sched_ule.c revision 279349
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 279349 2015-02-27 02:56:58Z imp $"); 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/limits.h> 50#include <sys/lock.h> 51#include <sys/mutex.h> 52#include <sys/proc.h> 53#include <sys/resource.h> 54#include <sys/resourcevar.h> 55#include <sys/sched.h> 56#include <sys/sdt.h> 57#include <sys/smp.h> 58#include <sys/sx.h> 59#include <sys/sysctl.h> 60#include <sys/sysproto.h> 61#include <sys/turnstile.h> 62#include <sys/umtx.h> 63#include <sys/vmmeter.h> 64#include <sys/cpuset.h> 65#include <sys/sbuf.h> 66 67#ifdef HWPMC_HOOKS 68#include <sys/pmckern.h> 69#endif 70 71#ifdef KDTRACE_HOOKS 72#include <sys/dtrace_bsd.h> 73int dtrace_vtime_active; 74dtrace_vtime_switch_func_t dtrace_vtime_switch_func; 75#endif 76 77#include <machine/cpu.h> 78#include <machine/smp.h> 79 80#define KTR_ULE 0 81 82#define TS_NAME_LEN (MAXCOMLEN + sizeof(" td ") + sizeof(__XSTRING(UINT_MAX))) 83#define TDQ_NAME_LEN (sizeof("sched lock ") + sizeof(__XSTRING(MAXCPU))) 84#define TDQ_LOADNAME_LEN (sizeof("CPU ") + sizeof(__XSTRING(MAXCPU)) - 1 + sizeof(" load")) 85 86/* 87 * Thread scheduler specific section. All fields are protected 88 * by the thread lock. 89 */ 90struct td_sched { 91 struct runq *ts_runq; /* Run-queue we're queued on. */ 92 short ts_flags; /* TSF_* flags. */ 93 int ts_cpu; /* CPU that we have affinity for. */ 94 int ts_rltick; /* Real last tick, for affinity. */ 95 int ts_slice; /* Ticks of slice remaining. */ 96 u_int ts_slptime; /* Number of ticks we vol. slept */ 97 u_int ts_runtime; /* Number of ticks we were running */ 98 int ts_ltick; /* Last tick that we were running on */ 99 int ts_ftick; /* First tick that we were running on */ 100 int ts_ticks; /* Tick count */ 101#ifdef KTR 102 char ts_name[TS_NAME_LEN]; 103#endif 104}; 105/* flags kept in ts_flags */ 106#define TSF_BOUND 0x0001 /* Thread can not migrate. */ 107#define TSF_XFERABLE 0x0002 /* Thread was added as transferable. */ 108 109static struct td_sched td_sched0; 110 111#define THREAD_CAN_MIGRATE(td) ((td)->td_pinned == 0) 112#define THREAD_CAN_SCHED(td, cpu) \ 113 CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask) 114 115/* 116 * Priority ranges used for interactive and non-interactive timeshare 117 * threads. The timeshare priorities are split up into four ranges. 118 * The first range handles interactive threads. The last three ranges 119 * (NHALF, x, and NHALF) handle non-interactive threads with the outer 120 * ranges supporting nice values. 121 */ 122#define PRI_TIMESHARE_RANGE (PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE + 1) 123#define PRI_INTERACT_RANGE ((PRI_TIMESHARE_RANGE - SCHED_PRI_NRESV) / 2) 124#define PRI_BATCH_RANGE (PRI_TIMESHARE_RANGE - PRI_INTERACT_RANGE) 125 126#define PRI_MIN_INTERACT PRI_MIN_TIMESHARE 127#define PRI_MAX_INTERACT (PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE - 1) 128#define PRI_MIN_BATCH (PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE) 129#define PRI_MAX_BATCH PRI_MAX_TIMESHARE 130 131/* 132 * Cpu percentage computation macros and defines. 133 * 134 * SCHED_TICK_SECS: Number of seconds to average the cpu usage across. 135 * SCHED_TICK_TARG: Number of hz ticks to average the cpu usage across. 136 * SCHED_TICK_MAX: Maximum number of ticks before scaling back. 137 * SCHED_TICK_SHIFT: Shift factor to avoid rounding away results. 138 * SCHED_TICK_HZ: Compute the number of hz ticks for a given ticks count. 139 * SCHED_TICK_TOTAL: Gives the amount of time we've been recording ticks. 140 */ 141#define SCHED_TICK_SECS 10 142#define SCHED_TICK_TARG (hz * SCHED_TICK_SECS) 143#define SCHED_TICK_MAX (SCHED_TICK_TARG + hz) 144#define SCHED_TICK_SHIFT 10 145#define SCHED_TICK_HZ(ts) ((ts)->ts_ticks >> SCHED_TICK_SHIFT) 146#define SCHED_TICK_TOTAL(ts) (max((ts)->ts_ltick - (ts)->ts_ftick, hz)) 147 148/* 149 * These macros determine priorities for non-interactive threads. They are 150 * assigned a priority based on their recent cpu utilization as expressed 151 * by the ratio of ticks to the tick total. NHALF priorities at the start 152 * and end of the MIN to MAX timeshare range are only reachable with negative 153 * or positive nice respectively. 154 * 155 * PRI_RANGE: Priority range for utilization dependent priorities. 156 * PRI_NRESV: Number of nice values. 157 * PRI_TICKS: Compute a priority in PRI_RANGE from the ticks count and total. 158 * PRI_NICE: Determines the part of the priority inherited from nice. 159 */ 160#define SCHED_PRI_NRESV (PRIO_MAX - PRIO_MIN) 161#define SCHED_PRI_NHALF (SCHED_PRI_NRESV / 2) 162#define SCHED_PRI_MIN (PRI_MIN_BATCH + SCHED_PRI_NHALF) 163#define SCHED_PRI_MAX (PRI_MAX_BATCH - SCHED_PRI_NHALF) 164#define SCHED_PRI_RANGE (SCHED_PRI_MAX - SCHED_PRI_MIN + 1) 165#define SCHED_PRI_TICKS(ts) \ 166 (SCHED_TICK_HZ((ts)) / \ 167 (roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE)) 168#define SCHED_PRI_NICE(nice) (nice) 169 170/* 171 * These determine the interactivity of a process. Interactivity differs from 172 * cpu utilization in that it expresses the voluntary time slept vs time ran 173 * while cpu utilization includes all time not running. This more accurately 174 * models the intent of the thread. 175 * 176 * SLP_RUN_MAX: Maximum amount of sleep time + run time we'll accumulate 177 * before throttling back. 178 * SLP_RUN_FORK: Maximum slp+run time to inherit at fork time. 179 * INTERACT_MAX: Maximum interactivity value. Smaller is better. 180 * INTERACT_THRESH: Threshold for placement on the current runq. 181 */ 182#define SCHED_SLP_RUN_MAX ((hz * 5) << SCHED_TICK_SHIFT) 183#define SCHED_SLP_RUN_FORK ((hz / 2) << SCHED_TICK_SHIFT) 184#define SCHED_INTERACT_MAX (100) 185#define SCHED_INTERACT_HALF (SCHED_INTERACT_MAX / 2) 186#define SCHED_INTERACT_THRESH (30) 187 188/* 189 * These parameters determine the slice behavior for batch work. 190 */ 191#define SCHED_SLICE_DEFAULT_DIVISOR 10 /* ~94 ms, 12 stathz ticks. */ 192#define SCHED_SLICE_MIN_DIVISOR 6 /* DEFAULT/MIN = ~16 ms. */ 193 194/* Flags kept in td_flags. */ 195#define TDF_SLICEEND TDF_SCHED2 /* Thread time slice is over. */ 196 197/* 198 * tickincr: Converts a stathz tick into a hz domain scaled by 199 * the shift factor. Without the shift the error rate 200 * due to rounding would be unacceptably high. 201 * realstathz: stathz is sometimes 0 and run off of hz. 202 * sched_slice: Runtime of each thread before rescheduling. 203 * preempt_thresh: Priority threshold for preemption and remote IPIs. 204 */ 205static int sched_interact = SCHED_INTERACT_THRESH; 206static int tickincr = 8 << SCHED_TICK_SHIFT; 207static int realstathz = 127; /* reset during boot. */ 208static int sched_slice = 10; /* reset during boot. */ 209static int sched_slice_min = 1; /* reset during boot. */ 210#ifdef PREEMPTION 211#ifdef FULL_PREEMPTION 212static int preempt_thresh = PRI_MAX_IDLE; 213#else 214static int preempt_thresh = PRI_MIN_KERN; 215#endif 216#else 217static int preempt_thresh = 0; 218#endif 219static int static_boost = PRI_MIN_BATCH; 220static int sched_idlespins = 10000; 221static int sched_idlespinthresh = -1; 222 223/* 224 * tdq - per processor runqs and statistics. All fields are protected by the 225 * tdq_lock. The load and lowpri may be accessed without to avoid excess 226 * locking in sched_pickcpu(); 227 */ 228struct tdq { 229 /* 230 * Ordered to improve efficiency of cpu_search() and switch(). 231 * tdq_lock is padded to avoid false sharing with tdq_load and 232 * tdq_cpu_idle. 233 */ 234 struct mtx_padalign tdq_lock; /* run queue lock. */ 235 struct cpu_group *tdq_cg; /* Pointer to cpu topology. */ 236 volatile int tdq_load; /* Aggregate load. */ 237 volatile int tdq_cpu_idle; /* cpu_idle() is active. */ 238 int tdq_sysload; /* For loadavg, !ITHD load. */ 239 int tdq_transferable; /* Transferable thread count. */ 240 short tdq_switchcnt; /* Switches this tick. */ 241 short tdq_oldswitchcnt; /* Switches last tick. */ 242 u_char tdq_lowpri; /* Lowest priority thread. */ 243 u_char tdq_ipipending; /* IPI pending. */ 244 u_char tdq_idx; /* Current insert index. */ 245 u_char tdq_ridx; /* Current removal index. */ 246 struct runq tdq_realtime; /* real-time run queue. */ 247 struct runq tdq_timeshare; /* timeshare run queue. */ 248 struct runq tdq_idle; /* Queue of IDLE threads. */ 249 char tdq_name[TDQ_NAME_LEN]; 250#ifdef KTR 251 char tdq_loadname[TDQ_LOADNAME_LEN]; 252#endif 253} __aligned(64); 254 255/* Idle thread states and config. */ 256#define TDQ_RUNNING 1 257#define TDQ_IDLE 2 258 259#ifdef SMP 260struct cpu_group *cpu_top; /* CPU topology */ 261 262#define SCHED_AFFINITY_DEFAULT (max(1, hz / 1000)) 263#define SCHED_AFFINITY(ts, t) ((ts)->ts_rltick > ticks - ((t) * affinity)) 264 265/* 266 * Run-time tunables. 267 */ 268static int rebalance = 1; 269static int balance_interval = 128; /* Default set in sched_initticks(). */ 270static int affinity; 271static int steal_idle = 1; 272static int steal_thresh = 2; 273 274/* 275 * One thread queue per processor. 276 */ 277static struct tdq tdq_cpu[MAXCPU]; 278static struct tdq *balance_tdq; 279static int balance_ticks; 280static DPCPU_DEFINE(uint32_t, randomval); 281 282#define TDQ_SELF() (&tdq_cpu[PCPU_GET(cpuid)]) 283#define TDQ_CPU(x) (&tdq_cpu[(x)]) 284#define TDQ_ID(x) ((int)((x) - tdq_cpu)) 285#else /* !SMP */ 286static struct tdq tdq_cpu; 287 288#define TDQ_ID(x) (0) 289#define TDQ_SELF() (&tdq_cpu) 290#define TDQ_CPU(x) (&tdq_cpu) 291#endif 292 293#define TDQ_LOCK_ASSERT(t, type) mtx_assert(TDQ_LOCKPTR((t)), (type)) 294#define TDQ_LOCK(t) mtx_lock_spin(TDQ_LOCKPTR((t))) 295#define TDQ_LOCK_FLAGS(t, f) mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f)) 296#define TDQ_UNLOCK(t) mtx_unlock_spin(TDQ_LOCKPTR((t))) 297#define TDQ_LOCKPTR(t) ((struct mtx *)(&(t)->tdq_lock)) 298 299static void sched_priority(struct thread *); 300static void sched_thread_priority(struct thread *, u_char); 301static int sched_interact_score(struct thread *); 302static void sched_interact_update(struct thread *); 303static void sched_interact_fork(struct thread *); 304static void sched_pctcpu_update(struct td_sched *, int); 305static int sched_random(void); 306 307/* Operations on per processor queues */ 308static struct thread *tdq_choose(struct tdq *); 309static void tdq_setup(struct tdq *); 310static void tdq_load_add(struct tdq *, struct thread *); 311static void tdq_load_rem(struct tdq *, struct thread *); 312static __inline void tdq_runq_add(struct tdq *, struct thread *, int); 313static __inline void tdq_runq_rem(struct tdq *, struct thread *); 314static inline int sched_shouldpreempt(int, int, int); 315void tdq_print(int cpu); 316static