sched_ule.c revision 171482
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 it's 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 171482 2007-07-17 22:53:23Z 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_slptick; /* Tick when we went to sleep. */ 92 int ts_slice; /* Ticks of slice remaining. */ 93 u_int ts_slptime; /* Number of ticks we vol. slept */ 94 u_int ts_runtime; /* Number of ticks we were running */ 95 /* The following variables are only used for pctcpu calculation */ 96 int ts_ltick; /* Last tick that we were running on */ 97 int ts_ftick; /* First tick that we were running on */ 98 int ts_ticks; /* Tick count */ 99#ifdef SMP 100 int ts_rltick; /* Real last tick, for affinity. */ 101#endif 102}; 103/* flags kept in ts_flags */ 104#define TSF_BOUND 0x0001 /* Thread can not migrate. */ 105#define TSF_XFERABLE 0x0002 /* Thread was added as transferable. */ 106 107static struct td_sched td_sched0; 108 109/* 110 * Cpu percentage computation macros and defines. 111 * 112 * SCHED_TICK_SECS: Number of seconds to average the cpu usage across. 113 * SCHED_TICK_TARG: Number of hz ticks to average the cpu usage across. 114 * SCHED_TICK_MAX: Maximum number of ticks before scaling back. 115 * SCHED_TICK_SHIFT: Shift factor to avoid rounding away results. 116 * SCHED_TICK_HZ: Compute the number of hz ticks for a given ticks count. 117 * SCHED_TICK_TOTAL: Gives the amount of time we've been recording ticks. 118 */ 119#define SCHED_TICK_SECS 10 120#define SCHED_TICK_TARG (hz * SCHED_TICK_SECS) 121#define SCHED_TICK_MAX (SCHED_TICK_TARG + hz) 122#define SCHED_TICK_SHIFT 10 123#define SCHED_TICK_HZ(ts) ((ts)->ts_ticks >> SCHED_TICK_SHIFT) 124#define SCHED_TICK_TOTAL(ts) (max((ts)->ts_ltick - (ts)->ts_ftick, hz)) 125 126/* 127 * These macros determine priorities for non-interactive threads. They are 128 * assigned a priority based on their recent cpu utilization as expressed 129 * by the ratio of ticks to the tick total. NHALF priorities at the start 130 * and end of the MIN to MAX timeshare range are only reachable with negative 131 * or positive nice respectively. 132 * 133 * PRI_RANGE: Priority range for utilization dependent priorities. 134 * PRI_NRESV: Number of nice values. 135 * PRI_TICKS: Compute a priority in PRI_RANGE from the ticks count and total. 136 * PRI_NICE: Determines the part of the priority inherited from nice. 137 */ 138#define SCHED_PRI_NRESV (PRIO_MAX - PRIO_MIN) 139#define SCHED_PRI_NHALF (SCHED_PRI_NRESV / 2) 140#define SCHED_PRI_MIN (PRI_MIN_TIMESHARE + SCHED_PRI_NHALF) 141#define SCHED_PRI_MAX (PRI_MAX_TIMESHARE - SCHED_PRI_NHALF) 142#define SCHED_PRI_RANGE (SCHED_PRI_MAX - SCHED_PRI_MIN) 143#define SCHED_PRI_TICKS(ts) \ 144 (SCHED_TICK_HZ((ts)) / \ 145 (roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE)) 146#define SCHED_PRI_NICE(nice) (nice) 147 148/* 149 * These determine the interactivity of a process. Interactivity differs from 150 * cpu utilization in that it expresses the voluntary time slept vs time ran 151 * while cpu utilization includes all time not running. This more accurately 152 * models the intent of the thread. 153 * 154 * SLP_RUN_MAX: Maximum amount of sleep time + run time we'll accumulate 155 * before throttling back. 156 * SLP_RUN_FORK: Maximum slp+run time to inherit at fork time. 157 * INTERACT_MAX: Maximum interactivity value. Smaller is better. 158 * INTERACT_THRESH: Threshhold for placement on the current runq. 159 */ 160#define SCHED_SLP_RUN_MAX ((hz * 5) << SCHED_TICK_SHIFT) 161#define SCHED_SLP_RUN_FORK ((hz / 2) << SCHED_TICK_SHIFT) 162#define SCHED_INTERACT_MAX (100) 163#define SCHED_INTERACT_HALF (SCHED_INTERACT_MAX / 2) 164#define SCHED_INTERACT_THRESH (30) 165 166/* 167 * tickincr: Converts a stathz tick into a hz domain scaled by 168 * the shift factor. Without the shift the error rate 169 * due to rounding would be unacceptably high. 170 * realstathz: stathz is sometimes 0 and run off of hz. 171 * sched_slice: Runtime of each thread before rescheduling. 172 * preempt_thresh: Priority threshold for preemption and remote IPIs. 173 */ 174static int sched_interact = SCHED_INTERACT_THRESH; 175static int realstathz; 176static int tickincr; 177static int sched_slice; 178static int preempt_thresh = PRI_MIN_KERN; 179 180#define SCHED_BAL_SECS 2 /* How often we run the rebalance algorithm. */ 181 182/* 183 * tdq - per processor runqs and statistics. All fields are protected by the 184 * tdq_lock. The load and lowpri may be accessed without to avoid excess 185 * locking in sched_pickcpu(); 186 */ 187struct tdq { 188 struct mtx tdq_lock; /* Protects all fields below. */ 189 struct runq tdq_realtime; /* real-time run queue. */ 190 struct runq tdq_timeshare; /* timeshare run queue. */ 191 struct runq tdq_idle; /* Queue of IDLE threads. */ 192 int tdq_load; /* Aggregate load. */ 193 u_char tdq_idx; /* Current insert index. */ 194 u_char tdq_ridx; /* Current removal index. */ 195#ifdef SMP 196 u_char tdq_lowpri; /* Lowest priority thread. */ 197 int tdq_transferable; /* Transferable thread count. */ 198 LIST_ENTRY(tdq) tdq_siblings; /* Next in tdq group. */ 199 struct tdq_group *tdq_group; /* Our processor group. */ 200#else 201 int tdq_sysload; /* For loadavg, !ITHD load. */ 202#endif 203 char tdq_name[16]; /* lock name. */ 204} __aligned(64); 205 206 207#ifdef SMP 208/* 209 * tdq groups are groups of processors which can cheaply share threads. When 210 * one processor in the group goes idle it will check the runqs of the other 211 * processors in its group prior to halting and waiting for an interrupt. 212 * These groups are suitable for SMT (Symetric Multi-Threading) and not NUMA. 213 * In a numa environment we'd want an idle bitmap per group and a two tiered 214 * load balancer. 215 */ 216struct tdq_group { 217 int tdg_cpus; /* Count of CPUs in this tdq group. */ 218 cpumask_t tdg_cpumask; /* Mask of cpus in this group. */ 219 cpumask_t tdg_idlemask; /* Idle cpus in this group. */ 220 cpumask_t tdg_mask; /* Bit mask for first cpu. */ 221 int tdg_load; /* Total load of this group. */ 222 int tdg_transferable; /* Transferable load of this group. */ 223 LIST_HEAD(, tdq) tdg_members; /* Linked list of all members. */ 224} __aligned(64); 225 226#define SCHED_AFFINITY_DEFAULT (max(1, hz / 300)) 227#define SCHED_AFFINITY(ts) ((ts)->ts_rltick > ticks - affinity) 228 229/* 230 * Run-time tunables. 231 */ 232static int rebalance = 0; 233static int pick_pri = 0; 234static int pick_zero = 0; 235static int affinity; 236static int tryself = 1; 237static int tryselfidle = 1; 238static int steal_htt = 0; 239static int steal_idle = 0; 240static int topology = 0; 241 242/* 243 * One thread queue per processor. 244 */ 245static volatile cpumask_t tdq_idle; 246static int tdg_maxid; 247static struct tdq tdq_cpu[MAXCPU]; 248static struct tdq_group tdq_groups[MAXCPU]; 249static struct callout balco; 250static struct callout gbalco; 251 252#define TDQ_SELF() (&tdq_cpu[PCPU_GET(cpuid)]) 253#define TDQ_CPU(x) (&tdq_cpu[(x)]) 254#define TDQ_ID(x) ((x) - tdq_cpu) 255#define TDQ_GROUP(x) (&tdq_groups[(x)]) 256#else /* !SMP */ 257static struct tdq tdq_cpu; 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 *); 301 302#define THREAD_CAN_MIGRATE(td) ((td)->td_pinned == 0) 303#endif 304 305static void sched_setup(void *dummy); 306SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL) 307 308static void sched_initticks(void *dummy); 309SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks, NULL) 310 311/* 312 * Print the threads waiting on a run-queue. 313 */ 314static void 315runq_print(struct runq *rq) 316{ 317 struct rqhead *rqh; 318 struct td_sched *ts; 319 int pri; 320 int j; 321 int i; 322 323 for (i = 0; i < RQB_LEN; i++) { 324 printf("\t\trunq bits %d 0x%zx\n", 325 i, rq->rq_status.rqb_bits[i]); 326 for (j = 0; j < RQB_BPW; j++) 327 if (rq->rq_status.rqb_bits[i] & (1ul << j)) { 328 pri = j + (i << RQB_L2BPW); 329 rqh = &rq->rq_queues[pri]; 330 TAILQ_FOREACH(ts, rqh, ts_procq) { 331 printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n", 332 ts->ts_thread, ts->ts_thread->td_proc->p_comm, ts->ts_thread->td_priority, ts->ts_rqindex, pri); 333 } 334 } 335 } 336} 337 338/* 339 * Print the status of a per-cpu thread queue. Should be a ddb show cmd. 340 */ 341void 342tdq_print(int cpu) 343{ 344 struct tdq *tdq; 345 346 tdq = TDQ_CPU(cpu); 347 348 printf("tdq:\n"); 349 printf("\tlockptr %p\n", TDQ_LOCKPTR(tdq)); 350 printf("\tlock name %s\n", tdq->tdq_name); 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#endif 364} 365 366#define TS_RQ_PPQ (((PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE) + 1) / RQ_NQS) 367/* 368 * Add a thread to the actual run-queue. Keeps transferable counts up to 369 * date with what is actually on the run-queue. Selects the correct 370 * queue position for timeshare threads. 