sched_4bsd.c revision 121127
1104964Sjeff/*- 2104964Sjeff * Copyright (c) 1982, 1986, 1990, 1991, 1993 3104964Sjeff * The Regents of the University of California. All rights reserved. 4104964Sjeff * (c) UNIX System Laboratories, Inc. 5104964Sjeff * All or some portions of this file are derived from material licensed 6104964Sjeff * to the University of California by American Telephone and Telegraph 7104964Sjeff * Co. or Unix System Laboratories, Inc. and are reproduced herein with 8104964Sjeff * the permission of UNIX System Laboratories, Inc. 9104964Sjeff * 10104964Sjeff * Redistribution and use in source and binary forms, with or without 11104964Sjeff * modification, are permitted provided that the following conditions 12104964Sjeff * are met: 13104964Sjeff * 1. Redistributions of source code must retain the above copyright 14104964Sjeff * notice, this list of conditions and the following disclaimer. 15104964Sjeff * 2. Redistributions in binary form must reproduce the above copyright 16104964Sjeff * notice, this list of conditions and the following disclaimer in the 17104964Sjeff * documentation and/or other materials provided with the distribution. 18104964Sjeff * 3. All advertising materials mentioning features or use of this software 19104964Sjeff * must display the following acknowledgement: 20104964Sjeff * This product includes software developed by the University of 21104964Sjeff * California, Berkeley and its contributors. 22104964Sjeff * 4. Neither the name of the University nor the names of its contributors 23104964Sjeff * may be used to endorse or promote products derived from this software 24104964Sjeff * without specific prior written permission. 25104964Sjeff * 26104964Sjeff * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 27104964Sjeff * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 28104964Sjeff * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 29104964Sjeff * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 30104964Sjeff * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 31104964Sjeff * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 32104964Sjeff * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 33104964Sjeff * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 34104964Sjeff * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 35104964Sjeff * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 36104964Sjeff * SUCH DAMAGE. 37104964Sjeff */ 38104964Sjeff 39116182Sobrien#include <sys/cdefs.h> 40116182Sobrien__FBSDID("$FreeBSD: head/sys/kern/sched_4bsd.c 121127 2003-10-16 08:39:15Z jeff $"); 41116182Sobrien 42104964Sjeff#include <sys/param.h> 43104964Sjeff#include <sys/systm.h> 44104964Sjeff#include <sys/kernel.h> 45104964Sjeff#include <sys/ktr.h> 46104964Sjeff#include <sys/lock.h> 47104964Sjeff#include <sys/mutex.h> 48104964Sjeff#include <sys/proc.h> 49104964Sjeff#include <sys/resourcevar.h> 50104964Sjeff#include <sys/sched.h> 51104964Sjeff#include <sys/smp.h> 52104964Sjeff#include <sys/sysctl.h> 53104964Sjeff#include <sys/sx.h> 54104964Sjeff 55107135Sjeff/* 56107135Sjeff * INVERSE_ESTCPU_WEIGHT is only suitable for statclock() frequencies in 57107135Sjeff * the range 100-256 Hz (approximately). 