sched_ule.c revision 172308
1/*- 2 * Copyright (c) 2002-2007, Jeffrey Roberson <jeff@freebsd.org> 3 * All rights reserved. 4 * 5 * Redistribution and use in source and binary forms, with or without 6 * modification, are permitted provided that the following conditions 7 * are met: 8 * 1. Redistributions of source code must retain the above copyright 9 * notice unmodified, this list of conditions, and the following 10 * disclaimer. 11 * 2. Redistributions in binary form must reproduce the above copyright 12 * notice, this list of conditions and the following disclaimer in the 13 * documentation and/or other materials provided with the distribution. 14 * 15 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR 16 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES 17 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. 18 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, 19 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT 20 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 21 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 22 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 23 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF 24 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 25 */ 26 27/* 28 * This file implements the ULE scheduler. ULE supports independent CPU 29 * run queues and fine grain locking. It has superior interactive 30 * performance under load even on uni-processor systems. 31 * 32 * etymology: 33 * ULE is the last three letters in schedule. It owes its name to a 34 * generic user created for a scheduling system by Paul Mikesell at 35 * Isilon Systems and a general lack of creativity on the part of the author. 36 */ 37 38#include <sys/cdefs.h> 39__FBSDID("$FreeBSD: head/sys/kern/sched_ule.c 172308 2007-09-24 00:28:54Z jeff $"); 40 41#include "opt_hwpmc_hooks.h" 42#include "opt_sched.h" 43 44#include <sys/param.h> 45#include <sys/systm.h> 46#include <sys/kdb.h> 47#include <sys/kernel.h> 48#include <sys/ktr.h> 49#include <sys/lock.h> 50#include <sys/mutex.h> 51#include <sys/proc.h> 52#include <sys/resource.h> 53#include <sys/resourcevar.h> 54#include <sys/sched.h> 55#include <sys/smp.h> 56#include <sys/sx.h> 57#include <sys/sysctl.h> 58#include <sys/sysproto.h> 59#include <sys/turnstile.h> 60#include <sys/umtx.h> 61#include <sys/vmmeter.h> 62#ifdef KTRACE 63#include <sys/uio.h> 64#include <sys/ktrace.h> 65#endif 66 67#ifdef HWPMC_HOOKS 68#include <sys/pmckern.h> 69#endif 70 71#include <machine/cpu.h> 72#include <machine/smp.h> 73 74#ifndef PREEMPTION 75#error "SCHED_ULE requires options PREEMPTION" 76#endif 77 78#define KTR_ULE 0 79 80/* 81 * Thread scheduler specific section. All fields are protected 82 * by the thread lock. 83 */ 84struct td_sched { 85 TAILQ_ENTRY(td_sched) ts_procq; /* Run queue. */ 86 struct thread *ts_thread; /* Active associated thread. */ 87 struct runq *ts_runq; /* Run-queue we're queued on. */ 88 short ts_flags; /* TSF_* flags. */ 89 u_char ts_rqindex; /* Run queue index. */ 90 u_char ts_cpu; /* CPU that we have affinity for. */ 91 int ts_slice; /* Ticks of slice remaining. */ 92 u_int ts_slptime; /* Number of ticks we vol. slept */ 93 u_int ts_runtime; /* Number of ticks we were running */ 94 /* The following variables are only used for pctcpu calculation */ 95 int ts_ltick; /* Last tick that we were running on */ 96 int ts_ftick; /* First tick that we were running on */ 97 int ts_ticks; /* Tick count */ 98#ifdef SMP 99 int ts_rltick; /* Real last tick, for affinity. */ 100#endif 101}; 102/* flags kept in ts_flags */ 103#define TSF_BOUND 0x0001 /* Thread can not migrate. */ 104#define TSF_XFERABLE 0x0002 /* Thread was added as transferable. */ 105 106static struct td_sched td_sched0; 107 108/* 109 * Cpu percentage computation macros and defines. 110 * 111 * SCHED_TICK_SECS: Number of seconds to average the cpu usage across. 112 * SCHED_TICK_TARG: Number of hz ticks to average the cpu usage across. 113 * SCHED_TICK_MAX: Maximum number of ticks before scaling back. 114 * SCHED_TICK_SHIFT: Shift factor to avoid rounding away results. 115 * SCHED_TICK_HZ: Compute the number of hz ticks for a given ticks count. 116 * SCHED_TICK_TOTAL: Gives the amount of time we've been recording ticks. 117 */ 118#define SCHED_TICK_SECS 10 119#define SCHED_TICK_TARG (hz * SCHED_TICK_SECS) 120#define SCHED_TICK_MAX (SCHED_TICK_TARG + hz) 121#define SCHED_TICK_SHIFT 10 122#define SCHED_TICK_HZ(ts) ((ts)->ts_ticks >> SCHED_TICK_SHIFT) 123#define SCHED_TICK_TOTAL(ts) (max((ts)->ts_ltick - (ts)->ts_ftick, hz)) 124 125/* 126 * These macros determine priorities for non-interactive threads. They are 127 * assigned a priority based on their recent cpu utilization as expressed 128 * by the ratio of ticks to the tick total. NHALF priorities at the start 129 * and end of the MIN to MAX timeshare range are only reachable with negative 130 * or positive nice respectively. 131 * 132 * PRI_RANGE: Priority range for utilization dependent priorities. 133 * PRI_NRESV: Number of nice values. 134 * PRI_TICKS: Compute a priority in PRI_RANGE from the ticks count and total. 135 * PRI_NICE: Determines the part of the priority inherited from nice. 136 */ 137#define SCHED_PRI_NRESV (PRIO_MAX - PRIO_MIN) 138#define SCHED_PRI_NHALF (SCHED_PRI_NRESV / 2) 139#define SCHED_PRI_MIN (PRI_MIN_TIMESHARE + SCHED_PRI_NHALF) 140#define SCHED_PRI_MAX (PRI_MAX_TIMESHARE - SCHED_PRI_NHALF) 141#define SCHED_PRI_RANGE (SCHED_PRI_MAX - SCHED_PRI_MIN) 142#define SCHED_PRI_TICKS(ts) \ 143 (SCHED_TICK_HZ((ts)) / \ 144 (roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE)) 145#define SCHED_PRI_NICE(nice) (nice) 146 147/* 148 * These determine the interactivity of a process. Interactivity differs from 149 * cpu utilization in that it expresses the voluntary time slept vs time ran 150 * while cpu utilization includes all time not running. This more accurately 151 * models the intent of the thread. 152 * 153 * SLP_RUN_MAX: Maximum amount of sleep time + run time we'll accumulate 154 * before throttling back. 155 * SLP_RUN_FORK: Maximum slp+run time to inherit at fork time. 156 * INTERACT_MAX: Maximum interactivity value. Smaller is better. 157 * INTERACT_THRESH: Threshhold for placement on the current runq. 158 */ 159#define SCHED_SLP_RUN_MAX ((hz * 5) << SCHED_TICK_SHIFT) 160#define SCHED_SLP_RUN_FORK ((hz / 2) << SCHED_TICK_SHIFT) 161#define SCHED_INTERACT_MAX (100) 162#define SCHED_INTERACT_HALF (SCHED_INTERACT_MAX / 2) 163#define SCHED_INTERACT_THRESH (30) 164 165/* 166 * tickincr: Converts a stathz tick into a hz domain scaled by 167 * the shift factor. Without the shift the error rate 168 * due to rounding would be unacceptably high. 169 * realstathz: stathz is sometimes 0 and run off of hz. 170 * sched_slice: Runtime of each thread before rescheduling. 171 * preempt_thresh: Priority threshold for preemption and remote IPIs. 172 */ 173static int sched_interact = SCHED_INTERACT_THRESH; 174static int realstathz; 175static int tickincr; 176static int sched_slice; 177static int preempt_thresh = PRI_MIN_KERN; 178 179/* 180 * tdq - per processor runqs and statistics. All fields are protected by the 181 * tdq_lock. The load and lowpri may be accessed without to avoid excess 182 * locking in sched_pickcpu(); 183 */ 184struct tdq { 185 struct mtx *tdq_lock; /* Pointer to group lock. */ 186 struct runq tdq_realtime; /* real-time run queue. */ 187 struct runq tdq_timeshare; /* timeshare run queue. */ 188 struct runq tdq_idle; /* Queue of IDLE threads. */ 189 int tdq_load; /* Aggregate load. */ 190 u_char tdq_idx; /* Current insert index. */ 191 u_char tdq_ridx; /* Current removal index. */ 192#ifdef SMP 193 u_char tdq_lowpri; /* Lowest priority thread. */ 194 int tdq_transferable; /* Transferable thread count. */ 195 LIST_ENTRY(tdq) tdq_siblings; /* Next in tdq group. */ 196 struct tdq_group *tdq_group; /* Our processor group. */ 197#else 198 int tdq_sysload; /* For loadavg, !ITHD load. */ 199#endif 200} __aligned(64); 201 202 203#ifdef SMP 204/* 205 * tdq groups are groups of processors which can cheaply share threads. When 206 * one processor in the group goes idle it will check the runqs of the other 207 * processors in its group prior to halting and waiting for an interrupt. 208 * These groups are suitable for SMT (Symetric Multi-Threading) and not NUMA. 209 * In a numa environment we'd want an idle bitmap per group and a two tiered 210 * load balancer. 211 */ 212struct tdq_group { 213 struct mtx tdg_lock; /* Protects all fields below. */ 214 int tdg_cpus; /* Count of CPUs in this tdq group. */ 215 cpumask_t tdg_cpumask; /* Mask of cpus in this group. */ 216 cpumask_t tdg_idlemask; /* Idle cpus in this group. */ 217 cpumask_t tdg_mask; /* Bit mask for first cpu. */ 218 int tdg_load; /* Total load of this group. */ 219 int tdg_transferable; /* Transferable load of this group. */ 220 LIST_HEAD(, tdq) tdg_members; /* Linked list of all members. */ 221 char tdg_name[16]; /* lock name. */ 222} __aligned(64); 223 224#define SCHED_AFFINITY_DEFAULT (max(1, hz / 300)) 225#define SCHED_AFFINITY(ts) ((ts)->ts_rltick > ticks - affinity) 226 227/* 228 * Run-time tunables. 229 */ 230static int rebalance = 1; 231static int balance_secs = 1; 232static int pick_pri = 1; 233static int affinity; 234static int tryself = 1; 235static int steal_htt = 0; 236static int steal_idle = 1; 237static int steal_thresh = 2; 238static int topology = 0; 239 240/* 241 * One thread queue per processor. 