371 */ 372static __inline void 373tdq_runq_add(struct tdq *tdq, struct td_sched *ts, int flags) 374{ 375 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 376 THREAD_LOCK_ASSERT(ts->ts_thread, MA_OWNED); 377#ifdef SMP 378 if (THREAD_CAN_MIGRATE(ts->ts_thread)) { 379 tdq->tdq_transferable++; 380 tdq->tdq_group->tdg_transferable++; 381 ts->ts_flags |= TSF_XFERABLE; 382 } 383#endif 384 if (ts->ts_runq == &tdq->tdq_timeshare) { 385 u_char pri; 386 387 pri = ts->ts_thread->td_priority; 388 KASSERT(pri <= PRI_MAX_TIMESHARE && pri >= PRI_MIN_TIMESHARE, 389 ("Invalid priority %d on timeshare runq", pri)); 390 /* 391 * This queue contains only priorities between MIN and MAX 392 * realtime. Use the whole queue to represent these values. 393 */ 394 if ((flags & SRQ_BORROWING) == 0) { 395 pri = (pri - PRI_MIN_TIMESHARE) / TS_RQ_PPQ; 396 pri = (pri + tdq->tdq_idx) % RQ_NQS; 397 /* 398 * This effectively shortens the queue by one so we 399 * can have a one slot difference between idx and 400 * ridx while we wait for threads to drain. 401 */ 402 if (tdq->tdq_ridx != tdq->tdq_idx && 403 pri == tdq->tdq_ridx) 404 pri = (unsigned char)(pri - 1) % RQ_NQS; 405 } else 406 pri = tdq->tdq_ridx; 407 runq_add_pri(ts->ts_runq, ts, pri, flags); 408 } else 409 runq_add(ts->ts_runq, ts, flags); 410} 411 412/* 413 * Remove a thread from a run-queue. This typically happens when a thread 414 * is selected to run. Running threads are not on the queue and the 415 * transferable count does not reflect them. 416 */ 417static __inline void 418tdq_runq_rem(struct tdq *tdq, struct td_sched *ts) 419{ 420 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 421 KASSERT(ts->ts_runq != NULL, 422 ("tdq_runq_remove: thread %p null ts_runq", ts->ts_thread)); 423#ifdef SMP 424 if (ts->ts_flags & TSF_XFERABLE) { 425 tdq->tdq_transferable--; 426 tdq->tdq_group->tdg_transferable--; 427 ts->ts_flags &= ~TSF_XFERABLE; 428 } 429#endif 430 if (ts->ts_runq == &tdq->tdq_timeshare) { 431 if (tdq->tdq_idx != tdq->tdq_ridx) 432 runq_remove_idx(ts->ts_runq, ts, &tdq->tdq_ridx); 433 else 434 runq_remove_idx(ts->ts_runq, ts, NULL); 435 /* 436 * For timeshare threads we update the priority here so 437 * the priority reflects the time we've been sleeping. 438 */ 439 ts->ts_ltick = ticks; 440 sched_pctcpu_update(ts); 441 sched_priority(ts->ts_thread); 442 } else 443 runq_remove(ts->ts_runq, ts); 444} 445 446/* 447 * Load is maintained for all threads RUNNING and ON_RUNQ. Add the load 448 * for this thread to the referenced thread queue. 449 */ 450static void 451tdq_load_add(struct tdq *tdq, struct td_sched *ts) 452{ 453 int class; 454 455 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 456 THREAD_LOCK_ASSERT(ts->ts_thread, MA_OWNED); 457 class = PRI_BASE(ts->ts_thread->td_pri_class); 458 tdq->tdq_load++; 459 CTR2(KTR_SCHED, "cpu %jd load: %d", TDQ_ID(tdq), tdq->tdq_load); 460 if (class != PRI_ITHD && 461 (ts->ts_thread->td_proc->p_flag & P_NOLOAD) == 0) 462#ifdef SMP 463 tdq->tdq_group->tdg_load++; 464#else 465 tdq->tdq_sysload++; 466#endif 467} 468 469/* 470 * Remove the load from a thread that is transitioning to a sleep state or 471 * exiting. 472 */ 473static void 474tdq_load_rem(struct tdq *tdq, struct td_sched *ts) 475{ 476 int class; 477 478 THREAD_LOCK_ASSERT(ts->ts_thread, MA_OWNED); 479 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 480 class = PRI_BASE(ts->ts_thread->td_pri_class); 481 if (class != PRI_ITHD && 482 (ts->ts_thread->td_proc->p_flag & P_NOLOAD) == 0) 483#ifdef SMP 484 tdq->tdq_group->tdg_load--; 485#else 486 tdq->tdq_sysload--; 487#endif 488 KASSERT(tdq->tdq_load != 0, 489 ("tdq_load_rem: Removing with 0 load on queue %d", (int)TDQ_ID(tdq))); 490 tdq->tdq_load--; 491 CTR1(KTR_SCHED, "load: %d", tdq->tdq_load); 492 ts->ts_runq = NULL; 493} 494 495#ifdef SMP 496/* 497 * sched_balance is a simple CPU load balancing algorithm. It operates by 498 * finding the least loaded and most loaded cpu and equalizing their load 499 * by migrating some processes. 500 * 501 * Dealing only with two CPUs at a time has two advantages. Firstly, most 502 * installations will only have 2 cpus. Secondly, load balancing too much at 503 * once can have an unpleasant effect on the system. The scheduler rarely has 504 * enough information to make perfect decisions. So this algorithm chooses 505 * simplicity and more gradual effects on load in larger systems. 506 * 507 */ 508static void 509sched_balance(void *arg) 510{ 511 struct tdq_group *high; 512 struct tdq_group *low; 513 struct tdq_group *tdg; 514 int cnt; 515 int i; 516 517 callout_reset(&balco, max(hz / 2, random() % (hz * SCHED_BAL_SECS)), 518 sched_balance, NULL); 519 if (smp_started == 0 || rebalance == 0) 520 return; 521 low = high = NULL; 522 i = random() % (tdg_maxid + 1); 523 for (cnt = 0; cnt <= tdg_maxid; cnt++) { 524 tdg = TDQ_GROUP(i); 525 /* 526 * Find the CPU with the highest load that has some 527 * threads to transfer. 528 */ 529 if ((high == NULL || tdg->tdg_load > high->tdg_load) 530 && tdg->tdg_transferable) 531 high = tdg; 532 if (low == NULL || tdg->tdg_load < low->tdg_load) 533 low = tdg; 534 if (++i > tdg_maxid) 535 i = 0; 536 } 537 if (low != NULL && high != NULL && high != low) 538 sched_balance_pair(LIST_FIRST(&high->tdg_members), 539 LIST_FIRST(&low->tdg_members)); 540} 541 542/* 543 * Balance load between CPUs in a group. Will only migrate within the group. 544 */ 545static void 546sched_balance_groups(void *arg) 547{ 548 int i; 549 550 callout_reset(&gbalco, max(hz / 2, random() % (hz * SCHED_BAL_SECS)), 551 sched_balance_groups, NULL); 552 if (smp_started == 0 || rebalance == 0) 553 return; 554 for (i = 0; i <= tdg_maxid; i++) 555 sched_balance_group(TDQ_GROUP(i)); 556} 557 558/* 559 * Finds the greatest imbalance between two tdqs in a group. 560 */ 561static void 562sched_balance_group(struct tdq_group *tdg) 563{ 564 struct tdq *tdq; 565 struct tdq *high; 566 struct tdq *low; 567 int load; 568 569 if (tdg->tdg_transferable == 0) 570 return; 571 low = NULL; 572 high = NULL; 573 LIST_FOREACH(tdq, &tdg->tdg_members, tdq_siblings) { 574 load = tdq->tdq_load; 575 if (high == NULL || load > high->tdq_load) 576 high = tdq; 577 if (low == NULL || load < low->tdq_load) 578 low = tdq; 579 } 580 if (high != NULL && low != NULL && high != low) 581 sched_balance_pair(high, low); 582} 583 584/* 585 * Lock two thread queues using their address to maintain lock order. 586 */ 587static void 588tdq_lock_pair(struct tdq *one, struct tdq *two) 589{ 590 if (one < two) { 591 TDQ_LOCK(one); 592 TDQ_LOCK_FLAGS(two, MTX_DUPOK); 593 } else { 594 TDQ_LOCK(two); 595 TDQ_LOCK_FLAGS(one, MTX_DUPOK); 596 } 597} 598 599/* 600 * Transfer load between two imbalanced thread queues. 601 */ 602static void 603sched_balance_pair(struct tdq *high, struct tdq *low) 604{ 605 int transferable; 606 int high_load; 607 int low_load; 608 int move; 609 int diff; 610 int i; 611 612 tdq_lock_pair(high, low); 613 /* 614 * If we're transfering within a group we have to use this specific 615 * tdq's transferable count, otherwise we can steal from other members 616 * of the group. 617 */ 618 if (high->tdq_group == low->tdq_group) { 619 transferable = high->tdq_transferable; 620 high_load = high->tdq_load; 621 low_load = low->tdq_load; 622 } else { 623 transferable = high->tdq_group->tdg_transferable; 624 high_load = high->tdq_group->tdg_load; 625 low_load = low->tdq_group->tdg_load; 626 } 627 /* 628 * Determine what the imbalance is and then adjust that to how many 629 * threads we actually have to give up (transferable). 630 */ 631 if (transferable != 0) { 632 diff = high_load - low_load; 633 move = diff / 2; 634 if (diff & 0x1) 635 move++; 636 move = min(move, transferable); 637 for (i = 0; i < move; i++) 638 tdq_move(high, low); 639 } 640 TDQ_UNLOCK(high); 641 TDQ_UNLOCK(low); 642 return; 643} 644 645/* 646 * Move a thread from one thread queue to another. 