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 189 190 /* 191 * tdq - per processor runqs and statistics. All fields are protected by the 192 * tdq_lock. The load and lowpri may be accessed without to avoid excess 193 * locking in sched_pickcpu(); 194 */ 195 struct tdq { 196 struct cpu_group *tdq_cg; /* Pointer to cpu topology. */ 197 struct mtx tdq_lock; /* run queue lock. */ 198 struct runq tdq_realtime; /* real-time run queue. */ 199 struct runq tdq_timeshare; /* timeshare run queue. */ 200 struct runq tdq_idle; /* Queue of IDLE threads. */ 201 int tdq_load; /* Aggregate load. */ 202 int tdq_sysload; /* For loadavg, !ITHD load. */ 203 u_char tdq_idx; /* Current insert index. */ 204 u_char tdq_ridx; /* Current removal index. */ 205 u_char tdq_lowpri; /* Lowest priority thread. */ 206 u_char tdq_ipipending; /* IPI pending. */ 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 affinity; 224 static int steal_htt = 1; 225 static int steal_idle = 1; 226 static int steal_thresh = 2; 227 228 /* 229 * One thread queue per processor. 230 */ 231 static struct tdq tdq_cpu[MAXCPU]; 232 static struct tdq *balance_tdq; 233 static int balance_ticks; 234 235 #define TDQ_SELF() (&tdq_cpu[PCPU_GET(cpuid)]) 236 #define TDQ_CPU(x) (&tdq_cpu[(x)]) 237 #define TDQ_ID(x) ((int)((x) - tdq_cpu)) 238 #else /* !SMP */ 239 static struct tdq tdq_cpu; 240 241 #define TDQ_ID(x) (0) 242 #define TDQ_SELF() (&tdq_cpu) 243 #define TDQ_CPU(x) (&tdq_cpu) 244 #endif 245 246 #define TDQ_LOCK_ASSERT(t, type) mtx_assert(TDQ_LOCKPTR((t)), (type)) 247 #define TDQ_LOCK(t) mtx_lock_spin(TDQ_LOCKPTR((t))) 248 #define TDQ_LOCK_FLAGS(t, f) mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f)) 249 #define TDQ_UNLOCK(t) mtx_unlock_spin(TDQ_LOCKPTR((t))) 250 #define TDQ_LOCKPTR(t) (&(t)->tdq_lock) 251 252 static void sched_priority(struct thread *); 253 static void sched_thread_priority(struct thread *, u_char); 254 static int sched_interact_score(struct thread *); 255 static void sched_interact_update(struct thread *); 256 static void sched_interact_fork(struct thread *); 257 static void sched_pctcpu_update(struct td_sched *); 258 259 /* Operations on per processor queues */ 260 static struct td_sched * tdq_choose(struct tdq *); 261 static void tdq_setup(struct tdq *); 262 static void tdq_load_add(struct tdq *, struct td_sched *); 263 static void tdq_load_rem(struct tdq *, struct td_sched *); 264 static __inline void tdq_runq_add(struct tdq *, struct td_sched *, int); 265 static __inline void tdq_runq_rem(struct tdq *, struct td_sched *); 266 static inline int sched_shouldpreempt(int, int, int); 267 void tdq_print(int cpu); 268 static void runq_print(struct runq *rq); 269 static void tdq_add(struct tdq *, struct thread *, int); 270 #ifdef SMP 271 static int tdq_move(struct tdq *, struct tdq *); 272 static int tdq_idled(struct tdq *); 273 static void tdq_notify(struct tdq *, struct td_sched *); 274 static struct td_sched *tdq_steal(struct tdq *, int); 275 static struct td_sched *runq_steal(struct runq *, int); 276 static int sched_pickcpu(struct td_sched *, int); 277 static void sched_balance(void); 278 static int sched_balance_pair(struct tdq *, struct tdq *); 279 static inline struct tdq *sched_setcpu(struct td_sched *, int, int); 280 static inline struct mtx *thread_block_switch(struct thread *); 281 static inline void thread_unblock_switch(struct thread *, struct mtx *); 282 static struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int); 283 #endif 284 285 static void sched_setup(void *dummy); 286 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL) 287 288 static void sched_initticks(void *dummy); 289 SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks, NULL) 290 291 /* 292 * Print the threads waiting on a run-queue. 293 */ 294 static void 295 runq_print(struct runq *rq) 296 { 297 struct rqhead *rqh; 298 struct td_sched *ts; 299 int pri; 300 int j; 301 int i; 302 303 for (i = 0; i < RQB_LEN; i++) { 304 printf("\t\trunq bits %d 0x%zx\n", 305 i, rq->rq_status.rqb_bits[i]); 306 for (j = 0; j < RQB_BPW; j++) 307 if (rq->rq_status.rqb_bits[i] & (1ul << j)) { 308 pri = j + (i << RQB_L2BPW); 309 rqh = &rq->rq_queues[pri]; 310 TAILQ_FOREACH(ts, rqh, ts_procq) { 311 printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n", 312 ts->ts_thread, ts->ts_thread->td_name, ts->ts_thread->td_priority, ts->ts_rqindex, pri); 313 } 314 } 315 } 316 } 317 318 /* 319 * Print the status of a per-cpu thread queue. Should be a ddb show cmd. 320 */ 321 void 322 tdq_print(int cpu) 323 { 324 struct tdq *tdq; 325 326 tdq = TDQ_CPU(cpu); 327 328 printf("tdq %d:\n", TDQ_ID(tdq)); 329 printf("\tlock %p\n", TDQ_LOCKPTR(tdq)); 330 printf("\tLock name: %s\n", tdq->tdq_name); 331 printf("\tload: %d\n", tdq->tdq_load); 332 printf("\ttimeshare idx: %d\n", tdq->tdq_idx); 333 printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx); 334 printf("\trealtime runq:\n"); 335 runq_print(&tdq->tdq_realtime); 336 printf("\ttimeshare runq:\n"); 337 runq_print(&tdq->tdq_timeshare); 338 printf("\tidle runq:\n"); 339 runq_print(&tdq->tdq_idle); 340 printf("\tload transferable: %d\n", tdq->tdq_transferable); 341 printf("\tlowest priority: %d\n", tdq->tdq_lowpri); 342 } 343 344 static inline int 345 sched_shouldpreempt(int pri, int cpri, int remote) 346 { 347 /* 348 * If the new priority is not better than the current priority there is 349 * nothing to do. 350 */ 351 if (pri >= cpri) 352 return (0); 353 /* 354 * Always preempt idle. 355 */ 356 if (cpri >= PRI_MIN_IDLE) 357 return (1); 358 /* 359 * If preemption is disabled don't preempt others. 360 */ 361 if (preempt_thresh == 0) 362 return (0); 363 /* 364 * Preempt if we exceed the threshold. 365 */ 366 if (pri <= preempt_thresh) 367 return (1); 368 /* 369 * If we're realtime or better and there is timeshare or worse running 370 * preempt only remote processors. 371 */ 372 if (remote && pri <= PRI_MAX_REALTIME && cpri > PRI_MAX_REALTIME) 373 return (1); 374 return (0); 375 } 376 377 #define TS_RQ_PPQ (((PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE) + 1) / RQ_NQS) 378 /* 379 * Add a thread to the actual run-queue. Keeps transferable counts up to 380 * date with what is actually on the run-queue. Selects the correct 381 * queue position for timeshare threads. 382 */ 383 static __inline void 384 tdq_runq_add(struct tdq *tdq, struct td_sched *ts, int flags) 385 { 386 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 387 THREAD_LOCK_ASSERT(ts->ts_thread, MA_OWNED); 388 if (THREAD_CAN_MIGRATE(ts->ts_thread)) { 389 tdq->tdq_transferable++; 390 ts->ts_flags |= TSF_XFERABLE; 391 } 392 if (ts->ts_runq == &tdq->tdq_timeshare) { 393 u_char pri; 394 395 pri = ts->ts_thread->td_priority; 396 KASSERT(pri <= PRI_MAX_TIMESHARE && pri >= PRI_MIN_TIMESHARE, 397 ("Invalid priority %d on timeshare runq", pri)); 398 /* 399 * This queue contains only priorities between MIN and MAX 400 * realtime. Use the whole queue to represent these values. 