242 */ 243static volatile cpumask_t tdq_idle; 244static int tdg_maxid; 245static struct tdq tdq_cpu[MAXCPU]; 246static struct tdq_group tdq_groups[MAXCPU]; 247static struct callout balco; 248static struct callout gbalco; 249 250#define TDQ_SELF() (&tdq_cpu[PCPU_GET(cpuid)]) 251#define TDQ_CPU(x) (&tdq_cpu[(x)]) 252#define TDQ_ID(x) ((int)((x) - tdq_cpu)) 253#define TDQ_GROUP(x) (&tdq_groups[(x)]) 254#define TDG_ID(x) ((int)((x) - tdq_groups)) 255#else /* !SMP */ 256static struct tdq tdq_cpu; 257static struct mtx tdq_lock; 258 259#define TDQ_ID(x) (0) 260#define TDQ_SELF() (&tdq_cpu) 261#define TDQ_CPU(x) (&tdq_cpu) 262#endif 263 264#define TDQ_LOCK_ASSERT(t, type) mtx_assert(TDQ_LOCKPTR((t)), (type)) 265#define TDQ_LOCK(t) mtx_lock_spin(TDQ_LOCKPTR((t))) 266#define TDQ_LOCK_FLAGS(t, f) mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f)) 267#define TDQ_UNLOCK(t) mtx_unlock_spin(TDQ_LOCKPTR((t))) 268#define TDQ_LOCKPTR(t) ((t)->tdq_lock) 269 270static void sched_priority(struct thread *); 271static void sched_thread_priority(struct thread *, u_char); 272static int sched_interact_score(struct thread *); 273static void sched_interact_update(struct thread *); 274static void sched_interact_fork(struct thread *); 275static void sched_pctcpu_update(struct td_sched *); 276 277/* Operations on per processor queues */ 278static struct td_sched * tdq_choose(struct tdq *); 279static void tdq_setup(struct tdq *); 280static void tdq_load_add(struct tdq *, struct td_sched *); 281static void tdq_load_rem(struct tdq *, struct td_sched *); 282static __inline void tdq_runq_add(struct tdq *, struct td_sched *, int); 283static __inline void tdq_runq_rem(struct tdq *, struct td_sched *); 284void tdq_print(int cpu); 285static void runq_print(struct runq *rq); 286static void tdq_add(struct tdq *, struct thread *, int); 287#ifdef SMP 288static void tdq_move(struct tdq *, struct tdq *); 289static int tdq_idled(struct tdq *); 290static void tdq_notify(struct td_sched *); 291static struct td_sched *tdq_steal(struct tdq *, int); 292static struct td_sched *runq_steal(struct runq *); 293static int sched_pickcpu(struct td_sched *, int); 294static void sched_balance(void *); 295static void sched_balance_groups(void *); 296static void sched_balance_group(struct tdq_group *); 297static void sched_balance_pair(struct tdq *, struct tdq *); 298static inline struct tdq *sched_setcpu(struct td_sched *, int, int); 299static inline struct mtx *thread_block_switch(struct thread *); 300static inline void thread_unblock_switch(struct thread *, struct mtx *); 301static struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int); 302 303#define THREAD_CAN_MIGRATE(td) ((td)->td_pinned == 0) 304#endif 305 306static void sched_setup(void *dummy); 307SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL) 308 309static void sched_initticks(void *dummy); 310SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks, NULL) 311 312/* 313 * Print the threads waiting on a run-queue. 314 */ 315static void 316runq_print(struct runq *rq) 317{ 318 struct rqhead *rqh; 319 struct td_sched *ts; 320 int pri; 321 int j; 322 int i; 323 324 for (i = 0; i < RQB_LEN; i++) { 325 printf("\t\trunq bits %d 0x%zx\n", 326 i, rq->rq_status.rqb_bits[i]); 327 for (j = 0; j < RQB_BPW; j++) 328 if (rq->rq_status.rqb_bits[i] & (1ul << j)) { 329 pri = j + (i << RQB_L2BPW); 330 rqh = &rq->rq_queues[pri]; 331 TAILQ_FOREACH(ts, rqh, ts_procq) { 332 printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n", 333 ts->ts_thread, ts->ts_thread->td_proc->p_comm, ts->ts_thread->td_priority, ts->ts_rqindex, pri); 334 } 335 } 336 } 337} 338 339/* 340 * Print the status of a per-cpu thread queue. Should be a ddb show cmd. 341 */ 342void 343tdq_print(int cpu) 344{ 345 struct tdq *tdq; 346 347 tdq = TDQ_CPU(cpu); 348 349 printf("tdq %d:\n", TDQ_ID(tdq)); 350 printf("\tlockptr %p\n", TDQ_LOCKPTR(tdq)); 351 printf("\tload: %d\n", tdq->tdq_load); 352 printf("\ttimeshare idx: %d\n", tdq->tdq_idx); 353 printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx); 354 printf("\trealtime runq:\n"); 355 runq_print(&tdq->tdq_realtime); 356 printf("\ttimeshare runq:\n"); 357 runq_print(&tdq->tdq_timeshare); 358 printf("\tidle runq:\n"); 359 runq_print(&tdq->tdq_idle); 360#ifdef SMP 361 printf("\tload transferable: %d\n", tdq->tdq_transferable); 362 printf("\tlowest priority: %d\n", tdq->tdq_lowpri); 363 printf("\tgroup: %d\n", TDG_ID(tdq->tdq_group)); 364 printf("\tLock name: %s\n", tdq->tdq_group->tdg_name); 365#endif 366} 367 368#define TS_RQ_PPQ (((PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE) + 1) / RQ_NQS) 369/* 370 * Add a thread to the actual run-queue. Keeps transferable counts up to 371 * date with what is actually on the run-queue. Selects the correct 372 * queue position for timeshare threads. 373 */ 374static __inline void 375tdq_runq_add(struct tdq *tdq, struct td_sched *ts, int flags) 376{ 377 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 378 THREAD_LOCK_ASSERT(ts->ts_thread, MA_OWNED); 379#ifdef SMP 380 if (THREAD_CAN_MIGRATE(ts->ts_thread)) { 381 tdq->tdq_transferable++; 382 tdq->tdq_group->tdg_transferable++; 383 ts->ts_flags |= TSF_XFERABLE; 384 } 385#endif 386 if (ts->ts_runq == &tdq->tdq_timeshare) { 387 u_char pri; 388 389 pri = ts->ts_thread->td_priority; 390 KASSERT(pri <= PRI_MAX_TIMESHARE && pri >= PRI_MIN_TIMESHARE, 391 ("Invalid priority %d on timeshare runq", pri)); 392 /* 393 * This queue contains only priorities between MIN and MAX 394 * realtime. Use the whole queue to represent these values. 395 */ 396 if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) { 397 pri = (pri - PRI_MIN_TIMESHARE) / TS_RQ_PPQ; 398 pri = (pri + tdq->tdq_idx) % RQ_NQS; 399 /* 400 * This effectively shortens the queue by one so we 401 * can have a one slot difference between idx and 402 * ridx while we wait for threads to drain. 403 */ 404 if (tdq->tdq_ridx != tdq->tdq_idx && 405 pri == tdq->tdq_ridx) 406 pri = (unsigned char)(pri - 1) % RQ_NQS; 407 } else 408 pri = tdq->tdq_ridx; 409 runq_add_pri(ts->ts_runq, ts, pri, flags); 410 } else 411 runq_add(ts->ts_runq, ts, flags); 412} 413 414/* 415 * Remove a thread from a run-queue. This typically happens when a thread 416 * is selected to run. Running threads are not on the queue and the 417 * transferable count does not reflect them. 418 */ 419static __inline void 420tdq_runq_rem(struct tdq *tdq, struct td_sched *ts) 421{ 422 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 423 KASSERT(ts->ts_runq != NULL, 424 ("tdq_runq_remove: thread %p null ts_runq", ts->ts_thread)); 425#ifdef SMP 426 if (ts->ts_flags & TSF_XFERABLE) { 427 tdq->tdq_transferable--; 428 tdq->tdq_group->tdg_transferable--; 429 ts->ts_flags &= ~TSF_XFERABLE; 430 } 431#endif 432 if (ts->ts_runq == &tdq->tdq_timeshare) { 433 if (tdq->tdq_idx != tdq->tdq_ridx) 434 runq_remove_idx(ts->ts_runq, ts, &tdq->tdq_ridx); 435 else 436 runq_remove_idx(ts->ts_runq, ts, NULL); 437 /* 438 * For timeshare threads we update the priority here so 439 * the priority reflects the time we've been sleeping. 440 */ 441 ts->ts_ltick = ticks; 442 sched_pctcpu_update(ts); 443 sched_priority(ts->ts_thread); 444 } else 445 runq_remove(ts->ts_runq, ts); 446} 447 448/* 449 * Load is maintained for all threads RUNNING and ON_RUNQ. Add the load 450 * for this thread to the referenced thread queue. 451 */ 452static void 453tdq_load_add(struct tdq *tdq, struct td_sched *ts) 454{ 455 int class; 456 457 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 458 THREAD_LOCK_ASSERT(ts->ts_thread, MA_OWNED); 459 class = PRI_BASE(ts->ts_thread->td_pri_class); 460 tdq->tdq_load++; 461 CTR2(KTR_SCHED, "cpu %d load: %d", TDQ_ID(tdq), tdq->tdq_load); 462 if (class != PRI_ITHD && 463 (ts->ts_thread->td_proc->p_flag & P_NOLOAD) == 0) 464#ifdef SMP 465 tdq->tdq_group->tdg_load++; 466#else 467 tdq->tdq_sysload++; 468#endif 469} 470 471/* 472 * Remove the load from a thread that is transitioning to a sleep state or 473 * exiting. 474 */ 475static void 476tdq_load_rem(struct tdq *tdq, struct td_sched *ts) 477{ 478 int class; 479 480 THREAD_LOCK_ASSERT(ts->ts_thread, MA_OWNED); 481 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 482 class = PRI_BASE(ts->ts_thread->td_pri_class); 483 if (class != PRI_ITHD && 484 (ts->ts_thread->td_proc->p_flag & P_NOLOAD) == 0) 485#ifdef SMP 486 tdq->tdq_group->tdg_load--; 487#else 488 tdq->tdq_sysload--; 489#endif 490 KASSERT(tdq->tdq_load != 0, 491 ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq))); 492 tdq->tdq_load--; 493 CTR1(KTR_SCHED, "load: %d", tdq->tdq_load); 494 ts->ts_runq = NULL; 495} 496 497#ifdef SMP 498/* 499 * sched_balance is a simple CPU load balancing algorithm. It operates by 500 * finding the least loaded and most loaded cpu and equalizing their load 501 * by migrating some processes. 502 * 503 * Dealing only with two CPUs at a time has two advantages. Firstly, most 504 * installations will only have 2 cpus. Secondly, load balancing too much at 505 * once can have an unpleasant effect on the system. The scheduler rarely has 506 * enough information to make perfect decisions. So this algorithm chooses 507 * simplicity and more gradual effects on load in larger systems. 508 * 509 */ 510static void 511sched_balance(void *arg) 512{ 513 struct tdq_group *high; 514 struct tdq_group *low; 515 struct tdq_group *tdg; 516 int cnt; 517 int i; 518 519 callout_reset(&balco, max(hz / 2, random() % (hz * balance_secs)), 520 sched_balance, NULL); 521 if (smp_started == 0 || rebalance == 0) 522 return; 523 low = high = NULL; 524 i = random() % (tdg_maxid + 1); 525 for (cnt = 0; cnt <= tdg_maxid; cnt++) { 526 tdg = TDQ_GROUP(i); 527 /* 528 * Find the CPU with the highest load that has some 529 * threads to transfer. 