sched_4bsd.c revision 145109
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 * 4. Neither the name of the University nor the names of its contributors 19104964Sjeff * may be used to endorse or promote products derived from this software 20104964Sjeff * without specific prior written permission. 21104964Sjeff * 22104964Sjeff * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 23104964Sjeff * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 24104964Sjeff * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 25104964Sjeff * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 26104964Sjeff * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 27104964Sjeff * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 28104964Sjeff * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 29104964Sjeff * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 30104964Sjeff * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 31104964Sjeff * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32104964Sjeff * SUCH DAMAGE. 33104964Sjeff */ 34104964Sjeff 35116182Sobrien#include <sys/cdefs.h> 36116182Sobrien__FBSDID("$FreeBSD: head/sys/kern/sched_4bsd.c 145109 2005-04-15 14:01:43Z maxim $"); 37116182Sobrien 38134791Sjulian#define kse td_sched 39134791Sjulian 40104964Sjeff#include <sys/param.h> 41104964Sjeff#include <sys/systm.h> 42104964Sjeff#include <sys/kernel.h> 43104964Sjeff#include <sys/ktr.h> 44104964Sjeff#include <sys/lock.h> 45123871Sjhb#include <sys/kthread.h> 46104964Sjeff#include <sys/mutex.h> 47104964Sjeff#include <sys/proc.h> 48104964Sjeff#include <sys/resourcevar.h> 49104964Sjeff#include <sys/sched.h> 50104964Sjeff#include <sys/smp.h> 51104964Sjeff#include <sys/sysctl.h> 52104964Sjeff#include <sys/sx.h> 53139453Sjhb#include <sys/turnstile.h> 54134689Sjulian#include <machine/smp.h> 55104964Sjeff 56107135Sjeff/* 57107135Sjeff * INVERSE_ESTCPU_WEIGHT is only suitable for statclock() frequencies in 58107135Sjeff * the range 100-256 Hz (approximately). 59107135Sjeff */ 60107135Sjeff#define ESTCPULIM(e) \ 61107135Sjeff min((e), INVERSE_ESTCPU_WEIGHT * (NICE_WEIGHT * (PRIO_MAX - PRIO_MIN) - \ 62107135Sjeff RQ_PPQ) + INVERSE_ESTCPU_WEIGHT - 1) 63122355Sbde#ifdef SMP 64122355Sbde#define INVERSE_ESTCPU_WEIGHT (8 * smp_cpus) 65122355Sbde#else 66107135Sjeff#define INVERSE_ESTCPU_WEIGHT 8 /* 1 / (priorities per estcpu level). */ 67122355Sbde#endif 68107135Sjeff#define NICE_WEIGHT 1 /* Priorities per nice level. */ 69107135Sjeff 70134791Sjulian/* 71134791Sjulian * The schedulable entity that can be given a context to run. 72134791Sjulian * A process may have several of these. Probably one per processor 73134791Sjulian * but posibly a few more. In this universe they are grouped 74134791Sjulian * with a KSEG that contains the priority and niceness 75134791Sjulian * for the group. 76134791Sjulian */ 77134791Sjulianstruct kse { 78134791Sjulian TAILQ_ENTRY(kse) ke_procq; /* (j/z) Run queue. */ 79134791Sjulian struct thread *ke_thread; /* (*) Active associated thread. */ 80134791Sjulian fixpt_t ke_pctcpu; /* (j) %cpu during p_swtime. */ 81134791Sjulian char ke_rqindex; /* (j) Run queue index. */ 82134791Sjulian enum { 83134791Sjulian KES_THREAD = 0x0, /* slaved to thread state */ 84134791Sjulian KES_ONRUNQ 85134791Sjulian } ke_state; /* (j) KSE status. */ 86134791Sjulian int ke_cpticks; /* (j) Ticks of cpu time. */ 87134791Sjulian struct runq *ke_runq; /* runq the kse is currently on */ 88109145Sjeff}; 89109145Sjeff 90134791Sjulian#define ke_proc ke_thread->td_proc 91134791Sjulian#define ke_ksegrp ke_thread->td_ksegrp 92134791Sjulian 93134791Sjulian#define td_kse td_sched 94134791Sjulian 95134791Sjulian/* flags kept in td_flags */ 96134791Sjulian#define TDF_DIDRUN TDF_SCHED0 /* KSE actually ran. */ 97134791Sjulian#define TDF_EXIT TDF_SCHED1 /* KSE is being killed. */ 98134791Sjulian#define TDF_BOUND TDF_SCHED2 99134791Sjulian 100134791Sjulian#define ke_flags ke_thread->td_flags 101134791Sjulian#define KEF_DIDRUN TDF_DIDRUN /* KSE actually ran. */ 102134791Sjulian#define KEF_EXIT TDF_EXIT /* KSE is being killed. */ 103134791Sjulian#define KEF_BOUND TDF_BOUND /* stuck to one CPU */ 104134791Sjulian 105124955Sjeff#define SKE_RUNQ_PCPU(ke) \ 106124955Sjeff ((ke)->ke_runq != 0 && (ke)->ke_runq != &runq) 107124955Sjeff 108134791Sjulianstruct kg_sched { 109134791Sjulian struct thread *skg_last_assigned; /* (j) Last thread assigned to */ 110134791Sjulian /* the system scheduler. */ 111134791Sjulian int skg_avail_opennings; /* (j) Num KSEs requested in group. */ 112134791Sjulian int skg_concurrency; /* (j) Num KSEs requested in group. */ 113134791Sjulian}; 114134791Sjulian#define kg_last_assigned kg_sched->skg_last_assigned 115134791Sjulian#define kg_avail_opennings kg_sched->skg_avail_opennings 116134791Sjulian#define kg_concurrency kg_sched->skg_concurrency 117134791Sjulian 118136167Sjulian#define SLOT_RELEASE(kg) \ 119136167Sjuliando { \ 120136167Sjulian kg->kg_avail_opennings++; \ 121136167Sjulian CTR3(KTR_RUNQ, "kg %p(%d) Slot released (->%d)", \ 122136167Sjulian kg, \ 123136167Sjulian kg->kg_concurrency, \ 124136167Sjulian kg->kg_avail_opennings); \ 125136167Sjulian/* KASSERT((kg->kg_avail_opennings <= kg->kg_concurrency), \ 126136167Sjulian ("slots out of whack"));*/ \ 127136167Sjulian} while (0) 128136167Sjulian 129136167Sjulian#define SLOT_USE(kg) \ 130136167Sjuliando { \ 131136167Sjulian kg->kg_avail_opennings--; \ 132136167Sjulian CTR3(KTR_RUNQ, "kg %p(%d) Slot used (->%d)", \ 133136167Sjulian kg, \ 134136167Sjulian kg->kg_concurrency, \ 135136167Sjulian kg->kg_avail_opennings); \ 136136167Sjulian/* KASSERT((kg->kg_avail_opennings >= 0), \ 137136167Sjulian ("slots out of whack"));*/ \ 138136167Sjulian} while (0) 139136167Sjulian 140124955Sjeff/* 141124955Sjeff * KSE_CAN_MIGRATE macro returns true if the kse can migrate between 142125295Sjeff * cpus. 143124955Sjeff */ 144124955Sjeff#define KSE_CAN_MIGRATE(ke) \ 145135076Sscottl ((ke)->ke_thread->td_pinned == 0 && ((ke)->ke_flags & KEF_BOUND) == 0) 146109145Sjeff 147134791Sjulianstatic struct kse kse0; 148134791Sjulianstatic struct kg_sched kg_sched0; 149104964Sjeff 150125288Sjeffstatic int sched_tdcnt; /* Total runnable threads in the system. */ 151104964Sjeffstatic int sched_quantum; /* Roundrobin scheduling quantum in ticks. */ 152112535Smux#define SCHED_QUANTUM (hz / 10) /* Default sched quantum */ 153104964Sjeff 154104964Sjeffstatic struct callout roundrobin_callout; 155104964Sjeff 156134791Sjulianstatic void slot_fill(struct ksegrp *kg); 157134791Sjulianstatic