401 */ 402 if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) { 403 pri = (pri - PRI_MIN_TIMESHARE) / TS_RQ_PPQ; 404 pri = (pri + tdq->tdq_idx) % RQ_NQS; 405 /* 406 * This effectively shortens the queue by one so we 407 * can have a one slot difference between idx and 408 * ridx while we wait for threads to drain. 409 */ 410 if (tdq->tdq_ridx != tdq->tdq_idx && 411 pri == tdq->tdq_ridx) 412 pri = (unsigned char)(pri - 1) % RQ_NQS; 413 } else 414 pri = tdq->tdq_ridx; 415 runq_add_pri(ts->ts_runq, ts, pri, flags); 416 } else 417 runq_add(ts->ts_runq, ts, flags); 418 } 419 420 /* 421 * Remove a thread from a run-queue. This typically happens when a thread 422 * is selected to run. Running threads are not on the queue and the 423 * transferable count does not reflect them. 424 */ 425 static __inline void 426 tdq_runq_rem(struct tdq *tdq, struct td_sched *ts) 427 { 428 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 429 KASSERT(ts->ts_runq != NULL, 430 ("tdq_runq_remove: thread %p null ts_runq", ts->ts_thread)); 431 if (ts->ts_flags & TSF_XFERABLE) { 432 tdq->tdq_transferable--; 433 ts->ts_flags &= ~TSF_XFERABLE; 434 } 435 if (ts->ts_runq == &tdq->tdq_timeshare) { 436 if (tdq->tdq_idx != tdq->tdq_ridx) 437 runq_remove_idx(ts->ts_runq, ts, &tdq->tdq_ridx); 438 else 439 runq_remove_idx(ts->ts_runq, ts, NULL); 440 /* 441 * For timeshare threads we update the priority here so 442 * the priority reflects the time we've been sleeping. 443 */ 444 ts->ts_ltick = ticks; 445 sched_pctcpu_update(ts); 446 sched_priority(ts->ts_thread); 447 } else 448 runq_remove(ts->ts_runq, ts); 449 } 450 451 /* 452 * Load is maintained for all threads RUNNING and ON_RUNQ. Add the load 453 * for this thread to the referenced thread queue. 454 */ 455 static void 456 tdq_load_add(struct tdq *tdq, struct td_sched *ts) 457 { 458 int class; 459 460 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 461 THREAD_LOCK_ASSERT(ts->ts_thread, MA_OWNED); 462 class = PRI_BASE(ts->ts_thread->td_pri_class); 463 tdq->tdq_load++; 464 CTR2(KTR_SCHED, "cpu %d load: %d", TDQ_ID(tdq), tdq->tdq_load); 465 if (class != PRI_ITHD && 466 (ts->ts_thread->td_proc->p_flag & P_NOLOAD) == 0) 467 tdq->tdq_sysload++; 468 } 469 470 /* 471 * Remove the load from a thread that is transitioning to a sleep state or 472 * exiting. 473 */ 474 static void 475 tdq_load_rem(struct tdq *tdq, struct td_sched *ts) 476 { 477 int class; 478 479 THREAD_LOCK_ASSERT(ts->ts_thread, MA_OWNED); 480 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 481 class = PRI_BASE(ts->ts_thread->td_pri_class); 482 if (class != PRI_ITHD && 483 (ts->ts_thread->td_proc->p_flag & P_NOLOAD) == 0) 484 tdq->tdq_sysload--; 485 KASSERT(tdq->tdq_load != 0, 486 ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq))); 487 tdq->tdq_load--; 488 CTR1(KTR_SCHED, "load: %d", tdq->tdq_load); 489 ts->ts_runq = NULL; 490 } 491 492 /* 493 * Set lowpri to its exact value by searching the run-queue and 494 * evaluating curthread. curthread may be passed as an optimization. 495 */ 496 static void 497 tdq_setlowpri(struct tdq *tdq, struct thread *ctd) 498 { 499 struct td_sched *ts; 500 struct thread *td; 501 502 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 503 if (ctd == NULL) 504 ctd = pcpu_find(TDQ_ID(tdq))->pc_curthread; 505 ts = tdq_choose(tdq); 506 if (ts) 507 td = ts->ts_thread; 508 if (ts == NULL || td->td_priority > ctd->td_priority) 509 tdq->tdq_lowpri = ctd->td_priority; 510 else 511 tdq->tdq_lowpri = td->td_priority; 512 } 513 514 #ifdef SMP 515 struct cpu_search { 516 cpumask_t cs_mask; /* Mask of valid cpus. */ 517 u_int cs_load; 518 u_int cs_cpu; 519 int cs_limit; /* Min priority for low min load for high. */ 520 }; 521 522 #define CPU_SEARCH_LOWEST 0x1 523 #define CPU_SEARCH_HIGHEST 0x2 524 #define CPU_SEARCH_BOTH (CPU_SEARCH_LOWEST|CPU_SEARCH_HIGHEST) 525 526 #define CPUMASK_FOREACH(cpu, mask) \ 527 for ((cpu) = 0; (cpu) < sizeof((mask)) * 8; (cpu)++) \ 528 if ((mask) & 1 << (cpu)) 529 530 __inline int cpu_search(struct cpu_group *cg, struct cpu_search *low, 531 struct cpu_search *high, const int match); 532 int cpu_search_lowest(struct cpu_group *cg, struct cpu_search *low); 533 int cpu_search_highest(struct cpu_group *cg, struct cpu_search *high); 534 int cpu_search_both(struct cpu_group *cg, struct cpu_search *low, 535 struct cpu_search *high); 536 537 /* 538 * This routine compares according to the match argument and should be 539 * reduced in actual instantiations via constant propagation and dead code 540 * elimination. 541 */ 542 static __inline int 543 cpu_compare(int cpu, struct cpu_search *low, struct cpu_search *high, 544 const int match) 545 { 546 struct tdq *tdq; 547 548 tdq = TDQ_CPU(cpu); 549 if (match & CPU_SEARCH_LOWEST) 550 if (low->cs_mask & (1 << cpu) && 551 tdq->tdq_load < low->cs_load && 552 tdq->tdq_lowpri > low->cs_limit) { 553 low->cs_cpu = cpu; 554 low->cs_load = tdq->tdq_load; 555 } 556 if (match & CPU_SEARCH_HIGHEST) 557 if (high->cs_mask & (1 << cpu) && 558 tdq->tdq_load >= high->cs_limit && 559 tdq->tdq_load > high->cs_load && 560 tdq->tdq_transferable) { 561 high->cs_cpu = cpu; 562 high->cs_load = tdq->tdq_load; 563 } 564 return (tdq->tdq_load); 565 } 566 567 /* 568 * Search the tree of cpu_groups for the lowest or highest loaded cpu 569 * according to the match argument. This routine actually compares the 570 * load on all paths through the tree and finds the least loaded cpu on 571 * the least loaded path, which may differ from the least loaded cpu in 572 * the system. This balances work among caches and busses. 573 * 574 * This inline is instantiated in three forms below using constants for the 575 * match argument. It is reduced to the minimum set for each case. It is 576 * also recursive to the depth of the tree. 577 */ 578 static inline int 579 cpu_search(struct cpu_group *cg, struct cpu_search *low, 580 struct cpu_search *high, const int match) 581 { 582 int total; 583 584 total = 0; 585 if (cg->cg_children) { 586 struct cpu_search lgroup; 587 struct cpu_search hgroup; 588 struct cpu_group *child; 589 u_int lload; 590 int hload; 591 int load; 592 int i; 593 594 lload = -1; 595 hload = -1; 596 for (i = 0; i < cg->cg_children; i++) { 597 child = &cg->cg_child[i]; 598 if (match & CPU_SEARCH_LOWEST) { 599 lgroup = *low; 600 lgroup.cs_load = -1; 601 } 602 if (match & CPU_SEARCH_HIGHEST) { 603 hgroup = *high; 604 lgroup.cs_load = 0; 605 } 606 switch (match) { 607 case CPU_SEARCH_LOWEST: 608 load = cpu_search_lowest(child, &lgroup); 609 break; 610 case CPU_SEARCH_HIGHEST: 611 load = cpu_search_highest(child, &hgroup); 612 break; 613 case CPU_SEARCH_BOTH: 614 load = cpu_search_both(child, &lgroup, &hgroup); 615 break; 616 } 617 total += load; 618 if (match & CPU_SEARCH_LOWEST) 619 if (load < lload || low->cs_cpu == -1) { 620 *low = lgroup; 621 lload = load; 622 } 623 if (match & CPU_SEARCH_HIGHEST) 624 if (load > hload || high->cs_cpu == -1) { 625 hload = load; 626 *high = hgroup; 627 } 628 } 629 } else { 630 int cpu; 631 632 CPUMASK_FOREACH(cpu, cg->cg_mask) 633 total += cpu_compare(cpu, low, high, match); 634 } 635 return (total); 636 } 637 638 /* 639 * cpu_search instantiations must pass constants to maintain the inline 640 * optimization. 