530 */ 531 if ((high == NULL || tdg->tdg_load > high->tdg_load) 532 && tdg->tdg_transferable) 533 high = tdg; 534 if (low == NULL || tdg->tdg_load < low->tdg_load) 535 low = tdg; 536 if (++i > tdg_maxid) 537 i = 0; 538 } 539 if (low != NULL && high != NULL && high != low) 540 sched_balance_pair(LIST_FIRST(&high->tdg_members), 541 LIST_FIRST(&low->tdg_members)); 542} 543 544/* 545 * Balance load between CPUs in a group. Will only migrate within the group. 546 */ 547static void 548sched_balance_groups(void *arg) 549{ 550 int i; 551 552 callout_reset(&gbalco, max(hz / 2, random() % (hz * balance_secs)), 553 sched_balance_groups, NULL); 554 if (smp_started == 0 || rebalance == 0) 555 return; 556 for (i = 0; i <= tdg_maxid; i++) 557 sched_balance_group(TDQ_GROUP(i)); 558} 559 560/* 561 * Finds the greatest imbalance between two tdqs in a group. 562 */ 563static void 564sched_balance_group(struct tdq_group *tdg) 565{ 566 struct tdq *tdq; 567 struct tdq *high; 568 struct tdq *low; 569 int load; 570 571 if (tdg->tdg_transferable == 0) 572 return; 573 low = NULL; 574 high = NULL; 575 LIST_FOREACH(tdq, &tdg->tdg_members, tdq_siblings) { 576 load = tdq->tdq_load; 577 if (high == NULL || load > high->tdq_load) 578 high = tdq; 579 if (low == NULL || load < low->tdq_load) 580 low = tdq; 581 } 582 if (high != NULL && low != NULL && high != low) 583 sched_balance_pair(high, low); 584} 585 586/* 587 * Lock two thread queues using their address to maintain lock order. 588 */ 589static void 590tdq_lock_pair(struct tdq *one, struct tdq *two) 591{ 592 if (one < two) { 593 TDQ_LOCK(one); 594 TDQ_LOCK_FLAGS(two, MTX_DUPOK); 595 } else { 596 TDQ_LOCK(two); 597 TDQ_LOCK_FLAGS(one, MTX_DUPOK); 598 } 599} 600 601/* 602 * Transfer load between two imbalanced thread queues. 603 */ 604static void 605sched_balance_pair(struct tdq *high, struct tdq *low) 606{ 607 int transferable; 608 int high_load; 609 int low_load; 610 int move; 611 int diff; 612 int i; 613 614 tdq_lock_pair(high, low); 615 /* 616 * If we're transfering within a group we have to use this specific 617 * tdq's transferable count, otherwise we can steal from other members 618 * of the group. 619 */ 620 if (high->tdq_group == low->tdq_group) { 621 transferable = high->tdq_transferable; 622 high_load = high->tdq_load; 623 low_load = low->tdq_load; 624 } else { 625 transferable = high->tdq_group->tdg_transferable; 626 high_load = high->tdq_group->tdg_load; 627 low_load = low->tdq_group->tdg_load; 628 } 629 /* 630 * Determine what the imbalance is and then adjust that to how many 631 * threads we actually have to give up (transferable). 632 */ 633 if (transferable != 0) { 634 diff = high_load - low_load; 635 move = diff / 2; 636 if (diff & 0x1) 637 move++; 638 move = min(move, transferable); 639 for (i = 0; i < move; i++) 640 tdq_move(high, low); 641 /* 642 * IPI the target cpu to force it to reschedule with the new 643 * workload. 644 */ 645 ipi_selected(1 << TDQ_ID(low), IPI_PREEMPT); 646 } 647 TDQ_UNLOCK(high); 648 TDQ_UNLOCK(low); 649 return; 650} 651 652/* 653 * Move a thread from one thread queue to another. 654 */ 655static void 656tdq_move(struct tdq *from, struct tdq *to) 657{ 658 struct td_sched *ts; 659 struct thread *td; 660 struct tdq *tdq; 661 int cpu; 662 663 tdq = from; 664 cpu = TDQ_ID(to); 665 ts = tdq_steal(tdq, 1); 666 if (ts == NULL) { 667 struct tdq_group *tdg; 668 669 tdg = tdq->tdq_group; 670 LIST_FOREACH(tdq, &tdg->tdg_members, tdq_siblings) { 671 if (tdq == from || tdq->tdq_transferable == 0) 672 continue; 673 ts = tdq_steal(tdq, 1); 674 break; 675 } 676 if (ts == NULL) 677 return; 678 } 679 if (tdq == to) 680 return; 681 td = ts->ts_thread; 682 /* 683 * Although the run queue is locked the thread may be blocked. Lock 684 * it to clear this. 685 */ 686 thread_lock(td); 687 /* Drop recursive lock on from. */ 688 TDQ_UNLOCK(from); 689 sched_rem(td); 690 ts->ts_cpu = cpu; 691 td->td_lock = TDQ_LOCKPTR(to); 692 tdq_add(to, td, SRQ_YIELDING); 693} 694 695/* 696 * This tdq has idled. Try to steal a thread from another cpu and switch 697 * to it. 698 */ 699static int 700tdq_idled(struct tdq *tdq) 701{ 702 struct tdq_group *tdg; 703 struct tdq *steal; 704 struct td_sched *ts; 705 struct thread *td; 706 int highload; 707 int highcpu; 708 int load; 709 int cpu; 710 711 /* We don't want to be preempted while we're iterating over tdqs */ 712 spinlock_enter(); 713 tdg = tdq->tdq_group; 714 /* 715 * If we're in a cpu group, try and steal threads from another cpu in 716 * the group before idling. 717 */ 718 if (steal_htt && tdg->tdg_cpus > 1 && tdg->tdg_transferable) { 719 LIST_FOREACH(steal, &tdg->tdg_members, tdq_siblings) { 720 if (steal == tdq || steal->tdq_transferable == 0) 721 continue; 722 TDQ_LOCK(steal); 723 ts = tdq_steal(steal, 0); 724 if (ts) 725 goto steal; 726 TDQ_UNLOCK(steal); 727 } 728 } 729 for (;;) { 730 if (steal_idle == 0) 731 break; 732 highcpu = 0; 733 highload = 0; 734 for (cpu = 0; cpu <= mp_maxid; cpu++) { 735 if (CPU_ABSENT(cpu)) 736 continue; 737 steal = TDQ_CPU(cpu); 738 load = TDQ_CPU(cpu)->tdq_transferable; 739 if (load < highload) 740 continue; 741 highload = load; 742 highcpu = cpu; 743 } 744 if (highload < steal_thresh) 745 break; 746 steal = TDQ_CPU(highcpu); 747 TDQ_LOCK(steal); 748 if (steal->tdq_transferable >= steal_thresh && 749 (ts = tdq_steal(steal, 1)) != NULL) 750 goto steal; 751 TDQ_UNLOCK(steal); 752 break; 753 } 754 spinlock_exit(); 755 return (1); 756steal: 757 td = ts->ts_thread; 758 thread_lock(td); 759 spinlock_exit(); 760 MPASS(td->td_lock == TDQ_LOCKPTR(steal)); 761 TDQ_UNLOCK(steal); 762 sched_rem(td); 763 sched_setcpu(ts, PCPU_GET(cpuid), SRQ_YIELDING); 764 tdq_add(tdq, td, SRQ_YIELDING); 765 MPASS(td->td_lock == curthread->td_lock); 766 mi_switch(SW_VOL, NULL); 767 thread_unlock(curthread); 768 769 return (0); 770} 771 772/* 773 * Notify a remote cpu of new work. Sends an IPI if criteria are met. 774 */ 775static void 776tdq_notify(struct td_sched *ts) 777{ 778 struct thread *ctd; 779 struct pcpu *pcpu; 780 int cpri; 781 int pri; 782 int cpu; 783 784 cpu = ts->ts_cpu; 785 pri = ts->ts_thread->td_priority; 786 pcpu = pcpu_find(cpu); 787 ctd = pcpu->pc_curthread; 788 cpri = ctd->td_priority; 789 790 /* 791 * If our priority is not better than the current priority there is 792 * nothing to do. 793 */ 794 if (pri > cpri) 795 return; 796 /* 797 * Always IPI idle. 798 */ 799 if (cpri > PRI_MIN_IDLE) 800 goto sendipi; 801 /* 802 * If we're realtime or better and there is timeshare or worse running 803 * send an IPI. 804 */ 805 if (pri < PRI_MAX_REALTIME && cpri > PRI_MAX_REALTIME) 806 goto sendipi; 807 /* 808 * Otherwise only IPI if we exceed the threshold. 809 */ 810 if (pri > preempt_thresh) 811 return; 812sendipi: 813 ctd->td_flags |= TDF_NEEDRESCHED; 814 ipi_selected(1 << cpu, IPI_PREEMPT); 815} 816 817/* 818 * Steals load from a timeshare queue. Honors the rotating queue head 819 * index. 820 */ 821static struct td_sched * 822runq_steal_from(struct runq *rq, u_char start) 823{ 824 struct td_sched *ts; 825 struct rqbits *rqb; 826 struct rqhead *rqh; 827 int first; 828 int bit; 829 int pri; 830 int i; 831 832 rqb = &rq->rq_status; 833 bit = start & (RQB_BPW -1); 834 pri = 0; 835 first = 0; 836again: 837 for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) { 838 if (rqb->rqb_bits[i] == 0) 839 continue; 840 if (bit != 0) { 841 for (pri = bit; pri < RQB_BPW; pri++) 842 if (rqb->rqb_bits[i] & (1ul << pri)) 843 break; 844 if (pri >= RQB_BPW) 845 continue; 846 } else 847 pri = RQB_FFS(rqb->rqb_bits[i]); 848 pri += (i << RQB_L2BPW); 849 rqh = &rq->rq_queues[pri]; 850 TAILQ_FOREACH(ts, rqh, ts_procq) { 851 if (first && THREAD_CAN_MIGRATE(ts->ts_thread)) 852 return (ts); 853 first = 1; 854 } 855 } 856 if (start != 0) { 857 start = 0; 858 goto again; 859 } 860 861 return (NULL); 862} 863 864/* 865 * Steals load from a standard linear queue. 866 */ 867static struct td_sched * 868runq_steal(struct runq *rq) 869{ 870 struct rqhead *rqh; 871 struct rqbits *rqb; 872 struct td_sched *ts; 873 int word; 874 int bit; 875 876 rqb = &rq->rq_status; 877 for (word = 0; word < RQB_LEN; word++) { 878 if (rqb->rqb_bits[word] == 0) 879 continue; 880 for (bit = 0; bit < RQB_BPW; bit++) { 881 if ((rqb->rqb_bits[word] & (1ul << bit)) == 0) 882 continue; 883 rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)]; 884 TAILQ_FOREACH(ts, rqh, ts_procq) 885 if (THREAD_CAN_MIGRATE(ts->ts_thread)) 886 return (ts); 887 } 888 } 889 return (NULL); 890} 891 892/* 893 * Attempt to steal a thread in priority order from a thread queue. 894 */ 895static struct td_sched * 896tdq_steal(struct tdq *tdq, int stealidle) 897{ 898 struct td_sched *ts; 899 900 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 901 if ((ts = runq_steal(&tdq->tdq_realtime)) != NULL) 902 return (ts); 903 if ((ts = runq_steal_from(&tdq->tdq_timeshare, tdq->tdq_ridx)) != NULL) 904 return (ts); 905 if (stealidle) 906 return (runq_steal(&tdq->tdq_idle)); 907 return (NULL); 908} 909 910/* 911 * Sets the thread lock and ts_cpu to match the requested cpu. Unlocks the 912 * current lock and returns with the assigned queue locked. If this is 913 * via sched_switch() we leave the thread in a blocked state as an 914 * optimization. 