647 */ 648static void 649tdq_move(struct tdq *from, struct tdq *to) 650{ 651 struct td_sched *ts; 652 struct thread *td; 653 struct tdq *tdq; 654 int cpu; 655 656 tdq = from; 657 cpu = TDQ_ID(to); 658 ts = tdq_steal(tdq, 1); 659 if (ts == NULL) { 660 struct tdq_group *tdg; 661 662 tdg = tdq->tdq_group; 663 LIST_FOREACH(tdq, &tdg->tdg_members, tdq_siblings) { 664 if (tdq == from || tdq->tdq_transferable == 0) 665 continue; 666 ts = tdq_steal(tdq, 1); 667 break; 668 } 669 if (ts == NULL) 670 return; 671 } 672 if (tdq == to) 673 return; 674 td = ts->ts_thread; 675 /* 676 * Although the run queue is locked the thread may be blocked. Lock 677 * it to clear this. 678 */ 679 thread_lock(td); 680 /* Drop recursive lock on from. */ 681 TDQ_UNLOCK(from); 682 sched_rem(td); 683 ts->ts_cpu = cpu; 684 td->td_lock = TDQ_LOCKPTR(to); 685 tdq_add(to, td, SRQ_YIELDING); 686} 687 688/* 689 * This tdq has idled. Try to steal a thread from another cpu and switch 690 * to it. 691 */ 692static int 693tdq_idled(struct tdq *tdq) 694{ 695 struct tdq_group *tdg; 696 struct tdq *steal; 697 struct td_sched *ts; 698 struct thread *td; 699 int highload; 700 int highcpu; 701 int load; 702 int cpu; 703 704 /* We don't want to be preempted while we're iterating over tdqs */ 705 spinlock_enter(); 706 tdg = tdq->tdq_group; 707 /* 708 * If we're in a cpu group, try and steal threads from another cpu in 709 * the group before idling. 710 */ 711 if (steal_htt && tdg->tdg_cpus > 1 && tdg->tdg_transferable) { 712 LIST_FOREACH(steal, &tdg->tdg_members, tdq_siblings) { 713 if (steal == tdq || steal->tdq_transferable == 0) 714 continue; 715 TDQ_LOCK(steal); 716 ts = tdq_steal(steal, 0); 717 if (ts) 718 goto steal; 719 TDQ_UNLOCK(steal); 720 } 721 } 722 for (;;) { 723 if (steal_idle == 0) 724 break; 725 highcpu = 0; 726 highload = 0; 727 for (cpu = 0; cpu <= mp_maxid; cpu++) { 728 if (CPU_ABSENT(cpu)) 729 continue; 730 steal = TDQ_CPU(cpu); 731 load = TDQ_CPU(cpu)->tdq_transferable; 732 if (load < highload) 733 continue; 734 highload = load; 735 highcpu = cpu; 736 } 737 if (highload < 2) 738 break; 739 steal = TDQ_CPU(highcpu); 740 TDQ_LOCK(steal); 741 if (steal->tdq_transferable > 1 && 742 (ts = tdq_steal(steal, 1)) != NULL) 743 goto steal; 744 TDQ_UNLOCK(steal); 745 break; 746 } 747 spinlock_exit(); 748 return (1); 749steal: 750 td = ts->ts_thread; 751 thread_lock(td); 752 spinlock_exit(); 753 MPASS(td->td_lock == TDQ_LOCKPTR(steal)); 754 TDQ_UNLOCK(steal); 755 sched_rem(td); 756 sched_setcpu(ts, PCPU_GET(cpuid), SRQ_YIELDING); 757 tdq_add(tdq, td, SRQ_YIELDING); 758 MPASS(td->td_lock == curthread->td_lock); 759 mi_switch(SW_VOL, NULL); 760 thread_unlock(curthread); 761 762 return (0); 763} 764 765/* 766 * Notify a remote cpu of new work. Sends an IPI if criteria are met. 767 */ 768static void 769tdq_notify(struct td_sched *ts) 770{ 771 struct thread *ctd; 772 struct pcpu *pcpu; 773 int cpri; 774 int pri; 775 int cpu; 776 777 cpu = ts->ts_cpu; 778 pri = ts->ts_thread->td_priority; 779 pcpu = pcpu_find(cpu); 780 ctd = pcpu->pc_curthread; 781 cpri = ctd->td_priority; 782 783 /* 784 * If our priority is not better than the current priority there is 785 * nothing to do. 786 */ 787 if (pri > cpri) 788 return; 789 /* 790 * Always IPI idle. 791 */ 792 if (cpri > PRI_MIN_IDLE) 793 goto sendipi; 794 /* 795 * If we're realtime or better and there is timeshare or worse running 796 * send an IPI. 797 */ 798 if (pri < PRI_MAX_REALTIME && cpri > PRI_MAX_REALTIME) 799 goto sendipi; 800 /* 801 * Otherwise only IPI if we exceed the threshold. 802 */ 803 if (pri > preempt_thresh) 804 return; 805sendipi: 806 ctd->td_flags |= TDF_NEEDRESCHED; 807 ipi_selected(1 << cpu, IPI_PREEMPT); 808} 809 810/* 811 * Steals load from a timeshare queue. Honors the rotating queue head 812 * index. 813 */ 814static struct td_sched * 815runq_steal_from(struct runq *rq, u_char start) 816{ 817 struct td_sched *ts; 818 struct rqbits *rqb; 819 struct rqhead *rqh; 820 int first; 821 int bit; 822 int pri; 823 int i; 824 825 rqb = &rq->rq_status; 826 bit = start & (RQB_BPW -1); 827 pri = 0; 828 first = 0; 829again: 830 for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) { 831 if (rqb->rqb_bits[i] == 0) 832 continue; 833 if (bit != 0) { 834 for (pri = bit; pri < RQB_BPW; pri++) 835 if (rqb->rqb_bits[i] & (1ul << pri)) 836 break; 837 if (pri >= RQB_BPW) 838 continue; 839 } else 840 pri = RQB_FFS(rqb->rqb_bits[i]); 841 pri += (i << RQB_L2BPW); 842 rqh = &rq->rq_queues[pri]; 843 TAILQ_FOREACH(ts, rqh, ts_procq) { 844 if (first && THREAD_CAN_MIGRATE(ts->ts_thread)) 845 return (ts); 846 first = 1; 847 } 848 } 849 if (start != 0) { 850 start = 0; 851 goto again; 852 } 853 854 return (NULL); 855} 856 857/* 858 * Steals load from a standard linear queue. 859 */ 860static struct td_sched * 861runq_steal(struct runq *rq) 862{ 863 struct rqhead *rqh; 864 struct rqbits *rqb; 865 struct td_sched *ts; 866 int first; 867 int word; 868 int bit; 869 870 first = 0; 871 rqb = &rq->rq_status; 872 for (word = 0; word < RQB_LEN; word++) { 873 if (rqb->rqb_bits[word] == 0) 874 continue; 875 for (bit = 0; bit < RQB_BPW; bit++) { 876 if ((rqb->rqb_bits[word] & (1ul << bit)) == 0) 877 continue; 878 rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)]; 879 TAILQ_FOREACH(ts, rqh, ts_procq) { 880 if (first && THREAD_CAN_MIGRATE(ts->ts_thread)) 881 return (ts); 882 first = 1; 883 } 884 } 885 } 886 return (NULL); 887} 888 889/* 890 * Attempt to steal a thread in priority order from a thread queue. 891 */ 892static struct td_sched * 893tdq_steal(struct tdq *tdq, int stealidle) 894{ 895 struct td_sched *ts; 896 897 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 898 if ((ts = runq_steal(&tdq->tdq_realtime)) != NULL) 899 return (ts); 900 if ((ts = runq_steal_from(&tdq->tdq_timeshare, tdq->tdq_ridx)) != NULL) 901 return (ts); 902 if (stealidle) 903 return (runq_steal(&tdq->tdq_idle)); 904 return (NULL); 905} 906 907/* 908 * Sets the thread lock and ts_cpu to match the requested cpu. Unlocks the 909 * current lock and returns with the assigned queue locked. If this is 910 * via sched_switch() we leave the thread in a blocked state as an 911 * optimization. 912 */ 913static inline struct tdq * 914sched_setcpu(struct td_sched *ts, int cpu, int flags) 915{ 916 struct thread *td; 917 struct tdq *tdq; 918 919 THREAD_LOCK_ASSERT(ts->ts_thread, MA_OWNED); 920 921 tdq = TDQ_CPU(cpu); 922 td = ts->ts_thread; 923 ts->ts_cpu = cpu; 924 if (td->td_lock == TDQ_LOCKPTR(tdq)) 925 return (tdq); 926#ifdef notyet 927 /* 928 * If the thread isn't running it's lockptr is a 929 * turnstile or a sleepqueue. We can just lock_set without 930 * blocking. 931 */ 932 if (TD_CAN_RUN(td)) { 933 TDQ_LOCK(tdq); 934 thread_lock_set(td, TDQ_LOCKPTR(tdq)); 935 return (tdq); 936 } 937#endif 938 /* 939 * The hard case, migration, we need to block the thread first to 940 * prevent order reversals with other cpus locks. 941 */ 942 thread_lock_block(td); 943 TDQ_LOCK(tdq); 944 /* Return to sched_switch() with the lock still blocked */ 945 if ((flags & SRQ_OURSELF) == 0) 946 thread_lock_unblock(td, TDQ_LOCKPTR(tdq)); 947 return (tdq); 948} 949 950/* 951 * Find the thread queue running the lowest priority thread. 952 */ 953static int 954tdq_lowestpri(void) 955{ 956 struct tdq *tdq; 957 int lowpri; 958 int lowcpu; 959 int lowload; 960 int load; 961 int cpu; 962 int pri; 963 964 lowload = 0; 965 lowpri = lowcpu = 0; 966 for (cpu = 0; cpu <= mp_maxid; cpu++) { 967 if (CPU_ABSENT(cpu)) 968 continue; 969 tdq = TDQ_CPU(cpu); 970 pri = tdq->tdq_lowpri; 971 load = TDQ_CPU(cpu)->tdq_load; 972 CTR4(KTR_ULE, 973 "cpu %d pri %d lowcpu %d lowpri %d", 974 cpu, pri, lowcpu, lowpri); 975 if (pri < lowpri) 976 continue; 977 if (lowpri && lowpri == pri && load > lowload) 978 continue; 979 lowpri = pri; 980 lowcpu = cpu; 981 lowload = load; 982 } 983 984 return (lowcpu); 985} 986 987/* 988 * Find the thread queue with the least load. 989 */ 990static int 991tdq_lowestload(void) 992{ 993 struct tdq *tdq; 994 int lowload; 995 int lowpri; 996 int lowcpu; 997 int load; 998 int cpu; 999 int pri; 1000 1001 lowcpu = 0; 1002 lowload = TDQ_CPU(0)->tdq_load; 1003 lowpri = TDQ_CPU(0)->tdq_lowpri; 1004 for (cpu = 1; cpu <= mp_maxid; cpu++) { 1005 if (CPU_ABSENT(cpu)) 1006 continue; 1007 tdq = TDQ_CPU(cpu); 1008 load = tdq->tdq_load; 1009 pri = tdq->tdq_lowpri; 1010 CTR4(KTR_ULE, "cpu %d load %d lowcpu %d lowload %d", 1011 cpu, load, lowcpu, lowload); 1012 if (load > lowload) 1013 continue; 1014 if (load == lowload && pri < lowpri) 1015 continue; 1016 lowcpu = cpu; 1017 lowload = load; 1018 lowpri = pri; 1019 } 1020 1021 return (lowcpu); 1022} 1023 1024/* 1025 * Pick the destination cpu for sched_add(). Respects affinity and makes 1026 * a determination based on load or priority of available processors. 