.1.6.1 ad * Called from hardclock every hz/10 == rrticks hardclock ticks.
139 1.1.6.1 ad */
140 1.1.6.1 ad /* ARGSUSED */
141 1.1.6.1 ad void
142 1.1.6.1 ad sched_tick(struct cpu_info *ci)
143 1.1.6.1 ad {
144 1.1.6.1 ad struct schedstate_percpu *spc = &ci->ci_schedstate;
145 1.1.6.1 ad
146 1.1.6.1 ad spc->spc_ticks = rrticks;
147 1.1.6.1 ad
148 1.1.6.1 ad if (!CURCPU_IDLE_P()) {
149 1.1.6.1 ad if (spc->spc_flags & SPCF_SEENRR) {
150 1.1.6.1 ad /*
151 1.1.6.1 ad * The process has already been through a roundrobin
152 1.1.6.1 ad * without switching and may be hogging the CPU.
153 1.1.6.1 ad * Indicate that the process should yield.
154 1.1.6.1 ad */
155 1.1.6.1 ad spc->spc_flags |= SPCF_SHOULDYIELD;
156 1.1.6.1 ad } else
157 1.1.6.1 ad spc->spc_flags |= SPCF_SEENRR;
158 1.1.6.1 ad }
159 1.1.6.1 ad cpu_need_resched(curcpu(), 0);
160 1.1.6.1 ad }
161 1.1.6.1 ad
162 1.1.6.1 ad #define NICE_WEIGHT 2 /* priorities per nice level */
163 1.1.6.1 ad
164 1.1.6.1 ad #define ESTCPU_SHIFT 11
165 1.1.6.1 ad #define ESTCPU_MAX ((NICE_WEIGHT * PRIO_MAX - PPQ) << ESTCPU_SHIFT)
166 1.1.6.1 ad #define ESTCPULIM(e) min((e), ESTCPU_MAX)
167 1.1.6.1 ad
168 1.1.6.1 ad /*
169 1.1.6.1 ad * Constants for digital decay and forget:
170 1.1.6.1 ad * 90% of (p_estcpu) usage in 5 * loadav time
171 1.1.6.1 ad * 95% of (p_pctcpu) usage in 60 seconds (load insensitive)
172 1.1.6.1 ad * Note that, as ps(1) mentions, this can let percentages
173 1.1.6.1 ad * total over 100% (I've seen 137.9% for 3 processes).
174 1.1.6.1 ad *
175 1.1.6.1 ad * Note that hardclock updates p_estcpu and p_cpticks independently.
176 1.1.6.1 ad *
177 1.1.6.1 ad * We wish to decay away 90% of p_estcpu in (5 * loadavg) seconds.
178 1.1.6.1 ad * That is, the system wants to compute a value of decay such
179 1.1.6.1 ad * that the following for loop:
180 1.1.6.1 ad * for (i = 0; i < (5 * loadavg); i++)
181 1.1.6.1 ad * p_estcpu *= decay;
182 1.1.6.1 ad * will compute
183 1.1.6.1 ad * p_estcpu *= 0.1;
184 1.1.6.1 ad * for all values of loadavg:
185 1.1.6.1 ad *
186 1.1.6.1 ad * Mathematically this loop can be expressed by saying:
187 1.1.6.1 ad * decay ** (5 * loadavg) ~= .1
188 1.1.6.1 ad *
189 1.1.6.1 ad * The system computes decay as:
190 1.1.6.1 ad * decay = (2 * loadavg) / (2 * loadavg + 1)
191 1.1.6.1 ad *
192 1.1.6.1 ad * We wish to prove that the system's computation of decay
193 1.1.6.1 ad * will always fulfill the equation:
194 1.1.6.1 ad * decay ** (5 * loadavg) ~= .1
195 1.1.6.1 ad *
196 1.1.6.1 ad * If we compute b as:
197 1.1.6.1 ad * b = 2 * loadavg
198 1.1.6.1 ad * then
199 1.1.6.1 ad * decay = b / (b + 1)
200 1.1.6.1 ad *
201 1.1.6.1 ad * We now need to prove two things:
202 1.1.6.1 ad * 1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1)
203 1.1.6.1 ad * 2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg)
204 1.1.6.1 ad *
205 1.1.6.1 ad * Facts:
206 1.1.6.1 ad * For x close to zero, exp(x) =~ 1 + x, since
207 1.1.6.1 ad * exp(x) = 0! + x**1/1! + x**2/2! + ... .
208 1.1.6.1 ad * therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b.
