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 #include <sys/cdefs.h> 28 __FBSDID("$FreeBSD$"); 29 30 #include "opt_hwpmc_hooks.h" 31 #include "opt_sched.h" 32 33 #include <sys/param.h> 34 #include <sys/systm.h> 35 #include <sys/kdb.h> 36 #include <sys/kernel.h> 37 #include <sys/ktr.h> 38 #include <sys/lock.h> 39 #include <sys/mutex.h> 40 #include <sys/proc.h> 41 #include <sys/resource.h> 42 #include <sys/resourcevar.h> 43 #include <sys/sched.h> 44 #include <sys/smp.h> 45 #include <sys/sx.h> 46 #include <sys/sysctl.h> 47 #include <sys/sysproto.h> 48 #include <sys/turnstile.h> 49 #include <sys/umtx.h> 50 #include <sys/vmmeter.h> 51 #ifdef KTRACE 52 #include <sys/uio.h> 53 #include <sys/ktrace.h> 54 #endif 55 56 #ifdef HWPMC_HOOKS 57 #include <sys/pmckern.h> 58 #endif 59 60 #include <machine/cpu.h> 61 #include <machine/smp.h> 62 63 #ifndef PREEMPTION 64 #error "SCHED_ULE requires options PREEMPTION" 65 #endif 66 67 /* 68 * TODO: 69 * Pick idle from affinity group or self group first. 70 * Implement pick_score. 71 */ 72 73 #define KTR_ULE 0x0 /* Enable for pickpri debugging. */ 74 75 /* 76 * Thread scheduler specific section. 77 */ 78 struct td_sched { 79 TAILQ_ENTRY(td_sched) ts_procq; /* (j/z) Run queue. */ 80 int ts_flags; /* (j) TSF_* flags. */ 81 struct thread *ts_thread; /* (*) Active associated thread. */ 82 u_char ts_rqindex; /* (j) Run queue index. */ 83 int ts_slptime; 84 int ts_slice; 85 struct runq *ts_runq; 86 u_char ts_cpu; /* CPU that we have affinity for. */ 87 /* The following variables are only used for pctcpu calculation */ 88 int ts_ltick; /* Last tick that we were running on */ 89 int ts_ftick; /* First tick that we were running on */ 90 int ts_ticks; /* Tick count */ 91 #ifdef SMP 92 int ts_rltick; /* Real last tick, for affinity. */ 93 #endif 94 95 /* originally from kg_sched */ 96 u_int skg_slptime; /* Number of ticks we vol. slept */ 97 u_int skg_runtime; /* Number of ticks we were running */ 98 }; 99 /* flags kept in ts_flags */ 100 #define TSF_BOUND 0x0001 /* Thread can not migrate. */ 101 #define TSF_XFERABLE 0x0002 /* Thread was added as transferable. */ 102 103 static struct td_sched td_sched0; 104 105 /* 106 * Cpu percentage computation macros and defines. 107 * 108 * SCHED_TICK_SECS: Number of seconds to average the cpu usage across. 109 * SCHED_TICK_TARG: Number of hz ticks to average the cpu usage across. 110 * SCHED_TICK_MAX: Maximum number of ticks before scaling back. 111 * SCHED_TICK_SHIFT: Shift factor to avoid rounding away results. 112 * SCHED_TICK_HZ: Compute the number of hz ticks for a given ticks count. 113 * SCHED_TICK_TOTAL: Gives the amount of time we've been recording ticks. 114 */ 115 #define SCHED_TICK_SECS 10 116 #define SCHED_TICK_TARG (hz * SCHED_TICK_SECS) 117 #define SCHED_TICK_MAX (SCHED_TICK_TARG + hz) 118 #define SCHED_TICK_SHIFT 10 119 #define SCHED_TICK_HZ(ts) ((ts)->ts_ticks >> SCHED_TICK_SHIFT) 120 #define SCHED_TICK_TOTAL(ts) (max((ts)->ts_ltick - (ts)->ts_ftick, hz)) 121 122 /* 123 * These macros determine priorities for non-interactive threads. They are 124 * assigned a priority based on their recent cpu utilization as expressed 125 * by the ratio of ticks to the tick total. NHALF priorities at the start 126 * and end of the MIN to MAX timeshare range are only reachable with negative 127 * or positive nice respectively. 128 * 129 * PRI_RANGE: Priority range for utilization dependent priorities. 130 * PRI_NRESV: Number of nice values. 131 * PRI_TICKS: Compute a priority in PRI_RANGE from the ticks count and total. 132 * PRI_NICE: Determines the part of the priority inherited from nice. 133 */ 134 #define SCHED_PRI_NRESV (PRIO_MAX - PRIO_MIN) 135 #define SCHED_PRI_NHALF (SCHED_PRI_NRESV / 2) 136 #define SCHED_PRI_MIN (PRI_MIN_TIMESHARE + SCHED_PRI_NHALF) 137 #define SCHED_PRI_MAX (PRI_MAX_TIMESHARE - SCHED_PRI_NHALF) 138 #define SCHED_PRI_RANGE (SCHED_PRI_MAX - SCHED_PRI_MIN) 139 #define SCHED_PRI_TICKS(ts) \ 140 (SCHED_TICK_HZ((ts)) / \ 141 (roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE)) 142 #define SCHED_PRI_NICE(nice) (nice) 143 144 /* 145 * These determine the interactivity of a process. Interactivity differs from 146 * cpu utilization in that it expresses the voluntary time slept vs time ran 147 * while cpu utilization includes all time not running. This more accurately 148 * models the intent of the thread. 149 * 150 * SLP_RUN_MAX: Maximum amount of sleep time + run time we'll accumulate 151 * before throttling back. 152 * SLP_RUN_FORK: Maximum slp+run time to inherit at fork time. 153 * INTERACT_MAX: Maximum interactivity value. Smaller is better. 154 * INTERACT_THRESH: Threshhold for placement on the current runq. 155 */ 156 #define SCHED_SLP_RUN_MAX ((hz * 5) << SCHED_TICK_SHIFT) 157 #define SCHED_SLP_RUN_FORK ((hz / 2) << SCHED_TICK_SHIFT) 158 #define SCHED_INTERACT_MAX (100) 159 #define SCHED_INTERACT_HALF (SCHED_INTERACT_MAX / 2) 160 #define SCHED_INTERACT_THRESH (30) 161 162 /* 163 * tickincr: Converts a stathz tick into a hz domain scaled by 164 * the shift factor. Without the shift the error rate 165 * due to rounding would be unacceptably high. 166 * realstathz: stathz is sometimes 0 and run off of hz. 167 * sched_slice: Runtime of each thread before rescheduling. 168 */ 169 static int sched_interact = SCHED_INTERACT_THRESH; 170 static int realstathz; 171 static int tickincr; 172 static int sched_slice; 173 174 /* 175 * tdq - per processor runqs and statistics. 176 */ 177 struct tdq { 178 struct runq tdq_idle; /* Queue of IDLE threads. */ 179 struct runq tdq_timeshare; /* timeshare run queue. */ 180 struct runq tdq_realtime; /* real-time run queue. */ 181 u_char tdq_idx; /* Current insert index. */ 182 u_char tdq_ridx; /* Current removal index. */ 183 short tdq_flags; /* Thread queue flags */ 184 int tdq_load; /* Aggregate load. */ 185 #ifdef SMP 186 int tdq_transferable; 187 LIST_ENTRY(tdq) tdq_siblings; /* Next in tdq group. */ 188 struct tdq_group *tdq_group; /* Our processor group. */ 189 #else 190 int tdq_sysload; /* For loadavg, !ITHD load. */ 191 #endif 192 }; 193 194 #define TDQF_BUSY 0x0001 /* Queue is marked as busy */ 195 196 #ifdef SMP 197 /* 198 * tdq groups are groups of processors which can cheaply share threads. When 199 * one processor in the group goes idle it will check the runqs of the other 200 * processors in its group prior to halting and waiting for an interrupt. 201 * These groups are suitable for SMT (Symetric Multi-Threading) and not NUMA. 202 * In a numa environment we'd want an idle bitmap per group and a two tiered 203 * load balancer. 204 */ 205 struct tdq_group { 206 int tdg_cpus; /* Count of CPUs in this tdq group. */ 207 cpumask_t tdg_cpumask; /* Mask of cpus in this group. */ 208 cpumask_t tdg_idlemask; /* Idle cpus in this group. */ 209 cpumask_t tdg_mask; /* Bit mask for first cpu. */ 210 int tdg_load; /* Total load of this group. */ 211 int tdg_transferable; /* Transferable load of this group. */ 212 LIST_HEAD(, tdq) tdg_members; /* Linked list of all members. */ 213 }; 214 215 #define SCHED_AFFINITY_DEFAULT (hz / 100) 216 #define SCHED_AFFINITY(ts) ((ts)->ts_rltick > ticks - affinity) 217 218 /* 219 * Run-time tunables. 220 */ 221 static int rebalance = 0; 222 static int pick_pri = 0; 223 static int affinity; 224 static int tryself = 1; 225 static int tryselfidle = 1; 226 static int ipi_ast = 0; 227 static int ipi_preempt = 1; 228 static int ipi_thresh = PRI_MIN_KERN; 229 static int steal_htt = 1; 230 static int steal_busy = 1; 231 static int busy_thresh = 4; 232 static int topology = 0; 233 234 /* 235 * One thread queue per processor. 