1027 */ 1028static int 1029sched_pickcpu(struct td_sched *ts, int flags) 1030{ 1031 struct tdq *tdq; 1032 int self; 1033 int pri; 1034 int cpu; 1035 1036 cpu = self = PCPU_GET(cpuid); 1037 if (smp_started == 0) 1038 return (self); 1039 pri = ts->ts_thread->td_priority; 1040 cpu = ts->ts_cpu; 1041 /* 1042 * Regardless of affinity, if the last cpu is idle send it there. 1043 */ 1044 tdq = TDQ_CPU(cpu); 1045 if (tdq->tdq_lowpri > PRI_MIN_IDLE) { 1046 CTR5(KTR_ULE, 1047 "ts_cpu %d idle, ltick %d ticks %d pri %d curthread %d", 1048 ts->ts_cpu, ts->ts_rltick, ticks, pri, 1049 tdq->tdq_lowpri); 1050 return (ts->ts_cpu); 1051 } 1052 /* 1053 * If we have affinity, try to place it on the cpu we last ran on. 1054 */ 1055 if (SCHED_AFFINITY(ts) && tdq->tdq_lowpri > pri) { 1056 CTR5(KTR_ULE, 1057 "affinity for %d, ltick %d ticks %d pri %d curthread %d", 1058 ts->ts_cpu, ts->ts_rltick, ticks, pri, 1059 tdq->tdq_lowpri); 1060 return (ts->ts_cpu); 1061 } 1062 /* 1063 * Try ourself first; If we're running something lower priority this 1064 * may have some locality with the waking thread and execute faster 1065 * here. 1066 */ 1067 if (tryself) { 1068 /* 1069 * If we're being awoken by an interrupt thread or the waker 1070 * is going right to sleep run here as well. 1071 */ 1072 if ((TDQ_SELF()->tdq_load <= 1) && (flags & (SRQ_YIELDING) || 1073 curthread->td_pri_class == PRI_ITHD)) { 1074 CTR2(KTR_ULE, "tryself load %d flags %d", 1075 TDQ_SELF()->tdq_load, flags); 1076 return (self); 1077 } 1078 } 1079 /* 1080 * Look for an idle group. 1081 */ 1082 CTR1(KTR_ULE, "tdq_idle %X", tdq_idle); 1083 cpu = ffs(tdq_idle); 1084 if (cpu) 1085 return (--cpu); 1086 if (tryselfidle && pri < curthread->td_priority) { 1087 CTR1(KTR_ULE, "tryselfidle %d", 1088 curthread->td_priority); 1089 return (self); 1090 } 1091 /* 1092 * XXX Under heavy load mysql performs way better if you 1093 * serialize the non-running threads on one cpu. This is 1094 * a horrible hack. 1095 */ 1096 if (pick_zero) 1097 return (0); 1098 /* 1099 * Now search for the cpu running the lowest priority thread with 1100 * the least load. 1101 */ 1102 if (pick_pri) 1103 cpu = tdq_lowestpri(); 1104 else 1105 cpu = tdq_lowestload(); 1106 return (cpu); 1107} 1108 1109#endif /* SMP */ 1110 1111/* 1112 * Pick the highest priority task we have and return it. 1113 */ 1114static struct td_sched * 1115tdq_choose(struct tdq *tdq) 1116{ 1117 struct td_sched *ts; 1118 1119 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 1120 ts = runq_choose(&tdq->tdq_realtime); 1121 if (ts != NULL) 1122 return (ts); 1123 ts = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx); 1124 if (ts != NULL) { 1125 KASSERT(ts->ts_thread->td_priority >= PRI_MIN_TIMESHARE, 1126 ("tdq_choose: Invalid priority on timeshare queue %d", 1127 ts->ts_thread->td_priority)); 1128 return (ts); 1129 } 1130 1131 ts = runq_choose(&tdq->tdq_idle); 1132 if (ts != NULL) { 1133 KASSERT(ts->ts_thread->td_priority >= PRI_MIN_IDLE, 1134 ("tdq_choose: Invalid priority on idle queue %d", 1135 ts->ts_thread->td_priority)); 1136 return (ts); 1137 } 1138 1139 return (NULL); 1140} 1141 1142/* 1143 * Initialize a thread queue. 1144 */ 1145static void 1146tdq_setup(struct tdq *tdq) 1147{ 1148 1149 snprintf(tdq->tdq_name, sizeof(tdq->tdq_name), 1150 "sched lock %d", (int)TDQ_ID(tdq)); 1151 mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock", 1152 MTX_SPIN | MTX_RECURSE); 1153 runq_init(&tdq->tdq_realtime); 1154 runq_init(&tdq->tdq_timeshare); 1155 runq_init(&tdq->tdq_idle); 1156 tdq->tdq_load = 0; 1157} 1158 1159/* 1160 * Setup the thread queues and initialize the topology based on MD 1161 * information. 1162 */ 1163static void 1164sched_setup(void *dummy) 1165{ 1166 struct tdq *tdq; 1167#ifdef SMP 1168 int balance_groups; 1169 int i; 1170 1171 balance_groups = 0; 1172 /* 1173 * Initialize the tdqs. 1174 */ 1175 for (i = 0; i < MAXCPU; i++) { 1176 tdq = &tdq_cpu[i]; 1177 tdq_setup(&tdq_cpu[i]); 1178 } 1179 if (smp_topology == NULL) { 1180 struct tdq_group *tdg; 1181 int cpus; 1182 1183 for (cpus = 0, i = 0; i < MAXCPU; i++) { 1184 if (CPU_ABSENT(i)) 1185 continue; 1186 tdq = &tdq_cpu[i]; 1187 tdg = &tdq_groups[cpus]; 1188 /* 1189 * Setup a tdq group with one member. 1190 */ 1191 tdq->tdq_transferable = 0; 1192 tdq->tdq_group = tdg; 1193 tdg->tdg_cpus = 1; 1194 tdg->tdg_idlemask = 0; 1195 tdg->tdg_cpumask = tdg->tdg_mask = 1 << i; 1196 tdg->tdg_load = 0; 1197 tdg->tdg_transferable = 0; 1198 LIST_INIT(&tdg->tdg_members); 1199 LIST_INSERT_HEAD(&tdg->tdg_members, tdq, tdq_siblings); 1200 cpus++; 1201 } 1202 tdg_maxid = cpus - 1; 1203 } else { 1204 struct tdq_group *tdg; 1205 struct cpu_group *cg; 1206 int j; 1207 1208 topology = 1; 1209 for (i = 0; i < smp_topology->ct_count; i++) { 1210 cg = &smp_topology->ct_group[i]; 1211 tdg = &tdq_groups[i]; 1212 /* 1213 * Initialize the group. 1214 */ 1215 tdg->tdg_idlemask = 0; 1216 tdg->tdg_load = 0; 1217 tdg->tdg_transferable = 0; 1218 tdg->tdg_cpus = cg->cg_count; 1219 tdg->tdg_cpumask = cg->cg_mask; 1220 LIST_INIT(&tdg->tdg_members); 1221 /* 1222 * Find all of the group members and add them. 1223 */ 1224 for (j = 0; j < MAXCPU; j++) { 1225 if ((cg->cg_mask & (1 << j)) != 0) { 1226 if (tdg->tdg_mask == 0) 1227 tdg->tdg_mask = 1 << j; 1228 tdq_cpu[j].tdq_transferable = 0; 1229 tdq_cpu[j].tdq_group = tdg; 1230 LIST_INSERT_HEAD(&tdg->tdg_members, 1231 &tdq_cpu[j], tdq_siblings); 1232 } 1233 } 1234 if (tdg->tdg_cpus > 1) 1235 balance_groups = 1; 1236 } 1237 tdg_maxid = smp_topology->ct_count - 1; 1238 } 1239 /* 1240 * Initialize long-term cpu balancing algorithm. 1241 */ 1242 callout_init(&balco, CALLOUT_MPSAFE); 1243 callout_init(&gbalco, CALLOUT_MPSAFE); 1244 sched_balance(NULL); 1245 if (balance_groups) 1246 sched_balance_groups(NULL); 1247 1248#else 1249 tdq_setup(TDQ_SELF()); 1250#endif 1251 /* 1252 * To avoid divide-by-zero, we set realstathz a dummy value 1253 * in case which sched_clock() called before sched_initticks(). 1254 */ 1255 realstathz = hz; 1256 sched_slice = (realstathz/10); /* ~100ms */ 1257 tickincr = 1 << SCHED_TICK_SHIFT; 1258 1259 /* Add thread0's load since it's running. */ 1260 tdq = TDQ_SELF(); 1261 TDQ_LOCK(tdq); 1262 tdq_load_add(tdq, &td_sched0); 1263 TDQ_UNLOCK(tdq); 1264} 1265 1266/* 1267 * This routine determines the tickincr after stathz and hz are setup. 1268 */ 1269/* ARGSUSED */ 1270static void 1271sched_initticks(void *dummy) 1272{ 1273 int incr; 1274 1275 realstathz = stathz ? stathz : hz; 1276 sched_slice = (realstathz/10); /* ~100ms */ 1277 1278 /* 1279 * tickincr is shifted out by 10 to avoid rounding errors due to 1280 * hz not being evenly divisible by stathz on all platforms. 1281 */ 1282 incr = (hz << SCHED_TICK_SHIFT) / realstathz; 1283 /* 1284 * This does not work for values of stathz that are more than 1285 * 1 << SCHED_TICK_SHIFT * hz. In practice this does not happen. 1286 */ 1287 if (incr == 0) 1288 incr = 1; 1289 tickincr = incr; 1290#ifdef SMP 1291 affinity = SCHED_AFFINITY_DEFAULT; 1292#endif 1293} 1294 1295 1296/* 1297 * This is the core of the interactivity algorithm. Determines a score based 1298 * on past behavior. It is the ratio of sleep time to run time scaled to 1299 * a [0, 100] integer. This is the voluntary sleep time of a process, which 1300 * differs from the cpu usage because it does not account for time spent 1301 * waiting on a run-queue. Would be prettier if we had floating point. 1302 */ 1303static int 1304sched_interact_score(struct thread *td) 1305{ 1306 struct td_sched *ts; 1307 int div; 1308 1309 ts = td->td_sched; 1310 /* 1311 * The score is only needed if this is likely to be an interactive 1312 * task. Don't go through the expense of computing it if there's 1313 * no chance. 1314 */ 1315 if (sched_interact <= SCHED_INTERACT_HALF && 1316 ts->ts_runtime >= ts->ts_slptime) 1317 return (SCHED_INTERACT_HALF); 1318 1319 if (ts->ts_runtime > ts->ts_slptime) { 1320 div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF); 1321 return (SCHED_INTERACT_HALF + 1322 (SCHED_INTERACT_HALF - (ts->ts_slptime / div))); 1323 } 1324 if (ts->ts_slptime > ts->ts_runtime) { 1325 div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF); 1326 return (ts->ts_runtime / div); 1327 } 1328 /* runtime == slptime */ 1329 if (ts->ts_runtime) 1330 return (SCHED_INTERACT_HALF); 1331 1332 /* 1333 * This can happen if slptime and runtime are 0. 1334 */ 1335 return (0); 1336 1337} 1338 1339/* 1340 * Scale the scheduling priority according to the "interactivity" of this 1341 * process. 