1b43179fbSJeff Roberson /*- 2b43179fbSJeff Roberson * Copyright (c) 1982, 1986, 1990, 1991, 1993 3b43179fbSJeff Roberson * The Regents of the University of California. All rights reserved. 4b43179fbSJeff Roberson * (c) UNIX System Laboratories, Inc. 5b43179fbSJeff Roberson * All or some portions of this file are derived from material licensed 6b43179fbSJeff Roberson * to the University of California by American Telephone and Telegraph 7b43179fbSJeff Roberson * Co. or Unix System Laboratories, Inc. and are reproduced herein with 8b43179fbSJeff Roberson * the permission of UNIX System Laboratories, Inc. 9b43179fbSJeff Roberson * 10b43179fbSJeff Roberson * Redistribution and use in source and binary forms, with or without 11b43179fbSJeff Roberson * modification, are permitted provided that the following conditions 12b43179fbSJeff Roberson * are met: 13b43179fbSJeff Roberson * 1. Redistributions of source code must retain the above copyright 14b43179fbSJeff Roberson * notice, this list of conditions and the following disclaimer. 15b43179fbSJeff Roberson * 2. Redistributions in binary form must reproduce the above copyright 16b43179fbSJeff Roberson * notice, this list of conditions and the following disclaimer in the 17b43179fbSJeff Roberson * documentation and/or other materials provided with the distribution. 18b43179fbSJeff Roberson * 3. All advertising materials mentioning features or use of this software 19b43179fbSJeff Roberson * must display the following acknowledgement: 20b43179fbSJeff Roberson * This product includes software developed by the University of 21b43179fbSJeff Roberson * California, Berkeley and its contributors. 22b43179fbSJeff Roberson * 4. Neither the name of the University nor the names of its contributors 23b43179fbSJeff Roberson * may be used to endorse or promote products derived from this software 24b43179fbSJeff Roberson * without specific prior written permission. 25b43179fbSJeff Roberson * 26b43179fbSJeff Roberson * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 27b43179fbSJeff Roberson * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 28b43179fbSJeff Roberson * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 29b43179fbSJeff Roberson * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 30b43179fbSJeff Roberson * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 31b43179fbSJeff Roberson * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 32b43179fbSJeff Roberson * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 33b43179fbSJeff Roberson * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 34b43179fbSJeff Roberson * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 35b43179fbSJeff Roberson * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 36b43179fbSJeff Roberson * SUCH DAMAGE. 37b43179fbSJeff Roberson * 38b43179fbSJeff Roberson * $FreeBSD$ 39b43179fbSJeff Roberson */ 40b43179fbSJeff Roberson 41b43179fbSJeff Roberson #include <sys/param.h> 42b43179fbSJeff Roberson #include <sys/systm.h> 43b43179fbSJeff Roberson #include <sys/kernel.h> 44b43179fbSJeff Roberson #include <sys/ktr.h> 45b43179fbSJeff Roberson #include <sys/lock.h> 46b43179fbSJeff Roberson #include <sys/mutex.h> 47b43179fbSJeff Roberson #include <sys/proc.h> 48b43179fbSJeff Roberson #include <sys/resourcevar.h> 49b43179fbSJeff Roberson #include <sys/sched.h> 50b43179fbSJeff Roberson #include <sys/smp.h> 51b43179fbSJeff Roberson #include <sys/sysctl.h> 52b43179fbSJeff Roberson #include <sys/sx.h> 53b43179fbSJeff Roberson 5406439a04SJeff Roberson /* 5506439a04SJeff Roberson * INVERSE_ESTCPU_WEIGHT is only suitable for statclock() frequencies in 5606439a04SJeff Roberson * the range 100-256 Hz (approximately). 