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(__i386__) && !defined(__amd64__) && !defined(__powerpc__) && !defined(__arm__) 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 TAILQ_ENTRY(td_sched) ts_procq; /* Run queue. */ 87 struct thread *ts_thread; /* Active associated thread. */ 88 struct runq *ts_runq; /* Run-queue we're queued on. */ 89 short ts_flags; /* TSF_* flags. */ 90 u_char ts_rqindex; /* Run queue index. */ 91 u_char ts_cpu; /* CPU that we have affinity for. */ 92 int ts_slice; /* Ticks of slice remaining. */ 93 u_int ts_slptime; /* Number of ticks we vol. slept */ 94 u_int ts_runtime; /* Number of ticks we were running */ 95 /* The following variables are only used for pctcpu calculation */ 96 int ts_ltick; /* Last tick that we were running on */ 97 int ts_ftick; /* First tick that we were running on */ 98 int ts_ticks; /* Tick count */ 99 int ts_rltick; /* Real last tick, for affinity. */ 100 }; 101 /* flags kept in ts_flags */ 102 #define TSF_BOUND 0x0001 /* Thread can not migrate. */ 103 #define TSF_XFERABLE 0x0002 /* Thread was added as transferable. */ 104 105 static struct td_sched td_sched0; 106 107 #define THREAD_CAN_MIGRATE(td) ((td)->td_pinned == 0) 108 #define THREAD_CAN_SCHED(td, cpu) \ 109 CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask) 110 111 /* 112 * Cpu percentage computation macros and defines. 113 * 114 * SCHED_TICK_SECS: Number of seconds to average the cpu usage across. 115 * SCHED_TICK_TARG: Number of hz ticks to average the cpu usage across. 116 * SCHED_TICK_MAX: Maximum number of ticks before scaling back. 117 * SCHED_TICK_SHIFT: Shift factor to avoid rounding away results. 118 * SCHED_TICK_HZ: Compute the number of hz ticks for a given ticks count. 119 * SCHED_TICK_TOTAL: Gives the amount of time we've been recording ticks. 120 */ 121 #define SCHED_TICK_SECS 10 122 #define SCHED_TICK_TARG (hz * SCHED_TICK_SECS) 123 #define SCHED_TICK_MAX (SCHED_TICK_TARG + hz) 124 #define SCHED_TICK_SHIFT 10 125 #define SCHED_TICK_HZ(ts) ((ts)->ts_ticks >> SCHED_TICK_SHIFT) 126 #define SCHED_TICK_TOTAL(ts) (max((ts)->ts_ltick - (ts)->ts_ftick, hz)) 127 128 /* 129 * These macros determine priorities for non-interactive threads. They are 130 * assigned a priority based on their recent cpu utilization as expressed 131 * by the ratio of ticks to the tick total. NHALF priorities at the start 132 * and end of the MIN to MAX timeshare range are only reachable with negative 133 * or positive nice respectively. 134 * 135 * PRI_RANGE: Priority range for utilization dependent priorities. 136 * PRI_NRESV: Number of nice values. 137 * PRI_TICKS: Compute a priority in PRI_RANGE from the ticks count and total. 138 * PRI_NICE: Determines the part of the priority inherited from nice. 139 */ 140 #define SCHED_PRI_NRESV (PRIO_MAX - PRIO_MIN) 141 #define SCHED_PRI_NHALF (SCHED_PRI_NRESV / 2) 142 #define SCHED_PRI_MIN (PRI_MIN_TIMESHARE + SCHED_PRI_NHALF) 143 #define SCHED_PRI_MAX (PRI_MAX_TIMESHARE - SCHED_PRI_NHALF) 144 #define SCHED_PRI_RANGE (SCHED_PRI_MAX - SCHED_PRI_MIN) 145 #define SCHED_PRI_TICKS(ts) \ 146 (SCHED_TICK_HZ((ts)) / \ 147 (roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE)) 148 #define SCHED_PRI_NICE(nice) (nice) 149 150 /* 151 * These determine the interactivity of a process. Interactivity differs from 152 * cpu utilization in that it expresses the voluntary time slept vs time ran 153 * while cpu utilization includes all time not running. This more accurately 154 * models the intent of the thread. 155 * 156 * SLP_RUN_MAX: Maximum amount of sleep time + run time we'll accumulate 157 * before throttling back. 158 * SLP_RUN_FORK: Maximum slp+run time to inherit at fork time. 159 * INTERACT_MAX: Maximum interactivity value. Smaller is better. 160 * INTERACT_THRESH: Threshhold for placement on the current runq. 161 */ 162 #define SCHED_SLP_RUN_MAX ((hz * 5) << SCHED_TICK_SHIFT) 163 #define SCHED_SLP_RUN_FORK ((hz / 2) << SCHED_TICK_SHIFT) 164 #define SCHED_INTERACT_MAX (100) 165 #define SCHED_INTERACT_HALF (SCHED_INTERACT_MAX / 2) 166 #define SCHED_INTERACT_THRESH (30) 167 168 /* 169 * tickincr: Converts a stathz tick into a hz domain scaled by 170 * the shift factor. Without the shift the error rate 171 * due to rounding would be unacceptably high. 172 * realstathz: stathz is sometimes 0 and run off of hz. 173 * sched_slice: Runtime of each thread before rescheduling. 174 * preempt_thresh: Priority threshold for preemption and remote IPIs. 175 */ 176 static int sched_interact = SCHED_INTERACT_THRESH; 177 static int realstathz; 178 static int tickincr; 179 static int sched_slice; 180 #ifdef PREEMPTION 181 #ifdef FULL_PREEMPTION 182 static int preempt_thresh = PRI_MAX_IDLE; 183 #else 184 static int preempt_thresh = PRI_MIN_KERN; 185 #endif 186 #else 187 static int preempt_thresh = 0; 188 #endif 189 static int lowpri_userret = 1; 190 191 /* 192 * tdq - per processor runqs and statistics. All fields are protected by the 193 * tdq_lock. The load and lowpri may be accessed without to avoid excess 194 * locking in sched_pickcpu(); 195 */ 196 struct tdq { 197 struct cpu_group *tdq_cg; /* Pointer to cpu topology. */ 198 struct mtx tdq_lock; /* run queue lock. */ 199 struct runq tdq_realtime; /* real-time run queue. */ 200 struct runq tdq_timeshare; /* timeshare run queue. */ 201 struct runq tdq_idle; /* Queue of IDLE threads. */ 202 int tdq_load; /* Aggregate load. */ 203 int tdq_sysload; /* For loadavg, !ITHD load. */ 204 u_char tdq_idx; /* Current insert index. */ 205 u_char tdq_ridx; /* Current removal index. */ 206 u_char tdq_lowpri; /* Lowest priority thread. */ 207 int tdq_transferable; /* Transferable thread count. */ 208 char tdq_name[sizeof("sched lock") + 6]; 209 } __aligned(64); 210 211 212 #ifdef SMP 213 struct cpu_group *cpu_top; 214 215 #define SCHED_AFFINITY_DEFAULT (max(1, hz / 1000)) 216 #define SCHED_AFFINITY(ts, t) ((ts)->ts_rltick > ticks - ((t) * affinity)) 217 218 /* 219 * Run-time tunables. 220 */ 221 static int rebalance = 1; 222 static int balance_interval = 128; /* Default set in sched_initticks(). */ 223 static int pick_pri = 1; 224 static int affinity; 225 static int tryself = 1; 226 static int oldtryself = 0; 227 static int steal_htt = 1; 228 static int steal_idle = 1; 229 static int steal_thresh = 2; 230 231 /* 232 * One thread queue per processor. 233 */ 234 static struct tdq tdq_cpu[MAXCPU]; 235 static struct tdq *balance_tdq; 236 static int balance_ticks; 237 238 #define TDQ_SELF() (&tdq_cpu[PCPU_GET(cpuid)]) 239 #define TDQ_CPU(x) (&tdq_cpu[(x)]) 240 #define TDQ_ID(x) ((int)((x) - tdq_cpu)) 241 #else /* !SMP */ 242 static struct tdq tdq_cpu; 243 244 #define TDQ_ID(x) (0) 245 #define TDQ_SELF() (&tdq_cpu) 246 #define TDQ_CPU(x) (&tdq_cpu) 247 #endif 248 249 #define TDQ_LOCK_ASSERT(t, type) mtx_assert(TDQ_LOCKPTR((t)), (type)) 250 #define TDQ_LOCK(t) mtx_lock_spin(TDQ_LOCKPTR((t))) 251 #define TDQ_LOCK_FLAGS(t, f) mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f)) 252 #define TDQ_UNLOCK(t) mtx_unlock_spin(TDQ_LOCKPTR((t))) 253 #define TDQ_LOCKPTR(t) (&(t)->tdq_lock) 254 255 static void sched_priority(struct thread *); 256 static void sched_thread_priority(struct thread *, u_char); 257 static int sched_interact_score(struct thread *); 258 static void sched_interact_update(struct thread *); 259 static void sched_interact_fork(struct thread *); 260 static void sched_pctcpu_update(struct td_sched *); 261 262 /* Operations on per processor queues */ 263 static struct td_sched * tdq_choose(struct tdq *); 264 static void tdq_setup(struct tdq *); 265 static void tdq_load_add(struct tdq *, struct td_sched *); 266 static void tdq_load_rem(struct tdq *, struct td_sched *); 267 static __inline void tdq_runq_add(struct tdq *, struct td_sched *, int); 268 static __inline void tdq_runq_rem(struct tdq *, struct td_sched *); 269 void tdq_print(int cpu); 270 static void runq_print(struct runq *rq); 271 static void tdq_add(struct tdq *, struct thread *, int); 272 #ifdef SMP 273 static int tdq_move(struct tdq *, struct tdq *); 274 static int tdq_idled(struct tdq *); 275 static void tdq_notify(struct td_sched *); 276 static struct td_sched *tdq_steal(struct tdq *, int); 277 static struct td_sched *runq_steal(struct runq *, int); 278 static int sched_pickcpu(struct td_sched *, int); 279 static void sched_balance(void); 280 static int sched_balance_pair(struct tdq *, struct tdq *); 281 static inline struct tdq *sched_setcpu(struct td_sched *, int, int); 282 static inline struct mtx *thread_block_switch(struct thread *); 283 static inline void thread_unblock_switch(struct thread *, struct mtx *); 284 static struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int); 285 #endif 286 287 static void sched_setup(void *dummy); 288 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL) 289 290 static void sched_initticks(void *dummy); 291 SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks, NULL) 292 293 /* 294 * Print the threads waiting on a run-queue. 