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("$FreeBSD$"); 40 41 #include "opt_hwpmc_hooks.h" 42 #include "opt_sched.h" 43 44 #include <sys/param.h> 45 #include <sys/systm.h> 46 #include <sys/kdb.h> 47 #include <sys/kernel.h> 48 #include <sys/ktr.h> 49 #include <sys/lock.h> 50 #include <sys/mutex.h> 51 #include <sys/proc.h> 52 #include <sys/resource.h> 53 #include <sys/resourcevar.h> 54 #include <sys/sched.h> 55 #include <sys/smp.h> 56 #include <sys/sx.h> 57 #include <sys/sysctl.h> 58 #include <sys/sysproto.h> 59 #include <sys/turnstile.h> 60 #include <sys/umtx.h> 61 #include <sys/vmmeter.h> 62 #include <sys/cpuset.h> 63 #ifdef KTRACE 64 #include <sys/uio.h> 65 #include <sys/ktrace.h> 66 #endif 67 68 #ifdef HWPMC_HOOKS 69 #include <sys/pmckern.h> 70 #endif 71 72 #include <machine/cpu.h> 73 #include <machine/smp.h> 74 75 #if defined(__sparc64__) || defined(__mips__) 76 #error "This architecture is not currently compatible with ULE" 77 #endif 78 79 #define KTR_ULE 0 80 81 /* 82 * Thread scheduler specific section. All fields are protected 83 * by the thread lock. 84 */ 85 struct td_sched { 86 struct runq *ts_runq; /* Run-queue we're queued on. */ 87 short ts_flags; /* TSF_* flags. */ 88 u_char ts_cpu; /* CPU that we have affinity for. */ 89 int ts_rltick; /* Real last tick, for affinity. */ 90 int ts_slice; /* Ticks of slice remaining. */ 91 u_int ts_slptime; /* Number of ticks we vol. slept */ 92 u_int ts_runtime; /* Number of ticks we were running */ 93 int ts_ltick; /* Last tick that we were running on */ 94 int ts_ftick; /* First tick that we were running on */ 95 int ts_ticks; /* Tick count */ 96 }; 97 /* flags kept in ts_flags */ 98 #define TSF_BOUND 0x0001 /* Thread can not migrate. */ 99 #define TSF_XFERABLE 0x0002 /* Thread was added as transferable. */ 100 101 static struct td_sched td_sched0; 102 103 #define THREAD_CAN_MIGRATE(td) ((td)->td_pinned == 0) 104 #define THREAD_CAN_SCHED(td, cpu) \ 105 CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask) 106 107 /* 108 * Cpu percentage computation macros and defines. 109 * 110 * SCHED_TICK_SECS: Number of seconds to average the cpu usage across. 111 * SCHED_TICK_TARG: Number of hz ticks to average the cpu usage across. 112 * SCHED_TICK_MAX: Maximum number of ticks before scaling back. 113 * SCHED_TICK_SHIFT: Shift factor to avoid rounding away results. 114 * SCHED_TICK_HZ: Compute the number of hz ticks for a given ticks count. 115 * SCHED_TICK_TOTAL: Gives the amount of time we've been recording ticks. 116 */ 117 #define SCHED_TICK_SECS 10 118 #define SCHED_TICK_TARG (hz * SCHED_TICK_SECS) 119 #define SCHED_TICK_MAX (SCHED_TICK_TARG + hz) 120 #define SCHED_TICK_SHIFT 10 121 #define SCHED_TICK_HZ(ts) ((ts)->ts_ticks >> SCHED_TICK_SHIFT) 122 #define SCHED_TICK_TOTAL(ts) (max((ts)->ts_ltick - (ts)->ts_ftick, hz)) 123 124 /* 125 * These macros determine priorities for non-interactive threads. They are 126 * assigned a priority based on their recent cpu utilization as expressed 127 * by the ratio of ticks to the tick total. NHALF priorities at the start 128 * and end of the MIN to MAX timeshare range are only reachable with negative 129 * or positive nice respectively. 130 * 131 * PRI_RANGE: Priority range for utilization dependent priorities. 132 * PRI_NRESV: Number of nice values. 133 * PRI_TICKS: Compute a priority in PRI_RANGE from the ticks count and total. 134 * PRI_NICE: Determines the part of the priority inherited from nice. 135 */ 136 #define SCHED_PRI_NRESV (PRIO_MAX - PRIO_MIN) 137 #define SCHED_PRI_NHALF (SCHED_PRI_NRESV / 2) 138 #define SCHED_PRI_MIN (PRI_MIN_TIMESHARE + SCHED_PRI_NHALF) 139 #define SCHED_PRI_MAX (PRI_MAX_TIMESHARE - SCHED_PRI_NHALF) 140 #define SCHED_PRI_RANGE (SCHED_PRI_MAX - SCHED_PRI_MIN) 141 #define SCHED_PRI_TICKS(ts) \ 142 (SCHED_TICK_HZ((ts)) / \ 143 (roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE)) 144 #define SCHED_PRI_NICE(nice) (nice) 145 146 /* 147 * These determine the interactivity of a process. Interactivity differs from 148 * cpu utilization in that it expresses the voluntary time slept vs time ran 149 * while cpu utilization includes all time not running. This more accurately 150 * models the intent of the thread. 151 * 152 * SLP_RUN_MAX: Maximum amount of sleep time + run time we'll accumulate 153 * before throttling back. 154 * SLP_RUN_FORK: Maximum slp+run time to inherit at fork time. 155 * INTERACT_MAX: Maximum interactivity value. Smaller is better. 156 * INTERACT_THRESH: Threshhold for placement on the current runq. 157 */ 158 #define SCHED_SLP_RUN_MAX ((hz * 5) << SCHED_TICK_SHIFT) 159 #define SCHED_SLP_RUN_FORK ((hz / 2) << SCHED_TICK_SHIFT) 160 #define SCHED_INTERACT_MAX (100) 161 #define SCHED_INTERACT_HALF (SCHED_INTERACT_MAX / 2) 162 #define SCHED_INTERACT_THRESH (30) 163 164 /* 165 * tickincr: Converts a stathz tick into a hz domain scaled by 166 * the shift factor. Without the shift the error rate 167 * due to rounding would be unacceptably high. 168 * realstathz: stathz is sometimes 0 and run off of hz. 169 * sched_slice: Runtime of each thread before rescheduling. 170 * preempt_thresh: Priority threshold for preemption and remote IPIs. 171 */ 172 static int sched_interact = SCHED_INTERACT_THRESH; 173 static int realstathz; 174 static int tickincr; 175 static int sched_slice = 1; 176 #ifdef PREEMPTION 177 #ifdef FULL_PREEMPTION 178 static int preempt_thresh = PRI_MAX_IDLE; 179 #else 180 static int preempt_thresh = PRI_MIN_KERN; 181 #endif 182 #else 183 static int preempt_thresh = 0; 184 #endif 185 static int static_boost = PRI_MIN_TIMESHARE; 186 static int sched_idlespins = 10000; 187 static int sched_idlespinthresh = 4; 188 189 /* 190 * tdq - per processor runqs and statistics. All fields are protected by the 191 * tdq_lock. The load and lowpri may be accessed without to avoid excess 192 * locking in sched_pickcpu(); 193 */ 194 struct tdq { 195 /* Ordered to improve efficiency of cpu_search() and switch(). */ 196 struct mtx tdq_lock; /* run queue lock. */ 197 struct cpu_group *tdq_cg; /* Pointer to cpu topology. */ 198 volatile int tdq_load; /* Aggregate load. */ 199 int tdq_sysload; /* For loadavg, !ITHD load. */ 200 int tdq_transferable; /* Transferable thread count. */ 201 volatile int tdq_idlestate; /* State of the idle thread. */ 202 short tdq_switchcnt; /* Switches this tick. */ 203 short tdq_oldswitchcnt; /* Switches last tick. */ 204 u_char tdq_lowpri; /* Lowest priority thread. */ 205 u_char tdq_ipipending; /* IPI pending. */ 206 u_char tdq_idx; /* Current insert index. */ 207 u_char tdq_ridx; /* Current removal index. */ 208 struct runq tdq_realtime; /* real-time run queue. */ 209 struct runq tdq_timeshare; /* timeshare run queue. */ 210 struct runq tdq_idle; /* Queue of IDLE threads. */ 211 char tdq_name[sizeof("sched lock") + 6]; 212 } __aligned(64); 213 214 /* Idle thread states and config. */ 215 #define TDQ_RUNNING 1 216 #define TDQ_IDLE 2 217 218 #ifdef SMP 219 struct cpu_group *cpu_top; 220 221 #define SCHED_AFFINITY_DEFAULT (max(1, hz / 1000)) 222 #define SCHED_AFFINITY(ts, t) ((ts)->ts_rltick > ticks - ((t) * affinity)) 223 224 /* 225 * Run-time tunables. 226 */ 227 static int rebalance = 1; 228 static int balance_interval = 128; /* Default set in sched_initticks(). */ 229 static int affinity; 230 static int steal_htt = 1; 231 static int steal_idle = 1; 232 static int steal_thresh = 2; 233 234 /* 235 * One thread queue per processor. 236 */ 237 static struct tdq tdq_cpu[MAXCPU]; 238 static struct tdq *balance_tdq; 239 static int balance_ticks; 240 241 #define TDQ_SELF() (&tdq_cpu[PCPU_GET(cpuid)]) 242 #define TDQ_CPU(x) (&tdq_cpu[(x)]) 243 #define TDQ_ID(x) ((int)((x) - tdq_cpu)) 244 #else /* !SMP */ 245 static struct tdq tdq_cpu; 246 247 #define TDQ_ID(x) (0) 248 #define TDQ_SELF() (&tdq_cpu) 249 #define TDQ_CPU(x) (&tdq_cpu) 250 #endif 251 252 #define TDQ_LOCK_ASSERT(t, type) mtx_assert(TDQ_LOCKPTR((t)), (type)) 253 #define TDQ_LOCK(t) mtx_lock_spin(TDQ_LOCKPTR((t))) 254 #define TDQ_LOCK_FLAGS(t, f) mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f)) 255 #define TDQ_UNLOCK(t) mtx_unlock_spin(TDQ_LOCKPTR((t))) 256 #define TDQ_LOCKPTR(t) (&(t)->tdq_lock) 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 *); 264 265 /* Operations on per processor queues */ 266 static struct thread *tdq_choose(struct tdq *); 267 static void tdq_setup(struct tdq *); 268 static void tdq_load_add(struct tdq *, struct thread *); 269 static void tdq_load_rem(struct tdq *, struct thread *); 270 static __inline void tdq_runq_add(struct tdq *, struct thread *, int); 271 static __inline void tdq_runq_rem(struct tdq *, struct thread *); 272 static inline int sched_shouldpreempt(int, int, int); 273 void tdq_print(int cpu); 274 static void runq_print(struct runq *rq); 275 static void tdq_add(struct tdq *, struct thread *, int); 276 #ifdef SMP 277 static int tdq_move(struct tdq *, struct tdq *); 278 static int tdq_idled(struct tdq *); 279 static void tdq_notify(struct tdq *, struct thread *); 280 static struct thread *tdq_steal(struct tdq *, int); 281 static struct thread *runq_steal(struct runq *, int); 282 static int sched_pickcpu(struct thread *, int); 283 static void sched_balance(void); 284 static int sched_balance_pair(struct tdq *, struct tdq *); 285 static inline struct tdq *sched_setcpu(struct thread *, int, int); 286 static inline struct mtx *thread_block_switch(struct thread *); 287 static inline void thread_unblock_switch(struct thread *, struct mtx *); 288 static struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int); 289 #endif 290 291 static void sched_setup(void *dummy); 292 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL); 293 294 static void sched_initticks(void *dummy); 295 SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks, 296 NULL); 297 298 /* 299 * Print the threads waiting on a run-queue. 