641 */ 642 int 643 cpu_search_lowest(struct cpu_group *cg, struct cpu_search *low) 644 { 645 return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST); 646 } 647 648 int 649 cpu_search_highest(struct cpu_group *cg, struct cpu_search *high) 650 { 651 return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST); 652 } 653 654 int 655 cpu_search_both(struct cpu_group *cg, struct cpu_search *low, 656 struct cpu_search *high) 657 { 658 return cpu_search(cg, low, high, CPU_SEARCH_BOTH); 659 } 660 661 /* 662 * Find the cpu with the least load via the least loaded path that has a 663 * lowpri greater than pri pri. A pri of -1 indicates any priority is 664 * acceptable. 665 */ 666 static inline int 667 sched_lowest(struct cpu_group *cg, cpumask_t mask, int pri) 668 { 669 struct cpu_search low; 670 671 low.cs_cpu = -1; 672 low.cs_load = -1; 673 low.cs_mask = mask; 674 low.cs_limit = pri; 675 cpu_search_lowest(cg, &low); 676 return low.cs_cpu; 677 } 678 679 /* 680 * Find the cpu with the highest load via the highest loaded path. 681 */ 682 static inline int 683 sched_highest(struct cpu_group *cg, cpumask_t mask, int minload) 684 { 685 struct cpu_search high; 686 687 high.cs_cpu = -1; 688 high.cs_load = 0; 689 high.cs_mask = mask; 690 high.cs_limit = minload; 691 cpu_search_highest(cg, &high); 692 return high.cs_cpu; 693 } 694 695 /* 696 * Simultaneously find the highest and lowest loaded cpu reachable via 697 * cg. 698 */ 699 static inline void 700 sched_both(struct cpu_group *cg, cpumask_t mask, int *lowcpu, int *highcpu) 701 { 702 struct cpu_search high; 703 struct cpu_search low; 704 705 low.cs_cpu = -1; 706 low.cs_limit = -1; 707 low.cs_load = -1; 708 low.cs_mask = mask; 709 high.cs_load = 0; 710 high.cs_cpu = -1; 711 high.cs_limit = -1; 712 high.cs_mask = mask; 713 cpu_search_both(cg, &low, &high); 714 *lowcpu = low.cs_cpu; 715 *highcpu = high.cs_cpu; 716 return; 717 } 718 719 static void 720 sched_balance_group(struct cpu_group *cg) 721 { 722 cpumask_t mask; 723 int high; 724 int low; 725 int i; 726 727 mask = -1; 728 for (;;) { 729 sched_both(cg, mask, &low, &high); 730 if (low == high || low == -1 || high == -1) 731 break; 732 if (sched_balance_pair(TDQ_CPU(high), TDQ_CPU(low))) 733 break; 734 /* 735 * If we failed to move any threads determine which cpu 736 * to kick out of the set and try again. 737 */ 738 if (TDQ_CPU(high)->tdq_transferable == 0) 739 mask &= ~(1 << high); 740 else 741 mask &= ~(1 << low); 742 } 743 744 for (i = 0; i < cg->cg_children; i++) 745 sched_balance_group(&cg->cg_child[i]); 746 } 747 748 static void 749 sched_balance() 750 { 751 struct tdq *tdq; 752 753 /* 754 * Select a random time between .5 * balance_interval and 755 * 1.5 * balance_interval. 756 */ 757 balance_ticks = max(balance_interval / 2, 1); 758 balance_ticks += random() % balance_interval; 759 if (smp_started == 0 || rebalance == 0) 760 return; 761 tdq = TDQ_SELF(); 762 TDQ_UNLOCK(tdq); 763 sched_balance_group(cpu_top); 764 TDQ_LOCK(tdq); 765 } 766 767 /* 768 * Lock two thread queues using their address to maintain lock order. 769 */ 770 static void 771 tdq_lock_pair(struct tdq *one, struct tdq *two) 772 { 773 if (one < two) { 774 TDQ_LOCK(one); 775 TDQ_LOCK_FLAGS(two, MTX_DUPOK); 776 } else { 777 TDQ_LOCK(two); 778 TDQ_LOCK_FLAGS(one, MTX_DUPOK); 779 } 780 } 781 782 /* 783 * Unlock two thread queues. Order is not important here. 784 */ 785 static void 786 tdq_unlock_pair(struct tdq *one, struct tdq *two) 787 { 788 TDQ_UNLOCK(one); 789 TDQ_UNLOCK(two); 790 } 791 792 /* 793 * Transfer load between two imbalanced thread queues. 794 */ 795 static int 796 sched_balance_pair(struct tdq *high, struct tdq *low) 797 { 798 int transferable; 799 int high_load; 800 int low_load; 801 int moved; 802 int move; 803 int diff; 804 int i; 805 806 tdq_lock_pair(high, low); 807 transferable = high->tdq_transferable; 808 high_load = high->tdq_load; 809 low_load = low->tdq_load; 810 moved = 0; 811 /* 812 * Determine what the imbalance is and then adjust that to how many 813 * threads we actually have to give up (transferable). 814 */ 815 if (transferable != 0) { 816 diff = high_load - low_load; 817 move = diff / 2; 818 if (diff & 0x1) 819 move++; 820 move = min(move, transferable); 821 for (i = 0; i < move; i++) 822 moved += tdq_move(high, low); 823 /* 824 * IPI the target cpu to force it to reschedule with the new 825 * workload. 826 */ 827 ipi_selected(1 << TDQ_ID(low), IPI_PREEMPT); 828 } 829 tdq_unlock_pair(high, low); 830 return (moved); 831 } 832 833 /* 834 * Move a thread from one thread queue to another. 835 */ 836 static int 837 tdq_move(struct tdq *from, struct tdq *to) 838 { 839 struct td_sched *ts; 840 struct thread *td; 841 struct tdq *tdq; 842 int cpu; 843 844 TDQ_LOCK_ASSERT(from, MA_OWNED); 845 TDQ_LOCK_ASSERT(to, MA_OWNED); 846 847 tdq = from; 848 cpu = TDQ_ID(to); 849 ts = tdq_steal(tdq, cpu); 850 if (ts == NULL) 851 return (0); 852 td = ts->ts_thread; 853 /* 854 * Although the run queue is locked the thread may be blocked. Lock 855 * it to clear this and acquire the run-queue lock. 856 */ 857 thread_lock(td); 858 /* Drop recursive lock on from acquired via thread_lock(). */ 859 TDQ_UNLOCK(from); 860 sched_rem(td); 861 ts->ts_cpu = cpu; 862 td->td_lock = TDQ_LOCKPTR(to); 863 tdq_add(to, td, SRQ_YIELDING); 864 return (1); 865 } 866 867 /* 868 * This tdq has idled. Try to steal a thread from another cpu and switch 869 * to it. 870 */ 871 static int 872 tdq_idled(struct tdq *tdq) 873 { 874 struct cpu_group *cg; 875 struct tdq *steal; 876 cpumask_t mask; 877 int thresh; 878 int cpu; 879 880 if (smp_started == 0 || steal_idle == 0) 881 return (1); 882 mask = -1; 883 mask &= ~PCPU_GET(cpumask); 884 /* We don't want to be preempted while we're iterating. */ 885 spinlock_enter(); 886 for (cg = tdq->tdq_cg; cg != NULL; ) { 887 if ((cg->cg_flags & (CG_FLAG_HTT | CG_FLAG_THREAD)) == 0) 888 thresh = steal_thresh; 889 else 890 thresh = 1; 891 cpu = sched_highest(cg, mask, thresh); 892 if (cpu == -1) { 893 cg = cg->cg_parent; 894 continue; 895 } 896 steal = TDQ_CPU(cpu); 897 mask &= ~(1 << cpu); 898 tdq_lock_pair(tdq, steal); 899 if (steal->tdq_load < thresh || steal->tdq_transferable == 0) { 900 tdq_unlock_pair(tdq, steal); 901 continue; 902 } 903 /* 904 * If a thread was added while interrupts were disabled don't 905 * steal one here. If we fail to acquire one due to affinity 906 * restrictions loop again with this cpu removed from the 907 * set. 908 */ 909 if (tdq->tdq_load == 0 && tdq_move(steal, tdq) == 0) { 910 tdq_unlock_pair(tdq, steal); 911 continue; 912 } 913 spinlock_exit(); 914 TDQ_UNLOCK(steal); 915 mi_switch(SW_VOL, NULL); 916 thread_unlock(curthread); 917 918 return (0); 919 } 920 spinlock_exit(); 921 return (1); 922 } 923 924 /* 925 * Notify a remote cpu of new work. Sends an IPI if criteria are met. 926 */ 927 static void 928 tdq_notify(struct tdq *tdq, struct td_sched *ts) 929 { 930 int cpri; 931 int pri; 932 int cpu; 933 934 if (tdq->tdq_ipipending) 935 return; 936 cpu = ts->ts_cpu; 937 pri = ts->ts_thread->td_priority; 938 cpri = pcpu_find(cpu)->pc_curthread->td_priority; 939 if (!sched_shouldpreempt(pri, cpri, 1)) 940 return; 941 tdq->tdq_ipipending = 1; 942 ipi_selected(1 << cpu, IPI_PREEMPT); 943 } 944 945 /* 946 * Steals load from a timeshare queue. Honors the rotating queue head 947 * index. 