58107135Sjeff */ 59107135Sjeff#define ESTCPULIM(e) \ 60107135Sjeff min((e), INVERSE_ESTCPU_WEIGHT * (NICE_WEIGHT * (PRIO_MAX - PRIO_MIN) - \ 61107135Sjeff RQ_PPQ) + INVERSE_ESTCPU_WEIGHT - 1) 62107135Sjeff#define INVERSE_ESTCPU_WEIGHT 8 /* 1 / (priorities per estcpu level). */ 63107135Sjeff#define NICE_WEIGHT 1 /* Priorities per nice level. */ 64107135Sjeff 65109145Sjeffstruct ke_sched { 66109145Sjeff int ske_cpticks; /* (j) Ticks of cpu time. */ 67109145Sjeff}; 68109145Sjeff 69114293Smarkmstatic struct ke_sched ke_sched; 70109145Sjeff 71109145Sjeffstruct ke_sched *kse0_sched = &ke_sched; 72107126Sjeffstruct kg_sched *ksegrp0_sched = NULL; 73107126Sjeffstruct p_sched *proc0_sched = NULL; 74107126Sjeffstruct td_sched *thread0_sched = NULL; 75104964Sjeff 76104964Sjeffstatic int sched_quantum; /* Roundrobin scheduling quantum in ticks. */ 77112535Smux#define SCHED_QUANTUM (hz / 10) /* Default sched quantum */ 78104964Sjeff 79104964Sjeffstatic struct callout schedcpu_callout; 80104964Sjeffstatic struct callout roundrobin_callout; 81104964Sjeff 82104964Sjeffstatic void roundrobin(void *arg); 83104964Sjeffstatic void schedcpu(void *arg); 84104964Sjeffstatic void sched_setup(void *dummy); 85104964Sjeffstatic void maybe_resched(struct thread *td); 86104964Sjeffstatic void updatepri(struct ksegrp *kg); 87104964Sjeffstatic void resetpriority(struct ksegrp *kg); 88104964Sjeff 89104964SjeffSYSINIT(sched_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, sched_setup, NULL) 90104964Sjeff 91104964Sjeff/* 92104964Sjeff * Global run queue. 93104964Sjeff */ 94104964Sjeffstatic struct runq runq; 95104964SjeffSYSINIT(runq, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, runq_init, &runq) 96104964Sjeff 97104964Sjeffstatic int 98104964Sjeffsysctl_kern_quantum(SYSCTL_HANDLER_ARGS) 99104964Sjeff{ 100104964Sjeff int error, new_val; 101104964Sjeff 102104964Sjeff new_val = sched_quantum * tick; 103104964Sjeff error = sysctl_handle_int(oidp, &new_val, 0, req); 104104964Sjeff if (error != 0 || req->newptr == NULL) 105104964Sjeff return (error); 106104964Sjeff if (new_val < tick) 107104964Sjeff return (EINVAL); 108104964Sjeff sched_quantum = new_val / tick; 109104964Sjeff hogticks = 2 * sched_quantum; 110104964Sjeff return (0); 111104964Sjeff} 112104964Sjeff 113104964SjeffSYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW, 114104964Sjeff 0, sizeof sched_quantum, sysctl_kern_quantum, "I", 115104964Sjeff "Roundrobin scheduling quantum in microseconds"); 116104964Sjeff 117104964Sjeff/* 118104964Sjeff * Arrange to reschedule if necessary, taking the priorities and 119104964Sjeff * schedulers into account. 120104964Sjeff */ 121104964Sjeffstatic void 122104964Sjeffmaybe_resched(struct thread *td) 123104964Sjeff{ 124104964Sjeff 125104964Sjeff mtx_assert(&sched_lock, MA_OWNED); 126108338Sjulian if (td->td_priority < curthread->td_priority && curthread->td_kse) 127111032Sjulian curthread->td_flags |= TDF_NEEDRESCHED; 128104964Sjeff} 129104964Sjeff 130104964Sjeff/* 131104964Sjeff * Force switch among equal priority processes every 100ms. 132104964Sjeff * We don't actually need to force a context switch of the current process. 133104964Sjeff * The act of firing the event triggers a context switch to softclock() and 134104964Sjeff * then switching back out again which is equivalent to a preemption, thus 135104964Sjeff * no further work is needed on the local CPU. 136104964Sjeff */ 137104964Sjeff/* ARGSUSED */ 138104964Sjeffstatic void 139104964Sjeffroundrobin(void *arg) 140104964Sjeff{ 141104964Sjeff 142104964Sjeff#ifdef SMP 143104964Sjeff mtx_lock_spin(&sched_lock); 144104964Sjeff forward_roundrobin(); 145104964Sjeff mtx_unlock_spin(&sched_lock); 146104964Sjeff#endif 147104964Sjeff 148104964Sjeff callout_reset(&roundrobin_callout, sched_quantum, roundrobin, NULL); 149104964Sjeff} 150104964Sjeff 151104964Sjeff/* 152104964Sjeff * Constants for digital decay and forget: 153118972Sjhb * 90% of (kg_estcpu) usage in 5 * loadav time 154118972Sjhb * 95% of (ke_pctcpu) usage in 60 seconds (load insensitive) 155104964Sjeff * Note that, as ps(1) mentions, this can let percentages 156104964Sjeff * total over 100% (I've seen 137.9% for 3 processes). 