300 */ 301 static void 302 runq_print(struct runq *rq) 303 { 304 struct rqhead *rqh; 305 struct thread *td; 306 int pri; 307 int j; 308 int i; 309 310 for (i = 0; i < RQB_LEN; i++) { 311 printf("\t\trunq bits %d 0x%zx\n", 312 i, rq->rq_status.rqb_bits[i]); 313 for (j = 0; j < RQB_BPW; j++) 314 if (rq->rq_status.rqb_bits[i] & (1ul << j)) { 315 pri = j + (i << RQB_L2BPW); 316 rqh = &rq->rq_queues[pri]; 317 TAILQ_FOREACH(td, rqh, td_runq) { 318 printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n", 319 td, td->td_name, td->td_priority, 320 td->td_rqindex, pri); 321 } 322 } 323 } 324 } 325 326 /* 327 * Print the status of a per-cpu thread queue. Should be a ddb show cmd. 328 */ 329 void 330 tdq_print(int cpu) 331 { 332 struct tdq *tdq; 333 334 tdq = TDQ_CPU(cpu); 335 336 printf("tdq %d:\n", TDQ_ID(tdq)); 337 printf("\tlock %p\n", TDQ_LOCKPTR(tdq)); 338 printf("\tLock name: %s\n", tdq->tdq_name); 339 printf("\tload: %d\n", tdq->tdq_load); 340 printf("\tswitch cnt: %d\n", tdq->tdq_switchcnt); 341 printf("\told switch cnt: %d\n", tdq->tdq_oldswitchcnt); 342 printf("\tidle state: %d\n", tdq->tdq_idlestate); 343 printf("\ttimeshare idx: %d\n", tdq->tdq_idx); 344 printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx); 345 printf("\tload transferable: %d\n", tdq->tdq_transferable); 346 printf("\tlowest priority: %d\n", tdq->tdq_lowpri); 347 printf("\trealtime runq:\n"); 348 runq_print(&tdq->tdq_realtime); 349 printf("\ttimeshare runq:\n"); 350 runq_print(&tdq->tdq_timeshare); 351 printf("\tidle runq:\n"); 352 runq_print(&tdq->tdq_idle); 353 } 354 355 static inline int 356 sched_shouldpreempt(int pri, int cpri, int remote) 357 { 358 /* 359 * If the new priority is not better than the current priority there is 360 * nothing to do. 361 */ 362 if (pri >= cpri) 363 return (0); 364 /* 365 * Always preempt idle. 366 */ 367 if (cpri >= PRI_MIN_IDLE) 368 return (1); 369 /* 370 * If preemption is disabled don't preempt others. 371 */ 372 if (preempt_thresh == 0) 373 return (0); 374 /* 375 * Preempt if we exceed the threshold. 376 */ 377 if (pri <= preempt_thresh) 378 return (1); 379 /* 380 * If we're realtime or better and there is timeshare or worse running 381 * preempt only remote processors. 382 */ 383 if (remote && pri <= PRI_MAX_REALTIME && cpri > PRI_MAX_REALTIME) 384 return (1); 385 return (0); 386 } 387 388 #define TS_RQ_PPQ (((PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE) + 1) / RQ_NQS) 389 /* 390 * Add a thread to the actual run-queue. Keeps transferable counts up to 391 * date with what is actually on the run-queue. Selects the correct 392 * queue position for timeshare threads. 393 */ 394 static __inline void 395 tdq_runq_add(struct tdq *tdq, struct thread *td, int flags) 396 { 397 struct td_sched *ts; 398 u_char pri; 399 400 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 401 THREAD_LOCK_ASSERT(td, MA_OWNED); 402 403 pri = td->td_priority; 404 ts = td->td_sched; 405 TD_SET_RUNQ(td); 406 if (THREAD_CAN_MIGRATE(td)) { 407 tdq->tdq_transferable++; 408 ts->ts_flags |= TSF_XFERABLE; 409 } 410 if (pri <= PRI_MAX_REALTIME) { 411 ts->ts_runq = &tdq->tdq_realtime; 412 } else if (pri <= PRI_MAX_TIMESHARE) { 413 ts->ts_runq = &tdq->tdq_timeshare; 414 KASSERT(pri <= PRI_MAX_TIMESHARE && pri >= PRI_MIN_TIMESHARE, 415 ("Invalid priority %d on timeshare runq", pri)); 416 /* 417 * This queue contains only priorities between MIN and MAX 418 * realtime. Use the whole queue to represent these values. 419 */ 420 if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) { 421 pri = (pri - PRI_MIN_TIMESHARE) / TS_RQ_PPQ; 422 pri = (pri + tdq->tdq_idx) % RQ_NQS; 423 /* 424 * This effectively shortens the queue by one so we 425 * can have a one slot difference between idx and 426 * ridx while we wait for threads to drain. 427 */ 428 if (tdq->tdq_ridx != tdq->tdq_idx && 429 pri == tdq->tdq_ridx) 430 pri = (unsigned char)(pri - 1) % RQ_NQS; 431 } else 432 pri = tdq->tdq_ridx; 433 runq_add_pri(ts->ts_runq, td, pri, flags); 434 return; 435 } else 436 ts->ts_runq = &tdq->tdq_idle; 437 runq_add(ts->ts_runq, td, flags); 438 } 439 440 /* 441 * Remove a thread from a run-queue. This typically happens when a thread 442 * is selected to run. Running threads are not on the queue and the 443 * transferable count does not reflect them. 444 */ 445 static __inline void 446 tdq_runq_rem(struct tdq *tdq, struct thread *td) 447 { 448 struct td_sched *ts; 449 450 ts = td->td_sched; 451 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 452 KASSERT(ts->ts_runq != NULL, 453 ("tdq_runq_remove: thread %p null ts_runq", td)); 454 if (ts->ts_flags & TSF_XFERABLE) { 455 tdq->tdq_transferable--; 456 ts->ts_flags &= ~TSF_XFERABLE; 457 } 458 if (ts->ts_runq == &tdq->tdq_timeshare) { 459 if (tdq->tdq_idx != tdq->tdq_ridx) 460 runq_remove_idx(ts->ts_runq, td, &tdq->tdq_ridx); 461 else 462 runq_remove_idx(ts->ts_runq, td, NULL); 463 } else 464 runq_remove(ts->ts_runq, td); 465 } 466 467 /* 468 * Load is maintained for all threads RUNNING and ON_RUNQ. Add the load 469 * for this thread to the referenced thread queue. 470 */ 471 static void 472 tdq_load_add(struct tdq *tdq, struct thread *td) 473 { 474 475 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 476 THREAD_LOCK_ASSERT(td, MA_OWNED); 477 478 tdq->tdq_load++; 479 if ((td->td_proc->p_flag & P_NOLOAD) == 0) 480 tdq->tdq_sysload++; 481 CTR2(KTR_SCHED, "cpu %d load: %d", TDQ_ID(tdq), tdq->tdq_load); 482 } 483 484 /* 485 * Remove the load from a thread that is transitioning to a sleep state or 486 * exiting. 487 */ 488 static void 489 tdq_load_rem(struct tdq *tdq, struct thread *td) 490 { 491 492 THREAD_LOCK_ASSERT(td, MA_OWNED); 493 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 494 KASSERT(tdq->tdq_load != 0, 495 ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq))); 496 497 tdq->tdq_load--; 498 if ((td->td_proc->p_flag & P_NOLOAD) == 0) 499 tdq->tdq_sysload--; 500 CTR1(KTR_SCHED, "load: %d", tdq->tdq_load); 501 } 502 503 /* 504 * Set lowpri to its exact value by searching the run-queue and 505 * evaluating curthread. curthread may be passed as an optimization. 506 */ 507 static void 508 tdq_setlowpri(struct tdq *tdq, struct thread *ctd) 509 { 510 struct thread *td; 511 512 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 513 if (ctd == NULL) 514 ctd = pcpu_find(TDQ_ID(tdq))->pc_curthread; 515 td = tdq_choose(tdq); 516 if (td == NULL || td->td_priority > ctd->td_priority) 517 tdq->tdq_lowpri = ctd->td_priority; 518 else 519 tdq->tdq_lowpri = td->td_priority; 520 } 521 522 #ifdef SMP 523 struct cpu_search { 524 cpumask_t cs_mask; /* Mask of valid cpus. */ 525 u_int cs_load; 526 u_int cs_cpu; 527 int cs_limit; /* Min priority for low min load for high. */ 528 }; 529 530 #define CPU_SEARCH_LOWEST 0x1 531 #define CPU_SEARCH_HIGHEST 0x2 532 #define CPU_SEARCH_BOTH (CPU_SEARCH_LOWEST|CPU_SEARCH_HIGHEST) 533 534 #define CPUMASK_FOREACH(cpu, mask) \ 535 for ((cpu) = 0; (cpu) < sizeof((mask)) * 8; (cpu)++) \ 536 if ((mask) & 1 << (cpu)) 537 538 static __inline int cpu_search(struct cpu_group *cg, struct cpu_search *low, 539 struct cpu_search *high, const int match); 540 int cpu_search_lowest(struct cpu_group *cg, struct cpu_search *low); 541 int cpu_search_highest(struct cpu_group *cg, struct cpu_search *high); 542 int cpu_search_both(struct cpu_group *cg, struct cpu_search *low, 543 struct cpu_search *high); 544 545 /* 546 * This routine compares according to the match argument and should be 547 * reduced in actual instantiations via constant propagation and dead code 548 * elimination. 549 */ 550 static __inline int 551 cpu_compare(int cpu, struct cpu_search *low, struct cpu_search *high, 552 const int match) 553 { 554 struct tdq *tdq; 555 556 tdq = TDQ_CPU(cpu); 557 if (match & CPU_SEARCH_LOWEST) 558 if (low->cs_mask & (1 << cpu) && 559 tdq->tdq_load < low->cs_load && 560 tdq->tdq_lowpri > low->cs_limit) { 561 low->cs_cpu = cpu; 562 low->cs_load = tdq->tdq_load; 563 } 564 if (match & CPU_SEARCH_HIGHEST) 565 if (high->cs_mask & (1 << cpu) && 566 tdq->tdq_load >= high->cs_limit && 567 tdq->tdq_load > high->cs_load && 568 tdq->tdq_transferable) { 569 high->cs_cpu = cpu; 570 high->cs_load = tdq->tdq_load; 571 } 572 return (tdq->tdq_load); 573 } 574 575 /* 576 * Search the tree of cpu_groups for the lowest or highest loaded cpu 577 * according to the match argument. This routine actually compares the 578 * load on all paths through the tree and finds the least loaded cpu on 579 * the least loaded path, which may differ from the least loaded cpu in 580 * the system. This balances work among caches and busses. 