236 */ 237 static volatile cpumask_t tdq_idle; 238 static volatile cpumask_t tdq_busy; 239 static int tdg_maxid; 240 static struct tdq tdq_cpu[MAXCPU]; 241 static struct tdq_group tdq_groups[MAXCPU]; 242 static int bal_tick; 243 static int gbal_tick; 244 static int balance_groups; 245 246 #define TDQ_SELF() (&tdq_cpu[PCPU_GET(cpuid)]) 247 #define TDQ_CPU(x) (&tdq_cpu[(x)]) 248 #define TDQ_ID(x) ((x) - tdq_cpu) 249 #define TDQ_GROUP(x) (&tdq_groups[(x)]) 250 #else /* !SMP */ 251 static struct tdq tdq_cpu; 252 253 #define TDQ_ID(x) (0) 254 #define TDQ_SELF() (&tdq_cpu) 255 #define TDQ_CPU(x) (&tdq_cpu) 256 #endif 257 258 static void sched_priority(struct thread *); 259 static void sched_thread_priority(struct thread *, u_char); 260 static int sched_interact_score(struct thread *); 261 static void sched_interact_update(struct thread *); 262 static void sched_interact_fork(struct thread *); 263 static void sched_pctcpu_update(struct td_sched *); 264 static inline void sched_pin_td(struct thread *td); 265 static inline void sched_unpin_td(struct thread *td); 266 267 /* Operations on per processor queues */ 268 static struct td_sched * tdq_choose(struct tdq *); 269 static void tdq_setup(struct tdq *); 270 static void tdq_load_add(struct tdq *, struct td_sched *); 271 static void tdq_load_rem(struct tdq *, struct td_sched *); 272 static __inline void tdq_runq_add(struct tdq *, struct td_sched *, int); 273 static __inline void tdq_runq_rem(struct tdq *, struct td_sched *); 274 void tdq_print(int cpu); 275 static void runq_print(struct runq *rq); 276 #ifdef SMP 277 static int tdq_pickidle(struct tdq *, struct td_sched *); 278 static int tdq_pickpri(struct tdq *, struct td_sched *, int); 279 static struct td_sched *runq_steal(struct runq *); 280 static void sched_balance(void); 281 static void sched_balance_groups(void); 282 static void sched_balance_group(struct tdq_group *); 283 static void sched_balance_pair(struct tdq *, struct tdq *); 284 static void sched_smp_tick(struct thread *); 285 static void tdq_move(struct tdq *, int); 286 static int tdq_idled(struct tdq *); 287 static void tdq_notify(struct td_sched *); 288 static struct td_sched *tdq_steal(struct tdq *, int); 289 290 #define THREAD_CAN_MIGRATE(td) ((td)->td_pinned == 0) 291 #endif 292 293 static void sched_setup(void *dummy); 294 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL) 295 296 static void sched_initticks(void *dummy); 297 SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks, NULL) 298 299 static inline void 300 sched_pin_td(struct thread *td) 301 { 302 td->td_pinned++; 303 } 304 305 static inline void 306 sched_unpin_td(struct thread *td) 307 { 308 td->td_pinned--; 309 } 310 311 static void 312 runq_print(struct runq *rq) 313 { 314 struct rqhead *rqh; 315 struct td_sched *ts; 316 int pri; 317 int j; 318 int i; 319 320 for (i = 0; i < RQB_LEN; i++) { 321 printf("\t\trunq bits %d 0x%zx\n", 322 i, rq->rq_status.rqb_bits[i]); 323 for (j = 0; j < RQB_BPW; j++) 324 if (rq->rq_status.rqb_bits[i] & (1ul << j)) { 325 pri = j + (i << RQB_L2BPW); 326 rqh = &rq->rq_queues[pri]; 327 TAILQ_FOREACH(ts, rqh, ts_procq) { 328 printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n", 329 ts->ts_thread, ts->ts_thread->td_proc->p_comm, ts->ts_thread->td_priority, ts->ts_rqindex, pri); 330 } 331 } 332 } 333 } 334 335 void 336 tdq_print(int cpu) 337 { 338 struct tdq *tdq; 339 340 tdq = TDQ_CPU(cpu); 341 342 printf("tdq:\n"); 343 printf("\tload: %d\n", tdq->tdq_load); 344 printf("\ttimeshare idx: %d\n", tdq->tdq_idx); 345 printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx); 346 printf("\trealtime runq:\n"); 347 runq_print(&tdq->tdq_realtime); 348 printf("\ttimeshare runq:\n"); 349 runq_print(&tdq->tdq_timeshare); 350 printf("\tidle runq:\n"); 351 runq_print(&tdq->tdq_idle); 352 #ifdef SMP 353 printf("\tload transferable: %d\n", tdq->tdq_transferable); 354 #endif 355 } 356 357 static __inline void 358 tdq_runq_add(struct tdq *tdq, struct td_sched *ts, int flags) 359 { 360 #ifdef SMP 361 if (THREAD_CAN_MIGRATE(ts->ts_thread)) { 362 tdq->tdq_transferable++; 363 tdq->tdq_group->tdg_transferable++; 364 ts->ts_flags |= TSF_XFERABLE; 365 if (tdq->tdq_transferable >= busy_thresh && 366 (tdq->tdq_flags & TDQF_BUSY) == 0) { 367 tdq->tdq_flags |= TDQF_BUSY; 368 atomic_set_int(&tdq_busy, 1 << TDQ_ID(tdq)); 369 } 370 } 371 #endif 372 if (ts->ts_runq == &tdq->tdq_timeshare) { 373 u_char pri; 374 375 pri = ts->ts_thread->td_priority; 376 KASSERT(pri <= PRI_MAX_TIMESHARE && pri >= PRI_MIN_TIMESHARE, 377 ("Invalid priority %d on timeshare runq", pri)); 378 /* 379 * This queue contains only priorities between MIN and MAX 380 * realtime. Use the whole queue to represent these values. 381 */ 382 #define TS_RQ_PPQ (((PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE) + 1) / RQ_NQS) 383 if ((flags & SRQ_BORROWING) == 0) { 384 pri = (pri - PRI_MIN_TIMESHARE) / TS_RQ_PPQ; 385 pri = (pri + tdq->tdq_idx) % RQ_NQS; 386 /* 387 * This effectively shortens the queue by one so we 388 * can have a one slot difference between idx and 389 * ridx while we wait for threads to drain. 390 */ 391 if (tdq->tdq_ridx != tdq->tdq_idx && 392 pri == tdq->tdq_ridx) 393 pri = (unsigned char)(pri - 1) % RQ_NQS; 394 } else 395 pri = tdq->tdq_ridx; 396 runq_add_pri(ts->ts_runq, ts, pri, flags); 397 } else 398 runq_add(ts->ts_runq, ts, flags); 399 } 400 401 static __inline void 402 tdq_runq_rem(struct tdq *tdq, struct td_sched *ts) 403 { 404 #ifdef SMP 405 if (ts->ts_flags & TSF_XFERABLE) { 406 tdq->tdq_transferable--; 407 tdq->tdq_group->tdg_transferable--; 408 ts->ts_flags &= ~TSF_XFERABLE; 409 if (tdq->tdq_transferable < busy_thresh && 410 (tdq->tdq_flags & TDQF_BUSY)) { 411 atomic_clear_int(&tdq_busy, 1 << TDQ_ID(tdq)); 412 tdq->tdq_flags &= ~TDQF_BUSY; 413 } 414 } 415 #endif 416 if (ts->ts_runq == &tdq->tdq_timeshare) { 417 if (tdq->tdq_idx != tdq->tdq_ridx) 418 runq_remove_idx(ts->ts_runq, ts, &tdq->tdq_ridx); 419 else 420 runq_remove_idx(ts->ts_runq, ts, NULL); 421 /* 422 * For timeshare threads we update the priority here so 423 * the priority reflects the time we've been sleeping. 424 */ 425 ts->ts_ltick = ticks; 426 sched_pctcpu_update(ts); 427 sched_priority(ts->ts_thread); 428 } else 429 runq_remove(ts->ts_runq, ts); 430 } 431 432 static void 433 tdq_load_add(struct tdq *tdq, struct td_sched *ts) 434 { 435 int class; 436 mtx_assert(&sched_lock, MA_OWNED); 437 class = PRI_BASE(ts->ts_thread->td_pri_class); 438 tdq->tdq_load++; 439 CTR2(KTR_SCHED, "cpu %jd load: %d", TDQ_ID(tdq), tdq->tdq_load); 440 if (class != PRI_ITHD && 441 (ts->ts_thread->td_proc->p_flag & P_NOLOAD) == 0) 442 #ifdef SMP 443 tdq->tdq_group->tdg_load++; 444 #else 445 tdq->tdq_sysload++; 446 #endif 447 } 448 449 static void 450 tdq_load_rem(struct tdq *tdq, struct td_sched *ts) 451 { 452 int class; 453 mtx_assert(&sched_lock, MA_OWNED); 454 class = PRI_BASE(ts->ts_thread->td_pri_class); 455 if (class != PRI_ITHD && 456 (ts->ts_thread->td_proc->p_flag & P_NOLOAD) == 0) 457 #ifdef SMP 458 tdq->tdq_group->tdg_load--; 459 #else 460 tdq->tdq_sysload--; 461 #endif 462 tdq->tdq_load--; 463 CTR1(KTR_SCHED, "load: %d", tdq->tdq_load); 464 ts->ts_runq = NULL; 465 } 466 467 #ifdef SMP 468 static void 469 sched_smp_tick(struct thread *td) 470 { 471 struct tdq *tdq; 472 473 tdq = TDQ_SELF(); 474 if (rebalance) { 475 if (ticks >= bal_tick) 476 sched_balance(); 477 if (ticks >= gbal_tick && balance_groups) 478 sched_balance_groups(); 479 } 480 td->td_sched->ts_rltick = ticks; 481 } 482 483 /* 484 * sched_balance is a simple CPU load balancing algorithm. It operates by 485 * finding the least loaded and most loaded cpu and equalizing their load 486 * by migrating some processes. 