5706439a04SJeff Roberson */ 5806439a04SJeff Roberson #define ESTCPULIM(e) \ 5906439a04SJeff Roberson min((e), INVERSE_ESTCPU_WEIGHT * (NICE_WEIGHT * (PRIO_MAX - PRIO_MIN) - \ 6006439a04SJeff Roberson RQ_PPQ) + INVERSE_ESTCPU_WEIGHT - 1) 6106439a04SJeff Roberson #define INVERSE_ESTCPU_WEIGHT 8 /* 1 / (priorities per estcpu level). */ 6206439a04SJeff Roberson #define NICE_WEIGHT 1 /* Priorities per nice level. */ 6306439a04SJeff Roberson 64de028f5aSJeff Roberson struct ke_sched *kse0_sched = NULL; 65de028f5aSJeff Roberson struct kg_sched *ksegrp0_sched = NULL; 66de028f5aSJeff Roberson struct p_sched *proc0_sched = NULL; 67de028f5aSJeff Roberson struct td_sched *thread0_sched = NULL; 68b43179fbSJeff Roberson 69b43179fbSJeff Roberson static int sched_quantum; /* Roundrobin scheduling quantum in ticks. */ 70b43179fbSJeff Roberson #define SCHED_QUANTUM (hz / 10); /* Default sched quantum */ 71b43179fbSJeff Roberson 72b43179fbSJeff Roberson static struct callout schedcpu_callout; 73b43179fbSJeff Roberson static struct callout roundrobin_callout; 74b43179fbSJeff Roberson 75b43179fbSJeff Roberson static void roundrobin(void *arg); 76b43179fbSJeff Roberson static void schedcpu(void *arg); 77b43179fbSJeff Roberson static void sched_setup(void *dummy); 78b43179fbSJeff Roberson static void maybe_resched(struct thread *td); 79b43179fbSJeff Roberson static void updatepri(struct ksegrp *kg); 80b43179fbSJeff Roberson static void resetpriority(struct ksegrp *kg); 81b43179fbSJeff Roberson 82b43179fbSJeff Roberson SYSINIT(sched_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, sched_setup, NULL) 83b43179fbSJeff Roberson 84b43179fbSJeff Roberson /* 85b43179fbSJeff Roberson * Global run queue. 86b43179fbSJeff Roberson */ 87b43179fbSJeff Roberson static struct runq runq; 88b43179fbSJeff Roberson SYSINIT(runq, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, runq_init, &runq) 89b43179fbSJeff Roberson 90b43179fbSJeff Roberson static int 91b43179fbSJeff Roberson sysctl_kern_quantum(SYSCTL_HANDLER_ARGS) 92b43179fbSJeff Roberson { 93b43179fbSJeff Roberson int error, new_val; 94b43179fbSJeff Roberson 95b43179fbSJeff Roberson new_val = sched_quantum * tick; 96b43179fbSJeff Roberson error = sysctl_handle_int(oidp, &new_val, 0, req); 97b43179fbSJeff Roberson if (error != 0 || req->newptr == NULL) 98b43179fbSJeff Roberson return (error); 99b43179fbSJeff Roberson if (new_val < tick) 100b43179fbSJeff Roberson return (EINVAL); 101b43179fbSJeff Roberson sched_quantum = new_val / tick; 102b43179fbSJeff Roberson hogticks = 2 * sched_quantum; 103b43179fbSJeff Roberson return (0); 104b43179fbSJeff Roberson } 105b43179fbSJeff Roberson 106b43179fbSJeff Roberson SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW, 107b43179fbSJeff Roberson 0, sizeof sched_quantum, sysctl_kern_quantum, "I", 108b43179fbSJeff Roberson "Roundrobin scheduling quantum in microseconds"); 109b43179fbSJeff Roberson 110b43179fbSJeff Roberson /* 111b43179fbSJeff Roberson * Arrange to reschedule if necessary, taking the priorities and 112b43179fbSJeff Roberson * schedulers into account. 113b43179fbSJeff Roberson */ 114b43179fbSJeff Roberson static void 115b43179fbSJeff Roberson maybe_resched(struct thread *td) 116b43179fbSJeff Roberson { 117b43179fbSJeff Roberson 118b43179fbSJeff Roberson mtx_assert(&sched_lock, MA_OWNED); 119b43179fbSJeff Roberson if (td->td_priority < curthread->td_priority) 120b43179fbSJeff Roberson curthread->td_kse->ke_flags |= KEF_NEEDRESCHED; 121b43179fbSJeff Roberson } 122b43179fbSJeff Roberson 123b43179fbSJeff Roberson /* 124b43179fbSJeff Roberson * Force switch among equal priority processes every 100ms. 125b43179fbSJeff Roberson * We don't actually need to force a context switch of the current process. 126b43179fbSJeff Roberson * The act of firing the event triggers a context switch to softclock() and 127b43179fbSJeff Roberson * then switching back out again which is equivalent to a preemption, thus 128b43179fbSJeff Roberson * no further work is needed on the local CPU. 129b43179fbSJeff Roberson */ 130b43179fbSJeff Roberson /* ARGSUSED */ 131b43179fbSJeff Roberson static void 132b43179fbSJeff Roberson roundrobin(void *arg) 133b43179fbSJeff Roberson { 134b43179fbSJeff Roberson 135b43179fbSJeff Roberson #ifdef SMP 136b43179fbSJeff Roberson mtx_lock_spin(&sched_lock); 137b43179fbSJeff Roberson forward_roundrobin(); 138b43179fbSJeff Roberson mtx_unlock_spin(&sched_lock); 139b43179fbSJeff Roberson #endif 140b43179fbSJeff Roberson 141b43179fbSJeff Roberson callout_reset(&roundrobin_callout, sched_quantum, roundrobin, NULL); 142b43179fbSJeff Roberson } 143b43179fbSJeff Roberson 144b43179fbSJeff Roberson /* 145b43179fbSJeff Roberson * Constants for digital decay and forget: 146b43179fbSJeff Roberson * 90% of (p_estcpu) usage in 5 * loadav time 147b43179fbSJeff Roberson * 95% of (p_pctcpu) usage in 60 seconds (load insensitive) 148b43179fbSJeff Roberson * Note that, as ps(1) mentions, this can let percentages 149b43179fbSJeff Roberson * total over 100% (I've seen 137.9% for 3 processes). 