915 */ 916static inline struct tdq * 917sched_setcpu(struct td_sched *ts, int cpu, int flags) 918{ 919 struct thread *td; 920 struct tdq *tdq; 921 922 THREAD_LOCK_ASSERT(ts->ts_thread, MA_OWNED); 923 924 tdq = TDQ_CPU(cpu); 925 td = ts->ts_thread; 926 ts->ts_cpu = cpu; 927 928 /* If the lock matches just return the queue. */ 929 if (td->td_lock == TDQ_LOCKPTR(tdq)) 930 return (tdq); 931#ifdef notyet 932 /* 933 * If the thread isn't running its lockptr is a 934 * turnstile or a sleepqueue. We can just lock_set without 935 * blocking. 936 */ 937 if (TD_CAN_RUN(td)) { 938 TDQ_LOCK(tdq); 939 thread_lock_set(td, TDQ_LOCKPTR(tdq)); 940 return (tdq); 941 } 942#endif 943 /* 944 * The hard case, migration, we need to block the thread first to 945 * prevent order reversals with other cpus locks. 946 */ 947 thread_lock_block(td); 948 TDQ_LOCK(tdq); 949 thread_lock_unblock(td, TDQ_LOCKPTR(tdq)); 950 return (tdq); 951} 952 953/* 954 * Find the thread queue running the lowest priority thread. 955 */ 956static int 957tdq_lowestpri(void) 958{ 959 struct tdq *tdq; 960 int lowpri; 961 int lowcpu; 962 int lowload; 963 int load; 964 int cpu; 965 int pri; 966 967 lowload = 0; 968 lowpri = lowcpu = 0; 969 for (cpu = 0; cpu <= mp_maxid; cpu++) { 970 if (CPU_ABSENT(cpu)) 971 continue; 972 tdq = TDQ_CPU(cpu); 973 pri = tdq->tdq_lowpri; 974 load = TDQ_CPU(cpu)->tdq_load; 975 CTR4(KTR_ULE, 976 "cpu %d pri %d lowcpu %d lowpri %d", 977 cpu, pri, lowcpu, lowpri); 978 if (pri < lowpri) 979 continue; 980 if (lowpri && lowpri == pri && load > lowload) 981 continue; 982 lowpri = pri; 983 lowcpu = cpu; 984 lowload = load; 985 } 986 987 return (lowcpu); 988} 989 990/* 991 * Find the thread queue with the least load. 992 */ 993static int 994tdq_lowestload(void) 995{ 996 struct tdq *tdq; 997 int lowload; 998 int lowpri; 999 int lowcpu; 1000 int load; 1001 int cpu; 1002 int pri; 1003 1004 lowcpu = 0; 1005 lowload = TDQ_CPU(0)->tdq_load; 1006 lowpri = TDQ_CPU(0)->tdq_lowpri; 1007 for (cpu = 1; cpu <= mp_maxid; cpu++) { 1008 if (CPU_ABSENT(cpu)) 1009 continue; 1010 tdq = TDQ_CPU(cpu); 1011 load = tdq->tdq_load; 1012 pri = tdq->tdq_lowpri; 1013 CTR4(KTR_ULE, "cpu %d load %d lowcpu %d lowload %d", 1014 cpu, load, lowcpu, lowload); 1015 if (load > lowload) 1016 continue; 1017 if (load == lowload && pri < lowpri) 1018 continue; 1019 lowcpu = cpu; 1020 lowload = load; 1021 lowpri = pri; 1022 } 1023 1024 return (lowcpu); 1025} 1026 1027/* 1028 * Pick the destination cpu for sched_add(). Respects affinity and makes 1029 * a determination based on load or priority of available processors. 1030 */ 1031static int 1032sched_pickcpu(struct td_sched *ts, int flags) 1033{ 1034 struct tdq *tdq; 1035 int self; 1036 int pri; 1037 int cpu; 1038 1039 cpu = self = PCPU_GET(cpuid); 1040 if (smp_started == 0) 1041 return (self); 1042 /* 1043 * Don't migrate a running thread from sched_switch(). 1044 */ 1045 if (flags & SRQ_OURSELF) { 1046 CTR1(KTR_ULE, "YIELDING %d", 1047 curthread->td_priority); 1048 return (self); 1049 } 1050 pri = ts->ts_thread->td_priority; 1051 cpu = ts->ts_cpu; 1052 /* 1053 * Regardless of affinity, if the last cpu is idle send it there. 1054 */ 1055 tdq = TDQ_CPU(cpu); 1056 if (tdq->tdq_lowpri > PRI_MIN_IDLE) { 1057 CTR5(KTR_ULE, 1058 "ts_cpu %d idle, ltick %d ticks %d pri %d curthread %d", 1059 ts->ts_cpu, ts->ts_rltick, ticks, pri, 1060 tdq->tdq_lowpri); 1061 return (ts->ts_cpu); 1062 } 1063 /* 1064 * If we have affinity, try to place it on the cpu we last ran on. 1065 */ 1066 if (SCHED_AFFINITY(ts) && tdq->tdq_lowpri > pri) { 1067 CTR5(KTR_ULE, 1068 "affinity for %d, ltick %d ticks %d pri %d curthread %d", 1069 ts->ts_cpu, ts->ts_rltick, ticks, pri, 1070 tdq->tdq_lowpri); 1071 return (ts->ts_cpu); 1072 } 1073 /* 1074 * Look for an idle group. 1075 */ 1076 CTR1(KTR_ULE, "tdq_idle %X", tdq_idle); 1077 cpu = ffs(tdq_idle); 1078 if (cpu) 1079 return (--cpu); 1080 /* 1081 * If there are no idle cores see if we can run the thread locally. This may 1082 * improve locality among sleepers and wakers when there is shared data. 1083 */ 1084 if (tryself && pri < curthread->td_priority) { 1085 CTR1(KTR_ULE, "tryself %d", 1086 curthread->td_priority); 1087 return (self); 1088 } 1089 /* 1090 * Now search for the cpu running the lowest priority thread with 1091 * the least load. 1092 */ 1093 if (pick_pri) 1094 cpu = tdq_lowestpri(); 1095 else 1096 cpu = tdq_lowestload(); 1097 return (cpu); 1098} 1099 1100#endif /* SMP */ 1101 1102/* 1103 * Pick the highest priority task we have and return it. 1104 */ 1105static struct td_sched * 1106tdq_choose(struct tdq *tdq) 1107{ 1108 struct td_sched *ts; 1109 1110 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 1111 ts = runq_choose(&tdq->tdq_realtime); 1112 if (ts != NULL) 1113 return (ts); 1114 ts = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx); 1115 if (ts != NULL) { 1116 KASSERT(ts->ts_thread->td_priority >= PRI_MIN_TIMESHARE, 1117 ("tdq_choose: Invalid priority on timeshare queue %d", 1118 ts->ts_thread->td_priority)); 1119 return (ts); 1120 } 1121 1122 ts = runq_choose(&tdq->tdq_idle); 1123 if (ts != NULL) { 1124 KASSERT(ts->ts_thread->td_priority >= PRI_MIN_IDLE, 1125 ("tdq_choose: Invalid priority on idle queue %d", 1126 ts->ts_thread->td_priority)); 1127 return (ts); 1128 } 1129 1130 return (NULL); 1131} 1132 1133/* 1134 * Initialize a thread queue. 1135 */ 1136static void 1137tdq_setup(struct tdq *tdq) 1138{ 1139 1140 if (bootverbose) 1141 printf("ULE: setup cpu %d\n", TDQ_ID(tdq)); 1142 runq_init(&tdq->tdq_realtime); 1143 runq_init(&tdq->tdq_timeshare); 1144 runq_init(&tdq->tdq_idle); 1145 tdq->tdq_load = 0; 1146} 1147 1148#ifdef SMP 1149static void 1150tdg_setup(struct tdq_group *tdg) 1151{ 1152 if (bootverbose) 1153 printf("ULE: setup cpu group %d\n", TDG_ID(tdg)); 1154 snprintf(tdg->tdg_name, sizeof(tdg->tdg_name), 1155 "sched lock %d", (int)TDG_ID(tdg)); 1156 mtx_init(&tdg->tdg_lock, tdg->tdg_name, "sched lock", 1157 MTX_SPIN | MTX_RECURSE); 1158 LIST_INIT(&tdg->tdg_members); 1159 tdg->tdg_load = 0; 1160 tdg->tdg_transferable = 0; 1161 tdg->tdg_cpus = 0; 1162 tdg->tdg_mask = 0; 1163 tdg->tdg_cpumask = 0; 1164 tdg->tdg_idlemask = 0; 1165} 1166 1167static void 1168tdg_add(struct tdq_group *tdg, struct tdq *tdq) 1169{ 1170 if (tdg->tdg_mask == 0) 1171 tdg->tdg_mask |= 1 << TDQ_ID(tdq); 1172 tdg->tdg_cpumask |= 1 << TDQ_ID(tdq); 1173 tdg->tdg_cpus++; 1174 tdq->tdq_group = tdg; 1175 tdq->tdq_lock = &tdg->tdg_lock; 1176 LIST_INSERT_HEAD(&tdg->tdg_members, tdq, tdq_siblings); 1177 if (bootverbose) 1178 printf("ULE: adding cpu %d to group %d: cpus %d mask 0x%X\n", 1179 TDQ_ID(tdq), TDG_ID(tdg), tdg->tdg_cpus, tdg->tdg_cpumask); 1180} 1181 1182static void 1183sched_setup_topology(void) 1184{ 1185 struct tdq_group *tdg; 1186 struct cpu_group *cg; 1187 int balance_groups; 1188 struct tdq *tdq; 1189 int i; 1190 int j; 1191 1192 topology = 1; 1193 balance_groups = 0; 1194 for (i = 0; i < smp_topology->ct_count; i++) { 1195 cg = &smp_topology->ct_group[i]; 1196 tdg = &tdq_groups[i]; 1197 /* 1198 * Initialize the group. 1199 */ 1200 tdg_setup(tdg); 1201 /* 1202 * Find all of the group members and add them. 1203 */ 1204 for (j = 0; j < MAXCPU; j++) { 1205 if ((cg->cg_mask & (1 << j)) != 0) { 1206 tdq = TDQ_CPU(j); 1207 tdq_setup(tdq); 1208 tdg_add(tdg, tdq); 1209 } 1210 } 1211 if (tdg->tdg_cpus > 1) 1212 balance_groups = 1; 1213 } 1214 tdg_maxid = smp_topology->ct_count - 1; 1215 if (balance_groups) 1216 sched_balance_groups(NULL); 1217} 1218 1219static void 1220sched_setup_smp(void) 1221{ 1222 struct tdq_group *tdg; 1223 struct tdq *tdq; 1224 int cpus; 1225 int i; 1226 1227 for (cpus = 0, i = 0; i < MAXCPU; i++) { 1228 if (CPU_ABSENT(i)) 1229 continue; 1230 tdq = &tdq_cpu[i]; 1231 tdg = &tdq_groups[i]; 1232 /* 1233 * Setup a tdq group with one member. 1234 */ 1235 tdg_setup(tdg); 1236 tdq_setup(tdq); 1237 tdg_add(tdg, tdq); 1238 cpus++; 1239 } 1240 tdg_maxid = cpus - 1; 1241} 1242 1243/* 1244 * Fake a topology with one group containing all CPUs. 1245 */ 1246static void 1247sched_fake_topo(void) 1248{ 1249#ifdef SCHED_FAKE_TOPOLOGY 1250 static struct cpu_top top; 1251 static struct cpu_group group; 1252 1253 top.ct_count = 1; 1254 top.ct_group = &group; 1255 group.cg_mask = all_cpus; 1256 group.cg_count = mp_ncpus; 1257 group.cg_children = 0; 1258 smp_topology = ⊤ 1259#endif 1260} 1261#endif 1262 1263/* 1264 * Setup the thread queues and initialize the topology based on MD 1265 * information. 1266 */ 1267static void 1268sched_setup(void *dummy) 1269{ 1270 struct tdq *tdq; 1271 1272 tdq = TDQ_SELF(); 1273#ifdef SMP 1274 /* 1275 * Initialize long-term cpu balancing algorithm. 1276 */ 1277 callout_init(&balco, CALLOUT_MPSAFE); 1278 callout_init(&gbalco, CALLOUT_MPSAFE); 1279 sched_fake_topo(); 1280 /* 1281 * Setup tdqs based on a topology configuration or vanilla SMP based 1282 * on mp_maxid. 1283 */ 1284 if (smp_topology == NULL) 1285 sched_setup_smp(); 1286 else 1287 sched_setup_topology(); 1288 sched_balance(NULL); 1289#else 1290 tdq_setup(tdq); 1291 mtx_init(&tdq_lock, "sched lock", "sched lock", MTX_SPIN | MTX_RECURSE); 1292 tdq->tdq_lock = &tdq_lock; 1293#endif 1294 /* 1295 * To avoid divide-by-zero, we set realstathz a dummy value 1296 * in case which sched_clock() called before sched_initticks(). 1297 */ 1298 realstathz = hz; 1299 sched_slice = (realstathz/10); /* ~100ms */ 1300 tickincr = 1 << SCHED_TICK_SHIFT; 1301 1302 /* Add thread0's load since it's running. */ 1303 TDQ_LOCK(tdq); 1304 thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF()); 1305 tdq_load_add(tdq, &td_sched0); 1306 TDQ_UNLOCK(tdq); 1307} 1308 1309/* 1310 * This routine determines the tickincr after stathz and hz are setup. 