948 */ 949 static struct td_sched * 950 runq_steal_from(struct runq *rq, int cpu, u_char start) 951 { 952 struct td_sched *ts; 953 struct rqbits *rqb; 954 struct rqhead *rqh; 955 int first; 956 int bit; 957 int pri; 958 int i; 959 960 rqb = &rq->rq_status; 961 bit = start & (RQB_BPW -1); 962 pri = 0; 963 first = 0; 964 again: 965 for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) { 966 if (rqb->rqb_bits[i] == 0) 967 continue; 968 if (bit != 0) { 969 for (pri = bit; pri < RQB_BPW; pri++) 970 if (rqb->rqb_bits[i] & (1ul << pri)) 971 break; 972 if (pri >= RQB_BPW) 973 continue; 974 } else 975 pri = RQB_FFS(rqb->rqb_bits[i]); 976 pri += (i << RQB_L2BPW); 977 rqh = &rq->rq_queues[pri]; 978 TAILQ_FOREACH(ts, rqh, ts_procq) { 979 if (first && THREAD_CAN_MIGRATE(ts->ts_thread) && 980 THREAD_CAN_SCHED(ts->ts_thread, cpu)) 981 return (ts); 982 first = 1; 983 } 984 } 985 if (start != 0) { 986 start = 0; 987 goto again; 988 } 989 990 return (NULL); 991 } 992 993 /* 994 * Steals load from a standard linear queue. 995 */ 996 static struct td_sched * 997 runq_steal(struct runq *rq, int cpu) 998 { 999 struct rqhead *rqh; 1000 struct rqbits *rqb; 1001 struct td_sched *ts; 1002 int word; 1003 int bit; 1004 1005 rqb = &rq->rq_status; 1006 for (word = 0; word < RQB_LEN; word++) { 1007 if (rqb->rqb_bits[word] == 0) 1008 continue; 1009 for (bit = 0; bit < RQB_BPW; bit++) { 1010 if ((rqb->rqb_bits[word] & (1ul << bit)) == 0) 1011 continue; 1012 rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)]; 1013 TAILQ_FOREACH(ts, rqh, ts_procq) 1014 if (THREAD_CAN_MIGRATE(ts->ts_thread) && 1015 THREAD_CAN_SCHED(ts->ts_thread, cpu)) 1016 return (ts); 1017 } 1018 } 1019 return (NULL); 1020 } 1021 1022 /* 1023 * Attempt to steal a thread in priority order from a thread queue. 1024 */ 1025 static struct td_sched * 1026 tdq_steal(struct tdq *tdq, int cpu) 1027 { 1028 struct td_sched *ts; 1029 1030 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 1031 if ((ts = runq_steal(&tdq->tdq_realtime, cpu)) != NULL) 1032 return (ts); 1033 if ((ts = runq_steal_from(&tdq->tdq_timeshare, cpu, tdq->tdq_ridx)) 1034 != NULL) 1035 return (ts); 1036 return (runq_steal(&tdq->tdq_idle, cpu)); 1037 } 1038 1039 /* 1040 * Sets the thread lock and ts_cpu to match the requested cpu. Unlocks the 1041 * current lock and returns with the assigned queue locked. 1042 */ 1043 static inline struct tdq * 1044 sched_setcpu(struct td_sched *ts, int cpu, int flags) 1045 { 1046 struct thread *td; 1047 struct tdq *tdq; 1048 1049 THREAD_LOCK_ASSERT(ts->ts_thread, MA_OWNED); 1050 1051 tdq = TDQ_CPU(cpu); 1052 td = ts->ts_thread; 1053 ts->ts_cpu = cpu; 1054 1055 /* If the lock matches just return the queue. */ 1056 if (td->td_lock == TDQ_LOCKPTR(tdq)) 1057 return (tdq); 1058 #ifdef notyet 1059 /* 1060 * If the thread isn't running its lockptr is a 1061 * turnstile or a sleepqueue. We can just lock_set without 1062 * blocking. 1063 */ 1064 if (TD_CAN_RUN(td)) { 1065 TDQ_LOCK(tdq); 1066 thread_lock_set(td, TDQ_LOCKPTR(tdq)); 1067 return (tdq); 1068 } 1069 #endif 1070 /* 1071 * The hard case, migration, we need to block the thread first to 1072 * prevent order reversals with other cpus locks. 1073 */ 1074 thread_lock_block(td); 1075 TDQ_LOCK(tdq); 1076 thread_lock_unblock(td, TDQ_LOCKPTR(tdq)); 1077 return (tdq); 1078 } 1079 1080 static int 1081 sched_pickcpu(struct td_sched *ts, int flags) 1082 { 1083 struct cpu_group *cg; 1084 struct thread *td; 1085 struct tdq *tdq; 1086 cpumask_t mask; 1087 int self; 1088 int pri; 1089 int cpu; 1090 1091 self = PCPU_GET(cpuid); 1092 td = ts->ts_thread; 1093 if (smp_started == 0) 1094 return (self); 1095 /* 1096 * Don't migrate a running thread from sched_switch(). 1097 */ 1098 if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td)) 1099 return (ts->ts_cpu); 1100 /* 1101 * Prefer to run interrupt threads on the processors that generate 1102 * the interrupt. 1103 */ 1104 if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) && 1105 curthread->td_intr_nesting_level) 1106 ts->ts_cpu = self; 1107 /* 1108 * If the thread can run on the last cpu and the affinity has not 1109 * expired or it is idle run it there. 1110 */ 1111 pri = td->td_priority; 1112 tdq = TDQ_CPU(ts->ts_cpu); 1113 if (THREAD_CAN_SCHED(td, ts->ts_cpu)) { 1114 if (tdq->tdq_lowpri > PRI_MIN_IDLE) 1115 return (ts->ts_cpu); 1116 if (SCHED_AFFINITY(ts, CG_SHARE_L2) && tdq->tdq_lowpri > pri) 1117 return (ts->ts_cpu); 1118 } 1119 /* 1120 * Search for the highest level in the tree that still has affinity. 1121 */ 1122 cg = NULL; 1123 for (cg = tdq->tdq_cg; cg != NULL; cg = cg->cg_parent) 1124 if (SCHED_AFFINITY(ts, cg->cg_level)) 1125 break; 1126 cpu = -1; 1127 mask = td->td_cpuset->cs_mask.__bits[0]; 1128 if (cg) 1129 cpu = sched_lowest(cg, mask, pri); 1130 if (cpu == -1) 1131 cpu = sched_lowest(cpu_top, mask, -1); 1132 /* 1133 * Compare the lowest loaded cpu to current cpu. 1134 */ 1135 if (THREAD_CAN_SCHED(td, self) && TDQ_CPU(self)->tdq_lowpri > pri && 1136 TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE) 1137 cpu = self; 1138 KASSERT(cpu != -1, ("sched_pickcpu: Failed to find a cpu.")); 1139 return (cpu); 1140 } 1141 #endif 1142 1143 /* 1144 * Pick the highest priority task we have and return it. 1145 */ 1146 static struct td_sched * 1147 tdq_choose(struct tdq *tdq) 1148 { 1149 struct td_sched *ts; 1150 1151 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 1152 ts = runq_choose(&tdq->tdq_realtime); 1153 if (ts != NULL) 1154 return (ts); 1155 ts = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx); 1156 if (ts != NULL) { 1157 KASSERT(ts->ts_thread->td_priority >= PRI_MIN_TIMESHARE, 1158 ("tdq_choose: Invalid priority on timeshare queue %d", 1159 ts->ts_thread->td_priority)); 1160 return (ts); 1161 } 1162 1163 ts = runq_choose(&tdq->tdq_idle); 1164 if (ts != NULL) { 1165 KASSERT(ts->ts_thread->td_priority >= PRI_MIN_IDLE, 1166 ("tdq_choose: Invalid priority on idle queue %d", 1167 ts->ts_thread->td_priority)); 1168 return (ts); 1169 } 1170 1171 return (NULL); 1172 } 1173 1174 /* 1175 * Initialize a thread queue. 1176 */ 1177 static void 1178 tdq_setup(struct tdq *tdq) 1179 { 1180 1181 if (bootverbose) 1182 printf("ULE: setup cpu %d\n", TDQ_ID(tdq)); 1183 runq_init(&tdq->tdq_realtime); 1184 runq_init(&tdq->tdq_timeshare); 1185 runq_init(&tdq->tdq_idle); 1186 snprintf(tdq->tdq_name, sizeof(tdq->tdq_name), 1187 "sched lock %d", (int)TDQ_ID(tdq)); 1188 mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock", 1189 MTX_SPIN | MTX_RECURSE); 1190 } 1191 1192 #ifdef SMP 1193 static void 1194 sched_setup_smp(void) 1195 { 1196 struct tdq *tdq; 1197 int i; 1198 1199 cpu_top = smp_topo(); 1200 for (i = 0; i < MAXCPU; i++) { 1201 if (CPU_ABSENT(i)) 1202 continue; 1203 tdq = TDQ_CPU(i); 1204 tdq_setup(tdq); 1205 tdq->tdq_cg = smp_topo_find(cpu_top, i); 1206 if (tdq->tdq_cg == NULL) 1207 panic("Can't find cpu group for %d\n", i); 1208 } 1209 balance_tdq = TDQ_SELF(); 1210 sched_balance(); 1211 } 1212 #endif 1213 1214 /* 1215 * Setup the thread queues and initialize the topology based on MD 1216 * information. 