150b43179fbSJeff Roberson * 151b43179fbSJeff Roberson * Note that schedclock() updates p_estcpu and p_cpticks asynchronously. 152b43179fbSJeff Roberson * 153b43179fbSJeff Roberson * We wish to decay away 90% of p_estcpu in (5 * loadavg) seconds. 154b43179fbSJeff Roberson * That is, the system wants to compute a value of decay such 155b43179fbSJeff Roberson * that the following for loop: 156b43179fbSJeff Roberson * for (i = 0; i < (5 * loadavg); i++) 157b43179fbSJeff Roberson * p_estcpu *= decay; 158b43179fbSJeff Roberson * will compute 159b43179fbSJeff Roberson * p_estcpu *= 0.1; 160b43179fbSJeff Roberson * for all values of loadavg: 161b43179fbSJeff Roberson * 162b43179fbSJeff Roberson * Mathematically this loop can be expressed by saying: 163b43179fbSJeff Roberson * decay ** (5 * loadavg) ~= .1 164b43179fbSJeff Roberson * 165b43179fbSJeff Roberson * The system computes decay as: 166b43179fbSJeff Roberson * decay = (2 * loadavg) / (2 * loadavg + 1) 167b43179fbSJeff Roberson * 168b43179fbSJeff Roberson * We wish to prove that the system's computation of decay 169b43179fbSJeff Roberson * will always fulfill the equation: 170b43179fbSJeff Roberson * decay ** (5 * loadavg) ~= .1 171b43179fbSJeff Roberson * 172b43179fbSJeff Roberson * If we compute b as: 173b43179fbSJeff Roberson * b = 2 * loadavg 174b43179fbSJeff Roberson * then 175b43179fbSJeff Roberson * decay = b / (b + 1) 176b43179fbSJeff Roberson * 177b43179fbSJeff Roberson * We now need to prove two things: 178b43179fbSJeff Roberson * 1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1) 179b43179fbSJeff Roberson * 2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg) 180b43179fbSJeff Roberson * 181b43179fbSJeff Roberson * Facts: 182b43179fbSJeff Roberson * For x close to zero, exp(x) =~ 1 + x, since 183b43179fbSJeff Roberson * exp(x) = 0! + x**1/1! + x**2/2! + ... . 184b43179fbSJeff Roberson * therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b. 185b43179fbSJeff Roberson * For x close to zero, ln(1+x) =~ x, since 186b43179fbSJeff Roberson * ln(1+x) = x - x**2/2 + x**3/3 - ... -1 < x < 1 187b43179fbSJeff Roberson * therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1). 188b43179fbSJeff Roberson * ln(.1) =~ -2.30 189b43179fbSJeff Roberson * 190b43179fbSJeff Roberson * Proof of (1): 191b43179fbSJeff Roberson * Solve (factor)**(power) =~ .1 given power (5*loadav): 192b43179fbSJeff Roberson * solving for factor, 193b43179fbSJeff Roberson * ln(factor) =~ (-2.30/5*loadav), or 194b43179fbSJeff Roberson * factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) = 195b43179fbSJeff Roberson * exp(-1/b) =~ (b-1)/b =~ b/(b+1). QED 196b43179fbSJeff Roberson * 197b43179fbSJeff Roberson * Proof of (2): 198b43179fbSJeff Roberson * Solve (factor)**(power) =~ .1 given factor == (b/(b+1)): 199b43179fbSJeff Roberson * solving for power, 200b43179fbSJeff Roberson * power*ln(b/(b+1)) =~ -2.30, or 201b43179fbSJeff Roberson * power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav. QED 202b43179fbSJeff Roberson * 203b43179fbSJeff Roberson * Actual power values for the implemented algorithm are as follows: 204b43179fbSJeff Roberson * loadav: 1 2 3 4 205b43179fbSJeff Roberson * power: 5.68 10.32 14.94 19.55 206b43179fbSJeff Roberson */ 207b43179fbSJeff Roberson 208b43179fbSJeff Roberson /* calculations for digital decay to forget 90% of usage in 5*loadav sec */ 209b43179fbSJeff Roberson #define loadfactor(loadav) (2 * (loadav)) 210b43179fbSJeff Roberson #define decay_cpu(loadfac, cpu) (((loadfac) * (cpu)) / ((loadfac) + FSCALE)) 211b43179fbSJeff Roberson 212b43179fbSJeff Roberson /* decay 95% of `p_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */ 213b43179fbSJeff Roberson static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */ 214b43179fbSJeff Roberson SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 215b43179fbSJeff Roberson 216b43179fbSJeff Roberson /* 217b43179fbSJeff Roberson * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the 218b43179fbSJeff Roberson * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below 219b43179fbSJeff Roberson * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT). 