void runq_print(struct runq *rq); 317static void tdq_add(struct tdq *, struct thread *, int); 318#ifdef SMP 319static int tdq_move(struct tdq *, struct tdq *); 320static int tdq_idled(struct tdq *); 321static void tdq_notify(struct tdq *, struct thread *); 322static struct thread *tdq_steal(struct tdq *, int); 323static struct thread *runq_steal(struct runq *, int); 324static int sched_pickcpu(struct thread *, int); 325static void sched_balance(void); 326static int sched_balance_pair(struct tdq *, struct tdq *); 327static inline struct tdq *sched_setcpu(struct thread *, int, int); 328static inline void thread_unblock_switch(struct thread *, struct mtx *); 329static struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int); 330static int sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS); 331static int sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, 332 struct cpu_group *cg, int indent); 333#endif 334 335static void sched_setup(void *dummy); 336SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL); 337 338static void sched_initticks(void *dummy); 339SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks, 340 NULL); 341 342SDT_PROVIDER_DEFINE(sched); 343 344SDT_PROBE_DEFINE3(sched, , , change__pri, "struct thread *", 345 "struct proc *", "uint8_t"); 346SDT_PROBE_DEFINE3(sched, , , dequeue, "struct thread *", 347 "struct proc *", "void *"); 348SDT_PROBE_DEFINE4(sched, , , enqueue, "struct thread *", 349 "struct proc *", "void *", "int"); 350SDT_PROBE_DEFINE4(sched, , , lend__pri, "struct thread *", 351 "struct proc *", "uint8_t", "struct thread *"); 352SDT_PROBE_DEFINE2(sched, , , load__change, "int", "int"); 353SDT_PROBE_DEFINE2(sched, , , off__cpu, "struct thread *", 354 "struct proc *"); 355SDT_PROBE_DEFINE(sched, , , on__cpu); 356SDT_PROBE_DEFINE(sched, , , remain__cpu); 357SDT_PROBE_DEFINE2(sched, , , surrender, "struct thread *", 358 "struct proc *"); 359 360/* 361 * We need some randomness. Implement the classic Linear Congruential 362 * generator X_{n+1}=(aX_n+c) mod m. These values are optimized for 363 * m = 2^32, a = 69069 and c = 5. This is signed so that we can get 364 * both positive and negative values from it by shifting the value 365 * right. 366 */ 367static int sched_random() 368{ 369 int rnd, *rndptr; 370 rndptr = DPCPU_PTR(randomval); 371 rnd = *rndptr * 69069 + 5; 372 *rndptr = rnd; 373 return(rnd); 374} 375 376/* 377 * Print the threads waiting on a run-queue. 378 */ 379static void 380runq_print(struct runq *rq) 381{ 382 struct rqhead *rqh; 383 struct thread *td; 384 int pri; 385 int j; 386 int i; 387 388 for (i = 0; i < RQB_LEN; i++) { 389 printf("\t\trunq bits %d 0x%zx\n", 390 i, rq->rq_status.rqb_bits[i]); 391 for (j = 0; j < RQB_BPW; j++) 392 if (rq->rq_status.rqb_bits[i] & (1ul << j)) { 393 pri = j + (i << RQB_L2BPW); 394 rqh = &rq->rq_queues[pri]; 395 TAILQ_FOREACH(td, rqh, td_runq) { 396 printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n", 397 td, td->td_name, td->td_priority, 398 td->td_rqindex, pri); 399 } 400 } 401 } 402} 403 404/* 405 * Print the status of a per-cpu thread queue. Should be a ddb show cmd. 406 */ 407void 408tdq_print(int cpu) 409{ 410 struct tdq *tdq; 411 412 tdq = TDQ_CPU(cpu); 413 414 printf("tdq %d:\n", TDQ_ID(tdq)); 415 printf("\tlock %p\n", TDQ_LOCKPTR(tdq)); 416 printf("\tLock name: %s\n", tdq->tdq_name); 417 printf("\tload: %d\n", tdq->tdq_load); 418 printf("\tswitch cnt: %d\n", tdq->tdq_switchcnt); 419 printf("\told switch cnt: %d\n", tdq->tdq_oldswitchcnt); 420 printf("\ttimeshare idx: %d\n", tdq->tdq_idx); 421 printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx); 422 printf("\tload transferable: %d\n", tdq->tdq_transferable); 423 printf("\tlowest priority: %d\n", tdq->tdq_lowpri); 424 printf("\trealtime runq:\n"); 425 runq_print(&tdq->tdq_realtime); 426 printf("\ttimeshare runq:\n"); 427 runq_print(&tdq->tdq_timeshare); 428 printf("\tidle runq:\n"); 429 runq_print(&tdq->tdq_idle); 430} 431 432static inline int 433sched_shouldpreempt(int pri, int cpri, int remote) 434{ 435 /* 436 * If the new priority is not better than the current priority there is 437 * nothing to do. 438 */ 439 if (pri >= cpri) 440 return (0); 441 /* 442 * Always preempt idle. 443 */ 444 if (cpri >= PRI_MIN_IDLE) 445 return (1); 446 /* 447 * If preemption is disabled don't preempt others. 448 */ 449 if (preempt_thresh == 0) 450 return (0); 451 /* 452 * Preempt if we exceed the threshold. 453 */ 454 if (pri <= preempt_thresh) 455 return (1); 456 /* 457 * If we're interactive or better and there is non-interactive 458 * or worse running preempt only remote processors. 459 */ 460 if (remote && pri <= PRI_MAX_INTERACT && cpri > PRI_MAX_INTERACT) 461 return (1); 462 return (0); 463} 464 465/* 466 * Add a thread to the actual run-queue. Keeps transferable counts up to 467 * date with what is actually on the run-queue. Selects the correct 468 * queue position for timeshare threads. 469 */ 470static __inline void 471tdq_runq_add(struct tdq *tdq, struct thread *td, int flags) 472{ 473 struct td_sched *ts; 474 u_char pri; 475 476 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 477 THREAD_LOCK_ASSERT(td, MA_OWNED); 478 479 pri = td->td_priority; 480 ts = td->td_sched; 481 TD_SET_RUNQ(td); 482 if (THREAD_CAN_MIGRATE(td)) { 483 tdq->tdq_transferable++; 484 ts->ts_flags |= TSF_XFERABLE; 485 } 486 if (pri < PRI_MIN_BATCH) { 487 ts->ts_runq = &tdq->tdq_realtime; 488 } else if (pri <= PRI_MAX_BATCH) { 489 ts->ts_runq = &tdq->tdq_timeshare; 490 KASSERT(pri <= PRI_MAX_BATCH && pri >= PRI_MIN_BATCH, 491 ("Invalid priority %d on timeshare runq", pri)); 492 /* 493 * This queue contains only priorities between MIN and MAX 494 * realtime. Use the whole queue to represent these values. 495 */ 496 if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) { 497 pri = RQ_NQS * (pri - PRI_MIN_BATCH) / PRI_BATCH_RANGE; 498 pri = (pri + tdq->tdq_idx) % RQ_NQS; 499 /* 500 * This effectively shortens the queue by one so we 501 * can have a one slot difference between idx and 502 * ridx while we wait for threads to drain. 503 */ 504 if (tdq->tdq_ridx != tdq->tdq_idx && 505 pri == tdq->tdq_ridx) 506 pri = (unsigned char)(pri - 1) % RQ_NQS; 507 } else 508 pri = tdq->tdq_ridx; 509 runq_add_pri(ts->ts_runq, td, pri, flags); 510 return; 511 } else 512 ts->ts_runq = &tdq->tdq_idle; 513 runq_add(ts->ts_runq, td, flags); 514} 515 516/* 517 * Remove a thread from a run-queue. This typically happens when a thread 518 * is selected to run. Running threads are not on the queue and the 519 * transferable count does not reflect them. 520 */ 521static __inline void 522tdq_runq_rem(struct tdq *tdq, struct thread *td) 523{ 524 struct td_sched *ts; 525 526 ts = td->td_sched; 527 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 528 KASSERT(ts->ts_runq != NULL, 529 ("tdq_runq_remove: thread %p null ts_runq", td)); 530 if (ts->ts_flags & TSF_XFERABLE) { 531 tdq->tdq_transferable--; 532 ts->ts_flags &= ~TSF_XFERABLE; 533 } 534 if (ts->ts_runq == &tdq->tdq_timeshare) { 535 if (tdq->tdq_idx != tdq->tdq_ridx) 536 runq_remove_idx(ts->ts_runq, td, &tdq->tdq_ridx); 537 else 538 runq_remove_idx(ts->ts_runq, td, NULL); 539 } else 540 runq_remove(ts->ts_runq, td); 541} 542 543/* 544 * Load is maintained for all threads RUNNING and ON_RUNQ. Add the load 545 * for this thread to the referenced thread queue. 546 */ 547static void 548tdq_load_add(struct tdq *tdq, struct thread *td) 549{ 550 551 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 552 THREAD_LOCK_ASSERT(td, MA_OWNED); 553 554 tdq->tdq_load++; 555 if ((td->td_flags & TDF_NOLOAD) == 0) 556 tdq->tdq_sysload++; 557 KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load); 558 SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load); 559} 560 561/* 562 * Remove the load from a thread that is transitioning to a sleep state or 563 * exiting. 564 */ 565static void 566tdq_load_rem(struct tdq *tdq, struct thread *td) 567{ 568 569 THREAD_LOCK_ASSERT(td, MA_OWNED); 570 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 571 KASSERT(tdq->tdq_load != 0, 572 ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq))); 573 574 tdq->tdq_load--; 575 if ((td->td_flags & TDF_NOLOAD) == 0) 576 tdq->tdq_sysload--; 577 KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load); 578 SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load); 579} 580 581/* 582 * Bound timeshare latency by decreasing slice size as load increases. We 583 * consider the maximum latency as the sum of the threads waiting to run 584 * aside from curthread and target no more than sched_slice latency but 585 * no less than sched_slice_min runtime. 586 */ 587static inline int 588tdq_slice(struct tdq *tdq) 589{ 590 int load; 591 592 /* 593 * It is safe to use sys_load here because this is called from 594 * contexts where timeshare threads are running and so there 595 * cannot be higher priority load in the system. 596 */ 597 load = tdq->tdq_sysload - 1; 598 if (load >= SCHED_SLICE_MIN_DIVISOR) 599 return (sched_slice_min); 600 if (load <= 1) 601 return (sched_slice); 602 return (sched_slice / load); 603} 604 605/* 606 * Set lowpri to its exact value by searching the run-queue and 607 * evaluating curthread. curthread may be passed as an optimization. 608 */ 609static void 610tdq_setlowpri(struct tdq *tdq, struct thread *ctd) 611{ 612 struct thread *td; 613 614 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 615 if (ctd == NULL) 616 ctd = pcpu_find(TDQ_ID(tdq))->pc_curthread; 617 td = tdq_choose(tdq); 618 if (td == NULL || td->td_priority > ctd->td_priority) 619 tdq->tdq_lowpri = ctd->td_priority; 620 else 621 tdq->tdq_lowpri = td->td_priority; 622} 623 624#ifdef SMP 625struct cpu_search { 626 cpuset_t cs_mask; 627 u_int cs_prefer; 628 int cs_pri; /* Min priority for low. */ 629 int cs_limit; /* Max load for low, min load for high. */ 630 int cs_cpu; 631 int cs_load; 632}; 633 634#define CPU_SEARCH_LOWEST 0x1 635#define CPU_SEARCH_HIGHEST 0x2 636#define CPU_SEARCH_BOTH (CPU_SEARCH_LOWEST|CPU_SEARCH_HIGHEST) 637 638#define CPUSET_FOREACH(cpu, mask) \ 639 for ((cpu) = 0; (cpu) <= mp_maxid; (cpu)++) \ 640 if (CPU_ISSET(cpu, &mask)) 641 642static __always_inline int cpu_search(const struct cpu_group *cg, 643 struct cpu_search *low, struct cpu_search *high, const int match); 644int __noinline cpu_search_lowest(const struct cpu_group *cg, 645 struct cpu_search *low); 646int __noinline cpu_search_highest(const struct cpu_group *cg, 647 struct cpu_search *high); 648int __noinline cpu_search_both(const struct cpu_group *cg, 649 struct cpu_search *low, struct cpu_search *high); 650 651/* 652 * Search the tree of cpu_groups for the lowest or highest loaded cpu 653 * according to the match argument. This routine actually compares the 654 * load on all paths through the tree and finds the least loaded cpu on 655 * the least loaded path, which may differ from the least loaded cpu in 656 * the system. This balances work among caches and busses. 