209 1.1.6.1 ad * For x close to zero, ln(1+x) =~ x, since
210 1.1.6.1 ad * ln(1+x) = x - x**2/2 + x**3/3 - ... -1 < x < 1
211 1.1.6.1 ad * therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1).
212 1.1.6.1 ad * ln(.1) =~ -2.30
213 1.1.6.1 ad *
214 1.1.6.1 ad * Proof of (1):
215 1.1.6.1 ad * Solve (factor)**(power) =~ .1 given power (5*loadav):
216 1.1.6.1 ad * solving for factor,
217 1.1.6.1 ad * ln(factor) =~ (-2.30/5*loadav), or
218 1.1.6.1 ad * factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) =
219 1.1.6.1 ad * exp(-1/b) =~ (b-1)/b =~ b/(b+1). QED
220 1.1.6.1 ad *
221 1.1.6.1 ad * Proof of (2):
222 1.1.6.1 ad * Solve (factor)**(power) =~ .1 given factor == (b/(b+1)):
223 1.1.6.1 ad * solving for power,
224 1.1.6.1 ad * power*ln(b/(b+1)) =~ -2.30, or
225 1.1.6.1 ad * power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav. QED
226 1.1.6.1 ad *
227 1.1.6.1 ad * Actual power values for the implemented algorithm are as follows:
228 1.1.6.1 ad * loadav: 1 2 3 4
229 1.1.6.1 ad * power: 5.68 10.32 14.94 19.55
230 1.1.6.1 ad */
231 1.1.6.1 ad
232 1.1.6.1 ad /* calculations for digital decay to forget 90% of usage in 5*loadav sec */
233 1.1.6.1 ad #define loadfactor(loadav) (2 * (loadav))
234 1.1.6.1 ad
235 1.1.6.1 ad static fixpt_t
236 1.1.6.1 ad decay_cpu(fixpt_t loadfac, fixpt_t estcpu)
237 1.1.6.1 ad {
238 1.1.6.1 ad
239 1.1.6.1 ad if (estcpu == 0) {
240 1.1.6.1 ad return 0;
241 1.1.6.1 ad }
242 1.1.6.1 ad
243 1.1.6.1 ad #if !defined(_LP64)
244 1.1.6.1 ad /* avoid 64bit arithmetics. */
245 1.1.6.1 ad #define FIXPT_MAX ((fixpt_t)((UINTMAX_C(1) << sizeof(fixpt_t) * CHAR_BIT) - 1))
246 1.1.6.1 ad if (__predict_true(loadfac <= FIXPT_MAX / ESTCPU_MAX)) {
247 1.1.6.1 ad return estcpu * loadfac / (loadfac + FSCALE);
248 1.1.6.1 ad }
249 1.1.6.1 ad #endif /* !defined(_LP64) */
250 1.1.6.1 ad
251 1.1.6.1 ad return (uint64_t)estcpu * loadfac / (loadfac + FSCALE);
252 1.1.6.1 ad }
253 1.1.6.1 ad
254 1.1.6.1 ad /*
255 1.1.6.1 ad * For all load averages >= 1 and max p_estcpu of (255 << ESTCPU_SHIFT),
256 1.1.6.1 ad * sleeping for at least seven times the loadfactor will decay p_estcpu to
257 1.1.6.1 ad * less than (1 << ESTCPU_SHIFT).
258 1.1.6.1 ad *
259 1.1.6.1 ad * note that our ESTCPU_MAX is actually much smaller than (255 << ESTCPU_SHIFT).
260 1.1.6.1 ad */
261 1.1.6.1 ad static fixpt_t
262 1.1.6.1 ad decay_cpu_batch(fixpt_t loadfac, fixpt_t estcpu, unsigned int n)
263 1.1.6.1 ad {
264 1.1.6.1 ad
265 1.1.6.1 ad if ((n << FSHIFT) >= 7 * loadfac) {
266 1.1.6.1 ad return 0;
267 1.1.6.1 ad }
268 1.1.6.1 ad
269 1.1.6.1 ad while (estcpu != 0 && n > 1) {
270 1.1.6.1 ad estcpu = decay_cpu(loadfac, estcpu);
271 1.1.6.1 ad n--;
272 1.1.6.1 ad }
273 1.1.6.1 ad
274 1.1.6.1 ad return estcpu;
275 1.1.6.1 ad }
276 1.1.6.1 ad
277 1.1.6.1 ad /*
278 1.1.6.1 ad * sched_pstats_hook:
279 1.1.6.1 ad *
280 1.1.6.1 ad * Periodically called from sched_pstats(); used to recalculate priorities.