struct kse *sched_choose(void); /* XXX Should be thread * */ 158134791Sjulian 159124955Sjeffstatic void setup_runqs(void); 160104964Sjeffstatic void roundrobin(void *arg); 161123871Sjhbstatic void schedcpu(void); 162124955Sjeffstatic void schedcpu_thread(void); 163139453Sjhbstatic void sched_priority(struct thread *td, u_char prio); 164104964Sjeffstatic void sched_setup(void *dummy); 165104964Sjeffstatic void maybe_resched(struct thread *td); 166104964Sjeffstatic void updatepri(struct ksegrp *kg); 167104964Sjeffstatic void resetpriority(struct ksegrp *kg); 168139453Sjhbstatic void resetpriority_thread(struct thread *td, struct ksegrp *kg); 169134694Sjulian#ifdef SMP 170134688Sjulianstatic int forward_wakeup(int cpunum); 171134694Sjulian#endif 172104964Sjeff 173124955Sjeffstatic struct kproc_desc sched_kp = { 174124955Sjeff "schedcpu", 175124955Sjeff schedcpu_thread, 176124955Sjeff NULL 177124955Sjeff}; 178124955SjeffSYSINIT(schedcpu, SI_SUB_RUN_SCHEDULER, SI_ORDER_FIRST, kproc_start, &sched_kp) 179124955SjeffSYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL) 180104964Sjeff 181104964Sjeff/* 182104964Sjeff * Global run queue. 183104964Sjeff */ 184104964Sjeffstatic struct runq runq; 185104964Sjeff 186124955Sjeff#ifdef SMP 187124955Sjeff/* 188124955Sjeff * Per-CPU run queues 189124955Sjeff */ 190124955Sjeffstatic struct runq runq_pcpu[MAXCPU]; 191124955Sjeff#endif 192124955Sjeff 193124955Sjeffstatic void 194124955Sjeffsetup_runqs(void) 195124955Sjeff{ 196124955Sjeff#ifdef SMP 197124955Sjeff int i; 198124955Sjeff 199124955Sjeff for (i = 0; i < MAXCPU; ++i) 200124955Sjeff runq_init(&runq_pcpu[i]); 201124955Sjeff#endif 202124955Sjeff 203124955Sjeff runq_init(&runq); 204124955Sjeff} 205124955Sjeff 206104964Sjeffstatic int 207104964Sjeffsysctl_kern_quantum(SYSCTL_HANDLER_ARGS) 208104964Sjeff{ 209104964Sjeff int error, new_val; 210104964Sjeff 211104964Sjeff new_val = sched_quantum * tick; 212104964Sjeff error = sysctl_handle_int(oidp, &new_val, 0, req); 213104964Sjeff if (error != 0 || req->newptr == NULL) 214104964Sjeff return (error); 215104964Sjeff if (new_val < tick) 216104964Sjeff return (EINVAL); 217104964Sjeff sched_quantum = new_val / tick; 218104964Sjeff hogticks = 2 * sched_quantum; 219104964Sjeff return (0); 220104964Sjeff} 221104964Sjeff 222132589SscottlSYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RD, 0, "Scheduler"); 223130881Sscottl 224132589SscottlSYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "4BSD", 0, 225132589Sscottl "Scheduler name"); 226130881Sscottl 227132589SscottlSYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW, 228132589Sscottl 0, sizeof sched_quantum, sysctl_kern_quantum, "I", 229132589Sscottl "Roundrobin scheduling quantum in microseconds"); 230104964Sjeff 231134693Sjulian#ifdef SMP 232134688Sjulian/* Enable forwarding of wakeups to all other cpus */ 233134688SjulianSYSCTL_NODE(_kern_sched, OID_AUTO, ipiwakeup, CTLFLAG_RD, NULL, "Kernel SMP"); 234134688Sjulian 235134792Sjulianstatic int forward_wakeup_enabled = 1; 236134688SjulianSYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, enabled, CTLFLAG_RW, 237134688Sjulian &forward_wakeup_enabled, 0, 238134688Sjulian "Forwarding of wakeup to idle CPUs"); 239134688Sjulian 240134688Sjulianstatic int forward_wakeups_requested = 0; 241134688SjulianSYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, requested, CTLFLAG_RD, 242134688Sjulian &forward_wakeups_requested, 0, 243134688Sjulian "Requests for Forwarding of wakeup to idle CPUs"); 244134688Sjulian 245134688Sjulianstatic int forward_wakeups_delivered = 0; 246134688SjulianSYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, delivered, CTLFLAG_RD, 247134688Sjulian &forward_wakeups_delivered, 0, 248134688Sjulian "Completed Forwarding of wakeup to idle CPUs"); 249134688Sjulian 250134792Sjulianstatic int forward_wakeup_use_mask = 1; 251134688SjulianSYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, usemask, CTLFLAG_RW, 252134688Sjulian &forward_wakeup_use_mask, 0, 253134688Sjulian "Use the mask of idle cpus"); 254134688Sjulian 255134688Sjulianstatic int forward_wakeup_use_loop = 0; 256134688SjulianSYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, useloop, CTLFLAG_RW, 257134688Sjulian &forward_wakeup_use_loop, 0, 258134688Sjulian "Use a loop to find idle cpus"); 259134688Sjulian 260134688Sjulianstatic int forward_wakeup_use_single = 0; 261134688SjulianSYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, onecpu, CTLFLAG_RW, 262134688Sjulian &forward_wakeup_use_single, 0, 263134688Sjulian "Only signal one idle cpu"); 264134688Sjulian 265134688Sjulianstatic int forward_wakeup_use_htt = 0; 266134688SjulianSYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, htt2, CTLFLAG_RW, 267134688Sjulian &forward_wakeup_use_htt, 0, 268134688Sjulian "account for htt"); 269135051Sjulian 270134693Sjulian#endif 271135051Sjulianstatic int sched_followon = 0; 272135051SjulianSYSCTL_INT(_kern_sched, OID_AUTO, followon, CTLFLAG_RW, 273135051Sjulian &sched_followon, 0, 274135051Sjulian "allow threads to share a quantum"); 275134688Sjulian 276135051Sjulianstatic int sched_pfollowons = 0; 277135051SjulianSYSCTL_INT(_kern_sched, OID_AUTO, pfollowons, CTLFLAG_RD, 278135051Sjulian &sched_pfollowons, 0, 279135051Sjulian "number of followons done to a different ksegrp"); 280135051Sjulian 281135051Sjulianstatic int sched_kgfollowons = 0; 282135051SjulianSYSCTL_INT(_kern_sched, OID_AUTO, kgfollowons, CTLFLAG_RD, 283135051Sjulian &sched_kgfollowons, 0, 284135051Sjulian "number of followons done in a ksegrp"); 285135051Sjulian 286139317Sjeffstatic __inline void 287139317Sjeffsched_load_add(void) 288139317Sjeff{ 289139317Sjeff sched_tdcnt++; 290139317Sjeff CTR1(KTR_SCHED, "global load: %d", sched_tdcnt); 291139317Sjeff} 292139317Sjeff 293139317Sjeffstatic __inline void 294139317Sjeffsched_load_rem(void) 295139317Sjeff{ 296139317Sjeff sched_tdcnt--; 297139317Sjeff CTR1(KTR_SCHED, "global load: %d", sched_tdcnt); 298139317Sjeff} 299104964Sjeff/* 300104964Sjeff * Arrange to reschedule if necessary, taking the priorities and 301104964Sjeff * schedulers into account. 302104964Sjeff */ 303104964Sjeffstatic void 304104964Sjeffmaybe_resched(struct thread *td) 305104964Sjeff{ 306104964Sjeff 307104964Sjeff mtx_assert(&sched_lock, MA_OWNED); 308134791Sjulian if (td->td_priority < curthread->td_priority) 309111032Sjulian curthread->td_flags |= TDF_NEEDRESCHED; 310104964Sjeff} 311104964Sjeff 312104964Sjeff/* 313104964Sjeff * Force switch among equal priority processes every 100ms. 314104964Sjeff * We don't actually need to force a context switch of the current process. 