657 * 658 * This inline is instantiated in three forms below using constants for the 659 * match argument. It is reduced to the minimum set for each case. It is 660 * also recursive to the depth of the tree. 661 */ 662static __always_inline int 663cpu_search(const struct cpu_group *cg, struct cpu_search *low, 664 struct cpu_search *high, const int match) 665{ 666 struct cpu_search lgroup; 667 struct cpu_search hgroup; 668 cpuset_t cpumask; 669 struct cpu_group *child; 670 struct tdq *tdq; 671 int cpu, i, hload, lload, load, total, rnd; 672 673 total = 0; 674 cpumask = cg->cg_mask; 675 if (match & CPU_SEARCH_LOWEST) { 676 lload = INT_MAX; 677 lgroup = *low; 678 } 679 if (match & CPU_SEARCH_HIGHEST) { 680 hload = INT_MIN; 681 hgroup = *high; 682 } 683 684 /* Iterate through the child CPU groups and then remaining CPUs. */ 685 for (i = cg->cg_children, cpu = mp_maxid; ; ) { 686 if (i == 0) { 687#ifdef HAVE_INLINE_FFSL 688 cpu = CPU_FFS(&cpumask) - 1; 689#else 690 while (cpu >= 0 && !CPU_ISSET(cpu, &cpumask)) 691 cpu--; 692#endif 693 if (cpu < 0) 694 break; 695 child = NULL; 696 } else 697 child = &cg->cg_child[i - 1]; 698 699 if (match & CPU_SEARCH_LOWEST) 700 lgroup.cs_cpu = -1; 701 if (match & CPU_SEARCH_HIGHEST) 702 hgroup.cs_cpu = -1; 703 if (child) { /* Handle child CPU group. */ 704 CPU_NAND(&cpumask, &child->cg_mask); 705 switch (match) { 706 case CPU_SEARCH_LOWEST: 707 load = cpu_search_lowest(child, &lgroup); 708 break; 709 case CPU_SEARCH_HIGHEST: 710 load = cpu_search_highest(child, &hgroup); 711 break; 712 case CPU_SEARCH_BOTH: 713 load = cpu_search_both(child, &lgroup, &hgroup); 714 break; 715 } 716 } else { /* Handle child CPU. */ 717 CPU_CLR(cpu, &cpumask); 718 tdq = TDQ_CPU(cpu); 719 load = tdq->tdq_load * 256; 720 rnd = sched_random() >> 26; /* -32 to +31 */ 721 if (match & CPU_SEARCH_LOWEST) { 722 if (cpu == low->cs_prefer) 723 load -= 64; 724 /* If that CPU is allowed and get data. */ 725 if (tdq->tdq_lowpri > lgroup.cs_pri && 726 tdq->tdq_load <= lgroup.cs_limit && 727 CPU_ISSET(cpu, &lgroup.cs_mask)) { 728 lgroup.cs_cpu = cpu; 729 lgroup.cs_load = load - rnd; 730 } 731 } 732 if (match & CPU_SEARCH_HIGHEST) 733 if (tdq->tdq_load >= hgroup.cs_limit && 734 tdq->tdq_transferable && 735 CPU_ISSET(cpu, &hgroup.cs_mask)) { 736 hgroup.cs_cpu = cpu; 737 hgroup.cs_load = load - rnd; 738 } 739 } 740 total += load; 741 742 /* We have info about child item. Compare it. */ 743 if (match & CPU_SEARCH_LOWEST) { 744 if (lgroup.cs_cpu >= 0 && 745 (load < lload || 746 (load == lload && lgroup.cs_load < low->cs_load))) { 747 lload = load; 748 low->cs_cpu = lgroup.cs_cpu; 749 low->cs_load = lgroup.cs_load; 750 } 751 } 752 if (match & CPU_SEARCH_HIGHEST) 753 if (hgroup.cs_cpu >= 0 && 754 (load > hload || 755 (load == hload && hgroup.cs_load > high->cs_load))) { 756 hload = load; 757 high->cs_cpu = hgroup.cs_cpu; 758 high->cs_load = hgroup.cs_load; 759 } 760 if (child) { 761 i--; 762 if (i == 0 && CPU_EMPTY(&cpumask)) 763 break; 764 } 765#ifndef HAVE_INLINE_FFSL 766 else 767 cpu--; 768#endif 769 } 770 return (total); 771} 772 773/* 774 * cpu_search instantiations must pass constants to maintain the inline 775 * optimization. 776 */ 777int 778cpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low) 779{ 780 return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST); 781} 782 783int 784cpu_search_highest(const struct cpu_group *cg, struct cpu_search *high) 785{ 786 return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST); 787} 788 789int 790cpu_search_both(const struct cpu_group *cg, struct cpu_search *low, 791 struct cpu_search *high) 792{ 793 return cpu_search(cg, low, high, CPU_SEARCH_BOTH); 794} 795 796/* 797 * Find the cpu with the least load via the least loaded path that has a 798 * lowpri greater than pri pri. A pri of -1 indicates any priority is 799 * acceptable. 800 */ 801static inline int 802sched_lowest(const struct cpu_group *cg, cpuset_t mask, int pri, int maxload, 803 int prefer) 804{ 805 struct cpu_search low; 806 807 low.cs_cpu = -1; 808 low.cs_prefer = prefer; 809 low.cs_mask = mask; 810 low.cs_pri = pri; 811 low.cs_limit = maxload; 812 cpu_search_lowest(cg, &low); 813 return low.cs_cpu; 814} 815 816/* 817 * Find the cpu with the highest load via the highest loaded path. 818 */ 819static inline int 820sched_highest(const struct cpu_group *cg, cpuset_t mask, int minload) 821{ 822 struct cpu_search high; 823 824 high.cs_cpu = -1; 825 high.cs_mask = mask; 826 high.cs_limit = minload; 827 cpu_search_highest(cg, &high); 828 return high.cs_cpu; 829} 830 831static void 832sched_balance_group(struct cpu_group *cg) 833{ 834 cpuset_t hmask, lmask; 835 int high, low, anylow; 836 837 CPU_FILL(&hmask); 838 for (;;) { 839 high = sched_highest(cg, hmask, 1); 840 /* Stop if there is no more CPU with transferrable threads. */ 841 if (high == -1) 842 break; 843 CPU_CLR(high, &hmask); 844 CPU_COPY(&hmask, &lmask); 845 /* Stop if there is no more CPU left for low. */ 846 if (CPU_EMPTY(&lmask)) 847 break; 848 anylow = 1; 849nextlow: 850 low = sched_lowest(cg, lmask, -1, 851 TDQ_CPU(high)->tdq_load - 1, high); 852 /* Stop if we looked well and found no less loaded CPU. */ 853 if (anylow && low == -1) 854 break; 855 /* Go to next high if we found no less loaded CPU. */ 856 if (low == -1) 857 continue; 858 /* Transfer thread from high to low. */ 859 if (sched_balance_pair(TDQ_CPU(high), TDQ_CPU(low))) { 860 /* CPU that got thread can no longer be a donor. */ 861 CPU_CLR(low, &hmask); 862 } else { 863 /* 864 * If failed, then there is no threads on high 865 * that can run on this low. Drop low from low 866 * mask and look for different one. 867 */ 868 CPU_CLR(low, &lmask); 869 anylow = 0; 870 goto nextlow; 871 } 872 } 873} 874 875static void 876sched_balance(void) 877{ 878 struct tdq *tdq; 879 880 if (smp_started == 0 || rebalance == 0) 881 return; 882 883 balance_ticks = max(balance_interval / 2, 1) + 884 ((sched_random() >> 16) % balance_interval); 885 tdq = TDQ_SELF(); 886 TDQ_UNLOCK(tdq); 887 sched_balance_group(cpu_top); 888 TDQ_LOCK(tdq); 889} 890 891/* 892 * Lock two thread queues using their address to maintain lock order. 893 */ 894static void 895tdq_lock_pair(struct tdq *one, struct tdq *two) 896{ 897 if (one < two) { 898 TDQ_LOCK(one); 899 TDQ_LOCK_FLAGS(two, MTX_DUPOK); 900 } else { 901 TDQ_LOCK(two); 902 TDQ_LOCK_FLAGS(one, MTX_DUPOK); 903 } 904} 905 906/* 907 * Unlock two thread queues. Order is not important here. 908 */ 909static void 910tdq_unlock_pair(struct tdq *one, struct tdq *two) 911{ 912 TDQ_UNLOCK(one); 913 TDQ_UNLOCK(two); 914} 915 916/* 917 * Transfer load between two imbalanced thread queues. 918 */ 919static int 920sched_balance_pair(struct tdq *high, struct tdq *low) 921{ 922 int moved; 923 int cpu; 924 925 tdq_lock_pair(high, low); 926 moved = 0; 927 /* 928 * Determine what the imbalance is and then adjust that to how many 929 * threads we actually have to give up (transferable). 930 */ 931 if (high->tdq_transferable != 0 && high->tdq_load > low->tdq_load && 932 (moved = tdq_move(high, low)) > 0) { 933 /* 934 * In case the target isn't the current cpu IPI it to force a 935 * reschedule with the new workload. 936 */ 937 cpu = TDQ_ID(low); 938 if (cpu != PCPU_GET(cpuid)) 939 ipi_cpu(cpu, IPI_PREEMPT); 940 } 941 tdq_unlock_pair(high, low); 942 return (moved); 943} 944 945/* 946 * Move a thread from one thread queue to another. 947 */ 948static int 949tdq_move(struct tdq *from, struct tdq *to) 950{ 951 struct td_sched *ts; 952 struct thread *td; 953 struct tdq *tdq; 954 int cpu; 955 956 TDQ_LOCK_ASSERT(from, MA_OWNED); 957 TDQ_LOCK_ASSERT(to, MA_OWNED); 958 959 tdq = from; 960 cpu = TDQ_ID(to); 961 td = tdq_steal(tdq, cpu); 962 if (td == NULL) 963 return (0); 964 ts = td->td_sched; 965 /* 966 * Although the run queue is locked the thread may be blocked. Lock 967 * it to clear this and acquire the run-queue lock. 968 */ 969 thread_lock(td); 970 /* Drop recursive lock on from acquired via thread_lock(). */ 971 TDQ_UNLOCK(from); 972 sched_rem(td); 973 ts->ts_cpu = cpu; 974 td->td_lock = TDQ_LOCKPTR(to); 975 tdq_add(to, td, SRQ_YIELDING); 976 return (1); 977} 978 979/* 980 * This tdq has idled. Try to steal a thread from another cpu and switch 981 * to it. 982 */ 983static int 984tdq_idled(struct tdq *tdq) 985{ 986 struct cpu_group *cg; 987 struct tdq *steal; 988 cpuset_t mask; 989 int thresh; 990 int cpu; 991 992 if (smp_started == 0 || steal_idle == 0) 993 return (1); 994 CPU_FILL(&mask); 995 CPU_CLR(PCPU_GET(cpuid), &mask); 996 /* We don't want to be preempted while we're iterating. */ 997 spinlock_enter(); 998 for (cg = tdq->tdq_cg; cg != NULL; ) { 999 if ((cg->cg_flags & CG_FLAG_THREAD) == 0) 1000 thresh = steal_thresh; 1001 else 1002 thresh = 1; 1003 cpu = sched_highest(cg, mask, thresh); 1004 if (cpu == -1) { 1005 cg = cg->cg_parent; 1006 continue; 1007 } 1008 steal = TDQ_CPU(cpu); 1009 CPU_CLR(cpu, &mask); 1010 tdq_lock_pair(tdq, steal); 1011 if (steal->tdq_load < thresh || steal->tdq_transferable == 0) { 1012 tdq_unlock_pair(tdq, steal); 1013 continue; 1014 } 1015 /* 1016 * If a thread was added while interrupts were disabled don't 1017 * steal one here. If we fail to acquire one due to affinity 1018 * restrictions loop again with this cpu removed from the 1019 * set. 1020 */ 1021 if (tdq->tdq_load == 0 && tdq_move(steal, tdq) == 0) { 1022 tdq_unlock_pair(tdq, steal); 1023 continue; 1024 } 1025 spinlock_exit(); 1026 TDQ_UNLOCK(steal); 1027 mi_switch(SW_VOL | SWT_IDLE, NULL); 1028 thread_unlock(curthread); 1029 1030 return (0); 1031 } 1032 spinlock_exit(); 1033 return (1); 1034} 1035 1036/* 1037 * Notify a remote cpu of new work. Sends an IPI if criteria are met. 1038 */ 1039static void 1040tdq_notify(struct tdq *tdq, struct thread *td) 1041{ 1042 struct thread *ctd; 1043 int pri; 1044 int cpu; 1045 1046 if (tdq->tdq_ipipending) 1047 return; 1048 cpu = td->td_sched->ts_cpu; 1049 pri = td->td_priority; 1050 ctd = pcpu_find(cpu)->pc_curthread; 1051 if (!sched_shouldpreempt(pri, ctd->td_priority, 1)) 1052 return; 1053 1054 /* 1055 * Make sure that our caller's earlier update to tdq_load is 1056 * globally visible before we read tdq_cpu_idle. Idle thread 1057 * accesses both of them without locks, and the order is important. 