581 * 582 * This inline is instantiated in three forms below using constants for the 583 * match argument. It is reduced to the minimum set for each case. It is 584 * also recursive to the depth of the tree. 585 */ 586 static __inline int 587 cpu_search(struct cpu_group *cg, struct cpu_search *low, 588 struct cpu_search *high, const int match) 589 { 590 int total; 591 592 total = 0; 593 if (cg->cg_children) { 594 struct cpu_search lgroup; 595 struct cpu_search hgroup; 596 struct cpu_group *child; 597 u_int lload; 598 int hload; 599 int load; 600 int i; 601 602 lload = -1; 603 hload = -1; 604 for (i = 0; i < cg->cg_children; i++) { 605 child = &cg->cg_child[i]; 606 if (match & CPU_SEARCH_LOWEST) { 607 lgroup = *low; 608 lgroup.cs_load = -1; 609 } 610 if (match & CPU_SEARCH_HIGHEST) { 611 hgroup = *high; 612 lgroup.cs_load = 0; 613 } 614 switch (match) { 615 case CPU_SEARCH_LOWEST: 616 load = cpu_search_lowest(child, &lgroup); 617 break; 618 case CPU_SEARCH_HIGHEST: 619 load = cpu_search_highest(child, &hgroup); 620 break; 621 case CPU_SEARCH_BOTH: 622 load = cpu_search_both(child, &lgroup, &hgroup); 623 break; 624 } 625 total += load; 626 if (match & CPU_SEARCH_LOWEST) 627 if (load < lload || low->cs_cpu == -1) { 628 *low = lgroup; 629 lload = load; 630 } 631 if (match & CPU_SEARCH_HIGHEST) 632 if (load > hload || high->cs_cpu == -1) { 633 hload = load; 634 *high = hgroup; 635 } 636 } 637 } else { 638 int cpu; 639 640 CPUMASK_FOREACH(cpu, cg->cg_mask) 641 total += cpu_compare(cpu, low, high, match); 642 } 643 return (total); 644 } 645 646 /* 647 * cpu_search instantiations must pass constants to maintain the inline 648 * optimization. 649 */ 650 int 651 cpu_search_lowest(struct cpu_group *cg, struct cpu_search *low) 652 { 653 return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST); 654 } 655 656 int 657 cpu_search_highest(struct cpu_group *cg, struct cpu_search *high) 658 { 659 return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST); 660 } 661 662 int 663 cpu_search_both(struct cpu_group *cg, struct cpu_search *low, 664 struct cpu_search *high) 665 { 666 return cpu_search(cg, low, high, CPU_SEARCH_BOTH); 667 } 668 669 /* 670 * Find the cpu with the least load via the least loaded path that has a 671 * lowpri greater than pri pri. A pri of -1 indicates any priority is 672 * acceptable. 673 */ 674 static inline int 675 sched_lowest(struct cpu_group *cg, cpumask_t mask, int pri) 676 { 677 struct cpu_search low; 678 679 low.cs_cpu = -1; 680 low.cs_load = -1; 681 low.cs_mask = mask; 682 low.cs_limit = pri; 683 cpu_search_lowest(cg, &low); 684 return low.cs_cpu; 685 } 686 687 /* 688 * Find the cpu with the highest load via the highest loaded path. 689 */ 690 static inline int 691 sched_highest(struct cpu_group *cg, cpumask_t mask, int minload) 692 { 693 struct cpu_search high; 694 695 high.cs_cpu = -1; 696 high.cs_load = 0; 697 high.cs_mask = mask; 698 high.cs_limit = minload; 699 cpu_search_highest(cg, &high); 700 return high.cs_cpu; 701 } 702 703 /* 704 * Simultaneously find the highest and lowest loaded cpu reachable via 705 * cg. 706 */ 707 static inline void 708 sched_both(struct cpu_group *cg, cpumask_t mask, int *lowcpu, int *highcpu) 709 { 710 struct cpu_search high; 711 struct cpu_search low; 712 713 low.cs_cpu = -1; 714 low.cs_limit = -1; 715 low.cs_load = -1; 716 low.cs_mask = mask; 717 high.cs_load = 0; 718 high.cs_cpu = -1; 719 high.cs_limit = -1; 720 high.cs_mask = mask; 721 cpu_search_both(cg, &low, &high); 722 *lowcpu = low.cs_cpu; 723 *highcpu = high.cs_cpu; 724 return; 725 } 726 727 static void 728 sched_balance_group(struct cpu_group *cg) 729 { 730 cpumask_t mask; 731 int high; 732 int low; 733 int i; 734 735 mask = -1; 736 for (;;) { 737 sched_both(cg, mask, &low, &high); 738 if (low == high || low == -1 || high == -1) 739 break; 740 if (sched_balance_pair(TDQ_CPU(high), TDQ_CPU(low))) 741 break; 742 /* 743 * If we failed to move any threads determine which cpu 744 * to kick out of the set and try again. 745 */ 746 if (TDQ_CPU(high)->tdq_transferable == 0) 747 mask &= ~(1 << high); 748 else 749 mask &= ~(1 << low); 750 } 751 752 for (i = 0; i < cg->cg_children; i++) 753 sched_balance_group(&cg->cg_child[i]); 754 } 755 756 static void 757 sched_balance() 758 { 759 struct tdq *tdq; 760 761 /* 762 * Select a random time between .5 * balance_interval and 763 * 1.5 * balance_interval. 764 */ 765 balance_ticks = max(balance_interval / 2, 1); 766 balance_ticks += random() % balance_interval; 767 if (smp_started == 0 || rebalance == 0) 768 return; 769 tdq = TDQ_SELF(); 770 TDQ_UNLOCK(tdq); 771 sched_balance_group(cpu_top); 772 TDQ_LOCK(tdq); 773 } 774 775 /* 776 * Lock two thread queues using their address to maintain lock order. 777 */ 778 static void 779 tdq_lock_pair(struct tdq *one, struct tdq *two) 780 { 781 if (one < two) { 782 TDQ_LOCK(one); 783 TDQ_LOCK_FLAGS(two, MTX_DUPOK); 784 } else { 785 TDQ_LOCK(two); 786 TDQ_LOCK_FLAGS(one, MTX_DUPOK); 787 } 788 } 789 790 /* 791 * Unlock two thread queues. Order is not important here. 792 */ 793 static void 794 tdq_unlock_pair(struct tdq *one, struct tdq *two) 795 { 796 TDQ_UNLOCK(one); 797 TDQ_UNLOCK(two); 798 } 799 800 /* 801 * Transfer load between two imbalanced thread queues. 802 */ 803 static int 804 sched_balance_pair(struct tdq *high, struct tdq *low) 805 { 806 int transferable; 807 int high_load; 808 int low_load; 809 int moved; 810 int move; 811 int diff; 812 int i; 813 814 tdq_lock_pair(high, low); 815 transferable = high->tdq_transferable; 816 high_load = high->tdq_load; 817 low_load = low->tdq_load; 818 moved = 0; 819 /* 820 * Determine what the imbalance is and then adjust that to how many 821 * threads we actually have to give up (transferable). 822 */ 823 if (transferable != 0) { 824 diff = high_load - low_load; 825 move = diff / 2; 826 if (diff & 0x1) 827 move++; 828 move = min(move, transferable); 829 for (i = 0; i < move; i++) 830 moved += tdq_move(high, low); 831 /* 832 * IPI the target cpu to force it to reschedule with the new 833 * workload. 834 */ 835 ipi_selected(1 << TDQ_ID(low), IPI_PREEMPT); 836 } 837 tdq_unlock_pair(high, low); 838 return (moved); 839 } 840 841 /* 842 * Move a thread from one thread queue to another. 843 */ 844 static int 845 tdq_move(struct tdq *from, struct tdq *to) 846 { 847 struct td_sched *ts; 848 struct thread *td; 849 struct tdq *tdq; 850 int cpu; 851 852 TDQ_LOCK_ASSERT(from, MA_OWNED); 853 TDQ_LOCK_ASSERT(to, MA_OWNED); 854 855 tdq = from; 856 cpu = TDQ_ID(to); 857 td = tdq_steal(tdq, cpu); 858 if (td == NULL) 859 return (0); 860 ts = td->td_sched; 861 /* 862 * Although the run queue is locked the thread may be blocked. Lock 863 * it to clear this and acquire the run-queue lock. 864 */ 865 thread_lock(td); 866 /* Drop recursive lock on from acquired via thread_lock(). */ 867 TDQ_UNLOCK(from); 868 sched_rem(td); 869 ts->ts_cpu = cpu; 870 td->td_lock = TDQ_LOCKPTR(to); 871 tdq_add(to, td, SRQ_YIELDING); 872 return (1); 873 } 874 875 /* 876 * This tdq has idled. Try to steal a thread from another cpu and switch 877 * to it. 878 */ 879 static int 880 tdq_idled(struct tdq *tdq) 881 { 882 struct cpu_group *cg; 883 struct tdq *steal; 884 cpumask_t mask; 885 int thresh; 886 int cpu; 887 888 if (smp_started == 0 || steal_idle == 0) 889 return (1); 890 mask = -1; 891 mask &= ~PCPU_GET(cpumask); 892 /* We don't want to be preempted while we're iterating. */ 893 spinlock_enter(); 894 for (cg = tdq->tdq_cg; cg != NULL; ) { 895 if ((cg->cg_flags & (CG_FLAG_HTT | CG_FLAG_THREAD)) == 0) 896 thresh = steal_thresh; 897 else 898 thresh = 1; 899 cpu = sched_highest(cg, mask, thresh); 900 if (cpu == -1) { 901 cg = cg->cg_parent; 902 continue; 903 } 904 steal = TDQ_CPU(cpu); 905 mask &= ~(1 << cpu); 906 tdq_lock_pair(tdq, steal); 907 if (steal->tdq_load < thresh || steal->tdq_transferable == 0) { 908 tdq_unlock_pair(tdq, steal); 909 continue; 910 } 911 /* 912 * If a thread was added while interrupts were disabled don't 913 * steal one here. If we fail to acquire one due to affinity 914 * restrictions loop again with this cpu removed from the 915 * set. 916 */ 917 if (tdq->tdq_load == 0 && tdq_move(steal, tdq) == 0) { 918 tdq_unlock_pair(tdq, steal); 919 continue; 920 } 921 spinlock_exit(); 922 TDQ_UNLOCK(steal); 923 mi_switch(SW_VOL | SWT_IDLE, NULL); 924 thread_unlock(curthread); 925 926 return (0); 927 } 928 spinlock_exit(); 929 return (1); 930 } 931 932 /* 933 * Notify a remote cpu of new work. Sends an IPI if criteria are met. 934 */ 935 static void 936 tdq_notify(struct tdq *tdq, struct thread *td) 937 { 938 int cpri; 939 int pri; 940 int cpu; 941 942 if (tdq->tdq_ipipending) 943 return; 944 cpu = td->td_sched->ts_cpu; 945 pri = td->td_priority; 946 cpri = pcpu_find(cpu)->pc_curthread->td_priority; 947 if (!sched_shouldpreempt(pri, cpri, 1)) 948 return; 949 if (TD_IS_IDLETHREAD(td)) { 950 /* 951 * If the idle thread is still 'running' it's probably 952 * waiting on us to release the tdq spinlock already. No 953 * need to ipi. 954 */ 955 if (tdq->tdq_idlestate == TDQ_RUNNING) 956 return; 957 } 958 tdq->tdq_ipipending = 1; 959 ipi_selected(1 << cpu, IPI_PREEMPT); 960 } 961 962 /* 963 * Steals load from a timeshare queue. Honors the rotating queue head 964 * index. 