315104964Sjeff * The act of firing the event triggers a context switch to softclock() and 316104964Sjeff * then switching back out again which is equivalent to a preemption, thus 317104964Sjeff * no further work is needed on the local CPU. 318104964Sjeff */ 319104964Sjeff/* ARGSUSED */ 320104964Sjeffstatic void 321104964Sjeffroundrobin(void *arg) 322104964Sjeff{ 323104964Sjeff 324104964Sjeff#ifdef SMP 325104964Sjeff mtx_lock_spin(&sched_lock); 326104964Sjeff forward_roundrobin(); 327104964Sjeff mtx_unlock_spin(&sched_lock); 328104964Sjeff#endif 329104964Sjeff 330104964Sjeff callout_reset(&roundrobin_callout, sched_quantum, roundrobin, NULL); 331104964Sjeff} 332104964Sjeff 333104964Sjeff/* 334104964Sjeff * Constants for digital decay and forget: 335118972Sjhb * 90% of (kg_estcpu) usage in 5 * loadav time 336118972Sjhb * 95% of (ke_pctcpu) usage in 60 seconds (load insensitive) 337104964Sjeff * Note that, as ps(1) mentions, this can let percentages 338104964Sjeff * total over 100% (I've seen 137.9% for 3 processes). 339104964Sjeff * 340118972Sjhb * Note that schedclock() updates kg_estcpu and p_cpticks asynchronously. 341104964Sjeff * 342118972Sjhb * We wish to decay away 90% of kg_estcpu in (5 * loadavg) seconds. 343104964Sjeff * That is, the system wants to compute a value of decay such 344104964Sjeff * that the following for loop: 345104964Sjeff * for (i = 0; i < (5 * loadavg); i++) 346118972Sjhb * kg_estcpu *= decay; 347104964Sjeff * will compute 348118972Sjhb * kg_estcpu *= 0.1; 349104964Sjeff * for all values of loadavg: 350104964Sjeff * 351104964Sjeff * Mathematically this loop can be expressed by saying: 352104964Sjeff * decay ** (5 * loadavg) ~= .1 353104964Sjeff * 354104964Sjeff * The system computes decay as: 355104964Sjeff * decay = (2 * loadavg) / (2 * loadavg + 1) 356104964Sjeff * 357104964Sjeff * We wish to prove that the system's computation of decay 358104964Sjeff * will always fulfill the equation: 359104964Sjeff * decay ** (5 * loadavg) ~= .1 360104964Sjeff * 361104964Sjeff * If we compute b as: 362104964Sjeff * b = 2 * loadavg 363104964Sjeff * then 364104964Sjeff * decay = b / (b + 1) 365104964Sjeff * 366104964Sjeff * We now need to prove two things: 367104964Sjeff * 1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1) 368104964Sjeff * 2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg) 369104964Sjeff * 370104964Sjeff * Facts: 371104964Sjeff * For x close to zero, exp(x) =~ 1 + x, since 372104964Sjeff * exp(x) = 0! + x**1/1! + x**2/2! + ... . 373104964Sjeff * therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b. 374104964Sjeff * For x close to zero, ln(1+x) =~ x, since 375104964Sjeff * ln(1+x) = x - x**2/2 + x**3/3 - ... -1 < x < 1 376104964Sjeff * therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1). 377104964Sjeff * ln(.1) =~ -2.30 378104964Sjeff * 379104964Sjeff * Proof of (1): 380104964Sjeff * Solve (factor)**(power) =~ .1 given power (5*loadav): 381104964Sjeff * solving for factor, 382104964Sjeff * ln(factor) =~ (-2.30/5*loadav), or 383104964Sjeff * factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) = 384104964Sjeff * exp(-1/b) =~ (b-1)/b =~ b/(b+1). QED 385104964Sjeff * 386104964Sjeff * Proof of (2): 387104964Sjeff * Solve (factor)**(power) =~ .1 given factor == (b/(b+1)): 388104964Sjeff * solving for power, 389104964Sjeff * power*ln(b/(b+1)) =~ -2.30, or 390104964Sjeff * power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav. QED 391104964Sjeff * 392104964Sjeff * Actual power values for the implemented algorithm are as follows: 393104964Sjeff * loadav: 1 2 3 4 394104964Sjeff * power: 5.68 10.32 14.94 19.55 395104964Sjeff */ 396104964Sjeff 397104964Sjeff/* calculations for digital decay to forget 90% of usage in 5*loadav sec */ 398104964Sjeff#define loadfactor(loadav) (2 * (loadav)) 399104964Sjeff#define decay_cpu(loadfac, cpu) (((loadfac) * (cpu)) / ((loadfac) + FSCALE)) 400104964Sjeff 401118972Sjhb/* decay 95% of `ke_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */ 402104964Sjeffstatic fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */ 403104964SjeffSYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 404104964Sjeff 405104964Sjeff/* 406104964Sjeff * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the 407104964Sjeff * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below 408104964Sjeff * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT). 409104964Sjeff * 410104964Sjeff * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used: 411104964Sjeff * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits). 412104964Sjeff * 413104964Sjeff * If you don't want to bother with the faster/more-accurate formula, you 414104964Sjeff * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate 415104964Sjeff * (more general) method of calculating the %age of CPU used by a process. 416104964Sjeff */ 417104964Sjeff#define CCPU_SHIFT 11 418104964Sjeff 419104964Sjeff/* 420104964Sjeff * Recompute process priorities, every hz ticks. 421104964Sjeff * MP-safe, called without the Giant mutex. 422104964Sjeff */ 423104964Sjeff/* ARGSUSED */ 424104964Sjeffstatic void 425123871Sjhbschedcpu(void) 426104964Sjeff{ 427104964Sjeff register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 428104964Sjeff struct thread *td; 429104964Sjeff struct proc *p; 430104964Sjeff struct kse *ke; 431104964Sjeff struct ksegrp *kg; 432118972Sjhb int awake, realstathz; 433104964Sjeff 434104964Sjeff realstathz = stathz ? stathz : hz; 435104964Sjeff sx_slock(&allproc_lock); 436104964Sjeff FOREACH_PROC_IN_SYSTEM(p) { 437118972Sjhb /* 438118972Sjhb * Prevent state changes and protect run queue. 439118972Sjhb */ 440104964Sjeff mtx_lock_spin(&sched_lock); 441118972Sjhb /* 442118972Sjhb * Increment time in/out of memory. We ignore overflow; with 443118972Sjhb * 16-bit int's (remember them?) overflow takes 45 days. 444118972Sjhb */ 445104964Sjeff p->p_swtime++; 446104964Sjeff FOREACH_KSEGRP_IN_PROC(p, kg) { 447104964Sjeff awake = 0; 448134791Sjulian FOREACH_THREAD_IN_GROUP(kg, td) { 449134791Sjulian ke = td->td_kse; 450104964Sjeff /* 451118972Sjhb * Increment sleep time (if sleeping). We 452118972Sjhb * ignore overflow, as above. 