1058 */ 1059 mb(); 1060 1061 if (TD_IS_IDLETHREAD(ctd)) { 1062 /* 1063 * If the MD code has an idle wakeup routine try that before 1064 * falling back to IPI. 1065 */ 1066 if (!tdq->tdq_cpu_idle || cpu_idle_wakeup(cpu)) 1067 return; 1068 } 1069 tdq->tdq_ipipending = 1; 1070 ipi_cpu(cpu, IPI_PREEMPT); 1071} 1072 1073/* 1074 * Steals load from a timeshare queue. Honors the rotating queue head 1075 * index. 1076 */ 1077static struct thread * 1078runq_steal_from(struct runq *rq, int cpu, u_char start) 1079{ 1080 struct rqbits *rqb; 1081 struct rqhead *rqh; 1082 struct thread *td, *first; 1083 int bit; 1084 int i; 1085 1086 rqb = &rq->rq_status; 1087 bit = start & (RQB_BPW -1); 1088 first = NULL; 1089again: 1090 for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) { 1091 if (rqb->rqb_bits[i] == 0) 1092 continue; 1093 if (bit == 0) 1094 bit = RQB_FFS(rqb->rqb_bits[i]); 1095 for (; bit < RQB_BPW; bit++) { 1096 if ((rqb->rqb_bits[i] & (1ul << bit)) == 0) 1097 continue; 1098 rqh = &rq->rq_queues[bit + (i << RQB_L2BPW)]; 1099 TAILQ_FOREACH(td, rqh, td_runq) { 1100 if (first && THREAD_CAN_MIGRATE(td) && 1101 THREAD_CAN_SCHED(td, cpu)) 1102 return (td); 1103 first = td; 1104 } 1105 } 1106 } 1107 if (start != 0) { 1108 start = 0; 1109 goto again; 1110 } 1111 1112 if (first && THREAD_CAN_MIGRATE(first) && 1113 THREAD_CAN_SCHED(first, cpu)) 1114 return (first); 1115 return (NULL); 1116} 1117 1118/* 1119 * Steals load from a standard linear queue. 1120 */ 1121static struct thread * 1122runq_steal(struct runq *rq, int cpu) 1123{ 1124 struct rqhead *rqh; 1125 struct rqbits *rqb; 1126 struct thread *td; 1127 int word; 1128 int bit; 1129 1130 rqb = &rq->rq_status; 1131 for (word = 0; word < RQB_LEN; word++) { 1132 if (rqb->rqb_bits[word] == 0) 1133 continue; 1134 for (bit = 0; bit < RQB_BPW; bit++) { 1135 if ((rqb->rqb_bits[word] & (1ul << bit)) == 0) 1136 continue; 1137 rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)]; 1138 TAILQ_FOREACH(td, rqh, td_runq) 1139 if (THREAD_CAN_MIGRATE(td) && 1140 THREAD_CAN_SCHED(td, cpu)) 1141 return (td); 1142 } 1143 } 1144 return (NULL); 1145} 1146 1147/* 1148 * Attempt to steal a thread in priority order from a thread queue. 1149 */ 1150static struct thread * 1151tdq_steal(struct tdq *tdq, int cpu) 1152{ 1153 struct thread *td; 1154 1155 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 1156 if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL) 1157 return (td); 1158 if ((td = runq_steal_from(&tdq->tdq_timeshare, 1159 cpu, tdq->tdq_ridx)) != NULL) 1160 return (td); 1161 return (runq_steal(&tdq->tdq_idle, cpu)); 1162} 1163 1164/* 1165 * Sets the thread lock and ts_cpu to match the requested cpu. Unlocks the 1166 * current lock and returns with the assigned queue locked. 1167 */ 1168static inline struct tdq * 1169sched_setcpu(struct thread *td, int cpu, int flags) 1170{ 1171 1172 struct tdq *tdq; 1173 1174 THREAD_LOCK_ASSERT(td, MA_OWNED); 1175 tdq = TDQ_CPU(cpu); 1176 td->td_sched->ts_cpu = cpu; 1177 /* 1178 * If the lock matches just return the queue. 1179 */ 1180 if (td->td_lock == TDQ_LOCKPTR(tdq)) 1181 return (tdq); 1182#ifdef notyet 1183 /* 1184 * If the thread isn't running its lockptr is a 1185 * turnstile or a sleepqueue. We can just lock_set without 1186 * blocking. 1187 */ 1188 if (TD_CAN_RUN(td)) { 1189 TDQ_LOCK(tdq); 1190 thread_lock_set(td, TDQ_LOCKPTR(tdq)); 1191 return (tdq); 1192 } 1193#endif 1194 /* 1195 * The hard case, migration, we need to block the thread first to 1196 * prevent order reversals with other cpus locks. 1197 */ 1198 spinlock_enter(); 1199 thread_lock_block(td); 1200 TDQ_LOCK(tdq); 1201 thread_lock_unblock(td, TDQ_LOCKPTR(tdq)); 1202 spinlock_exit(); 1203 return (tdq); 1204} 1205 1206SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding"); 1207SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity"); 1208SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity"); 1209SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load"); 1210SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu"); 1211SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration"); 1212 1213static int 1214sched_pickcpu(struct thread *td, int flags) 1215{ 1216 struct cpu_group *cg, *ccg; 1217 struct td_sched *ts; 1218 struct tdq *tdq; 1219 cpuset_t mask; 1220 int cpu, pri, self; 1221 1222 self = PCPU_GET(cpuid); 1223 ts = td->td_sched; 1224 if (smp_started == 0) 1225 return (self); 1226 /* 1227 * Don't migrate a running thread from sched_switch(). 1228 */ 1229 if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td)) 1230 return (ts->ts_cpu); 1231 /* 1232 * Prefer to run interrupt threads on the processors that generate 1233 * the interrupt. 1234 */ 1235 pri = td->td_priority; 1236 if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) && 1237 curthread->td_intr_nesting_level && ts->ts_cpu != self) { 1238 SCHED_STAT_INC(pickcpu_intrbind); 1239 ts->ts_cpu = self; 1240 if (TDQ_CPU(self)->tdq_lowpri > pri) { 1241 SCHED_STAT_INC(pickcpu_affinity); 1242 return (ts->ts_cpu); 1243 } 1244 } 1245 /* 1246 * If the thread can run on the last cpu and the affinity has not 1247 * expired or it is idle run it there. 1248 */ 1249 tdq = TDQ_CPU(ts->ts_cpu); 1250 cg = tdq->tdq_cg; 1251 if (THREAD_CAN_SCHED(td, ts->ts_cpu) && 1252 tdq->tdq_lowpri >= PRI_MIN_IDLE && 1253 SCHED_AFFINITY(ts, CG_SHARE_L2)) { 1254 if (cg->cg_flags & CG_FLAG_THREAD) { 1255 CPUSET_FOREACH(cpu, cg->cg_mask) { 1256 if (TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE) 1257 break; 1258 } 1259 } else 1260 cpu = INT_MAX; 1261 if (cpu > mp_maxid) { 1262 SCHED_STAT_INC(pickcpu_idle_affinity); 1263 return (ts->ts_cpu); 1264 } 1265 } 1266 /* 1267 * Search for the last level cache CPU group in the tree. 1268 * Skip caches with expired affinity time and SMT groups. 1269 * Affinity to higher level caches will be handled less aggressively. 1270 */ 1271 for (ccg = NULL; cg != NULL; cg = cg->cg_parent) { 1272 if (cg->cg_flags & CG_FLAG_THREAD) 1273 continue; 1274 if (!SCHED_AFFINITY(ts, cg->cg_level)) 1275 continue; 1276 ccg = cg; 1277 } 1278 if (ccg != NULL) 1279 cg = ccg; 1280 cpu = -1; 1281 /* Search the group for the less loaded idle CPU we can run now. */ 1282 mask = td->td_cpuset->cs_mask; 1283 if (cg != NULL && cg != cpu_top && 1284 CPU_CMP(&cg->cg_mask, &cpu_top->cg_mask) != 0) 1285 cpu = sched_lowest(cg, mask, max(pri, PRI_MAX_TIMESHARE), 1286 INT_MAX, ts->ts_cpu); 1287 /* Search globally for the less loaded CPU we can run now. */ 1288 if (cpu == -1) 1289 cpu = sched_lowest(cpu_top, mask, pri, INT_MAX, ts->ts_cpu); 1290 /* Search globally for the less loaded CPU. */ 1291 if (cpu == -1) 1292 cpu = sched_lowest(cpu_top, mask, -1, INT_MAX, ts->ts_cpu); 1293 KASSERT(cpu != -1, ("sched_pickcpu: Failed to find a cpu.")); 1294 /* 1295 * Compare the lowest loaded cpu to current cpu. 1296 */ 1297 if (THREAD_CAN_SCHED(td, self) && TDQ_CPU(self)->tdq_lowpri > pri && 1298 TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE && 1299 TDQ_CPU(self)->tdq_load <= TDQ_CPU(cpu)->tdq_load + 1) { 1300 SCHED_STAT_INC(pickcpu_local); 1301 cpu = self; 1302 } else 1303 SCHED_STAT_INC(pickcpu_lowest); 1304 if (cpu != ts->ts_cpu) 1305 SCHED_STAT_INC(pickcpu_migration); 1306 return (cpu); 1307} 1308#endif 1309 1310/* 1311 * Pick the highest priority task we have and return it. 1312 */ 1313static struct thread * 1314tdq_choose(struct tdq *tdq) 1315{ 1316 struct thread *td; 1317 1318 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 1319 td = runq_choose(&tdq->tdq_realtime); 1320 if (td != NULL) 1321 return (td); 1322 td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx); 1323 if (td != NULL) { 1324 KASSERT(td->td_priority >= PRI_MIN_BATCH, 1325 ("tdq_choose: Invalid priority on timeshare queue %d", 1326 td->td_priority)); 1327 return (td); 1328 } 1329 td = runq_choose(&tdq->tdq_idle); 1330 if (td != NULL) { 1331 KASSERT(td->td_priority >= PRI_MIN_IDLE, 1332 ("tdq_choose: Invalid priority on idle queue %d", 1333 td->td_priority)); 1334 return (td); 1335 } 1336 1337 return (NULL); 1338} 1339 1340/* 1341 * Initialize a thread queue. 1342 */ 1343static void 1344tdq_setup(struct tdq *tdq) 1345{ 1346 1347 if (bootverbose) 1348 printf("ULE: setup cpu %d\n", TDQ_ID(tdq)); 1349 runq_init(&tdq->tdq_realtime); 1350 runq_init(&tdq->tdq_timeshare); 1351 runq_init(&tdq->tdq_idle); 1352 snprintf(tdq->tdq_name, sizeof(tdq->tdq_name), 1353 "sched lock %d", (int)TDQ_ID(tdq)); 1354 mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock", 1355 MTX_SPIN | MTX_RECURSE); 1356#ifdef KTR 1357 snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname), 1358 "CPU %d load", (int)TDQ_ID(tdq)); 1359#endif 1360} 1361 1362#ifdef SMP 1363static void 1364sched_setup_smp(void) 1365{ 1366 struct tdq *tdq; 1367 int i; 1368 1369 cpu_top = smp_topo(); 1370 CPU_FOREACH(i) { 1371 tdq = TDQ_CPU(i); 1372 tdq_setup(tdq); 1373 tdq->tdq_cg = smp_topo_find(cpu_top, i); 1374 if (tdq->tdq_cg == NULL) 1375 panic("Can't find cpu group for %d\n", i); 1376 } 1377 balance_tdq = TDQ_SELF(); 1378 sched_balance(); 1379} 1380#endif 1381 1382/* 1383 * Setup the thread queues and initialize the topology based on MD 1384 * information. 1385 */ 1386static void 1387sched_setup(void *dummy) 1388{ 1389 struct tdq *tdq; 1390 1391 tdq = TDQ_SELF(); 1392#ifdef SMP 1393 sched_setup_smp(); 1394#else 1395 tdq_setup(tdq); 1396#endif 1397 1398 /* Add thread0's load since it's running. */ 1399 TDQ_LOCK(tdq); 1400 thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF()); 1401 tdq_load_add(tdq, &thread0); 1402 tdq->tdq_lowpri = thread0.td_priority; 1403 TDQ_UNLOCK(tdq); 1404} 1405 1406/* 1407 * This routine determines time constants after stathz and hz are setup. 1408 */ 1409/* ARGSUSED */ 1410static void 1411sched_initticks(void *dummy) 1412{ 1413 int incr; 1414 1415 realstathz = stathz ? stathz : hz; 1416 sched_slice = realstathz / SCHED_SLICE_DEFAULT_DIVISOR; 1417 sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR; 1418 hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) / 1419 realstathz); 1420 1421 /* 1422 * tickincr is shifted out by 10 to avoid rounding errors due to 1423 * hz not being evenly divisible by stathz on all platforms. 1424 */ 1425 incr = (hz << SCHED_TICK_SHIFT) / realstathz; 1426 /* 1427 * This does not work for values of stathz that are more than 1428 * 1 << SCHED_TICK_SHIFT * hz. In practice this does not happen. 1429 */ 1430 if (incr == 0) 1431 incr = 1; 1432 tickincr = incr; 1433#ifdef SMP 1434 /* 1435 * Set the default balance interval now that we know 1436 * what realstathz is. 1437 */ 1438 balance_interval = realstathz; 1439 affinity = SCHED_AFFINITY_DEFAULT; 1440#endif 1441 if (sched_idlespinthresh < 0) 1442 sched_idlespinthresh = 2 * max(10000, 6 * hz) / realstathz; 1443} 1444 1445 1446/* 1447 * This is the core of the interactivity algorithm. Determines a score based 1448 * on past behavior. It is the ratio of sleep time to run time scaled to 1449 * a [0, 100] integer. This is the voluntary sleep time of a process, which 1450 * differs from the cpu usage because it does not account for time spent 1451 * waiting on a run-queue. Would be prettier if we had floating point. 