281 1.1.6.1 ad */
282 1.1.6.1 ad void
283 1.1.6.1 ad sched_pstats_hook(struct proc *p, int minslp)
284 1.1.6.1 ad {
285 1.1.6.1 ad struct lwp *l;
286 1.1.6.1 ad fixpt_t loadfac = loadfactor(averunnable.ldavg[0]);
287 1.1.6.1 ad
288 1.1.6.1 ad /*
289 1.1.6.1 ad * If the process has slept the entire second,
290 1.1.6.1 ad * stop recalculating its priority until it wakes up.
291 1.1.6.1 ad */
292 1.1.6.1 ad if (minslp <= 1) {
293 1.1.6.1 ad p->p_estcpu = decay_cpu(loadfac, p->p_estcpu);
294 1.1.6.1 ad
295 1.1.6.1 ad LIST_FOREACH(l, &p->p_lwps, l_sibling) {
296 1.1.6.1 ad if ((l->l_flag & LW_IDLE) != 0)
297 1.1.6.1 ad continue;
298 1.1.6.1 ad lwp_lock(l);
299 1.1.6.1 ad if (l->l_slptime <= 1 && l->l_priority >= PUSER)
300 1.1.6.1 ad resetpriority(l);
301 1.1.6.1 ad lwp_unlock(l);
302 1.1.6.1 ad }
303 1.1.6.1 ad }
304 1.1.6.1 ad }
305 1.1.6.1 ad
306 1.1.6.1 ad /*
307 1.1.6.1 ad * Recalculate the priority of a process after it has slept for a while.
308 1.1.6.1 ad */
309 1.1.6.1 ad static void
310 1.1.6.1 ad updatepri(struct lwp *l)
311 1.1.6.1 ad {
312 1.1.6.1 ad struct proc *p = l->l_proc;
313 1.1.6.1 ad fixpt_t loadfac;
314 1.1.6.1 ad
315 1.1.6.1 ad KASSERT(lwp_locked(l, NULL));
316 1.1.6.1 ad KASSERT(l->l_slptime > 1);
317 1.1.6.1 ad
318 1.1.6.1 ad loadfac = loadfactor(averunnable.ldavg[0]);
319 1.1.6.1 ad
320 1.1.6.1 ad l->l_slptime--; /* the first time was done in sched_pstats */
321 1.1.6.1 ad /* XXX NJWLWP */
322 1.1.6.1 ad /* XXXSMP occasionally unlocked, should be per-LWP */
323 1.1.6.1 ad p->p_estcpu = decay_cpu_batch(loadfac, p->p_estcpu, l->l_slptime);
324 1.1.6.1 ad resetpriority(l);
325 1.1.6.1 ad }
326 1.1.6.1 ad
327 1.1.6.1 ad /*
328 1.1.6.1 ad * On some architectures, it's faster to use a MSB ordering for the priorites
329 1.1.6.1 ad * than the traditional LSB ordering.
330 1.1.6.1 ad */
331 1.1.6.1 ad #define RQMASK(n) (0x00000001 << (n))
332 1.1.6.1 ad
333 1.1.6.1 ad /*
334 1.1.6.1 ad * The primitives that manipulate the run queues. whichqs tells which
335 1.1.6.1 ad * of the 32 queues qs have processes in them. sched_enqueue() puts processes
336 1.1.6.1 ad * into queues, sched_dequeue removes them from queues. The running process is
337 1.1.6.1 ad * on no queue, other processes are on a queue related to p->p_priority,
338 1.1.6.1 ad * divided by 4 actually to shrink the 0-127 range of priorities into the 32
339 1.1.6.1 ad * available queues.