295 */ 296 static void 297 runq_print(struct runq *rq) 298 { 299 struct rqhead *rqh; 300 struct td_sched *ts; 301 int pri; 302 int j; 303 int i; 304 305 for (i = 0; i < RQB_LEN; i++) { 306 printf("\t\trunq bits %d 0x%zx\n", 307 i, rq->rq_status.rqb_bits[i]); 308 for (j = 0; j < RQB_BPW; j++) 309 if (rq->rq_status.rqb_bits[i] & (1ul << j)) { 310 pri = j + (i << RQB_L2BPW); 311 rqh = &rq->rq_queues[pri]; 312 TAILQ_FOREACH(ts, rqh, ts_procq) { 313 printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n", 314 ts->ts_thread, ts->ts_thread->td_name, ts->ts_thread->td_priority, ts->ts_rqindex, pri); 315 } 316 } 317 } 318 } 319 320 /* 321 * Print the status of a per-cpu thread queue. Should be a ddb show cmd. 322 */ 323 void 324 tdq_print(int cpu) 325 { 326 struct tdq *tdq; 327 328 tdq = TDQ_CPU(cpu); 329 330 printf("tdq %d:\n", TDQ_ID(tdq)); 331 printf("\tlock %p\n", TDQ_LOCKPTR(tdq)); 332 printf("\tLock name: %s\n", tdq->tdq_name); 333 printf("\tload: %d\n", tdq->tdq_load); 334 printf("\ttimeshare idx: %d\n", tdq->tdq_idx); 335 printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx); 336 printf("\trealtime runq:\n"); 337 runq_print(&tdq->tdq_realtime); 338 printf("\ttimeshare runq:\n"); 339 runq_print(&tdq->tdq_timeshare); 340 printf("\tidle runq:\n"); 341 runq_print(&tdq->tdq_idle); 342 printf("\tload transferable: %d\n", tdq->tdq_transferable); 343 printf("\tlowest priority: %d\n", tdq->tdq_lowpri); 344 } 345 346 #define TS_RQ_PPQ (((PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE) + 1) / RQ_NQS) 347 /* 348 * Add a thread to the actual run-queue. Keeps transferable counts up to 349 * date with what is actually on the run-queue. Selects the correct 350 * queue position for timeshare threads. 351 */ 352 static __inline void 353 tdq_runq_add(struct tdq *tdq, struct td_sched *ts, int flags) 354 { 355 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 356 THREAD_LOCK_ASSERT(ts->ts_thread, MA_OWNED); 357 if (THREAD_CAN_MIGRATE(ts->ts_thread)) { 358 tdq->tdq_transferable++; 359 ts->ts_flags |= TSF_XFERABLE; 360 } 361 if (ts->ts_runq == &tdq->tdq_timeshare) { 362 u_char pri; 363 364 pri = ts->ts_thread->td_priority; 365 KASSERT(pri <= PRI_MAX_TIMESHARE && pri >= PRI_MIN_TIMESHARE, 366 ("Invalid priority %d on timeshare runq", pri)); 367 /* 368 * This queue contains only priorities between MIN and MAX 369 * realtime. Use the whole queue to represent these values. 370 */ 371 if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) { 372 pri = (pri - PRI_MIN_TIMESHARE) / TS_RQ_PPQ; 373 pri = (pri + tdq->tdq_idx) % RQ_NQS; 374 /* 375 * This effectively shortens the queue by one so we 376 * can have a one slot difference between idx and 377 * ridx while we wait for threads to drain. 378 */ 379 if (tdq->tdq_ridx != tdq->tdq_idx && 380 pri == tdq->tdq_ridx) 381 pri = (unsigned char)(pri - 1) % RQ_NQS; 382 } else 383 pri = tdq->tdq_ridx; 384 runq_add_pri(ts->ts_runq, ts, pri, flags); 385 } else 386 runq_add(ts->ts_runq, ts, flags); 387 } 388 389 /* 390 * Remove a thread from a run-queue. This typically happens when a thread 391 * is selected to run. Running threads are not on the queue and the 392 * transferable count does not reflect them. 393 */ 394 static __inline void 395 tdq_runq_rem(struct tdq *tdq, struct td_sched *ts) 396 { 397 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 398 KASSERT(ts->ts_runq != NULL, 399 ("tdq_runq_remove: thread %p null ts_runq", ts->ts_thread)); 400 if (ts->ts_flags & TSF_XFERABLE) { 401 tdq->tdq_transferable--; 402 ts->ts_flags &= ~TSF_XFERABLE; 403 } 404 if (ts->ts_runq == &tdq->tdq_timeshare) { 405 if (tdq->tdq_idx != tdq->tdq_ridx) 406 runq_remove_idx(ts->ts_runq, ts, &tdq->tdq_ridx); 407 else 408 runq_remove_idx(ts->ts_runq, ts, NULL); 409 /* 410 * For timeshare threads we update the priority here so 411 * the priority reflects the time we've been sleeping. 412 */ 413 ts->ts_ltick = ticks; 414 sched_pctcpu_update(ts); 415 sched_priority(ts->ts_thread); 416 } else 417 runq_remove(ts->ts_runq, ts); 418 } 419 420 /* 421 * Load is maintained for all threads RUNNING and ON_RUNQ. Add the load 422 * for this thread to the referenced thread queue. 423 */ 424 static void 425 tdq_load_add(struct tdq *tdq, struct td_sched *ts) 426 { 427 int class; 428 429 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 430 THREAD_LOCK_ASSERT(ts->ts_thread, MA_OWNED); 431 class = PRI_BASE(ts->ts_thread->td_pri_class); 432 tdq->tdq_load++; 433 CTR2(KTR_SCHED, "cpu %d load: %d", TDQ_ID(tdq), tdq->tdq_load); 434 if (class != PRI_ITHD && 435 (ts->ts_thread->td_proc->p_flag & P_NOLOAD) == 0) 436 tdq->tdq_sysload++; 437 } 438 439 /* 440 * Remove the load from a thread that is transitioning to a sleep state or 441 * exiting. 442 */ 443 static void 444 tdq_load_rem(struct tdq *tdq, struct td_sched *ts) 445 { 446 int class; 447 448 THREAD_LOCK_ASSERT(ts->ts_thread, MA_OWNED); 449 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 450 class = PRI_BASE(ts->ts_thread->td_pri_class); 451 if (class != PRI_ITHD && 452 (ts->ts_thread->td_proc->p_flag & P_NOLOAD) == 0) 453 tdq->tdq_sysload--; 454 KASSERT(tdq->tdq_load != 0, 455 ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq))); 456 tdq->tdq_load--; 457 CTR1(KTR_SCHED, "load: %d", tdq->tdq_load); 458 ts->ts_runq = NULL; 459 } 460 461 /* 462 * Set lowpri to its exact value by searching the run-queue and 463 * evaluating curthread. curthread may be passed as an optimization. 464 */ 465 static void 466 tdq_setlowpri(struct tdq *tdq, struct thread *ctd) 467 { 468 struct td_sched *ts; 469 struct thread *td; 470 471 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 472 if (ctd == NULL) 473 ctd = pcpu_find(TDQ_ID(tdq))->pc_curthread; 474 ts = tdq_choose(tdq); 475 if (ts) 476 td = ts->ts_thread; 477 if (ts == NULL || td->td_priority > ctd->td_priority) 478 tdq->tdq_lowpri = ctd->td_priority; 479 else 480 tdq->tdq_lowpri = td->td_priority; 481 } 482 483 #ifdef SMP 484 struct cpu_search { 485 cpumask_t cs_mask; /* Mask of valid cpus. */ 486 u_int cs_load; 487 u_int cs_cpu; 488 int cs_limit; /* Min priority for low min load for high. */ 489 }; 490 491 #define CPU_SEARCH_LOWEST 0x1 492 #define CPU_SEARCH_HIGHEST 0x2 493 #define CPU_SEARCH_BOTH (CPU_SEARCH_LOWEST|CPU_SEARCH_HIGHEST) 494 495 #define CPUMASK_FOREACH(cpu, mask) \ 496 for ((cpu) = 0; (cpu) < sizeof((mask)) * 8; (cpu)++) \ 497 if ((mask) & 1 << (cpu)) 498 499 __inline int cpu_search(struct cpu_group *cg, struct cpu_search *low, 500 struct cpu_search *high, const int match); 501 int cpu_search_lowest(struct cpu_group *cg, struct cpu_search *low); 502 int cpu_search_highest(struct cpu_group *cg, struct cpu_search *high); 503 int cpu_search_both(struct cpu_group *cg, struct cpu_search *low, 504 struct cpu_search *high); 505 506 /* 507 * This routine compares according to the match argument and should be 508 * reduced in actual instantiations via constant propagation and dead code 509 * elimination. 510 */ 511 static __inline int 512 cpu_compare(int cpu, struct cpu_search *low, struct cpu_search *high, 513 const int match) 514 { 515 struct tdq *tdq; 516 517 tdq = TDQ_CPU(cpu); 518 if (match & CPU_SEARCH_LOWEST) 519 if (low->cs_mask & (1 << cpu) && 520 tdq->tdq_load < low->cs_load && 521 tdq->tdq_lowpri > low->cs_limit) { 522 low->cs_cpu = cpu; 523 low->cs_load = tdq->tdq_load; 524 } 525 if (match & CPU_SEARCH_HIGHEST) 526 if (high->cs_mask & (1 << cpu) && 527 tdq->tdq_load >= high->cs_limit && 528 tdq->tdq_load > high->cs_load && 529 tdq->tdq_transferable) { 530 high->cs_cpu = cpu; 531 high->cs_load = tdq->tdq_load; 532 } 533 return (tdq->tdq_load); 534 } 535 536 /* 537 * Search the tree of cpu_groups for the lowest or highest loaded cpu 538 * according to the match argument. This routine actually compares the 539 * load on all paths through the tree and finds the least loaded cpu on 540 * the least loaded path, which may differ from the least loaded cpu in 541 * the system. This balances work among caches and busses. 542 * 543 * This inline is instantiated in three forms below using constants for the 544 * match argument. It is reduced to the minimum set for each case. It is 545 * also recursive to the depth of the tree. 546 */ 547 static inline int 548 cpu_search(struct cpu_group *cg, struct cpu_search *low, 549 struct cpu_search *high, const int match) 550 { 551 int total; 552 553 total = 0; 554 if (cg->cg_children) { 555 struct cpu_search lgroup; 556 struct cpu_search hgroup; 557 struct cpu_group *child; 558 u_int lload; 559 int hload; 560 int load; 561 int i; 562 563 lload = -1; 564 hload = -1; 565 for (i = 0; i < cg->cg_children; i++) { 566 child = &cg->cg_child[i]; 567 if (match & CPU_SEARCH_LOWEST) { 568 lgroup = *low; 569 lgroup.cs_load = -1; 570 } 571 if (match & CPU_SEARCH_HIGHEST) { 572 hgroup = *high; 573 lgroup.cs_load = 0; 574 } 575 switch (match) { 576 case CPU_SEARCH_LOWEST: 577 load = cpu_search_lowest(child, &lgroup); 578 break; 579 case CPU_SEARCH_HIGHEST: 580 load = cpu_search_highest(child, &hgroup); 581 break; 582 case CPU_SEARCH_BOTH: 583 load = cpu_search_both(child, &lgroup, &hgroup); 584 break; 585 } 586 total += load; 587 if (match & CPU_SEARCH_LOWEST) 588 if (load < lload || low->cs_cpu == -1) { 589 *low = lgroup; 590 lload = load; 591 } 592 if (match & CPU_SEARCH_HIGHEST) 593 if (load > hload || high->cs_cpu == -1) { 594 hload = load; 595 *high = hgroup; 596 } 597 } 598 } else { 599 int cpu; 600 601 CPUMASK_FOREACH(cpu, cg->cg_mask) 602 total += cpu_compare(cpu, low, high, match); 603 } 604 return (total); 605 } 606 607 /* 608 * cpu_search instantiations must pass constants to maintain the inline 609 * optimization. 