1311 */ 1312/* ARGSUSED */ 1313static void 1314sched_initticks(void *dummy) 1315{ 1316 int incr; 1317 1318 realstathz = stathz ? stathz : hz; 1319 sched_slice = (realstathz/10); /* ~100ms */ 1320 1321 /* 1322 * tickincr is shifted out by 10 to avoid rounding errors due to 1323 * hz not being evenly divisible by stathz on all platforms. 1324 */ 1325 incr = (hz << SCHED_TICK_SHIFT) / realstathz; 1326 /* 1327 * This does not work for values of stathz that are more than 1328 * 1 << SCHED_TICK_SHIFT * hz. In practice this does not happen. 1329 */ 1330 if (incr == 0) 1331 incr = 1; 1332 tickincr = incr; 1333#ifdef SMP 1334 /* 1335 * Set steal thresh to log2(mp_ncpu) but no greater than 4. This 1336 * prevents excess thrashing on large machines and excess idle on 1337 * smaller machines. 1338 */ 1339 steal_thresh = min(ffs(mp_ncpus) - 1, 4); 1340 affinity = SCHED_AFFINITY_DEFAULT; 1341#endif 1342} 1343 1344 1345/* 1346 * This is the core of the interactivity algorithm. Determines a score based 1347 * on past behavior. It is the ratio of sleep time to run time scaled to 1348 * a [0, 100] integer. This is the voluntary sleep time of a process, which 1349 * differs from the cpu usage because it does not account for time spent 1350 * waiting on a run-queue. Would be prettier if we had floating point. 1351 */ 1352static int 1353sched_interact_score(struct thread *td) 1354{ 1355 struct td_sched *ts; 1356 int div; 1357 1358 ts = td->td_sched; 1359 /* 1360 * The score is only needed if this is likely to be an interactive 1361 * task. Don't go through the expense of computing it if there's 1362 * no chance. 1363 */ 1364 if (sched_interact <= SCHED_INTERACT_HALF && 1365 ts->ts_runtime >= ts->ts_slptime) 1366 return (SCHED_INTERACT_HALF); 1367 1368 if (ts->ts_runtime > ts->ts_slptime) { 1369 div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF); 1370 return (SCHED_INTERACT_HALF + 1371 (SCHED_INTERACT_HALF - (ts->ts_slptime / div))); 1372 } 1373 if (ts->ts_slptime > ts->ts_runtime) { 1374 div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF); 1375 return (ts->ts_runtime / div); 1376 } 1377 /* runtime == slptime */ 1378 if (ts->ts_runtime) 1379 return (SCHED_INTERACT_HALF); 1380 1381 /* 1382 * This can happen if slptime and runtime are 0. 1383 */ 1384 return (0); 1385 1386} 1387 1388/* 1389 * Scale the scheduling priority according to the "interactivity" of this 1390 * process. 1391 */ 1392static void 1393sched_priority(struct thread *td) 1394{ 1395 int score; 1396 int pri; 1397 1398 if (td->td_pri_class != PRI_TIMESHARE) 1399 return; 1400 /* 1401 * If the score is interactive we place the thread in the realtime 1402 * queue with a priority that is less than kernel and interrupt 1403 * priorities. These threads are not subject to nice restrictions. 1404 * 1405 * Scores greater than this are placed on the normal timeshare queue 1406 * where the priority is partially decided by the most recent cpu 1407 * utilization and the rest is decided by nice value. 1408 * 1409 * The nice value of the process has a linear effect on the calculated 1410 * score. Negative nice values make it easier for a thread to be 1411 * considered interactive. 1412 */ 1413 score = imax(0, sched_interact_score(td) - td->td_proc->p_nice); 1414 if (score < sched_interact) { 1415 pri = PRI_MIN_REALTIME; 1416 pri += ((PRI_MAX_REALTIME - PRI_MIN_REALTIME) / sched_interact) 1417 * score; 1418 KASSERT(pri >= PRI_MIN_REALTIME && pri <= PRI_MAX_REALTIME, 1419 ("sched_priority: invalid interactive priority %d score %d", 1420 pri, score)); 1421 } else { 1422 pri = SCHED_PRI_MIN; 1423 if (td->td_sched->ts_ticks) 1424 pri += SCHED_PRI_TICKS(td->td_sched); 1425 pri += SCHED_PRI_NICE(td->td_proc->p_nice); 1426 KASSERT(pri >= PRI_MIN_TIMESHARE && pri <= PRI_MAX_TIMESHARE, 1427 ("sched_priority: invalid priority %d: nice %d, " 1428 "ticks %d ftick %d ltick %d tick pri %d", 1429 pri, td->td_proc->p_nice, td->td_sched->ts_ticks, 1430 td->td_sched->ts_ftick, td->td_sched->ts_ltick, 1431 SCHED_PRI_TICKS(td->td_sched))); 1432 } 1433 sched_user_prio(td, pri); 1434 1435 return; 1436} 1437 1438/* 1439 * This routine enforces a maximum limit on the amount of scheduling history 1440 * kept. It is called after either the slptime or runtime is adjusted. This 1441 * function is ugly due to integer math. 1442 */ 1443static void 1444sched_interact_update(struct thread *td) 1445{ 1446 struct td_sched *ts; 1447 u_int sum; 1448 1449 ts = td->td_sched; 1450 sum = ts->ts_runtime + ts->ts_slptime; 1451 if (sum < SCHED_SLP_RUN_MAX) 1452 return; 1453 /* 1454 * This only happens from two places: 1455 * 1) We have added an unusual amount of run time from fork_exit. 1456 * 2) We have added an unusual amount of sleep time from sched_sleep(). 1457 */ 1458 if (sum > SCHED_SLP_RUN_MAX * 2) { 1459 if (ts->ts_runtime > ts->ts_slptime) { 1460 ts->ts_runtime = SCHED_SLP_RUN_MAX; 1461 ts->ts_slptime = 1; 1462 } else { 1463 ts->ts_slptime = SCHED_SLP_RUN_MAX; 1464 ts->ts_runtime = 1; 1465 } 1466 return; 1467 } 1468 /* 1469 * If we have exceeded by more than 1/5th then the algorithm below 1470 * will not bring us back into range. Dividing by two here forces 1471 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX] 1472 */ 1473 if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) { 1474 ts->ts_runtime /= 2; 1475 ts->ts_slptime /= 2; 1476 return; 1477 } 1478 ts->ts_runtime = (ts->ts_runtime / 5) * 4; 1479 ts->ts_slptime = (ts->ts_slptime / 5) * 4; 1480} 1481 1482/* 1483 * Scale back the interactivity history when a child thread is created. The 1484 * history is inherited from the parent but the thread may behave totally 1485 * differently. For example, a shell spawning a compiler process. We want 1486 * to learn that the compiler is behaving badly very quickly. 1487 */ 1488static void 1489sched_interact_fork(struct thread *td) 1490{ 1491 int ratio; 1492 int sum; 1493 1494 sum = td->td_sched->ts_runtime + td->td_sched->ts_slptime; 1495 if (sum > SCHED_SLP_RUN_FORK) { 1496 ratio = sum / SCHED_SLP_RUN_FORK; 1497 td->td_sched->ts_runtime /= ratio; 1498 td->td_sched->ts_slptime /= ratio; 1499 } 1500} 1501 1502/* 1503 * Called from proc0_init() to setup the scheduler fields. 1504 */ 1505void 1506schedinit(void) 1507{ 1508 1509 /* 1510 * Set up the scheduler specific parts of proc0. 1511 */ 1512 proc0.p_sched = NULL; /* XXX */ 1513 thread0.td_sched = &td_sched0; 1514 td_sched0.ts_ltick = ticks; 1515 td_sched0.ts_ftick = ticks; 1516 td_sched0.ts_thread = &thread0; 1517} 1518 1519/* 1520 * This is only somewhat accurate since given many processes of the same 1521 * priority they will switch when their slices run out, which will be 1522 * at most sched_slice stathz ticks. 1523 */ 1524int 1525sched_rr_interval(void) 1526{ 1527 1528 /* Convert sched_slice to hz */ 1529 return (hz/(realstathz/sched_slice)); 1530} 1531 1532/* 1533 * Update the percent cpu tracking information when it is requested or 1534 * the total history exceeds the maximum. We keep a sliding history of 1535 * tick counts that slowly decays. This is less precise than the 4BSD 1536 * mechanism since it happens with less regular and frequent events. 1537 */ 1538static void 1539sched_pctcpu_update(struct td_sched *ts) 1540{ 1541 1542 if (ts->ts_ticks == 0) 1543 return; 1544 if (ticks - (hz / 10) < ts->ts_ltick && 1545 SCHED_TICK_TOTAL(ts) < SCHED_TICK_MAX) 1546 return; 1547 /* 1548 * Adjust counters and watermark for pctcpu calc. 1549 */ 1550 if (ts->ts_ltick > ticks - SCHED_TICK_TARG) 1551 ts->ts_ticks = (ts->ts_ticks / (ticks - ts->ts_ftick)) * 1552 SCHED_TICK_TARG; 1553 else 1554 ts->ts_ticks = 0; 1555 ts->ts_ltick = ticks; 1556 ts->ts_ftick = ts->ts_ltick - SCHED_TICK_TARG; 1557} 1558 1559/* 1560 * Adjust the priority of a thread. Move it to the appropriate run-queue 1561 * if necessary. This is the back-end for several priority related 1562 * functions. 1563 */ 1564static void 1565sched_thread_priority(struct thread *td, u_char prio) 1566{ 1567 struct td_sched *ts; 1568 1569 CTR6(KTR_SCHED, "sched_prio: %p(%s) prio %d newprio %d by %p(%s)", 1570 td, td->td_proc->p_comm, td->td_priority, prio, curthread, 1571 curthread->td_proc->p_comm); 1572 ts = td->td_sched; 1573 THREAD_LOCK_ASSERT(td, MA_OWNED); 1574 if (td->td_priority == prio) 1575 return; 1576 1577 if (TD_ON_RUNQ(td) && prio < td->td_priority) { 1578 /* 1579 * If the priority has been elevated due to priority 1580 * propagation, we may have to move ourselves to a new 1581 * queue. This could be optimized to not re-add in some 1582 * cases. 1583 */ 1584 sched_rem(td); 1585 td->td_priority = prio; 1586 sched_add(td, SRQ_BORROWING); 1587 } else { 1588#ifdef SMP 1589 struct tdq *tdq; 1590 1591 tdq = TDQ_CPU(ts->ts_cpu); 1592 if (prio < tdq->tdq_lowpri) 1593 tdq->tdq_lowpri = prio; 1594#endif 1595 td->td_priority = prio; 1596 } 1597} 1598 1599/* 1600 * Update a thread's priority when it is lent another thread's 1601 * priority. 1602 */ 1603void 1604sched_lend_prio(struct thread *td, u_char prio) 1605{ 1606 1607 td->td_flags |= TDF_BORROWING; 1608 sched_thread_priority(td, prio); 1609} 1610 1611/* 1612 * Restore a thread's priority when priority propagation is 1613 * over. The prio argument is the minimum priority the thread 1614 * needs to have to satisfy other possible priority lending 1615 * requests. If the thread's regular priority is less 1616 * important than prio, the thread will keep a priority boost 1617 * of prio. 1618 */ 1619void 1620sched_unlend_prio(struct thread *td, u_char prio) 1621{ 1622 u_char base_pri; 1623 1624 if (td->td_base_pri >= PRI_MIN_TIMESHARE && 1625 td->td_base_pri <= PRI_MAX_TIMESHARE) 1626 base_pri = td->td_user_pri; 1627 else 1628 base_pri = td->td_base_pri; 1629 if (prio >= base_pri) { 1630 td->td_flags &= ~TDF_BORROWING; 1631 sched_thread_priority(td, base_pri); 1632 } else 1633 sched_lend_prio(td, prio); 1634} 1635 1636/* 1637 * Standard entry for setting the priority to an absolute value. 