1217 */ 1218 static void 1219 sched_setup(void *dummy) 1220 { 1221 struct tdq *tdq; 1222 1223 tdq = TDQ_SELF(); 1224 #ifdef SMP 1225 sched_setup_smp(); 1226 #else 1227 tdq_setup(tdq); 1228 #endif 1229 /* 1230 * To avoid divide-by-zero, we set realstathz a dummy value 1231 * in case which sched_clock() called before sched_initticks(). 1232 */ 1233 realstathz = hz; 1234 sched_slice = (realstathz/10); /* ~100ms */ 1235 tickincr = 1 << SCHED_TICK_SHIFT; 1236 1237 /* Add thread0's load since it's running. */ 1238 TDQ_LOCK(tdq); 1239 thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF()); 1240 tdq_load_add(tdq, &td_sched0); 1241 tdq->tdq_lowpri = thread0.td_priority; 1242 TDQ_UNLOCK(tdq); 1243 } 1244 1245 /* 1246 * This routine determines the tickincr after stathz and hz are setup. 1247 */ 1248 /* ARGSUSED */ 1249 static void 1250 sched_initticks(void *dummy) 1251 { 1252 int incr; 1253 1254 realstathz = stathz ? stathz : hz; 1255 sched_slice = (realstathz/10); /* ~100ms */ 1256 1257 /* 1258 * tickincr is shifted out by 10 to avoid rounding errors due to 1259 * hz not being evenly divisible by stathz on all platforms. 1260 */ 1261 incr = (hz << SCHED_TICK_SHIFT) / realstathz; 1262 /* 1263 * This does not work for values of stathz that are more than 1264 * 1 << SCHED_TICK_SHIFT * hz. In practice this does not happen. 1265 */ 1266 if (incr == 0) 1267 incr = 1; 1268 tickincr = incr; 1269 #ifdef SMP 1270 /* 1271 * Set the default balance interval now that we know 1272 * what realstathz is. 1273 */ 1274 balance_interval = realstathz; 1275 /* 1276 * Set steal thresh to log2(mp_ncpu) but no greater than 4. This 1277 * prevents excess thrashing on large machines and excess idle on 1278 * smaller machines. 1279 */ 1280 steal_thresh = min(ffs(mp_ncpus) - 1, 3); 1281 affinity = SCHED_AFFINITY_DEFAULT; 1282 #endif 1283 } 1284 1285 1286 /* 1287 * This is the core of the interactivity algorithm. Determines a score based 1288 * on past behavior. It is the ratio of sleep time to run time scaled to 1289 * a [0, 100] integer. This is the voluntary sleep time of a process, which 1290 * differs from the cpu usage because it does not account for time spent 1291 * waiting on a run-queue. Would be prettier if we had floating point. 1292 */ 1293 static int 1294 sched_interact_score(struct thread *td) 1295 { 1296 struct td_sched *ts; 1297 int div; 1298 1299 ts = td->td_sched; 1300 /* 1301 * The score is only needed if this is likely to be an interactive 1302 * task. Don't go through the expense of computing it if there's 1303 * no chance. 1304 */ 1305 if (sched_interact <= SCHED_INTERACT_HALF && 1306 ts->ts_runtime >= ts->ts_slptime) 1307 return (SCHED_INTERACT_HALF); 1308 1309 if (ts->ts_runtime > ts->ts_slptime) { 1310 div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF); 1311 return (SCHED_INTERACT_HALF + 1312 (SCHED_INTERACT_HALF - (ts->ts_slptime / div))); 1313 } 1314 if (ts->ts_slptime > ts->ts_runtime) { 1315 div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF); 1316 return (ts->ts_runtime / div); 1317 } 1318 /* runtime == slptime */ 1319 if (ts->ts_runtime) 1320 return (SCHED_INTERACT_HALF); 1321 1322 /* 1323 * This can happen if slptime and runtime are 0. 1324 */ 1325 return (0); 1326 1327 } 1328 1329 /* 1330 * Scale the scheduling priority according to the "interactivity" of this 1331 * process. 1332 */ 1333 static void 1334 sched_priority(struct thread *td) 1335 { 1336 int score; 1337 int pri; 1338 1339 if (td->td_pri_class != PRI_TIMESHARE) 1340 return; 1341 /* 1342 * If the score is interactive we place the thread in the realtime 1343 * queue with a priority that is less than kernel and interrupt 1344 * priorities. These threads are not subject to nice restrictions. 1345 * 1346 * Scores greater than this are placed on the normal timeshare queue 1347 * where the priority is partially decided by the most recent cpu 1348 * utilization and the rest is decided by nice value. 1349 * 1350 * The nice value of the process has a linear effect on the calculated 1351 * score. Negative nice values make it easier for a thread to be 1352 * considered interactive. 1353 */ 1354 score = imax(0, sched_interact_score(td) - td->td_proc->p_nice); 1355 if (score < sched_interact) { 1356 pri = PRI_MIN_REALTIME; 1357 pri += ((PRI_MAX_REALTIME - PRI_MIN_REALTIME) / sched_interact) 1358 * score; 1359 KASSERT(pri >= PRI_MIN_REALTIME && pri <= PRI_MAX_REALTIME, 1360 ("sched_priority: invalid interactive priority %d score %d", 1361 pri, score)); 1362 } else { 1363 pri = SCHED_PRI_MIN; 1364 if (td->td_sched->ts_ticks) 1365 pri += SCHED_PRI_TICKS(td->td_sched); 1366 pri += SCHED_PRI_NICE(td->td_proc->p_nice); 1367 KASSERT(pri >= PRI_MIN_TIMESHARE && pri <= PRI_MAX_TIMESHARE, 1368 ("sched_priority: invalid priority %d: nice %d, " 1369 "ticks %d ftick %d ltick %d tick pri %d", 1370 pri, td->td_proc->p_nice, td->td_sched->ts_ticks, 1371 td->td_sched->ts_ftick, td->td_sched->ts_ltick, 1372 SCHED_PRI_TICKS(td->td_sched))); 1373 } 1374 sched_user_prio(td, pri); 1375 1376 return; 1377 } 1378 1379 /* 1380 * This routine enforces a maximum limit on the amount of scheduling history 1381 * kept. It is called after either the slptime or runtime is adjusted. This 1382 * function is ugly due to integer math. 1383 */ 1384 static void 1385 sched_interact_update(struct thread *td) 1386 { 1387 struct td_sched *ts; 1388 u_int sum; 1389 1390 ts = td->td_sched; 1391 sum = ts->ts_runtime + ts->ts_slptime; 1392 if (sum < SCHED_SLP_RUN_MAX) 1393 return; 1394 /* 1395 * This only happens from two places: 1396 * 1) We have added an unusual amount of run time from fork_exit. 1397 * 2) We have added an unusual amount of sleep time from sched_sleep(). 1398 */ 1399 if (sum > SCHED_SLP_RUN_MAX * 2) { 1400 if (ts->ts_runtime > ts->ts_slptime) { 1401 ts->ts_runtime = SCHED_SLP_RUN_MAX; 1402 ts->ts_slptime = 1; 1403 } else { 1404 ts->ts_slptime = SCHED_SLP_RUN_MAX; 1405 ts->ts_runtime = 1; 1406 } 1407 return; 1408 } 1409 /* 1410 * If we have exceeded by more than 1/5th then the algorithm below 1411 * will not bring us back into range. Dividing by two here forces 1412 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX] 1413 */ 1414 if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) { 1415 ts->ts_runtime /= 2; 1416 ts->ts_slptime /= 2; 1417 return; 1418 } 1419 ts->ts_runtime = (ts->ts_runtime / 5) * 4; 1420 ts->ts_slptime = (ts->ts_slptime / 5) * 4; 1421 } 1422 1423 /* 1424 * Scale back the interactivity history when a child thread is created. The 1425 * history is inherited from the parent but the thread may behave totally 1426 * differently. For example, a shell spawning a compiler process. We want 1427 * to learn that the compiler is behaving badly very quickly. 1428 */ 1429 static void 1430 sched_interact_fork(struct thread *td) 1431 { 1432 int ratio; 1433 int sum; 1434 1435 sum = td->td_sched->ts_runtime + td->td_sched->ts_slptime; 1436 if (sum > SCHED_SLP_RUN_FORK) { 1437 ratio = sum / SCHED_SLP_RUN_FORK; 1438 td->td_sched->ts_runtime /= ratio; 1439 td->td_sched->ts_slptime /= ratio; 1440 } 1441 } 1442 1443 /* 1444 * Called from proc0_init() to setup the scheduler fields. 