610 */ 611 int 612 cpu_search_lowest(struct cpu_group *cg, struct cpu_search *low) 613 { 614 return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST); 615 } 616 617 int 618 cpu_search_highest(struct cpu_group *cg, struct cpu_search *high) 619 { 620 return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST); 621 } 622 623 int 624 cpu_search_both(struct cpu_group *cg, struct cpu_search *low, 625 struct cpu_search *high) 626 { 627 return cpu_search(cg, low, high, CPU_SEARCH_BOTH); 628 } 629 630 /* 631 * Find the cpu with the least load via the least loaded path that has a 632 * lowpri greater than pri pri. A pri of -1 indicates any priority is 633 * acceptable. 634 */ 635 static inline int 636 sched_lowest(struct cpu_group *cg, cpumask_t mask, int pri) 637 { 638 struct cpu_search low; 639 640 low.cs_cpu = -1; 641 low.cs_load = -1; 642 low.cs_mask = mask; 643 low.cs_limit = pri; 644 cpu_search_lowest(cg, &low); 645 return low.cs_cpu; 646 } 647 648 /* 649 * Find the cpu with the highest load via the highest loaded path. 650 */ 651 static inline int 652 sched_highest(struct cpu_group *cg, cpumask_t mask, int minload) 653 { 654 struct cpu_search high; 655 656 high.cs_cpu = -1; 657 high.cs_load = 0; 658 high.cs_mask = mask; 659 high.cs_limit = minload; 660 cpu_search_highest(cg, &high); 661 return high.cs_cpu; 662 } 663 664 /* 665 * Simultaneously find the highest and lowest loaded cpu reachable via 666 * cg. 667 */ 668 static inline void 669 sched_both(struct cpu_group *cg, cpumask_t mask, int *lowcpu, int *highcpu) 670 { 671 struct cpu_search high; 672 struct cpu_search low; 673 674 low.cs_cpu = -1; 675 low.cs_limit = -1; 676 low.cs_load = -1; 677 low.cs_mask = mask; 678 high.cs_load = 0; 679 high.cs_cpu = -1; 680 high.cs_limit = -1; 681 high.cs_mask = mask; 682 cpu_search_both(cg, &low, &high); 683 *lowcpu = low.cs_cpu; 684 *highcpu = high.cs_cpu; 685 return; 686 } 687 688 static void 689 sched_balance_group(struct cpu_group *cg) 690 { 691 cpumask_t mask; 692 int high; 693 int low; 694 int i; 695 696 mask = -1; 697 for (;;) { 698 sched_both(cg, mask, &low, &high); 699 if (low == high || low == -1 || high == -1) 700 break; 701 if (sched_balance_pair(TDQ_CPU(high), TDQ_CPU(low))) 702 break; 703 /* 704 * If we failed to move any threads determine which cpu 705 * to kick out of the set and try again. 706 */ 707 if (TDQ_CPU(high)->tdq_transferable == 0) 708 mask &= ~(1 << high); 709 else 710 mask &= ~(1 << low); 711 } 712 713 for (i = 0; i < cg->cg_children; i++) 714 sched_balance_group(&cg->cg_child[i]); 715 } 716 717 static void 718 sched_balance() 719 { 720 struct tdq *tdq; 721 722 /* 723 * Select a random time between .5 * balance_interval and 724 * 1.5 * balance_interval. 725 */ 726 balance_ticks = max(balance_interval / 2, 1); 727 balance_ticks += random() % balance_interval; 728 if (smp_started == 0 || rebalance == 0) 729 return; 730 tdq = TDQ_SELF(); 731 TDQ_UNLOCK(tdq); 732 sched_balance_group(cpu_top); 733 TDQ_LOCK(tdq); 734 } 735 736 /* 737 * Lock two thread queues using their address to maintain lock order. 738 */ 739 static void 740 tdq_lock_pair(struct tdq *one, struct tdq *two) 741 { 742 if (one < two) { 743 TDQ_LOCK(one); 744 TDQ_LOCK_FLAGS(two, MTX_DUPOK); 745 } else { 746 TDQ_LOCK(two); 747 TDQ_LOCK_FLAGS(one, MTX_DUPOK); 748 } 749 } 750 751 /* 752 * Unlock two thread queues. Order is not important here. 753 */ 754 static void 755 tdq_unlock_pair(struct tdq *one, struct tdq *two) 756 { 757 TDQ_UNLOCK(one); 758 TDQ_UNLOCK(two); 759 } 760 761 /* 762 * Transfer load between two imbalanced thread queues. 763 */ 764 static int 765 sched_balance_pair(struct tdq *high, struct tdq *low) 766 { 767 int transferable; 768 int high_load; 769 int low_load; 770 int moved; 771 int move; 772 int diff; 773 int i; 774 775 tdq_lock_pair(high, low); 776 transferable = high->tdq_transferable; 777 high_load = high->tdq_load; 778 low_load = low->tdq_load; 779 moved = 0; 780 /* 781 * Determine what the imbalance is and then adjust that to how many 782 * threads we actually have to give up (transferable). 783 */ 784 if (transferable != 0) { 785 diff = high_load - low_load; 786 move = diff / 2; 787 if (diff & 0x1) 788 move++; 789 move = min(move, transferable); 790 for (i = 0; i < move; i++) 791 moved += tdq_move(high, low); 792 /* 793 * IPI the target cpu to force it to reschedule with the new 794 * workload. 795 */ 796 ipi_selected(1 << TDQ_ID(low), IPI_PREEMPT); 797 } 798 tdq_unlock_pair(high, low); 799 return (moved); 800 } 801 802 /* 803 * Move a thread from one thread queue to another. 804 */ 805 static int 806 tdq_move(struct tdq *from, struct tdq *to) 807 { 808 struct td_sched *ts; 809 struct thread *td; 810 struct tdq *tdq; 811 int cpu; 812 813 TDQ_LOCK_ASSERT(from, MA_OWNED); 814 TDQ_LOCK_ASSERT(to, MA_OWNED); 815 816 tdq = from; 817 cpu = TDQ_ID(to); 818 ts = tdq_steal(tdq, cpu); 819 if (ts == NULL) 820 return (0); 821 td = ts->ts_thread; 822 /* 823 * Although the run queue is locked the thread may be blocked. Lock 824 * it to clear this and acquire the run-queue lock. 825 */ 826 thread_lock(td); 827 /* Drop recursive lock on from acquired via thread_lock(). */ 828 TDQ_UNLOCK(from); 829 sched_rem(td); 830 ts->ts_cpu = cpu; 831 td->td_lock = TDQ_LOCKPTR(to); 832 tdq_add(to, td, SRQ_YIELDING); 833 return (1); 834 } 835 836 /* 837 * This tdq has idled. Try to steal a thread from another cpu and switch 838 * to it. 839 */ 840 static int 841 tdq_idled(struct tdq *tdq) 842 { 843 struct cpu_group *cg; 844 struct tdq *steal; 845 cpumask_t mask; 846 int thresh; 847 int cpu; 848 849 if (smp_started == 0 || steal_idle == 0) 850 return (1); 851 mask = -1; 852 mask &= ~PCPU_GET(cpumask); 853 /* We don't want to be preempted while we're iterating. */ 854 spinlock_enter(); 855 for (cg = tdq->tdq_cg; cg != NULL; ) { 856 if ((cg->cg_flags & (CG_FLAG_HTT | CG_FLAG_THREAD)) == 0) 857 thresh = steal_thresh; 858 else 859 thresh = 1; 860 cpu = sched_highest(cg, mask, thresh); 861 if (cpu == -1) { 862 cg = cg->cg_parent; 863 continue; 864 } 865 steal = TDQ_CPU(cpu); 866 mask &= ~(1 << cpu); 867 tdq_lock_pair(tdq, steal); 868 if (steal->tdq_load < thresh || steal->tdq_transferable == 0) { 869 tdq_unlock_pair(tdq, steal); 870 continue; 871 } 872 /* 873 * If a thread was added while interrupts were disabled don't 874 * steal one here. If we fail to acquire one due to affinity 875 * restrictions loop again with this cpu removed from the 876 * set. 877 */ 878 if (tdq->tdq_load == 0 && tdq_move(steal, tdq) == 0) { 879 tdq_unlock_pair(tdq, steal); 880 continue; 881 } 882 spinlock_exit(); 883 TDQ_UNLOCK(steal); 884 mi_switch(SW_VOL, NULL); 885 thread_unlock(curthread); 886 887 return (0); 888 } 889 spinlock_exit(); 890 return (1); 891 } 892 893 /* 894 * Notify a remote cpu of new work. Sends an IPI if criteria are met. 895 */ 896 static void 897 tdq_notify(struct td_sched *ts) 898 { 899 struct thread *ctd; 900 struct pcpu *pcpu; 901 int cpri; 902 int pri; 903 int cpu; 904 905 cpu = ts->ts_cpu; 906 pri = ts->ts_thread->td_priority; 907 pcpu = pcpu_find(cpu); 908 ctd = pcpu->pc_curthread; 909 cpri = ctd->td_priority; 910 911 /* 912 * If our priority is not better than the current priority there is 913 * nothing to do. 914 */ 915 if (pri > cpri) 916 return; 917 /* 918 * Always IPI idle. 919 */ 920 if (cpri > PRI_MIN_IDLE) 921 goto sendipi; 922 /* 923 * If we're realtime or better and there is timeshare or worse running 924 * send an IPI. 925 */ 926 if (pri < PRI_MAX_REALTIME && cpri > PRI_MAX_REALTIME) 927 goto sendipi; 928 /* 929 * Otherwise only IPI if we exceed the threshold. 930 */ 931 if (pri > preempt_thresh) 932 return; 933 sendipi: 934 ctd->td_flags |= TDF_NEEDRESCHED; 935 ipi_selected(1 << cpu, IPI_PREEMPT); 936 } 937 938 /* 939 * Steals load from a timeshare queue. Honors the rotating queue head 940 * index. 941 */ 942 static struct td_sched * 943 runq_steal_from(struct runq *rq, int cpu, u_char start) 944 { 945 struct td_sched *ts; 946 struct rqbits *rqb; 947 struct rqhead *rqh; 948 int first; 949 int bit; 950 int pri; 951 int i; 952 953 rqb = &rq->rq_status; 954 bit = start & (RQB_BPW -1); 955 pri = 0; 956 first = 0; 957 again: 958 for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) { 959 if (rqb->rqb_bits[i] == 0) 960 continue; 961 if (bit != 0) { 962 for (pri = bit; pri < RQB_BPW; pri++) 963 if (rqb->rqb_bits[i] & (1ul << pri)) 964 break; 965 if (pri >= RQB_BPW) 966 continue; 967 } else 968 pri = RQB_FFS(rqb->rqb_bits[i]); 969 pri += (i << RQB_L2BPW); 970 rqh = &rq->rq_queues[pri]; 971 TAILQ_FOREACH(ts, rqh, ts_procq) { 972 if (first && THREAD_CAN_MIGRATE(ts->ts_thread) && 973 THREAD_CAN_SCHED(ts->ts_thread, cpu)) 974 return (ts); 975 first = 1; 976 } 977 } 978 if (start != 0) { 979 start = 0; 980 goto again; 981 } 982 983 return (NULL); 984 } 985 986 /* 987 * Steals load from a standard linear queue. 