1638 */ 1639void 1640sched_prio(struct thread *td, u_char prio) 1641{ 1642 u_char oldprio; 1643 1644 /* First, update the base priority. */ 1645 td->td_base_pri = prio; 1646 1647 /* 1648 * If the thread is borrowing another thread's priority, don't 1649 * ever lower the priority. 1650 */ 1651 if (td->td_flags & TDF_BORROWING && td->td_priority < prio) 1652 return; 1653 1654 /* Change the real priority. */ 1655 oldprio = td->td_priority; 1656 sched_thread_priority(td, prio); 1657 1658 /* 1659 * If the thread is on a turnstile, then let the turnstile update 1660 * its state. 1661 */ 1662 if (TD_ON_LOCK(td) && oldprio != prio) 1663 turnstile_adjust(td, oldprio); 1664} 1665 1666/* 1667 * Set the base user priority, does not effect current running priority. 1668 */ 1669void 1670sched_user_prio(struct thread *td, u_char prio) 1671{ 1672 u_char oldprio; 1673 1674 td->td_base_user_pri = prio; 1675 if (td->td_flags & TDF_UBORROWING && td->td_user_pri <= prio) 1676 return; 1677 oldprio = td->td_user_pri; 1678 td->td_user_pri = prio; 1679 1680 if (TD_ON_UPILOCK(td) && oldprio != prio) 1681 umtx_pi_adjust(td, oldprio); 1682} 1683 1684void 1685sched_lend_user_prio(struct thread *td, u_char prio) 1686{ 1687 u_char oldprio; 1688 1689 td->td_flags |= TDF_UBORROWING; 1690 1691 oldprio = td->td_user_pri; 1692 td->td_user_pri = prio; 1693 1694 if (TD_ON_UPILOCK(td) && oldprio != prio) 1695 umtx_pi_adjust(td, oldprio); 1696} 1697 1698void 1699sched_unlend_user_prio(struct thread *td, u_char prio) 1700{ 1701 u_char base_pri; 1702 1703 base_pri = td->td_base_user_pri; 1704 if (prio >= base_pri) { 1705 td->td_flags &= ~TDF_UBORROWING; 1706 sched_user_prio(td, base_pri); 1707 } else 1708 sched_lend_user_prio(td, prio); 1709} 1710 1711/* 1712 * Add the thread passed as 'newtd' to the run queue before selecting 1713 * the next thread to run. This is only used for KSE. 1714 */ 1715static void 1716sched_switchin(struct tdq *tdq, struct thread *td) 1717{ 1718#ifdef SMP 1719 spinlock_enter(); 1720 TDQ_UNLOCK(tdq); 1721 thread_lock(td); 1722 spinlock_exit(); 1723 sched_setcpu(td->td_sched, TDQ_ID(tdq), SRQ_YIELDING); 1724#else 1725 td->td_lock = TDQ_LOCKPTR(tdq); 1726#endif 1727 tdq_add(tdq, td, SRQ_YIELDING); 1728 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1729} 1730 1731/* 1732 * Handle migration from sched_switch(). This happens only for 1733 * cpu binding. 1734 */ 1735static struct mtx * 1736sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags) 1737{ 1738 struct tdq *tdn; 1739 1740 tdn = TDQ_CPU(td->td_sched->ts_cpu); 1741#ifdef SMP 1742 /* 1743 * Do the lock dance required to avoid LOR. We grab an extra 1744 * spinlock nesting to prevent preemption while we're 1745 * not holding either run-queue lock. 1746 */ 1747 spinlock_enter(); 1748 thread_block_switch(td); /* This releases the lock on tdq. */ 1749 TDQ_LOCK(tdn); 1750 tdq_add(tdn, td, flags); 1751 tdq_notify(td->td_sched); 1752 /* 1753 * After we unlock tdn the new cpu still can't switch into this 1754 * thread until we've unblocked it in cpu_switch(). The lock 1755 * pointers may match in the case of HTT cores. Don't unlock here 1756 * or we can deadlock when the other CPU runs the IPI handler. 1757 */ 1758 if (TDQ_LOCKPTR(tdn) != TDQ_LOCKPTR(tdq)) { 1759 TDQ_UNLOCK(tdn); 1760 TDQ_LOCK(tdq); 1761 } 1762 spinlock_exit(); 1763#endif 1764 return (TDQ_LOCKPTR(tdn)); 1765} 1766 1767/* 1768 * Block a thread for switching. Similar to thread_block() but does not 1769 * bump the spin count. 1770 */ 1771static inline struct mtx * 1772thread_block_switch(struct thread *td) 1773{ 1774 struct mtx *lock; 1775 1776 THREAD_LOCK_ASSERT(td, MA_OWNED); 1777 lock = td->td_lock; 1778 td->td_lock = &blocked_lock; 1779 mtx_unlock_spin(lock); 1780 1781 return (lock); 1782} 1783 1784/* 1785 * Release a thread that was blocked with thread_block_switch(). 1786 */ 1787static inline void 1788thread_unblock_switch(struct thread *td, struct mtx *mtx) 1789{ 1790 atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock, 1791 (uintptr_t)mtx); 1792} 1793 1794/* 1795 * Switch threads. This function has to handle threads coming in while 1796 * blocked for some reason, running, or idle. It also must deal with 1797 * migrating a thread from one queue to another as running threads may 1798 * be assigned elsewhere via binding. 1799 */ 1800void 1801sched_switch(struct thread *td, struct thread *newtd, int flags) 1802{ 1803 struct tdq *tdq; 1804 struct td_sched *ts; 1805 struct mtx *mtx; 1806 int srqflag; 1807 int cpuid; 1808 1809 THREAD_LOCK_ASSERT(td, MA_OWNED); 1810 1811 cpuid = PCPU_GET(cpuid); 1812 tdq = TDQ_CPU(cpuid); 1813 ts = td->td_sched; 1814 mtx = td->td_lock; 1815#ifdef SMP 1816 ts->ts_rltick = ticks; 1817 if (newtd && newtd->td_priority < tdq->tdq_lowpri) 1818 tdq->tdq_lowpri = newtd->td_priority; 1819#endif 1820 td->td_lastcpu = td->td_oncpu; 1821 td->td_oncpu = NOCPU; 1822 td->td_flags &= ~TDF_NEEDRESCHED; 1823 td->td_owepreempt = 0; 1824 /* 1825 * The lock pointer in an idle thread should never change. Reset it 1826 * to CAN_RUN as well. 1827 */ 1828 if (TD_IS_IDLETHREAD(td)) { 1829 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1830 TD_SET_CAN_RUN(td); 1831 } else if (TD_IS_RUNNING(td)) { 1832 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1833 tdq_load_rem(tdq, ts); 1834 srqflag = (flags & SW_PREEMPT) ? 1835 SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED : 1836 SRQ_OURSELF|SRQ_YIELDING; 1837 if (ts->ts_cpu == cpuid) 1838 tdq_add(tdq, td, srqflag); 1839 else 1840 mtx = sched_switch_migrate(tdq, td, srqflag); 1841 } else { 1842 /* This thread must be going to sleep. */ 1843 TDQ_LOCK(tdq); 1844 mtx = thread_block_switch(td); 1845 tdq_load_rem(tdq, ts); 1846 } 1847 /* 1848 * We enter here with the thread blocked and assigned to the 1849 * appropriate cpu run-queue or sleep-queue and with the current 1850 * thread-queue locked. 1851 */ 1852 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED); 1853 /* 1854 * If KSE assigned a new thread just add it here and let choosethread 1855 * select the best one. 1856 */ 1857 if (newtd != NULL) 1858 sched_switchin(tdq, newtd); 1859 newtd = choosethread(); 1860 /* 1861 * Call the MD code to switch contexts if necessary. 1862 */ 1863 if (td != newtd) { 1864#ifdef HWPMC_HOOKS 1865 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1866 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT); 1867#endif 1868 cpu_switch(td, newtd, mtx); 1869 /* 1870 * We may return from cpu_switch on a different cpu. However, 1871 * we always return with td_lock pointing to the current cpu's 1872 * run queue lock. 1873 */ 1874 cpuid = PCPU_GET(cpuid); 1875 tdq = TDQ_CPU(cpuid); 1876 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)td; 1877#ifdef HWPMC_HOOKS 1878 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1879 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN); 1880#endif 1881 } else 1882 thread_unblock_switch(td, mtx); 1883 /* 1884 * Assert that all went well and return. 1885 */ 1886#ifdef SMP 1887 /* We should always get here with the lowest priority td possible */ 1888 tdq->tdq_lowpri = td->td_priority; 1889#endif 1890 TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED); 1891 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1892 td->td_oncpu = cpuid; 1893} 1894 1895/* 1896 * Adjust thread priorities as a result of a nice request. 1897 */ 1898void 1899sched_nice(struct proc *p, int nice) 1900{ 1901 struct thread *td; 1902 1903 PROC_LOCK_ASSERT(p, MA_OWNED); 1904 PROC_SLOCK_ASSERT(p, MA_OWNED); 1905 1906 p->p_nice = nice; 1907 FOREACH_THREAD_IN_PROC(p, td) { 1908 thread_lock(td); 1909 sched_priority(td); 1910 sched_prio(td, td->td_base_user_pri); 1911 thread_unlock(td); 1912 } 1913} 1914 1915/* 1916 * Record the sleep time for the interactivity scorer. 1917 */ 1918void 1919sched_sleep(struct thread *td) 1920{ 1921 1922 THREAD_LOCK_ASSERT(td, MA_OWNED); 1923 1924 td->td_slptick = ticks; 1925} 1926 1927/* 1928 * Schedule a thread to resume execution and record how long it voluntarily 1929 * slept. We also update the pctcpu, interactivity, and priority. 1930 */ 1931void 1932sched_wakeup(struct thread *td) 1933{ 1934 struct td_sched *ts; 1935 int slptick; 1936 1937 THREAD_LOCK_ASSERT(td, MA_OWNED); 1938 ts = td->td_sched; 1939 /* 1940 * If we slept for more than a tick update our interactivity and 1941 * priority. 1942 */ 1943 slptick = td->td_slptick; 1944 td->td_slptick = 0; 1945 if (slptick && slptick != ticks) { 1946 u_int hzticks; 1947 1948 hzticks = (ticks - slptick) << SCHED_TICK_SHIFT; 1949 ts->ts_slptime += hzticks; 1950 sched_interact_update(td); 1951 sched_pctcpu_update(ts); 1952 sched_priority(td); 1953 } 1954 /* Reset the slice value after we sleep. */ 1955 ts->ts_slice = sched_slice; 1956 sched_add(td, SRQ_BORING); 1957} 1958 1959/* 1960 * Penalize the parent for creating a new child and initialize the child's 1961 * priority. 1962 */ 1963void 1964sched_fork(struct thread *td, struct thread *child) 1965{ 1966 THREAD_LOCK_ASSERT(td, MA_OWNED); 1967 sched_fork_thread(td, child); 1968 /* 1969 * Penalize the parent and child for forking. 