1445 */ 1446 void 1447 schedinit(void) 1448 { 1449 1450 /* 1451 * Set up the scheduler specific parts of proc0. 1452 */ 1453 proc0.p_sched = NULL; /* XXX */ 1454 thread0.td_sched = &td_sched0; 1455 td_sched0.ts_ltick = ticks; 1456 td_sched0.ts_ftick = ticks; 1457 td_sched0.ts_thread = &thread0; 1458 } 1459 1460 /* 1461 * This is only somewhat accurate since given many processes of the same 1462 * priority they will switch when their slices run out, which will be 1463 * at most sched_slice stathz ticks. 1464 */ 1465 int 1466 sched_rr_interval(void) 1467 { 1468 1469 /* Convert sched_slice to hz */ 1470 return (hz/(realstathz/sched_slice)); 1471 } 1472 1473 /* 1474 * Update the percent cpu tracking information when it is requested or 1475 * the total history exceeds the maximum. We keep a sliding history of 1476 * tick counts that slowly decays. This is less precise than the 4BSD 1477 * mechanism since it happens with less regular and frequent events. 1478 */ 1479 static void 1480 sched_pctcpu_update(struct td_sched *ts) 1481 { 1482 1483 if (ts->ts_ticks == 0) 1484 return; 1485 if (ticks - (hz / 10) < ts->ts_ltick && 1486 SCHED_TICK_TOTAL(ts) < SCHED_TICK_MAX) 1487 return; 1488 /* 1489 * Adjust counters and watermark for pctcpu calc. 1490 */ 1491 if (ts->ts_ltick > ticks - SCHED_TICK_TARG) 1492 ts->ts_ticks = (ts->ts_ticks / (ticks - ts->ts_ftick)) * 1493 SCHED_TICK_TARG; 1494 else 1495 ts->ts_ticks = 0; 1496 ts->ts_ltick = ticks; 1497 ts->ts_ftick = ts->ts_ltick - SCHED_TICK_TARG; 1498 } 1499 1500 /* 1501 * Adjust the priority of a thread. Move it to the appropriate run-queue 1502 * if necessary. This is the back-end for several priority related 1503 * functions. 1504 */ 1505 static void 1506 sched_thread_priority(struct thread *td, u_char prio) 1507 { 1508 struct td_sched *ts; 1509 1510 CTR6(KTR_SCHED, "sched_prio: %p(%s) prio %d newprio %d by %p(%s)", 1511 td, td->td_name, td->td_priority, prio, curthread, 1512 curthread->td_name); 1513 ts = td->td_sched; 1514 THREAD_LOCK_ASSERT(td, MA_OWNED); 1515 if (td->td_priority == prio) 1516 return; 1517 1518 if (TD_ON_RUNQ(td) && prio < td->td_priority) { 1519 /* 1520 * If the priority has been elevated due to priority 1521 * propagation, we may have to move ourselves to a new 1522 * queue. This could be optimized to not re-add in some 1523 * cases. 1524 */ 1525 sched_rem(td); 1526 td->td_priority = prio; 1527 sched_add(td, SRQ_BORROWING); 1528 } else if (TD_IS_RUNNING(td)) { 1529 struct tdq *tdq; 1530 int oldpri; 1531 1532 tdq = TDQ_CPU(ts->ts_cpu); 1533 oldpri = td->td_priority; 1534 td->td_priority = prio; 1535 if (prio < tdq->tdq_lowpri) 1536 tdq->tdq_lowpri = prio; 1537 else if (tdq->tdq_lowpri == oldpri) 1538 tdq_setlowpri(tdq, td); 1539 } else 1540 td->td_priority = prio; 1541 } 1542 1543 /* 1544 * Update a thread's priority when it is lent another thread's 1545 * priority. 1546 */ 1547 void 1548 sched_lend_prio(struct thread *td, u_char prio) 1549 { 1550 1551 td->td_flags |= TDF_BORROWING; 1552 sched_thread_priority(td, prio); 1553 } 1554 1555 /* 1556 * Restore a thread's priority when priority propagation is 1557 * over. The prio argument is the minimum priority the thread 1558 * needs to have to satisfy other possible priority lending 1559 * requests. If the thread's regular priority is less 1560 * important than prio, the thread will keep a priority boost 1561 * of prio. 1562 */ 1563 void 1564 sched_unlend_prio(struct thread *td, u_char prio) 1565 { 1566 u_char base_pri; 1567 1568 if (td->td_base_pri >= PRI_MIN_TIMESHARE && 1569 td->td_base_pri <= PRI_MAX_TIMESHARE) 1570 base_pri = td->td_user_pri; 1571 else 1572 base_pri = td->td_base_pri; 1573 if (prio >= base_pri) { 1574 td->td_flags &= ~TDF_BORROWING; 1575 sched_thread_priority(td, base_pri); 1576 } else 1577 sched_lend_prio(td, prio); 1578 } 1579 1580 /* 1581 * Standard entry for setting the priority to an absolute value. 1582 */ 1583 void 1584 sched_prio(struct thread *td, u_char prio) 1585 { 1586 u_char oldprio; 1587 1588 /* First, update the base priority. */ 1589 td->td_base_pri = prio; 1590 1591 /* 1592 * If the thread is borrowing another thread's priority, don't 1593 * ever lower the priority. 1594 */ 1595 if (td->td_flags & TDF_BORROWING && td->td_priority < prio) 1596 return; 1597 1598 /* Change the real priority. */ 1599 oldprio = td->td_priority; 1600 sched_thread_priority(td, prio); 1601 1602 /* 1603 * If the thread is on a turnstile, then let the turnstile update 1604 * its state. 1605 */ 1606 if (TD_ON_LOCK(td) && oldprio != prio) 1607 turnstile_adjust(td, oldprio); 1608 } 1609 1610 /* 1611 * Set the base user priority, does not effect current running priority. 1612 */ 1613 void 1614 sched_user_prio(struct thread *td, u_char prio) 1615 { 1616 u_char oldprio; 1617 1618 td->td_base_user_pri = prio; 1619 if (td->td_flags & TDF_UBORROWING && td->td_user_pri <= prio) 1620 return; 1621 oldprio = td->td_user_pri; 1622 td->td_user_pri = prio; 1623 } 1624 1625 void 1626 sched_lend_user_prio(struct thread *td, u_char prio) 1627 { 1628 u_char oldprio; 1629 1630 THREAD_LOCK_ASSERT(td, MA_OWNED); 1631 td->td_flags |= TDF_UBORROWING; 1632 oldprio = td->td_user_pri; 1633 td->td_user_pri = prio; 1634 } 1635 1636 void 1637 sched_unlend_user_prio(struct thread *td, u_char prio) 1638 { 1639 u_char base_pri; 1640 1641 THREAD_LOCK_ASSERT(td, MA_OWNED); 1642 base_pri = td->td_base_user_pri; 1643 if (prio >= base_pri) { 1644 td->td_flags &= ~TDF_UBORROWING; 1645 sched_user_prio(td, base_pri); 1646 } else { 1647 sched_lend_user_prio(td, prio); 1648 } 1649 } 1650 1651 /* 1652 * Add the thread passed as 'newtd' to the run queue before selecting 1653 * the next thread to run. This is only used for KSE. 1654 */ 1655 static void 1656 sched_switchin(struct tdq *tdq, struct thread *td) 1657 { 1658 #ifdef SMP 1659 spinlock_enter(); 1660 TDQ_UNLOCK(tdq); 1661 thread_lock(td); 1662 spinlock_exit(); 1663 sched_setcpu(td->td_sched, TDQ_ID(tdq), SRQ_YIELDING); 1664 #else 1665 td->td_lock = TDQ_LOCKPTR(tdq); 1666 #endif 1667 tdq_add(tdq, td, SRQ_YIELDING); 1668 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1669 } 1670 1671 /* 1672 * Block a thread for switching. Similar to thread_block() but does not 1673 * bump the spin count. 1674 */ 1675 static inline struct mtx * 1676 thread_block_switch(struct thread *td) 1677 { 1678 struct mtx *lock; 1679 1680 THREAD_LOCK_ASSERT(td, MA_OWNED); 1681 lock = td->td_lock; 1682 td->td_lock = &blocked_lock; 1683 mtx_unlock_spin(lock); 1684 1685 return (lock); 1686 } 1687 1688 /* 1689 * Handle migration from sched_switch(). This happens only for 1690 * cpu binding. 