988 */ 989 static struct td_sched * 990 runq_steal(struct runq *rq, int cpu) 991 { 992 struct rqhead *rqh; 993 struct rqbits *rqb; 994 struct td_sched *ts; 995 int word; 996 int bit; 997 998 rqb = &rq->rq_status; 999 for (word = 0; word < RQB_LEN; word++) { 1000 if (rqb->rqb_bits[word] == 0) 1001 continue; 1002 for (bit = 0; bit < RQB_BPW; bit++) { 1003 if ((rqb->rqb_bits[word] & (1ul << bit)) == 0) 1004 continue; 1005 rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)]; 1006 TAILQ_FOREACH(ts, rqh, ts_procq) 1007 if (THREAD_CAN_MIGRATE(ts->ts_thread) && 1008 THREAD_CAN_SCHED(ts->ts_thread, cpu)) 1009 return (ts); 1010 } 1011 } 1012 return (NULL); 1013 } 1014 1015 /* 1016 * Attempt to steal a thread in priority order from a thread queue. 1017 */ 1018 static struct td_sched * 1019 tdq_steal(struct tdq *tdq, int cpu) 1020 { 1021 struct td_sched *ts; 1022 1023 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 1024 if ((ts = runq_steal(&tdq->tdq_realtime, cpu)) != NULL) 1025 return (ts); 1026 if ((ts = runq_steal_from(&tdq->tdq_timeshare, cpu, tdq->tdq_ridx)) 1027 != NULL) 1028 return (ts); 1029 return (runq_steal(&tdq->tdq_idle, cpu)); 1030 } 1031 1032 /* 1033 * Sets the thread lock and ts_cpu to match the requested cpu. Unlocks the 1034 * current lock and returns with the assigned queue locked. 1035 */ 1036 static inline struct tdq * 1037 sched_setcpu(struct td_sched *ts, int cpu, int flags) 1038 { 1039 struct thread *td; 1040 struct tdq *tdq; 1041 1042 THREAD_LOCK_ASSERT(ts->ts_thread, MA_OWNED); 1043 1044 tdq = TDQ_CPU(cpu); 1045 td = ts->ts_thread; 1046 ts->ts_cpu = cpu; 1047 1048 /* If the lock matches just return the queue. */ 1049 if (td->td_lock == TDQ_LOCKPTR(tdq)) 1050 return (tdq); 1051 #ifdef notyet 1052 /* 1053 * If the thread isn't running its lockptr is a 1054 * turnstile or a sleepqueue. We can just lock_set without 1055 * blocking. 1056 */ 1057 if (TD_CAN_RUN(td)) { 1058 TDQ_LOCK(tdq); 1059 thread_lock_set(td, TDQ_LOCKPTR(tdq)); 1060 return (tdq); 1061 } 1062 #endif 1063 /* 1064 * The hard case, migration, we need to block the thread first to 1065 * prevent order reversals with other cpus locks. 1066 */ 1067 thread_lock_block(td); 1068 TDQ_LOCK(tdq); 1069 thread_lock_unblock(td, TDQ_LOCKPTR(tdq)); 1070 return (tdq); 1071 } 1072 1073 static int 1074 sched_pickcpu(struct td_sched *ts, int flags) 1075 { 1076 struct cpu_group *cg; 1077 struct thread *td; 1078 struct tdq *tdq; 1079 cpumask_t mask; 1080 int self; 1081 int pri; 1082 int cpu; 1083 1084 self = PCPU_GET(cpuid); 1085 td = ts->ts_thread; 1086 if (smp_started == 0) 1087 return (self); 1088 /* 1089 * Don't migrate a running thread from sched_switch(). 1090 */ 1091 if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td)) 1092 return (ts->ts_cpu); 1093 /* 1094 * Prefer to run interrupt threads on the processors that generate 1095 * the interrupt. 1096 */ 1097 if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) && 1098 curthread->td_intr_nesting_level) 1099 ts->ts_cpu = self; 1100 /* 1101 * If the thread can run on the last cpu and the affinity has not 1102 * expired or it is idle run it there. 1103 */ 1104 pri = td->td_priority; 1105 tdq = TDQ_CPU(ts->ts_cpu); 1106 if (THREAD_CAN_SCHED(td, ts->ts_cpu)) { 1107 if (tdq->tdq_lowpri > PRI_MIN_IDLE) 1108 return (ts->ts_cpu); 1109 if (SCHED_AFFINITY(ts, CG_SHARE_L2) && tdq->tdq_lowpri > pri) 1110 return (ts->ts_cpu); 1111 } 1112 /* 1113 * Search for the highest level in the tree that still has affinity. 1114 */ 1115 cg = NULL; 1116 for (cg = tdq->tdq_cg; cg != NULL; cg = cg->cg_parent) 1117 if (SCHED_AFFINITY(ts, cg->cg_level)) 1118 break; 1119 cpu = -1; 1120 mask = td->td_cpuset->cs_mask.__bits[0]; 1121 if (cg) 1122 cpu = sched_lowest(cg, mask, pri); 1123 if (cpu == -1) 1124 cpu = sched_lowest(cpu_top, mask, -1); 1125 /* 1126 * Compare the lowest loaded cpu to current cpu. 1127 */ 1128 if (THREAD_CAN_SCHED(td, self) && 1129 TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE) { 1130 if (tryself && TDQ_CPU(self)->tdq_lowpri > pri) 1131 cpu = self; 1132 else if (oldtryself && curthread->td_priority > pri) 1133 cpu = self; 1134 } 1135 if (cpu == -1) { 1136 panic("cpu == -1, mask 0x%X cpu top %p", mask, cpu_top); 1137 } 1138 return (cpu); 1139 } 1140 #endif 1141 1142 /* 1143 * Pick the highest priority task we have and return it. 1144 */ 1145 static struct td_sched * 1146 tdq_choose(struct tdq *tdq) 1147 { 1148 struct td_sched *ts; 1149 1150 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 1151 ts = runq_choose(&tdq->tdq_realtime); 1152 if (ts != NULL) 1153 return (ts); 1154 ts = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx); 1155 if (ts != NULL) { 1156 KASSERT(ts->ts_thread->td_priority >= PRI_MIN_TIMESHARE, 1157 ("tdq_choose: Invalid priority on timeshare queue %d", 1158 ts->ts_thread->td_priority)); 1159 return (ts); 1160 } 1161 1162 ts = runq_choose(&tdq->tdq_idle); 1163 if (ts != NULL) { 1164 KASSERT(ts->ts_thread->td_priority >= PRI_MIN_IDLE, 1165 ("tdq_choose: Invalid priority on idle queue %d", 1166 ts->ts_thread->td_priority)); 1167 return (ts); 1168 } 1169 1170 return (NULL); 1171 } 1172 1173 /* 1174 * Initialize a thread queue. 1175 */ 1176 static void 1177 tdq_setup(struct tdq *tdq) 1178 { 1179 1180 if (bootverbose) 1181 printf("ULE: setup cpu %d\n", TDQ_ID(tdq)); 1182 runq_init(&tdq->tdq_realtime); 1183 runq_init(&tdq->tdq_timeshare); 1184 runq_init(&tdq->tdq_idle); 1185 snprintf(tdq->tdq_name, sizeof(tdq->tdq_name), 1186 "sched lock %d", (int)TDQ_ID(tdq)); 1187 mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock", 1188 MTX_SPIN | MTX_RECURSE); 1189 } 1190 1191 #ifdef SMP 1192 static void 1193 sched_setup_smp(void) 1194 { 1195 struct tdq *tdq; 1196 int i; 1197 1198 cpu_top = smp_topo(); 1199 for (i = 0; i < MAXCPU; i++) { 1200 if (CPU_ABSENT(i)) 1201 continue; 1202 tdq = TDQ_CPU(i); 1203 tdq_setup(tdq); 1204 tdq->tdq_cg = smp_topo_find(cpu_top, i); 1205 if (tdq->tdq_cg == NULL) 1206 panic("Can't find cpu group for %d\n", i); 1207 } 1208 balance_tdq = TDQ_SELF(); 1209 sched_balance(); 1210 } 1211 #endif 1212 1213 /* 1214 * Setup the thread queues and initialize the topology based on MD 1215 * information. 1216 */ 1217 static void 1218 sched_setup(void *dummy) 1219 { 1220 struct tdq *tdq; 1221 1222 tdq = TDQ_SELF(); 1223 #ifdef SMP 1224 sched_setup_smp(); 1225 #else 1226 tdq_setup(tdq); 1227 #endif 1228 /* 1229 * To avoid divide-by-zero, we set realstathz a dummy value 1230 * in case which sched_clock() called before sched_initticks(). 1231 */ 1232 realstathz = hz; 1233 sched_slice = (realstathz/10); /* ~100ms */ 1234 tickincr = 1 << SCHED_TICK_SHIFT; 1235 1236 /* Add thread0's load since it's running. */ 1237 TDQ_LOCK(tdq); 1238 thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF()); 1239 tdq_load_add(tdq, &td_sched0); 1240 tdq->tdq_lowpri = thread0.td_priority; 1241 TDQ_UNLOCK(tdq); 1242 } 1243 1244 /* 1245 * This routine determines the tickincr after stathz and hz are setup. 1246 */ 1247 /* ARGSUSED */ 1248 static void 1249 sched_initticks(void *dummy) 1250 { 1251 int incr; 1252 1253 realstathz = stathz ? stathz : hz; 1254 sched_slice = (realstathz/10); /* ~100ms */ 1255 1256 /* 1257 * tickincr is shifted out by 10 to avoid rounding errors due to 1258 * hz not being evenly divisible by stathz on all platforms. 1259 */ 1260 incr = (hz << SCHED_TICK_SHIFT) / realstathz; 1261 /* 1262 * This does not work for values of stathz that are more than 1263 * 1 << SCHED_TICK_SHIFT * hz. In practice this does not happen. 1264 */ 1265 if (incr == 0) 1266 incr = 1; 1267 tickincr = incr; 1268 #ifdef SMP 1269 /* 1270 * Set the default balance interval now that we know 1271 * what realstathz is. 1272 */ 1273 balance_interval = realstathz; 1274 /* 1275 * Set steal thresh to log2(mp_ncpu) but no greater than 4. This 1276 * prevents excess thrashing on large machines and excess idle on 1277 * smaller machines. 1278 */ 1279 steal_thresh = min(ffs(mp_ncpus) - 1, 3); 1280 affinity = SCHED_AFFINITY_DEFAULT; 1281 #endif 1282 } 1283 1284 1285 /* 1286 * This is the core of the interactivity algorithm. Determines a score based 1287 * on past behavior. It is the ratio of sleep time to run time scaled to 1288 * a [0, 100] integer. This is the voluntary sleep time of a process, which 1289 * differs from the cpu usage because it does not account for time spent 1290 * waiting on a run-queue. Would be prettier if we had floating point. 