965 */ 966 static struct thread * 967 runq_steal_from(struct runq *rq, int cpu, u_char start) 968 { 969 struct rqbits *rqb; 970 struct rqhead *rqh; 971 struct thread *td; 972 int first; 973 int bit; 974 int pri; 975 int i; 976 977 rqb = &rq->rq_status; 978 bit = start & (RQB_BPW -1); 979 pri = 0; 980 first = 0; 981 again: 982 for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) { 983 if (rqb->rqb_bits[i] == 0) 984 continue; 985 if (bit != 0) { 986 for (pri = bit; pri < RQB_BPW; pri++) 987 if (rqb->rqb_bits[i] & (1ul << pri)) 988 break; 989 if (pri >= RQB_BPW) 990 continue; 991 } else 992 pri = RQB_FFS(rqb->rqb_bits[i]); 993 pri += (i << RQB_L2BPW); 994 rqh = &rq->rq_queues[pri]; 995 TAILQ_FOREACH(td, rqh, td_runq) { 996 if (first && THREAD_CAN_MIGRATE(td) && 997 THREAD_CAN_SCHED(td, cpu)) 998 return (td); 999 first = 1; 1000 } 1001 } 1002 if (start != 0) { 1003 start = 0; 1004 goto again; 1005 } 1006 1007 return (NULL); 1008 } 1009 1010 /* 1011 * Steals load from a standard linear queue. 1012 */ 1013 static struct thread * 1014 runq_steal(struct runq *rq, int cpu) 1015 { 1016 struct rqhead *rqh; 1017 struct rqbits *rqb; 1018 struct thread *td; 1019 int word; 1020 int bit; 1021 1022 rqb = &rq->rq_status; 1023 for (word = 0; word < RQB_LEN; word++) { 1024 if (rqb->rqb_bits[word] == 0) 1025 continue; 1026 for (bit = 0; bit < RQB_BPW; bit++) { 1027 if ((rqb->rqb_bits[word] & (1ul << bit)) == 0) 1028 continue; 1029 rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)]; 1030 TAILQ_FOREACH(td, rqh, td_runq) 1031 if (THREAD_CAN_MIGRATE(td) && 1032 THREAD_CAN_SCHED(td, cpu)) 1033 return (td); 1034 } 1035 } 1036 return (NULL); 1037 } 1038 1039 /* 1040 * Attempt to steal a thread in priority order from a thread queue. 1041 */ 1042 static struct thread * 1043 tdq_steal(struct tdq *tdq, int cpu) 1044 { 1045 struct thread *td; 1046 1047 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 1048 if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL) 1049 return (td); 1050 if ((td = runq_steal_from(&tdq->tdq_timeshare, 1051 cpu, tdq->tdq_ridx)) != NULL) 1052 return (td); 1053 return (runq_steal(&tdq->tdq_idle, cpu)); 1054 } 1055 1056 /* 1057 * Sets the thread lock and ts_cpu to match the requested cpu. Unlocks the 1058 * current lock and returns with the assigned queue locked. 1059 */ 1060 static inline struct tdq * 1061 sched_setcpu(struct thread *td, int cpu, int flags) 1062 { 1063 1064 struct tdq *tdq; 1065 1066 THREAD_LOCK_ASSERT(td, MA_OWNED); 1067 tdq = TDQ_CPU(cpu); 1068 td->td_sched->ts_cpu = cpu; 1069 /* 1070 * If the lock matches just return the queue. 1071 */ 1072 if (td->td_lock == TDQ_LOCKPTR(tdq)) 1073 return (tdq); 1074 #ifdef notyet 1075 /* 1076 * If the thread isn't running its lockptr is a 1077 * turnstile or a sleepqueue. We can just lock_set without 1078 * blocking. 1079 */ 1080 if (TD_CAN_RUN(td)) { 1081 TDQ_LOCK(tdq); 1082 thread_lock_set(td, TDQ_LOCKPTR(tdq)); 1083 return (tdq); 1084 } 1085 #endif 1086 /* 1087 * The hard case, migration, we need to block the thread first to 1088 * prevent order reversals with other cpus locks. 1089 */ 1090 thread_lock_block(td); 1091 TDQ_LOCK(tdq); 1092 thread_lock_unblock(td, TDQ_LOCKPTR(tdq)); 1093 return (tdq); 1094 } 1095 1096 SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding"); 1097 SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity"); 1098 SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity"); 1099 SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load"); 1100 SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu"); 1101 SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration"); 1102 1103 static int 1104 sched_pickcpu(struct thread *td, int flags) 1105 { 1106 struct cpu_group *cg; 1107 struct td_sched *ts; 1108 struct tdq *tdq; 1109 cpumask_t mask; 1110 int self; 1111 int pri; 1112 int cpu; 1113 1114 self = PCPU_GET(cpuid); 1115 ts = td->td_sched; 1116 if (smp_started == 0) 1117 return (self); 1118 /* 1119 * Don't migrate a running thread from sched_switch(). 1120 */ 1121 if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td)) 1122 return (ts->ts_cpu); 1123 /* 1124 * Prefer to run interrupt threads on the processors that generate 1125 * the interrupt. 1126 */ 1127 if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) && 1128 curthread->td_intr_nesting_level && ts->ts_cpu != self) { 1129 SCHED_STAT_INC(pickcpu_intrbind); 1130 ts->ts_cpu = self; 1131 } 1132 /* 1133 * If the thread can run on the last cpu and the affinity has not 1134 * expired or it is idle run it there. 1135 */ 1136 pri = td->td_priority; 1137 tdq = TDQ_CPU(ts->ts_cpu); 1138 if (THREAD_CAN_SCHED(td, ts->ts_cpu)) { 1139 if (tdq->tdq_lowpri > PRI_MIN_IDLE) { 1140 SCHED_STAT_INC(pickcpu_idle_affinity); 1141 return (ts->ts_cpu); 1142 } 1143 if (SCHED_AFFINITY(ts, CG_SHARE_L2) && tdq->tdq_lowpri > pri) { 1144 SCHED_STAT_INC(pickcpu_affinity); 1145 return (ts->ts_cpu); 1146 } 1147 } 1148 /* 1149 * Search for the highest level in the tree that still has affinity. 1150 */ 1151 cg = NULL; 1152 for (cg = tdq->tdq_cg; cg != NULL; cg = cg->cg_parent) 1153 if (SCHED_AFFINITY(ts, cg->cg_level)) 1154 break; 1155 cpu = -1; 1156 mask = td->td_cpuset->cs_mask.__bits[0]; 1157 if (cg) 1158 cpu = sched_lowest(cg, mask, pri); 1159 if (cpu == -1) 1160 cpu = sched_lowest(cpu_top, mask, -1); 1161 /* 1162 * Compare the lowest loaded cpu to current cpu. 1163 */ 1164 if (THREAD_CAN_SCHED(td, self) && TDQ_CPU(self)->tdq_lowpri > pri && 1165 TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE) { 1166 SCHED_STAT_INC(pickcpu_local); 1167 cpu = self; 1168 } else 1169 SCHED_STAT_INC(pickcpu_lowest); 1170 if (cpu != ts->ts_cpu) 1171 SCHED_STAT_INC(pickcpu_migration); 1172 KASSERT(cpu != -1, ("sched_pickcpu: Failed to find a cpu.")); 1173 return (cpu); 1174 } 1175 #endif 1176 1177 /* 1178 * Pick the highest priority task we have and return it. 1179 */ 1180 static struct thread * 1181 tdq_choose(struct tdq *tdq) 1182 { 1183 struct thread *td; 1184 1185 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 1186 td = runq_choose(&tdq->tdq_realtime); 1187 if (td != NULL) 1188 return (td); 1189 td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx); 1190 if (td != NULL) { 1191 KASSERT(td->td_priority >= PRI_MIN_TIMESHARE, 1192 ("tdq_choose: Invalid priority on timeshare queue %d", 1193 td->td_priority)); 1194 return (td); 1195 } 1196 td = runq_choose(&tdq->tdq_idle); 1197 if (td != NULL) { 1198 KASSERT(td->td_priority >= PRI_MIN_IDLE, 1199 ("tdq_choose: Invalid priority on idle queue %d", 1200 td->td_priority)); 1201 return (td); 1202 } 1203 1204 return (NULL); 1205 } 1206 1207 /* 1208 * Initialize a thread queue. 1209 */ 1210 static void 1211 tdq_setup(struct tdq *tdq) 1212 { 1213 1214 if (bootverbose) 1215 printf("ULE: setup cpu %d\n", TDQ_ID(tdq)); 1216 runq_init(&tdq->tdq_realtime); 1217 runq_init(&tdq->tdq_timeshare); 1218 runq_init(&tdq->tdq_idle); 1219 snprintf(tdq->tdq_name, sizeof(tdq->tdq_name), 1220 "sched lock %d", (int)TDQ_ID(tdq)); 1221 mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock", 1222 MTX_SPIN | MTX_RECURSE); 1223 } 1224 1225 #ifdef SMP 1226 static void 1227 sched_setup_smp(void) 1228 { 1229 struct tdq *tdq; 1230 int i; 1231 1232 cpu_top = smp_topo(); 1233 for (i = 0; i < MAXCPU; i++) { 1234 if (CPU_ABSENT(i)) 1235 continue; 1236 tdq = TDQ_CPU(i); 1237 tdq_setup(tdq); 1238 tdq->tdq_cg = smp_topo_find(cpu_top, i); 1239 if (tdq->tdq_cg == NULL) 1240 panic("Can't find cpu group for %d\n", i); 1241 } 1242 balance_tdq = TDQ_SELF(); 1243 sched_balance(); 1244 } 1245 #endif 1246 1247 /* 1248 * Setup the thread queues and initialize the topology based on MD 1249 * information. 1250 */ 1251 static void 1252 sched_setup(void *dummy) 1253 { 1254 struct tdq *tdq; 1255 1256 tdq = TDQ_SELF(); 1257 #ifdef SMP 1258 sched_setup_smp(); 1259 #else 1260 tdq_setup(tdq); 1261 #endif 1262 /* 1263 * To avoid divide-by-zero, we set realstathz a dummy value 1264 * in case which sched_clock() called before sched_initticks(). 1265 */ 1266 realstathz = hz; 1267 sched_slice = (realstathz/10); /* ~100ms */ 1268 tickincr = 1 << SCHED_TICK_SHIFT; 1269 1270 /* Add thread0's load since it's running. */ 1271 TDQ_LOCK(tdq); 1272 thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF()); 1273 tdq_load_add(tdq, &thread0); 1274 tdq->tdq_lowpri = thread0.td_priority; 1275 TDQ_UNLOCK(tdq); 1276 } 1277 1278 /* 1279 * This routine determines the tickincr after stathz and hz are setup. 1280 */ 1281 /* ARGSUSED */ 1282 static void 1283 sched_initticks(void *dummy) 1284 { 1285 int incr; 1286 1287 realstathz = stathz ? stathz : hz; 1288 sched_slice = (realstathz/10); /* ~100ms */ 1289 1290 /* 1291 * tickincr is shifted out by 10 to avoid rounding errors due to 1292 * hz not being evenly divisible by stathz on all platforms. 1293 */ 1294 incr = (hz << SCHED_TICK_SHIFT) / realstathz; 1295 /* 1296 * This does not work for values of stathz that are more than 1297 * 1 << SCHED_TICK_SHIFT * hz. In practice this does not happen. 1298 */ 1299 if (incr == 0) 1300 incr = 1; 1301 tickincr = incr; 1302 #ifdef SMP 1303 /* 1304 * Set the default balance interval now that we know 1305 * what realstathz is. 1306 */ 1307 balance_interval = realstathz; 1308 /* 1309 * Set steal thresh to log2(mp_ncpu) but no greater than 4. This 1310 * prevents excess thrashing on large machines and excess idle on 1311 * smaller machines. 