220b43179fbSJeff Roberson * 221b43179fbSJeff Roberson * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used: 222b43179fbSJeff Roberson * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits). 223b43179fbSJeff Roberson * 224b43179fbSJeff Roberson * If you don't want to bother with the faster/more-accurate formula, you 225b43179fbSJeff Roberson * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate 226b43179fbSJeff Roberson * (more general) method of calculating the %age of CPU used by a process. 227b43179fbSJeff Roberson */ 228b43179fbSJeff Roberson #define CCPU_SHIFT 11 229b43179fbSJeff Roberson 230b43179fbSJeff Roberson /* 231b43179fbSJeff Roberson * Recompute process priorities, every hz ticks. 232b43179fbSJeff Roberson * MP-safe, called without the Giant mutex. 233b43179fbSJeff Roberson */ 234b43179fbSJeff Roberson /* ARGSUSED */ 235b43179fbSJeff Roberson static void 236b43179fbSJeff Roberson schedcpu(void *arg) 237b43179fbSJeff Roberson { 238b43179fbSJeff Roberson register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 239b43179fbSJeff Roberson struct thread *td; 240b43179fbSJeff Roberson struct proc *p; 241b43179fbSJeff Roberson struct kse *ke; 242b43179fbSJeff Roberson struct ksegrp *kg; 243b43179fbSJeff Roberson int realstathz; 244b43179fbSJeff Roberson int awake; 245b43179fbSJeff Roberson 246b43179fbSJeff Roberson realstathz = stathz ? stathz : hz; 247b43179fbSJeff Roberson sx_slock(&allproc_lock); 248b43179fbSJeff Roberson FOREACH_PROC_IN_SYSTEM(p) { 249b43179fbSJeff Roberson mtx_lock_spin(&sched_lock); 250b43179fbSJeff Roberson p->p_swtime++; 251b43179fbSJeff Roberson FOREACH_KSEGRP_IN_PROC(p, kg) { 252b43179fbSJeff Roberson awake = 0; 253b43179fbSJeff Roberson FOREACH_KSE_IN_GROUP(kg, ke) { 254b43179fbSJeff Roberson /* 255b43179fbSJeff Roberson * Increment time in/out of memory and sleep 256b43179fbSJeff Roberson * time (if sleeping). We ignore overflow; 257b43179fbSJeff Roberson * with 16-bit int's (remember them?) 258b43179fbSJeff Roberson * overflow takes 45 days. 259b43179fbSJeff Roberson */ 260b43179fbSJeff Roberson /* 261b43179fbSJeff Roberson * The kse slptimes are not touched in wakeup 262b43179fbSJeff Roberson * because the thread may not HAVE a KSE. 263b43179fbSJeff Roberson */ 264b43179fbSJeff Roberson if (ke->ke_state == KES_ONRUNQ) { 265b43179fbSJeff Roberson awake = 1; 266b43179fbSJeff Roberson ke->ke_flags &= ~KEF_DIDRUN; 267b43179fbSJeff Roberson } else if ((ke->ke_state == KES_THREAD) && 268b43179fbSJeff Roberson (TD_IS_RUNNING(ke->ke_thread))) { 269b43179fbSJeff Roberson awake = 1; 270b43179fbSJeff Roberson /* Do not clear KEF_DIDRUN */ 271b43179fbSJeff Roberson } else if (ke->ke_flags & KEF_DIDRUN) { 272b43179fbSJeff Roberson awake = 1; 273b43179fbSJeff Roberson ke->ke_flags &= ~KEF_DIDRUN; 274b43179fbSJeff Roberson } 275b43179fbSJeff Roberson 276b43179fbSJeff Roberson /* 277b43179fbSJeff Roberson * pctcpu is only for ps? 278b43179fbSJeff Roberson * Do it per kse.. and add them up at the end? 279b43179fbSJeff Roberson * XXXKSE 280b43179fbSJeff Roberson */ 281b43179fbSJeff Roberson ke->ke_pctcpu 282b43179fbSJeff Roberson = (ke->ke_pctcpu * ccpu) >> FSHIFT; 283b43179fbSJeff Roberson /* 284b43179fbSJeff Roberson * If the kse has been idle the entire second, 285b43179fbSJeff Roberson * stop recalculating its priority until 286b43179fbSJeff Roberson * it wakes up. 287b43179fbSJeff Roberson */ 288b43179fbSJeff Roberson if (ke->ke_cpticks == 0) 289b43179fbSJeff Roberson continue; 290b43179fbSJeff Roberson #if (FSHIFT >= CCPU_SHIFT) 291b43179fbSJeff Roberson ke->ke_pctcpu += (realstathz == 100) ? 292b43179fbSJeff Roberson ((fixpt_t) ke->ke_cpticks) << 293b43179fbSJeff Roberson (FSHIFT - CCPU_SHIFT) : 294b43179fbSJeff Roberson 100 * (((fixpt_t) ke->ke_cpticks) << 295b43179fbSJeff Roberson (FSHIFT - CCPU_SHIFT)) / realstathz; 296b43179fbSJeff Roberson #else 297b43179fbSJeff Roberson ke->ke_pctcpu += ((FSCALE - ccpu) * 298b43179fbSJeff Roberson (ke->ke_cpticks * FSCALE / realstathz)) >> 299b43179fbSJeff Roberson FSHIFT; 300b43179fbSJeff Roberson #endif 301b43179fbSJeff Roberson ke->ke_cpticks = 0; 302b43179fbSJeff Roberson } /* end of kse loop */ 303b43179fbSJeff Roberson /* 304b43179fbSJeff Roberson * If there are ANY running threads in this KSEGRP, 305b43179fbSJeff Roberson * then don't count it as sleeping. 