487 * 488 * Dealing only with two CPUs at a time has two advantages. Firstly, most 489 * installations will only have 2 cpus. Secondly, load balancing too much at 490 * once can have an unpleasant effect on the system. The scheduler rarely has 491 * enough information to make perfect decisions. So this algorithm chooses 492 * algorithm simplicity and more gradual effects on load in larger systems. 493 * 494 * It could be improved by considering the priorities and slices assigned to 495 * each task prior to balancing them. There are many pathological cases with 496 * any approach and so the semi random algorithm below may work as well as any. 497 * 498 */ 499 static void 500 sched_balance(void) 501 { 502 struct tdq_group *high; 503 struct tdq_group *low; 504 struct tdq_group *tdg; 505 int cnt; 506 int i; 507 508 bal_tick = ticks + (random() % (hz * 2)); 509 if (smp_started == 0) 510 return; 511 low = high = NULL; 512 i = random() % (tdg_maxid + 1); 513 for (cnt = 0; cnt <= tdg_maxid; cnt++) { 514 tdg = TDQ_GROUP(i); 515 /* 516 * Find the CPU with the highest load that has some 517 * threads to transfer. 518 */ 519 if ((high == NULL || tdg->tdg_load > high->tdg_load) 520 && tdg->tdg_transferable) 521 high = tdg; 522 if (low == NULL || tdg->tdg_load < low->tdg_load) 523 low = tdg; 524 if (++i > tdg_maxid) 525 i = 0; 526 } 527 if (low != NULL && high != NULL && high != low) 528 sched_balance_pair(LIST_FIRST(&high->tdg_members), 529 LIST_FIRST(&low->tdg_members)); 530 } 531 532 static void 533 sched_balance_groups(void) 534 { 535 int i; 536 537 gbal_tick = ticks + (random() % (hz * 2)); 538 mtx_assert(&sched_lock, MA_OWNED); 539 if (smp_started) 540 for (i = 0; i <= tdg_maxid; i++) 541 sched_balance_group(TDQ_GROUP(i)); 542 } 543 544 static void 545 sched_balance_group(struct tdq_group *tdg) 546 { 547 struct tdq *tdq; 548 struct tdq *high; 549 struct tdq *low; 550 int load; 551 552 if (tdg->tdg_transferable == 0) 553 return; 554 low = NULL; 555 high = NULL; 556 LIST_FOREACH(tdq, &tdg->tdg_members, tdq_siblings) { 557 load = tdq->tdq_load; 558 if (high == NULL || load > high->tdq_load) 559 high = tdq; 560 if (low == NULL || load < low->tdq_load) 561 low = tdq; 562 } 563 if (high != NULL && low != NULL && high != low) 564 sched_balance_pair(high, low); 565 } 566 567 static void 568 sched_balance_pair(struct tdq *high, struct tdq *low) 569 { 570 int transferable; 571 int high_load; 572 int low_load; 573 int move; 574 int diff; 575 int i; 576 577 /* 578 * If we're transfering within a group we have to use this specific 579 * tdq's transferable count, otherwise we can steal from other members 580 * of the group. 581 */ 582 if (high->tdq_group == low->tdq_group) { 583 transferable = high->tdq_transferable; 584 high_load = high->tdq_load; 585 low_load = low->tdq_load; 586 } else { 587 transferable = high->tdq_group->tdg_transferable; 588 high_load = high->tdq_group->tdg_load; 589 low_load = low->tdq_group->tdg_load; 590 } 591 if (transferable == 0) 592 return; 593 /* 594 * Determine what the imbalance is and then adjust that to how many 595 * threads we actually have to give up (transferable). 596 */ 597 diff = high_load - low_load; 598 move = diff / 2; 599 if (diff & 0x1) 600 move++; 601 move = min(move, transferable); 602 for (i = 0; i < move; i++) 603 tdq_move(high, TDQ_ID(low)); 604 return; 605 } 606 607 static void 608 tdq_move(struct tdq *from, int cpu) 609 { 610 struct tdq *tdq; 611 struct tdq *to; 612 struct td_sched *ts; 613 614 tdq = from; 615 to = TDQ_CPU(cpu); 616 ts = tdq_steal(tdq, 1); 617 if (ts == NULL) { 618 struct tdq_group *tdg; 619 620 tdg = tdq->tdq_group; 621 LIST_FOREACH(tdq, &tdg->tdg_members, tdq_siblings) { 622 if (tdq == from || tdq->tdq_transferable == 0) 623 continue; 624 ts = tdq_steal(tdq, 1); 625 break; 626 } 627 if (ts == NULL) 628 panic("tdq_move: No threads available with a " 629 "transferable count of %d\n", 630 tdg->tdg_transferable); 631 } 632 if (tdq == to) 633 return; 634 sched_rem(ts->ts_thread); 635 ts->ts_cpu = cpu; 636 sched_pin_td(ts->ts_thread); 637 sched_add(ts->ts_thread, SRQ_YIELDING); 638 sched_unpin_td(ts->ts_thread); 639 } 640 641 static int 642 tdq_idled(struct tdq *tdq) 643 { 644 struct tdq_group *tdg; 645 struct tdq *steal; 646 struct td_sched *ts; 647 648 tdg = tdq->tdq_group; 649 /* 650 * If we're in a cpu group, try and steal threads from another cpu in 651 * the group before idling. 652 */ 653 if (steal_htt && tdg->tdg_cpus > 1 && tdg->tdg_transferable) { 654 LIST_FOREACH(steal, &tdg->tdg_members, tdq_siblings) { 655 if (steal == tdq || steal->tdq_transferable == 0) 656 continue; 657 ts = tdq_steal(steal, 0); 658 if (ts) 659 goto steal; 660 } 661 } 662 if (steal_busy) { 663 while (tdq_busy) { 664 int cpu; 665 666 cpu = ffs(tdq_busy); 667 if (cpu == 0) 668 break; 669 cpu--; 670 steal = TDQ_CPU(cpu); 671 if (steal->tdq_transferable == 0) 672 continue; 673 ts = tdq_steal(steal, 1); 674 if (ts == NULL) 675 continue; 676 CTR5(KTR_ULE, 677 "tdq_idled: stealing td %p(%s) pri %d from %d busy 0x%X", 678 ts->ts_thread, ts->ts_thread->td_proc->p_comm, 679 ts->ts_thread->td_priority, cpu, tdq_busy); 680 goto steal; 681 } 682 } 683 /* 684 * We only set the idled bit when all of the cpus in the group are 685 * idle. Otherwise we could get into a situation where a thread bounces 686 * back and forth between two idle cores on seperate physical CPUs. 687 */ 688 tdg->tdg_idlemask |= PCPU_GET(cpumask); 689 if (tdg->tdg_idlemask == tdg->tdg_cpumask) 690 atomic_set_int(&tdq_idle, tdg->tdg_mask); 691 return (1); 692 steal: 693 sched_rem(ts->ts_thread); 694 ts->ts_cpu = PCPU_GET(cpuid); 695 sched_pin_td(ts->ts_thread); 696 sched_add(ts->ts_thread, SRQ_YIELDING); 697 sched_unpin_td(ts->ts_thread); 698 699 return (0); 700 } 701 702 static void 703 tdq_notify(struct td_sched *ts) 704 { 705 struct thread *ctd; 706 struct pcpu *pcpu; 707 int cpri; 708 int pri; 709 int cpu; 710 711 cpu = ts->ts_cpu; 712 pri = ts->ts_thread->td_priority; 713 pcpu = pcpu_find(cpu); 714 ctd = pcpu->pc_curthread; 715 cpri = ctd->td_priority; 716 717 /* 718 * If our priority is not better than the current priority there is 719 * nothing to do. 720 */ 721 if (pri > cpri) 722 return; 723 /* 724 * Always IPI idle. 725 */ 726 if (cpri > PRI_MIN_IDLE) 727 goto sendipi; 728 /* 729 * If we're realtime or better and there is timeshare or worse running 730 * send an IPI. 731 */ 732 if (pri < PRI_MAX_REALTIME && cpri > PRI_MAX_REALTIME) 733 goto sendipi; 734 /* 735 * Otherwise only IPI if we exceed the threshold. 736 */ 737 if (pri > ipi_thresh) 738 return; 739 sendipi: 740 ctd->td_flags |= TDF_NEEDRESCHED; 741 if (cpri < PRI_MIN_IDLE) { 742 if (ipi_ast) 743 ipi_selected(1 << cpu, IPI_AST); 744 else if (ipi_preempt) 745 ipi_selected(1 << cpu, IPI_PREEMPT); 746 } else 747 ipi_selected(1 << cpu, IPI_PREEMPT); 748 } 749 750 static struct td_sched * 751 runq_steal(struct runq *rq) 752 { 753 struct rqhead *rqh; 754 struct rqbits *rqb; 755 struct td_sched *ts; 756 int word; 757 int bit; 758 759 mtx_assert(&sched_lock, MA_OWNED); 760 rqb = &rq->rq_status; 761 for (word = 0; word < RQB_LEN; word++) { 762 if (rqb->rqb_bits[word] == 0) 763 continue; 764 for (bit = 0; bit < RQB_BPW; bit++) { 765 if ((rqb->rqb_bits[word] & (1ul << bit)) == 0) 766 continue; 767 rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)]; 768 TAILQ_FOREACH(ts, rqh, ts_procq) { 769 if (THREAD_CAN_MIGRATE(ts->ts_thread)) 770 return (ts); 771 } 772 } 773 } 774 return (NULL); 775 } 776 777 static struct td_sched * 778 tdq_steal(struct tdq *tdq, int stealidle) 779 { 780 struct td_sched *ts; 781 782 /* 783 * Steal from next first to try to get a non-interactive task that 784 * may not have run for a while. 