1291 */ 1292 static int 1293 sched_interact_score(struct thread *td) 1294 { 1295 struct td_sched *ts; 1296 int div; 1297 1298 ts = td->td_sched; 1299 /* 1300 * The score is only needed if this is likely to be an interactive 1301 * task. Don't go through the expense of computing it if there's 1302 * no chance. 1303 */ 1304 if (sched_interact <= SCHED_INTERACT_HALF && 1305 ts->ts_runtime >= ts->ts_slptime) 1306 return (SCHED_INTERACT_HALF); 1307 1308 if (ts->ts_runtime > ts->ts_slptime) { 1309 div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF); 1310 return (SCHED_INTERACT_HALF + 1311 (SCHED_INTERACT_HALF - (ts->ts_slptime / div))); 1312 } 1313 if (ts->ts_slptime > ts->ts_runtime) { 1314 div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF); 1315 return (ts->ts_runtime / div); 1316 } 1317 /* runtime == slptime */ 1318 if (ts->ts_runtime) 1319 return (SCHED_INTERACT_HALF); 1320 1321 /* 1322 * This can happen if slptime and runtime are 0. 1323 */ 1324 return (0); 1325 1326 } 1327 1328 /* 1329 * Scale the scheduling priority according to the "interactivity" of this 1330 * process. 1331 */ 1332 static void 1333 sched_priority(struct thread *td) 1334 { 1335 int score; 1336 int pri; 1337 1338 if (td->td_pri_class != PRI_TIMESHARE) 1339 return; 1340 /* 1341 * If the score is interactive we place the thread in the realtime 1342 * queue with a priority that is less than kernel and interrupt 1343 * priorities. These threads are not subject to nice restrictions. 1344 * 1345 * Scores greater than this are placed on the normal timeshare queue 1346 * where the priority is partially decided by the most recent cpu 1347 * utilization and the rest is decided by nice value. 1348 * 1349 * The nice value of the process has a linear effect on the calculated 1350 * score. Negative nice values make it easier for a thread to be 1351 * considered interactive. 1352 */ 1353 score = imax(0, sched_interact_score(td) - td->td_proc->p_nice); 1354 if (score < sched_interact) { 1355 pri = PRI_MIN_REALTIME; 1356 pri += ((PRI_MAX_REALTIME - PRI_MIN_REALTIME) / sched_interact) 1357 * score; 1358 KASSERT(pri >= PRI_MIN_REALTIME && pri <= PRI_MAX_REALTIME, 1359 ("sched_priority: invalid interactive priority %d score %d", 1360 pri, score)); 1361 } else { 1362 pri = SCHED_PRI_MIN; 1363 if (td->td_sched->ts_ticks) 1364 pri += SCHED_PRI_TICKS(td->td_sched); 1365 pri += SCHED_PRI_NICE(td->td_proc->p_nice); 1366 KASSERT(pri >= PRI_MIN_TIMESHARE && pri <= PRI_MAX_TIMESHARE, 1367 ("sched_priority: invalid priority %d: nice %d, " 1368 "ticks %d ftick %d ltick %d tick pri %d", 1369 pri, td->td_proc->p_nice, td->td_sched->ts_ticks, 1370 td->td_sched->ts_ftick, td->td_sched->ts_ltick, 1371 SCHED_PRI_TICKS(td->td_sched))); 1372 } 1373 sched_user_prio(td, pri); 1374 1375 return; 1376 } 1377 1378 /* 1379 * This routine enforces a maximum limit on the amount of scheduling history 1380 * kept. It is called after either the slptime or runtime is adjusted. This 1381 * function is ugly due to integer math. 1382 */ 1383 static void 1384 sched_interact_update(struct thread *td) 1385 { 1386 struct td_sched *ts; 1387 u_int sum; 1388 1389 ts = td->td_sched; 1390 sum = ts->ts_runtime + ts->ts_slptime; 1391 if (sum < SCHED_SLP_RUN_MAX) 1392 return; 1393 /* 1394 * This only happens from two places: 1395 * 1) We have added an unusual amount of run time from fork_exit. 1396 * 2) We have added an unusual amount of sleep time from sched_sleep(). 1397 */ 1398 if (sum > SCHED_SLP_RUN_MAX * 2) { 1399 if (ts->ts_runtime > ts->ts_slptime) { 1400 ts->ts_runtime = SCHED_SLP_RUN_MAX; 1401 ts->ts_slptime = 1; 1402 } else { 1403 ts->ts_slptime = SCHED_SLP_RUN_MAX; 1404 ts->ts_runtime = 1; 1405 } 1406 return; 1407 } 1408 /* 1409 * If we have exceeded by more than 1/5th then the algorithm below 1410 * will not bring us back into range. Dividing by two here forces 1411 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX] 1412 */ 1413 if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) { 1414 ts->ts_runtime /= 2; 1415 ts->ts_slptime /= 2; 1416 return; 1417 } 1418 ts->ts_runtime = (ts->ts_runtime / 5) * 4; 1419 ts->ts_slptime = (ts->ts_slptime / 5) * 4; 1420 } 1421 1422 /* 1423 * Scale back the interactivity history when a child thread is created. The 1424 * history is inherited from the parent but the thread may behave totally 1425 * differently. For example, a shell spawning a compiler process. We want 1426 * to learn that the compiler is behaving badly very quickly. 1427 */ 1428 static void 1429 sched_interact_fork(struct thread *td) 1430 { 1431 int ratio; 1432 int sum; 1433 1434 sum = td->td_sched->ts_runtime + td->td_sched->ts_slptime; 1435 if (sum > SCHED_SLP_RUN_FORK) { 1436 ratio = sum / SCHED_SLP_RUN_FORK; 1437 td->td_sched->ts_runtime /= ratio; 1438 td->td_sched->ts_slptime /= ratio; 1439 } 1440 } 1441 1442 /* 1443 * Called from proc0_init() to setup the scheduler fields. 1444 */ 1445 void 1446 schedinit(void) 1447 { 1448 1449 /* 1450 * Set up the scheduler specific parts of proc0. 1451 */ 1452 proc0.p_sched = NULL; /* XXX */ 1453 thread0.td_sched = &td_sched0; 1454 td_sched0.ts_ltick = ticks; 1455 td_sched0.ts_ftick = ticks; 1456 td_sched0.ts_thread = &thread0; 1457 } 1458 1459 /* 1460 * This is only somewhat accurate since given many processes of the same 1461 * priority they will switch when their slices run out, which will be 1462 * at most sched_slice stathz ticks. 1463 */ 1464 int 1465 sched_rr_interval(void) 1466 { 1467 1468 /* Convert sched_slice to hz */ 1469 return (hz/(realstathz/sched_slice)); 1470 } 1471 1472 /* 1473 * Update the percent cpu tracking information when it is requested or 1474 * the total history exceeds the maximum. We keep a sliding history of 1475 * tick counts that slowly decays. This is less precise than the 4BSD 1476 * mechanism since it happens with less regular and frequent events. 1477 */ 1478 static void 1479 sched_pctcpu_update(struct td_sched *ts) 1480 { 1481 1482 if (ts->ts_ticks == 0) 1483 return; 1484 if (ticks - (hz / 10) < ts->ts_ltick && 1485 SCHED_TICK_TOTAL(ts) < SCHED_TICK_MAX) 1486 return; 1487 /* 1488 * Adjust counters and watermark for pctcpu calc. 1489 */ 1490 if (ts->ts_ltick > ticks - SCHED_TICK_TARG) 1491 ts->ts_ticks = (ts->ts_ticks / (ticks - ts->ts_ftick)) * 1492 SCHED_TICK_TARG; 1493 else 1494 ts->ts_ticks = 0; 1495 ts->ts_ltick = ticks; 1496 ts->ts_ftick = ts->ts_ltick - SCHED_TICK_TARG; 1497 } 1498 1499 /* 1500 * Adjust the priority of a thread. Move it to the appropriate run-queue 1501 * if necessary. This is the back-end for several priority related 1502 * functions. 1503 */ 1504 static void 1505 sched_thread_priority(struct thread *td, u_char prio) 1506 { 1507 struct td_sched *ts; 1508 1509 CTR6(KTR_SCHED, "sched_prio: %p(%s) prio %d newprio %d by %p(%s)", 1510 td, td->td_name, td->td_priority, prio, curthread, 1511 curthread->td_name); 1512 ts = td->td_sched; 1513 THREAD_LOCK_ASSERT(td, MA_OWNED); 1514 if (td->td_priority == prio) 1515 return; 1516 1517 if (TD_ON_RUNQ(td) && prio < td->td_priority) { 1518 /* 1519 * If the priority has been elevated due to priority 1520 * propagation, we may have to move ourselves to a new 1521 * queue. This could be optimized to not re-add in some 1522 * cases. 1523 */ 1524 sched_rem(td); 1525 td->td_priority = prio; 1526 sched_add(td, SRQ_BORROWING); 1527 } else if (TD_IS_RUNNING(td)) { 1528 struct tdq *tdq; 1529 int oldpri; 1530 1531 tdq = TDQ_CPU(ts->ts_cpu); 1532 oldpri = td->td_priority; 1533 td->td_priority = prio; 1534 if (prio < tdq->tdq_lowpri) 1535 tdq->tdq_lowpri = prio; 1536 else if (tdq->tdq_lowpri == oldpri) 1537 tdq_setlowpri(tdq, td); 1538 } else 1539 td->td_priority = prio; 1540 } 1541 1542 /* 1543 * Update a thread's priority when it is lent another thread's 1544 * priority. 1545 */ 1546 void 1547 sched_lend_prio(struct thread *td, u_char prio) 1548 { 1549 1550 td->td_flags |= TDF_BORROWING; 1551 sched_thread_priority(td, prio); 1552 } 1553 1554 /* 1555 * Restore a thread's priority when priority propagation is 1556 * over. The prio argument is the minimum priority the thread 1557 * needs to have to satisfy other possible priority lending 1558 * requests. If the thread's regular priority is less 1559 * important than prio, the thread will keep a priority boost 1560 * of prio. 1561 */ 1562 void 1563 sched_unlend_prio(struct thread *td, u_char prio) 1564 { 1565 u_char base_pri; 1566 1567 if (td->td_base_pri >= PRI_MIN_TIMESHARE && 1568 td->td_base_pri <= PRI_MAX_TIMESHARE) 1569 base_pri = td->td_user_pri; 1570 else 1571 base_pri = td->td_base_pri; 1572 if (prio >= base_pri) { 1573 td->td_flags &= ~TDF_BORROWING; 1574 sched_thread_priority(td, base_pri); 1575 } else 1576 sched_lend_prio(td, prio); 1577 } 1578 1579 /* 1580 * Standard entry for setting the priority to an absolute value. 1581 */ 1582 void 1583 sched_prio(struct thread *td, u_char prio) 1584 { 1585 u_char oldprio; 1586 1587 /* First, update the base priority. */ 1588 td->td_base_pri = prio; 1589 1590 /* 1591 * If the thread is borrowing another thread's priority, don't 1592 * ever lower the priority. 