453104964Sjeff */ 454104964Sjeff /* 455104964Sjeff * The kse slptimes are not touched in wakeup 456104964Sjeff * because the thread may not HAVE a KSE. 457104964Sjeff */ 458104964Sjeff if (ke->ke_state == KES_ONRUNQ) { 459104964Sjeff awake = 1; 460104964Sjeff ke->ke_flags &= ~KEF_DIDRUN; 461104964Sjeff } else if ((ke->ke_state == KES_THREAD) && 462134791Sjulian (TD_IS_RUNNING(td))) { 463104964Sjeff awake = 1; 464104964Sjeff /* Do not clear KEF_DIDRUN */ 465104964Sjeff } else if (ke->ke_flags & KEF_DIDRUN) { 466104964Sjeff awake = 1; 467104964Sjeff ke->ke_flags &= ~KEF_DIDRUN; 468104964Sjeff } 469104964Sjeff 470104964Sjeff /* 471118972Sjhb * ke_pctcpu is only for ps and ttyinfo(). 472118972Sjhb * Do it per kse, and add them up at the end? 473104964Sjeff * XXXKSE 474104964Sjeff */ 475118972Sjhb ke->ke_pctcpu = (ke->ke_pctcpu * ccpu) >> 476109145Sjeff FSHIFT; 477104964Sjeff /* 478104964Sjeff * If the kse has been idle the entire second, 479104964Sjeff * stop recalculating its priority until 480104964Sjeff * it wakes up. 481104964Sjeff */ 482134145Sjulian if (ke->ke_cpticks == 0) 483104964Sjeff continue; 484104964Sjeff#if (FSHIFT >= CCPU_SHIFT) 485109157Sjeff ke->ke_pctcpu += (realstathz == 100) 486134145Sjulian ? ((fixpt_t) ke->ke_cpticks) << 487104964Sjeff (FSHIFT - CCPU_SHIFT) : 488134145Sjulian 100 * (((fixpt_t) ke->ke_cpticks) 489109145Sjeff << (FSHIFT - CCPU_SHIFT)) / realstathz; 490104964Sjeff#else 491109157Sjeff ke->ke_pctcpu += ((FSCALE - ccpu) * 492134145Sjulian (ke->ke_cpticks * 493109145Sjeff FSCALE / realstathz)) >> FSHIFT; 494104964Sjeff#endif 495134145Sjulian ke->ke_cpticks = 0; 496104964Sjeff } /* end of kse loop */ 497104964Sjeff /* 498104964Sjeff * If there are ANY running threads in this KSEGRP, 499104964Sjeff * then don't count it as sleeping. 500104964Sjeff */ 501104964Sjeff if (awake) { 502104964Sjeff if (kg->kg_slptime > 1) { 503104964Sjeff /* 504104964Sjeff * In an ideal world, this should not 505104964Sjeff * happen, because whoever woke us 506104964Sjeff * up from the long sleep should have 507104964Sjeff * unwound the slptime and reset our 508104964Sjeff * priority before we run at the stale 509104964Sjeff * priority. Should KASSERT at some 510104964Sjeff * point when all the cases are fixed. 511104964Sjeff */ 512104964Sjeff updatepri(kg); 513104964Sjeff } 514104964Sjeff kg->kg_slptime = 0; 515118972Sjhb } else 516104964Sjeff kg->kg_slptime++; 517104964Sjeff if (kg->kg_slptime > 1) 518104964Sjeff continue; 519104964Sjeff kg->kg_estcpu = decay_cpu(loadfac, kg->kg_estcpu); 520104964Sjeff resetpriority(kg); 521104964Sjeff FOREACH_THREAD_IN_GROUP(kg, td) { 522139453Sjhb resetpriority_thread(td, kg); 523104964Sjeff } 524104964Sjeff } /* end of ksegrp loop */ 525104964Sjeff mtx_unlock_spin(&sched_lock); 526104964Sjeff } /* end of process loop */ 527104964Sjeff sx_sunlock(&allproc_lock); 528104964Sjeff} 529104964Sjeff 530104964Sjeff/* 531123871Sjhb * Main loop for a kthread that executes schedcpu once a second. 532123871Sjhb */ 533123871Sjhbstatic void 534124955Sjeffschedcpu_thread(void) 535123871Sjhb{ 536123871Sjhb int nowake; 537123871Sjhb 538123871Sjhb for (;;) { 539123871Sjhb schedcpu(); 540123871Sjhb tsleep(&nowake, curthread->td_priority, "-", hz); 541123871Sjhb } 542123871Sjhb} 543123871Sjhb 544123871Sjhb/* 545104964Sjeff * Recalculate the priority of a process after it has slept for a while. 546118972Sjhb * For all load averages >= 1 and max kg_estcpu of 255, sleeping for at 547118972Sjhb * least six times the loadfactor will decay kg_estcpu to zero. 548104964Sjeff */ 549104964Sjeffstatic void 550104964Sjeffupdatepri(struct ksegrp *kg) 551104964Sjeff{ 552118972Sjhb register fixpt_t loadfac; 553104964Sjeff register unsigned int newcpu; 554104964Sjeff 555118972Sjhb loadfac = loadfactor(averunnable.ldavg[0]); 556104964Sjeff if (kg->kg_slptime > 5 * loadfac) 557104964Sjeff kg->kg_estcpu = 0; 558104964Sjeff else { 559118972Sjhb newcpu = kg->kg_estcpu; 560118972Sjhb kg->kg_slptime--; /* was incremented in schedcpu() */ 561104964Sjeff while (newcpu && --kg->kg_slptime) 562104964Sjeff newcpu = decay_cpu(loadfac, newcpu); 563104964Sjeff kg->kg_estcpu = newcpu; 564104964Sjeff } 565104964Sjeff} 566104964Sjeff 567104964Sjeff/* 568104964Sjeff * Compute the priority of a process when running in user mode. 569104964Sjeff * Arrange to reschedule if the resulting priority is better 570104964Sjeff * than that of the current process. 571104964Sjeff */ 572104964Sjeffstatic void 573104964Sjeffresetpriority(struct ksegrp *kg) 574104964Sjeff{ 575104964Sjeff register unsigned int newpriority; 576104964Sjeff 577104964Sjeff if (kg->kg_pri_class == PRI_TIMESHARE) { 578104964Sjeff newpriority = PUSER + kg->kg_estcpu / INVERSE_ESTCPU_WEIGHT + 579130551Sjulian NICE_WEIGHT * (kg->kg_proc->p_nice - PRIO_MIN); 580104964Sjeff newpriority = min(max(newpriority, PRI_MIN_TIMESHARE), 581104964Sjeff PRI_MAX_TIMESHARE); 582104964Sjeff kg->kg_user_pri = newpriority; 583104964Sjeff } 584104964Sjeff} 585104964Sjeff 586139453Sjhb/* 587139453Sjhb * Update the thread's priority when the associated ksegroup's user 588139453Sjhb * priority changes. 589139453Sjhb */ 590139453Sjhbstatic void 591139453Sjhbresetpriority_thread(struct thread *td, struct ksegrp *kg) 592139453Sjhb{ 593139453Sjhb 594139453Sjhb /* Only change threads with a time sharing user priority. */ 595139453Sjhb if (td->td_priority < PRI_MIN_TIMESHARE || 596139453Sjhb td->td_priority > PRI_MAX_TIMESHARE) 597139453Sjhb return; 598139453Sjhb 599139453Sjhb /* XXX the whole needresched thing is broken, but not silly. */ 600139453Sjhb maybe_resched(td); 601139453Sjhb 602139453Sjhb sched_prio(td, kg->kg_user_pri); 603139453Sjhb} 604139453Sjhb 605104964Sjeff/* ARGSUSED */ 606104964Sjeffstatic void 607104964Sjeffsched_setup(void *dummy) 608104964Sjeff{ 609124955Sjeff setup_runqs(); 610118972Sjhb 611104964Sjeff if (sched_quantum == 0) 612104964Sjeff sched_quantum = SCHED_QUANTUM; 613104964Sjeff hogticks = 2 * sched_quantum; 614104964Sjeff 615126665Srwatson callout_init(&roundrobin_callout, CALLOUT_MPSAFE); 616104964Sjeff 617104964Sjeff /* Kick off timeout driven events by calling first time. */ 618104964Sjeff roundrobin(NULL); 619125288Sjeff 620125288Sjeff /* Account for thread0. */ 621139317Sjeff sched_load_add(); 622104964Sjeff} 623104964Sjeff 624104964Sjeff/* External interfaces start here */ 625134791Sjulian/* 626134791Sjulian * Very early in the boot some setup of scheduler-specific 627145109Smaxim * parts of proc0 and of some scheduler resources needs to be done. 628134791Sjulian * Called from: 629134791Sjulian * proc0_init() 630134791Sjulian */ 631134791Sjulianvoid 632134791Sjulianschedinit(void) 633134791Sjulian{ 634134791Sjulian /* 635134791Sjulian * Set up the scheduler specific parts of proc0. 