785 * XXX Need to effect steal order for timeshare threads. 786 */ 787 if ((ts = runq_steal(&tdq->tdq_realtime)) != NULL) 788 return (ts); 789 if ((ts = runq_steal(&tdq->tdq_timeshare)) != NULL) 790 return (ts); 791 if (stealidle) 792 return (runq_steal(&tdq->tdq_idle)); 793 return (NULL); 794 } 795 796 int 797 tdq_pickidle(struct tdq *tdq, struct td_sched *ts) 798 { 799 struct tdq_group *tdg; 800 int self; 801 int cpu; 802 803 self = PCPU_GET(cpuid); 804 if (smp_started == 0) 805 return (self); 806 /* 807 * If the current CPU has idled, just run it here. 808 */ 809 if ((tdq->tdq_group->tdg_idlemask & PCPU_GET(cpumask)) != 0) 810 return (self); 811 /* 812 * Try the last group we ran on. 813 */ 814 tdg = TDQ_CPU(ts->ts_cpu)->tdq_group; 815 cpu = ffs(tdg->tdg_idlemask); 816 if (cpu) 817 return (cpu - 1); 818 /* 819 * Search for an idle group. 820 */ 821 cpu = ffs(tdq_idle); 822 if (cpu) 823 return (cpu - 1); 824 /* 825 * XXX If there are no idle groups, check for an idle core. 826 */ 827 /* 828 * No idle CPUs? 829 */ 830 return (self); 831 } 832 833 static int 834 tdq_pickpri(struct tdq *tdq, struct td_sched *ts, int flags) 835 { 836 struct pcpu *pcpu; 837 int lowpri; 838 int lowcpu; 839 int lowload; 840 int load; 841 int self; 842 int pri; 843 int cpu; 844 845 self = PCPU_GET(cpuid); 846 if (smp_started == 0) 847 return (self); 848 849 pri = ts->ts_thread->td_priority; 850 /* 851 * Regardless of affinity, if the last cpu is idle send it there. 852 */ 853 pcpu = pcpu_find(ts->ts_cpu); 854 if (pcpu->pc_curthread->td_priority > PRI_MIN_IDLE) { 855 CTR5(KTR_ULE, 856 "ts_cpu %d idle, ltick %d ticks %d pri %d curthread %d", 857 ts->ts_cpu, ts->ts_rltick, ticks, pri, 858 pcpu->pc_curthread->td_priority); 859 return (ts->ts_cpu); 860 } 861 /* 862 * If we have affinity, try to place it on the cpu we last ran on. 863 */ 864 if (SCHED_AFFINITY(ts) && pcpu->pc_curthread->td_priority > pri) { 865 CTR5(KTR_ULE, 866 "affinity for %d, ltick %d ticks %d pri %d curthread %d", 867 ts->ts_cpu, ts->ts_rltick, ticks, pri, 868 pcpu->pc_curthread->td_priority); 869 return (ts->ts_cpu); 870 } 871 /* 872 * Try ourself first; If we're running something lower priority this 873 * may have some locality with the waking thread and execute faster 874 * here. 875 */ 876 if (tryself) { 877 /* 878 * If we're being awoken by an interrupt thread or the waker 879 * is going right to sleep run here as well. 880 */ 881 if ((TDQ_SELF()->tdq_load == 1) && (flags & SRQ_YIELDING || 882 curthread->td_pri_class == PRI_ITHD)) { 883 CTR2(KTR_ULE, "tryself load %d flags %d", 884 TDQ_SELF()->tdq_load, flags); 885 return (self); 886 } 887 } 888 /* 889 * Look for an idle group. 890 */ 891 CTR1(KTR_ULE, "tdq_idle %X", tdq_idle); 892 cpu = ffs(tdq_idle); 893 if (cpu) 894 return (cpu - 1); 895 if (tryselfidle && pri < curthread->td_priority) { 896 CTR1(KTR_ULE, "tryself %d", 897 curthread->td_priority); 898 return (self); 899 } 900 /* 901 * Now search for the cpu running the lowest priority thread with 902 * the least load. 903 */ 904 lowload = 0; 905 lowpri = lowcpu = 0; 906 for (cpu = 0; cpu <= mp_maxid; cpu++) { 907 if (CPU_ABSENT(cpu)) 908 continue; 909 pcpu = pcpu_find(cpu); 910 pri = pcpu->pc_curthread->td_priority; 911 CTR4(KTR_ULE, 912 "cpu %d pri %d lowcpu %d lowpri %d", 913 cpu, pri, lowcpu, lowpri); 914 if (pri < lowpri) 915 continue; 916 load = TDQ_CPU(cpu)->tdq_load; 917 if (lowpri && lowpri == pri && load > lowload) 918 continue; 919 lowpri = pri; 920 lowcpu = cpu; 921 lowload = load; 922 } 923 924 return (lowcpu); 925 } 926 927 #endif /* SMP */ 928 929 /* 930 * Pick the highest priority task we have and return it. 931 */ 932 933 static struct td_sched * 934 tdq_choose(struct tdq *tdq) 935 { 936 struct td_sched *ts; 937 938 mtx_assert(&sched_lock, MA_OWNED); 939 940 ts = runq_choose(&tdq->tdq_realtime); 941 if (ts != NULL) 942 return (ts); 943 ts = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx); 944 if (ts != NULL) { 945 KASSERT(ts->ts_thread->td_priority >= PRI_MIN_TIMESHARE, 946 ("tdq_choose: Invalid priority on timeshare queue %d", 947 ts->ts_thread->td_priority)); 948 return (ts); 949 } 950 951 ts = runq_choose(&tdq->tdq_idle); 952 if (ts != NULL) { 953 KASSERT(ts->ts_thread->td_priority >= PRI_MIN_IDLE, 954 ("tdq_choose: Invalid priority on idle queue %d", 955 ts->ts_thread->td_priority)); 956 return (ts); 957 } 958 959 return (NULL); 960 } 961 962 static void 963 tdq_setup(struct tdq *tdq) 964 { 965 runq_init(&tdq->tdq_realtime); 966 runq_init(&tdq->tdq_timeshare); 967 runq_init(&tdq->tdq_idle); 968 tdq->tdq_load = 0; 969 } 970 971 static void 972 sched_setup(void *dummy) 973 { 974 #ifdef SMP 975 int i; 976 #endif 977 978 /* 979 * To avoid divide-by-zero, we set realstathz a dummy value 980 * in case which sched_clock() called before sched_initticks(). 981 */ 982 realstathz = hz; 983 sched_slice = (realstathz/10); /* ~100ms */ 984 tickincr = 1 << SCHED_TICK_SHIFT; 985 986 #ifdef SMP 987 balance_groups = 0; 988 /* 989 * Initialize the tdqs. 990 */ 991 for (i = 0; i < MAXCPU; i++) { 992 struct tdq *tdq; 993 994 tdq = &tdq_cpu[i]; 995 tdq_setup(&tdq_cpu[i]); 996 } 997 if (smp_topology == NULL) { 998 struct tdq_group *tdg; 999 struct tdq *tdq; 1000 int cpus; 1001 1002 for (cpus = 0, i = 0; i < MAXCPU; i++) { 1003 if (CPU_ABSENT(i)) 1004 continue; 1005 tdq = &tdq_cpu[i]; 1006 tdg = &tdq_groups[cpus]; 1007 /* 1008 * Setup a tdq group with one member. 1009 */ 1010 tdq->tdq_transferable = 0; 1011 tdq->tdq_group = tdg; 1012 tdg->tdg_cpus = 1; 1013 tdg->tdg_idlemask = 0; 1014 tdg->tdg_cpumask = tdg->tdg_mask = 1 << i; 1015 tdg->tdg_load = 0; 1016 tdg->tdg_transferable = 0; 1017 LIST_INIT(&tdg->tdg_members); 1018 LIST_INSERT_HEAD(&tdg->tdg_members, tdq, tdq_siblings); 1019 cpus++; 1020 } 1021 tdg_maxid = cpus - 1; 1022 } else { 1023 struct tdq_group *tdg; 1024 struct cpu_group *cg; 1025 int j; 1026 1027 topology = 1; 1028 for (i = 0; i < smp_topology->ct_count; i++) { 1029 cg = &smp_topology->ct_group[i]; 1030 tdg = &tdq_groups[i]; 1031 /* 1032 * Initialize the group. 1033 */ 1034 tdg->tdg_idlemask = 0; 1035 tdg->tdg_load = 0; 1036 tdg->tdg_transferable = 0; 1037 tdg->tdg_cpus = cg->cg_count; 1038 tdg->tdg_cpumask = cg->cg_mask; 1039 LIST_INIT(&tdg->tdg_members); 1040 /* 1041 * Find all of the group members and add them. 1042 */ 1043 for (j = 0; j < MAXCPU; j++) { 1044 if ((cg->cg_mask & (1 << j)) != 0) { 1045 if (tdg->tdg_mask == 0) 1046 tdg->tdg_mask = 1 << j; 1047 tdq_cpu[j].tdq_transferable = 0; 1048 tdq_cpu[j].tdq_group = tdg; 1049 LIST_INSERT_HEAD(&tdg->tdg_members, 1050 &tdq_cpu[j], tdq_siblings); 1051 } 1052 } 1053 if (tdg->tdg_cpus > 1) 1054 balance_groups = 1; 1055 } 1056 tdg_maxid = smp_topology->ct_count - 1; 1057 } 1058 /* 1059 * Stagger the group and global load balancer so they do not 1060 * interfere with each other. 1061 */ 1062 bal_tick = ticks + hz; 1063 if (balance_groups) 1064 gbal_tick = ticks + (hz / 2); 1065 #else 1066 tdq_setup(TDQ_SELF()); 1067 #endif 1068 mtx_lock_spin(&sched_lock); 1069 tdq_load_add(TDQ_SELF(), &td_sched0); 1070 mtx_unlock_spin(&sched_lock); 1071 } 1072 1073 /* ARGSUSED */ 1074 static void 1075 sched_initticks(void *dummy) 1076 { 1077 mtx_lock_spin(&sched_lock); 1078 realstathz = stathz ? stathz : hz; 1079 sched_slice = (realstathz/10); /* ~100ms */ 1080 1081 /* 1082 * tickincr is shifted out by 10 to avoid rounding errors due to 1083 * hz not being evenly divisible by stathz on all platforms. 