1452 */ 1453static int 1454sched_interact_score(struct thread *td) 1455{ 1456 struct td_sched *ts; 1457 int div; 1458 1459 ts = td->td_sched; 1460 /* 1461 * The score is only needed if this is likely to be an interactive 1462 * task. Don't go through the expense of computing it if there's 1463 * no chance. 1464 */ 1465 if (sched_interact <= SCHED_INTERACT_HALF && 1466 ts->ts_runtime >= ts->ts_slptime) 1467 return (SCHED_INTERACT_HALF); 1468 1469 if (ts->ts_runtime > ts->ts_slptime) { 1470 div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF); 1471 return (SCHED_INTERACT_HALF + 1472 (SCHED_INTERACT_HALF - (ts->ts_slptime / div))); 1473 } 1474 if (ts->ts_slptime > ts->ts_runtime) { 1475 div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF); 1476 return (ts->ts_runtime / div); 1477 } 1478 /* runtime == slptime */ 1479 if (ts->ts_runtime) 1480 return (SCHED_INTERACT_HALF); 1481 1482 /* 1483 * This can happen if slptime and runtime are 0. 1484 */ 1485 return (0); 1486 1487} 1488 1489/* 1490 * Scale the scheduling priority according to the "interactivity" of this 1491 * process. 1492 */ 1493static void 1494sched_priority(struct thread *td) 1495{ 1496 int score; 1497 int pri; 1498 1499 if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE) 1500 return; 1501 /* 1502 * If the score is interactive we place the thread in the realtime 1503 * queue with a priority that is less than kernel and interrupt 1504 * priorities. These threads are not subject to nice restrictions. 1505 * 1506 * Scores greater than this are placed on the normal timeshare queue 1507 * where the priority is partially decided by the most recent cpu 1508 * utilization and the rest is decided by nice value. 1509 * 1510 * The nice value of the process has a linear effect on the calculated 1511 * score. Negative nice values make it easier for a thread to be 1512 * considered interactive. 1513 */ 1514 score = imax(0, sched_interact_score(td) + td->td_proc->p_nice); 1515 if (score < sched_interact) { 1516 pri = PRI_MIN_INTERACT; 1517 pri += ((PRI_MAX_INTERACT - PRI_MIN_INTERACT + 1) / 1518 sched_interact) * score; 1519 KASSERT(pri >= PRI_MIN_INTERACT && pri <= PRI_MAX_INTERACT, 1520 ("sched_priority: invalid interactive priority %d score %d", 1521 pri, score)); 1522 } else { 1523 pri = SCHED_PRI_MIN; 1524 if (td->td_sched->ts_ticks) 1525 pri += min(SCHED_PRI_TICKS(td->td_sched), 1526 SCHED_PRI_RANGE - 1); 1527 pri += SCHED_PRI_NICE(td->td_proc->p_nice); 1528 KASSERT(pri >= PRI_MIN_BATCH && pri <= PRI_MAX_BATCH, 1529 ("sched_priority: invalid priority %d: nice %d, " 1530 "ticks %d ftick %d ltick %d tick pri %d", 1531 pri, td->td_proc->p_nice, td->td_sched->ts_ticks, 1532 td->td_sched->ts_ftick, td->td_sched->ts_ltick, 1533 SCHED_PRI_TICKS(td->td_sched))); 1534 } 1535 sched_user_prio(td, pri); 1536 1537 return; 1538} 1539 1540/* 1541 * This routine enforces a maximum limit on the amount of scheduling history 1542 * kept. It is called after either the slptime or runtime is adjusted. This 1543 * function is ugly due to integer math. 1544 */ 1545static void 1546sched_interact_update(struct thread *td) 1547{ 1548 struct td_sched *ts; 1549 u_int sum; 1550 1551 ts = td->td_sched; 1552 sum = ts->ts_runtime + ts->ts_slptime; 1553 if (sum < SCHED_SLP_RUN_MAX) 1554 return; 1555 /* 1556 * This only happens from two places: 1557 * 1) We have added an unusual amount of run time from fork_exit. 1558 * 2) We have added an unusual amount of sleep time from sched_sleep(). 1559 */ 1560 if (sum > SCHED_SLP_RUN_MAX * 2) { 1561 if (ts->ts_runtime > ts->ts_slptime) { 1562 ts->ts_runtime = SCHED_SLP_RUN_MAX; 1563 ts->ts_slptime = 1; 1564 } else { 1565 ts->ts_slptime = SCHED_SLP_RUN_MAX; 1566 ts->ts_runtime = 1; 1567 } 1568 return; 1569 } 1570 /* 1571 * If we have exceeded by more than 1/5th then the algorithm below 1572 * will not bring us back into range. Dividing by two here forces 1573 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX] 1574 */ 1575 if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) { 1576 ts->ts_runtime /= 2; 1577 ts->ts_slptime /= 2; 1578 return; 1579 } 1580 ts->ts_runtime = (ts->ts_runtime / 5) * 4; 1581 ts->ts_slptime = (ts->ts_slptime / 5) * 4; 1582} 1583 1584/* 1585 * Scale back the interactivity history when a child thread is created. The 1586 * history is inherited from the parent but the thread may behave totally 1587 * differently. For example, a shell spawning a compiler process. We want 1588 * to learn that the compiler is behaving badly very quickly. 1589 */ 1590static void 1591sched_interact_fork(struct thread *td) 1592{ 1593 int ratio; 1594 int sum; 1595 1596 sum = td->td_sched->ts_runtime + td->td_sched->ts_slptime; 1597 if (sum > SCHED_SLP_RUN_FORK) { 1598 ratio = sum / SCHED_SLP_RUN_FORK; 1599 td->td_sched->ts_runtime /= ratio; 1600 td->td_sched->ts_slptime /= ratio; 1601 } 1602} 1603 1604/* 1605 * Called from proc0_init() to setup the scheduler fields. 1606 */ 1607void 1608schedinit(void) 1609{ 1610 1611 /* 1612 * Set up the scheduler specific parts of proc0. 1613 */ 1614 proc0.p_sched = NULL; /* XXX */ 1615 thread0.td_sched = &td_sched0; 1616 td_sched0.ts_ltick = ticks; 1617 td_sched0.ts_ftick = ticks; 1618 td_sched0.ts_slice = 0; 1619} 1620 1621/* 1622 * This is only somewhat accurate since given many processes of the same 1623 * priority they will switch when their slices run out, which will be 1624 * at most sched_slice stathz ticks. 1625 */ 1626int 1627sched_rr_interval(void) 1628{ 1629 1630 /* Convert sched_slice from stathz to hz. */ 1631 return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz)); 1632} 1633 1634/* 1635 * Update the percent cpu tracking information when it is requested or 1636 * the total history exceeds the maximum. We keep a sliding history of 1637 * tick counts that slowly decays. This is less precise than the 4BSD 1638 * mechanism since it happens with less regular and frequent events. 1639 */ 1640static void 1641sched_pctcpu_update(struct td_sched *ts, int run) 1642{ 1643 int t = ticks; 1644 1645 if (t - ts->ts_ltick >= SCHED_TICK_TARG) { 1646 ts->ts_ticks = 0; 1647 ts->ts_ftick = t - SCHED_TICK_TARG; 1648 } else if (t - ts->ts_ftick >= SCHED_TICK_MAX) { 1649 ts->ts_ticks = (ts->ts_ticks / (ts->ts_ltick - ts->ts_ftick)) * 1650 (ts->ts_ltick - (t - SCHED_TICK_TARG)); 1651 ts->ts_ftick = t - SCHED_TICK_TARG; 1652 } 1653 if (run) 1654 ts->ts_ticks += (t - ts->ts_ltick) << SCHED_TICK_SHIFT; 1655 ts->ts_ltick = t; 1656} 1657 1658/* 1659 * Adjust the priority of a thread. Move it to the appropriate run-queue 1660 * if necessary. This is the back-end for several priority related 1661 * functions. 1662 */ 1663static void 1664sched_thread_priority(struct thread *td, u_char prio) 1665{ 1666 struct td_sched *ts; 1667 struct tdq *tdq; 1668 int oldpri; 1669 1670 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio", 1671 "prio:%d", td->td_priority, "new prio:%d", prio, 1672 KTR_ATTR_LINKED, sched_tdname(curthread)); 1673 SDT_PROBE3(sched, , , change__pri, td, td->td_proc, prio); 1674 if (td != curthread && prio < td->td_priority) { 1675 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread), 1676 "lend prio", "prio:%d", td->td_priority, "new prio:%d", 1677 prio, KTR_ATTR_LINKED, sched_tdname(td)); 1678 SDT_PROBE4(sched, , , lend__pri, td, td->td_proc, prio, 1679 curthread); 1680 } 1681 ts = td->td_sched; 1682 THREAD_LOCK_ASSERT(td, MA_OWNED); 1683 if (td->td_priority == prio) 1684 return; 1685 /* 1686 * If the priority has been elevated due to priority 1687 * propagation, we may have to move ourselves to a new 1688 * queue. This could be optimized to not re-add in some 1689 * cases. 1690 */ 1691 if (TD_ON_RUNQ(td) && prio < td->td_priority) { 1692 sched_rem(td); 1693 td->td_priority = prio; 1694 sched_add(td, SRQ_BORROWING); 1695 return; 1696 } 1697 /* 1698 * If the thread is currently running we may have to adjust the lowpri 1699 * information so other cpus are aware of our current priority. 1700 */ 1701 if (TD_IS_RUNNING(td)) { 1702 tdq = TDQ_CPU(ts->ts_cpu); 1703 oldpri = td->td_priority; 1704 td->td_priority = prio; 1705 if (prio < tdq->tdq_lowpri) 1706 tdq->tdq_lowpri = prio; 1707 else if (tdq->tdq_lowpri == oldpri) 1708 tdq_setlowpri(tdq, td); 1709 return; 1710 } 1711 td->td_priority = prio; 1712} 1713 1714/* 1715 * Update a thread's priority when it is lent another thread's 1716 * priority. 1717 */ 1718void 1719sched_lend_prio(struct thread *td, u_char prio) 1720{ 1721 1722 td->td_flags |= TDF_BORROWING; 1723 sched_thread_priority(td, prio); 1724} 1725 1726/* 1727 * Restore a thread's priority when priority propagation is 1728 * over. The prio argument is the minimum priority the thread 1729 * needs to have to satisfy other possible priority lending 1730 * requests. If the thread's regular priority is less 1731 * important than prio, the thread will keep a priority boost 1732 * of prio. 1733 */ 1734void 1735sched_unlend_prio(struct thread *td, u_char prio) 1736{ 1737 u_char base_pri; 1738 1739 if (td->td_base_pri >= PRI_MIN_TIMESHARE && 1740 td->td_base_pri <= PRI_MAX_TIMESHARE) 1741 base_pri = td->td_user_pri; 1742 else 1743 base_pri = td->td_base_pri; 1744 if (prio >= base_pri) { 1745 td->td_flags &= ~TDF_BORROWING; 1746 sched_thread_priority(td, base_pri); 1747 } else 1748 sched_lend_prio(td, prio); 1749} 1750 1751/* 1752 * Standard entry for setting the priority to an absolute value. 1753 */ 1754void 1755sched_prio(struct thread *td, u_char prio) 1756{ 1757 u_char oldprio; 1758 1759 /* First, update the base priority. */ 1760 td->td_base_pri = prio; 1761 1762 /* 1763 * If the thread is borrowing another thread's priority, don't 1764 * ever lower the priority. 1765 */ 1766 if (td->td_flags & TDF_BORROWING && td->td_priority < prio) 1767 return; 1768 1769 /* Change the real priority. */ 1770 oldprio = td->td_priority; 1771 sched_thread_priority(td, prio); 1772 1773 /* 1774 * If the thread is on a turnstile, then let the turnstile update 1775 * its state. 1776 */ 1777 if (TD_ON_LOCK(td) && oldprio != prio) 1778 turnstile_adjust(td, oldprio); 1779} 1780 1781/* 1782 * Set the base user priority, does not effect current running priority. 