340 1.1.6.1 ad */
341 1.1.6.1 ad #ifdef RQDEBUG
342 1.1.6.1 ad static void
343 1.1.6.1 ad runqueue_check(const runqueue_t *rq, int whichq, struct lwp *l)
344 1.1.6.1 ad {
345 1.1.6.1 ad const subqueue_t * const sq = &rq->rq_subqueues[whichq];
346 1.1.6.1 ad const uint32_t bitmap = rq->rq_bitmap;
347 1.1.6.1 ad struct lwp *l2;
348 1.1.6.1 ad int found = 0;
349 1.1.6.1 ad int die = 0;
350 1.1.6.1 ad int empty = 1;
351 1.1.6.1 ad
352 1.1.6.1 ad TAILQ_FOREACH(l2, &sq->sq_queue, l_runq) {
353 1.1.6.1 ad if (l2->l_stat != LSRUN) {
354 1.1.6.1 ad printf("runqueue_check[%d]: lwp %p state (%d) "
355 1.1.6.1 ad " != LSRUN\n", whichq, l2, l2->l_stat);
356 1.1.6.1 ad }
357 1.1.6.1 ad if (l2 == l)
358 1.1.6.1 ad found = 1;
359 1.1.6.1 ad empty = 0;
360 1.1.6.1 ad }
361 1.1.6.1 ad if (empty && (bitmap & RQMASK(whichq)) != 0) {
362 1.1.6.1 ad printf("runqueue_check[%d]: bit set for empty run-queue %p\n",
363 1.1.6.1 ad whichq, rq);
364 1.1.6.1 ad die = 1;
365 1.1.6.1 ad } else if (!empty && (bitmap & RQMASK(whichq)) == 0) {
366 1.1.6.1 ad printf("runqueue_check[%d]: bit clear for non-empty "
367 1.1.6.1 ad "run-queue %p\n", whichq, rq);
368 1.1.6.1 ad die = 1;
369 1.1.6.1 ad }
370 1.1.6.1 ad if (l != NULL && (bitmap & RQMASK(whichq)) == 0) {
371 1.1.6.1 ad printf("runqueue_check[%d]: bit clear for active lwp %p\n",
372 1.1.6.1 ad whichq, l);
373 1.1.6.1 ad die = 1;
374 1.1.6.1 ad }
375 1.1.6.1 ad if (l != NULL && empty) {
376 1.1.6.1 ad printf("runqueue_check[%d]: empty run-queue %p with "
377 1.1.6.1 ad "active lwp %p\n", whichq, rq, l);
378 1.1.6.1 ad die = 1;
379 1.1.6.1 ad }
380 1.1.6.1 ad if (l != NULL && !found) {
381 1.1.6.1 ad printf("runqueue_check[%d]: lwp %p not in runqueue %p!",
382 1.1.6.1 ad whichq, l, rq);
383 1.1.6.1 ad die = 1;
384 1.1.6.1 ad }
385 1.1.6.1 ad if (die)
386 1.1.6.1 ad panic("runqueue_check: inconsistency found");
387 1.1.6.1 ad }
388 1.1.6.1 ad #else /* RQDEBUG */
389 1.1.6.1 ad #define runqueue_check(a, b, c) /* nothing */
390 1.1.6.1 ad #endif /* RQDEBUG */
391 1.1.6.1 ad
392 1.1.6.1 ad static void
393 1.1.6.1 ad runqueue_init(runqueue_t *rq)
394 1.1.6.1 ad {
395 1.1.6.1 ad int i;
396 1.1.6.1 ad
397 1.1.6.1 ad for (i = 0; i < RUNQUE_NQS; i++)
398 1.1.6.1 ad TAILQ_INIT(&rq->rq_subqueues[i].sq_queue);
399 1.1.6.1 ad }
400 1.1.6.1 ad
401 1.1.6.1 ad static void
402 1.1.6.1 ad runqueue_enqueue(runqueue_t *rq, struct lwp *l)
403 1.1.6.1 ad {
404 1.1.6.1 ad subqueue_t *sq;
405 1.1.6.1 ad const int whichq = lwp_eprio(l) / PPQ;
406 1.1.6.1 ad