1342 */ 1343static void 1344sched_priority(struct thread *td) 1345{ 1346 int score; 1347 int pri; 1348 1349 if (td->td_pri_class != PRI_TIMESHARE) 1350 return; 1351 /* 1352 * If the score is interactive we place the thread in the realtime 1353 * queue with a priority that is less than kernel and interrupt 1354 * priorities. These threads are not subject to nice restrictions. 1355 * 1356 * Scores greater than this are placed on the normal timeshare queue 1357 * where the priority is partially decided by the most recent cpu 1358 * utilization and the rest is decided by nice value. 1359 */ 1360 score = sched_interact_score(td); 1361 if (score < sched_interact) { 1362 pri = PRI_MIN_REALTIME; 1363 pri += ((PRI_MAX_REALTIME - PRI_MIN_REALTIME) / sched_interact) 1364 * score; 1365 KASSERT(pri >= PRI_MIN_REALTIME && pri <= PRI_MAX_REALTIME, 1366 ("sched_priority: invalid interactive priority %d score %d", 1367 pri, score)); 1368 } else { 1369 pri = SCHED_PRI_MIN; 1370 if (td->td_sched->ts_ticks) 1371 pri += SCHED_PRI_TICKS(td->td_sched); 1372 pri += SCHED_PRI_NICE(td->td_proc->p_nice); 1373 KASSERT(pri >= PRI_MIN_TIMESHARE && pri <= PRI_MAX_TIMESHARE, 1374 ("sched_priority: invalid priority %d: nice %d, " 1375 "ticks %d ftick %d ltick %d tick pri %d", 1376 pri, td->td_proc->p_nice, td->td_sched->ts_ticks, 1377 td->td_sched->ts_ftick, td->td_sched->ts_ltick, 1378 SCHED_PRI_TICKS(td->td_sched))); 1379 } 1380 sched_user_prio(td, pri); 1381 1382 return; 1383} 1384 1385/* 1386 * This routine enforces a maximum limit on the amount of scheduling history 1387 * kept. It is called after either the slptime or runtime is adjusted. This 1388 * function is ugly due to integer math. 1389 */ 1390static void 1391sched_interact_update(struct thread *td) 1392{ 1393 struct td_sched *ts; 1394 u_int sum; 1395 1396 ts = td->td_sched; 1397 sum = ts->ts_runtime + ts->ts_slptime; 1398 if (sum < SCHED_SLP_RUN_MAX) 1399 return; 1400 /* 1401 * This only happens from two places: 1402 * 1) We have added an unusual amount of run time from fork_exit. 1403 * 2) We have added an unusual amount of sleep time from sched_sleep(). 1404 */ 1405 if (sum > SCHED_SLP_RUN_MAX * 2) { 1406 if (ts->ts_runtime > ts->ts_slptime) { 1407 ts->ts_runtime = SCHED_SLP_RUN_MAX; 1408 ts->ts_slptime = 1; 1409 } else { 1410 ts->ts_slptime = SCHED_SLP_RUN_MAX; 1411 ts->ts_runtime = 1; 1412 } 1413 return; 1414 } 1415 /* 1416 * If we have exceeded by more than 1/5th then the algorithm below 1417 * will not bring us back into range. Dividing by two here forces 1418 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX] 1419 */ 1420 if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) { 1421 ts->ts_runtime /= 2; 1422 ts->ts_slptime /= 2; 1423 return; 1424 } 1425 ts->ts_runtime = (ts->ts_runtime / 5) * 4; 1426 ts->ts_slptime = (ts->ts_slptime / 5) * 4; 1427} 1428 1429/* 1430 * Scale back the interactivity history when a child thread is created. The 1431 * history is inherited from the parent but the thread may behave totally 1432 * differently. For example, a shell spawning a compiler process. We want 1433 * to learn that the compiler is behaving badly very quickly. 1434 */ 1435static void 1436sched_interact_fork(struct thread *td) 1437{ 1438 int ratio; 1439 int sum; 1440 1441 sum = td->td_sched->ts_runtime + td->td_sched->ts_slptime; 1442 if (sum > SCHED_SLP_RUN_FORK) { 1443 ratio = sum / SCHED_SLP_RUN_FORK; 1444 td->td_sched->ts_runtime /= ratio; 1445 td->td_sched->ts_slptime /= ratio; 1446 } 1447} 1448 1449/* 1450 * Called from proc0_init() to setup the scheduler fields. 1451 */ 1452void 1453schedinit(void) 1454{ 1455 1456 /* 1457 * Set up the scheduler specific parts of proc0. 1458 */ 1459 proc0.p_sched = NULL; /* XXX */ 1460 thread0.td_sched = &td_sched0; 1461 thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF()); 1462 td_sched0.ts_ltick = ticks; 1463 td_sched0.ts_ftick = ticks; 1464 td_sched0.ts_thread = &thread0; 1465} 1466 1467/* 1468 * This is only somewhat accurate since given many processes of the same 1469 * priority they will switch when their slices run out, which will be 1470 * at most sched_slice stathz ticks. 1471 */ 1472int 1473sched_rr_interval(void) 1474{ 1475 1476 /* Convert sched_slice to hz */ 1477 return (hz/(realstathz/sched_slice)); 1478} 1479 1480/* 1481 * Update the percent cpu tracking information when it is requested or 1482 * the total history exceeds the maximum. We keep a sliding history of 1483 * tick counts that slowly decays. This is less precise than the 4BSD 1484 * mechanism since it happens with less regular and frequent events. 1485 */ 1486static void 1487sched_pctcpu_update(struct td_sched *ts) 1488{ 1489 1490 if (ts->ts_ticks == 0) 1491 return; 1492 if (ticks - (hz / 10) < ts->ts_ltick && 1493 SCHED_TICK_TOTAL(ts) < SCHED_TICK_MAX) 1494 return; 1495 /* 1496 * Adjust counters and watermark for pctcpu calc. 1497 */ 1498 if (ts->ts_ltick > ticks - SCHED_TICK_TARG) 1499 ts->ts_ticks = (ts->ts_ticks / (ticks - ts->ts_ftick)) * 1500 SCHED_TICK_TARG; 1501 else 1502 ts->ts_ticks = 0; 1503 ts->ts_ltick = ticks; 1504 ts->ts_ftick = ts->ts_ltick - SCHED_TICK_TARG; 1505} 1506 1507/* 1508 * Adjust the priority of a thread. Move it to the appropriate run-queue 1509 * if necessary. This is the back-end for several priority related 1510 * functions. 1511 */ 1512static void 1513sched_thread_priority(struct thread *td, u_char prio) 1514{ 1515 struct td_sched *ts; 1516 1517 CTR6(KTR_SCHED, "sched_prio: %p(%s) prio %d newprio %d by %p(%s)", 1518 td, td->td_proc->p_comm, td->td_priority, prio, curthread, 1519 curthread->td_proc->p_comm); 1520 ts = td->td_sched; 1521 THREAD_LOCK_ASSERT(td, MA_OWNED); 1522 if (td->td_priority == prio) 1523 return; 1524 1525 if (TD_ON_RUNQ(td) && prio < td->td_priority) { 1526 /* 1527 * If the priority has been elevated due to priority 1528 * propagation, we may have to move ourselves to a new 1529 * queue. This could be optimized to not re-add in some 1530 * cases. 1531 */ 1532 sched_rem(td); 1533 td->td_priority = prio; 1534 sched_add(td, SRQ_BORROWING); 1535 } else { 1536#ifdef SMP 1537 struct tdq *tdq; 1538 1539 tdq = TDQ_CPU(ts->ts_cpu); 1540 if (prio < tdq->tdq_lowpri) 1541 tdq->tdq_lowpri = prio; 1542#endif 1543 td->td_priority = prio; 1544 } 1545} 1546 1547/* 1548 * Update a thread's priority when it is lent another thread's 1549 * priority. 1550 */ 1551void 1552sched_lend_prio(struct thread *td, u_char prio) 1553{ 1554 1555 td->td_flags |= TDF_BORROWING; 1556 sched_thread_priority(td, prio); 1557} 1558 1559/* 1560 * Restore a thread's priority when priority propagation is 1561 * over. The prio argument is the minimum priority the thread 1562 * needs to have to satisfy other possible priority lending 1563 * requests. If the thread's regular priority is less 1564 * important than prio, the thread will keep a priority boost 1565 * of prio. 1566 */ 1567void 1568sched_unlend_prio(struct thread *td, u_char prio) 1569{ 1570 u_char base_pri; 1571 1572 if (td->td_base_pri >= PRI_MIN_TIMESHARE && 1573 td->td_base_pri <= PRI_MAX_TIMESHARE) 1574 base_pri = td->td_user_pri; 1575 else 1576 base_pri = td->td_base_pri; 1577 if (prio >= base_pri) { 1578 td->td_flags &= ~TDF_BORROWING; 1579 sched_thread_priority(td, base_pri); 1580 } else 1581 sched_lend_prio(td, prio); 1582} 1583 1584/* 1585 * Standard entry for setting the priority to an absolute value. 1586 */ 1587void 1588sched_prio(struct thread *td, u_char prio) 1589{ 1590 u_char oldprio; 1591 1592 /* First, update the base priority. */ 1593 td->td_base_pri = prio; 1594 1595 /* 1596 * If the thread is borrowing another thread's priority, don't 1597 * ever lower the priority. 1598 */ 1599 if (td->td_flags & TDF_BORROWING && td->td_priority < prio) 1600 return; 1601 1602 /* Change the real priority. */ 1603 oldprio = td->td_priority; 1604 sched_thread_priority(td, prio); 1605 1606 /* 1607 * If the thread is on a turnstile, then let the turnstile update 1608 * its state. 1609 */ 1610 if (TD_ON_LOCK(td) && oldprio != prio) 1611 turnstile_adjust(td, oldprio); 1612} 1613 1614/* 1615 * Set the base user priority, does not effect current running priority. 1616 */ 1617void 1618sched_user_prio(struct thread *td, u_char prio) 1619{ 1620 u_char oldprio; 1621 1622 td->td_base_user_pri = prio; 1623 if (td->td_flags & TDF_UBORROWING && td->td_user_pri <= prio) 1624 return; 1625 oldprio = td->td_user_pri; 1626 td->td_user_pri = prio; 1627 1628 if (TD_ON_UPILOCK(td) && oldprio != prio) 1629 umtx_pi_adjust(td, oldprio); 1630} 1631 1632void 1633sched_lend_user_prio(struct thread *td, u_char prio) 1634{ 1635 u_char oldprio; 1636 1637 td->td_flags |= TDF_UBORROWING; 1638 1639 oldprio = td->td_user_pri; 1640 td->td_user_pri = prio; 1641 1642 if (TD_ON_UPILOCK(td) && oldprio != prio) 1643 umtx_pi_adjust(td, oldprio); 1644} 1645 1646void 1647sched_unlend_user_prio(struct thread *td, u_char prio) 1648{ 1649 u_char base_pri; 1650 1651 base_pri = td->td_base_user_pri; 1652 if (prio >= base_pri) { 1653 td->td_flags &= ~TDF_UBORROWING; 1654 sched_user_prio(td, base_pri); 1655 } else 1656 sched_lend_user_prio(td, prio); 1657} 1658 1659/* 1660 * Block a thread for switching. Similar to thread_block() but does not 1661 * bump the spin count. 