636134791Sjulian */ 637134791Sjulian proc0.p_sched = NULL; /* XXX */ 638134791Sjulian ksegrp0.kg_sched = &kg_sched0; 639134791Sjulian thread0.td_sched = &kse0; 640134791Sjulian kse0.ke_thread = &thread0; 641134791Sjulian kse0.ke_state = KES_THREAD; 642134791Sjulian kg_sched0.skg_concurrency = 1; 643134791Sjulian kg_sched0.skg_avail_opennings = 0; /* we are already running */ 644134791Sjulian} 645134791Sjulian 646104964Sjeffint 647104964Sjeffsched_runnable(void) 648104964Sjeff{ 649124955Sjeff#ifdef SMP 650124955Sjeff return runq_check(&runq) + runq_check(&runq_pcpu[PCPU_GET(cpuid)]); 651124955Sjeff#else 652124955Sjeff return runq_check(&runq); 653124955Sjeff#endif 654104964Sjeff} 655104964Sjeff 656104964Sjeffint 657104964Sjeffsched_rr_interval(void) 658104964Sjeff{ 659104964Sjeff if (sched_quantum == 0) 660104964Sjeff sched_quantum = SCHED_QUANTUM; 661104964Sjeff return (sched_quantum); 662104964Sjeff} 663104964Sjeff 664104964Sjeff/* 665104964Sjeff * We adjust the priority of the current process. The priority of 666104964Sjeff * a process gets worse as it accumulates CPU time. The cpu usage 667118972Sjhb * estimator (kg_estcpu) is increased here. resetpriority() will 668118972Sjhb * compute a different priority each time kg_estcpu increases by 669104964Sjeff * INVERSE_ESTCPU_WEIGHT 670104964Sjeff * (until MAXPRI is reached). The cpu usage estimator ramps up 671104964Sjeff * quite quickly when the process is running (linearly), and decays 672104964Sjeff * away exponentially, at a rate which is proportionally slower when 673104964Sjeff * the system is busy. The basic principle is that the system will 674104964Sjeff * 90% forget that the process used a lot of CPU time in 5 * loadav 675104964Sjeff * seconds. This causes the system to favor processes which haven't 676104964Sjeff * run much recently, and to round-robin among other processes. 677104964Sjeff */ 678104964Sjeffvoid 679121127Sjeffsched_clock(struct thread *td) 680104964Sjeff{ 681104964Sjeff struct ksegrp *kg; 682121127Sjeff struct kse *ke; 683104964Sjeff 684113923Sjhb mtx_assert(&sched_lock, MA_OWNED); 685121127Sjeff kg = td->td_ksegrp; 686121127Sjeff ke = td->td_kse; 687113356Sjeff 688134145Sjulian ke->ke_cpticks++; 689104964Sjeff kg->kg_estcpu = ESTCPULIM(kg->kg_estcpu + 1); 690104964Sjeff if ((kg->kg_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) { 691104964Sjeff resetpriority(kg); 692139453Sjhb resetpriority_thread(td, kg); 693104964Sjeff } 694104964Sjeff} 695118972Sjhb 696104964Sjeff/* 697104964Sjeff * charge childs scheduling cpu usage to parent. 698104964Sjeff * 699104964Sjeff * XXXKSE assume only one thread & kse & ksegrp keep estcpu in each ksegrp. 700104964Sjeff * Charge it to the ksegrp that did the wait since process estcpu is sum of 701104964Sjeff * all ksegrps, this is strictly as expected. Assume that the child process 702104964Sjeff * aggregated all the estcpu into the 'built-in' ksegrp. 703104964Sjeff */ 704104964Sjeffvoid 705132372Sjuliansched_exit(struct proc *p, struct thread *td) 706104964Sjeff{ 707132372Sjulian sched_exit_ksegrp(FIRST_KSEGRP_IN_PROC(p), td); 708132372Sjulian sched_exit_thread(FIRST_THREAD_IN_PROC(p), td); 709113356Sjeff} 710113356Sjeff 711113356Sjeffvoid 712132372Sjuliansched_exit_ksegrp(struct ksegrp *kg, struct thread *childtd) 713113356Sjeff{ 714113923Sjhb 715113923Sjhb mtx_assert(&sched_lock, MA_OWNED); 716132372Sjulian kg->kg_estcpu = ESTCPULIM(kg->kg_estcpu + childtd->td_ksegrp->kg_estcpu); 717104964Sjeff} 718104964Sjeff 719104964Sjeffvoid 720113356Sjeffsched_exit_thread(struct thread *td, struct thread *child) 721104964Sjeff{ 722139317Sjeff CTR3(KTR_SCHED, "sched_exit_thread: %p(%s) prio %d", 723139317Sjeff child, child->td_proc->p_comm, child->td_priority); 724127894Sdfr if ((child->td_proc->p_flag & P_NOLOAD) == 0) 725139317Sjeff sched_load_rem(); 726113356Sjeff} 727109145Sjeff 728113356Sjeffvoid 729134791Sjuliansched_fork(struct thread *td, struct thread *childtd) 730113356Sjeff{ 731134791Sjulian sched_fork_ksegrp(td, childtd->td_ksegrp); 732134791Sjulian sched_fork_thread(td, childtd); 733113356Sjeff} 734113356Sjeff 735113356Sjeffvoid 736132372Sjuliansched_fork_ksegrp(struct thread *td, struct ksegrp *child) 737113356Sjeff{ 738113923Sjhb mtx_assert(&sched_lock, MA_OWNED); 739132372Sjulian child->kg_estcpu = td->td_ksegrp->kg_estcpu; 740113356Sjeff} 741109145Sjeff 742113356Sjeffvoid 743134791Sjuliansched_fork_thread(struct thread *td, struct thread *childtd) 744113356Sjeff{ 745134791Sjulian sched_newthread(childtd); 746104964Sjeff} 747104964Sjeff 748104964Sjeffvoid 749130551Sjuliansched_nice(struct proc *p, int nice) 750104964Sjeff{ 751130551Sjulian struct ksegrp *kg; 752139453Sjhb struct thread *td; 753113873Sjhb 754130551Sjulian PROC_LOCK_ASSERT(p, MA_OWNED); 755113873Sjhb mtx_assert(&sched_lock, MA_OWNED); 756130551Sjulian p->p_nice = nice; 757130551Sjulian FOREACH_KSEGRP_IN_PROC(p, kg) { 758130551Sjulian resetpriority(kg); 759139453Sjhb FOREACH_THREAD_IN_GROUP(kg, td) { 760139453Sjhb resetpriority_thread(td, kg); 761139453Sjhb } 762130551Sjulian } 763104964Sjeff} 764104964Sjeff 765113356Sjeffvoid 766113356Sjeffsched_class(struct ksegrp *kg, int class) 767113356Sjeff{ 768113923Sjhb mtx_assert(&sched_lock, MA_OWNED); 769113356Sjeff kg->kg_pri_class = class; 770113356Sjeff} 771113356Sjeff 772105127Sjulian/* 773105127Sjulian * Adjust the priority of a thread. 774105127Sjulian * This may include moving the thread within the KSEGRP, 775105127Sjulian * changing the assignment of a kse to the thread, 776105127Sjulian * and moving a KSE in the system run queue. 777105127Sjulian */ 778139453Sjhbstatic void 779139453Sjhbsched_priority(struct thread *td, u_char prio) 780104964Sjeff{ 781139317Sjeff CTR6(KTR_SCHED, "sched_prio: %p(%s) prio %d newprio %d by %p(%s)", 782139317Sjeff td, td->td_proc->p_comm, td->td_priority, prio, curthread, 783139317Sjeff curthread->td_proc->p_comm); 784104964Sjeff 785113923Sjhb mtx_assert(&sched_lock, MA_OWNED); 786139453Sjhb if (td->td_priority == prio) 787139453Sjhb return; 788104964Sjeff if (TD_ON_RUNQ(td)) { 789105127Sjulian adjustrunqueue(td, prio); 790105127Sjulian } else { 791105127Sjulian td->td_priority = prio; 792104964Sjeff } 793104964Sjeff} 794104964Sjeff 795139453Sjhb/* 796139453Sjhb * Update a thread's priority when it is lent another thread's 797139453Sjhb * priority. 