1783 */ 1784void 1785sched_user_prio(struct thread *td, u_char prio) 1786{ 1787 1788 td->td_base_user_pri = prio; 1789 if (td->td_lend_user_pri <= prio) 1790 return; 1791 td->td_user_pri = prio; 1792} 1793 1794void 1795sched_lend_user_prio(struct thread *td, u_char prio) 1796{ 1797 1798 THREAD_LOCK_ASSERT(td, MA_OWNED); 1799 td->td_lend_user_pri = prio; 1800 td->td_user_pri = min(prio, td->td_base_user_pri); 1801 if (td->td_priority > td->td_user_pri) 1802 sched_prio(td, td->td_user_pri); 1803 else if (td->td_priority != td->td_user_pri) 1804 td->td_flags |= TDF_NEEDRESCHED; 1805} 1806 1807/* 1808 * Handle migration from sched_switch(). This happens only for 1809 * cpu binding. 1810 */ 1811static struct mtx * 1812sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags) 1813{ 1814 struct tdq *tdn; 1815 1816 tdn = TDQ_CPU(td->td_sched->ts_cpu); 1817#ifdef SMP 1818 tdq_load_rem(tdq, td); 1819 /* 1820 * Do the lock dance required to avoid LOR. We grab an extra 1821 * spinlock nesting to prevent preemption while we're 1822 * not holding either run-queue lock. 1823 */ 1824 spinlock_enter(); 1825 thread_lock_block(td); /* This releases the lock on tdq. */ 1826 1827 /* 1828 * Acquire both run-queue locks before placing the thread on the new 1829 * run-queue to avoid deadlocks created by placing a thread with a 1830 * blocked lock on the run-queue of a remote processor. The deadlock 1831 * occurs when a third processor attempts to lock the two queues in 1832 * question while the target processor is spinning with its own 1833 * run-queue lock held while waiting for the blocked lock to clear. 1834 */ 1835 tdq_lock_pair(tdn, tdq); 1836 tdq_add(tdn, td, flags); 1837 tdq_notify(tdn, td); 1838 TDQ_UNLOCK(tdn); 1839 spinlock_exit(); 1840#endif 1841 return (TDQ_LOCKPTR(tdn)); 1842} 1843 1844/* 1845 * Variadic version of thread_lock_unblock() that does not assume td_lock 1846 * is blocked. 1847 */ 1848static inline void 1849thread_unblock_switch(struct thread *td, struct mtx *mtx) 1850{ 1851 atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock, 1852 (uintptr_t)mtx); 1853} 1854 1855/* 1856 * Switch threads. This function has to handle threads coming in while 1857 * blocked for some reason, running, or idle. It also must deal with 1858 * migrating a thread from one queue to another as running threads may 1859 * be assigned elsewhere via binding. 1860 */ 1861void 1862sched_switch(struct thread *td, struct thread *newtd, int flags) 1863{ 1864 struct tdq *tdq; 1865 struct td_sched *ts; 1866 struct mtx *mtx; 1867 int srqflag; 1868 int cpuid, preempted; 1869 1870 THREAD_LOCK_ASSERT(td, MA_OWNED); 1871 KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument")); 1872 1873 cpuid = PCPU_GET(cpuid); 1874 tdq = TDQ_CPU(cpuid); 1875 ts = td->td_sched; 1876 mtx = td->td_lock; 1877 sched_pctcpu_update(ts, 1); 1878 ts->ts_rltick = ticks; 1879 td->td_lastcpu = td->td_oncpu; 1880 td->td_oncpu = NOCPU; 1881 preempted = !((td->td_flags & TDF_SLICEEND) || 1882 (flags & SWT_RELINQUISH)); 1883 td->td_flags &= ~(TDF_NEEDRESCHED | TDF_SLICEEND); 1884 td->td_owepreempt = 0; 1885 if (!TD_IS_IDLETHREAD(td)) 1886 tdq->tdq_switchcnt++; 1887 /* 1888 * The lock pointer in an idle thread should never change. Reset it 1889 * to CAN_RUN as well. 1890 */ 1891 if (TD_IS_IDLETHREAD(td)) { 1892 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1893 TD_SET_CAN_RUN(td); 1894 } else if (TD_IS_RUNNING(td)) { 1895 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1896 srqflag = preempted ? 1897 SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED : 1898 SRQ_OURSELF|SRQ_YIELDING; 1899#ifdef SMP 1900 if (THREAD_CAN_MIGRATE(td) && !THREAD_CAN_SCHED(td, ts->ts_cpu)) 1901 ts->ts_cpu = sched_pickcpu(td, 0); 1902#endif 1903 if (ts->ts_cpu == cpuid) 1904 tdq_runq_add(tdq, td, srqflag); 1905 else { 1906 KASSERT(THREAD_CAN_MIGRATE(td) || 1907 (ts->ts_flags & TSF_BOUND) != 0, 1908 ("Thread %p shouldn't migrate", td)); 1909 mtx = sched_switch_migrate(tdq, td, srqflag); 1910 } 1911 } else { 1912 /* This thread must be going to sleep. */ 1913 TDQ_LOCK(tdq); 1914 mtx = thread_lock_block(td); 1915 tdq_load_rem(tdq, td); 1916 } 1917 /* 1918 * We enter here with the thread blocked and assigned to the 1919 * appropriate cpu run-queue or sleep-queue and with the current 1920 * thread-queue locked. 1921 */ 1922 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED); 1923 newtd = choosethread(); 1924 /* 1925 * Call the MD code to switch contexts if necessary. 1926 */ 1927 if (td != newtd) { 1928#ifdef HWPMC_HOOKS 1929 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1930 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT); 1931#endif 1932 SDT_PROBE2(sched, , , off__cpu, newtd, newtd->td_proc); 1933 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object); 1934 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd; 1935 sched_pctcpu_update(newtd->td_sched, 0); 1936 1937#ifdef KDTRACE_HOOKS 1938 /* 1939 * If DTrace has set the active vtime enum to anything 1940 * other than INACTIVE (0), then it should have set the 1941 * function to call. 1942 */ 1943 if (dtrace_vtime_active) 1944 (*dtrace_vtime_switch_func)(newtd); 1945#endif 1946 1947 cpu_switch(td, newtd, mtx); 1948 /* 1949 * We may return from cpu_switch on a different cpu. However, 1950 * we always return with td_lock pointing to the current cpu's 1951 * run queue lock. 1952 */ 1953 cpuid = PCPU_GET(cpuid); 1954 tdq = TDQ_CPU(cpuid); 1955 lock_profile_obtain_lock_success( 1956 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__); 1957 1958 SDT_PROBE0(sched, , , on__cpu); 1959#ifdef HWPMC_HOOKS 1960 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1961 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN); 1962#endif 1963 } else { 1964 thread_unblock_switch(td, mtx); 1965 SDT_PROBE0(sched, , , remain__cpu); 1966 } 1967 /* 1968 * Assert that all went well and return. 1969 */ 1970 TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED); 1971 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1972 td->td_oncpu = cpuid; 1973} 1974 1975/* 1976 * Adjust thread priorities as a result of a nice request. 1977 */ 1978void 1979sched_nice(struct proc *p, int nice) 1980{ 1981 struct thread *td; 1982 1983 PROC_LOCK_ASSERT(p, MA_OWNED); 1984 1985 p->p_nice = nice; 1986 FOREACH_THREAD_IN_PROC(p, td) { 1987 thread_lock(td); 1988 sched_priority(td); 1989 sched_prio(td, td->td_base_user_pri); 1990 thread_unlock(td); 1991 } 1992} 1993 1994/* 1995 * Record the sleep time for the interactivity scorer. 1996 */ 1997void 1998sched_sleep(struct thread *td, int prio) 1999{ 2000 2001 THREAD_LOCK_ASSERT(td, MA_OWNED); 2002 2003 td->td_slptick = ticks; 2004 if (TD_IS_SUSPENDED(td) || prio >= PSOCK) 2005 td->td_flags |= TDF_CANSWAP; 2006 if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE) 2007 return; 2008 if (static_boost == 1 && prio) 2009 sched_prio(td, prio); 2010 else if (static_boost && td->td_priority > static_boost) 2011 sched_prio(td, static_boost); 2012} 2013 2014/* 2015 * Schedule a thread to resume execution and record how long it voluntarily 2016 * slept. We also update the pctcpu, interactivity, and priority. 2017 */ 2018void 2019sched_wakeup(struct thread *td) 2020{ 2021 struct td_sched *ts; 2022 int slptick; 2023 2024 THREAD_LOCK_ASSERT(td, MA_OWNED); 2025 ts = td->td_sched; 2026 td->td_flags &= ~TDF_CANSWAP; 2027 /* 2028 * If we slept for more than a tick update our interactivity and 2029 * priority. 2030 */ 2031 slptick = td->td_slptick; 2032 td->td_slptick = 0; 2033 if (slptick && slptick != ticks) { 2034 ts->ts_slptime += (ticks - slptick) << SCHED_TICK_SHIFT; 2035 sched_interact_update(td); 2036 sched_pctcpu_update(ts, 0); 2037 } 2038 /* 2039 * Reset the slice value since we slept and advanced the round-robin. 2040 */ 2041 ts->ts_slice = 0; 2042 sched_add(td, SRQ_BORING); 2043} 2044 2045/* 2046 * Penalize the parent for creating a new child and initialize the child's 2047 * priority. 2048 */ 2049void 2050sched_fork(struct thread *td, struct thread *child) 2051{ 2052 THREAD_LOCK_ASSERT(td, MA_OWNED); 2053 sched_pctcpu_update(td->td_sched, 1); 2054 sched_fork_thread(td, child); 2055 /* 2056 * Penalize the parent and child for forking. 2057 */ 2058 sched_interact_fork(child); 2059 sched_priority(child); 2060 td->td_sched->ts_runtime += tickincr; 2061 sched_interact_update(td); 2062 sched_priority(td); 2063} 2064 2065/* 2066 * Fork a new thread, may be within the same process. 2067 */ 2068void 2069sched_fork_thread(struct thread *td, struct thread *child) 2070{ 2071 struct td_sched *ts; 2072 struct td_sched *ts2; 2073 struct tdq *tdq; 2074 2075 tdq = TDQ_SELF(); 2076 THREAD_LOCK_ASSERT(td, MA_OWNED); 2077 /* 2078 * Initialize child. 2079 */ 2080 ts = td->td_sched; 2081 ts2 = child->td_sched; 2082 child->td_lock = TDQ_LOCKPTR(tdq); 2083 child->td_cpuset = cpuset_ref(td->td_cpuset); 2084 ts2->ts_cpu = ts->ts_cpu; 2085 ts2->ts_flags = 0; 2086 /* 2087 * Grab our parents cpu estimation information. 2088 */ 2089 ts2->ts_ticks = ts->ts_ticks; 2090 ts2->ts_ltick = ts->ts_ltick; 2091 ts2->ts_ftick = ts->ts_ftick; 2092 /* 2093 * Do not inherit any borrowed priority from the parent. 2094 */ 2095 child->td_priority = child->td_base_pri; 2096 /* 2097 * And update interactivity score. 2098 */ 2099 ts2->ts_slptime = ts->ts_slptime; 2100 ts2->ts_runtime = ts->ts_runtime; 2101 /* Attempt to quickly learn interactivity. */ 2102 ts2->ts_slice = tdq_slice(tdq) - sched_slice_min; 2103#ifdef KTR 2104 bzero(ts2->ts_name, sizeof(ts2->ts_name)); 2105#endif 2106} 2107 2108/* 2109 * Adjust the priority class of a thread. 2110 */ 2111void 2112sched_class(struct thread *td, int class) 2113{ 2114 2115 THREAD_LOCK_ASSERT(td, MA_OWNED); 2116 if (td->td_pri_class == class) 2117 return; 2118 td->td_pri_class = class; 2119} 2120 2121/* 2122 * Return some of the child's priority and interactivity to the parent. 2123 */ 2124void 2125sched_exit(struct proc *p, struct thread *child) 2126{ 2127 struct thread *td; 2128 2129 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit", 2130 "prio:%d", child->td_priority); 2131 PROC_LOCK_ASSERT(p, MA_OWNED); 2132 td = FIRST_THREAD_IN_PROC(p); 2133 sched_exit_thread(td, child); 2134} 2135 2136/* 2137 * Penalize another thread for the time spent on this one. This helps to 2138 * worsen the priority and interactivity of processes which schedule batch 2139 * jobs such as make. This has little effect on the make process itself but 2140 * causes new processes spawned by it to receive worse scores immediately. 2141 */ 2142void 2143sched_exit_thread(struct thread *td, struct thread *child) 2144{ 2145 2146 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit", 2147 "prio:%d", child->td_priority); 2148 /* 2149 * Give the child's runtime to the parent without returning the 2150 * sleep time as a penalty to the parent. This causes shells that 2151 * launch expensive things to mark their children as expensive. 