407 1.1.6.1 ad KASSERT(lwp_locked(l, l->l_cpu->ci_schedstate.spc_mutex));
408 1.1.6.1 ad
409 1.1.6.1 ad runqueue_check(rq, whichq, NULL);
410 1.1.6.1 ad rq->rq_bitmap |= RQMASK(whichq);
411 1.1.6.1 ad sq = &rq->rq_subqueues[whichq];
412 1.1.6.1 ad TAILQ_INSERT_TAIL(&sq->sq_queue, l, l_runq);
413 1.1.6.1 ad runqueue_check(rq, whichq, l);
414 1.1.6.1 ad }
415 1.1.6.1 ad
416 1.1.6.1 ad static void
417 1.1.6.1 ad runqueue_dequeue(runqueue_t *rq, struct lwp *l)
418 1.1.6.1 ad {
419 1.1.6.1 ad subqueue_t *sq;
420 1.1.6.1 ad const int whichq = lwp_eprio(l) / PPQ;
421 1.1.6.1 ad
422 1.1.6.1 ad KASSERT(lwp_locked(l, l->l_cpu->ci_schedstate.spc_mutex));
423 1.1.6.1 ad
424 1.1.6.1 ad runqueue_check(rq, whichq, l);
425 1.1.6.1 ad KASSERT((rq->rq_bitmap & RQMASK(whichq)) != 0);
426 1.1.6.1 ad sq = &rq->rq_subqueues[whichq];
427 1.1.6.1 ad TAILQ_REMOVE(&sq->sq_queue, l, l_runq);
428 1.1.6.1 ad if (TAILQ_EMPTY(&sq->sq_queue))
429 1.1.6.1 ad rq->rq_bitmap &= ~RQMASK(whichq);
430 1.1.6.1 ad runqueue_check(rq, whichq, NULL);
431 1.1.6.1 ad }
432 1.1.6.1 ad
433 1.1.6.1 ad static struct lwp *
434 1.1.6.1 ad runqueue_nextlwp(runqueue_t *rq)
435 1.1.6.1 ad {
436 1.1.6.1 ad const uint32_t bitmap = rq->rq_bitmap;
437 1.1.6.1 ad int whichq;
438 1.1.6.1 ad
439 1.1.6.1 ad if (bitmap == 0) {
440 1.1.6.1 ad return NULL;
441 1.1.6.1 ad }
442 1.1.6.1 ad whichq = ffs(bitmap) - 1;
443 1.1.6.1 ad return TAILQ_FIRST(&rq->rq_subqueues[whichq].sq_queue);
444 1.1.6.1 ad }
445 1.1.6.1 ad
446 1.1.6.1 ad #if defined(DDB)
447 1.1.6.1 ad static void
448 1.1.6.1 ad runqueue_print(const runqueue_t *rq, void (*pr)(const char *, ...))
449 1.1.6.1 ad {
450 1.1.6.1 ad const uint32_t bitmap = rq->rq_bitmap;
451 1.1.6.1 ad struct lwp *l;
452 1.1.6.1 ad int i, first;
453 1.1.6.1 ad
454 1.1.6.1 ad for (i = 0; i < RUNQUE_NQS; i++) {
455 1.1.6.1 ad const subqueue_t *sq;
456 1.1.6.1 ad first = 1;
457 1.1.6.1 ad sq = &rq->rq_subqueues[i];
458 1.1.6.1 ad TAILQ_FOREACH(l, &sq->sq_queue, l_runq) {
459 1.1.6.1 ad if (first) {
460 1.1.6.1 ad (*pr)("%c%d",
461 1.1.6.1 ad (bitmap & RQMASK(i)) ? ' ' : '!', i);
462 1.1.6.1 ad first = 0;
463 1.1.6.1 ad }
464 1.1.6.1 ad (*pr)("\t%d.%d (%s) pri=%d usrpri=%d\n",
465 1.1.6.1 ad l->l_proc->p_pid,
466 1.1.6.1 ad l->l_lid, l->l_proc->p_comm,
467 1.1.6.1 ad (int)l->l_priority, (int)l->l_usrpri);
468 1.1.6.1 ad }
469 1.1.6.1 ad }
470 1.1.6.1 ad }
471 1.1.6.1 ad #endif /* defined(DDB) */
472 1.1.6.1 ad #undef RQMASK
473 1.1.6.1 ad
474 1.1.6.1 ad /*
475 1.1.6.1 ad * Initialize the (doubly-linked) run queues
476 1.1.6.1 ad * to be empty.