798139453Sjhb */ 799104964Sjeffvoid 800139453Sjhbsched_lend_prio(struct thread *td, u_char prio) 801139453Sjhb{ 802139453Sjhb 803139453Sjhb td->td_flags |= TDF_BORROWING; 804139453Sjhb sched_priority(td, prio); 805139453Sjhb} 806139453Sjhb 807139453Sjhb/* 808139453Sjhb * Restore a thread's priority when priority propagation is 809139453Sjhb * over. The prio argument is the minimum priority the thread 810139453Sjhb * needs to have to satisfy other possible priority lending 811139453Sjhb * requests. If the thread's regulary priority is less 812139453Sjhb * important than prio the thread will keep a priority boost 813139453Sjhb * of prio. 814139453Sjhb */ 815139453Sjhbvoid 816139453Sjhbsched_unlend_prio(struct thread *td, u_char prio) 817139453Sjhb{ 818139453Sjhb u_char base_pri; 819139453Sjhb 820139453Sjhb if (td->td_base_pri >= PRI_MIN_TIMESHARE && 821139453Sjhb td->td_base_pri <= PRI_MAX_TIMESHARE) 822139453Sjhb base_pri = td->td_ksegrp->kg_user_pri; 823139453Sjhb else 824139453Sjhb base_pri = td->td_base_pri; 825139453Sjhb if (prio >= base_pri) { 826139453Sjhb td->td_flags &= ~TDF_BORROWING; 827139453Sjhb sched_prio(td, base_pri); 828139453Sjhb } else 829139453Sjhb sched_lend_prio(td, prio); 830139453Sjhb} 831139453Sjhb 832139453Sjhbvoid 833139453Sjhbsched_prio(struct thread *td, u_char prio) 834139453Sjhb{ 835139453Sjhb u_char oldprio; 836139453Sjhb 837139453Sjhb /* First, update the base priority. */ 838139453Sjhb td->td_base_pri = prio; 839139453Sjhb 840139453Sjhb /* 841139453Sjhb * If the thread is borrowing another thread's priority, don't ever 842139453Sjhb * lower the priority. 843139453Sjhb */ 844139453Sjhb if (td->td_flags & TDF_BORROWING && td->td_priority < prio) 845139453Sjhb return; 846139453Sjhb 847139453Sjhb /* Change the real priority. */ 848139453Sjhb oldprio = td->td_priority; 849139453Sjhb sched_priority(td, prio); 850139453Sjhb 851139453Sjhb /* 852139453Sjhb * If the thread is on a turnstile, then let the turnstile update 853139453Sjhb * its state. 854139453Sjhb */ 855139453Sjhb if (TD_ON_LOCK(td) && oldprio != prio) 856139453Sjhb turnstile_adjust(td, oldprio); 857139453Sjhb} 858139453Sjhb 859139453Sjhbvoid 860126326Sjhbsched_sleep(struct thread *td) 861104964Sjeff{ 862113923Sjhb 863113923Sjhb mtx_assert(&sched_lock, MA_OWNED); 864104964Sjeff td->td_ksegrp->kg_slptime = 0; 865104964Sjeff} 866104964Sjeff 867135051Sjulianstatic void remrunqueue(struct thread *td); 868135051Sjulian 869104964Sjeffvoid 870135051Sjuliansched_switch(struct thread *td, struct thread *newtd, int flags) 871104964Sjeff{ 872104964Sjeff struct kse *ke; 873135051Sjulian struct ksegrp *kg; 874104964Sjeff struct proc *p; 875104964Sjeff 876104964Sjeff ke = td->td_kse; 877104964Sjeff p = td->td_proc; 878104964Sjeff 879113923Sjhb mtx_assert(&sched_lock, MA_OWNED); 880104964Sjeff 881125295Sjeff if ((p->p_flag & P_NOLOAD) == 0) 882139317Sjeff sched_load_rem(); 883134791Sjulian /* 884135051Sjulian * We are volunteering to switch out so we get to nominate 885135051Sjulian * a successor for the rest of our quantum 886135051Sjulian * First try another thread in our ksegrp, and then look for 887135051Sjulian * other ksegrps in our process. 888135051Sjulian */ 889135051Sjulian if (sched_followon && 890135051Sjulian (p->p_flag & P_HADTHREADS) && 891135051Sjulian (flags & SW_VOL) && 892135051Sjulian newtd == NULL) { 893135051Sjulian /* lets schedule another thread from this process */ 894135051Sjulian kg = td->td_ksegrp; 895135051Sjulian if ((newtd = TAILQ_FIRST(&kg->kg_runq))) { 896135051Sjulian remrunqueue(newtd); 897135051Sjulian sched_kgfollowons++; 898135051Sjulian } else { 899135051Sjulian FOREACH_KSEGRP_IN_PROC(p, kg) { 900135051Sjulian if ((newtd = TAILQ_FIRST(&kg->kg_runq))) { 901135051Sjulian sched_pfollowons++; 902135051Sjulian remrunqueue(newtd); 903135051Sjulian break; 904135051Sjulian } 905135051Sjulian } 906135051Sjulian } 907135051Sjulian } 908135051Sjulian 909138527Sups if (newtd) 910138527Sups newtd->td_flags |= (td->td_flags & TDF_NEEDRESCHED); 911138527Sups 912113339Sjulian td->td_lastcpu = td->td_oncpu; 913132266Sjhb td->td_flags &= ~TDF_NEEDRESCHED; 914144777Sups td->td_owepreempt = 0; 915113339Sjulian td->td_oncpu = NOCPU; 916104964Sjeff /* 917104964Sjeff * At the last moment, if this thread is still marked RUNNING, 918104964Sjeff * then put it back on the run queue as it has not been suspended 919131473Sjhb * or stopped or any thing else similar. We never put the idle 920131473Sjhb * threads on the run queue, however. 921104964Sjeff */ 922131473Sjhb if (td == PCPU_GET(idlethread)) 923131473Sjhb TD_SET_CAN_RUN(td); 924134791Sjulian else { 925136170Sjulian SLOT_RELEASE(td->td_ksegrp); 926134791Sjulian if (TD_IS_RUNNING(td)) { 927134791Sjulian /* Put us back on the run queue (kse and all). */ 928136170Sjulian setrunqueue(td, (flags & SW_PREEMPT) ? 929136170Sjulian SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED : 930136170Sjulian SRQ_OURSELF|SRQ_YIELDING); 931134791Sjulian } else if (p->p_flag & P_HADTHREADS) { 932134791Sjulian /* 933134791Sjulian * We will not be on the run queue. So we must be 934134791Sjulian * sleeping or similar. As it's available, 935134791Sjulian * someone else can use the KSE if they need it. 936136170Sjulian * It's NOT available if we are about to need it 937134791Sjulian */ 938136170Sjulian if (newtd == NULL || newtd->td_ksegrp != td->td_ksegrp) 939136170Sjulian slot_fill(td->td_ksegrp); 940134791Sjulian } 941104964Sjeff } 942136170Sjulian if (newtd) { 943136170Sjulian /* 944136170Sjulian * The thread we are about to run needs to be counted 945136170Sjulian * as if it had been added to the run queue and selected. 946136170Sjulian * It came from: 947136170Sjulian * * A preemption 948136170Sjulian * * An upcall 949136170Sjulian * * A followon 950136170Sjulian */ 951136170Sjulian KASSERT((newtd->td_inhibitors == 0), 952136170Sjulian ("trying to run inhibitted thread")); 953136170Sjulian SLOT_USE(newtd->td_ksegrp); 954136170Sjulian newtd->td_kse->ke_flags |= KEF_DIDRUN; 955136170Sjulian TD_SET_RUNNING(newtd); 956136170Sjulian if ((newtd->td_proc->p_flag & P_NOLOAD) == 0) 957139317Sjeff sched_load_add(); 958136170Sjulian } else { 959131473Sjhb newtd = choosethread(); 960136170Sjulian } 961136170Sjulian 962121128Sjeff if (td != newtd) 963121128Sjeff cpu_switch(td, newtd); 964121128Sjeff sched_lock.mtx_lock = (uintptr_t)td; 965121128Sjeff td->td_oncpu = PCPU_GET(cpuid); 966104964Sjeff} 967104964Sjeff 968104964Sjeffvoid 969104964Sjeffsched_wakeup(struct thread *td) 970104964Sjeff{ 971104964Sjeff struct ksegrp *kg; 972104964Sjeff 973113923Sjhb mtx_assert(&sched_lock, MA_OWNED); 974104964Sjeff kg = td->td_ksegrp; 975139453Sjhb if (kg->kg_slptime > 1) { 976104964Sjeff updatepri(kg); 977139453Sjhb resetpriority(kg); 978139453Sjhb } 979104964Sjeff kg->kg_slptime = 0; 980134586Sjulian setrunqueue(td, SRQ_BORING); 981104964Sjeff} 982104964Sjeff 983134693Sjulian#ifdef SMP 984134688Sjulian/* enable HTT_2 if you have a 2-way HTT cpu.