157104964Sjeff * 158118972Sjhb * Note that schedclock() updates kg_estcpu and p_cpticks asynchronously. 159104964Sjeff * 160118972Sjhb * We wish to decay away 90% of kg_estcpu in (5 * loadavg) seconds. 161104964Sjeff * That is, the system wants to compute a value of decay such 162104964Sjeff * that the following for loop: 163104964Sjeff * for (i = 0; i < (5 * loadavg); i++) 164118972Sjhb * kg_estcpu *= decay; 165104964Sjeff * will compute 166118972Sjhb * kg_estcpu *= 0.1; 167104964Sjeff * for all values of loadavg: 168104964Sjeff * 169104964Sjeff * Mathematically this loop can be expressed by saying: 170104964Sjeff * decay ** (5 * loadavg) ~= .1 171104964Sjeff * 172104964Sjeff * The system computes decay as: 173104964Sjeff * decay = (2 * loadavg) / (2 * loadavg + 1) 174104964Sjeff * 175104964Sjeff * We wish to prove that the system's computation of decay 176104964Sjeff * will always fulfill the equation: 177104964Sjeff * decay ** (5 * loadavg) ~= .1 178104964Sjeff * 179104964Sjeff * If we compute b as: 180104964Sjeff * b = 2 * loadavg 181104964Sjeff * then 182104964Sjeff * decay = b / (b + 1) 183104964Sjeff * 184104964Sjeff * We now need to prove two things: 185104964Sjeff * 1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1) 186104964Sjeff * 2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg) 187104964Sjeff * 188104964Sjeff * Facts: 189104964Sjeff * For x close to zero, exp(x) =~ 1 + x, since 190104964Sjeff * exp(x) = 0! + x**1/1! + x**2/2! + ... . 191104964Sjeff * therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b. 192104964Sjeff * For x close to zero, ln(1+x) =~ x, since 193104964Sjeff * ln(1+x) = x - x**2/2 + x**3/3 - ... -1 < x < 1 194104964Sjeff * therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1). 195104964Sjeff * ln(.1) =~ -2.30 196104964Sjeff * 197104964Sjeff * Proof of (1): 198104964Sjeff * Solve (factor)**(power) =~ .1 given power (5*loadav): 199104964Sjeff * solving for factor, 200104964Sjeff * ln(factor) =~ (-2.30/5*loadav), or 201104964Sjeff * factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) = 202104964Sjeff * exp(-1/b) =~ (b-1)/b =~ b/(b+1). QED 203104964Sjeff * 204104964Sjeff * Proof of (2): 205104964Sjeff * Solve (factor)**(power) =~ .1 given factor == (b/(b+1)): 206104964Sjeff * solving for power, 207104964Sjeff * power*ln(b/(b+1)) =~ -2.30, or 208104964Sjeff * power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav. QED 209104964Sjeff * 210104964Sjeff * Actual power values for the implemented algorithm are as follows: 211104964Sjeff * loadav: 1 2 3 4 212104964Sjeff * power: 5.68 10.32 14.94 19.55 213104964Sjeff */ 214104964Sjeff 215104964Sjeff/* calculations for digital decay to forget 90% of usage in 5*loadav sec */ 216104964Sjeff#define loadfactor(loadav) (2 * (loadav)) 217104964Sjeff#define decay_cpu(loadfac, cpu) (((loadfac) * (cpu)) / ((loadfac) + FSCALE)) 218104964Sjeff 219118972Sjhb/* decay 95% of `ke_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */ 220104964Sjeffstatic fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */ 221104964SjeffSYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 222104964Sjeff 223104964Sjeff/* 224104964Sjeff * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the 225104964Sjeff * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below 226104964Sjeff * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT). 227104964Sjeff * 228104964Sjeff * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used: 229104964Sjeff * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits). 