306b43179fbSJeff Roberson */ 307b43179fbSJeff Roberson if (awake) { 308b43179fbSJeff Roberson if (kg->kg_slptime > 1) { 309b43179fbSJeff Roberson /* 310b43179fbSJeff Roberson * In an ideal world, this should not 311b43179fbSJeff Roberson * happen, because whoever woke us 312b43179fbSJeff Roberson * up from the long sleep should have 313b43179fbSJeff Roberson * unwound the slptime and reset our 314b43179fbSJeff Roberson * priority before we run at the stale 315b43179fbSJeff Roberson * priority. Should KASSERT at some 316b43179fbSJeff Roberson * point when all the cases are fixed. 317b43179fbSJeff Roberson */ 318b43179fbSJeff Roberson updatepri(kg); 319b43179fbSJeff Roberson } 320b43179fbSJeff Roberson kg->kg_slptime = 0; 321b43179fbSJeff Roberson } else { 322b43179fbSJeff Roberson kg->kg_slptime++; 323b43179fbSJeff Roberson } 324b43179fbSJeff Roberson if (kg->kg_slptime > 1) 325b43179fbSJeff Roberson continue; 326b43179fbSJeff Roberson kg->kg_estcpu = decay_cpu(loadfac, kg->kg_estcpu); 327b43179fbSJeff Roberson resetpriority(kg); 328b43179fbSJeff Roberson FOREACH_THREAD_IN_GROUP(kg, td) { 329b43179fbSJeff Roberson if (td->td_priority >= PUSER) { 3301f955e2dSJulian Elischer sched_prio(td, kg->kg_user_pri); 331b43179fbSJeff Roberson } 332b43179fbSJeff Roberson } 333b43179fbSJeff Roberson } /* end of ksegrp loop */ 334b43179fbSJeff Roberson mtx_unlock_spin(&sched_lock); 335b43179fbSJeff Roberson } /* end of process loop */ 336b43179fbSJeff Roberson sx_sunlock(&allproc_lock); 337b43179fbSJeff Roberson callout_reset(&schedcpu_callout, hz, schedcpu, NULL); 338b43179fbSJeff Roberson } 339b43179fbSJeff Roberson 340b43179fbSJeff Roberson /* 341b43179fbSJeff Roberson * Recalculate the priority of a process after it has slept for a while. 342b43179fbSJeff Roberson * For all load averages >= 1 and max p_estcpu of 255, sleeping for at 343b43179fbSJeff Roberson * least six times the loadfactor will decay p_estcpu to zero. 344b43179fbSJeff Roberson */ 345b43179fbSJeff Roberson static void 346b43179fbSJeff Roberson updatepri(struct ksegrp *kg) 347b43179fbSJeff Roberson { 348b43179fbSJeff Roberson register unsigned int newcpu; 349b43179fbSJeff Roberson register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 350b43179fbSJeff Roberson 351b43179fbSJeff Roberson newcpu = kg->kg_estcpu; 352b43179fbSJeff Roberson if (kg->kg_slptime > 5 * loadfac) 353b43179fbSJeff Roberson kg->kg_estcpu = 0; 354b43179fbSJeff Roberson else { 355b43179fbSJeff Roberson kg->kg_slptime--; /* the first time was done in schedcpu */ 356b43179fbSJeff Roberson while (newcpu && --kg->kg_slptime) 357b43179fbSJeff Roberson newcpu = decay_cpu(loadfac, newcpu); 358b43179fbSJeff Roberson kg->kg_estcpu = newcpu; 359b43179fbSJeff Roberson } 360b43179fbSJeff Roberson resetpriority(kg); 361b43179fbSJeff Roberson } 362b43179fbSJeff Roberson 363b43179fbSJeff Roberson /* 364b43179fbSJeff Roberson * Compute the priority of a process when running in user mode. 365b43179fbSJeff Roberson * Arrange to reschedule if the resulting priority is better 366b43179fbSJeff Roberson * than that of the current process. 367b43179fbSJeff Roberson */ 368b43179fbSJeff Roberson static void 369b43179fbSJeff Roberson resetpriority(struct ksegrp *kg) 370b43179fbSJeff Roberson { 371b43179fbSJeff Roberson register unsigned int newpriority; 372b43179fbSJeff Roberson struct thread *td; 373b43179fbSJeff Roberson 374b43179fbSJeff Roberson mtx_lock_spin(&sched_lock); 375b43179fbSJeff Roberson if (kg->kg_pri_class == PRI_TIMESHARE) { 376b43179fbSJeff Roberson newpriority = PUSER + kg->kg_estcpu / INVERSE_ESTCPU_WEIGHT + 377b43179fbSJeff Roberson NICE_WEIGHT * (kg->kg_nice - PRIO_MIN); 378b43179fbSJeff Roberson newpriority = min(max(newpriority, PRI_MIN_TIMESHARE), 379b43179fbSJeff Roberson PRI_MAX_TIMESHARE); 380b43179fbSJeff Roberson kg->kg_user_pri = newpriority; 381b43179fbSJeff Roberson } 382b43179fbSJeff Roberson FOREACH_THREAD_IN_GROUP(kg, td) { 383b43179fbSJeff Roberson maybe_resched(td); /* XXXKSE silly */ 384b43179fbSJeff Roberson } 385b43179fbSJeff Roberson mtx_unlock_spin(&sched_lock); 386b43179fbSJeff