1662 */ 1663static inline struct mtx * 1664thread_block_switch(struct thread *td) 1665{ 1666 struct mtx *lock; 1667 1668 THREAD_LOCK_ASSERT(td, MA_OWNED); 1669 lock = td->td_lock; 1670 td->td_lock = &blocked_lock; 1671 mtx_unlock_spin(lock); 1672 1673 return (lock); 1674} 1675 1676/* 1677 * Release a thread that was blocked with thread_block_switch(). 1678 */ 1679static inline void 1680thread_unblock_switch(struct thread *td, struct mtx *mtx) 1681{ 1682 atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock, 1683 (uintptr_t)mtx); 1684} 1685 1686/* 1687 * Switch threads. This function has to handle threads coming in while 1688 * blocked for some reason, running, or idle. It also must deal with 1689 * migrating a thread from one queue to another as running threads may 1690 * be assigned elsewhere via binding. 1691 */ 1692void 1693sched_switch(struct thread *td, struct thread *newtd, int flags) 1694{ 1695 struct tdq *tdq; 1696 struct td_sched *ts; 1697 struct mtx *mtx; 1698 int cpuid; 1699 1700 THREAD_LOCK_ASSERT(td, MA_OWNED); 1701 1702 cpuid = PCPU_GET(cpuid); 1703 tdq = TDQ_CPU(cpuid); 1704 ts = td->td_sched; 1705 mtx = TDQ_LOCKPTR(tdq); 1706#ifdef SMP 1707 ts->ts_rltick = ticks; 1708 if (newtd && newtd->td_priority < tdq->tdq_lowpri) 1709 tdq->tdq_lowpri = newtd->td_priority; 1710#endif 1711 td->td_lastcpu = td->td_oncpu; 1712 td->td_oncpu = NOCPU; 1713 td->td_flags &= ~TDF_NEEDRESCHED; 1714 td->td_owepreempt = 0; 1715 /* 1716 * The lock pointer in an idle thread should never change. Reset it 1717 * to CAN_RUN as well. 1718 */ 1719 if (TD_IS_IDLETHREAD(td)) { 1720 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1721 TD_SET_CAN_RUN(td); 1722 } else if (TD_IS_RUNNING(td)) { 1723 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1724 /* Remove our load so the selection algorithm is not biased. */ 1725 tdq_load_rem(tdq, ts); 1726 sched_add(td, (flags & SW_PREEMPT) ? 1727 SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED : 1728 SRQ_OURSELF|SRQ_YIELDING); 1729 /* 1730 * When migrating we return from sched_add with an extra 1731 * spinlock nesting, the tdq locked, and a blocked thread. 1732 * This is to optimize out an extra block/unblock cycle here. 1733 */ 1734 if (ts->ts_cpu != cpuid) { 1735 mtx = TDQ_LOCKPTR(TDQ_CPU(ts->ts_cpu)); 1736 mtx_unlock_spin(mtx); 1737 TDQ_LOCK(tdq); 1738 spinlock_exit(); 1739 } 1740 } else { 1741 /* This thread must be going to sleep. */ 1742 TDQ_LOCK(tdq); 1743 mtx = thread_block_switch(td); 1744 tdq_load_rem(tdq, ts); 1745 } 1746 /* 1747 * We enter here with the thread blocked and assigned to the 1748 * appropriate cpu run-queue or sleep-queue and with the current 1749 * thread-queue locked. 1750 */ 1751 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED); 1752 /* 1753 * If KSE assigned a new thread just add it here and pick the best one. 1754 */ 1755 if (newtd != NULL) { 1756 /* XXX This is bogus. What if the thread is locked elsewhere? */ 1757 td->td_lock = TDQ_LOCKPTR(tdq); 1758 td->td_sched->ts_cpu = cpuid; 1759 tdq_add(tdq, td, SRQ_YIELDING); 1760 } 1761 newtd = choosethread(); 1762 /* 1763 * Call the MD code to switch contexts if necessary. 1764 */ 1765 if (td != newtd) { 1766#ifdef HWPMC_HOOKS 1767 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1768 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT); 1769#endif 1770 cpu_switch(td, newtd, mtx); 1771 /* 1772 * We may return from cpu_switch on a different cpu. However, 1773 * we always return with td_lock pointing to the current cpu's 1774 * run queue lock. 1775 */ 1776 cpuid = PCPU_GET(cpuid); 1777 tdq = TDQ_CPU(cpuid); 1778 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)td; 1779#ifdef HWPMC_HOOKS 1780 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1781 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN); 1782#endif 1783 } else 1784 thread_unblock_switch(td, mtx); 1785 /* 1786 * Assert that all went well and return. 1787 */ 1788#ifdef SMP 1789 /* We should always get here with the lowest priority td possible */ 1790 tdq->tdq_lowpri = td->td_priority; 1791#endif 1792 TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED); 1793 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1794 td->td_oncpu = cpuid; 1795} 1796 1797/* 1798 * Adjust thread priorities as a result of a nice request. 1799 */ 1800void 1801sched_nice(struct proc *p, int nice) 1802{ 1803 struct thread *td; 1804 1805 PROC_LOCK_ASSERT(p, MA_OWNED); 1806 PROC_SLOCK_ASSERT(p, MA_OWNED); 1807 1808 p->p_nice = nice; 1809 FOREACH_THREAD_IN_PROC(p, td) { 1810 thread_lock(td); 1811 sched_priority(td); 1812 sched_prio(td, td->td_base_user_pri); 1813 thread_unlock(td); 1814 } 1815} 1816 1817/* 1818 * Record the sleep time for the interactivity scorer. 1819 */ 1820void 1821sched_sleep(struct thread *td) 1822{ 1823 1824 THREAD_LOCK_ASSERT(td, MA_OWNED); 1825 1826 td->td_sched->ts_slptick = ticks; 1827} 1828 1829/* 1830 * Schedule a thread to resume execution and record how long it voluntarily 1831 * slept. We also update the pctcpu, interactivity, and priority. 1832 */ 1833void 1834sched_wakeup(struct thread *td) 1835{ 1836 struct td_sched *ts; 1837 int slptick; 1838 1839 THREAD_LOCK_ASSERT(td, MA_OWNED); 1840 ts = td->td_sched; 1841 /* 1842 * If we slept for more than a tick update our interactivity and 1843 * priority. 1844 */ 1845 slptick = ts->ts_slptick; 1846 ts->ts_slptick = 0; 1847 if (slptick && slptick != ticks) { 1848 u_int hzticks; 1849 1850 hzticks = (ticks - slptick) << SCHED_TICK_SHIFT; 1851 ts->ts_slptime += hzticks; 1852 sched_interact_update(td); 1853 sched_pctcpu_update(ts); 1854 sched_priority(td); 1855 } 1856 /* Reset the slice value after we sleep. */ 1857 ts->ts_slice = sched_slice; 1858 sched_add(td, SRQ_BORING); 1859} 1860 1861/* 1862 * Penalize the parent for creating a new child and initialize the child's 1863 * priority. 1864 */ 1865void 1866sched_fork(struct thread *td, struct thread *child) 1867{ 1868 THREAD_LOCK_ASSERT(td, MA_OWNED); 1869 sched_fork_thread(td, child); 1870 /* 1871 * Penalize the parent and child for forking. 1872 */ 1873 sched_interact_fork(child); 1874 sched_priority(child); 1875 td->td_sched->ts_runtime += tickincr; 1876 sched_interact_update(td); 1877 sched_priority(td); 1878} 1879 1880/* 1881 * Fork a new thread, may be within the same process. 1882 */ 1883void 1884sched_fork_thread(struct thread *td, struct thread *child) 1885{ 1886 struct td_sched *ts; 1887 struct td_sched *ts2; 1888 1889 /* 1890 * Initialize child. 1891 */ 1892 THREAD_LOCK_ASSERT(td, MA_OWNED); 1893 sched_newthread(child); 1894 child->td_lock = TDQ_LOCKPTR(TDQ_SELF()); 1895 ts = td->td_sched; 1896 ts2 = child->td_sched; 1897 ts2->ts_cpu = ts->ts_cpu; 1898 ts2->ts_runq = NULL; 1899 /* 1900 * Grab our parents cpu estimation information and priority. 1901 */ 1902 ts2->ts_ticks = ts->ts_ticks; 1903 ts2->ts_ltick = ts->ts_ltick; 1904 ts2->ts_ftick = ts->ts_ftick; 1905 child->td_user_pri = td->td_user_pri; 1906 child->td_base_user_pri = td->td_base_user_pri; 1907 /* 1908 * And update interactivity score. 1909 */ 1910 ts2->ts_slptime = ts->ts_slptime; 1911 ts2->ts_runtime = ts->ts_runtime; 1912 ts2->ts_slice = 1; /* Attempt to quickly learn interactivity. */ 1913} 1914 1915/* 1916 * Adjust the priority class of a thread. 1917 */ 1918void 1919sched_class(struct thread *td, int class) 1920{ 1921 1922 THREAD_LOCK_ASSERT(td, MA_OWNED); 1923 if (td->td_pri_class == class) 1924 return; 1925 1926#ifdef SMP 1927 /* 1928 * On SMP if we're on the RUNQ we must adjust the transferable 1929 * count because could be changing to or from an interrupt 1930 * class. 1931 */ 1932 if (TD_ON_RUNQ(td)) { 1933 struct tdq *tdq; 1934 1935 tdq = TDQ_CPU(td->td_sched->ts_cpu); 1936 if (THREAD_CAN_MIGRATE(td)) { 1937 tdq->tdq_transferable--; 1938 tdq->tdq_group->tdg_transferable--; 1939 } 1940 td->td_pri_class = class; 1941 if (THREAD_CAN_MIGRATE(td)) { 1942 tdq->tdq_transferable++; 1943 tdq->tdq_group->tdg_transferable++; 1944 } 1945 } 1946#endif 1947 td->td_pri_class = class; 1948} 1949 1950/* 1951 * Return some of the child's priority and interactivity to the parent. 