230104964Sjeff * 231104964Sjeff * If you don't want to bother with the faster/more-accurate formula, you 232104964Sjeff * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate 233104964Sjeff * (more general) method of calculating the %age of CPU used by a process. 234104964Sjeff */ 235104964Sjeff#define CCPU_SHIFT 11 236104964Sjeff 237104964Sjeff/* 238104964Sjeff * Recompute process priorities, every hz ticks. 239104964Sjeff * MP-safe, called without the Giant mutex. 240104964Sjeff */ 241104964Sjeff/* ARGSUSED */ 242104964Sjeffstatic void 243104964Sjeffschedcpu(void *arg) 244104964Sjeff{ 245104964Sjeff register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 246104964Sjeff struct thread *td; 247104964Sjeff struct proc *p; 248104964Sjeff struct kse *ke; 249104964Sjeff struct ksegrp *kg; 250118972Sjhb int awake, realstathz; 251104964Sjeff 252104964Sjeff realstathz = stathz ? stathz : hz; 253104964Sjeff sx_slock(&allproc_lock); 254104964Sjeff FOREACH_PROC_IN_SYSTEM(p) { 255118972Sjhb /* 256118972Sjhb * Prevent state changes and protect run queue. 257118972Sjhb */ 258104964Sjeff mtx_lock_spin(&sched_lock); 259118972Sjhb /* 260118972Sjhb * Increment time in/out of memory. We ignore overflow; with 261118972Sjhb * 16-bit int's (remember them?) overflow takes 45 days. 262118972Sjhb */ 263104964Sjeff p->p_swtime++; 264104964Sjeff FOREACH_KSEGRP_IN_PROC(p, kg) { 265104964Sjeff awake = 0; 266104964Sjeff FOREACH_KSE_IN_GROUP(kg, ke) { 267104964Sjeff /* 268118972Sjhb * Increment sleep time (if sleeping). We 269118972Sjhb * ignore overflow, as above. 270104964Sjeff */ 271104964Sjeff /* 272104964Sjeff * The kse slptimes are not touched in wakeup 273104964Sjeff * because the thread may not HAVE a KSE. 274104964Sjeff */ 275104964Sjeff if (ke->ke_state == KES_ONRUNQ) { 276104964Sjeff awake = 1; 277104964Sjeff ke->ke_flags &= ~KEF_DIDRUN; 278104964Sjeff } else if ((ke->ke_state == KES_THREAD) && 279104964Sjeff (TD_IS_RUNNING(ke->ke_thread))) { 280104964Sjeff awake = 1; 281104964Sjeff /* Do not clear KEF_DIDRUN */ 282104964Sjeff } else if (ke->ke_flags & KEF_DIDRUN) { 283104964Sjeff awake = 1; 284104964Sjeff ke->ke_flags &= ~KEF_DIDRUN; 285104964Sjeff } 286104964Sjeff 287104964Sjeff /* 288118972Sjhb * ke_pctcpu is only for ps and ttyinfo(). 289118972Sjhb * Do it per kse, and add them up at the end? 290104964Sjeff * XXXKSE 291104964Sjeff */ 292118972Sjhb ke->ke_pctcpu = (ke->ke_pctcpu * ccpu) >> 293109145Sjeff FSHIFT; 294104964Sjeff /* 295104964Sjeff * If the kse has been idle the entire second, 296104964Sjeff * stop recalculating its priority until 297104964Sjeff * it wakes up. 298104964Sjeff */ 299109145Sjeff if (ke->ke_sched->ske_cpticks == 0) 300104964Sjeff continue; 301104964Sjeff#if (FSHIFT >= CCPU_SHIFT) 302109157Sjeff ke->ke_pctcpu += (realstathz == 100) 303109145Sjeff ? ((fixpt_t) ke->ke_sched->ske_cpticks) << 304104964Sjeff (FSHIFT - CCPU_SHIFT) : 305109145Sjeff 100 * (((fixpt_t) ke->ke_sched->ske_cpticks) 306109145Sjeff << (FSHIFT - CCPU_SHIFT)) / realstathz; 307104964Sjeff#else 308109157Sjeff ke->ke_pctcpu += ((FSCALE - ccpu) * 309109145Sjeff (ke->ke_sched->ske_cpticks * 310109145Sjeff FSCALE / realstathz)) >> FSHIFT; 311104964Sjeff#endif 312109145Sjeff ke->ke_sched->ske_cpticks = 0; 313104964Sjeff } /* end of kse loop */ 314104964Sjeff /* 315104964Sjeff * If there are ANY running threads in this KSEGRP, 316104964Sjeff * then don't count it as sleeping. 317104964Sjeff */ 318104964Sjeff if (awake) { 319104964Sjeff if (kg->kg_slptime > 1) { 320104964Sjeff /* 321104964Sjeff * In an ideal world, this should not 322104964Sjeff * happen, because whoever woke us 323104964Sjeff * up from the long sleep should have 324104964Sjeff * unwound the slptime and reset our 325104964Sjeff * priority before we run at the stale 326104964Sjeff * priority. Should KASSERT at some 327104964Sjeff * point when all the cases are fixed. 