1593 */ 1594 if (td->td_flags & TDF_BORROWING && td->td_priority < prio) 1595 return; 1596 1597 /* Change the real priority. */ 1598 oldprio = td->td_priority; 1599 sched_thread_priority(td, prio); 1600 1601 /* 1602 * If the thread is on a turnstile, then let the turnstile update 1603 * its state. 1604 */ 1605 if (TD_ON_LOCK(td) && oldprio != prio) 1606 turnstile_adjust(td, oldprio); 1607 } 1608 1609 /* 1610 * Set the base user priority, does not effect current running priority. 1611 */ 1612 void 1613 sched_user_prio(struct thread *td, u_char prio) 1614 { 1615 u_char oldprio; 1616 1617 td->td_base_user_pri = prio; 1618 if (td->td_flags & TDF_UBORROWING && td->td_user_pri <= prio) 1619 return; 1620 oldprio = td->td_user_pri; 1621 td->td_user_pri = prio; 1622 } 1623 1624 void 1625 sched_lend_user_prio(struct thread *td, u_char prio) 1626 { 1627 u_char oldprio; 1628 1629 THREAD_LOCK_ASSERT(td, MA_OWNED); 1630 td->td_flags |= TDF_UBORROWING; 1631 oldprio = td->td_user_pri; 1632 td->td_user_pri = prio; 1633 } 1634 1635 void 1636 sched_unlend_user_prio(struct thread *td, u_char prio) 1637 { 1638 u_char base_pri; 1639 1640 THREAD_LOCK_ASSERT(td, MA_OWNED); 1641 base_pri = td->td_base_user_pri; 1642 if (prio >= base_pri) { 1643 td->td_flags &= ~TDF_UBORROWING; 1644 sched_user_prio(td, base_pri); 1645 } else { 1646 sched_lend_user_prio(td, prio); 1647 } 1648 } 1649 1650 /* 1651 * Add the thread passed as 'newtd' to the run queue before selecting 1652 * the next thread to run. This is only used for KSE. 1653 */ 1654 static void 1655 sched_switchin(struct tdq *tdq, struct thread *td) 1656 { 1657 #ifdef SMP 1658 spinlock_enter(); 1659 TDQ_UNLOCK(tdq); 1660 thread_lock(td); 1661 spinlock_exit(); 1662 sched_setcpu(td->td_sched, TDQ_ID(tdq), SRQ_YIELDING); 1663 #else 1664 td->td_lock = TDQ_LOCKPTR(tdq); 1665 #endif 1666 tdq_add(tdq, td, SRQ_YIELDING); 1667 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1668 } 1669 1670 /* 1671 * Block a thread for switching. Similar to thread_block() but does not 1672 * bump the spin count. 1673 */ 1674 static inline struct mtx * 1675 thread_block_switch(struct thread *td) 1676 { 1677 struct mtx *lock; 1678 1679 THREAD_LOCK_ASSERT(td, MA_OWNED); 1680 lock = td->td_lock; 1681 td->td_lock = &blocked_lock; 1682 mtx_unlock_spin(lock); 1683 1684 return (lock); 1685 } 1686 1687 /* 1688 * Handle migration from sched_switch(). This happens only for 1689 * cpu binding. 1690 */ 1691 static struct mtx * 1692 sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags) 1693 { 1694 struct tdq *tdn; 1695 1696 tdn = TDQ_CPU(td->td_sched->ts_cpu); 1697 #ifdef SMP 1698 /* 1699 * Do the lock dance required to avoid LOR. We grab an extra 1700 * spinlock nesting to prevent preemption while we're 1701 * not holding either run-queue lock. 1702 */ 1703 spinlock_enter(); 1704 thread_block_switch(td); /* This releases the lock on tdq. */ 1705 TDQ_LOCK(tdn); 1706 tdq_add(tdn, td, flags); 1707 tdq_notify(td->td_sched); 1708 /* 1709 * After we unlock tdn the new cpu still can't switch into this 1710 * thread until we've unblocked it in cpu_switch(). The lock 1711 * pointers may match in the case of HTT cores. Don't unlock here 1712 * or we can deadlock when the other CPU runs the IPI handler. 1713 */ 1714 if (TDQ_LOCKPTR(tdn) != TDQ_LOCKPTR(tdq)) { 1715 TDQ_UNLOCK(tdn); 1716 TDQ_LOCK(tdq); 1717 } 1718 spinlock_exit(); 1719 #endif 1720 return (TDQ_LOCKPTR(tdn)); 1721 } 1722 1723 /* 1724 * Release a thread that was blocked with thread_block_switch(). 1725 */ 1726 static inline void 1727 thread_unblock_switch(struct thread *td, struct mtx *mtx) 1728 { 1729 atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock, 1730 (uintptr_t)mtx); 1731 } 1732 1733 /* 1734 * Switch threads. This function has to handle threads coming in while 1735 * blocked for some reason, running, or idle. It also must deal with 1736 * migrating a thread from one queue to another as running threads may 1737 * be assigned elsewhere via binding. 1738 */ 1739 void 1740 sched_switch(struct thread *td, struct thread *newtd, int flags) 1741 { 1742 struct tdq *tdq; 1743 struct td_sched *ts; 1744 struct mtx *mtx; 1745 int srqflag; 1746 int cpuid; 1747 1748 THREAD_LOCK_ASSERT(td, MA_OWNED); 1749 1750 cpuid = PCPU_GET(cpuid); 1751 tdq = TDQ_CPU(cpuid); 1752 ts = td->td_sched; 1753 mtx = td->td_lock; 1754 ts->ts_rltick = ticks; 1755 td->td_lastcpu = td->td_oncpu; 1756 td->td_oncpu = NOCPU; 1757 td->td_flags &= ~TDF_NEEDRESCHED; 1758 td->td_owepreempt = 0; 1759 /* 1760 * The lock pointer in an idle thread should never change. Reset it 1761 * to CAN_RUN as well. 1762 */ 1763 if (TD_IS_IDLETHREAD(td)) { 1764 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1765 TD_SET_CAN_RUN(td); 1766 } else if (TD_IS_RUNNING(td)) { 1767 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1768 tdq_load_rem(tdq, ts); 1769 srqflag = (flags & SW_PREEMPT) ? 1770 SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED : 1771 SRQ_OURSELF|SRQ_YIELDING; 1772 if (ts->ts_cpu == cpuid) 1773 tdq_add(tdq, td, srqflag); 1774 else 1775 mtx = sched_switch_migrate(tdq, td, srqflag); 1776 } else { 1777 /* This thread must be going to sleep. */ 1778 TDQ_LOCK(tdq); 1779 mtx = thread_block_switch(td); 1780 tdq_load_rem(tdq, ts); 1781 } 1782 /* 1783 * We enter here with the thread blocked and assigned to the 1784 * appropriate cpu run-queue or sleep-queue and with the current 1785 * thread-queue locked. 1786 */ 1787 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED); 1788 /* 1789 * If KSE assigned a new thread just add it here and let choosethread 1790 * select the best one. 1791 */ 1792 if (newtd != NULL) 1793 sched_switchin(tdq, newtd); 1794 newtd = choosethread(); 1795 /* 1796 * Call the MD code to switch contexts if necessary. 1797 */ 1798 if (td != newtd) { 1799 #ifdef HWPMC_HOOKS 1800 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1801 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT); 1802 #endif 1803 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object); 1804 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd; 1805 cpu_switch(td, newtd, mtx); 1806 /* 1807 * We may return from cpu_switch on a different cpu. However, 1808 * we always return with td_lock pointing to the current cpu's 1809 * run queue lock. 1810 */ 1811 cpuid = PCPU_GET(cpuid); 1812 tdq = TDQ_CPU(cpuid); 1813 lock_profile_obtain_lock_success( 1814 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__); 1815 #ifdef HWPMC_HOOKS 1816 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1817 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN); 1818 #endif 1819 } else 1820 thread_unblock_switch(td, mtx); 1821 /* 1822 * We should always get here with the lowest priority td possible. 1823 */ 1824 tdq->tdq_lowpri = td->td_priority; 1825 /* 1826 * Assert that all went well and return. 1827 */ 1828 TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED); 1829 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1830 td->td_oncpu = cpuid; 1831 } 1832 1833 /* 1834 * Adjust thread priorities as a result of a nice request. 1835 */ 1836 void 1837 sched_nice(struct proc *p, int nice) 1838 { 1839 struct thread *td; 1840 1841 PROC_LOCK_ASSERT(p, MA_OWNED); 1842 PROC_SLOCK_ASSERT(p, MA_OWNED); 1843 1844 p->p_nice = nice; 1845 FOREACH_THREAD_IN_PROC(p, td) { 1846 thread_lock(td); 1847 sched_priority(td); 1848 sched_prio(td, td->td_base_user_pri); 1849 thread_unlock(td); 1850 } 1851 } 1852 1853 /* 1854 * Record the sleep time for the interactivity scorer. 1855 */ 1856 void 1857 sched_sleep(struct thread *td) 1858 { 1859 1860 THREAD_LOCK_ASSERT(td, MA_OWNED); 1861 1862 td->td_slptick = ticks; 1863 } 1864 1865 /* 1866 * Schedule a thread to resume execution and record how long it voluntarily 1867 * slept. We also update the pctcpu, interactivity, and priority. 1868 */ 1869 void 1870 sched_wakeup(struct thread *td) 1871 { 1872 struct td_sched *ts; 1873 int slptick; 1874 1875 THREAD_LOCK_ASSERT(td, MA_OWNED); 1876 ts = td->td_sched; 1877 /* 1878 * If we slept for more than a tick update our interactivity and 1879 * priority. 1880 */ 1881 slptick = td->td_slptick; 1882 td->td_slptick = 0; 1883 if (slptick && slptick != ticks) { 1884 u_int hzticks; 1885 1886 hzticks = (ticks - slptick) << SCHED_TICK_SHIFT; 1887 ts->ts_slptime += hzticks; 1888 sched_interact_update(td); 1889 sched_pctcpu_update(ts); 1890 sched_priority(td); 1891 } 1892 /* Reset the slice value after we sleep. */ 1893 ts->ts_slice = sched_slice; 1894 sched_add(td, SRQ_BORING); 1895 } 1896 1897 /* 1898 * Penalize the parent for creating a new child and initialize the child's 1899 * priority. 1900 */ 1901 void 1902 sched_fork(struct thread *td, struct thread *child) 1903 { 1904 THREAD_LOCK_ASSERT(td, MA_OWNED); 1905 sched_fork_thread(td, child); 1906 /* 1907 * Penalize the parent and child for forking. 1908 */ 1909 sched_interact_fork(child); 1910 sched_priority(child); 1911 td->td_sched->ts_runtime += tickincr; 1912 sched_interact_update(td); 1913 sched_priority(td); 1914 } 1915 1916 /* 1917 * Fork a new thread, may be within the same process. 