2152 */ 2153 thread_lock(td); 2154 td->td_sched->ts_runtime += child->td_sched->ts_runtime; 2155 sched_interact_update(td); 2156 sched_priority(td); 2157 thread_unlock(td); 2158} 2159 2160void 2161sched_preempt(struct thread *td) 2162{ 2163 struct tdq *tdq; 2164 2165 SDT_PROBE2(sched, , , surrender, td, td->td_proc); 2166 2167 thread_lock(td); 2168 tdq = TDQ_SELF(); 2169 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2170 tdq->tdq_ipipending = 0; 2171 if (td->td_priority > tdq->tdq_lowpri) { 2172 int flags; 2173 2174 flags = SW_INVOL | SW_PREEMPT; 2175 if (td->td_critnest > 1) 2176 td->td_owepreempt = 1; 2177 else if (TD_IS_IDLETHREAD(td)) 2178 mi_switch(flags | SWT_REMOTEWAKEIDLE, NULL); 2179 else 2180 mi_switch(flags | SWT_REMOTEPREEMPT, NULL); 2181 } 2182 thread_unlock(td); 2183} 2184 2185/* 2186 * Fix priorities on return to user-space. Priorities may be elevated due 2187 * to static priorities in msleep() or similar. 2188 */ 2189void 2190sched_userret(struct thread *td) 2191{ 2192 /* 2193 * XXX we cheat slightly on the locking here to avoid locking in 2194 * the usual case. Setting td_priority here is essentially an 2195 * incomplete workaround for not setting it properly elsewhere. 2196 * Now that some interrupt handlers are threads, not setting it 2197 * properly elsewhere can clobber it in the window between setting 2198 * it here and returning to user mode, so don't waste time setting 2199 * it perfectly here. 2200 */ 2201 KASSERT((td->td_flags & TDF_BORROWING) == 0, 2202 ("thread with borrowed priority returning to userland")); 2203 if (td->td_priority != td->td_user_pri) { 2204 thread_lock(td); 2205 td->td_priority = td->td_user_pri; 2206 td->td_base_pri = td->td_user_pri; 2207 tdq_setlowpri(TDQ_SELF(), td); 2208 thread_unlock(td); 2209 } 2210} 2211 2212/* 2213 * Handle a stathz tick. This is really only relevant for timeshare 2214 * threads. 2215 */ 2216void 2217sched_clock(struct thread *td) 2218{ 2219 struct tdq *tdq; 2220 struct td_sched *ts; 2221 2222 THREAD_LOCK_ASSERT(td, MA_OWNED); 2223 tdq = TDQ_SELF(); 2224#ifdef SMP 2225 /* 2226 * We run the long term load balancer infrequently on the first cpu. 2227 */ 2228 if (balance_tdq == tdq) { 2229 if (balance_ticks && --balance_ticks == 0) 2230 sched_balance(); 2231 } 2232#endif 2233 /* 2234 * Save the old switch count so we have a record of the last ticks 2235 * activity. Initialize the new switch count based on our load. 2236 * If there is some activity seed it to reflect that. 2237 */ 2238 tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt; 2239 tdq->tdq_switchcnt = tdq->tdq_load; 2240 /* 2241 * Advance the insert index once for each tick to ensure that all 2242 * threads get a chance to run. 2243 */ 2244 if (tdq->tdq_idx == tdq->tdq_ridx) { 2245 tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS; 2246 if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx])) 2247 tdq->tdq_ridx = tdq->tdq_idx; 2248 } 2249 ts = td->td_sched; 2250 sched_pctcpu_update(ts, 1); 2251 if (td->td_pri_class & PRI_FIFO_BIT) 2252 return; 2253 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) { 2254 /* 2255 * We used a tick; charge it to the thread so 2256 * that we can compute our interactivity. 2257 */ 2258 td->td_sched->ts_runtime += tickincr; 2259 sched_interact_update(td); 2260 sched_priority(td); 2261 } 2262 2263 /* 2264 * Force a context switch if the current thread has used up a full 2265 * time slice (default is 100ms). 2266 */ 2267 if (!TD_IS_IDLETHREAD(td) && ++ts->ts_slice >= tdq_slice(tdq)) { 2268 ts->ts_slice = 0; 2269 td->td_flags |= TDF_NEEDRESCHED | TDF_SLICEEND; 2270 } 2271} 2272 2273/* 2274 * Called once per hz tick. 2275 */ 2276void 2277sched_tick(int cnt) 2278{ 2279 2280} 2281 2282/* 2283 * Return whether the current CPU has runnable tasks. Used for in-kernel 2284 * cooperative idle threads. 2285 */ 2286int 2287sched_runnable(void) 2288{ 2289 struct tdq *tdq; 2290 int load; 2291 2292 load = 1; 2293 2294 tdq = TDQ_SELF(); 2295 if ((curthread->td_flags & TDF_IDLETD) != 0) { 2296 if (tdq->tdq_load > 0) 2297 goto out; 2298 } else 2299 if (tdq->tdq_load - 1 > 0) 2300 goto out; 2301 load = 0; 2302out: 2303 return (load); 2304} 2305 2306/* 2307 * Choose the highest priority thread to run. The thread is removed from 2308 * the run-queue while running however the load remains. For SMP we set 2309 * the tdq in the global idle bitmask if it idles here. 2310 */ 2311struct thread * 2312sched_choose(void) 2313{ 2314 struct thread *td; 2315 struct tdq *tdq; 2316 2317 tdq = TDQ_SELF(); 2318 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2319 td = tdq_choose(tdq); 2320 if (td) { 2321 tdq_runq_rem(tdq, td); 2322 tdq->tdq_lowpri = td->td_priority; 2323 return (td); 2324 } 2325 tdq->tdq_lowpri = PRI_MAX_IDLE; 2326 return (PCPU_GET(idlethread)); 2327} 2328 2329/* 2330 * Set owepreempt if necessary. Preemption never happens directly in ULE, 2331 * we always request it once we exit a critical section. 2332 */ 2333static inline void 2334sched_setpreempt(struct thread *td) 2335{ 2336 struct thread *ctd; 2337 int cpri; 2338 int pri; 2339 2340 THREAD_LOCK_ASSERT(curthread, MA_OWNED); 2341 2342 ctd = curthread; 2343 pri = td->td_priority; 2344 cpri = ctd->td_priority; 2345 if (pri < cpri) 2346 ctd->td_flags |= TDF_NEEDRESCHED; 2347 if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd)) 2348 return; 2349 if (!sched_shouldpreempt(pri, cpri, 0)) 2350 return; 2351 ctd->td_owepreempt = 1; 2352} 2353 2354/* 2355 * Add a thread to a thread queue. Select the appropriate runq and add the 2356 * thread to it. This is the internal function called when the tdq is 2357 * predetermined. 2358 */ 2359void 2360tdq_add(struct tdq *tdq, struct thread *td, int flags) 2361{ 2362 2363 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2364 KASSERT((td->td_inhibitors == 0), 2365 ("sched_add: trying to run inhibited thread")); 2366 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)), 2367 ("sched_add: bad thread state")); 2368 KASSERT(td->td_flags & TDF_INMEM, 2369 ("sched_add: thread swapped out")); 2370 2371 if (td->td_priority < tdq->tdq_lowpri) 2372 tdq->tdq_lowpri = td->td_priority; 2373 tdq_runq_add(tdq, td, flags); 2374 tdq_load_add(tdq, td); 2375} 2376 2377/* 2378 * Select the target thread queue and add a thread to it. Request 2379 * preemption or IPI a remote processor if required. 2380 */ 2381void 2382sched_add(struct thread *td, int flags) 2383{ 2384 struct tdq *tdq; 2385#ifdef SMP 2386 int cpu; 2387#endif 2388 2389 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add", 2390 "prio:%d", td->td_priority, KTR_ATTR_LINKED, 2391 sched_tdname(curthread)); 2392 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup", 2393 KTR_ATTR_LINKED, sched_tdname(td)); 2394 SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL, 2395 flags & SRQ_PREEMPTED); 2396 THREAD_LOCK_ASSERT(td, MA_OWNED); 2397 /* 2398 * Recalculate the priority before we select the target cpu or 2399 * run-queue. 2400 */ 2401 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) 2402 sched_priority(td); 2403#ifdef SMP 2404 /* 2405 * Pick the destination cpu and if it isn't ours transfer to the 2406 * target cpu. 2407 */ 2408 cpu = sched_pickcpu(td, flags); 2409 tdq = sched_setcpu(td, cpu, flags); 2410 tdq_add(tdq, td, flags); 2411 if (cpu != PCPU_GET(cpuid)) { 2412 tdq_notify(tdq, td); 2413 return; 2414 } 2415#else 2416 tdq = TDQ_SELF(); 2417 TDQ_LOCK(tdq); 2418 /* 2419 * Now that the thread is moving to the run-queue, set the lock 2420 * to the scheduler's lock. 2421 */ 2422 thread_lock_set(td, TDQ_LOCKPTR(tdq)); 2423 tdq_add(tdq, td, flags); 2424#endif 2425 if (!(flags & SRQ_YIELDING)) 2426 sched_setpreempt(td); 2427} 2428 2429/* 2430 * Remove a thread from a run-queue without running it. This is used 2431 * when we're stealing a thread from a remote queue. Otherwise all threads 2432 * exit by calling sched_exit_thread() and sched_throw() themselves. 2433 */ 2434void 2435sched_rem(struct thread *td) 2436{ 2437 struct tdq *tdq; 2438 2439 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem", 2440 "prio:%d", td->td_priority); 2441 SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL); 2442 tdq = TDQ_CPU(td->td_sched->ts_cpu); 2443 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2444 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2445 KASSERT(TD_ON_RUNQ(td), 2446 ("sched_rem: thread not on run queue")); 2447 tdq_runq_rem(tdq, td); 2448 tdq_load_rem(tdq, td); 2449 TD_SET_CAN_RUN(td); 2450 if (td->td_priority == tdq->tdq_lowpri) 2451 tdq_setlowpri(tdq, NULL); 2452} 2453 2454/* 2455 * Fetch cpu utilization information. Updates on demand. 2456 */ 2457fixpt_t 2458sched_pctcpu(struct thread *td) 2459{ 2460 fixpt_t pctcpu; 2461 struct td_sched *ts; 2462 2463 pctcpu = 0; 2464 ts = td->td_sched; 2465 if (ts == NULL) 2466 return (0); 2467 2468 THREAD_LOCK_ASSERT(td, MA_OWNED); 2469 sched_pctcpu_update(ts, TD_IS_RUNNING(td)); 2470 if (ts->ts_ticks) { 2471 int rtick; 2472 2473 /* How many rtick per second ? */ 2474 rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz); 2475 pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT; 2476 } 2477 2478 return (pctcpu); 2479} 2480 2481/* 2482 * Enforce affinity settings for a thread. Called after adjustments to 2483 * cpumask. 2484 */ 2485void 2486sched_affinity(struct thread *td) 2487{ 2488#ifdef SMP 2489 struct td_sched *ts; 2490 2491 THREAD_LOCK_ASSERT(td, MA_OWNED); 2492 ts = td->td_sched; 2493 if (THREAD_CAN_SCHED(td, ts->ts_cpu)) 2494 return; 2495 if (TD_ON_RUNQ(td)) { 2496 sched_rem(td); 2497 sched_add(td, SRQ_BORING); 2498 return; 2499 } 2500 if (!TD_IS_RUNNING(td)) 2501 return; 2502 /* 2503 * Force a switch before returning to userspace. If the 2504 * target thread is not running locally send an ipi to force 2505 * the issue. 2506 */ 2507 td->td_flags |= TDF_NEEDRESCHED; 2508 if (td != curthread) 2509 ipi_cpu(ts->ts_cpu, IPI_PREEMPT); 2510#endif 2511} 2512 2513/* 2514 * Bind a thread to a target cpu. 2515 */ 2516void 2517sched_bind(struct thread *td, int cpu) 2518{ 2519 struct td_sched *ts; 2520 2521 THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED); 2522 KASSERT(td == curthread, ("sched_bind: can only bind curthread")); 2523 ts = td->td_sched; 2524 if (ts->ts_flags & TSF_BOUND) 2525 sched_unbind(td); 2526 KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td)); 2527 ts->ts_flags |= TSF_BOUND; 2528 sched_pin(); 2529 if (PCPU_GET(cpuid) == cpu) 2530 return; 2531 ts->ts_cpu = cpu; 2532 /* When we return from mi_switch we'll be on the correct cpu. */ 2533 mi_switch(SW_VOL, NULL); 2534} 2535 2536/* 2537 * Release a bound thread. 