328104964Sjeff */ 329104964Sjeff updatepri(kg); 330104964Sjeff } 331104964Sjeff kg->kg_slptime = 0; 332118972Sjhb } else 333104964Sjeff kg->kg_slptime++; 334104964Sjeff if (kg->kg_slptime > 1) 335104964Sjeff continue; 336104964Sjeff kg->kg_estcpu = decay_cpu(loadfac, kg->kg_estcpu); 337104964Sjeff resetpriority(kg); 338104964Sjeff FOREACH_THREAD_IN_GROUP(kg, td) { 339104964Sjeff if (td->td_priority >= PUSER) { 340105127Sjulian sched_prio(td, kg->kg_user_pri); 341104964Sjeff } 342104964Sjeff } 343104964Sjeff } /* end of ksegrp loop */ 344104964Sjeff mtx_unlock_spin(&sched_lock); 345104964Sjeff } /* end of process loop */ 346104964Sjeff sx_sunlock(&allproc_lock); 347104964Sjeff callout_reset(&schedcpu_callout, hz, schedcpu, NULL); 348104964Sjeff} 349104964Sjeff 350104964Sjeff/* 351104964Sjeff * Recalculate the priority of a process after it has slept for a while. 352118972Sjhb * For all load averages >= 1 and max kg_estcpu of 255, sleeping for at 353118972Sjhb * least six times the loadfactor will decay kg_estcpu to zero. 354104964Sjeff */ 355104964Sjeffstatic void 356104964Sjeffupdatepri(struct ksegrp *kg) 357104964Sjeff{ 358118972Sjhb register fixpt_t loadfac; 359104964Sjeff register unsigned int newcpu; 360104964Sjeff 361118972Sjhb loadfac = loadfactor(averunnable.ldavg[0]); 362104964Sjeff if (kg->kg_slptime > 5 * loadfac) 363104964Sjeff kg->kg_estcpu = 0; 364104964Sjeff else { 365118972Sjhb newcpu = kg->kg_estcpu; 366118972Sjhb kg->kg_slptime--; /* was incremented in schedcpu() */ 367104964Sjeff while (newcpu && --kg->kg_slptime) 368104964Sjeff newcpu = decay_cpu(loadfac, newcpu); 369104964Sjeff kg->kg_estcpu = newcpu; 370104964Sjeff } 371104964Sjeff resetpriority(kg); 372104964Sjeff} 373104964Sjeff 374104964Sjeff/* 375104964Sjeff * Compute the priority of a process when running in user mode. 376104964Sjeff * Arrange to reschedule if the resulting priority is better 377104964Sjeff * than that of the current process. 378104964Sjeff */ 379104964Sjeffstatic void 380104964Sjeffresetpriority(struct ksegrp *kg) 381104964Sjeff{ 382104964Sjeff register unsigned int newpriority; 383104964Sjeff struct thread *td; 384104964Sjeff 385104964Sjeff if (kg->kg_pri_class == PRI_TIMESHARE) { 386104964Sjeff newpriority = PUSER + kg->kg_estcpu / INVERSE_ESTCPU_WEIGHT + 387104964Sjeff NICE_WEIGHT * (kg->kg_nice - PRIO_MIN); 388104964Sjeff newpriority = min(max(newpriority, PRI_MIN_TIMESHARE), 389104964Sjeff PRI_MAX_TIMESHARE); 390104964Sjeff kg->kg_user_pri = newpriority; 391104964Sjeff } 392104964Sjeff FOREACH_THREAD_IN_GROUP(kg, td) { 393104964Sjeff maybe_resched(td); /* XXXKSE silly */ 394104964Sjeff } 395104964Sjeff} 396104964Sjeff 397104964Sjeff/* ARGSUSED */ 398104964Sjeffstatic void 399104964Sjeffsched_setup(void *dummy) 400104964Sjeff{ 401118972Sjhb 402104964Sjeff if (sched_quantum == 0) 403104964Sjeff sched_quantum = SCHED_QUANTUM; 404104964Sjeff hogticks = 2 * sched_quantum; 405104964Sjeff 406119137Ssam callout_init(&schedcpu_callout, CALLOUT_MPSAFE); 407104964Sjeff callout_init(&roundrobin_callout, 0); 408104964Sjeff 409104964Sjeff /* Kick off timeout driven events by calling first time. */ 410104964Sjeff roundrobin(NULL); 411104964Sjeff schedcpu(NULL); 412104964Sjeff} 413104964Sjeff 414104964Sjeff/* External interfaces start here */ 415104964Sjeffint 416104964Sjeffsched_runnable(void) 417104964Sjeff{ 418104964Sjeff return runq_check(&runq); 419104964Sjeff} 420104964Sjeff 421104964Sjeffint 422104964Sjeffsched_rr_interval(void) 423104964Sjeff{ 424104964Sjeff if (sched_quantum == 0) 425104964Sjeff sched_quantum = SCHED_QUANTUM; 426104964Sjeff return (sched_quantum); 