1970 */ 1971 sched_interact_fork(child); 1972 sched_priority(child); 1973 td->td_sched->ts_runtime += tickincr; 1974 sched_interact_update(td); 1975 sched_priority(td); 1976} 1977 1978/* 1979 * Fork a new thread, may be within the same process. 1980 */ 1981void 1982sched_fork_thread(struct thread *td, struct thread *child) 1983{ 1984 struct td_sched *ts; 1985 struct td_sched *ts2; 1986 1987 /* 1988 * Initialize child. 1989 */ 1990 THREAD_LOCK_ASSERT(td, MA_OWNED); 1991 sched_newthread(child); 1992 child->td_lock = TDQ_LOCKPTR(TDQ_SELF()); 1993 ts = td->td_sched; 1994 ts2 = child->td_sched; 1995 ts2->ts_cpu = ts->ts_cpu; 1996 ts2->ts_runq = NULL; 1997 /* 1998 * Grab our parents cpu estimation information and priority. 1999 */ 2000 ts2->ts_ticks = ts->ts_ticks; 2001 ts2->ts_ltick = ts->ts_ltick; 2002 ts2->ts_ftick = ts->ts_ftick; 2003 child->td_user_pri = td->td_user_pri; 2004 child->td_base_user_pri = td->td_base_user_pri; 2005 /* 2006 * And update interactivity score. 2007 */ 2008 ts2->ts_slptime = ts->ts_slptime; 2009 ts2->ts_runtime = ts->ts_runtime; 2010 ts2->ts_slice = 1; /* Attempt to quickly learn interactivity. */ 2011} 2012 2013/* 2014 * Adjust the priority class of a thread. 2015 */ 2016void 2017sched_class(struct thread *td, int class) 2018{ 2019 2020 THREAD_LOCK_ASSERT(td, MA_OWNED); 2021 if (td->td_pri_class == class) 2022 return; 2023 2024#ifdef SMP 2025 /* 2026 * On SMP if we're on the RUNQ we must adjust the transferable 2027 * count because could be changing to or from an interrupt 2028 * class. 2029 */ 2030 if (TD_ON_RUNQ(td)) { 2031 struct tdq *tdq; 2032 2033 tdq = TDQ_CPU(td->td_sched->ts_cpu); 2034 if (THREAD_CAN_MIGRATE(td)) { 2035 tdq->tdq_transferable--; 2036 tdq->tdq_group->tdg_transferable--; 2037 } 2038 td->td_pri_class = class; 2039 if (THREAD_CAN_MIGRATE(td)) { 2040 tdq->tdq_transferable++; 2041 tdq->tdq_group->tdg_transferable++; 2042 } 2043 } 2044#endif 2045 td->td_pri_class = class; 2046} 2047 2048/* 2049 * Return some of the child's priority and interactivity to the parent. 2050 */ 2051void 2052sched_exit(struct proc *p, struct thread *child) 2053{ 2054 struct thread *td; 2055 2056 CTR3(KTR_SCHED, "sched_exit: %p(%s) prio %d", 2057 child, child->td_proc->p_comm, child->td_priority); 2058 2059 PROC_SLOCK_ASSERT(p, MA_OWNED); 2060 td = FIRST_THREAD_IN_PROC(p); 2061 sched_exit_thread(td, child); 2062} 2063 2064/* 2065 * Penalize another thread for the time spent on this one. This helps to 2066 * worsen the priority and interactivity of processes which schedule batch 2067 * jobs such as make. This has little effect on the make process itself but 2068 * causes new processes spawned by it to receive worse scores immediately. 2069 */ 2070void 2071sched_exit_thread(struct thread *td, struct thread *child) 2072{ 2073 2074 CTR3(KTR_SCHED, "sched_exit_thread: %p(%s) prio %d", 2075 child, child->td_proc->p_comm, child->td_priority); 2076 2077#ifdef KSE 2078 /* 2079 * KSE forks and exits so often that this penalty causes short-lived 2080 * threads to always be non-interactive. This causes mozilla to 2081 * crawl under load. 2082 */ 2083 if ((td->td_pflags & TDP_SA) && td->td_proc == child->td_proc) 2084 return; 2085#endif 2086 /* 2087 * Give the child's runtime to the parent without returning the 2088 * sleep time as a penalty to the parent. This causes shells that 2089 * launch expensive things to mark their children as expensive. 2090 */ 2091 thread_lock(td); 2092 td->td_sched->ts_runtime += child->td_sched->ts_runtime; 2093 sched_interact_update(td); 2094 sched_priority(td); 2095 thread_unlock(td); 2096} 2097 2098/* 2099 * Fix priorities on return to user-space. Priorities may be elevated due 2100 * to static priorities in msleep() or similar. 2101 */ 2102void 2103sched_userret(struct thread *td) 2104{ 2105 /* 2106 * XXX we cheat slightly on the locking here to avoid locking in 2107 * the usual case. Setting td_priority here is essentially an 2108 * incomplete workaround for not setting it properly elsewhere. 2109 * Now that some interrupt handlers are threads, not setting it 2110 * properly elsewhere can clobber it in the window between setting 2111 * it here and returning to user mode, so don't waste time setting 2112 * it perfectly here. 2113 */ 2114 KASSERT((td->td_flags & TDF_BORROWING) == 0, 2115 ("thread with borrowed priority returning to userland")); 2116 if (td->td_priority != td->td_user_pri) { 2117 thread_lock(td); 2118 td->td_priority = td->td_user_pri; 2119 td->td_base_pri = td->td_user_pri; 2120 thread_unlock(td); 2121 } 2122} 2123 2124/* 2125 * Handle a stathz tick. This is really only relevant for timeshare 2126 * threads. 2127 */ 2128void 2129sched_clock(struct thread *td) 2130{ 2131 struct tdq *tdq; 2132 struct td_sched *ts; 2133 2134 THREAD_LOCK_ASSERT(td, MA_OWNED); 2135 tdq = TDQ_SELF(); 2136 /* 2137 * Advance the insert index once for each tick to ensure that all 2138 * threads get a chance to run. 2139 */ 2140 if (tdq->tdq_idx == tdq->tdq_ridx) { 2141 tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS; 2142 if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx])) 2143 tdq->tdq_ridx = tdq->tdq_idx; 2144 } 2145 ts = td->td_sched; 2146 /* 2147 * We only do slicing code for TIMESHARE threads. 2148 */ 2149 if (td->td_pri_class != PRI_TIMESHARE) 2150 return; 2151 /* 2152 * We used a tick; charge it to the thread so that we can compute our 2153 * interactivity. 2154 */ 2155 td->td_sched->ts_runtime += tickincr; 2156 sched_interact_update(td); 2157 /* 2158 * We used up one time slice. 2159 */ 2160 if (--ts->ts_slice > 0) 2161 return; 2162 /* 2163 * We're out of time, recompute priorities and requeue. 2164 */ 2165 sched_priority(td); 2166 td->td_flags |= TDF_NEEDRESCHED; 2167} 2168 2169/* 2170 * Called once per hz tick. Used for cpu utilization information. This 2171 * is easier than trying to scale based on stathz. 2172 */ 2173void 2174sched_tick(void) 2175{ 2176 struct td_sched *ts; 2177 2178 ts = curthread->td_sched; 2179 /* Adjust ticks for pctcpu */ 2180 ts->ts_ticks += 1 << SCHED_TICK_SHIFT; 2181 ts->ts_ltick = ticks; 2182 /* 2183 * Update if we've exceeded our desired tick threshhold by over one 2184 * second. 2185 */ 2186 if (ts->ts_ftick + SCHED_TICK_MAX < ts->ts_ltick) 2187 sched_pctcpu_update(ts); 2188} 2189 2190/* 2191 * Return whether the current CPU has runnable tasks. Used for in-kernel 2192 * cooperative idle threads. 2193 */ 2194int 2195sched_runnable(void) 2196{ 2197 struct tdq *tdq; 2198 int load; 2199 2200 load = 1; 2201 2202 tdq = TDQ_SELF(); 2203 if ((curthread->td_flags & TDF_IDLETD) != 0) { 2204 if (tdq->tdq_load > 0) 2205 goto out; 2206 } else 2207 if (tdq->tdq_load - 1 > 0) 2208 goto out; 2209 load = 0; 2210out: 2211 return (load); 2212} 2213 2214/* 2215 * Choose the highest priority thread to run. The thread is removed from 2216 * the run-queue while running however the load remains. For SMP we set 2217 * the tdq in the global idle bitmask if it idles here. 2218 */ 2219struct thread * 2220sched_choose(void) 2221{ 2222#ifdef SMP 2223 struct tdq_group *tdg; 2224#endif 2225 struct td_sched *ts; 2226 struct tdq *tdq; 2227 2228 tdq = TDQ_SELF(); 2229 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2230 ts = tdq_choose(tdq); 2231 if (ts) { 2232 tdq_runq_rem(tdq, ts); 2233 return (ts->ts_thread); 2234 } 2235#ifdef SMP 2236 /* 2237 * We only set the idled bit when all of the cpus in the group are 2238 * idle. Otherwise we could get into a situation where a thread bounces 2239 * back and forth between two idle cores on seperate physical CPUs. 2240 */ 2241 tdg = tdq->tdq_group; 2242 tdg->tdg_idlemask |= PCPU_GET(cpumask); 2243 if (tdg->tdg_idlemask == tdg->tdg_cpumask) 2244 atomic_set_int(&tdq_idle, tdg->tdg_mask); 2245 tdq->tdq_lowpri = PRI_MAX_IDLE; 2246#endif 2247 return (PCPU_GET(idlethread)); 2248} 2249 2250/* 2251 * Set owepreempt if necessary. Preemption never happens directly in ULE, 2252 * we always request it once we exit a critical section. 2253 */ 2254static inline void 2255sched_setpreempt(struct thread *td) 2256{ 2257 struct thread *ctd; 2258 int cpri; 2259 int pri; 2260 2261 ctd = curthread; 2262 pri = td->td_priority; 2263 cpri = ctd->td_priority; 2264 if (td->td_priority < ctd->td_priority) 2265 curthread->td_flags |= TDF_NEEDRESCHED; 2266 if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd)) 2267 return; 2268 /* 2269 * Always preempt IDLE threads. Otherwise only if the preempting 2270 * thread is an ithread. 2271 */ 2272 if (pri > preempt_thresh && cpri < PRI_MIN_IDLE) 2273 return; 2274 ctd->td_owepreempt = 1; 2275 return; 2276} 2277 2278/* 2279 * Add a thread to a thread queue. Initializes priority, slice, runq, and 2280 * add it to the appropriate queue. This is the internal function called 2281 * when the tdq is predetermined. 2282 */ 2283void 2284tdq_add(struct tdq *tdq, struct thread *td, int flags) 2285{ 2286 struct td_sched *ts; 2287 int class; 2288#ifdef SMP 2289 int cpumask; 2290#endif 2291 2292 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2293 KASSERT((td->td_inhibitors == 0), 2294 ("sched_add: trying to run inhibited thread")); 2295 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)), 2296 ("sched_add: bad thread state")); 2297 KASSERT(td->td_flags & TDF_INMEM, 2298 ("sched_add: thread swapped out")); 2299 2300 ts = td->td_sched; 2301 class = PRI_BASE(td->td_pri_class); 2302 TD_SET_RUNQ(td); 2303 if (ts->ts_slice == 0) 2304 ts->ts_slice = sched_slice; 2305 /* 2306 * Pick the run queue based on priority. 