*/ 985134688Sjulianstatic int 986134688Sjulianforward_wakeup(int cpunum) 987134688Sjulian{ 988134688Sjulian cpumask_t map, me, dontuse; 989134688Sjulian cpumask_t map2; 990134688Sjulian struct pcpu *pc; 991134688Sjulian cpumask_t id, map3; 992134688Sjulian 993134688Sjulian mtx_assert(&sched_lock, MA_OWNED); 994134688Sjulian 995134791Sjulian CTR0(KTR_RUNQ, "forward_wakeup()"); 996134688Sjulian 997134688Sjulian if ((!forward_wakeup_enabled) || 998134688Sjulian (forward_wakeup_use_mask == 0 && forward_wakeup_use_loop == 0)) 999134688Sjulian return (0); 1000134688Sjulian if (!smp_started || cold || panicstr) 1001134688Sjulian return (0); 1002134688Sjulian 1003134688Sjulian forward_wakeups_requested++; 1004134688Sjulian 1005134688Sjulian/* 1006134688Sjulian * check the idle mask we received against what we calculated before 1007134688Sjulian * in the old version. 1008134688Sjulian */ 1009134688Sjulian me = PCPU_GET(cpumask); 1010134688Sjulian /* 1011134688Sjulian * don't bother if we should be doing it ourself.. 1012134688Sjulian */ 1013134688Sjulian if ((me & idle_cpus_mask) && (cpunum == NOCPU || me == (1 << cpunum))) 1014134688Sjulian return (0); 1015134688Sjulian 1016134688Sjulian dontuse = me | stopped_cpus | hlt_cpus_mask; 1017134688Sjulian map3 = 0; 1018134688Sjulian if (forward_wakeup_use_loop) { 1019134688Sjulian SLIST_FOREACH(pc, &cpuhead, pc_allcpu) { 1020134688Sjulian id = pc->pc_cpumask; 1021134688Sjulian if ( (id & dontuse) == 0 && 1022134688Sjulian pc->pc_curthread == pc->pc_idlethread) { 1023134688Sjulian map3 |= id; 1024134688Sjulian } 1025134688Sjulian } 1026134688Sjulian } 1027134688Sjulian 1028134688Sjulian if (forward_wakeup_use_mask) { 1029134688Sjulian map = 0; 1030134688Sjulian map = idle_cpus_mask & ~dontuse; 1031134688Sjulian 1032134688Sjulian /* If they are both on, compare and use loop if different */ 1033134688Sjulian if (forward_wakeup_use_loop) { 1034134688Sjulian if (map != map3) { 1035134688Sjulian printf("map (%02X) != map3 (%02X)\n", 1036134688Sjulian map, map3); 1037134688Sjulian map = map3; 1038134688Sjulian } 1039134688Sjulian } 1040134688Sjulian } else { 1041134688Sjulian map = map3; 1042134688Sjulian } 1043134688Sjulian /* If we only allow a specific CPU, then mask off all the others */ 1044134688Sjulian if (cpunum != NOCPU) { 1045134688Sjulian KASSERT((cpunum <= mp_maxcpus),("forward_wakeup: bad cpunum.")); 1046134688Sjulian map &= (1 << cpunum); 1047134688Sjulian } else { 1048134688Sjulian /* Try choose an idle die. */ 1049134688Sjulian if (forward_wakeup_use_htt) { 1050134688Sjulian map2 = (map & (map >> 1)) & 0x5555; 1051134688Sjulian if (map2) { 1052134688Sjulian map = map2; 1053134688Sjulian } 1054134688Sjulian } 1055134688Sjulian 1056134688Sjulian /* set only one bit */ 1057134688Sjulian if (forward_wakeup_use_single) { 1058134688Sjulian map = map & ((~map) + 1); 1059134688Sjulian } 1060134688Sjulian } 1061134688Sjulian if (map) { 1062134688Sjulian forward_wakeups_delivered++; 1063134688Sjulian ipi_selected(map, IPI_AST); 1064134688Sjulian return (1); 1065134688Sjulian } 1066134688Sjulian if (cpunum == NOCPU) 1067134688Sjulian printf("forward_wakeup: Idle processor not found\n"); 1068134688Sjulian return (0); 1069134688Sjulian} 1070134693Sjulian#endif 1071134688Sjulian 1072104964Sjeffvoid 1073134586Sjuliansched_add(struct thread *td, int flags) 1074104964Sjeff{ 1075121127Sjeff struct kse *ke; 1076134591Sjulian#ifdef SMP 1077134591Sjulian int forwarded = 0; 1078134591Sjulian int cpu; 1079134591Sjulian#endif 1080121127Sjeff 1081121127Sjeff ke = td->td_kse; 1082104964Sjeff mtx_assert(&sched_lock, MA_OWNED); 1083104964Sjeff KASSERT(ke->ke_state != KES_ONRUNQ, 1084124957Sjeff ("sched_add: kse %p (%s) already in run queue", ke, 1085104964Sjeff ke->ke_proc->p_comm)); 1086104964Sjeff KASSERT(ke->ke_proc->p_sflag & PS_INMEM, 1087124957Sjeff ("sched_add: process swapped out")); 1088139317Sjeff CTR5(KTR_SCHED, "sched_add: %p(%s) prio %d by %p(%s)", 1089139317Sjeff td, td->td_proc->p_comm, td->td_priority, curthread, 1090139317Sjeff curthread->td_proc->p_comm); 1091131481Sjhb 1092131481Sjhb#ifdef SMP 1093124955Sjeff if (KSE_CAN_MIGRATE(ke)) { 1094134591Sjulian CTR2(KTR_RUNQ, 1095134591Sjulian "sched_add: adding kse:%p (td:%p) to gbl runq", ke, td); 1096134591Sjulian cpu = NOCPU; 1097124955Sjeff ke->ke_runq = &runq; 1098124955Sjeff } else { 1099124955Sjeff if (!SKE_RUNQ_PCPU(ke)) 1100134591Sjulian ke->ke_runq = &runq_pcpu[(cpu = PCPU_GET(cpuid))]; 1101134591Sjulian else 1102134591Sjulian cpu = td->td_lastcpu; 1103134591Sjulian CTR3(KTR_RUNQ, 1104134591Sjulian "sched_add: Put kse:%p(td:%p) on cpu%d runq", ke, td, cpu); 1105124955Sjeff } 1106124955Sjeff#else 1107133396Sjulian CTR2(KTR_RUNQ, "sched_add: adding kse:%p (td:%p) to runq", ke, td); 1108124955Sjeff ke->ke_runq = &runq; 1109134591Sjulian 1110124955Sjeff#endif 1111134591Sjulian /* 1112134591Sjulian * If we are yielding (on the way out anyhow) 1113134591Sjulian * or the thread being saved is US, 1114134591Sjulian * then don't try be smart about preemption 1115134591Sjulian * or kicking off another CPU 1116134591Sjulian * as it won't help and may hinder. 1117134591Sjulian * In the YIEDLING case, we are about to run whoever is 1118134591Sjulian * being put in the queue anyhow, and in the 1119134591Sjulian * OURSELF case, we are puting ourself on the run queue 1120134591Sjulian * which also only happens when we are about to yield. 1121134591Sjulian */ 1122134591Sjulian if((flags & SRQ_YIELDING) == 0) { 1123134591Sjulian#ifdef SMP 1124134591Sjulian cpumask_t me = PCPU_GET(cpumask); 1125134591Sjulian int idle = idle_cpus_mask & me; 1126134591Sjulian /* 1127134591Sjulian * Only try to kick off another CPU if 1128134591Sjulian * the thread is unpinned 1129134591Sjulian * or pinned to another cpu, 1130134591Sjulian * and there are other available and idle CPUs. 1131134837Sjulian * if we are idle, or it's an interrupt, 1132134837Sjulian * then skip straight to preemption. 