427104964Sjeff} 428104964Sjeff 429104964Sjeff/* 430104964Sjeff * We adjust the priority of the current process. The priority of 431104964Sjeff * a process gets worse as it accumulates CPU time. The cpu usage 432118972Sjhb * estimator (kg_estcpu) is increased here. resetpriority() will 433118972Sjhb * compute a different priority each time kg_estcpu increases by 434104964Sjeff * INVERSE_ESTCPU_WEIGHT 435104964Sjeff * (until MAXPRI is reached). The cpu usage estimator ramps up 436104964Sjeff * quite quickly when the process is running (linearly), and decays 437104964Sjeff * away exponentially, at a rate which is proportionally slower when 438104964Sjeff * the system is busy. The basic principle is that the system will 439104964Sjeff * 90% forget that the process used a lot of CPU time in 5 * loadav 440104964Sjeff * seconds. This causes the system to favor processes which haven't 441104964Sjeff * run much recently, and to round-robin among other processes. 442104964Sjeff */ 443104964Sjeffvoid 444121127Sjeffsched_clock(struct thread *td) 445104964Sjeff{ 446104964Sjeff struct ksegrp *kg; 447121127Sjeff struct kse *ke; 448104964Sjeff 449113923Sjhb mtx_assert(&sched_lock, MA_OWNED); 450121127Sjeff kg = td->td_ksegrp; 451121127Sjeff ke = td->td_kse; 452113356Sjeff 453109145Sjeff ke->ke_sched->ske_cpticks++; 454104964Sjeff kg->kg_estcpu = ESTCPULIM(kg->kg_estcpu + 1); 455104964Sjeff if ((kg->kg_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) { 456104964Sjeff resetpriority(kg); 457104964Sjeff if (td->td_priority >= PUSER) 458104964Sjeff td->td_priority = kg->kg_user_pri; 459104964Sjeff } 460104964Sjeff} 461118972Sjhb 462104964Sjeff/* 463104964Sjeff * charge childs scheduling cpu usage to parent. 464104964Sjeff * 465104964Sjeff * XXXKSE assume only one thread & kse & ksegrp keep estcpu in each ksegrp. 466104964Sjeff * Charge it to the ksegrp that did the wait since process estcpu is sum of 467104964Sjeff * all ksegrps, this is strictly as expected. Assume that the child process 468104964Sjeff * aggregated all the estcpu into the 'built-in' ksegrp. 469104964Sjeff */ 470104964Sjeffvoid 471113356Sjeffsched_exit(struct proc *p, struct proc *p1) 472104964Sjeff{ 473113356Sjeff sched_exit_kse(FIRST_KSE_IN_PROC(p), FIRST_KSE_IN_PROC(p1)); 474113356Sjeff sched_exit_ksegrp(FIRST_KSEGRP_IN_PROC(p), FIRST_KSEGRP_IN_PROC(p1)); 475113356Sjeff sched_exit_thread(FIRST_THREAD_IN_PROC(p), FIRST_THREAD_IN_PROC(p1)); 476113356Sjeff} 477113356Sjeff 478113356Sjeffvoid 479113356Sjeffsched_exit_kse(struct kse *ke, struct kse *child) 480113356Sjeff{ 481113356Sjeff} 482113356Sjeff 483113356Sjeffvoid 484113356Sjeffsched_exit_ksegrp(struct ksegrp *kg, struct ksegrp *child) 485113356Sjeff{ 486113923Sjhb 487113923Sjhb mtx_assert(&sched_lock, MA_OWNED); 488104964Sjeff kg->kg_estcpu = ESTCPULIM(kg->kg_estcpu + child->kg_estcpu); 489104964Sjeff} 490104964Sjeff 491104964Sjeffvoid 492113356Sjeffsched_exit_thread(struct thread *td, struct thread *child) 493104964Sjeff{ 494113356Sjeff} 495109145Sjeff 496113356Sjeffvoid 497113356Sjeffsched_fork(struct proc *p, struct proc *p1) 498113356Sjeff{ 499113356Sjeff sched_fork_kse(FIRST_KSE_IN_PROC(p), FIRST_KSE_IN_PROC(p1)); 500113356Sjeff sched_fork_ksegrp(FIRST_KSEGRP_IN_PROC(p), FIRST_KSEGRP_IN_PROC(p1)); 501113356Sjeff sched_fork_thread(FIRST_THREAD_IN_PROC(p), FIRST_THREAD_IN_PROC(p1)); 502113356Sjeff} 503113356Sjeff 504113356Sjeffvoid 505113356Sjeffsched_fork_kse(struct kse *ke, struct kse *child) 506113356Sjeff{ 507113356Sjeff child->ke_sched->ske_cpticks = 0; 508113356Sjeff} 509113356Sjeff 510113356Sjeffvoid 511113356Sjeffsched_fork_ksegrp(struct ksegrp *kg, struct ksegrp *child) 512113356Sjeff{ 513113923Sjhb