1312 */ 1313 steal_thresh = min(ffs(mp_ncpus) - 1, 3); 1314 affinity = SCHED_AFFINITY_DEFAULT; 1315 #endif 1316 } 1317 1318 1319 /* 1320 * This is the core of the interactivity algorithm. Determines a score based 1321 * on past behavior. It is the ratio of sleep time to run time scaled to 1322 * a [0, 100] integer. This is the voluntary sleep time of a process, which 1323 * differs from the cpu usage because it does not account for time spent 1324 * waiting on a run-queue. Would be prettier if we had floating point. 1325 */ 1326 static int 1327 sched_interact_score(struct thread *td) 1328 { 1329 struct td_sched *ts; 1330 int div; 1331 1332 ts = td->td_sched; 1333 /* 1334 * The score is only needed if this is likely to be an interactive 1335 * task. Don't go through the expense of computing it if there's 1336 * no chance. 1337 */ 1338 if (sched_interact <= SCHED_INTERACT_HALF && 1339 ts->ts_runtime >= ts->ts_slptime) 1340 return (SCHED_INTERACT_HALF); 1341 1342 if (ts->ts_runtime > ts->ts_slptime) { 1343 div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF); 1344 return (SCHED_INTERACT_HALF + 1345 (SCHED_INTERACT_HALF - (ts->ts_slptime / div))); 1346 } 1347 if (ts->ts_slptime > ts->ts_runtime) { 1348 div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF); 1349 return (ts->ts_runtime / div); 1350 } 1351 /* runtime == slptime */ 1352 if (ts->ts_runtime) 1353 return (SCHED_INTERACT_HALF); 1354 1355 /* 1356 * This can happen if slptime and runtime are 0. 1357 */ 1358 return (0); 1359 1360 } 1361 1362 /* 1363 * Scale the scheduling priority according to the "interactivity" of this 1364 * process. 1365 */ 1366 static void 1367 sched_priority(struct thread *td) 1368 { 1369 int score; 1370 int pri; 1371 1372 if (td->td_pri_class != PRI_TIMESHARE) 1373 return; 1374 /* 1375 * If the score is interactive we place the thread in the realtime 1376 * queue with a priority that is less than kernel and interrupt 1377 * priorities. These threads are not subject to nice restrictions. 1378 * 1379 * Scores greater than this are placed on the normal timeshare queue 1380 * where the priority is partially decided by the most recent cpu 1381 * utilization and the rest is decided by nice value. 1382 * 1383 * The nice value of the process has a linear effect on the calculated 1384 * score. Negative nice values make it easier for a thread to be 1385 * considered interactive. 1386 */ 1387 score = imax(0, sched_interact_score(td) - td->td_proc->p_nice); 1388 if (score < sched_interact) { 1389 pri = PRI_MIN_REALTIME; 1390 pri += ((PRI_MAX_REALTIME - PRI_MIN_REALTIME) / sched_interact) 1391 * score; 1392 KASSERT(pri >= PRI_MIN_REALTIME && pri <= PRI_MAX_REALTIME, 1393 ("sched_priority: invalid interactive priority %d score %d", 1394 pri, score)); 1395 } else { 1396 pri = SCHED_PRI_MIN; 1397 if (td->td_sched->ts_ticks) 1398 pri += SCHED_PRI_TICKS(td->td_sched); 1399 pri += SCHED_PRI_NICE(td->td_proc->p_nice); 1400 KASSERT(pri >= PRI_MIN_TIMESHARE && pri <= PRI_MAX_TIMESHARE, 1401 ("sched_priority: invalid priority %d: nice %d, " 1402 "ticks %d ftick %d ltick %d tick pri %d", 1403 pri, td->td_proc->p_nice, td->td_sched->ts_ticks, 1404 td->td_sched->ts_ftick, td->td_sched->ts_ltick, 1405 SCHED_PRI_TICKS(td->td_sched))); 1406 } 1407 sched_user_prio(td, pri); 1408 1409 return; 1410 } 1411 1412 /* 1413 * This routine enforces a maximum limit on the amount of scheduling history 1414 * kept. It is called after either the slptime or runtime is adjusted. This 1415 * function is ugly due to integer math. 1416 */ 1417 static void 1418 sched_interact_update(struct thread *td) 1419 { 1420 struct td_sched *ts; 1421 u_int sum; 1422 1423 ts = td->td_sched; 1424 sum = ts->ts_runtime + ts->ts_slptime; 1425 if (sum < SCHED_SLP_RUN_MAX) 1426 return; 1427 /* 1428 * This only happens from two places: 1429 * 1) We have added an unusual amount of run time from fork_exit. 1430 * 2) We have added an unusual amount of sleep time from sched_sleep(). 1431 */ 1432 if (sum > SCHED_SLP_RUN_MAX * 2) { 1433 if (ts->ts_runtime > ts->ts_slptime) { 1434 ts->ts_runtime = SCHED_SLP_RUN_MAX; 1435 ts->ts_slptime = 1; 1436 } else { 1437 ts->ts_slptime = SCHED_SLP_RUN_MAX; 1438 ts->ts_runtime = 1; 1439 } 1440 return; 1441 } 1442 /* 1443 * If we have exceeded by more than 1/5th then the algorithm below 1444 * will not bring us back into range. Dividing by two here forces 1445 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX] 1446 */ 1447 if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) { 1448 ts->ts_runtime /= 2; 1449 ts->ts_slptime /= 2; 1450 return; 1451 } 1452 ts->ts_runtime = (ts->ts_runtime / 5) * 4; 1453 ts->ts_slptime = (ts->ts_slptime / 5) * 4; 1454 } 1455 1456 /* 1457 * Scale back the interactivity history when a child thread is created. The 1458 * history is inherited from the parent but the thread may behave totally 1459 * differently. For example, a shell spawning a compiler process. We want 1460 * to learn that the compiler is behaving badly very quickly. 1461 */ 1462 static void 1463 sched_interact_fork(struct thread *td) 1464 { 1465 int ratio; 1466 int sum; 1467 1468 sum = td->td_sched->ts_runtime + td->td_sched->ts_slptime; 1469 if (sum > SCHED_SLP_RUN_FORK) { 1470 ratio = sum / SCHED_SLP_RUN_FORK; 1471 td->td_sched->ts_runtime /= ratio; 1472 td->td_sched->ts_slptime /= ratio; 1473 } 1474 } 1475 1476 /* 1477 * Called from proc0_init() to setup the scheduler fields. 1478 */ 1479 void 1480 schedinit(void) 1481 { 1482 1483 /* 1484 * Set up the scheduler specific parts of proc0. 1485 */ 1486 proc0.p_sched = NULL; /* XXX */ 1487 thread0.td_sched = &td_sched0; 1488 td_sched0.ts_ltick = ticks; 1489 td_sched0.ts_ftick = ticks; 1490 td_sched0.ts_slice = sched_slice; 1491 } 1492 1493 /* 1494 * This is only somewhat accurate since given many processes of the same 1495 * priority they will switch when their slices run out, which will be 1496 * at most sched_slice stathz ticks. 1497 */ 1498 int 1499 sched_rr_interval(void) 1500 { 1501 1502 /* Convert sched_slice to hz */ 1503 return (hz/(realstathz/sched_slice)); 1504 } 1505 1506 /* 1507 * Update the percent cpu tracking information when it is requested or 1508 * the total history exceeds the maximum. We keep a sliding history of 1509 * tick counts that slowly decays. This is less precise than the 4BSD 1510 * mechanism since it happens with less regular and frequent events. 1511 */ 1512 static void 1513 sched_pctcpu_update(struct td_sched *ts) 1514 { 1515 1516 if (ts->ts_ticks == 0) 1517 return; 1518 if (ticks - (hz / 10) < ts->ts_ltick && 1519 SCHED_TICK_TOTAL(ts) < SCHED_TICK_MAX) 1520 return; 1521 /* 1522 * Adjust counters and watermark for pctcpu calc. 1523 */ 1524 if (ts->ts_ltick > ticks - SCHED_TICK_TARG) 1525 ts->ts_ticks = (ts->ts_ticks / (ticks - ts->ts_ftick)) * 1526 SCHED_TICK_TARG; 1527 else 1528 ts->ts_ticks = 0; 1529 ts->ts_ltick = ticks; 1530 ts->ts_ftick = ts->ts_ltick - SCHED_TICK_TARG; 1531 } 1532 1533 /* 1534 * Adjust the priority of a thread. Move it to the appropriate run-queue 1535 * if necessary. This is the back-end for several priority related 1536 * functions. 1537 */ 1538 static void 1539 sched_thread_priority(struct thread *td, u_char prio) 1540 { 1541 struct td_sched *ts; 1542 struct tdq *tdq; 1543 int oldpri; 1544 1545 CTR6(KTR_SCHED, "sched_prio: %p(%s) prio %d newprio %d by %p(%s)", 1546 td, td->td_name, td->td_priority, prio, curthread, 1547 curthread->td_name); 1548 ts = td->td_sched; 1549 THREAD_LOCK_ASSERT(td, MA_OWNED); 1550 if (td->td_priority == prio) 1551 return; 1552 /* 1553 * If the priority has been elevated due to priority 1554 * propagation, we may have to move ourselves to a new 1555 * queue. This could be optimized to not re-add in some 1556 * cases. 1557 */ 1558 if (TD_ON_RUNQ(td) && prio < td->td_priority) { 1559 sched_rem(td); 1560 td->td_priority = prio; 1561 sched_add(td, SRQ_BORROWING); 1562 return; 1563 } 1564 /* 1565 * If the thread is currently running we may have to adjust the lowpri 1566 * information so other cpus are aware of our current priority. 1567 */ 1568 if (TD_IS_RUNNING(td)) { 1569 tdq = TDQ_CPU(ts->ts_cpu); 1570 oldpri = td->td_priority; 1571 td->td_priority = prio; 1572 if (prio < tdq->tdq_lowpri) 1573 tdq->tdq_lowpri = prio; 1574 else if (tdq->tdq_lowpri == oldpri) 1575 tdq_setlowpri(tdq, td); 1576 return; 1577 } 1578 td->td_priority = prio; 1579 } 1580 1581 /* 1582 * Update a thread's priority when it is lent another thread's 1583 * priority. 1584 */ 1585 void 1586 sched_lend_prio(struct thread *td, u_char prio) 1587 { 1588 1589 td->td_flags |= TDF_BORROWING; 1590 sched_thread_priority(td, prio); 1591 } 1592 1593 /* 1594 * Restore a thread's priority when priority propagation is 1595 * over. The prio argument is the minimum priority the thread 1596 * needs to have to satisfy other possible priority lending 1597 * requests. If the thread's regular priority is less 1598 * important than prio, the thread will keep a priority boost 1599 * of prio. 1600 */ 1601 void 1602 sched_unlend_prio(struct thread *td, u_char prio) 1603 { 1604 u_char base_pri; 1605 1606 if (td->td_base_pri >= PRI_MIN_TIMESHARE && 1607 td->td_base_pri <= PRI_MAX_TIMESHARE) 1608 base_pri = td->td_user_pri; 1609 else 1610 base_pri = td->td_base_pri; 1611 if (prio >= base_pri) { 1612 td->td_flags &= ~TDF_BORROWING; 1613 sched_thread_priority(td, base_pri); 1614 } else 1615 sched_lend_prio(td, prio); 1616 } 1617 1618 /* 1619 * Standard entry for setting the priority to an absolute value. 