477 1.1.6.1 ad */
478 1.1.6.1 ad void
479 1.1.6.1 ad sched_rqinit()
480 1.1.6.1 ad {
481 1.1.6.1 ad
482 1.1.6.1 ad runqueue_init(&global_queue);
483 1.1.6.1 ad mutex_init(&sched_mutex, MUTEX_SPIN, IPL_SCHED);
484 1.1.6.1 ad /* Initialize the lock pointer for lwp0 */
485 1.1.6.1 ad lwp0.l_mutex = &curcpu()->ci_schedstate.spc_lwplock;
486 1.1.6.1 ad }
487 1.1.6.1 ad
488 1.1.6.1 ad void
489 1.1.6.1 ad sched_cpuattach(struct cpu_info *ci)
490 1.1.6.1 ad {
491 1.1.6.1 ad runqueue_t *rq;
492 1.1.6.1 ad
493 1.1.6.1 ad ci->ci_schedstate.spc_mutex = &sched_mutex;
494 1.1.6.1 ad rq = kmem_zalloc(sizeof(*rq), KM_NOSLEEP);
495 1.1.6.1 ad runqueue_init(rq);
496 1.1.6.1 ad ci->ci_schedstate.spc_sched_info = rq;
497 1.1.6.1 ad }
498 1.1.6.1 ad
499 1.1.6.1 ad void
500 1.1.6.1 ad sched_setup()
501 1.1.6.1 ad {
502 1.1.6.1 ad
503 1.1.6.1 ad rrticks = hz / 10;
504 1.1.6.1 ad sched_pstats(NULL);
505 1.1.6.1 ad }
506 1.1.6.1 ad
507 1.1.6.1 ad void
508 1.1.6.1 ad sched_setrunnable(struct lwp *l)
509 1.1.6.1 ad {
510 1.1.6.1 ad
511 1.1.6.1 ad if (l->l_slptime > 1)
512 1.1.6.1 ad updatepri(l);
513 1.1.6.1 ad }
514 1.1.6.1 ad
515 1.1.6.1 ad bool
516 1.1.6.1 ad sched_curcpu_runnable_p(void)
517 1.1.6.1 ad {
518 1.1.6.1 ad runqueue_t *rq = curcpu()->ci_schedstate.spc_sched_info;
519 1.1.6.1 ad
520 1.1.6.1 ad return (global_queue.rq_bitmap | rq->rq_bitmap) != 0;
521 1.1.6.1 ad }
522 1.1.6.1 ad
523 1.1.6.1 ad void
524 1.1.6.1 ad sched_nice(struct proc *chgp, int n)
525 1.1.6.1 ad {
526 1.1.6.1 ad
527 1.1.6.1 ad chgp->p_nice = n;
528 1.1.6.1 ad (void)resetprocpriority(chgp);
529 1.1.6.1 ad }
530 1.1.6.1 ad
531 1.1.6.1 ad /*
532 1.1.6.1 ad * Compute the priority of a process when running in user mode.
533 1.1.6.1 ad * Arrange to reschedule if the resulting priority is better
534 1.1.6.1 ad * than that of the current process.
535 1.1.6.1 ad */
536 1.1.6.1 ad static void
537 1.1.6.1 ad resetpriority(struct lwp *l)
538 1.1.6.1 ad {
539 1.1.6.1 ad unsigned int newpriority;
540 1.1.6.1 ad struct proc *p = l->l_proc;
541 1.1.6.1 ad
542 1.1.6.1 ad /* XXXSMP KASSERT(mutex_owned(&p->p_stmutex)); */
543 1.1.6.1 ad KASSERT(lwp_locked(l, NULL));
544 1.1.6.1 ad
545 1.1.6.1 ad if ((l->l_flag & LW_SYSTEM) != 0)
546 1.1.6.1 ad return;
547 1.1.6.1 ad
548 1.1.6.1 ad newpriority = PUSER + (p->p_estcpu >> ESTCPU_SHIFT) +
549 1.1.6.1 ad NICE_WEIGHT * (p->p_nice - NZERO);
550 1.1.6.1 ad newpriority = min(newpriority, MAXPRI);
551 1.1.6.1 ad lwp_changepri(l, newpriority);
552 1.1.6.1 ad }
553 1.1.6.1 ad
554 1.1.6.1 ad /*
555 1.1.6.1 ad * Recompute priority for all LWPs in a process.
556 1.1.6.1 ad */
557 1.1.6.1 ad static void
558 1.1.6.1 ad resetprocpriority(struct proc *p)
559 1.1.6.1 ad {
560 1.1.6.1 ad struct lwp *l;
561 1.1.6.1 ad
562 1.1.6.1 ad KASSERT(mutex_owned(&p->p_stmutex));
563 1.1.6.1 ad
564 1.1.6.1 ad LIST_FOREACH(l, &p->p_lwps, l_sibling) {
565 1.1.6.1 ad lwp_lock(l);
566 1.1.6.1 ad resetpriority(l);
567 1.1.6.1 ad lwp_unlock(l);
568 1.1.6.1 ad }
569 1.1.6.1 ad }
570 1.1.6.1 ad
571 1.1.6.1 ad /*
572 1.1.6.1 ad * We adjust the priority of the current process. The priority of a process
573 1.1.6.1 ad * gets worse as it accumulates CPU time. The CPU usage estimator (p_estcpu)
574 1.1.6.1 ad * is increased here. The formula for computing priorities (in kern_synch.c)
575 1.1.6.1 ad * will compute a different value each time p_estcpu increases. This can
576 1.1.6.1 ad * cause a switch, but unless the priority crosses a PPQ boundary the actual
577 1.1.6.1 ad * queue will not change. The CPU usage estimator ramps up quite quickly
578 1.1.6.1 ad * when the process is running (linearly), and decays away exponentially, at
579 1.1.6.1 ad * a rate which is proportionally slower when the system is busy. The basic
580 1.1.6.1 ad * principle is that the system will 90% forget that the process used a lot
581 1.1.6.1 ad * of CPU time in 5 * loadav seconds. This causes the system to favor
582 1.1.6.1 ad * processes which haven't run much recently, and to round-robin among other
583 1.1.6.1 ad * processes.