mtx_assert(&sched_lock, MA_OWNED); 514104964Sjeff child->kg_estcpu = kg->kg_estcpu; 515113356Sjeff} 516109145Sjeff 517113356Sjeffvoid 518113356Sjeffsched_fork_thread(struct thread *td, struct thread *child) 519113356Sjeff{ 520104964Sjeff} 521104964Sjeff 522104964Sjeffvoid 523104964Sjeffsched_nice(struct ksegrp *kg, int nice) 524104964Sjeff{ 525113873Sjhb 526113873Sjhb PROC_LOCK_ASSERT(kg->kg_proc, MA_OWNED); 527113873Sjhb mtx_assert(&sched_lock, MA_OWNED); 528104964Sjeff kg->kg_nice = nice; 529104964Sjeff resetpriority(kg); 530104964Sjeff} 531104964Sjeff 532113356Sjeffvoid 533113356Sjeffsched_class(struct ksegrp *kg, int class) 534113356Sjeff{ 535113923Sjhb mtx_assert(&sched_lock, MA_OWNED); 536113356Sjeff kg->kg_pri_class = class; 537113356Sjeff} 538113356Sjeff 539105127Sjulian/* 540105127Sjulian * Adjust the priority of a thread. 541105127Sjulian * This may include moving the thread within the KSEGRP, 542105127Sjulian * changing the assignment of a kse to the thread, 543105127Sjulian * and moving a KSE in the system run queue. 544105127Sjulian */ 545104964Sjeffvoid 546104964Sjeffsched_prio(struct thread *td, u_char prio) 547104964Sjeff{ 548104964Sjeff 549113923Sjhb mtx_assert(&sched_lock, MA_OWNED); 550104964Sjeff if (TD_ON_RUNQ(td)) { 551105127Sjulian adjustrunqueue(td, prio); 552105127Sjulian } else { 553105127Sjulian td->td_priority = prio; 554104964Sjeff } 555104964Sjeff} 556104964Sjeff 557104964Sjeffvoid 558104964Sjeffsched_sleep(struct thread *td, u_char prio) 559104964Sjeff{ 560113923Sjhb 561113923Sjhb mtx_assert(&sched_lock, MA_OWNED); 562104964Sjeff td->td_ksegrp->kg_slptime = 0; 563104964Sjeff td->td_priority = prio; 564104964Sjeff} 565104964Sjeff 566104964Sjeffvoid 567104964Sjeffsched_switchin(struct thread *td) 568104964Sjeff{ 569113923Sjhb 570113923Sjhb mtx_assert(&sched_lock, MA_OWNED); 571113339Sjulian td->td_oncpu = PCPU_GET(cpuid); 572104964Sjeff} 573104964Sjeff 574104964Sjeffvoid 575104964Sjeffsched_switchout(struct thread *td) 576104964Sjeff{ 577104964Sjeff struct kse *ke; 578104964Sjeff struct proc *p; 579104964Sjeff 580104964Sjeff ke = td->td_kse; 581104964Sjeff p = td->td_proc; 582104964Sjeff 583113923Sjhb mtx_assert(&sched_lock, MA_OWNED); 584104964Sjeff KASSERT((ke->ke_state == KES_THREAD), ("mi_switch: kse state?")); 585104964Sjeff 586113339Sjulian td->td_lastcpu = td->td_oncpu; 587105127Sjulian td->td_last_kse = ke; 588113339Sjulian td->td_oncpu = NOCPU; 589111032Sjulian td->td_flags &= ~TDF_NEEDRESCHED; 590104964Sjeff /* 591104964Sjeff * At the last moment, if this thread is still marked RUNNING, 592104964Sjeff * then put it back on the run queue as it has not been suspended 593104964Sjeff * or stopped or any thing else similar. 594104964Sjeff */ 595104964Sjeff if (TD_IS_RUNNING(td)) { 596104964Sjeff /* Put us back on the run queue (kse and all). */ 597104964Sjeff setrunqueue(td); 598116361Sdavidxu } else if (p->p_flag & P_SA) { 599104964Sjeff /* 600104964Sjeff * We will not be on the run queue. So we must be 601104964Sjeff * sleeping or similar. As it's available, 602104964Sjeff * someone else can use the KSE if they need it. 603104964Sjeff */ 604104964Sjeff kse_reassign(ke); 605104964Sjeff } 606104964Sjeff} 607104964Sjeff 608104964Sjeffvoid 609104964Sjeffsched_wakeup(struct thread *td) 610104964Sjeff{ 611104964Sjeff struct ksegrp *kg; 612104964Sjeff 613113923Sjhb mtx_assert(&sched_lock, MA_OWNED); 614104964Sjeff kg = td->td_ksegrp; 615104964Sjeff if (kg->kg_slptime > 1) 616104964Sjeff updatepri(kg); 617104964Sjeff kg->kg_slptime = 0; 618104964Sjeff setrunqueue(td); 619104964Sjeff maybe_resched(td); 620104964Sjeff} 621104964Sjeff 622104964Sjeffvoid 623121127Sjeffsched_add(struct thread *td) 