1691 */ 1692 static struct mtx * 1693 sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags) 1694 { 1695 struct tdq *tdn; 1696 1697 tdn = TDQ_CPU(td->td_sched->ts_cpu); 1698 #ifdef SMP 1699 /* 1700 * Do the lock dance required to avoid LOR. We grab an extra 1701 * spinlock nesting to prevent preemption while we're 1702 * not holding either run-queue lock. 1703 */ 1704 spinlock_enter(); 1705 thread_block_switch(td); /* This releases the lock on tdq. */ 1706 TDQ_LOCK(tdn); 1707 tdq_add(tdn, td, flags); 1708 tdq_notify(tdn, td->td_sched); 1709 /* 1710 * After we unlock tdn the new cpu still can't switch into this 1711 * thread until we've unblocked it in cpu_switch(). The lock 1712 * pointers may match in the case of HTT cores. Don't unlock here 1713 * or we can deadlock when the other CPU runs the IPI handler. 1714 */ 1715 if (TDQ_LOCKPTR(tdn) != TDQ_LOCKPTR(tdq)) { 1716 TDQ_UNLOCK(tdn); 1717 TDQ_LOCK(tdq); 1718 } 1719 spinlock_exit(); 1720 #endif 1721 return (TDQ_LOCKPTR(tdn)); 1722 } 1723 1724 /* 1725 * Release a thread that was blocked with thread_block_switch(). 1726 */ 1727 static inline void 1728 thread_unblock_switch(struct thread *td, struct mtx *mtx) 1729 { 1730 atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock, 1731 (uintptr_t)mtx); 1732 } 1733 1734 /* 1735 * Switch threads. This function has to handle threads coming in while 1736 * blocked for some reason, running, or idle. It also must deal with 1737 * migrating a thread from one queue to another as running threads may 1738 * be assigned elsewhere via binding. 1739 */ 1740 void 1741 sched_switch(struct thread *td, struct thread *newtd, int flags) 1742 { 1743 struct tdq *tdq; 1744 struct td_sched *ts; 1745 struct mtx *mtx; 1746 int srqflag; 1747 int cpuid; 1748 1749 THREAD_LOCK_ASSERT(td, MA_OWNED); 1750 1751 cpuid = PCPU_GET(cpuid); 1752 tdq = TDQ_CPU(cpuid); 1753 ts = td->td_sched; 1754 mtx = td->td_lock; 1755 ts->ts_rltick = ticks; 1756 td->td_lastcpu = td->td_oncpu; 1757 td->td_oncpu = NOCPU; 1758 td->td_flags &= ~TDF_NEEDRESCHED; 1759 td->td_owepreempt = 0; 1760 /* 1761 * The lock pointer in an idle thread should never change. Reset it 1762 * to CAN_RUN as well. 1763 */ 1764 if (TD_IS_IDLETHREAD(td)) { 1765 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1766 TD_SET_CAN_RUN(td); 1767 } else if (TD_IS_RUNNING(td)) { 1768 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1769 tdq_load_rem(tdq, ts); 1770 srqflag = (flags & SW_PREEMPT) ? 1771 SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED : 1772 SRQ_OURSELF|SRQ_YIELDING; 1773 if (ts->ts_cpu == cpuid) 1774 tdq_add(tdq, td, srqflag); 1775 else 1776 mtx = sched_switch_migrate(tdq, td, srqflag); 1777 } else { 1778 /* This thread must be going to sleep. */ 1779 TDQ_LOCK(tdq); 1780 mtx = thread_block_switch(td); 1781 tdq_load_rem(tdq, ts); 1782 } 1783 /* 1784 * We enter here with the thread blocked and assigned to the 1785 * appropriate cpu run-queue or sleep-queue and with the current 1786 * thread-queue locked. 1787 */ 1788 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED); 1789 /* 1790 * If KSE assigned a new thread just add it here and let choosethread 1791 * select the best one. 1792 */ 1793 if (newtd != NULL) 1794 sched_switchin(tdq, newtd); 1795 newtd = choosethread(); 1796 /* 1797 * Call the MD code to switch contexts if necessary. 1798 */ 1799 if (td != newtd) { 1800 #ifdef HWPMC_HOOKS 1801 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1802 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT); 1803 #endif 1804 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object); 1805 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd; 1806 cpu_switch(td, newtd, mtx); 1807 /* 1808 * We may return from cpu_switch on a different cpu. However, 1809 * we always return with td_lock pointing to the current cpu's 1810 * run queue lock. 1811 */ 1812 cpuid = PCPU_GET(cpuid); 1813 tdq = TDQ_CPU(cpuid); 1814 lock_profile_obtain_lock_success( 1815 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__); 1816 #ifdef HWPMC_HOOKS 1817 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1818 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN); 1819 #endif 1820 } else 1821 thread_unblock_switch(td, mtx); 1822 /* 1823 * We should always get here with the lowest priority td possible. 1824 */ 1825 tdq->tdq_lowpri = td->td_priority; 1826 /* 1827 * Assert that all went well and return. 1828 */ 1829 TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED); 1830 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1831 td->td_oncpu = cpuid; 1832 } 1833 1834 /* 1835 * Adjust thread priorities as a result of a nice request. 1836 */ 1837 void 1838 sched_nice(struct proc *p, int nice) 1839 { 1840 struct thread *td; 1841 1842 PROC_LOCK_ASSERT(p, MA_OWNED); 1843 PROC_SLOCK_ASSERT(p, MA_OWNED); 1844 1845 p->p_nice = nice; 1846 FOREACH_THREAD_IN_PROC(p, td) { 1847 thread_lock(td); 1848 sched_priority(td); 1849 sched_prio(td, td->td_base_user_pri); 1850 thread_unlock(td); 1851 } 1852 } 1853 1854 /* 1855 * Record the sleep time for the interactivity scorer. 1856 */ 1857 void 1858 sched_sleep(struct thread *td) 1859 { 1860 1861 THREAD_LOCK_ASSERT(td, MA_OWNED); 1862 1863 td->td_slptick = ticks; 1864 } 1865 1866 /* 1867 * Schedule a thread to resume execution and record how long it voluntarily 1868 * slept. We also update the pctcpu, interactivity, and priority. 1869 */ 1870 void 1871 sched_wakeup(struct thread *td) 1872 { 1873 struct td_sched *ts; 1874 int slptick; 1875 1876 THREAD_LOCK_ASSERT(td, MA_OWNED); 1877 ts = td->td_sched; 1878 /* 1879 * If we slept for more than a tick update our interactivity and 1880 * priority. 1881 */ 1882 slptick = td->td_slptick; 1883 td->td_slptick = 0; 1884 if (slptick && slptick != ticks) { 1885 u_int hzticks; 1886 1887 hzticks = (ticks - slptick) << SCHED_TICK_SHIFT; 1888 ts->ts_slptime += hzticks; 1889 sched_interact_update(td); 1890 sched_pctcpu_update(ts); 1891 sched_priority(td); 1892 } 1893 /* Reset the slice value after we sleep. */ 1894 ts->ts_slice = sched_slice; 1895 sched_add(td, SRQ_BORING); 1896 } 1897 1898 /* 1899 * Penalize the parent for creating a new child and initialize the child's 1900 * priority. 1901 */ 1902 void 1903 sched_fork(struct thread *td, struct thread *child) 1904 { 1905 THREAD_LOCK_ASSERT(td, MA_OWNED); 1906 sched_fork_thread(td, child); 1907 /* 1908 * Penalize the parent and child for forking. 1909 */ 1910 sched_interact_fork(child); 1911 sched_priority(child); 1912 td->td_sched->ts_runtime += tickincr; 1913 sched_interact_update(td); 1914 sched_priority(td); 1915 } 1916 1917 /* 1918 * Fork a new thread, may be within the same process. 1919 */ 1920 void 1921 sched_fork_thread(struct thread *td, struct thread *child) 1922 { 1923 struct td_sched *ts; 1924 struct td_sched *ts2; 1925 1926 /* 1927 * Initialize child. 1928 */ 1929 THREAD_LOCK_ASSERT(td, MA_OWNED); 1930 sched_newthread(child); 1931 child->td_lock = TDQ_LOCKPTR(TDQ_SELF()); 1932 child->td_cpuset = cpuset_ref(td->td_cpuset); 1933 ts = td->td_sched; 1934 ts2 = child->td_sched; 1935 ts2->ts_cpu = ts->ts_cpu; 1936 ts2->ts_runq = NULL; 1937 /* 1938 * Grab our parents cpu estimation information and priority. 1939 */ 1940 ts2->ts_ticks = ts->ts_ticks; 1941 ts2->ts_ltick = ts->ts_ltick; 1942 ts2->ts_ftick = ts->ts_ftick; 1943 child->td_user_pri = td->td_user_pri; 1944 child->td_base_user_pri = td->td_base_user_pri; 1945 /* 1946 * And update interactivity score. 