1952 */ 1953void 1954sched_exit(struct proc *p, struct thread *child) 1955{ 1956 struct thread *td; 1957 1958 CTR3(KTR_SCHED, "sched_exit: %p(%s) prio %d", 1959 child, child->td_proc->p_comm, child->td_priority); 1960 1961 PROC_SLOCK_ASSERT(p, MA_OWNED); 1962 td = FIRST_THREAD_IN_PROC(p); 1963 sched_exit_thread(td, child); 1964} 1965 1966/* 1967 * Penalize another thread for the time spent on this one. This helps to 1968 * worsen the priority and interactivity of processes which schedule batch 1969 * jobs such as make. This has little effect on the make process itself but 1970 * causes new processes spawned by it to receive worse scores immediately. 1971 */ 1972void 1973sched_exit_thread(struct thread *td, struct thread *child) 1974{ 1975 1976 CTR3(KTR_SCHED, "sched_exit_thread: %p(%s) prio %d", 1977 child, child->td_proc->p_comm, child->td_priority); 1978 1979#ifdef KSE 1980 /* 1981 * KSE forks and exits so often that this penalty causes short-lived 1982 * threads to always be non-interactive. This causes mozilla to 1983 * crawl under load. 1984 */ 1985 if ((td->td_pflags & TDP_SA) && td->td_proc == child->td_proc) 1986 return; 1987#endif 1988 /* 1989 * Give the child's runtime to the parent without returning the 1990 * sleep time as a penalty to the parent. This causes shells that 1991 * launch expensive things to mark their children as expensive. 1992 */ 1993 thread_lock(td); 1994 td->td_sched->ts_runtime += child->td_sched->ts_runtime; 1995 sched_interact_update(td); 1996 sched_priority(td); 1997 thread_unlock(td); 1998} 1999 2000/* 2001 * Fix priorities on return to user-space. Priorities may be elevated due 2002 * to static priorities in msleep() or similar. 2003 */ 2004void 2005sched_userret(struct thread *td) 2006{ 2007 /* 2008 * XXX we cheat slightly on the locking here to avoid locking in 2009 * the usual case. Setting td_priority here is essentially an 2010 * incomplete workaround for not setting it properly elsewhere. 2011 * Now that some interrupt handlers are threads, not setting it 2012 * properly elsewhere can clobber it in the window between setting 2013 * it here and returning to user mode, so don't waste time setting 2014 * it perfectly here. 2015 */ 2016 KASSERT((td->td_flags & TDF_BORROWING) == 0, 2017 ("thread with borrowed priority returning to userland")); 2018 if (td->td_priority != td->td_user_pri) { 2019 thread_lock(td); 2020 td->td_priority = td->td_user_pri; 2021 td->td_base_pri = td->td_user_pri; 2022 thread_unlock(td); 2023 } 2024} 2025 2026/* 2027 * Handle a stathz tick. This is really only relevant for timeshare 2028 * threads. 2029 */ 2030void 2031sched_clock(struct thread *td) 2032{ 2033 struct tdq *tdq; 2034 struct td_sched *ts; 2035 2036 THREAD_LOCK_ASSERT(td, MA_OWNED); 2037 tdq = TDQ_SELF(); 2038 /* 2039 * Advance the insert index once for each tick to ensure that all 2040 * threads get a chance to run. 2041 */ 2042 if (tdq->tdq_idx == tdq->tdq_ridx) { 2043 tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS; 2044 if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx])) 2045 tdq->tdq_ridx = tdq->tdq_idx; 2046 } 2047 ts = td->td_sched; 2048 /* 2049 * We only do slicing code for TIMESHARE threads. 2050 */ 2051 if (td->td_pri_class != PRI_TIMESHARE) 2052 return; 2053 /* 2054 * We used a tick; charge it to the thread so that we can compute our 2055 * interactivity. 2056 */ 2057 td->td_sched->ts_runtime += tickincr; 2058 sched_interact_update(td); 2059 /* 2060 * We used up one time slice. 2061 */ 2062 if (--ts->ts_slice > 0) 2063 return; 2064 /* 2065 * We're out of time, recompute priorities and requeue. 2066 */ 2067 sched_priority(td); 2068 td->td_flags |= TDF_NEEDRESCHED; 2069} 2070 2071/* 2072 * Called once per hz tick. Used for cpu utilization information. This 2073 * is easier than trying to scale based on stathz. 2074 */ 2075void 2076sched_tick(void) 2077{ 2078 struct td_sched *ts; 2079 2080 ts = curthread->td_sched; 2081 /* Adjust ticks for pctcpu */ 2082 ts->ts_ticks += 1 << SCHED_TICK_SHIFT; 2083 ts->ts_ltick = ticks; 2084 /* 2085 * Update if we've exceeded our desired tick threshhold by over one 2086 * second. 2087 */ 2088 if (ts->ts_ftick + SCHED_TICK_MAX < ts->ts_ltick) 2089 sched_pctcpu_update(ts); 2090} 2091 2092/* 2093 * Return whether the current CPU has runnable tasks. Used for in-kernel 2094 * cooperative idle threads. 2095 */ 2096int 2097sched_runnable(void) 2098{ 2099 struct tdq *tdq; 2100 int load; 2101 2102 load = 1; 2103 2104 tdq = TDQ_SELF(); 2105 if ((curthread->td_flags & TDF_IDLETD) != 0) { 2106 if (tdq->tdq_load > 0) 2107 goto out; 2108 } else 2109 if (tdq->tdq_load - 1 > 0) 2110 goto out; 2111 load = 0; 2112out: 2113 return (load); 2114} 2115 2116/* 2117 * Choose the highest priority thread to run. The thread is removed from 2118 * the run-queue while running however the load remains. For SMP we set 2119 * the tdq in the global idle bitmask if it idles here. 2120 */ 2121struct thread * 2122sched_choose(void) 2123{ 2124#ifdef SMP 2125 struct tdq_group *tdg; 2126#endif 2127 struct td_sched *ts; 2128 struct tdq *tdq; 2129 2130 tdq = TDQ_SELF(); 2131 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2132 ts = tdq_choose(tdq); 2133 if (ts) { 2134 tdq_runq_rem(tdq, ts); 2135 return (ts->ts_thread); 2136 } 2137#ifdef SMP 2138 /* 2139 * We only set the idled bit when all of the cpus in the group are 2140 * idle. Otherwise we could get into a situation where a thread bounces 2141 * back and forth between two idle cores on seperate physical CPUs. 2142 */ 2143 tdg = tdq->tdq_group; 2144 tdg->tdg_idlemask |= PCPU_GET(cpumask); 2145 if (tdg->tdg_idlemask == tdg->tdg_cpumask) 2146 atomic_set_int(&tdq_idle, tdg->tdg_mask); 2147 tdq->tdq_lowpri = PRI_MAX_IDLE; 2148#endif 2149 return (PCPU_GET(idlethread)); 2150} 2151 2152/* 2153 * Set owepreempt if necessary. Preemption never happens directly in ULE, 2154 * we always request it once we exit a critical section. 2155 */ 2156static inline void 2157sched_setpreempt(struct thread *td) 2158{ 2159 struct thread *ctd; 2160 int cpri; 2161 int pri; 2162 2163 ctd = curthread; 2164 pri = td->td_priority; 2165 cpri = ctd->td_priority; 2166 if (td->td_priority < ctd->td_priority) 2167 curthread->td_flags |= TDF_NEEDRESCHED; 2168 if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd)) 2169 return; 2170 /* 2171 * Always preempt IDLE threads. Otherwise only if the preempting 2172 * thread is an ithread. 2173 */ 2174 if (pri > preempt_thresh && cpri < PRI_MIN_IDLE) 2175 return; 2176 ctd->td_owepreempt = 1; 2177 return; 2178} 2179 2180/* 2181 * Add a thread to a thread queue. Initializes priority, slice, runq, and 2182 * add it to the appropriate queue. This is the internal function called 2183 * when the tdq is predetermined. 2184 */ 2185void 2186tdq_add(struct tdq *tdq, struct thread *td, int flags) 2187{ 2188 struct td_sched *ts; 2189 int class; 2190#ifdef SMP 2191 int cpumask; 2192#endif 2193 2194 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2195 KASSERT((td->td_inhibitors == 0), 2196 ("sched_add: trying to run inhibited thread")); 2197 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)), 2198 ("sched_add: bad thread state")); 2199 KASSERT(td->td_proc->p_sflag & PS_INMEM, 2200 ("sched_add: process swapped out")); 2201 2202 ts = td->td_sched; 2203 class = PRI_BASE(td->td_pri_class); 2204 TD_SET_RUNQ(td); 2205 if (ts->ts_slice == 0) 2206 ts->ts_slice = sched_slice; 2207 /* 2208 * Pick the run queue based on priority. 2209 */ 2210 if (td->td_priority <= PRI_MAX_REALTIME) 2211 ts->ts_runq = &tdq->tdq_realtime; 2212 else if (td->td_priority <= PRI_MAX_TIMESHARE) 2213 ts->ts_runq = &tdq->tdq_timeshare; 2214 else 2215 ts->ts_runq = &tdq->tdq_idle; 2216#ifdef SMP 2217 cpumask = 1 << ts->ts_cpu; 2218 /* 2219 * If we had been idle, clear our bit in the group and potentially 2220 * the global bitmap. 2221 */ 2222 if ((class != PRI_IDLE && class != PRI_ITHD) && 2223 (tdq->tdq_group->tdg_idlemask & cpumask) != 0) { 2224 /* 2225 * Check to see if our group is unidling, and if so, remove it 2226 * from the global idle mask. 2227 */ 2228 if (tdq->tdq_group->tdg_idlemask == 2229 tdq->tdq_group->tdg_cpumask) 2230 atomic_clear_int(&tdq_idle, tdq->tdq_group->tdg_mask); 2231 /* 2232 * Now remove ourselves from the group specific idle mask. 2233 */ 2234 tdq->tdq_group->tdg_idlemask &= ~cpumask; 2235 } 2236 if (td->td_priority < tdq->tdq_lowpri) 2237 tdq->tdq_lowpri = td->td_priority; 2238#endif 2239 tdq_runq_add(tdq, ts, flags); 2240 tdq_load_add(tdq, ts); 2241} 2242 2243/* 2244 * Select the target thread queue and add a thread to it. Request 2245 * preemption or IPI a remote processor if required. 