2307 */ 2308 if (td->td_priority <= PRI_MAX_REALTIME) 2309 ts->ts_runq = &tdq->tdq_realtime; 2310 else if (td->td_priority <= PRI_MAX_TIMESHARE) 2311 ts->ts_runq = &tdq->tdq_timeshare; 2312 else 2313 ts->ts_runq = &tdq->tdq_idle; 2314#ifdef SMP 2315 cpumask = 1 << ts->ts_cpu; 2316 /* 2317 * If we had been idle, clear our bit in the group and potentially 2318 * the global bitmap. 2319 */ 2320 if ((class != PRI_IDLE && class != PRI_ITHD) && 2321 (tdq->tdq_group->tdg_idlemask & cpumask) != 0) { 2322 /* 2323 * Check to see if our group is unidling, and if so, remove it 2324 * from the global idle mask. 2325 */ 2326 if (tdq->tdq_group->tdg_idlemask == 2327 tdq->tdq_group->tdg_cpumask) 2328 atomic_clear_int(&tdq_idle, tdq->tdq_group->tdg_mask); 2329 /* 2330 * Now remove ourselves from the group specific idle mask. 2331 */ 2332 tdq->tdq_group->tdg_idlemask &= ~cpumask; 2333 } 2334 if (td->td_priority < tdq->tdq_lowpri) 2335 tdq->tdq_lowpri = td->td_priority; 2336#endif 2337 tdq_runq_add(tdq, ts, flags); 2338 tdq_load_add(tdq, ts); 2339} 2340 2341/* 2342 * Select the target thread queue and add a thread to it. Request 2343 * preemption or IPI a remote processor if required. 2344 */ 2345void 2346sched_add(struct thread *td, int flags) 2347{ 2348 struct td_sched *ts; 2349 struct tdq *tdq; 2350#ifdef SMP 2351 int cpuid; 2352 int cpu; 2353#endif 2354 CTR5(KTR_SCHED, "sched_add: %p(%s) prio %d by %p(%s)", 2355 td, td->td_proc->p_comm, td->td_priority, curthread, 2356 curthread->td_proc->p_comm); 2357 THREAD_LOCK_ASSERT(td, MA_OWNED); 2358 ts = td->td_sched; 2359 /* 2360 * Recalculate the priority before we select the target cpu or 2361 * run-queue. 2362 */ 2363 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) 2364 sched_priority(td); 2365#ifdef SMP 2366 cpuid = PCPU_GET(cpuid); 2367 /* 2368 * Pick the destination cpu and if it isn't ours transfer to the 2369 * target cpu. 2370 */ 2371 if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_MIGRATE(td)) 2372 cpu = cpuid; 2373 else if (!THREAD_CAN_MIGRATE(td)) 2374 cpu = ts->ts_cpu; 2375 else 2376 cpu = sched_pickcpu(ts, flags); 2377 tdq = sched_setcpu(ts, cpu, flags); 2378 tdq_add(tdq, td, flags); 2379 if (cpu != cpuid) { 2380 tdq_notify(ts); 2381 return; 2382 } 2383#else 2384 tdq = TDQ_SELF(); 2385 TDQ_LOCK(tdq); 2386 /* 2387 * Now that the thread is moving to the run-queue, set the lock 2388 * to the scheduler's lock. 2389 */ 2390 thread_lock_set(td, TDQ_LOCKPTR(tdq)); 2391 tdq_add(tdq, td, flags); 2392#endif 2393 if (!(flags & SRQ_YIELDING)) 2394 sched_setpreempt(td); 2395} 2396 2397/* 2398 * Remove a thread from a run-queue without running it. This is used 2399 * when we're stealing a thread from a remote queue. Otherwise all threads 2400 * exit by calling sched_exit_thread() and sched_throw() themselves. 2401 */ 2402void 2403sched_rem(struct thread *td) 2404{ 2405 struct tdq *tdq; 2406 struct td_sched *ts; 2407 2408 CTR5(KTR_SCHED, "sched_rem: %p(%s) prio %d by %p(%s)", 2409 td, td->td_proc->p_comm, td->td_priority, curthread, 2410 curthread->td_proc->p_comm); 2411 ts = td->td_sched; 2412 tdq = TDQ_CPU(ts->ts_cpu); 2413 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2414 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2415 KASSERT(TD_ON_RUNQ(td), 2416 ("sched_rem: thread not on run queue")); 2417 tdq_runq_rem(tdq, ts); 2418 tdq_load_rem(tdq, ts); 2419 TD_SET_CAN_RUN(td); 2420} 2421 2422/* 2423 * Fetch cpu utilization information. Updates on demand. 2424 */ 2425fixpt_t 2426sched_pctcpu(struct thread *td) 2427{ 2428 fixpt_t pctcpu; 2429 struct td_sched *ts; 2430 2431 pctcpu = 0; 2432 ts = td->td_sched; 2433 if (ts == NULL) 2434 return (0); 2435 2436 thread_lock(td); 2437 if (ts->ts_ticks) { 2438 int rtick; 2439 2440 sched_pctcpu_update(ts); 2441 /* How many rtick per second ? */ 2442 rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz); 2443 pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT; 2444 } 2445 thread_unlock(td); 2446 2447 return (pctcpu); 2448} 2449 2450/* 2451 * Bind a thread to a target cpu. 2452 */ 2453void 2454sched_bind(struct thread *td, int cpu) 2455{ 2456 struct td_sched *ts; 2457 2458 THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED); 2459 ts = td->td_sched; 2460 if (ts->ts_flags & TSF_BOUND) 2461 sched_unbind(td); 2462 ts->ts_flags |= TSF_BOUND; 2463#ifdef SMP 2464 sched_pin(); 2465 if (PCPU_GET(cpuid) == cpu) 2466 return; 2467 ts->ts_cpu = cpu; 2468 /* When we return from mi_switch we'll be on the correct cpu. */ 2469 mi_switch(SW_VOL, NULL); 2470#endif 2471} 2472 2473/* 2474 * Release a bound thread. 2475 */ 2476void 2477sched_unbind(struct thread *td) 2478{ 2479 struct td_sched *ts; 2480 2481 THREAD_LOCK_ASSERT(td, MA_OWNED); 2482 ts = td->td_sched; 2483 if ((ts->ts_flags & TSF_BOUND) == 0) 2484 return; 2485 ts->ts_flags &= ~TSF_BOUND; 2486#ifdef SMP 2487 sched_unpin(); 2488#endif 2489} 2490 2491int 2492sched_is_bound(struct thread *td) 2493{ 2494 THREAD_LOCK_ASSERT(td, MA_OWNED); 2495 return (td->td_sched->ts_flags & TSF_BOUND); 2496} 2497 2498/* 2499 * Basic yield call. 2500 */ 2501void 2502sched_relinquish(struct thread *td) 2503{ 2504 thread_lock(td); 2505 if (td->td_pri_class == PRI_TIMESHARE) 2506 sched_prio(td, PRI_MAX_TIMESHARE); 2507 SCHED_STAT_INC(switch_relinquish); 2508 mi_switch(SW_VOL, NULL); 2509 thread_unlock(td); 2510} 2511 2512/* 2513 * Return the total system load. 2514 */ 2515int 2516sched_load(void) 2517{ 2518#ifdef SMP 2519 int total; 2520 int i; 2521 2522 total = 0; 2523 for (i = 0; i <= tdg_maxid; i++) 2524 total += TDQ_GROUP(i)->tdg_load; 2525 return (total); 2526#else 2527 return (TDQ_SELF()->tdq_sysload); 2528#endif 2529} 2530 2531int 2532sched_sizeof_proc(void) 2533{ 2534 return (sizeof(struct proc)); 2535} 2536 2537int 2538sched_sizeof_thread(void) 2539{ 2540 return (sizeof(struct thread) + sizeof(struct td_sched)); 2541} 2542 2543/* 2544 * The actual idle process. 2545 */ 2546void 2547sched_idletd(void *dummy) 2548{ 2549 struct thread *td; 2550 struct tdq *tdq; 2551 2552 td = curthread; 2553 tdq = TDQ_SELF(); 2554 mtx_assert(&Giant, MA_NOTOWNED); 2555 /* ULE relies on preemption for idle interruption. */ 2556 for (;;) { 2557#ifdef SMP 2558 if (tdq_idled(tdq)) 2559 cpu_idle(); 2560#else 2561 cpu_idle(); 2562#endif 2563 } 2564} 2565 2566/* 2567 * A CPU is entering for the first time or a thread is exiting. 2568 */ 2569void 2570sched_throw(struct thread *td) 2571{ 2572 struct tdq *tdq; 2573 2574 tdq = TDQ_SELF(); 2575 if (td == NULL) { 2576 /* Correct spinlock nesting and acquire the correct lock. */ 2577 TDQ_LOCK(tdq); 2578 spinlock_exit(); 2579 } else { 2580 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2581 tdq_load_rem(tdq, td->td_sched); 2582 } 2583 KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count")); 2584 PCPU_SET(switchtime, cpu_ticks()); 2585 PCPU_SET(switchticks, ticks); 2586 cpu_throw(td, choosethread()); /* doesn't return */ 2587} 2588 2589/* 2590 * This is called from fork_exit(). Just acquire the correct locks and 2591 * let fork do the rest of the work. 2592 */ 2593void 2594sched_fork_exit(struct thread *td) 2595{ 2596 struct td_sched *ts; 2597 struct tdq *tdq; 2598 int cpuid; 2599 2600 /* 2601 * Finish setting up thread glue so that it begins execution in a 2602 * non-nested critical section with the scheduler lock held. 2603 */ 2604 cpuid = PCPU_GET(cpuid); 2605 tdq = TDQ_CPU(cpuid); 2606 ts = td->td_sched; 2607 if (TD_IS_IDLETHREAD(td)) 2608 td->td_lock = TDQ_LOCKPTR(tdq); 2609 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2610 td->td_oncpu = cpuid; 2611 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)td; 2612 THREAD_LOCK_ASSERT(td, MA_OWNED | MA_NOTRECURSED); 2613} 2614 2615static SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, 2616 "Scheduler"); 2617SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0, 2618 "Scheduler name"); 2619SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0, 2620 "Slice size for timeshare threads"); 2621SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0, 2622 "Interactivity score threshold"); 2623SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW, &preempt_thresh, 2624 0,"Min priority for preemption, lower priorities have greater precedence"); 2625#ifdef SMP 2626SYSCTL_INT(_kern_sched, OID_AUTO, pick_pri, CTLFLAG_RW, &pick_pri, 0, 2627 "Pick the target cpu based on priority rather than load."); 2628SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0, 2629 "Number of hz ticks to keep thread affinity for"); 2630SYSCTL_INT(_kern_sched, OID_AUTO, tryself, CTLFLAG_RW, &tryself, 0, ""); 2631SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0, 2632 "Enables the long-term load balancer"); 2633SYSCTL_INT(_kern_sched, OID_AUTO, balance_secs, CTLFLAG_RW, &balance_secs, 0, 2634 "Average frequence in seconds to run the long-term balancer"); 2635SYSCTL_INT(_kern_sched, OID_AUTO, steal_htt, CTLFLAG_RW, &steal_htt, 0, 2636 "Steals work from another hyper-threaded core on idle"); 2637SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0, 2638 "Attempts to steal work from other cores before idling"); 2639SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0, 2640 "Minimum load on remote cpu before we'll steal"); 2641SYSCTL_INT(_kern_sched, OID_AUTO, topology, CTLFLAG_RD, &topology, 0, 2642 "True when a topology has been specified by the MD code."); 2643#endif 2644 2645/* ps compat. All cpu percentages from ULE are weighted. */ 2646static int ccpu = 0; 2647SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 2648 2649 2650#define KERN_SWITCH_INCLUDE 1 2651#include "kern/kern_switch.c" 2652