1084 */ 1085 tickincr = (hz << SCHED_TICK_SHIFT) / realstathz; 1086 /* 1087 * This does not work for values of stathz that are more than 1088 * 1 << SCHED_TICK_SHIFT * hz. In practice this does not happen. 1089 */ 1090 if (tickincr == 0) 1091 tickincr = 1; 1092 #ifdef SMP 1093 affinity = SCHED_AFFINITY_DEFAULT; 1094 #endif 1095 mtx_unlock_spin(&sched_lock); 1096 } 1097 1098 1099 /* 1100 * Scale the scheduling priority according to the "interactivity" of this 1101 * process. 1102 */ 1103 static void 1104 sched_priority(struct thread *td) 1105 { 1106 int score; 1107 int pri; 1108 1109 if (td->td_pri_class != PRI_TIMESHARE) 1110 return; 1111 /* 1112 * If the score is interactive we place the thread in the realtime 1113 * queue with a priority that is less than kernel and interrupt 1114 * priorities. These threads are not subject to nice restrictions. 1115 * 1116 * Scores greater than this are placed on the normal realtime queue 1117 * where the priority is partially decided by the most recent cpu 1118 * utilization and the rest is decided by nice value. 1119 */ 1120 score = sched_interact_score(td); 1121 if (score < sched_interact) { 1122 pri = PRI_MIN_REALTIME; 1123 pri += ((PRI_MAX_REALTIME - PRI_MIN_REALTIME) / sched_interact) 1124 * score; 1125 KASSERT(pri >= PRI_MIN_REALTIME && pri <= PRI_MAX_REALTIME, 1126 ("sched_priority: invalid interactive priority %d score %d", 1127 pri, score)); 1128 } else { 1129 pri = SCHED_PRI_MIN; 1130 if (td->td_sched->ts_ticks) 1131 pri += SCHED_PRI_TICKS(td->td_sched); 1132 pri += SCHED_PRI_NICE(td->td_proc->p_nice); 1133 if (!(pri >= PRI_MIN_TIMESHARE && pri <= PRI_MAX_TIMESHARE)) { 1134 static int once = 1; 1135 if (once) { 1136 printf("sched_priority: invalid priority %d", 1137 pri); 1138 printf("nice %d, ticks %d ftick %d ltick %d tick pri %d\n", 1139 td->td_proc->p_nice, 1140 td->td_sched->ts_ticks, 1141 td->td_sched->ts_ftick, 1142 td->td_sched->ts_ltick, 1143 SCHED_PRI_TICKS(td->td_sched)); 1144 once = 0; 1145 } 1146 pri = min(max(pri, PRI_MIN_TIMESHARE), 1147 PRI_MAX_TIMESHARE); 1148 } 1149 } 1150 sched_user_prio(td, pri); 1151 1152 return; 1153 } 1154 1155 /* 1156 * This routine enforces a maximum limit on the amount of scheduling history 1157 * kept. It is called after either the slptime or runtime is adjusted. 1158 */ 1159 static void 1160 sched_interact_update(struct thread *td) 1161 { 1162 struct td_sched *ts; 1163 u_int sum; 1164 1165 ts = td->td_sched; 1166 sum = ts->skg_runtime + ts->skg_slptime; 1167 if (sum < SCHED_SLP_RUN_MAX) 1168 return; 1169 /* 1170 * This only happens from two places: 1171 * 1) We have added an unusual amount of run time from fork_exit. 1172 * 2) We have added an unusual amount of sleep time from sched_sleep(). 1173 */ 1174 if (sum > SCHED_SLP_RUN_MAX * 2) { 1175 if (ts->skg_runtime > ts->skg_slptime) { 1176 ts->skg_runtime = SCHED_SLP_RUN_MAX; 1177 ts->skg_slptime = 1; 1178 } else { 1179 ts->skg_slptime = SCHED_SLP_RUN_MAX; 1180 ts->skg_runtime = 1; 1181 } 1182 return; 1183 } 1184 /* 1185 * If we have exceeded by more than 1/5th then the algorithm below 1186 * will not bring us back into range. Dividing by two here forces 1187 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX] 1188 */ 1189 if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) { 1190 ts->skg_runtime /= 2; 1191 ts->skg_slptime /= 2; 1192 return; 1193 } 1194 ts->skg_runtime = (ts->skg_runtime / 5) * 4; 1195 ts->skg_slptime = (ts->skg_slptime / 5) * 4; 1196 } 1197 1198 static void 1199 sched_interact_fork(struct thread *td) 1200 { 1201 int ratio; 1202 int sum; 1203 1204 sum = td->td_sched->skg_runtime + td->td_sched->skg_slptime; 1205 if (sum > SCHED_SLP_RUN_FORK) { 1206 ratio = sum / SCHED_SLP_RUN_FORK; 1207 td->td_sched->skg_runtime /= ratio; 1208 td->td_sched->skg_slptime /= ratio; 1209 } 1210 } 1211 1212 static int 1213 sched_interact_score(struct thread *td) 1214 { 1215 int div; 1216 1217 if (td->td_sched->skg_runtime > td->td_sched->skg_slptime) { 1218 div = max(1, td->td_sched->skg_runtime / SCHED_INTERACT_HALF); 1219 return (SCHED_INTERACT_HALF + 1220 (SCHED_INTERACT_HALF - (td->td_sched->skg_slptime / div))); 1221 } 1222 if (td->td_sched->skg_slptime > td->td_sched->skg_runtime) { 1223 div = max(1, td->td_sched->skg_slptime / SCHED_INTERACT_HALF); 1224 return (td->td_sched->skg_runtime / div); 1225 } 1226 /* runtime == slptime */ 1227 if (td->td_sched->skg_runtime) 1228 return (SCHED_INTERACT_HALF); 1229 1230 /* 1231 * This can happen if slptime and runtime are 0. 1232 */ 1233 return (0); 1234 1235 } 1236 1237 /* 1238 * Called from proc0_init() to bootstrap the scheduler. 1239 */ 1240 void 1241 schedinit(void) 1242 { 1243 1244 /* 1245 * Set up the scheduler specific parts of proc0. 1246 */ 1247 proc0.p_sched = NULL; /* XXX */ 1248 thread0.td_sched = &td_sched0; 1249 thread0.td_lock = &sched_lock; 1250 td_sched0.ts_ltick = ticks; 1251 td_sched0.ts_ftick = ticks; 1252 td_sched0.ts_thread = &thread0; 1253 } 1254 1255 /* 1256 * This is only somewhat accurate since given many processes of the same 1257 * priority they will switch when their slices run out, which will be 1258 * at most sched_slice stathz ticks. 1259 */ 1260 int 1261 sched_rr_interval(void) 1262 { 1263 1264 /* Convert sched_slice to hz */ 1265 return (hz/(realstathz/sched_slice)); 1266 } 1267 1268 static void 1269 sched_pctcpu_update(struct td_sched *ts) 1270 { 1271 1272 if (ts->ts_ticks == 0) 1273 return; 1274 if (ticks - (hz / 10) < ts->ts_ltick && 1275 SCHED_TICK_TOTAL(ts) < SCHED_TICK_MAX) 1276 return; 1277 /* 1278 * Adjust counters and watermark for pctcpu calc. 1279 */ 1280 if (ts->ts_ltick > ticks - SCHED_TICK_TARG) 1281 ts->ts_ticks = (ts->ts_ticks / (ticks - ts->ts_ftick)) * 1282 SCHED_TICK_TARG; 1283 else 1284 ts->ts_ticks = 0; 1285 ts->ts_ltick = ticks; 1286 ts->ts_ftick = ts->ts_ltick - SCHED_TICK_TARG; 1287 } 1288 1289 static void 1290 sched_thread_priority(struct thread *td, u_char prio) 1291 { 1292 struct td_sched *ts; 1293 1294 CTR6(KTR_SCHED, "sched_prio: %p(%s) prio %d newprio %d by %p(%s)", 1295 td, td->td_proc->p_comm, td->td_priority, prio, curthread, 1296 curthread->td_proc->p_comm); 1297 ts = td->td_sched; 1298 THREAD_LOCK_ASSERT(td, MA_OWNED); 1299 if (td->td_priority == prio) 1300 return; 1301 1302 if (TD_ON_RUNQ(td) && prio < td->td_priority) { 1303 /* 1304 * If the priority has been elevated due to priority 1305 * propagation, we may have to move ourselves to a new 1306 * queue. This could be optimized to not re-add in some 1307 * cases. 1308 */ 1309 MPASS(td->td_lock == &sched_lock); 1310 sched_rem(td); 1311 td->td_priority = prio; 1312 sched_add(td, SRQ_BORROWING|SRQ_OURSELF); 1313 } else 1314 td->td_priority = prio; 1315 } 1316 1317 /* 1318 * Update a thread's priority when it is lent another thread's 1319 * priority. 1320 */ 1321 void 1322 sched_lend_prio(struct thread *td, u_char prio) 1323 { 1324 1325 td->td_flags |= TDF_BORROWING; 1326 sched_thread_priority(td, prio); 1327 } 1328 1329 /* 1330 * Restore a thread's priority when priority propagation is 1331 * over. The prio argument is the minimum priority the thread 1332 * needs to have to satisfy other possible priority lending 1333 * requests. If the thread's regular priority is less 1334 * important than prio, the thread will keep a priority boost 1335 * of prio. 1336 */ 1337 void 1338 sched_unlend_prio(struct thread *td, u_char prio) 1339 { 1340 u_char base_pri; 1341 1342 if (td->td_base_pri >= PRI_MIN_TIMESHARE && 1343 td->td_base_pri <= PRI_MAX_TIMESHARE) 1344 base_pri = td->td_user_pri; 1345 else 1346 base_pri = td->td_base_pri; 1347 if (prio >= base_pri) { 1348 td->td_flags &= ~TDF_BORROWING; 1349 sched_thread_priority(td, base_pri); 1350 } else 1351 sched_lend_prio(td, prio); 1352 } 1353 1354 void 1355 sched_prio(struct thread *td, u_char prio) 1356 { 1357 u_char oldprio; 1358 1359 /* First, update the base priority. */ 1360 td->td_base_pri = prio; 1361 1362 /* 1363 * If the thread is borrowing another thread's priority, don't 1364 * ever lower the priority. 