2246 */ 2247void 2248sched_add(struct thread *td, int flags) 2249{ 2250 struct td_sched *ts; 2251 struct tdq *tdq; 2252#ifdef SMP 2253 int cpuid; 2254 int cpu; 2255#endif 2256 CTR5(KTR_SCHED, "sched_add: %p(%s) prio %d by %p(%s)", 2257 td, td->td_proc->p_comm, td->td_priority, curthread, 2258 curthread->td_proc->p_comm); 2259 THREAD_LOCK_ASSERT(td, MA_OWNED); 2260 ts = td->td_sched; 2261 /* 2262 * Recalculate the priority before we select the target cpu or 2263 * run-queue. 2264 */ 2265 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) 2266 sched_priority(td); 2267#ifdef SMP 2268 cpuid = PCPU_GET(cpuid); 2269 /* 2270 * Pick the destination cpu and if it isn't ours transfer to the 2271 * target cpu. 2272 */ 2273 if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_MIGRATE(td)) 2274 cpu = cpuid; 2275 else if (!THREAD_CAN_MIGRATE(td)) 2276 cpu = ts->ts_cpu; 2277 else 2278 cpu = sched_pickcpu(ts, flags); 2279 tdq = sched_setcpu(ts, cpu, flags); 2280 tdq_add(tdq, td, flags); 2281 if (cpu != cpuid) { 2282 tdq_notify(ts); 2283 return; 2284 } 2285#else 2286 tdq = TDQ_SELF(); 2287 TDQ_LOCK(tdq); 2288 /* 2289 * Now that the thread is moving to the run-queue, set the lock 2290 * to the scheduler's lock. 2291 */ 2292 thread_lock_set(td, TDQ_LOCKPTR(tdq)); 2293 tdq_add(tdq, td, flags); 2294#endif 2295 if (!(flags & SRQ_YIELDING)) 2296 sched_setpreempt(td); 2297} 2298 2299/* 2300 * Remove a thread from a run-queue without running it. This is used 2301 * when we're stealing a thread from a remote queue. Otherwise all threads 2302 * exit by calling sched_exit_thread() and sched_throw() themselves. 2303 */ 2304void 2305sched_rem(struct thread *td) 2306{ 2307 struct tdq *tdq; 2308 struct td_sched *ts; 2309 2310 CTR5(KTR_SCHED, "sched_rem: %p(%s) prio %d by %p(%s)", 2311 td, td->td_proc->p_comm, td->td_priority, curthread, 2312 curthread->td_proc->p_comm); 2313 ts = td->td_sched; 2314 tdq = TDQ_CPU(ts->ts_cpu); 2315 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2316 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2317 KASSERT(TD_ON_RUNQ(td), 2318 ("sched_rem: thread not on run queue")); 2319 tdq_runq_rem(tdq, ts); 2320 tdq_load_rem(tdq, ts); 2321 TD_SET_CAN_RUN(td); 2322} 2323 2324/* 2325 * Fetch cpu utilization information. Updates on demand. 2326 */ 2327fixpt_t 2328sched_pctcpu(struct thread *td) 2329{ 2330 fixpt_t pctcpu; 2331 struct td_sched *ts; 2332 2333 pctcpu = 0; 2334 ts = td->td_sched; 2335 if (ts == NULL) 2336 return (0); 2337 2338 thread_lock(td); 2339 if (ts->ts_ticks) { 2340 int rtick; 2341 2342 sched_pctcpu_update(ts); 2343 /* How many rtick per second ? */ 2344 rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz); 2345 pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT; 2346 } 2347 td->td_proc->p_swtime = ts->ts_ltick - ts->ts_ftick; 2348 thread_unlock(td); 2349 2350 return (pctcpu); 2351} 2352 2353/* 2354 * Bind a thread to a target cpu. 2355 */ 2356void 2357sched_bind(struct thread *td, int cpu) 2358{ 2359 struct td_sched *ts; 2360 2361 THREAD_LOCK_ASSERT(td, MA_OWNED); 2362 ts = td->td_sched; 2363 if (ts->ts_flags & TSF_BOUND) 2364 sched_unbind(td); 2365 ts->ts_flags |= TSF_BOUND; 2366#ifdef SMP 2367 sched_pin(); 2368 if (PCPU_GET(cpuid) == cpu) 2369 return; 2370 ts->ts_cpu = cpu; 2371 /* When we return from mi_switch we'll be on the correct cpu. */ 2372 mi_switch(SW_VOL, NULL); 2373#endif 2374} 2375 2376/* 2377 * Release a bound thread. 2378 */ 2379void 2380sched_unbind(struct thread *td) 2381{ 2382 struct td_sched *ts; 2383 2384 THREAD_LOCK_ASSERT(td, MA_OWNED); 2385 ts = td->td_sched; 2386 if ((ts->ts_flags & TSF_BOUND) == 0) 2387 return; 2388 ts->ts_flags &= ~TSF_BOUND; 2389#ifdef SMP 2390 sched_unpin(); 2391#endif 2392} 2393 2394int 2395sched_is_bound(struct thread *td) 2396{ 2397 THREAD_LOCK_ASSERT(td, MA_OWNED); 2398 return (td->td_sched->ts_flags & TSF_BOUND); 2399} 2400 2401/* 2402 * Basic yield call. 2403 */ 2404void 2405sched_relinquish(struct thread *td) 2406{ 2407 thread_lock(td); 2408 if (td->td_pri_class == PRI_TIMESHARE) 2409 sched_prio(td, PRI_MAX_TIMESHARE); 2410 SCHED_STAT_INC(switch_relinquish); 2411 mi_switch(SW_VOL, NULL); 2412 thread_unlock(td); 2413} 2414 2415/* 2416 * Return the total system load. 2417 */ 2418int 2419sched_load(void) 2420{ 2421#ifdef SMP 2422 int total; 2423 int i; 2424 2425 total = 0; 2426 for (i = 0; i <= tdg_maxid; i++) 2427 total += TDQ_GROUP(i)->tdg_load; 2428 return (total); 2429#else 2430 return (TDQ_SELF()->tdq_sysload); 2431#endif 2432} 2433 2434int 2435sched_sizeof_proc(void) 2436{ 2437 return (sizeof(struct proc)); 2438} 2439 2440int 2441sched_sizeof_thread(void) 2442{ 2443 return (sizeof(struct thread) + sizeof(struct td_sched)); 2444} 2445 2446/* 2447 * The actual idle process. 2448 */ 2449void 2450sched_idletd(void *dummy) 2451{ 2452 struct thread *td; 2453 struct tdq *tdq; 2454 2455 td = curthread; 2456 tdq = TDQ_SELF(); 2457 mtx_assert(&Giant, MA_NOTOWNED); 2458 /* ULE relies on preemption for idle interruption. */ 2459 for (;;) { 2460#ifdef SMP 2461 if (tdq_idled(tdq)) 2462 cpu_idle(); 2463#else 2464 cpu_idle(); 2465#endif 2466 } 2467} 2468 2469/* 2470 * A CPU is entering for the first time or a thread is exiting. 2471 */ 2472void 2473sched_throw(struct thread *td) 2474{ 2475 struct tdq *tdq; 2476 2477 tdq = TDQ_SELF(); 2478 if (td == NULL) { 2479 /* Correct spinlock nesting and acquire the correct lock. */ 2480 TDQ_LOCK(tdq); 2481 spinlock_exit(); 2482 } else { 2483 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2484 tdq_load_rem(tdq, td->td_sched); 2485 } 2486 KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count")); 2487 PCPU_SET(switchtime, cpu_ticks()); 2488 PCPU_SET(switchticks, ticks); 2489 cpu_throw(td, choosethread()); /* doesn't return */ 2490} 2491 2492/* 2493 * This is called from fork_exit(). Just acquire the correct locks and 2494 * let fork do the rest of the work. 2495 */ 2496void 2497sched_fork_exit(struct thread *td) 2498{ 2499 struct td_sched *ts; 2500 struct tdq *tdq; 2501 int cpuid; 2502 2503 /* 2504 * Finish setting up thread glue so that it begins execution in a 2505 * non-nested critical section with the scheduler lock held. 2506 */ 2507 cpuid = PCPU_GET(cpuid); 2508 tdq = TDQ_CPU(cpuid); 2509 ts = td->td_sched; 2510 if (TD_IS_IDLETHREAD(td)) 2511 td->td_lock = TDQ_LOCKPTR(tdq); 2512 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2513 td->td_oncpu = cpuid; 2514 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)td; 2515 THREAD_LOCK_ASSERT(td, MA_OWNED | MA_NOTRECURSED); 2516} 2517 2518static SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, 2519 "Scheduler"); 2520SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0, 2521 "Scheduler name"); 2522SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0, 2523 "Slice size for timeshare threads"); 2524SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0, 2525 "Interactivity score threshold"); 2526SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW, &preempt_thresh, 2527 0,"Min priority for preemption, lower priorities have greater precedence"); 2528#ifdef SMP 2529SYSCTL_INT(_kern_sched, OID_AUTO, pick_pri, CTLFLAG_RW, &pick_pri, 0, 2530 "Pick the target cpu based on priority rather than load."); 2531SYSCTL_INT(_kern_sched, OID_AUTO, pick_zero, CTLFLAG_RW, &pick_zero, 0, 2532 "If there are no idle cpus pick cpu0"); 2533SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0, 2534 "Number of hz ticks to keep thread affinity for"); 2535SYSCTL_INT(_kern_sched, OID_AUTO, tryself, CTLFLAG_RW, &tryself, 0, ""); 2536SYSCTL_INT(_kern_sched, OID_AUTO, tryselfidle, CTLFLAG_RW, 2537 &tryselfidle, 0, ""); 2538SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0, 2539 "Enables the long-term load balancer"); 2540SYSCTL_INT(_kern_sched, OID_AUTO, steal_htt, CTLFLAG_RW, &steal_htt, 0, 2541 "Steals work from another hyper-threaded core on idle"); 2542SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0, 2543 "Attempts to steal work from other cores before idling"); 2544SYSCTL_INT(_kern_sched, OID_AUTO, topology, CTLFLAG_RD, &topology, 0, 2545 "True when a topology has been specified by the MD code."); 2546#endif 2547 2548/* ps compat */ 2549static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */ 2550SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 2551 2552 2553#define KERN_SWITCH_INCLUDE 1 2554#include "kern/kern_switch.c" 2555