2538 */ 2539void 2540sched_unbind(struct thread *td) 2541{ 2542 struct td_sched *ts; 2543 2544 THREAD_LOCK_ASSERT(td, MA_OWNED); 2545 KASSERT(td == curthread, ("sched_unbind: can only bind curthread")); 2546 ts = td->td_sched; 2547 if ((ts->ts_flags & TSF_BOUND) == 0) 2548 return; 2549 ts->ts_flags &= ~TSF_BOUND; 2550 sched_unpin(); 2551} 2552 2553int 2554sched_is_bound(struct thread *td) 2555{ 2556 THREAD_LOCK_ASSERT(td, MA_OWNED); 2557 return (td->td_sched->ts_flags & TSF_BOUND); 2558} 2559 2560/* 2561 * Basic yield call. 2562 */ 2563void 2564sched_relinquish(struct thread *td) 2565{ 2566 thread_lock(td); 2567 mi_switch(SW_VOL | SWT_RELINQUISH, NULL); 2568 thread_unlock(td); 2569} 2570 2571/* 2572 * Return the total system load. 2573 */ 2574int 2575sched_load(void) 2576{ 2577#ifdef SMP 2578 int total; 2579 int i; 2580 2581 total = 0; 2582 CPU_FOREACH(i) 2583 total += TDQ_CPU(i)->tdq_sysload; 2584 return (total); 2585#else 2586 return (TDQ_SELF()->tdq_sysload); 2587#endif 2588} 2589 2590int 2591sched_sizeof_proc(void) 2592{ 2593 return (sizeof(struct proc)); 2594} 2595 2596int 2597sched_sizeof_thread(void) 2598{ 2599 return (sizeof(struct thread) + sizeof(struct td_sched)); 2600} 2601 2602#ifdef SMP 2603#define TDQ_IDLESPIN(tdq) \ 2604 ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0) 2605#else 2606#define TDQ_IDLESPIN(tdq) 1 2607#endif 2608 2609/* 2610 * The actual idle process. 2611 */ 2612void 2613sched_idletd(void *dummy) 2614{ 2615 struct thread *td; 2616 struct tdq *tdq; 2617 int oldswitchcnt, switchcnt; 2618 int i; 2619 2620 mtx_assert(&Giant, MA_NOTOWNED); 2621 td = curthread; 2622 tdq = TDQ_SELF(); 2623 THREAD_NO_SLEEPING(); 2624 oldswitchcnt = -1; 2625 for (;;) { 2626 if (tdq->tdq_load) { 2627 thread_lock(td); 2628 mi_switch(SW_VOL | SWT_IDLE, NULL); 2629 thread_unlock(td); 2630 } 2631 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt; 2632#ifdef SMP 2633 if (switchcnt != oldswitchcnt) { 2634 oldswitchcnt = switchcnt; 2635 if (tdq_idled(tdq) == 0) 2636 continue; 2637 } 2638 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt; 2639#else 2640 oldswitchcnt = switchcnt; 2641#endif 2642 /* 2643 * If we're switching very frequently, spin while checking 2644 * for load rather than entering a low power state that 2645 * may require an IPI. However, don't do any busy 2646 * loops while on SMT machines as this simply steals 2647 * cycles from cores doing useful work. 2648 */ 2649 if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) { 2650 for (i = 0; i < sched_idlespins; i++) { 2651 if (tdq->tdq_load) 2652 break; 2653 cpu_spinwait(); 2654 } 2655 } 2656 2657 /* If there was context switch during spin, restart it. */ 2658 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt; 2659 if (tdq->tdq_load != 0 || switchcnt != oldswitchcnt) 2660 continue; 2661 2662 /* Run main MD idle handler. */ 2663 tdq->tdq_cpu_idle = 1; 2664 /* 2665 * Make sure that tdq_cpu_idle update is globally visible 2666 * before cpu_idle() read tdq_load. The order is important 2667 * to avoid race with tdq_notify. 2668 */ 2669 mb(); 2670 cpu_idle(switchcnt * 4 > sched_idlespinthresh); 2671 tdq->tdq_cpu_idle = 0; 2672 2673 /* 2674 * Account thread-less hardware interrupts and 2675 * other wakeup reasons equal to context switches. 2676 */ 2677 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt; 2678 if (switchcnt != oldswitchcnt) 2679 continue; 2680 tdq->tdq_switchcnt++; 2681 oldswitchcnt++; 2682 } 2683} 2684 2685/* 2686 * A CPU is entering for the first time or a thread is exiting. 2687 */ 2688void 2689sched_throw(struct thread *td) 2690{ 2691 struct thread *newtd; 2692 struct tdq *tdq; 2693 2694 tdq = TDQ_SELF(); 2695 if (td == NULL) { 2696 /* Correct spinlock nesting and acquire the correct lock. */ 2697 TDQ_LOCK(tdq); 2698 spinlock_exit(); 2699 PCPU_SET(switchtime, cpu_ticks()); 2700 PCPU_SET(switchticks, ticks); 2701 } else { 2702 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2703 tdq_load_rem(tdq, td); 2704 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object); 2705 } 2706 KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count")); 2707 newtd = choosethread(); 2708 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd; 2709 cpu_throw(td, newtd); /* doesn't return */ 2710} 2711 2712/* 2713 * This is called from fork_exit(). Just acquire the correct locks and 2714 * let fork do the rest of the work. 2715 */ 2716void 2717sched_fork_exit(struct thread *td) 2718{ 2719 struct tdq *tdq; 2720 int cpuid; 2721 2722 /* 2723 * Finish setting up thread glue so that it begins execution in a 2724 * non-nested critical section with the scheduler lock held. 2725 */ 2726 cpuid = PCPU_GET(cpuid); 2727 tdq = TDQ_CPU(cpuid); 2728 if (TD_IS_IDLETHREAD(td)) 2729 td->td_lock = TDQ_LOCKPTR(tdq); 2730 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2731 td->td_oncpu = cpuid; 2732 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED); 2733 lock_profile_obtain_lock_success( 2734 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__); 2735} 2736 2737/* 2738 * Create on first use to catch odd startup conditons. 2739 */ 2740char * 2741sched_tdname(struct thread *td) 2742{ 2743#ifdef KTR 2744 struct td_sched *ts; 2745 2746 ts = td->td_sched; 2747 if (ts->ts_name[0] == '\0') 2748 snprintf(ts->ts_name, sizeof(ts->ts_name), 2749 "%s tid %d", td->td_name, td->td_tid); 2750 return (ts->ts_name); 2751#else 2752 return (td->td_name); 2753#endif 2754} 2755 2756#ifdef KTR 2757void 2758sched_clear_tdname(struct thread *td) 2759{ 2760 struct td_sched *ts; 2761 2762 ts = td->td_sched; 2763 ts->ts_name[0] = '\0'; 2764} 2765#endif 2766 2767#ifdef SMP 2768 2769/* 2770 * Build the CPU topology dump string. Is recursively called to collect 2771 * the topology tree. 2772 */ 2773static int 2774sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg, 2775 int indent) 2776{ 2777 char cpusetbuf[CPUSETBUFSIZ]; 2778 int i, first; 2779 2780 sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent, 2781 "", 1 + indent / 2, cg->cg_level); 2782 sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"%s\">", indent, "", 2783 cg->cg_count, cpusetobj_strprint(cpusetbuf, &cg->cg_mask)); 2784 first = TRUE; 2785 for (i = 0; i < MAXCPU; i++) { 2786 if (CPU_ISSET(i, &cg->cg_mask)) { 2787 if (!first) 2788 sbuf_printf(sb, ", "); 2789 else 2790 first = FALSE; 2791 sbuf_printf(sb, "%d", i); 2792 } 2793 } 2794 sbuf_printf(sb, "</cpu>\n"); 2795 2796 if (cg->cg_flags != 0) { 2797 sbuf_printf(sb, "%*s <flags>", indent, ""); 2798 if ((cg->cg_flags & CG_FLAG_HTT) != 0) 2799 sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>"); 2800 if ((cg->cg_flags & CG_FLAG_THREAD) != 0) 2801 sbuf_printf(sb, "<flag name=\"THREAD\">THREAD group</flag>"); 2802 if ((cg->cg_flags & CG_FLAG_SMT) != 0) 2803 sbuf_printf(sb, "<flag name=\"SMT\">SMT group</flag>"); 2804 sbuf_printf(sb, "</flags>\n"); 2805 } 2806 2807 if (cg->cg_children > 0) { 2808 sbuf_printf(sb, "%*s <children>\n", indent, ""); 2809 for (i = 0; i < cg->cg_children; i++) 2810 sysctl_kern_sched_topology_spec_internal(sb, 2811 &cg->cg_child[i], indent+2); 2812 sbuf_printf(sb, "%*s </children>\n", indent, ""); 2813 } 2814 sbuf_printf(sb, "%*s</group>\n", indent, ""); 2815 return (0); 2816} 2817 2818/* 2819 * Sysctl handler for retrieving topology dump. It's a wrapper for 2820 * the recursive sysctl_kern_smp_topology_spec_internal(). 2821 */ 2822static int 2823sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS) 2824{ 2825 struct sbuf *topo; 2826 int err; 2827 2828 KASSERT(cpu_top != NULL, ("cpu_top isn't initialized")); 2829 2830 topo = sbuf_new(NULL, NULL, 500, SBUF_AUTOEXTEND); 2831 if (topo == NULL) 2832 return (ENOMEM); 2833 2834 sbuf_printf(topo, "<groups>\n"); 2835 err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1); 2836 sbuf_printf(topo, "</groups>\n"); 2837 2838 if (err == 0) { 2839 sbuf_finish(topo); 2840 err = SYSCTL_OUT(req, sbuf_data(topo), sbuf_len(topo)); 2841 } 2842 sbuf_delete(topo); 2843 return (err); 2844} 2845 2846#endif 2847 2848static int 2849sysctl_kern_quantum(SYSCTL_HANDLER_ARGS) 2850{ 2851 int error, new_val, period; 2852 2853 period = 1000000 / realstathz; 2854 new_val = period * sched_slice; 2855 error = sysctl_handle_int(oidp, &new_val, 0, req); 2856 if (error != 0 || req->newptr == NULL) 2857 return (error); 2858 if (new_val <= 0) 2859 return (EINVAL); 2860 sched_slice = imax(1, (new_val + period / 2) / period); 2861 sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR; 2862 hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) / 2863 realstathz); 2864 return (0); 2865} 2866 2867SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler"); 2868SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0, 2869 "Scheduler name"); 2870SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW, 2871 NULL, 0, sysctl_kern_quantum, "I", 2872 "Quantum for timeshare threads in microseconds"); 2873SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0, 2874 "Quantum for timeshare threads in stathz ticks"); 2875SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0, 2876 "Interactivity score threshold"); 2877SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW, 2878 &preempt_thresh, 0, 2879 "Maximal (lowest) priority for preemption"); 2880SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost, 0, 2881 "Assign static kernel priorities to sleeping threads"); 2882SYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins, 0, 2883 "Number of times idle thread will spin waiting for new work"); 2884SYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW, 2885 &sched_idlespinthresh, 0, 2886 "Threshold before we will permit idle thread spinning"); 2887#ifdef SMP 2888SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0, 2889 "Number of hz ticks to keep thread affinity for"); 2890SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0, 2891 "Enables the long-term load balancer"); 2892SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW, 2893 &balance_interval, 0, 2894 "Average period in stathz ticks to run the long-term balancer"); 2895SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0, 2896 "Attempts to steal work from other cores before idling"); 2897SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0, 2898 "Minimum load on remote CPU before we'll steal"); 2899SYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING | 2900 CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A", 2901 "XML dump of detected CPU topology"); 2902#endif 2903 2904/* ps compat. All cpu percentages from ULE are weighted. */ 2905static int ccpu = 0; 2906SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 2907