1365 */ 1366 if (td->td_flags & TDF_BORROWING && td->td_priority < prio) 1367 return; 1368 1369 /* Change the real priority. */ 1370 oldprio = td->td_priority; 1371 sched_thread_priority(td, prio); 1372 1373 /* 1374 * If the thread is on a turnstile, then let the turnstile update 1375 * its state. 1376 */ 1377 if (TD_ON_LOCK(td) && oldprio != prio) 1378 turnstile_adjust(td, oldprio); 1379 } 1380 1381 void 1382 sched_user_prio(struct thread *td, u_char prio) 1383 { 1384 u_char oldprio; 1385 1386 td->td_base_user_pri = prio; 1387 if (td->td_flags & TDF_UBORROWING && td->td_user_pri <= prio) 1388 return; 1389 oldprio = td->td_user_pri; 1390 td->td_user_pri = prio; 1391 1392 if (TD_ON_UPILOCK(td) && oldprio != prio) 1393 umtx_pi_adjust(td, oldprio); 1394 } 1395 1396 void 1397 sched_lend_user_prio(struct thread *td, u_char prio) 1398 { 1399 u_char oldprio; 1400 1401 td->td_flags |= TDF_UBORROWING; 1402 1403 oldprio = td->td_user_pri; 1404 td->td_user_pri = prio; 1405 1406 if (TD_ON_UPILOCK(td) && oldprio != prio) 1407 umtx_pi_adjust(td, oldprio); 1408 } 1409 1410 void 1411 sched_unlend_user_prio(struct thread *td, u_char prio) 1412 { 1413 u_char base_pri; 1414 1415 base_pri = td->td_base_user_pri; 1416 if (prio >= base_pri) { 1417 td->td_flags &= ~TDF_UBORROWING; 1418 sched_user_prio(td, base_pri); 1419 } else 1420 sched_lend_user_prio(td, prio); 1421 } 1422 1423 void 1424 sched_switch(struct thread *td, struct thread *newtd, int flags) 1425 { 1426 struct tdq *tdq; 1427 struct td_sched *ts; 1428 int preempt; 1429 1430 THREAD_LOCK_ASSERT(td, MA_OWNED); 1431 1432 preempt = flags & SW_PREEMPT; 1433 tdq = TDQ_SELF(); 1434 ts = td->td_sched; 1435 td->td_lastcpu = td->td_oncpu; 1436 td->td_oncpu = NOCPU; 1437 td->td_flags &= ~TDF_NEEDRESCHED; 1438 td->td_owepreempt = 0; 1439 /* 1440 * If the thread has been assigned it may be in the process of switching 1441 * to the new cpu. This is the case in sched_bind(). 1442 */ 1443 /* 1444 * Switch to the sched lock to fix things up and pick 1445 * a new thread. 1446 */ 1447 if (td->td_lock != &sched_lock) { 1448 mtx_lock_spin(&sched_lock); 1449 thread_unlock(td); 1450 } 1451 if (TD_IS_IDLETHREAD(td)) { 1452 MPASS(td->td_lock == &sched_lock); 1453 TD_SET_CAN_RUN(td); 1454 } else if (TD_IS_RUNNING(td)) { 1455 /* 1456 * Don't allow the thread to migrate 1457 * from a preemption. 1458 */ 1459 tdq_load_rem(tdq, ts); 1460 if (preempt) 1461 sched_pin_td(td); 1462 sched_add(td, preempt ? 1463 SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED : 1464 SRQ_OURSELF|SRQ_YIELDING); 1465 if (preempt) 1466 sched_unpin_td(td); 1467 } else 1468 tdq_load_rem(tdq, ts); 1469 mtx_assert(&sched_lock, MA_OWNED); 1470 if (newtd != NULL) { 1471 /* 1472 * If we bring in a thread account for it as if it had been 1473 * added to the run queue and then chosen. 1474 */ 1475 TD_SET_RUNNING(newtd); 1476 tdq_load_add(TDQ_SELF(), newtd->td_sched); 1477 } else 1478 newtd = choosethread(); 1479 if (td != newtd) { 1480 #ifdef HWPMC_HOOKS 1481 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1482 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT); 1483 #endif 1484 1485 cpu_switch(td, newtd, td->td_lock); 1486 #ifdef HWPMC_HOOKS 1487 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1488 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN); 1489 #endif 1490 } 1491 sched_lock.mtx_lock = (uintptr_t)td; 1492 td->td_oncpu = PCPU_GET(cpuid); 1493 MPASS(td->td_lock == &sched_lock); 1494 } 1495 1496 void 1497 sched_nice(struct proc *p, int nice) 1498 { 1499 struct thread *td; 1500 1501 PROC_LOCK_ASSERT(p, MA_OWNED); 1502 PROC_SLOCK_ASSERT(p, MA_OWNED); 1503 1504 p->p_nice = nice; 1505 FOREACH_THREAD_IN_PROC(p, td) { 1506 thread_lock(td); 1507 sched_priority(td); 1508 sched_prio(td, td->td_base_user_pri); 1509 thread_unlock(td); 1510 } 1511 } 1512 1513 void 1514 sched_sleep(struct thread *td) 1515 { 1516 1517 THREAD_LOCK_ASSERT(td, MA_OWNED); 1518 1519 td->td_sched->ts_slptime = ticks; 1520 } 1521 1522 void 1523 sched_wakeup(struct thread *td) 1524 { 1525 struct td_sched *ts; 1526 int slptime; 1527 1528 THREAD_LOCK_ASSERT(td, MA_OWNED); 1529 ts = td->td_sched; 1530 /* 1531 * If we slept for more than a tick update our interactivity and 1532 * priority. 1533 */ 1534 slptime = ts->ts_slptime; 1535 ts->ts_slptime = 0; 1536 if (slptime && slptime != ticks) { 1537 u_int hzticks; 1538 1539 hzticks = (ticks - slptime) << SCHED_TICK_SHIFT; 1540 ts->skg_slptime += hzticks; 1541 sched_interact_update(td); 1542 sched_pctcpu_update(ts); 1543 sched_priority(td); 1544 } 1545 /* Reset the slice value after we sleep. */ 1546 ts->ts_slice = sched_slice; 1547 sched_add(td, SRQ_BORING); 1548 } 1549 1550 /* 1551 * Penalize the parent for creating a new child and initialize the child's 1552 * priority. 1553 */ 1554 void 1555 sched_fork(struct thread *td, struct thread *child) 1556 { 1557 THREAD_LOCK_ASSERT(td, MA_OWNED); 1558 sched_fork_thread(td, child); 1559 /* 1560 * Penalize the parent and child for forking. 1561 */ 1562 sched_interact_fork(child); 1563 sched_priority(child); 1564 td->td_sched->skg_runtime += tickincr; 1565 sched_interact_update(td); 1566 sched_priority(td); 1567 } 1568 1569 void 1570 sched_fork_thread(struct thread *td, struct thread *child) 1571 { 1572 struct td_sched *ts; 1573 struct td_sched *ts2; 1574 1575 /* 1576 * Initialize child. 1577 */ 1578 THREAD_LOCK_ASSERT(td, MA_OWNED); 1579 sched_newthread(child); 1580 child->td_lock = &sched_lock; 1581 ts = td->td_sched; 1582 ts2 = child->td_sched; 1583 ts2->ts_cpu = ts->ts_cpu; 1584 ts2->ts_runq = NULL; 1585 /* 1586 * Grab our parents cpu estimation information and priority. 1587 */ 1588 ts2->ts_ticks = ts->ts_ticks; 1589 ts2->ts_ltick = ts->ts_ltick; 1590 ts2->ts_ftick = ts->ts_ftick; 1591 child->td_user_pri = td->td_user_pri; 1592 child->td_base_user_pri = td->td_base_user_pri; 1593 /* 1594 * And update interactivity score. 1595 */ 1596 ts2->skg_slptime = ts->skg_slptime; 1597 ts2->skg_runtime = ts->skg_runtime; 1598 ts2->ts_slice = 1; /* Attempt to quickly learn interactivity. */ 1599 } 1600 1601 void 1602 sched_class(struct thread *td, int class) 1603 { 1604 1605 THREAD_LOCK_ASSERT(td, MA_OWNED); 1606 if (td->td_pri_class == class) 1607 return; 1608 1609 #ifdef SMP 1610 /* 1611 * On SMP if we're on the RUNQ we must adjust the transferable 1612 * count because could be changing to or from an interrupt 1613 * class. 1614 */ 1615 if (TD_ON_RUNQ(td)) { 1616 struct tdq *tdq; 1617 1618 tdq = TDQ_CPU(td->td_sched->ts_cpu); 1619 if (THREAD_CAN_MIGRATE(td)) { 1620 tdq->tdq_transferable--; 1621 tdq->tdq_group->tdg_transferable--; 1622 } 1623 td->td_pri_class = class; 1624 if (THREAD_CAN_MIGRATE(td)) { 1625 tdq->tdq_transferable++; 1626 tdq->tdq_group->tdg_transferable++; 1627 } 1628 } 1629 #endif 1630 td->td_pri_class = class; 1631 } 1632 1633 /* 1634 * Return some of the child's priority and interactivity to the parent. 