584 1.1.6.1 ad */
585 1.1.6.1 ad
586 1.1.6.1 ad void
587 1.1.6.1 ad sched_schedclock(struct lwp *l)
588 1.1.6.1 ad {
589 1.1.6.1 ad struct proc *p = l->l_proc;
590 1.1.6.1 ad
591 1.1.6.1 ad KASSERT(!CURCPU_IDLE_P());
592 1.1.6.1 ad mutex_spin_enter(&p->p_stmutex);
593 1.1.6.1 ad p->p_estcpu = ESTCPULIM(p->p_estcpu + (1 << ESTCPU_SHIFT));
594 1.1.6.1 ad lwp_lock(l);
595 1.1.6.1 ad resetpriority(l);
596 1.1.6.1 ad mutex_spin_exit(&p->p_stmutex);
597 1.1.6.1 ad if ((l->l_flag & LW_SYSTEM) == 0 && l->l_priority >= PUSER)
598 1.1.6.1 ad l->l_priority = l->l_usrpri;
599 1.1.6.1 ad lwp_unlock(l);
600 1.1.6.1 ad }
601 1.1.6.1 ad
602 1.1.6.1 ad /*
603 1.1.6.1 ad * sched_proc_fork:
604 1.1.6.1 ad *
605 1.1.6.1 ad * Inherit the parent's scheduler history.
606 1.1.6.1 ad */
607 1.1.6.1 ad void
608 1.1.6.1 ad sched_proc_fork(struct proc *parent, struct proc *child)
609 1.1.6.1 ad {
610 1.1.6.1 ad
611 1.1.6.1 ad KASSERT(mutex_owned(&parent->p_smutex));
612 1.1.6.1 ad
613 1.1.6.1 ad child->p_estcpu = child->p_estcpu_inherited = parent->p_estcpu;
614 1.1.6.1 ad child->p_forktime = sched_pstats_ticks;
615 1.1.6.1 ad }
616 1.1.6.1 ad
617 1.1.6.1 ad /*
618 1.1.6.1 ad * sched_proc_exit:
619 1.1.6.1 ad *
620 1.1.6.1 ad * Chargeback parents for the sins of their children.
621 1.1.6.1 ad */
622 1.1.6.1 ad void
623 1.1.6.1 ad sched_proc_exit(struct proc *parent, struct proc *child)
624 1.1.6.1 ad {
625 1.1.6.1 ad fixpt_t loadfac = loadfactor(averunnable.ldavg[0]);
626 1.1.6.1 ad fixpt_t estcpu;
627 1.1.6.1 ad
628 1.1.6.1 ad /* XXX Only if parent != init?? */
629 1.1.6.1 ad
630 1.1.6.1 ad mutex_spin_enter(&parent->p_stmutex);
631 1.1.6.1 ad estcpu = decay_cpu_batch(loadfac, child->p_estcpu_inherited,
632 1.1.6.1 ad sched_pstats_ticks - child->p_forktime);
633 1.1.6.1 ad if (child->p_estcpu > estcpu)
634 1.1.6.1 ad parent->p_estcpu =
635 1.1.6.1 ad ESTCPULIM(parent->p_estcpu + child->p_estcpu - estcpu);
636 1.1.6.1 ad mutex_spin_exit(&parent->p_stmutex);
637 1.1.6.1 ad }
638 1.1.6.1 ad
639 1.1.6.1 ad void
640 1.1.6.1 ad sched_enqueue(struct lwp *l, bool ctxswitch)
641 1.1.6.1 ad {
642 1.1.6.1 ad
643 1.1.6.1 ad if ((l->l_flag & LW_BOUND) != 0)
644 1.1.6.1 ad runqueue_enqueue(l->l_cpu->ci_schedstate.spc_sched_info, l);
645 1.1.6.1 ad else
646 1.1.6.1 ad runqueue_enqueue(&global_queue, l);
647 1.1.6.1 ad }
648 1.1.6.1 ad
649 1.1.6.1 ad /*
650 1.1.6.1 ad * XXXSMP When LWP dispatch (cpu_switch()) is changed to use sched_dequeue(),
651 1.1.6.1 ad * drop of the effective priority level from kernel to user needs to be
652 1.1.6.1 ad * moved here from userret(). The assignment in userret() is currently
653 1.1.6.1 ad * done unlocked.