624104964Sjeff{ 625121127Sjeff struct kse *ke; 626121127Sjeff 627121127Sjeff ke = td->td_kse; 628104964Sjeff mtx_assert(&sched_lock, MA_OWNED); 629104964Sjeff KASSERT((ke->ke_thread != NULL), ("runq_add: No thread on KSE")); 630104964Sjeff KASSERT((ke->ke_thread->td_kse != NULL), 631104964Sjeff ("runq_add: No KSE on thread")); 632104964Sjeff KASSERT(ke->ke_state != KES_ONRUNQ, 633104964Sjeff ("runq_add: kse %p (%s) already in run queue", ke, 634104964Sjeff ke->ke_proc->p_comm)); 635104964Sjeff KASSERT(ke->ke_proc->p_sflag & PS_INMEM, 636104964Sjeff ("runq_add: process swapped out")); 637104964Sjeff ke->ke_ksegrp->kg_runq_kses++; 638104964Sjeff ke->ke_state = KES_ONRUNQ; 639104964Sjeff 640104964Sjeff runq_add(&runq, ke); 641104964Sjeff} 642104964Sjeff 643104964Sjeffvoid 644121127Sjeffsched_rem(struct thread *td) 645104964Sjeff{ 646121127Sjeff struct kse *ke; 647121127Sjeff 648121127Sjeff ke = td->td_kse; 649104964Sjeff KASSERT(ke->ke_proc->p_sflag & PS_INMEM, 650104964Sjeff ("runq_remove: process swapped out")); 651104964Sjeff KASSERT((ke->ke_state == KES_ONRUNQ), ("KSE not on run queue")); 652104964Sjeff mtx_assert(&sched_lock, MA_OWNED); 653104964Sjeff 654104964Sjeff runq_remove(&runq, ke); 655104964Sjeff ke->ke_state = KES_THREAD; 656104964Sjeff ke->ke_ksegrp->kg_runq_kses--; 657104964Sjeff} 658104964Sjeff 659104964Sjeffstruct kse * 660104964Sjeffsched_choose(void) 661104964Sjeff{ 662104964Sjeff struct kse *ke; 663104964Sjeff 664104964Sjeff ke = runq_choose(&runq); 665104964Sjeff 666104964Sjeff if (ke != NULL) { 667104964Sjeff runq_remove(&runq, ke); 668104964Sjeff ke->ke_state = KES_THREAD; 669104964Sjeff 670104964Sjeff KASSERT((ke->ke_thread != NULL), 671104964Sjeff ("runq_choose: No thread on KSE")); 672104964Sjeff KASSERT((ke->ke_thread->td_kse != NULL), 673104964Sjeff ("runq_choose: No KSE on thread")); 674104964Sjeff KASSERT(ke->ke_proc->p_sflag & PS_INMEM, 675104964Sjeff ("runq_choose: process swapped out")); 676104964Sjeff } 677104964Sjeff return (ke); 678104964Sjeff} 679104964Sjeff 680104964Sjeffvoid 681104964Sjeffsched_userret(struct thread *td) 682104964Sjeff{ 683104964Sjeff struct ksegrp *kg; 684104964Sjeff /* 685104964Sjeff * XXX we cheat slightly on the locking here to avoid locking in 686104964Sjeff * the usual case. Setting td_priority here is essentially an 687104964Sjeff * incomplete workaround for not setting it properly elsewhere. 688104964Sjeff * Now that some interrupt handlers are threads, not setting it 689104964Sjeff * properly elsewhere can clobber it in the window between setting 690104964Sjeff * it here and returning to user mode, so don't waste time setting 691104964Sjeff * it perfectly here. 692104964Sjeff */ 693104964Sjeff kg = td->td_ksegrp; 694104964Sjeff if (td->td_priority != kg->kg_user_pri) { 695104964Sjeff mtx_lock_spin(&sched_lock); 696104964Sjeff td->td_priority = kg->kg_user_pri; 697104964Sjeff mtx_unlock_spin(&sched_lock); 698104964Sjeff } 699104964Sjeff} 700107126Sjeff 701107126Sjeffint 702107126Sjeffsched_sizeof_kse(void) 703107126Sjeff{ 704109145Sjeff return (sizeof(struct kse) + sizeof(struct ke_sched)); 705107126Sjeff} 706107126Sjeffint 707107126Sjeffsched_sizeof_ksegrp(void) 708107126Sjeff{ 709107126Sjeff return (sizeof(struct ksegrp)); 710107126Sjeff} 711107126Sjeffint 712107126Sjeffsched_sizeof_proc(void) 713107126Sjeff{ 714107126Sjeff return (sizeof(struct proc)); 715107126Sjeff} 716107126Sjeffint 717107126Sjeffsched_sizeof_thread(void) 718107126Sjeff{ 719107126Sjeff return (sizeof(struct thread)); 720107126Sjeff} 721107137Sjeff 722107137Sjefffixpt_t 723121127Sjeffsched_pctcpu(struct thread *td) 724107137Sjeff{ 725121127Sjeff return (td->td_kse->ke_pctcpu); 726107137Sjeff} 727