1635 */ 1636 void 1637 sched_exit(struct proc *p, struct thread *child) 1638 { 1639 struct thread *td; 1640 1641 CTR3(KTR_SCHED, "sched_exit: %p(%s) prio %d", 1642 child, child->td_proc->p_comm, child->td_priority); 1643 1644 PROC_SLOCK_ASSERT(p, MA_OWNED); 1645 td = FIRST_THREAD_IN_PROC(p); 1646 sched_exit_thread(td, child); 1647 } 1648 1649 void 1650 sched_exit_thread(struct thread *td, struct thread *child) 1651 { 1652 1653 CTR3(KTR_SCHED, "sched_exit_thread: %p(%s) prio %d", 1654 child, child->td_proc->p_comm, child->td_priority); 1655 1656 thread_lock(child); 1657 tdq_load_rem(TDQ_CPU(child->td_sched->ts_cpu), child->td_sched); 1658 thread_unlock(child); 1659 #ifdef KSE 1660 /* 1661 * KSE forks and exits so often that this penalty causes short-lived 1662 * threads to always be non-interactive. This causes mozilla to 1663 * crawl under load. 1664 */ 1665 if ((td->td_pflags & TDP_SA) && td->td_proc == child->td_proc) 1666 return; 1667 #endif 1668 /* 1669 * Give the child's runtime to the parent without returning the 1670 * sleep time as a penalty to the parent. This causes shells that 1671 * launch expensive things to mark their children as expensive. 1672 */ 1673 thread_lock(td); 1674 td->td_sched->skg_runtime += child->td_sched->skg_runtime; 1675 sched_interact_update(td); 1676 sched_priority(td); 1677 thread_unlock(td); 1678 } 1679 1680 void 1681 sched_userret(struct thread *td) 1682 { 1683 /* 1684 * XXX we cheat slightly on the locking here to avoid locking in 1685 * the usual case. Setting td_priority here is essentially an 1686 * incomplete workaround for not setting it properly elsewhere. 1687 * Now that some interrupt handlers are threads, not setting it 1688 * properly elsewhere can clobber it in the window between setting 1689 * it here and returning to user mode, so don't waste time setting 1690 * it perfectly here. 1691 */ 1692 KASSERT((td->td_flags & TDF_BORROWING) == 0, 1693 ("thread with borrowed priority returning to userland")); 1694 if (td->td_priority != td->td_user_pri) { 1695 thread_lock(td); 1696 td->td_priority = td->td_user_pri; 1697 td->td_base_pri = td->td_user_pri; 1698 thread_unlock(td); 1699 } 1700 } 1701 1702 void 1703 sched_clock(struct thread *td) 1704 { 1705 struct tdq *tdq; 1706 struct td_sched *ts; 1707 1708 mtx_assert(&sched_lock, MA_OWNED); 1709 #ifdef SMP 1710 sched_smp_tick(td); 1711 #endif 1712 tdq = TDQ_SELF(); 1713 /* 1714 * Advance the insert index once for each tick to ensure that all 1715 * threads get a chance to run. 1716 */ 1717 if (tdq->tdq_idx == tdq->tdq_ridx) { 1718 tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS; 1719 if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx])) 1720 tdq->tdq_ridx = tdq->tdq_idx; 1721 } 1722 ts = td->td_sched; 1723 /* 1724 * We only do slicing code for TIMESHARE threads. 1725 */ 1726 if (td->td_pri_class != PRI_TIMESHARE) 1727 return; 1728 /* 1729 * We used a tick; charge it to the thread so that we can compute our 1730 * interactivity. 1731 */ 1732 td->td_sched->skg_runtime += tickincr; 1733 sched_interact_update(td); 1734 /* 1735 * We used up one time slice. 1736 */ 1737 if (--ts->ts_slice > 0) 1738 return; 1739 /* 1740 * We're out of time, recompute priorities and requeue. 1741 */ 1742 sched_priority(td); 1743 td->td_flags |= TDF_NEEDRESCHED; 1744 } 1745 1746 int 1747 sched_runnable(void) 1748 { 1749 struct tdq *tdq; 1750 int load; 1751 1752 load = 1; 1753 1754 tdq = TDQ_SELF(); 1755 #ifdef SMP 1756 if (tdq_busy) 1757 goto out; 1758 #endif 1759 if ((curthread->td_flags & TDF_IDLETD) != 0) { 1760 if (tdq->tdq_load > 0) 1761 goto out; 1762 } else 1763 if (tdq->tdq_load - 1 > 0) 1764 goto out; 1765 load = 0; 1766 out: 1767 return (load); 1768 } 1769 1770 struct thread * 1771 sched_choose(void) 1772 { 1773 struct tdq *tdq; 1774 struct td_sched *ts; 1775 1776 mtx_assert(&sched_lock, MA_OWNED); 1777 tdq = TDQ_SELF(); 1778 #ifdef SMP 1779 restart: 1780 #endif 1781 ts = tdq_choose(tdq); 1782 if (ts) { 1783 #ifdef SMP 1784 if (ts->ts_thread->td_priority > PRI_MIN_IDLE) 1785 if (tdq_idled(tdq) == 0) 1786 goto restart; 1787 #endif 1788 tdq_runq_rem(tdq, ts); 1789 return (ts->ts_thread); 1790 } 1791 #ifdef SMP 1792 if (tdq_idled(tdq) == 0) 1793 goto restart; 1794 #endif 1795 return (PCPU_GET(idlethread)); 1796 } 1797 1798 static int 1799 sched_preempt(struct thread *td) 1800 { 1801 struct thread *ctd; 1802 int cpri; 1803 int pri; 1804 1805 ctd = curthread; 1806 pri = td->td_priority; 1807 cpri = ctd->td_priority; 1808 if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd)) 1809 return (0); 1810 /* 1811 * Always preempt IDLE threads. Otherwise only if the preempting 1812 * thread is an ithread. 1813 */ 1814 if (pri > PRI_MAX_ITHD && cpri < PRI_MIN_IDLE) 1815 return (0); 1816 if (ctd->td_critnest > 1) { 1817 CTR1(KTR_PROC, "sched_preempt: in critical section %d", 1818 ctd->td_critnest); 1819 ctd->td_owepreempt = 1; 1820 return (0); 1821 } 1822 /* 1823 * Thread is runnable but not yet put on system run queue. 1824 */ 1825 MPASS(TD_ON_RUNQ(td)); 1826 TD_SET_RUNNING(td); 1827 MPASS(ctd->td_lock == &sched_lock); 1828 MPASS(td->td_lock == &sched_lock); 1829 CTR3(KTR_PROC, "preempting to thread %p (pid %d, %s)\n", td, 1830 td->td_proc->p_pid, td->td_proc->p_comm); 1831 /* 1832 * We enter the switch with two runnable threads that both have 1833 * the same lock. When we return td may be sleeping so we need 1834 * to switch locks to make sure he's locked correctly. 1835 */ 1836 SCHED_STAT_INC(switch_preempt); 1837 mi_switch(SW_INVOL|SW_PREEMPT, td); 1838 spinlock_enter(); 1839 thread_unlock(ctd); 1840 thread_lock(td); 1841 spinlock_exit(); 1842 1843 return (1); 1844 } 1845 1846 void 1847 sched_add(struct thread *td, int flags) 1848 { 1849 struct tdq *tdq; 1850 struct td_sched *ts; 1851 int preemptive; 1852 int class; 1853 #ifdef SMP 1854 int cpuid; 1855 int cpumask; 1856 #endif 1857 ts = td->td_sched; 1858 1859 THREAD_LOCK_ASSERT(td, MA_OWNED); 1860 CTR5(KTR_SCHED, "sched_add: %p(%s) prio %d by %p(%s)", 1861 td, td->td_proc->p_comm, td->td_priority, curthread, 1862 curthread->td_proc->p_comm); 1863 KASSERT((td->td_inhibitors == 0), 1864 ("sched_add: trying to run inhibited thread")); 1865 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)), 1866 ("sched_add: bad thread state")); 1867 KASSERT(td->td_proc->p_sflag & PS_INMEM, 1868 ("sched_add: process swapped out")); 1869 /* 1870 * Now that the thread is moving to the run-queue, set the lock 1871 * to the scheduler's lock. 1872 */ 1873 if (td->td_lock != &sched_lock) { 1874 mtx_lock_spin(&sched_lock); 1875 thread_lock_set(td, &sched_lock); 1876 } 1877 mtx_assert(&sched_lock, MA_OWNED); 1878 TD_SET_RUNQ(td); 1879 tdq = TDQ_SELF(); 1880 class = PRI_BASE(td->td_pri_class); 1881 preemptive = !(flags & SRQ_YIELDING); 1882 /* 1883 * Recalculate the priority before we select the target cpu or 1884 * run-queue. 1885 */ 1886 if (class == PRI_TIMESHARE) 1887 sched_priority(td); 1888 if (ts->ts_slice == 0) 1889 ts->ts_slice = sched_slice; 1890 #ifdef SMP 1891 cpuid = PCPU_GET(cpuid); 1892 /* 1893 * Pick the destination cpu and if it isn't ours transfer to the 1894 * target cpu. 1895 */ 1896 if (THREAD_CAN_MIGRATE(td)) { 1897 if (td->td_priority <= PRI_MAX_ITHD) { 1898 CTR2(KTR_ULE, "ithd %d < %d", 1899 td->td_priority, PRI_MAX_ITHD); 1900 ts->ts_cpu = cpuid; 1901 } else if (pick_pri) 1902 ts->ts_cpu = tdq_pickpri(tdq, ts, flags); 1903 else 1904 ts->ts_cpu = tdq_pickidle(tdq, ts); 1905 } else 1906 CTR1(KTR_ULE, "pinned %d", td->td_pinned); 1907 if (ts->ts_cpu != cpuid) 1908 preemptive = 0; 1909 tdq = TDQ_CPU(ts->ts_cpu); 1910 cpumask = 1 << ts->ts_cpu; 1911 /* 1912 * If we had been idle, clear our bit in the group and potentially 1913 * the global bitmap. 