Roberson } 387b43179fbSJeff Roberson 388b43179fbSJeff Roberson /* ARGSUSED */ 389b43179fbSJeff Roberson static void 390b43179fbSJeff Roberson sched_setup(void *dummy) 391b43179fbSJeff Roberson { 392b43179fbSJeff Roberson if (sched_quantum == 0) 393b43179fbSJeff Roberson sched_quantum = SCHED_QUANTUM; 394b43179fbSJeff Roberson hogticks = 2 * sched_quantum; 395b43179fbSJeff Roberson 396b43179fbSJeff Roberson callout_init(&schedcpu_callout, 1); 397b43179fbSJeff Roberson callout_init(&roundrobin_callout, 0); 398b43179fbSJeff Roberson 399b43179fbSJeff Roberson /* Kick off timeout driven events by calling first time. */ 400b43179fbSJeff Roberson roundrobin(NULL); 401b43179fbSJeff Roberson schedcpu(NULL); 402b43179fbSJeff Roberson } 403b43179fbSJeff Roberson 404b43179fbSJeff Roberson /* External interfaces start here */ 405b43179fbSJeff Roberson int 406b43179fbSJeff Roberson sched_runnable(void) 407b43179fbSJeff Roberson { 408b43179fbSJeff Roberson return runq_check(&runq); 409b43179fbSJeff Roberson } 410b43179fbSJeff Roberson 411b43179fbSJeff Roberson int 412b43179fbSJeff Roberson sched_rr_interval(void) 413b43179fbSJeff Roberson { 414b43179fbSJeff Roberson if (sched_quantum == 0) 415b43179fbSJeff Roberson sched_quantum = SCHED_QUANTUM; 416b43179fbSJeff Roberson return (sched_quantum); 417b43179fbSJeff Roberson } 418b43179fbSJeff Roberson 419b43179fbSJeff Roberson /* 420b43179fbSJeff Roberson * We adjust the priority of the current process. The priority of 421b43179fbSJeff Roberson * a process gets worse as it accumulates CPU time. The cpu usage 422b43179fbSJeff Roberson * estimator (p_estcpu) is increased here. resetpriority() will 423b43179fbSJeff Roberson * compute a different priority each time p_estcpu increases by 424b43179fbSJeff Roberson * INVERSE_ESTCPU_WEIGHT 425b43179fbSJeff Roberson * (until MAXPRI is reached). The cpu usage estimator ramps up 426b43179fbSJeff Roberson * quite quickly when the process is running (linearly), and decays 427b43179fbSJeff Roberson * away exponentially, at a rate which is proportionally slower when 428b43179fbSJeff Roberson * the system is busy. The basic principle is that the system will 429b43179fbSJeff Roberson * 90% forget that the process used a lot of CPU time in 5 * loadav 430b43179fbSJeff Roberson * seconds. This causes the system to favor processes which haven't 431b43179fbSJeff Roberson * run much recently, and to round-robin among other processes. 432b43179fbSJeff Roberson */ 433b43179fbSJeff Roberson void 434b43179fbSJeff Roberson sched_clock(struct thread *td) 435b43179fbSJeff Roberson { 436b43179fbSJeff Roberson struct kse *ke; 437b43179fbSJeff Roberson struct ksegrp *kg; 438b43179fbSJeff Roberson 439b43179fbSJeff Roberson KASSERT((td != NULL), ("schedclock: null thread pointer")); 440b43179fbSJeff Roberson ke = td->td_kse; 441b43179fbSJeff Roberson kg = td->td_ksegrp; 442b43179fbSJeff Roberson ke->ke_cpticks++; 443b43179fbSJeff Roberson kg->kg_estcpu = ESTCPULIM(kg->kg_estcpu + 1); 444b43179fbSJeff Roberson if ((kg->kg_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) { 445b43179fbSJeff Roberson resetpriority(kg); 446b43179fbSJeff Roberson if (td->td_priority >= PUSER) 447b43179fbSJeff Roberson td->td_priority = kg->kg_user_pri; 448b43179fbSJeff Roberson } 449b43179fbSJeff Roberson } 450b43179fbSJeff Roberson /* 451b43179fbSJeff Roberson * charge childs scheduling cpu usage to parent. 452b43179fbSJeff Roberson * 453b43179fbSJeff Roberson * XXXKSE assume only one thread & kse & ksegrp keep estcpu in each ksegrp. 454b43179fbSJeff Roberson * Charge it to the ksegrp that did the wait since process estcpu is sum of 455b43179fbSJeff Roberson * all ksegrps, this is strictly as expected. Assume that the child process 456b43179fbSJeff Roberson * aggregated all the estcpu into the 'built-in' ksegrp. 457b43179fbSJeff Roberson */ 458b43179fbSJeff Roberson void 459b43179fbSJeff Roberson sched_exit(struct ksegrp *kg, struct ksegrp *child) 460b43179fbSJeff Roberson { 461b43179fbSJeff Roberson kg->kg_estcpu = ESTCPULIM(kg->kg_estcpu + child->kg_estcpu); 462b43179fbSJeff Roberson } 463b43179fbSJeff Roberson 464b43179fbSJeff Roberson void 465b43179fbSJeff Roberson sched_fork(struct ksegrp *kg, struct ksegrp *child) 466b43179fbSJeff Roberson { 467b43179fbSJeff Roberson /* 468b43179fbSJeff Roberson * set priority of child to be that of parent. 