654 1.1.6.1 ad */
655 1.1.6.1 ad void
656 1.1.6.1 ad sched_dequeue(struct lwp *l)
657 1.1.6.1 ad {
658 1.1.6.1 ad
659 1.1.6.1 ad if ((l->l_flag & LW_BOUND) != 0)
660 1.1.6.1 ad runqueue_dequeue(l->l_cpu->ci_schedstate.spc_sched_info, l);
661 1.1.6.1 ad else
662 1.1.6.1 ad runqueue_dequeue(&global_queue, l);
663 1.1.6.1 ad }
664 1.1.6.1 ad
665 1.1.6.1 ad struct lwp *
666 1.1.6.1 ad sched_nextlwp(void)
667 1.1.6.1 ad {
668 1.1.6.1 ad lwp_t *l1, *l2;
669 1.1.6.1 ad
670 1.1.6.1 ad /* For now, just pick the highest priority LWP. */
671 1.1.6.1 ad l1 = runqueue_nextlwp(curcpu()->ci_schedstate.spc_sched_info);
672 1.1.6.1 ad l2 = runqueue_nextlwp(&global_queue);
673 1.1.6.1 ad
674 1.1.6.1 ad if (l1 == NULL)
675 1.1.6.1 ad return l2;
676 1.1.6.1 ad if (l2 == NULL)
677 1.1.6.1 ad return l1;
678 1.1.6.1 ad if (lwp_eprio(l2) < lwp_eprio(l1))
679 1.1.6.1 ad return l2;
680 1.1.6.1 ad else
681 1.1.6.1 ad return l1;
682 1.1.6.1 ad }
683 1.1.6.1 ad
684 1.1.6.1 ad /* Dummy */
685 1.1.6.1 ad void
686 1.1.6.1 ad sched_lwp_fork(struct lwp *l)
687 1.1.6.1 ad {
688 1.1.6.1 ad
689 1.1.6.1 ad }
690 1.1.6.1 ad
691 1.1.6.1 ad void
692 1.1.6.1 ad sched_lwp_exit(struct lwp *l)
693 1.1.6.1 ad {
694 1.1.6.1 ad
695 1.1.6.1 ad }
696 1.1.6.1 ad
697 1.1.6.1 ad /* SysCtl */
698 1.1.6.1 ad
699 1.1.6.1 ad SYSCTL_SETUP(sysctl_sched_setup, "sysctl kern.sched subtree setup")
700 1.1.6.1 ad {
701 1.1.6.1 ad const struct sysctlnode *node = NULL;
702 1.1.6.1 ad
703 1.1.6.1 ad sysctl_createv(clog, 0, NULL, NULL,
704 1.1.6.1 ad CTLFLAG_PERMANENT,
705 1.1.6.1 ad CTLTYPE_NODE, "kern", NULL,
706 1.1.6.1 ad NULL, 0, NULL, 0,
707 1.1.6.1 ad CTL_KERN, CTL_EOL);
708 1.1.6.1 ad sysctl_createv(clog, 0, NULL, &node,
709 1.1.6.1 ad CTLFLAG_PERMANENT,
710 1.1.6.1 ad CTLTYPE_NODE, "sched",
711 1.1.6.1 ad SYSCTL_DESCR("Scheduler options"),
712 1.1.6.1 ad NULL, 0, NULL, 0,
713 1.1.6.1 ad CTL_KERN, CTL_CREATE, CTL_EOL);
714 1.1.6.1 ad
715 1.1.6.1 ad if (node != NULL) {
716 1.1.6.1 ad sysctl_createv(clog, 0, &node, NULL,
717 1.1.6.1 ad CTLFLAG_PERMANENT,
718 1.1.6.1 ad CTLTYPE_STRING, "name", NULL,
719 1.1.6.1 ad NULL, 0, __UNCONST("4.4BSD"), 0,
720 1.1.6.1 ad CTL_CREATE, CTL_EOL);
721 1.1.6.1 ad }
722 1.1.6.1 ad }
723 1.1.6.1 ad
724 1.1.6.1 ad #if defined(DDB)
725 1.1.6.1 ad void
726 1.1.6.1 ad sched_print_runqueue(void (*pr)(const char *, ...))
727 1.1.6.1 ad {
728 1.1.6.1 ad
729 1.1.6.1 ad runqueue_print(&global_queue, pr);
730 1.1.6.1 ad }
731 1.1.6.1 ad #endif /* defined(DDB) */
732