1918 */ 1919 void 1920 sched_fork_thread(struct thread *td, struct thread *child) 1921 { 1922 struct td_sched *ts; 1923 struct td_sched *ts2; 1924 1925 /* 1926 * Initialize child. 1927 */ 1928 THREAD_LOCK_ASSERT(td, MA_OWNED); 1929 sched_newthread(child); 1930 child->td_lock = TDQ_LOCKPTR(TDQ_SELF()); 1931 child->td_cpuset = cpuset_ref(td->td_cpuset); 1932 ts = td->td_sched; 1933 ts2 = child->td_sched; 1934 ts2->ts_cpu = ts->ts_cpu; 1935 ts2->ts_runq = NULL; 1936 /* 1937 * Grab our parents cpu estimation information and priority. 1938 */ 1939 ts2->ts_ticks = ts->ts_ticks; 1940 ts2->ts_ltick = ts->ts_ltick; 1941 ts2->ts_ftick = ts->ts_ftick; 1942 child->td_user_pri = td->td_user_pri; 1943 child->td_base_user_pri = td->td_base_user_pri; 1944 /* 1945 * And update interactivity score. 1946 */ 1947 ts2->ts_slptime = ts->ts_slptime; 1948 ts2->ts_runtime = ts->ts_runtime; 1949 ts2->ts_slice = 1; /* Attempt to quickly learn interactivity. */ 1950 } 1951 1952 /* 1953 * Adjust the priority class of a thread. 1954 */ 1955 void 1956 sched_class(struct thread *td, int class) 1957 { 1958 1959 THREAD_LOCK_ASSERT(td, MA_OWNED); 1960 if (td->td_pri_class == class) 1961 return; 1962 /* 1963 * On SMP if we're on the RUNQ we must adjust the transferable 1964 * count because could be changing to or from an interrupt 1965 * class. 1966 */ 1967 if (TD_ON_RUNQ(td)) { 1968 struct tdq *tdq; 1969 1970 tdq = TDQ_CPU(td->td_sched->ts_cpu); 1971 if (THREAD_CAN_MIGRATE(td)) 1972 tdq->tdq_transferable--; 1973 td->td_pri_class = class; 1974 if (THREAD_CAN_MIGRATE(td)) 1975 tdq->tdq_transferable++; 1976 } 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_SLOCK_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 #ifdef KSE 2010 /* 2011 * KSE forks and exits so often that this penalty causes short-lived 2012 * threads to always be non-interactive. This causes mozilla to 2013 * crawl under load. 2014 */ 2015 if ((td->td_pflags & TDP_SA) && td->td_proc == child->td_proc) 2016 return; 2017 #endif 2018 /* 2019 * Give the child's runtime to the parent without returning the 2020 * sleep time as a penalty to the parent. This causes shells that 2021 * launch expensive things to mark their children as expensive. 2022 */ 2023 thread_lock(td); 2024 td->td_sched->ts_runtime += child->td_sched->ts_runtime; 2025 sched_interact_update(td); 2026 sched_priority(td); 2027 thread_unlock(td); 2028 } 2029 2030 /* 2031 * Fix priorities on return to user-space. Priorities may be elevated due 2032 * to static priorities in msleep() or similar. 2033 */ 2034 void 2035 sched_userret(struct thread *td) 2036 { 2037 /* 2038 * XXX we cheat slightly on the locking here to avoid locking in 2039 * the usual case. Setting td_priority here is essentially an 2040 * incomplete workaround for not setting it properly elsewhere. 2041 * Now that some interrupt handlers are threads, not setting it 2042 * properly elsewhere can clobber it in the window between setting 2043 * it here and returning to user mode, so don't waste time setting 2044 * it perfectly here. 2045 */ 2046 KASSERT((td->td_flags & TDF_BORROWING) == 0, 2047 ("thread with borrowed priority returning to userland")); 2048 if (td->td_priority != td->td_user_pri) { 2049 thread_lock(td); 2050 td->td_priority = td->td_user_pri; 2051 td->td_base_pri = td->td_user_pri; 2052 if (lowpri_userret) 2053 tdq_setlowpri(TDQ_SELF(), td); 2054 thread_unlock(td); 2055 } 2056 } 2057 2058 /* 2059 * Handle a stathz tick. This is really only relevant for timeshare 2060 * threads. 2061 */ 2062 void 2063 sched_clock(struct thread *td) 2064 { 2065 struct tdq *tdq; 2066 struct td_sched *ts; 2067 2068 THREAD_LOCK_ASSERT(td, MA_OWNED); 2069 tdq = TDQ_SELF(); 2070 #ifdef SMP 2071 /* 2072 * We run the long term load balancer infrequently on the first cpu. 2073 */ 2074 if (balance_tdq == tdq) { 2075 if (balance_ticks && --balance_ticks == 0) 2076 sched_balance(); 2077 } 2078 #endif 2079 /* 2080 * Advance the insert index once for each tick to ensure that all 2081 * threads get a chance to run. 2082 */ 2083 if (tdq->tdq_idx == tdq->tdq_ridx) { 2084 tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS; 2085 if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx])) 2086 tdq->tdq_ridx = tdq->tdq_idx; 2087 } 2088 ts = td->td_sched; 2089 if (td->td_pri_class & PRI_FIFO_BIT) 2090 return; 2091 if (td->td_pri_class == PRI_TIMESHARE) { 2092 /* 2093 * We used a tick; charge it to the thread so 2094 * that we can compute our interactivity. 2095 */ 2096 td->td_sched->ts_runtime += tickincr; 2097 sched_interact_update(td); 2098 } 2099 /* 2100 * We used up one time slice. 2101 */ 2102 if (--ts->ts_slice > 0) 2103 return; 2104 /* 2105 * We're out of time, recompute priorities and requeue. 2106 */ 2107 sched_priority(td); 2108 td->td_flags |= TDF_NEEDRESCHED; 2109 } 2110 2111 /* 2112 * Called once per hz tick. Used for cpu utilization information. This 2113 * is easier than trying to scale based on stathz. 2114 */ 2115 void 2116 sched_tick(void) 2117 { 2118 struct td_sched *ts; 2119 2120 ts = curthread->td_sched; 2121 /* Adjust ticks for pctcpu */ 2122 ts->ts_ticks += 1 << SCHED_TICK_SHIFT; 2123 ts->ts_ltick = ticks; 2124 /* 2125 * Update if we've exceeded our desired tick threshhold by over one 2126 * second. 2127 */ 2128 if (ts->ts_ftick + SCHED_TICK_MAX < ts->ts_ltick) 2129 sched_pctcpu_update(ts); 2130 } 2131 2132 /* 2133 * Return whether the current CPU has runnable tasks. Used for in-kernel 2134 * cooperative idle threads. 2135 */ 2136 int 2137 sched_runnable(void) 2138 { 2139 struct tdq *tdq; 2140 int load; 2141 2142 load = 1; 2143 2144 tdq = TDQ_SELF(); 2145 if ((curthread->td_flags & TDF_IDLETD) != 0) { 2146 if (tdq->tdq_load > 0) 2147 goto out; 2148 } else 2149 if (tdq->tdq_load - 1 > 0) 2150 goto out; 2151 load = 0; 2152 out: 2153 return (load); 2154 } 2155 2156 /* 2157 * Choose the highest priority thread to run. The thread is removed from 2158 * the run-queue while running however the load remains. For SMP we set 2159 * the tdq in the global idle bitmask if it idles here. 2160 */ 2161 struct thread * 2162 sched_choose(void) 2163 { 2164 struct td_sched *ts; 2165 struct tdq *tdq; 2166 2167 tdq = TDQ_SELF(); 2168 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2169 ts = tdq_choose(tdq); 2170 if (ts) { 2171 tdq_runq_rem(tdq, ts); 2172 return (ts->ts_thread); 2173 } 2174 return (PCPU_GET(idlethread)); 2175 } 2176 2177 /* 2178 * Set owepreempt if necessary. Preemption never happens directly in ULE, 2179 * we always request it once we exit a critical section. 2180 */ 2181 static inline void 2182 sched_setpreempt(struct thread *td) 2183 { 2184 struct thread *ctd; 2185 int cpri; 2186 int pri; 2187 2188 ctd = curthread; 2189 pri = td->td_priority; 2190 cpri = ctd->td_priority; 2191 if (td->td_priority < cpri) 2192 curthread->td_flags |= TDF_NEEDRESCHED; 2193 if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd)) 2194 return; 2195 /* 2196 * Always preempt IDLE threads. Otherwise only if the preempting 2197 * thread is an ithread. 2198 */ 2199 if (pri > preempt_thresh && cpri < PRI_MIN_IDLE) 2200 return; 2201 ctd->td_owepreempt = 1; 2202 return; 2203 } 2204 2205 /* 2206 * Add a thread to a thread queue. Initializes priority, slice, runq, and 2207 * add it to the appropriate queue. This is the internal function called 2208 * when the tdq is predetermined. 2209 */ 2210 void 2211 tdq_add(struct tdq *tdq, struct thread *td, int flags) 2212 { 2213 struct td_sched *ts; 2214 int class; 2215 2216 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2217 KASSERT((td->td_inhibitors == 0), 2218 ("sched_add: trying to run inhibited thread")); 2219 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)), 2220 ("sched_add: bad thread state")); 2221 KASSERT(td->td_flags & TDF_INMEM, 2222 ("sched_add: thread swapped out")); 2223 2224 ts = td->td_sched; 2225 class = PRI_BASE(td->td_pri_class); 2226 TD_SET_RUNQ(td); 2227 if (ts->ts_slice == 0) 2228 ts->ts_slice = sched_slice; 2229 /* 2230 * Pick the run queue based on priority. 2231 */ 2232 if (td->td_priority <= PRI_MAX_REALTIME) 2233 ts->ts_runq = &tdq->tdq_realtime; 2234 else if (td->td_priority <= PRI_MAX_TIMESHARE) 2235 ts->ts_runq = &tdq->tdq_timeshare; 2236 else 2237 ts->ts_runq = &tdq->tdq_idle; 2238 if (td->td_priority < tdq->tdq_lowpri) 2239 tdq->tdq_lowpri = td->td_priority; 2240 tdq_runq_add(tdq, ts, flags); 2241 tdq_load_add(tdq, ts); 2242 } 2243 2244 /* 2245 * Select the target thread queue and add a thread to it. Request 2246 * preemption or IPI a remote processor if required. 2247 */ 2248 void 2249 sched_add(struct thread *td, int flags) 2250 { 2251 struct td_sched *ts; 2252 struct tdq *tdq; 2253 #ifdef SMP 2254 int cpuid; 2255 int cpu; 2256 #endif 2257 CTR5(KTR_SCHED, "sched_add: %p(%s) prio %d by %p(%s)", 2258 td, td->td_name, td->td_priority, curthread, 2259 curthread->td_name); 2260 THREAD_LOCK_ASSERT(td, MA_OWNED); 2261 ts = td->td_sched; 2262 /* 2263 * Recalculate the priority before we select the target cpu or 2264 * run-queue. 