1620 */ 1621 void 1622 sched_prio(struct thread *td, u_char prio) 1623 { 1624 u_char oldprio; 1625 1626 /* First, update the base priority. */ 1627 td->td_base_pri = prio; 1628 1629 /* 1630 * If the thread is borrowing another thread's priority, don't 1631 * ever lower the priority. 1632 */ 1633 if (td->td_flags & TDF_BORROWING && td->td_priority < prio) 1634 return; 1635 1636 /* Change the real priority. */ 1637 oldprio = td->td_priority; 1638 sched_thread_priority(td, prio); 1639 1640 /* 1641 * If the thread is on a turnstile, then let the turnstile update 1642 * its state. 1643 */ 1644 if (TD_ON_LOCK(td) && oldprio != prio) 1645 turnstile_adjust(td, oldprio); 1646 } 1647 1648 /* 1649 * Set the base user priority, does not effect current running priority. 1650 */ 1651 void 1652 sched_user_prio(struct thread *td, u_char prio) 1653 { 1654 u_char oldprio; 1655 1656 td->td_base_user_pri = prio; 1657 if (td->td_flags & TDF_UBORROWING && td->td_user_pri <= prio) 1658 return; 1659 oldprio = td->td_user_pri; 1660 td->td_user_pri = prio; 1661 } 1662 1663 void 1664 sched_lend_user_prio(struct thread *td, u_char prio) 1665 { 1666 u_char oldprio; 1667 1668 THREAD_LOCK_ASSERT(td, MA_OWNED); 1669 td->td_flags |= TDF_UBORROWING; 1670 oldprio = td->td_user_pri; 1671 td->td_user_pri = prio; 1672 } 1673 1674 void 1675 sched_unlend_user_prio(struct thread *td, u_char prio) 1676 { 1677 u_char base_pri; 1678 1679 THREAD_LOCK_ASSERT(td, MA_OWNED); 1680 base_pri = td->td_base_user_pri; 1681 if (prio >= base_pri) { 1682 td->td_flags &= ~TDF_UBORROWING; 1683 sched_user_prio(td, base_pri); 1684 } else { 1685 sched_lend_user_prio(td, prio); 1686 } 1687 } 1688 1689 /* 1690 * Block a thread for switching. Similar to thread_block() but does not 1691 * bump the spin count. 1692 */ 1693 static inline struct mtx * 1694 thread_block_switch(struct thread *td) 1695 { 1696 struct mtx *lock; 1697 1698 THREAD_LOCK_ASSERT(td, MA_OWNED); 1699 lock = td->td_lock; 1700 td->td_lock = &blocked_lock; 1701 mtx_unlock_spin(lock); 1702 1703 return (lock); 1704 } 1705 1706 /* 1707 * Handle migration from sched_switch(). This happens only for 1708 * cpu binding. 1709 */ 1710 static struct mtx * 1711 sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags) 1712 { 1713 struct tdq *tdn; 1714 1715 tdn = TDQ_CPU(td->td_sched->ts_cpu); 1716 #ifdef SMP 1717 tdq_load_rem(tdq, td); 1718 /* 1719 * Do the lock dance required to avoid LOR. We grab an extra 1720 * spinlock nesting to prevent preemption while we're 1721 * not holding either run-queue lock. 1722 */ 1723 spinlock_enter(); 1724 thread_block_switch(td); /* This releases the lock on tdq. */ 1725 TDQ_LOCK(tdn); 1726 tdq_add(tdn, td, flags); 1727 tdq_notify(tdn, td); 1728 /* 1729 * After we unlock tdn the new cpu still can't switch into this 1730 * thread until we've unblocked it in cpu_switch(). The lock 1731 * pointers may match in the case of HTT cores. Don't unlock here 1732 * or we can deadlock when the other CPU runs the IPI handler. 1733 */ 1734 if (TDQ_LOCKPTR(tdn) != TDQ_LOCKPTR(tdq)) { 1735 TDQ_UNLOCK(tdn); 1736 TDQ_LOCK(tdq); 1737 } 1738 spinlock_exit(); 1739 #endif 1740 return (TDQ_LOCKPTR(tdn)); 1741 } 1742 1743 /* 1744 * Release a thread that was blocked with thread_block_switch(). 1745 */ 1746 static inline void 1747 thread_unblock_switch(struct thread *td, struct mtx *mtx) 1748 { 1749 atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock, 1750 (uintptr_t)mtx); 1751 } 1752 1753 /* 1754 * Switch threads. This function has to handle threads coming in while 1755 * blocked for some reason, running, or idle. It also must deal with 1756 * migrating a thread from one queue to another as running threads may 1757 * be assigned elsewhere via binding. 1758 */ 1759 void 1760 sched_switch(struct thread *td, struct thread *newtd, int flags) 1761 { 1762 struct tdq *tdq; 1763 struct td_sched *ts; 1764 struct mtx *mtx; 1765 int srqflag; 1766 int cpuid; 1767 1768 THREAD_LOCK_ASSERT(td, MA_OWNED); 1769 KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument")); 1770 1771 cpuid = PCPU_GET(cpuid); 1772 tdq = TDQ_CPU(cpuid); 1773 ts = td->td_sched; 1774 mtx = td->td_lock; 1775 ts->ts_rltick = ticks; 1776 td->td_lastcpu = td->td_oncpu; 1777 td->td_oncpu = NOCPU; 1778 td->td_flags &= ~TDF_NEEDRESCHED; 1779 td->td_owepreempt = 0; 1780 tdq->tdq_switchcnt++; 1781 /* 1782 * The lock pointer in an idle thread should never change. Reset it 1783 * to CAN_RUN as well. 1784 */ 1785 if (TD_IS_IDLETHREAD(td)) { 1786 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1787 TD_SET_CAN_RUN(td); 1788 } else if (TD_IS_RUNNING(td)) { 1789 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1790 srqflag = (flags & SW_PREEMPT) ? 1791 SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED : 1792 SRQ_OURSELF|SRQ_YIELDING; 1793 if (ts->ts_cpu == cpuid) 1794 tdq_runq_add(tdq, td, srqflag); 1795 else 1796 mtx = sched_switch_migrate(tdq, td, srqflag); 1797 } else { 1798 /* This thread must be going to sleep. */ 1799 TDQ_LOCK(tdq); 1800 mtx = thread_block_switch(td); 1801 tdq_load_rem(tdq, td); 1802 } 1803 /* 1804 * We enter here with the thread blocked and assigned to the 1805 * appropriate cpu run-queue or sleep-queue and with the current 1806 * thread-queue locked. 1807 */ 1808 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED); 1809 newtd = choosethread(); 1810 /* 1811 * Call the MD code to switch contexts if necessary. 1812 */ 1813 if (td != newtd) { 1814 #ifdef HWPMC_HOOKS 1815 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1816 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT); 1817 #endif 1818 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object); 1819 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd; 1820 cpu_switch(td, newtd, mtx); 1821 /* 1822 * We may return from cpu_switch on a different cpu. However, 1823 * we always return with td_lock pointing to the current cpu's 1824 * run queue lock. 1825 */ 1826 cpuid = PCPU_GET(cpuid); 1827 tdq = TDQ_CPU(cpuid); 1828 lock_profile_obtain_lock_success( 1829 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__); 1830 #ifdef HWPMC_HOOKS 1831 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1832 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN); 1833 #endif 1834 } else 1835 thread_unblock_switch(td, mtx); 1836 /* 1837 * Assert that all went well and return. 1838 */ 1839 TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED); 1840 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1841 td->td_oncpu = cpuid; 1842 } 1843 1844 /* 1845 * Adjust thread priorities as a result of a nice request. 1846 */ 1847 void 1848 sched_nice(struct proc *p, int nice) 1849 { 1850 struct thread *td; 1851 1852 PROC_LOCK_ASSERT(p, MA_OWNED); 1853 1854 p->p_nice = nice; 1855 FOREACH_THREAD_IN_PROC(p, td) { 1856 thread_lock(td); 1857 sched_priority(td); 1858 sched_prio(td, td->td_base_user_pri); 1859 thread_unlock(td); 1860 } 1861 } 1862 1863 /* 1864 * Record the sleep time for the interactivity scorer. 1865 */ 1866 void 1867 sched_sleep(struct thread *td, int prio) 1868 { 1869 1870 THREAD_LOCK_ASSERT(td, MA_OWNED); 1871 1872 td->td_slptick = ticks; 1873 if (TD_IS_SUSPENDED(td) || prio <= PSOCK) 1874 td->td_flags |= TDF_CANSWAP; 1875 if (static_boost == 1 && prio) 1876 sched_prio(td, prio); 1877 else if (static_boost && td->td_priority > static_boost) 1878 sched_prio(td, static_boost); 1879 } 1880 1881 /* 1882 * Schedule a thread to resume execution and record how long it voluntarily 1883 * slept. We also update the pctcpu, interactivity, and priority. 1884 */ 1885 void 1886 sched_wakeup(struct thread *td) 1887 { 1888 struct td_sched *ts; 1889 int slptick; 1890 1891 THREAD_LOCK_ASSERT(td, MA_OWNED); 1892 ts = td->td_sched; 1893 td->td_flags &= ~TDF_CANSWAP; 1894 /* 1895 * If we slept for more than a tick update our interactivity and 1896 * priority. 1897 */ 1898 slptick = td->td_slptick; 1899 td->td_slptick = 0; 1900 if (slptick && slptick != ticks) { 1901 u_int hzticks; 1902 1903 hzticks = (ticks - slptick) << SCHED_TICK_SHIFT; 1904 ts->ts_slptime += hzticks; 1905 sched_interact_update(td); 1906 sched_pctcpu_update(ts); 1907 } 1908 /* Reset the slice value after we sleep. */ 1909 ts->ts_slice = sched_slice; 1910 sched_add(td, SRQ_BORING); 1911 } 1912 1913 /* 1914 * Penalize the parent for creating a new child and initialize the child's 1915 * priority. 1916 */ 1917 void 1918 sched_fork(struct thread *td, struct thread *child) 1919 { 1920 THREAD_LOCK_ASSERT(td, MA_OWNED); 1921 sched_fork_thread(td, child); 1922 /* 1923 * Penalize the parent and child for forking. 1924 */ 1925 sched_interact_fork(child); 1926 sched_priority(child); 1927 td->td_sched->ts_runtime += tickincr; 1928 sched_interact_update(td); 1929 sched_priority(td); 1930 } 1931 1932 /* 1933 * Fork a new thread, may be within the same process. 1934 */ 1935 void 1936 sched_fork_thread(struct thread *td, struct thread *child) 1937 { 1938 struct td_sched *ts; 1939 struct td_sched *ts2; 1940 1941 THREAD_LOCK_ASSERT(td, MA_OWNED); 1942 /* 1943 * Initialize child. 1944 */ 1945 ts = td->td_sched; 1946 ts2 = child->td_sched; 1947 child->td_lock = TDQ_LOCKPTR(TDQ_SELF()); 1948 child->td_cpuset = cpuset_ref(td->td_cpuset); 1949 ts2->ts_cpu = ts->ts_cpu; 1950 ts2->ts_flags = 0; 1951 /* 1952 * Grab our parents cpu estimation information and priority. 