469b43179fbSJeff Roberson * XXXKSE this needs redefining.. 470b43179fbSJeff Roberson */ 471b43179fbSJeff Roberson child->kg_estcpu = kg->kg_estcpu; 472b43179fbSJeff Roberson } 473b43179fbSJeff Roberson 474b43179fbSJeff Roberson void 475b43179fbSJeff Roberson sched_nice(struct ksegrp *kg, int nice) 476b43179fbSJeff Roberson { 477b43179fbSJeff Roberson kg->kg_nice = nice; 478b43179fbSJeff Roberson resetpriority(kg); 479b43179fbSJeff Roberson } 480b43179fbSJeff Roberson 4811f955e2dSJulian Elischer /* 4821f955e2dSJulian Elischer * Adjust the priority of a thread. 4831f955e2dSJulian Elischer * This may include moving the thread within the KSEGRP, 4841f955e2dSJulian Elischer * changing the assignment of a kse to the thread, 4851f955e2dSJulian Elischer * and moving a KSE in the system run queue. 4861f955e2dSJulian Elischer */ 487b43179fbSJeff Roberson void 488b43179fbSJeff Roberson sched_prio(struct thread *td, u_char prio) 489b43179fbSJeff Roberson { 490b43179fbSJeff Roberson 491b43179fbSJeff Roberson if (TD_ON_RUNQ(td)) { 4921f955e2dSJulian Elischer adjustrunqueue(td, prio); 4931f955e2dSJulian Elischer } else { 4941f955e2dSJulian Elischer td->td_priority = prio; 495b43179fbSJeff Roberson } 496b43179fbSJeff Roberson } 497b43179fbSJeff Roberson 498b43179fbSJeff Roberson void 499b43179fbSJeff Roberson sched_sleep(struct thread *td, u_char prio) 500b43179fbSJeff Roberson { 501b43179fbSJeff Roberson td->td_ksegrp->kg_slptime = 0; 502b43179fbSJeff Roberson td->td_priority = prio; 503b43179fbSJeff Roberson } 504b43179fbSJeff Roberson 505b43179fbSJeff Roberson void 506b43179fbSJeff Roberson sched_switchin(struct thread *td) 507b43179fbSJeff Roberson { 508b43179fbSJeff Roberson td->td_kse->ke_oncpu = PCPU_GET(cpuid); 509b43179fbSJeff Roberson } 510b43179fbSJeff Roberson 511b43179fbSJeff Roberson void 512b43179fbSJeff Roberson sched_switchout(struct thread *td) 513b43179fbSJeff Roberson { 514b43179fbSJeff Roberson struct kse *ke; 515b43179fbSJeff Roberson struct proc *p; 516b43179fbSJeff Roberson 517b43179fbSJeff Roberson ke = td->td_kse; 518b43179fbSJeff Roberson p = td->td_proc; 519b43179fbSJeff Roberson 520b43179fbSJeff Roberson KASSERT((ke->ke_state == KES_THREAD), ("mi_switch: kse state?")); 521b43179fbSJeff Roberson 522b43179fbSJeff Roberson td->td_lastcpu = ke->ke_oncpu; 5231f955e2dSJulian Elischer td->td_last_kse = ke; 524b43179fbSJeff Roberson ke->ke_oncpu = NOCPU; 525b43179fbSJeff Roberson ke->ke_flags &= ~KEF_NEEDRESCHED; 526b43179fbSJeff Roberson /* 527b43179fbSJeff Roberson * At the last moment, if this thread is still marked RUNNING, 528b43179fbSJeff Roberson * then put it back on the run queue as it has not been suspended 529b43179fbSJeff Roberson * or stopped or any thing else similar. 530b43179fbSJeff Roberson */ 531b43179fbSJeff Roberson if (TD_IS_RUNNING(td)) { 532b43179fbSJeff Roberson /* Put us back on the run queue (kse and all). */ 533b43179fbSJeff Roberson setrunqueue(td); 534b43179fbSJeff Roberson } else if (p->p_flag & P_KSES) { 535b43179fbSJeff Roberson /* 536b43179fbSJeff Roberson * We will not be on the run queue. So we must be 537b43179fbSJeff Roberson * sleeping or similar. As it's available, 538b43179fbSJeff Roberson * someone else can use the KSE if they need it. 539b43179fbSJeff Roberson * (If bound LOANING can still occur). 540b43179fbSJeff Roberson */ 541b43179fbSJeff Roberson kse_reassign(ke); 542b43179fbSJeff Roberson } 543b43179fbSJeff Roberson } 544b43179fbSJeff Roberson 545b43179fbSJeff Roberson void 546b43179fbSJeff Roberson sched_wakeup(struct thread *td) 547b43179fbSJeff Roberson { 548b43179fbSJeff Roberson struct ksegrp *kg; 549b43179fbSJeff Roberson 550b43179fbSJeff Roberson kg = td->td_ksegrp; 551b43179fbSJeff Roberson if (kg->kg_slptime > 1) 552b43179fbSJeff Roberson updatepri(kg); 553b43179fbSJeff Roberson kg->kg_slptime = 0; 554b43179fbSJeff Roberson setrunqueue(td); 555b43179fbSJeff Roberson maybe_resched(td); 556b43179fbSJeff Roberson } 557b43179fbSJeff Roberson 558b43179fbSJeff Roberson void 559b43179fbSJeff Roberson sched_add(struct kse *ke) 560b43179fbSJeff Roberson { 561b43179fbSJeff Roberson mtx_assert(&sched_lock, MA_OWNED); 562b43179fbSJeff Roberson