1914 */ 1915 if ((class != PRI_IDLE && class != PRI_ITHD) && 1916 (tdq->tdq_group->tdg_idlemask & cpumask) != 0) { 1917 /* 1918 * Check to see if our group is unidling, and if so, remove it 1919 * from the global idle mask. 1920 */ 1921 if (tdq->tdq_group->tdg_idlemask == 1922 tdq->tdq_group->tdg_cpumask) 1923 atomic_clear_int(&tdq_idle, tdq->tdq_group->tdg_mask); 1924 /* 1925 * Now remove ourselves from the group specific idle mask. 1926 */ 1927 tdq->tdq_group->tdg_idlemask &= ~cpumask; 1928 } 1929 #endif 1930 /* 1931 * Pick the run queue based on priority. 1932 */ 1933 if (td->td_priority <= PRI_MAX_REALTIME) 1934 ts->ts_runq = &tdq->tdq_realtime; 1935 else if (td->td_priority <= PRI_MAX_TIMESHARE) 1936 ts->ts_runq = &tdq->tdq_timeshare; 1937 else 1938 ts->ts_runq = &tdq->tdq_idle; 1939 if (preemptive && sched_preempt(td)) 1940 return; 1941 tdq_runq_add(tdq, ts, flags); 1942 tdq_load_add(tdq, ts); 1943 #ifdef SMP 1944 if (ts->ts_cpu != cpuid) { 1945 tdq_notify(ts); 1946 return; 1947 } 1948 #endif 1949 if (td->td_priority < curthread->td_priority) 1950 curthread->td_flags |= TDF_NEEDRESCHED; 1951 } 1952 1953 void 1954 sched_rem(struct thread *td) 1955 { 1956 struct tdq *tdq; 1957 struct td_sched *ts; 1958 1959 CTR5(KTR_SCHED, "sched_rem: %p(%s) prio %d by %p(%s)", 1960 td, td->td_proc->p_comm, td->td_priority, curthread, 1961 curthread->td_proc->p_comm); 1962 THREAD_LOCK_ASSERT(td, MA_OWNED); 1963 ts = td->td_sched; 1964 KASSERT(TD_ON_RUNQ(td), 1965 ("sched_rem: thread not on run queue")); 1966 1967 tdq = TDQ_CPU(ts->ts_cpu); 1968 tdq_runq_rem(tdq, ts); 1969 tdq_load_rem(tdq, ts); 1970 TD_SET_CAN_RUN(td); 1971 } 1972 1973 fixpt_t 1974 sched_pctcpu(struct thread *td) 1975 { 1976 fixpt_t pctcpu; 1977 struct td_sched *ts; 1978 1979 pctcpu = 0; 1980 ts = td->td_sched; 1981 if (ts == NULL) 1982 return (0); 1983 1984 thread_lock(td); 1985 if (ts->ts_ticks) { 1986 int rtick; 1987 1988 sched_pctcpu_update(ts); 1989 /* How many rtick per second ? */ 1990 rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz); 1991 pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT; 1992 } 1993 td->td_proc->p_swtime = ts->ts_ltick - ts->ts_ftick; 1994 thread_unlock(td); 1995 1996 return (pctcpu); 1997 } 1998 1999 void 2000 sched_bind(struct thread *td, int cpu) 2001 { 2002 struct td_sched *ts; 2003 2004 THREAD_LOCK_ASSERT(td, MA_OWNED); 2005 ts = td->td_sched; 2006 if (ts->ts_flags & TSF_BOUND) 2007 sched_unbind(td); 2008 ts->ts_flags |= TSF_BOUND; 2009 #ifdef SMP 2010 sched_pin(); 2011 if (PCPU_GET(cpuid) == cpu) 2012 return; 2013 ts->ts_cpu = cpu; 2014 /* When we return from mi_switch we'll be on the correct cpu. */ 2015 mi_switch(SW_VOL, NULL); 2016 #endif 2017 } 2018 2019 void 2020 sched_unbind(struct thread *td) 2021 { 2022 struct td_sched *ts; 2023 2024 THREAD_LOCK_ASSERT(td, MA_OWNED); 2025 ts = td->td_sched; 2026 if ((ts->ts_flags & TSF_BOUND) == 0) 2027 return; 2028 ts->ts_flags &= ~TSF_BOUND; 2029 #ifdef SMP 2030 sched_unpin(); 2031 #endif 2032 } 2033 2034 int 2035 sched_is_bound(struct thread *td) 2036 { 2037 THREAD_LOCK_ASSERT(td, MA_OWNED); 2038 return (td->td_sched->ts_flags & TSF_BOUND); 2039 } 2040 2041 void 2042 sched_relinquish(struct thread *td) 2043 { 2044 thread_lock(td); 2045 if (td->td_pri_class == PRI_TIMESHARE) 2046 sched_prio(td, PRI_MAX_TIMESHARE); 2047 SCHED_STAT_INC(switch_relinquish); 2048 mi_switch(SW_VOL, NULL); 2049 thread_unlock(td); 2050 } 2051 2052 int 2053 sched_load(void) 2054 { 2055 #ifdef SMP 2056 int total; 2057 int i; 2058 2059 total = 0; 2060 for (i = 0; i <= tdg_maxid; i++) 2061 total += TDQ_GROUP(i)->tdg_load; 2062 return (total); 2063 #else 2064 return (TDQ_SELF()->tdq_sysload); 2065 #endif 2066 } 2067 2068 int 2069 sched_sizeof_proc(void) 2070 { 2071 return (sizeof(struct proc)); 2072 } 2073 2074 int 2075 sched_sizeof_thread(void) 2076 { 2077 return (sizeof(struct thread) + sizeof(struct td_sched)); 2078 } 2079 2080 void 2081 sched_tick(void) 2082 { 2083 struct td_sched *ts; 2084 2085 ts = curthread->td_sched; 2086 /* Adjust ticks for pctcpu */ 2087 ts->ts_ticks += 1 << SCHED_TICK_SHIFT; 2088 ts->ts_ltick = ticks; 2089 /* 2090 * Update if we've exceeded our desired tick threshhold by over one 2091 * second. 2092 */ 2093 if (ts->ts_ftick + SCHED_TICK_MAX < ts->ts_ltick) 2094 sched_pctcpu_update(ts); 2095 } 2096 2097 /* 2098 * The actual idle process. 2099 */ 2100 void 2101 sched_idletd(void *dummy) 2102 { 2103 struct proc *p; 2104 struct thread *td; 2105 2106 td = curthread; 2107 p = td->td_proc; 2108 mtx_assert(&Giant, MA_NOTOWNED); 2109 /* ULE Relies on preemption for idle interruption. */ 2110 for (;;) 2111 cpu_idle(); 2112 } 2113 2114 /* 2115 * A CPU is entering for the first time or a thread is exiting. 2116 */ 2117 void 2118 sched_throw(struct thread *td) 2119 { 2120 /* 2121 * Correct spinlock nesting. The idle thread context that we are 2122 * borrowing was created so that it would start out with a single 2123 * spin lock (sched_lock) held in fork_trampoline(). Since we've 2124 * explicitly acquired locks in this function, the nesting count 2125 * is now 2 rather than 1. Since we are nested, calling 2126 * spinlock_exit() will simply adjust the counts without allowing 2127 * spin lock using code to interrupt us. 2128 */ 2129 if (td == NULL) { 2130 mtx_lock_spin(&sched_lock); 2131 spinlock_exit(); 2132 } else { 2133 MPASS(td->td_lock == &sched_lock); 2134 } 2135 mtx_assert(&sched_lock, MA_OWNED); 2136 KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count")); 2137 PCPU_SET(switchtime, cpu_ticks()); 2138 PCPU_SET(switchticks, ticks); 2139 cpu_throw(td, choosethread()); /* doesn't return */ 2140 } 2141 2142 void 2143 sched_fork_exit(struct thread *td) 2144 { 2145 2146 /* 2147 * Finish setting up thread glue so that it begins execution in a 2148 * non-nested critical section with sched_lock held but not recursed. 2149 */ 2150 td->td_oncpu = PCPU_GET(cpuid); 2151 sched_lock.mtx_lock = (uintptr_t)td; 2152 THREAD_LOCK_ASSERT(td, MA_OWNED | MA_NOTRECURSED); 2153 } 2154 2155 static SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler"); 2156 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ule", 0, 2157 "Scheduler name"); 2158 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0, ""); 2159 SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0, ""); 2160 SYSCTL_INT(_kern_sched, OID_AUTO, tickincr, CTLFLAG_RD, &tickincr, 0, ""); 2161 SYSCTL_INT(_kern_sched, OID_AUTO, realstathz, CTLFLAG_RD, &realstathz, 0, ""); 2162 #ifdef SMP 2163 SYSCTL_INT(_kern_sched, OID_AUTO, pick_pri, CTLFLAG_RW, &pick_pri, 0, ""); 2164 SYSCTL_INT(_kern_sched, OID_AUTO, pick_pri_affinity, CTLFLAG_RW, 2165 &affinity, 0, ""); 2166 SYSCTL_INT(_kern_sched, OID_AUTO, pick_pri_tryself, CTLFLAG_RW, 2167 &tryself, 0, ""); 2168 SYSCTL_INT(_kern_sched, OID_AUTO, pick_pri_tryselfidle, CTLFLAG_RW, 2169 &tryselfidle, 0, ""); 2170 SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0, ""); 2171 SYSCTL_INT(_kern_sched, OID_AUTO, ipi_preempt, CTLFLAG_RW, &ipi_preempt, 0, ""); 2172 SYSCTL_INT(_kern_sched, OID_AUTO, ipi_ast, CTLFLAG_RW, &ipi_ast, 0, ""); 2173 SYSCTL_INT(_kern_sched, OID_AUTO, ipi_thresh, CTLFLAG_RW, &ipi_thresh, 0, ""); 2174 SYSCTL_INT(_kern_sched, OID_AUTO, steal_htt, CTLFLAG_RW, &steal_htt, 0, ""); 2175 SYSCTL_INT(_kern_sched, OID_AUTO, steal_busy, CTLFLAG_RW, &steal_busy, 0, ""); 2176 SYSCTL_INT(_kern_sched, OID_AUTO, busy_thresh, CTLFLAG_RW, &busy_thresh, 0, ""); 2177 SYSCTL_INT(_kern_sched, OID_AUTO, topology, CTLFLAG_RD, &topology, 0, ""); 2178 #endif 2179 2180 /* ps compat */ 2181 static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */ 2182 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 2183 2184 2185 #define KERN_SWITCH_INCLUDE 1 2186 #include "kern/kern_switch.c" 2187