1953 */ 1954 ts2->ts_ticks = ts->ts_ticks; 1955 ts2->ts_ltick = ts->ts_ltick; 1956 ts2->ts_ftick = ts->ts_ftick; 1957 child->td_user_pri = td->td_user_pri; 1958 child->td_base_user_pri = td->td_base_user_pri; 1959 /* 1960 * And update interactivity score. 1961 */ 1962 ts2->ts_slptime = ts->ts_slptime; 1963 ts2->ts_runtime = ts->ts_runtime; 1964 ts2->ts_slice = 1; /* Attempt to quickly learn interactivity. */ 1965 } 1966 1967 /* 1968 * Adjust the priority class of a thread. 1969 */ 1970 void 1971 sched_class(struct thread *td, int class) 1972 { 1973 1974 THREAD_LOCK_ASSERT(td, MA_OWNED); 1975 if (td->td_pri_class == class) 1976 return; 1977 td->td_pri_class = class; 1978 } 1979 1980 /* 1981 * Return some of the child's priority and interactivity to the parent. 1982 */ 1983 void 1984 sched_exit(struct proc *p, struct thread *child) 1985 { 1986 struct thread *td; 1987 1988 CTR3(KTR_SCHED, "sched_exit: %p(%s) prio %d", 1989 child, child->td_name, child->td_priority); 1990 1991 PROC_LOCK_ASSERT(p, MA_OWNED); 1992 td = FIRST_THREAD_IN_PROC(p); 1993 sched_exit_thread(td, child); 1994 } 1995 1996 /* 1997 * Penalize another thread for the time spent on this one. This helps to 1998 * worsen the priority and interactivity of processes which schedule batch 1999 * jobs such as make. This has little effect on the make process itself but 2000 * causes new processes spawned by it to receive worse scores immediately. 2001 */ 2002 void 2003 sched_exit_thread(struct thread *td, struct thread *child) 2004 { 2005 2006 CTR3(KTR_SCHED, "sched_exit_thread: %p(%s) prio %d", 2007 child, child->td_name, child->td_priority); 2008 2009 /* 2010 * Give the child's runtime to the parent without returning the 2011 * sleep time as a penalty to the parent. This causes shells that 2012 * launch expensive things to mark their children as expensive. 2013 */ 2014 thread_lock(td); 2015 td->td_sched->ts_runtime += child->td_sched->ts_runtime; 2016 sched_interact_update(td); 2017 sched_priority(td); 2018 thread_unlock(td); 2019 } 2020 2021 void 2022 sched_preempt(struct thread *td) 2023 { 2024 struct tdq *tdq; 2025 2026 thread_lock(td); 2027 tdq = TDQ_SELF(); 2028 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2029 tdq->tdq_ipipending = 0; 2030 if (td->td_priority > tdq->tdq_lowpri) { 2031 int flags; 2032 2033 flags = SW_INVOL | SW_PREEMPT; 2034 if (td->td_critnest > 1) 2035 td->td_owepreempt = 1; 2036 else if (TD_IS_IDLETHREAD(td)) 2037 mi_switch(flags | SWT_REMOTEWAKEIDLE, NULL); 2038 else 2039 mi_switch(flags | SWT_REMOTEPREEMPT, NULL); 2040 } 2041 thread_unlock(td); 2042 } 2043 2044 /* 2045 * Fix priorities on return to user-space. Priorities may be elevated due 2046 * to static priorities in msleep() or similar. 2047 */ 2048 void 2049 sched_userret(struct thread *td) 2050 { 2051 /* 2052 * XXX we cheat slightly on the locking here to avoid locking in 2053 * the usual case. Setting td_priority here is essentially an 2054 * incomplete workaround for not setting it properly elsewhere. 2055 * Now that some interrupt handlers are threads, not setting it 2056 * properly elsewhere can clobber it in the window between setting 2057 * it here and returning to user mode, so don't waste time setting 2058 * it perfectly here. 2059 */ 2060 KASSERT((td->td_flags & TDF_BORROWING) == 0, 2061 ("thread with borrowed priority returning to userland")); 2062 if (td->td_priority != td->td_user_pri) { 2063 thread_lock(td); 2064 td->td_priority = td->td_user_pri; 2065 td->td_base_pri = td->td_user_pri; 2066 tdq_setlowpri(TDQ_SELF(), td); 2067 thread_unlock(td); 2068 } 2069 } 2070 2071 /* 2072 * Handle a stathz tick. This is really only relevant for timeshare 2073 * threads. 2074 */ 2075 void 2076 sched_clock(struct thread *td) 2077 { 2078 struct tdq *tdq; 2079 struct td_sched *ts; 2080 2081 THREAD_LOCK_ASSERT(td, MA_OWNED); 2082 tdq = TDQ_SELF(); 2083 #ifdef SMP 2084 /* 2085 * We run the long term load balancer infrequently on the first cpu. 2086 */ 2087 if (balance_tdq == tdq) { 2088 if (balance_ticks && --balance_ticks == 0) 2089 sched_balance(); 2090 } 2091 #endif 2092 /* 2093 * Save the old switch count so we have a record of the last ticks 2094 * activity. Initialize the new switch count based on our load. 2095 * If there is some activity seed it to reflect that. 2096 */ 2097 tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt; 2098 if (tdq->tdq_load) 2099 tdq->tdq_switchcnt = 2; 2100 else 2101 tdq->tdq_switchcnt = 0; 2102 /* 2103 * Advance the insert index once for each tick to ensure that all 2104 * threads get a chance to run. 2105 */ 2106 if (tdq->tdq_idx == tdq->tdq_ridx) { 2107 tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS; 2108 if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx])) 2109 tdq->tdq_ridx = tdq->tdq_idx; 2110 } 2111 ts = td->td_sched; 2112 if (td->td_pri_class & PRI_FIFO_BIT) 2113 return; 2114 if (td->td_pri_class == PRI_TIMESHARE) { 2115 /* 2116 * We used a tick; charge it to the thread so 2117 * that we can compute our interactivity. 2118 */ 2119 td->td_sched->ts_runtime += tickincr; 2120 sched_interact_update(td); 2121 sched_priority(td); 2122 } 2123 /* 2124 * We used up one time slice. 2125 */ 2126 if (--ts->ts_slice > 0) 2127 return; 2128 /* 2129 * We're out of time, force a requeue at userret(). 2130 */ 2131 ts->ts_slice = sched_slice; 2132 td->td_flags |= TDF_NEEDRESCHED; 2133 } 2134 2135 /* 2136 * Called once per hz tick. Used for cpu utilization information. This 2137 * is easier than trying to scale based on stathz. 2138 */ 2139 void 2140 sched_tick(void) 2141 { 2142 struct td_sched *ts; 2143 2144 ts = curthread->td_sched; 2145 /* Adjust ticks for pctcpu */ 2146 ts->ts_ticks += 1 << SCHED_TICK_SHIFT; 2147 ts->ts_ltick = ticks; 2148 /* 2149 * Update if we've exceeded our desired tick threshhold by over one 2150 * second. 2151 */ 2152 if (ts->ts_ftick + SCHED_TICK_MAX < ts->ts_ltick) 2153 sched_pctcpu_update(ts); 2154 } 2155 2156 /* 2157 * Return whether the current CPU has runnable tasks. Used for in-kernel 2158 * cooperative idle threads. 2159 */ 2160 int 2161 sched_runnable(void) 2162 { 2163 struct tdq *tdq; 2164 int load; 2165 2166 load = 1; 2167 2168 tdq = TDQ_SELF(); 2169 if ((curthread->td_flags & TDF_IDLETD) != 0) { 2170 if (tdq->tdq_load > 0) 2171 goto out; 2172 } else 2173 if (tdq->tdq_load - 1 > 0) 2174 goto out; 2175 load = 0; 2176 out: 2177 return (load); 2178 } 2179 2180 /* 2181 * Choose the highest priority thread to run. The thread is removed from 2182 * the run-queue while running however the load remains. For SMP we set 2183 * the tdq in the global idle bitmask if it idles here. 2184 */ 2185 struct thread * 2186 sched_choose(void) 2187 { 2188 struct thread *td; 2189 struct tdq *tdq; 2190 2191 tdq = TDQ_SELF(); 2192 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2193 td = tdq_choose(tdq); 2194 if (td) { 2195 td->td_sched->ts_ltick = ticks; 2196 tdq_runq_rem(tdq, td); 2197 tdq->tdq_lowpri = td->td_priority; 2198 return (td); 2199 } 2200 tdq->tdq_lowpri = PRI_MAX_IDLE; 2201 return (PCPU_GET(idlethread)); 2202 } 2203 2204 /* 2205 * Set owepreempt if necessary. Preemption never happens directly in ULE, 2206 * we always request it once we exit a critical section. 2207 */ 2208 static inline void 2209 sched_setpreempt(struct thread *td) 2210 { 2211 struct thread *ctd; 2212 int cpri; 2213 int pri; 2214 2215 THREAD_LOCK_ASSERT(curthread, MA_OWNED); 2216 2217 ctd = curthread; 2218 pri = td->td_priority; 2219 cpri = ctd->td_priority; 2220 if (pri < cpri) 2221 ctd->td_flags |= TDF_NEEDRESCHED; 2222 if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd)) 2223 return; 2224 if (!sched_shouldpreempt(pri, cpri, 0)) 2225 return; 2226 ctd->td_owepreempt = 1; 2227 } 2228 2229 /* 2230 * Add a thread to a thread queue. Select the appropriate runq and add the 2231 * thread to it. This is the internal function called when the tdq is 2232 * predetermined. 2233 */ 2234 void 2235 tdq_add(struct tdq *tdq, struct thread *td, int flags) 2236 { 2237 2238 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2239 KASSERT((td->td_inhibitors == 0), 2240 ("sched_add: trying to run inhibited thread")); 2241 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)), 2242 ("sched_add: bad thread state")); 2243 KASSERT(td->td_flags & TDF_INMEM, 2244 ("sched_add: thread swapped out")); 2245 2246 if (td->td_priority < tdq->tdq_lowpri) 2247 tdq->tdq_lowpri = td->td_priority; 2248 tdq_runq_add(tdq, td, flags); 2249 tdq_load_add(tdq, td); 2250 } 2251 2252 /* 2253 * Select the target thread queue and add a thread to it. Request 2254 * preemption or IPI a remote processor if required. 2255 */ 2256 void 2257 sched_add(struct thread *td, int flags) 2258 { 2259 struct tdq *tdq; 2260 #ifdef SMP 2261 int cpu; 2262 #endif 2263 CTR5(KTR_SCHED, "sched_add: %p(%s) prio %d by %p(%s)", 2264 td, td->td_name, td->td_priority, curthread, 2265 curthread->td_name); 2266 THREAD_LOCK_ASSERT(td, MA_OWNED); 2267 /* 2268 * Recalculate the priority before we select the target cpu or 2269 * run-queue. 