KASSERT((ke->ke_thread != NULL), ("runq_add: No thread on KSE")); 563b43179fbSJeff Roberson KASSERT((ke->ke_thread->td_kse != NULL), 564b43179fbSJeff Roberson ("runq_add: No KSE on thread")); 565b43179fbSJeff Roberson KASSERT(ke->ke_state != KES_ONRUNQ, 566b43179fbSJeff Roberson ("runq_add: kse %p (%s) already in run queue", ke, 567b43179fbSJeff Roberson ke->ke_proc->p_comm)); 568b43179fbSJeff Roberson KASSERT(ke->ke_proc->p_sflag & PS_INMEM, 569b43179fbSJeff Roberson ("runq_add: process swapped out")); 570b43179fbSJeff Roberson ke->ke_ksegrp->kg_runq_kses++; 571b43179fbSJeff Roberson ke->ke_state = KES_ONRUNQ; 572b43179fbSJeff Roberson 573b43179fbSJeff Roberson runq_add(&runq, ke); 574b43179fbSJeff Roberson } 575b43179fbSJeff Roberson 576b43179fbSJeff Roberson void 577b43179fbSJeff Roberson sched_rem(struct kse *ke) 578b43179fbSJeff Roberson { 579b43179fbSJeff Roberson KASSERT(ke->ke_proc->p_sflag & PS_INMEM, 580b43179fbSJeff Roberson ("runq_remove: process swapped out")); 581b43179fbSJeff Roberson KASSERT((ke->ke_state == KES_ONRUNQ), ("KSE not on run queue")); 582b43179fbSJeff Roberson mtx_assert(&sched_lock, MA_OWNED); 583b43179fbSJeff Roberson 584b43179fbSJeff Roberson runq_remove(&runq, ke); 585b43179fbSJeff Roberson ke->ke_state = KES_THREAD; 586b43179fbSJeff Roberson ke->ke_ksegrp->kg_runq_kses--; 587b43179fbSJeff Roberson } 588b43179fbSJeff Roberson 589b43179fbSJeff Roberson struct kse * 590b43179fbSJeff Roberson sched_choose(void) 591b43179fbSJeff Roberson { 592b43179fbSJeff Roberson struct kse *ke; 593b43179fbSJeff Roberson 594b43179fbSJeff Roberson ke = runq_choose(&runq); 595b43179fbSJeff Roberson 596b43179fbSJeff Roberson if (ke != NULL) { 597b43179fbSJeff Roberson runq_remove(&runq, ke); 598b43179fbSJeff Roberson ke->ke_state = KES_THREAD; 599b43179fbSJeff Roberson 600b43179fbSJeff Roberson KASSERT((ke->ke_thread != NULL), 601b43179fbSJeff Roberson ("runq_choose: No thread on KSE")); 602b43179fbSJeff Roberson KASSERT((ke->ke_thread->td_kse != NULL), 603b43179fbSJeff Roberson ("runq_choose: No KSE on thread")); 604b43179fbSJeff Roberson KASSERT(ke->ke_proc->p_sflag & PS_INMEM, 605b43179fbSJeff Roberson ("runq_choose: process swapped out")); 606b43179fbSJeff Roberson } 607b43179fbSJeff Roberson return (ke); 608b43179fbSJeff Roberson } 609b43179fbSJeff Roberson 610b43179fbSJeff Roberson void 611b43179fbSJeff Roberson sched_userret(struct thread *td) 612b43179fbSJeff Roberson { 613b43179fbSJeff Roberson struct ksegrp *kg; 614b43179fbSJeff Roberson /* 615b43179fbSJeff Roberson * XXX we cheat slightly on the locking here to avoid locking in 616b43179fbSJeff Roberson * the usual case. Setting td_priority here is essentially an 617b43179fbSJeff Roberson * incomplete workaround for not setting it properly elsewhere. 618b43179fbSJeff Roberson * Now that some interrupt handlers are threads, not setting it 619b43179fbSJeff Roberson * properly elsewhere can clobber it in the window between setting 620b43179fbSJeff Roberson * it here and returning to user mode, so don't waste time setting 621b43179fbSJeff Roberson * it perfectly here. 622b43179fbSJeff Roberson */ 623b43179fbSJeff Roberson kg = td->td_ksegrp; 624b43179fbSJeff Roberson if (td->td_priority != kg->kg_user_pri) { 625b43179fbSJeff Roberson mtx_lock_spin(&sched_lock); 626b43179fbSJeff Roberson td->td_priority = kg->kg_user_pri; 627b43179fbSJeff Roberson mtx_unlock_spin(&sched_lock); 628b43179fbSJeff Roberson } 629b43179fbSJeff Roberson } 630de028f5aSJeff Roberson 631de028f5aSJeff Roberson int 632de028f5aSJeff Roberson sched_sizeof_kse(void) 633de028f5aSJeff Roberson { 634de028f5aSJeff Roberson return (sizeof(struct kse)); 635de028f5aSJeff Roberson } 636de028f5aSJeff Roberson int 637de028f5aSJeff Roberson sched_sizeof_ksegrp(void) 638de028f5aSJeff Roberson { 639de028f5aSJeff Roberson return (sizeof(struct ksegrp)); 640de028f5aSJeff Roberson } 641de028f5aSJeff Roberson int 642de028f5aSJeff Roberson sched_sizeof_proc(void) 643de028f5aSJeff Roberson { 644de028f5aSJeff Roberson return (sizeof(struct proc)); 645de028f5aSJeff Roberson } 646de028f5aSJeff Roberson int 647de028f5aSJeff Roberson sched_sizeof_thread(void) 648de028f5aSJeff Roberson { 649de028f5aSJeff Roberson return (sizeof(struct thread)); 650de028f5aSJeff Roberson } 651