2265 */ 2266 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) 2267 sched_priority(td); 2268 #ifdef SMP 2269 cpuid = PCPU_GET(cpuid); 2270 /* 2271 * Pick the destination cpu and if it isn't ours transfer to the 2272 * target cpu. 2273 */ 2274 cpu = sched_pickcpu(ts, flags); 2275 tdq = sched_setcpu(ts, cpu, flags); 2276 tdq_add(tdq, td, flags); 2277 if (cpu != cpuid) { 2278 tdq_notify(ts); 2279 return; 2280 } 2281 #else 2282 tdq = TDQ_SELF(); 2283 TDQ_LOCK(tdq); 2284 /* 2285 * Now that the thread is moving to the run-queue, set the lock 2286 * to the scheduler's lock. 2287 */ 2288 thread_lock_set(td, TDQ_LOCKPTR(tdq)); 2289 tdq_add(tdq, td, flags); 2290 #endif 2291 if (!(flags & SRQ_YIELDING)) 2292 sched_setpreempt(td); 2293 } 2294 2295 /* 2296 * Remove a thread from a run-queue without running it. This is used 2297 * when we're stealing a thread from a remote queue. Otherwise all threads 2298 * exit by calling sched_exit_thread() and sched_throw() themselves. 2299 */ 2300 void 2301 sched_rem(struct thread *td) 2302 { 2303 struct tdq *tdq; 2304 struct td_sched *ts; 2305 2306 CTR5(KTR_SCHED, "sched_rem: %p(%s) prio %d by %p(%s)", 2307 td, td->td_name, td->td_priority, curthread, 2308 curthread->td_name); 2309 ts = td->td_sched; 2310 tdq = TDQ_CPU(ts->ts_cpu); 2311 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2312 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2313 KASSERT(TD_ON_RUNQ(td), 2314 ("sched_rem: thread not on run queue")); 2315 tdq_runq_rem(tdq, ts); 2316 tdq_load_rem(tdq, ts); 2317 TD_SET_CAN_RUN(td); 2318 if (td->td_priority == tdq->tdq_lowpri) 2319 tdq_setlowpri(tdq, NULL); 2320 } 2321 2322 /* 2323 * Fetch cpu utilization information. Updates on demand. 2324 */ 2325 fixpt_t 2326 sched_pctcpu(struct thread *td) 2327 { 2328 fixpt_t pctcpu; 2329 struct td_sched *ts; 2330 2331 pctcpu = 0; 2332 ts = td->td_sched; 2333 if (ts == NULL) 2334 return (0); 2335 2336 thread_lock(td); 2337 if (ts->ts_ticks) { 2338 int rtick; 2339 2340 sched_pctcpu_update(ts); 2341 /* How many rtick per second ? */ 2342 rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz); 2343 pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT; 2344 } 2345 thread_unlock(td); 2346 2347 return (pctcpu); 2348 } 2349 2350 /* 2351 * Enforce affinity settings for a thread. Called after adjustments to 2352 * cpumask. 2353 */ 2354 void 2355 sched_affinity(struct thread *td) 2356 { 2357 #ifdef SMP 2358 struct td_sched *ts; 2359 int cpu; 2360 2361 THREAD_LOCK_ASSERT(td, MA_OWNED); 2362 ts = td->td_sched; 2363 if (THREAD_CAN_SCHED(td, ts->ts_cpu)) 2364 return; 2365 if (!TD_IS_RUNNING(td)) 2366 return; 2367 td->td_flags |= TDF_NEEDRESCHED; 2368 if (!THREAD_CAN_MIGRATE(td)) 2369 return; 2370 /* 2371 * Assign the new cpu and force a switch before returning to 2372 * userspace. If the target thread is not running locally send 2373 * an ipi to force the issue. 2374 */ 2375 cpu = ts->ts_cpu; 2376 ts->ts_cpu = sched_pickcpu(ts, 0); 2377 if (cpu != PCPU_GET(cpuid)) 2378 ipi_selected(1 << cpu, IPI_PREEMPT); 2379 #endif 2380 } 2381 2382 /* 2383 * Bind a thread to a target cpu. 2384 */ 2385 void 2386 sched_bind(struct thread *td, int cpu) 2387 { 2388 struct td_sched *ts; 2389 2390 THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED); 2391 ts = td->td_sched; 2392 if (ts->ts_flags & TSF_BOUND) 2393 sched_unbind(td); 2394 ts->ts_flags |= TSF_BOUND; 2395 sched_pin(); 2396 if (PCPU_GET(cpuid) == cpu) 2397 return; 2398 ts->ts_cpu = cpu; 2399 /* When we return from mi_switch we'll be on the correct cpu. */ 2400 mi_switch(SW_VOL, NULL); 2401 } 2402 2403 /* 2404 * Release a bound thread. 2405 */ 2406 void 2407 sched_unbind(struct thread *td) 2408 { 2409 struct td_sched *ts; 2410 2411 THREAD_LOCK_ASSERT(td, MA_OWNED); 2412 ts = td->td_sched; 2413 if ((ts->ts_flags & TSF_BOUND) == 0) 2414 return; 2415 ts->ts_flags &= ~TSF_BOUND; 2416 sched_unpin(); 2417 } 2418 2419 int 2420 sched_is_bound(struct thread *td) 2421 { 2422 THREAD_LOCK_ASSERT(td, MA_OWNED); 2423 return (td->td_sched->ts_flags & TSF_BOUND); 2424 } 2425 2426 /* 2427 * Basic yield call. 2428 */ 2429 void 2430 sched_relinquish(struct thread *td) 2431 { 2432 thread_lock(td); 2433 SCHED_STAT_INC(switch_relinquish); 2434 mi_switch(SW_VOL, NULL); 2435 thread_unlock(td); 2436 } 2437 2438 /* 2439 * Return the total system load. 2440 */ 2441 int 2442 sched_load(void) 2443 { 2444 #ifdef SMP 2445 int total; 2446 int i; 2447 2448 total = 0; 2449 for (i = 0; i <= mp_maxid; i++) 2450 total += TDQ_CPU(i)->tdq_sysload; 2451 return (total); 2452 #else 2453 return (TDQ_SELF()->tdq_sysload); 2454 #endif 2455 } 2456 2457 int 2458 sched_sizeof_proc(void) 2459 { 2460 return (sizeof(struct proc)); 2461 } 2462 2463 int 2464 sched_sizeof_thread(void) 2465 { 2466 return (sizeof(struct thread) + sizeof(struct td_sched)); 2467 } 2468 2469 /* 2470 * The actual idle process. 2471 */ 2472 void 2473 sched_idletd(void *dummy) 2474 { 2475 struct thread *td; 2476 struct tdq *tdq; 2477 2478 td = curthread; 2479 tdq = TDQ_SELF(); 2480 mtx_assert(&Giant, MA_NOTOWNED); 2481 /* ULE relies on preemption for idle interruption. */ 2482 for (;;) { 2483 #ifdef SMP 2484 if (tdq_idled(tdq)) 2485 cpu_idle(); 2486 #else 2487 cpu_idle(); 2488 #endif 2489 } 2490 } 2491 2492 /* 2493 * A CPU is entering for the first time or a thread is exiting. 2494 */ 2495 void 2496 sched_throw(struct thread *td) 2497 { 2498 struct thread *newtd; 2499 struct tdq *tdq; 2500 2501 tdq = TDQ_SELF(); 2502 if (td == NULL) { 2503 /* Correct spinlock nesting and acquire the correct lock. */ 2504 TDQ_LOCK(tdq); 2505 spinlock_exit(); 2506 } else { 2507 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2508 tdq_load_rem(tdq, td->td_sched); 2509 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object); 2510 } 2511 KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count")); 2512 newtd = choosethread(); 2513 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd; 2514 PCPU_SET(switchtime, cpu_ticks()); 2515 PCPU_SET(switchticks, ticks); 2516 cpu_throw(td, newtd); /* doesn't return */ 2517 } 2518 2519 /* 2520 * This is called from fork_exit(). Just acquire the correct locks and 2521 * let fork do the rest of the work. 2522 */ 2523 void 2524 sched_fork_exit(struct thread *td) 2525 { 2526 struct td_sched *ts; 2527 struct tdq *tdq; 2528 int cpuid; 2529 2530 /* 2531 * Finish setting up thread glue so that it begins execution in a 2532 * non-nested critical section with the scheduler lock held. 2533 */ 2534 cpuid = PCPU_GET(cpuid); 2535 tdq = TDQ_CPU(cpuid); 2536 ts = td->td_sched; 2537 if (TD_IS_IDLETHREAD(td)) 2538 td->td_lock = TDQ_LOCKPTR(tdq); 2539 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2540 td->td_oncpu = cpuid; 2541 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED); 2542 lock_profile_obtain_lock_success( 2543 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__); 2544 tdq->tdq_lowpri = td->td_priority; 2545 } 2546 2547 static SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, 2548 "Scheduler"); 2549 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0, 2550 "Scheduler name"); 2551 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0, 2552 "Slice size for timeshare threads"); 2553 SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0, 2554 "Interactivity score threshold"); 2555 SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW, &preempt_thresh, 2556 0,"Min priority for preemption, lower priorities have greater precedence"); 2557 #ifdef SMP 2558 SYSCTL_INT(_kern_sched, OID_AUTO, pick_pri, CTLFLAG_RW, &pick_pri, 0, 2559 "Pick the target cpu based on priority rather than load."); 2560 SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0, 2561 "Number of hz ticks to keep thread affinity for"); 2562 SYSCTL_INT(_kern_sched, OID_AUTO, tryself, CTLFLAG_RW, &tryself, 0, ""); 2563 SYSCTL_INT(_kern_sched, OID_AUTO, userret, CTLFLAG_RW, &lowpri_userret, 0, ""); 2564 SYSCTL_INT(_kern_sched, OID_AUTO, oldtryself, CTLFLAG_RW, &oldtryself, 0, ""); 2565 SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0, 2566 "Enables the long-term load balancer"); 2567 SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW, 2568 &balance_interval, 0, 2569 "Average frequency in stathz ticks to run the long-term balancer"); 2570 SYSCTL_INT(_kern_sched, OID_AUTO, steal_htt, CTLFLAG_RW, &steal_htt, 0, 2571 "Steals work from another hyper-threaded core on idle"); 2572 SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0, 2573 "Attempts to steal work from other cores before idling"); 2574 SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0, 2575 "Minimum load on remote cpu before we'll steal"); 2576 #endif 2577 2578 /* ps compat. All cpu percentages from ULE are weighted. */ 2579 static int ccpu = 0; 2580 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 2581 2582 2583 #define KERN_SWITCH_INCLUDE 1 2584 #include "kern/kern_switch.c" 2585