2270 */ 2271 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) 2272 sched_priority(td); 2273 #ifdef SMP 2274 /* 2275 * Pick the destination cpu and if it isn't ours transfer to the 2276 * target cpu. 2277 */ 2278 cpu = sched_pickcpu(td, flags); 2279 tdq = sched_setcpu(td, cpu, flags); 2280 tdq_add(tdq, td, flags); 2281 if (cpu != PCPU_GET(cpuid)) { 2282 tdq_notify(tdq, td); 2283 return; 2284 } 2285 #else 2286 tdq = TDQ_SELF(); 2287 TDQ_LOCK(tdq); 2288 /* 2289 * Now that the thread is moving to the run-queue, set the lock 2290 * to the scheduler's lock. 2291 */ 2292 thread_lock_set(td, TDQ_LOCKPTR(tdq)); 2293 tdq_add(tdq, td, flags); 2294 #endif 2295 if (!(flags & SRQ_YIELDING)) 2296 sched_setpreempt(td); 2297 } 2298 2299 /* 2300 * Remove a thread from a run-queue without running it. This is used 2301 * when we're stealing a thread from a remote queue. Otherwise all threads 2302 * exit by calling sched_exit_thread() and sched_throw() themselves. 2303 */ 2304 void 2305 sched_rem(struct thread *td) 2306 { 2307 struct tdq *tdq; 2308 2309 CTR5(KTR_SCHED, "sched_rem: %p(%s) prio %d by %p(%s)", 2310 td, td->td_name, td->td_priority, curthread, 2311 curthread->td_name); 2312 tdq = TDQ_CPU(td->td_sched->ts_cpu); 2313 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2314 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2315 KASSERT(TD_ON_RUNQ(td), 2316 ("sched_rem: thread not on run queue")); 2317 tdq_runq_rem(tdq, td); 2318 tdq_load_rem(tdq, td); 2319 TD_SET_CAN_RUN(td); 2320 if (td->td_priority == tdq->tdq_lowpri) 2321 tdq_setlowpri(tdq, NULL); 2322 } 2323 2324 /* 2325 * Fetch cpu utilization information. Updates on demand. 2326 */ 2327 fixpt_t 2328 sched_pctcpu(struct thread *td) 2329 { 2330 fixpt_t pctcpu; 2331 struct td_sched *ts; 2332 2333 pctcpu = 0; 2334 ts = td->td_sched; 2335 if (ts == NULL) 2336 return (0); 2337 2338 thread_lock(td); 2339 if (ts->ts_ticks) { 2340 int rtick; 2341 2342 sched_pctcpu_update(ts); 2343 /* How many rtick per second ? */ 2344 rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz); 2345 pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT; 2346 } 2347 thread_unlock(td); 2348 2349 return (pctcpu); 2350 } 2351 2352 /* 2353 * Enforce affinity settings for a thread. Called after adjustments to 2354 * cpumask. 2355 */ 2356 void 2357 sched_affinity(struct thread *td) 2358 { 2359 #ifdef SMP 2360 struct td_sched *ts; 2361 int cpu; 2362 2363 THREAD_LOCK_ASSERT(td, MA_OWNED); 2364 ts = td->td_sched; 2365 if (THREAD_CAN_SCHED(td, ts->ts_cpu)) 2366 return; 2367 if (!TD_IS_RUNNING(td)) 2368 return; 2369 td->td_flags |= TDF_NEEDRESCHED; 2370 if (!THREAD_CAN_MIGRATE(td)) 2371 return; 2372 /* 2373 * Assign the new cpu and force a switch before returning to 2374 * userspace. If the target thread is not running locally send 2375 * an ipi to force the issue. 2376 */ 2377 cpu = ts->ts_cpu; 2378 ts->ts_cpu = sched_pickcpu(td, 0); 2379 if (cpu != PCPU_GET(cpuid)) 2380 ipi_selected(1 << cpu, IPI_PREEMPT); 2381 #endif 2382 } 2383 2384 /* 2385 * Bind a thread to a target cpu. 2386 */ 2387 void 2388 sched_bind(struct thread *td, int cpu) 2389 { 2390 struct td_sched *ts; 2391 2392 THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED); 2393 ts = td->td_sched; 2394 if (ts->ts_flags & TSF_BOUND) 2395 sched_unbind(td); 2396 ts->ts_flags |= TSF_BOUND; 2397 sched_pin(); 2398 if (PCPU_GET(cpuid) == cpu) 2399 return; 2400 ts->ts_cpu = cpu; 2401 /* When we return from mi_switch we'll be on the correct cpu. */ 2402 mi_switch(SW_VOL, NULL); 2403 } 2404 2405 /* 2406 * Release a bound thread. 2407 */ 2408 void 2409 sched_unbind(struct thread *td) 2410 { 2411 struct td_sched *ts; 2412 2413 THREAD_LOCK_ASSERT(td, MA_OWNED); 2414 ts = td->td_sched; 2415 if ((ts->ts_flags & TSF_BOUND) == 0) 2416 return; 2417 ts->ts_flags &= ~TSF_BOUND; 2418 sched_unpin(); 2419 } 2420 2421 int 2422 sched_is_bound(struct thread *td) 2423 { 2424 THREAD_LOCK_ASSERT(td, MA_OWNED); 2425 return (td->td_sched->ts_flags & TSF_BOUND); 2426 } 2427 2428 /* 2429 * Basic yield call. 2430 */ 2431 void 2432 sched_relinquish(struct thread *td) 2433 { 2434 thread_lock(td); 2435 mi_switch(SW_VOL | SWT_RELINQUISH, NULL); 2436 thread_unlock(td); 2437 } 2438 2439 /* 2440 * Return the total system load. 2441 */ 2442 int 2443 sched_load(void) 2444 { 2445 #ifdef SMP 2446 int total; 2447 int i; 2448 2449 total = 0; 2450 for (i = 0; i <= mp_maxid; i++) 2451 total += TDQ_CPU(i)->tdq_sysload; 2452 return (total); 2453 #else 2454 return (TDQ_SELF()->tdq_sysload); 2455 #endif 2456 } 2457 2458 int 2459 sched_sizeof_proc(void) 2460 { 2461 return (sizeof(struct proc)); 2462 } 2463 2464 int 2465 sched_sizeof_thread(void) 2466 { 2467 return (sizeof(struct thread) + sizeof(struct td_sched)); 2468 } 2469 2470 /* 2471 * The actual idle process. 2472 */ 2473 void 2474 sched_idletd(void *dummy) 2475 { 2476 struct thread *td; 2477 struct tdq *tdq; 2478 int switchcnt; 2479 int i; 2480 2481 td = curthread; 2482 tdq = TDQ_SELF(); 2483 mtx_assert(&Giant, MA_NOTOWNED); 2484 /* ULE relies on preemption for idle interruption. */ 2485 for (;;) { 2486 tdq->tdq_idlestate = TDQ_RUNNING; 2487 #ifdef SMP 2488 if (tdq_idled(tdq) == 0) 2489 continue; 2490 #endif 2491 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt; 2492 /* 2493 * If we're switching very frequently, spin while checking 2494 * for load rather than entering a low power state that 2495 * requires an IPI. 2496 */ 2497 if (switchcnt > sched_idlespinthresh) { 2498 for (i = 0; i < sched_idlespins; i++) { 2499 if (tdq->tdq_load) 2500 break; 2501 cpu_spinwait(); 2502 } 2503 } 2504 /* 2505 * We must set our state to IDLE before checking 2506 * tdq_load for the last time to avoid a race with 2507 * tdq_notify(). 2508 */ 2509 if (tdq->tdq_load == 0) { 2510 tdq->tdq_idlestate = TDQ_IDLE; 2511 if (tdq->tdq_load == 0) 2512 cpu_idle(); 2513 } 2514 if (tdq->tdq_load) { 2515 thread_lock(td); 2516 mi_switch(SW_VOL | SWT_IDLE, NULL); 2517 thread_unlock(td); 2518 } 2519 } 2520 } 2521 2522 /* 2523 * A CPU is entering for the first time or a thread is exiting. 2524 */ 2525 void 2526 sched_throw(struct thread *td) 2527 { 2528 struct thread *newtd; 2529 struct tdq *tdq; 2530 2531 tdq = TDQ_SELF(); 2532 if (td == NULL) { 2533 /* Correct spinlock nesting and acquire the correct lock. */ 2534 TDQ_LOCK(tdq); 2535 spinlock_exit(); 2536 } else { 2537 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2538 tdq_load_rem(tdq, td); 2539 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object); 2540 } 2541 KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count")); 2542 newtd = choosethread(); 2543 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd; 2544 PCPU_SET(switchtime, cpu_ticks()); 2545 PCPU_SET(switchticks, ticks); 2546 cpu_throw(td, newtd); /* doesn't return */ 2547 } 2548 2549 /* 2550 * This is called from fork_exit(). Just acquire the correct locks and 2551 * let fork do the rest of the work. 2552 */ 2553 void 2554 sched_fork_exit(struct thread *td) 2555 { 2556 struct td_sched *ts; 2557 struct tdq *tdq; 2558 int cpuid; 2559 2560 /* 2561 * Finish setting up thread glue so that it begins execution in a 2562 * non-nested critical section with the scheduler lock held. 2563 */ 2564 cpuid = PCPU_GET(cpuid); 2565 tdq = TDQ_CPU(cpuid); 2566 ts = td->td_sched; 2567 if (TD_IS_IDLETHREAD(td)) 2568 td->td_lock = TDQ_LOCKPTR(tdq); 2569 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2570 td->td_oncpu = cpuid; 2571 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED); 2572 lock_profile_obtain_lock_success( 2573 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__); 2574 } 2575 2576 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler"); 2577 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0, 2578 "Scheduler name"); 2579 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0, 2580 "Slice size for timeshare threads"); 2581 SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0, 2582 "Interactivity score threshold"); 2583 SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW, &preempt_thresh, 2584 0,"Min priority for preemption, lower priorities have greater precedence"); 2585 SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost, 2586 0,"Controls whether static kernel priorities are assigned to sleeping threads."); 2587 SYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins, 2588 0,"Number of times idle will spin waiting for new work."); 2589 SYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW, &sched_idlespinthresh, 2590 0,"Threshold before we will permit idle spinning."); 2591 #ifdef SMP 2592 SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0, 2593 "Number of hz ticks to keep thread affinity for"); 2594 SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0, 2595 "Enables the long-term load balancer"); 2596 SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW, 2597 &balance_interval, 0, 2598 "Average frequency in stathz ticks to run the long-term balancer"); 2599 SYSCTL_INT(_kern_sched, OID_AUTO, steal_htt, CTLFLAG_RW, &steal_htt, 0, 2600 "Steals work from another hyper-threaded core on idle"); 2601 SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0, 2602 "Attempts to steal work from other cores before idling"); 2603 SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0, 2604 "Minimum load on remote cpu before we'll steal"); 2605 #endif 2606 2607 /* ps compat. All cpu percentages from ULE are weighted. */ 2608 static int ccpu = 0; 2609 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 2610