1133134591Sjulian */ 1134134837Sjulian if ( (! idle) && ((flags & SRQ_INTR) == 0) && 1135134591Sjulian (idle_cpus_mask & ~(hlt_cpus_mask | me)) && 1136134591Sjulian ( KSE_CAN_MIGRATE(ke) || 1137134591Sjulian ke->ke_runq != &runq_pcpu[PCPU_GET(cpuid)])) { 1138134591Sjulian forwarded = forward_wakeup(cpu); 1139134591Sjulian } 1140134591Sjulian /* 1141134591Sjulian * If we failed to kick off another cpu, then look to 1142134591Sjulian * see if we should preempt this CPU. Only allow this 1143134591Sjulian * if it is not pinned or IS pinned to this CPU. 1144134591Sjulian * If we are the idle thread, we also try do preempt. 1145134591Sjulian * as it will be quicker and being idle, we won't 1146134591Sjulian * lose in doing so.. 1147134591Sjulian */ 1148134591Sjulian if ((!forwarded) && 1149134591Sjulian (ke->ke_runq == &runq || 1150134591Sjulian ke->ke_runq == &runq_pcpu[PCPU_GET(cpuid)])) 1151134591Sjulian#endif 1152134591Sjulian 1153134591Sjulian { 1154134591Sjulian if (maybe_preempt(td)) 1155134591Sjulian return; 1156134591Sjulian } 1157134591Sjulian } 1158125295Sjeff if ((td->td_proc->p_flag & P_NOLOAD) == 0) 1159139317Sjeff sched_load_add(); 1160136170Sjulian SLOT_USE(td->td_ksegrp); 1161136170Sjulian runq_add(ke->ke_runq, ke, flags); 1162133520Sjulian ke->ke_state = KES_ONRUNQ; 1163132118Sjhb maybe_resched(td); 1164104964Sjeff} 1165104964Sjeff 1166104964Sjeffvoid 1167121127Sjeffsched_rem(struct thread *td) 1168104964Sjeff{ 1169121127Sjeff struct kse *ke; 1170121127Sjeff 1171121127Sjeff ke = td->td_kse; 1172104964Sjeff KASSERT(ke->ke_proc->p_sflag & PS_INMEM, 1173124957Sjeff ("sched_rem: process swapped out")); 1174124957Sjeff KASSERT((ke->ke_state == KES_ONRUNQ), 1175124957Sjeff ("sched_rem: KSE not on run queue")); 1176104964Sjeff mtx_assert(&sched_lock, MA_OWNED); 1177139317Sjeff CTR5(KTR_SCHED, "sched_rem: %p(%s) prio %d by %p(%s)", 1178139317Sjeff td, td->td_proc->p_comm, td->td_priority, curthread, 1179139317Sjeff curthread->td_proc->p_comm); 1180104964Sjeff 1181125295Sjeff if ((td->td_proc->p_flag & P_NOLOAD) == 0) 1182139317Sjeff sched_load_rem(); 1183136167Sjulian SLOT_RELEASE(td->td_ksegrp); 1184134145Sjulian runq_remove(ke->ke_runq, ke); 1185124955Sjeff 1186104964Sjeff ke->ke_state = KES_THREAD; 1187104964Sjeff} 1188104964Sjeff 1189135295Sjulian/* 1190135295Sjulian * Select threads to run. 1191135295Sjulian * Notice that the running threads still consume a slot. 1192135295Sjulian */ 1193104964Sjeffstruct kse * 1194104964Sjeffsched_choose(void) 1195104964Sjeff{ 1196104964Sjeff struct kse *ke; 1197124955Sjeff struct runq *rq; 1198104964Sjeff 1199124955Sjeff#ifdef SMP 1200124955Sjeff struct kse *kecpu; 1201124955Sjeff 1202124955Sjeff rq = &runq; 1203104964Sjeff ke = runq_choose(&runq); 1204124955Sjeff kecpu = runq_choose(&runq_pcpu[PCPU_GET(cpuid)]); 1205104964Sjeff 1206124955Sjeff if (ke == NULL || 1207124955Sjeff (kecpu != NULL && 1208124955Sjeff kecpu->ke_thread->td_priority < ke->ke_thread->td_priority)) { 1209133396Sjulian CTR2(KTR_RUNQ, "choosing kse %p from pcpu runq %d", kecpu, 1210124955Sjeff PCPU_GET(cpuid)); 1211124955Sjeff ke = kecpu; 1212124955Sjeff rq = &runq_pcpu[PCPU_GET(cpuid)]; 1213124955Sjeff } else { 1214133396Sjulian CTR1(KTR_RUNQ, "choosing kse %p from main runq", ke); 1215124955Sjeff } 1216124955Sjeff 1217124955Sjeff#else 1218124955Sjeff rq = &runq; 1219124955Sjeff ke = runq_choose(&runq); 1220124955Sjeff#endif 1221124955Sjeff 1222104964Sjeff if (ke != NULL) { 1223124955Sjeff runq_remove(rq, ke); 1224104964Sjeff ke->ke_state = KES_THREAD; 1225104964Sjeff 1226104964Sjeff KASSERT(ke->ke_proc->p_sflag & PS_INMEM, 1227124957Sjeff ("sched_choose: process swapped out")); 1228104964Sjeff } 1229104964Sjeff return (ke); 1230104964Sjeff} 1231104964Sjeff 1232104964Sjeffvoid 1233104964Sjeffsched_userret(struct thread *td) 1234104964Sjeff{ 1235104964Sjeff struct ksegrp *kg; 1236104964Sjeff /* 1237104964Sjeff * XXX we cheat slightly on the locking here to avoid locking in 1238104964Sjeff * the usual case. Setting td_priority here is essentially an 1239104964Sjeff * incomplete workaround for not setting it properly elsewhere. 1240104964Sjeff * Now that some interrupt handlers are threads, not setting it 1241104964Sjeff * properly elsewhere can clobber it in the window between setting 1242104964Sjeff * it here and returning to user mode, so don't waste time setting 1243104964Sjeff * it perfectly here. 1244104964Sjeff */ 1245139453Sjhb KASSERT((td->td_flags & TDF_BORROWING) == 0, 1246139453Sjhb ("thread with borrowed priority returning to userland")); 1247104964Sjeff kg = td->td_ksegrp; 1248104964Sjeff if (td->td_priority != kg->kg_user_pri) { 1249104964Sjeff mtx_lock_spin(&sched_lock); 1250104964Sjeff td->td_priority = kg->kg_user_pri; 1251139453Sjhb td->td_base_pri = kg->kg_user_pri; 1252104964Sjeff mtx_unlock_spin(&sched_lock); 1253104964Sjeff } 1254104964Sjeff} 1255107126Sjeff 1256124955Sjeffvoid 1257124955Sjeffsched_bind(struct thread *td, int cpu) 1258124955Sjeff{ 1259124955Sjeff struct kse *ke; 1260124955Sjeff 1261124955Sjeff mtx_assert(&sched_lock, MA_OWNED); 1262124955Sjeff KASSERT(TD_IS_RUNNING(td), 1263124955Sjeff ("sched_bind: cannot bind non-running thread")); 1264124955Sjeff 1265124955Sjeff ke = td->td_kse; 1266124955Sjeff 1267124955Sjeff ke->ke_flags |= KEF_BOUND; 1268124955Sjeff#ifdef SMP 1269124955Sjeff ke->ke_runq = &runq_pcpu[cpu]; 1270124955Sjeff if (PCPU_GET(cpuid) == cpu) 1271124955Sjeff return; 1272124955Sjeff 1273124955Sjeff ke->ke_state = KES_THREAD; 1274124955Sjeff 1275131473Sjhb mi_switch(SW_VOL, NULL); 1276124955Sjeff#endif 1277124955Sjeff} 1278124955Sjeff 1279124955Sjeffvoid 1280124955Sjeffsched_unbind(struct thread* td) 1281124955Sjeff{ 1282124955Sjeff mtx_assert(&sched_lock, MA_OWNED); 1283124955Sjeff td->td_kse->ke_flags &= ~KEF_BOUND; 1284124955Sjeff} 1285124955Sjeff 1286107126Sjeffint 1287125288Sjeffsched_load(void) 1288125288Sjeff{ 1289125288Sjeff return (sched_tdcnt); 1290125288Sjeff} 1291125288Sjeff 1292125288Sjeffint 1293107126Sjeffsched_sizeof_ksegrp(void) 1294107126Sjeff{ 1295134791Sjulian return (sizeof(struct ksegrp) + sizeof(struct kg_sched)); 1296107126Sjeff} 1297107126Sjeffint 1298107126Sjeffsched_sizeof_proc(void) 1299107126Sjeff{ 1300107126Sjeff return (sizeof(struct proc)); 1301107126Sjeff} 1302107126Sjeffint 1303107126Sjeffsched_sizeof_thread(void) 1304107126Sjeff{ 1305134791Sjulian return (sizeof(struct thread) + sizeof(struct kse)); 1306107126Sjeff} 1307107137Sjeff 1308107137Sjefffixpt_t 1309121127Sjeffsched_pctcpu(struct thread *td) 1310107137Sjeff{ 1311121147Sjeff struct kse *ke; 1312121147Sjeff 1313121147Sjeff ke = td->td_kse; 1314134791Sjulian return (ke->ke_pctcpu); 1315121147Sjeff 1316121147Sjeff return (0); 1317107137Sjeff} 1318134791Sjulian#define KERN_SWITCH_INCLUDE 1 1319134791Sjulian#include "kern/kern_switch.c" 1320