1947 */ 1948 ts2->ts_slptime = ts->ts_slptime; 1949 ts2->ts_runtime = ts->ts_runtime; 1950 ts2->ts_slice = 1; /* Attempt to quickly learn interactivity. */ 1951 } 1952 1953 /* 1954 * Adjust the priority class of a thread. 1955 */ 1956 void 1957 sched_class(struct thread *td, int class) 1958 { 1959 1960 THREAD_LOCK_ASSERT(td, MA_OWNED); 1961 if (td->td_pri_class == class) 1962 return; 1963 /* 1964 * On SMP if we're on the RUNQ we must adjust the transferable 1965 * count because could be changing to or from an interrupt 1966 * class. 1967 */ 1968 if (TD_ON_RUNQ(td)) { 1969 struct tdq *tdq; 1970 1971 tdq = TDQ_CPU(td->td_sched->ts_cpu); 1972 if (THREAD_CAN_MIGRATE(td)) 1973 tdq->tdq_transferable--; 1974 td->td_pri_class = class; 1975 if (THREAD_CAN_MIGRATE(td)) 1976 tdq->tdq_transferable++; 1977 } 1978 td->td_pri_class = class; 1979 } 1980 1981 /* 1982 * Return some of the child's priority and interactivity to the parent. 1983 */ 1984 void 1985 sched_exit(struct proc *p, struct thread *child) 1986 { 1987 struct thread *td; 1988 1989 CTR3(KTR_SCHED, "sched_exit: %p(%s) prio %d", 1990 child, child->td_name, child->td_priority); 1991 1992 PROC_SLOCK_ASSERT(p, MA_OWNED); 1993 td = FIRST_THREAD_IN_PROC(p); 1994 sched_exit_thread(td, child); 1995 } 1996 1997 /* 1998 * Penalize another thread for the time spent on this one. This helps to 1999 * worsen the priority and interactivity of processes which schedule batch 2000 * jobs such as make. This has little effect on the make process itself but 2001 * causes new processes spawned by it to receive worse scores immediately. 2002 */ 2003 void 2004 sched_exit_thread(struct thread *td, struct thread *child) 2005 { 2006 2007 CTR3(KTR_SCHED, "sched_exit_thread: %p(%s) prio %d", 2008 child, child->td_name, child->td_priority); 2009 2010 #ifdef KSE 2011 /* 2012 * KSE forks and exits so often that this penalty causes short-lived 2013 * threads to always be non-interactive. This causes mozilla to 2014 * crawl under load. 2015 */ 2016 if ((td->td_pflags & TDP_SA) && td->td_proc == child->td_proc) 2017 return; 2018 #endif 2019 /* 2020 * Give the child's runtime to the parent without returning the 2021 * sleep time as a penalty to the parent. This causes shells that 2022 * launch expensive things to mark their children as expensive. 2023 */ 2024 thread_lock(td); 2025 td->td_sched->ts_runtime += child->td_sched->ts_runtime; 2026 sched_interact_update(td); 2027 sched_priority(td); 2028 thread_unlock(td); 2029 } 2030 2031 void 2032 sched_preempt(struct thread *td) 2033 { 2034 struct tdq *tdq; 2035 2036 thread_lock(td); 2037 tdq = TDQ_SELF(); 2038 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2039 tdq->tdq_ipipending = 0; 2040 if (td->td_priority > tdq->tdq_lowpri) { 2041 if (td->td_critnest > 1) 2042 td->td_owepreempt = 1; 2043 else 2044 mi_switch(SW_INVOL | SW_PREEMPT, NULL); 2045 } 2046 thread_unlock(td); 2047 } 2048 2049 /* 2050 * Fix priorities on return to user-space. Priorities may be elevated due 2051 * to static priorities in msleep() or similar. 2052 */ 2053 void 2054 sched_userret(struct thread *td) 2055 { 2056 /* 2057 * XXX we cheat slightly on the locking here to avoid locking in 2058 * the usual case. Setting td_priority here is essentially an 2059 * incomplete workaround for not setting it properly elsewhere. 2060 * Now that some interrupt handlers are threads, not setting it 2061 * properly elsewhere can clobber it in the window between setting 2062 * it here and returning to user mode, so don't waste time setting 2063 * it perfectly here. 2064 */ 2065 KASSERT((td->td_flags & TDF_BORROWING) == 0, 2066 ("thread with borrowed priority returning to userland")); 2067 if (td->td_priority != td->td_user_pri) { 2068 thread_lock(td); 2069 td->td_priority = td->td_user_pri; 2070 td->td_base_pri = td->td_user_pri; 2071 tdq_setlowpri(TDQ_SELF(), td); 2072 thread_unlock(td); 2073 } 2074 } 2075 2076 /* 2077 * Handle a stathz tick. This is really only relevant for timeshare 2078 * threads. 2079 */ 2080 void 2081 sched_clock(struct thread *td) 2082 { 2083 struct tdq *tdq; 2084 struct td_sched *ts; 2085 2086 THREAD_LOCK_ASSERT(td, MA_OWNED); 2087 tdq = TDQ_SELF(); 2088 #ifdef SMP 2089 /* 2090 * We run the long term load balancer infrequently on the first cpu. 2091 */ 2092 if (balance_tdq == tdq) { 2093 if (balance_ticks && --balance_ticks == 0) 2094 sched_balance(); 2095 } 2096 #endif 2097 /* 2098 * Advance the insert index once for each tick to ensure that all 2099 * threads get a chance to run. 2100 */ 2101 if (tdq->tdq_idx == tdq->tdq_ridx) { 2102 tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS; 2103 if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx])) 2104 tdq->tdq_ridx = tdq->tdq_idx; 2105 } 2106 ts = td->td_sched; 2107 if (td->td_pri_class & PRI_FIFO_BIT) 2108 return; 2109 if (td->td_pri_class == PRI_TIMESHARE) { 2110 /* 2111 * We used a tick; charge it to the thread so 2112 * that we can compute our interactivity. 2113 */ 2114 td->td_sched->ts_runtime += tickincr; 2115 sched_interact_update(td); 2116 } 2117 /* 2118 * We used up one time slice. 2119 */ 2120 if (--ts->ts_slice > 0) 2121 return; 2122 /* 2123 * We're out of time, recompute priorities and requeue. 2124 */ 2125 sched_priority(td); 2126 td->td_flags |= TDF_NEEDRESCHED; 2127 } 2128 2129 /* 2130 * Called once per hz tick. Used for cpu utilization information. This 2131 * is easier than trying to scale based on stathz. 2132 */ 2133 void 2134 sched_tick(void) 2135 { 2136 struct td_sched *ts; 2137 2138 ts = curthread->td_sched; 2139 /* Adjust ticks for pctcpu */ 2140 ts->ts_ticks += 1 << SCHED_TICK_SHIFT; 2141 ts->ts_ltick = ticks; 2142 /* 2143 * Update if we've exceeded our desired tick threshhold by over one 2144 * second. 2145 */ 2146 if (ts->ts_ftick + SCHED_TICK_MAX < ts->ts_ltick) 2147 sched_pctcpu_update(ts); 2148 } 2149 2150 /* 2151 * Return whether the current CPU has runnable tasks. Used for in-kernel 2152 * cooperative idle threads. 2153 */ 2154 int 2155 sched_runnable(void) 2156 { 2157 struct tdq *tdq; 2158 int load; 2159 2160 load = 1; 2161 2162 tdq = TDQ_SELF(); 2163 if ((curthread->td_flags & TDF_IDLETD) != 0) { 2164 if (tdq->tdq_load > 0) 2165 goto out; 2166 } else 2167 if (tdq->tdq_load - 1 > 0) 2168 goto out; 2169 load = 0; 2170 out: 2171 return (load); 2172 } 2173 2174 /* 2175 * Choose the highest priority thread to run. The thread is removed from 2176 * the run-queue while running however the load remains. For SMP we set 2177 * the tdq in the global idle bitmask if it idles here. 2178 */ 2179 struct thread * 2180 sched_choose(void) 2181 { 2182 struct td_sched *ts; 2183 struct tdq *tdq; 2184 2185 tdq = TDQ_SELF(); 2186 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2187 ts = tdq_choose(tdq); 2188 if (ts) { 2189 tdq_runq_rem(tdq, ts); 2190 return (ts->ts_thread); 2191 } 2192 return (PCPU_GET(idlethread)); 2193 } 2194 2195 /* 2196 * Set owepreempt if necessary. Preemption never happens directly in ULE, 2197 * we always request it once we exit a critical section. 2198 */ 2199 static inline void 2200 sched_setpreempt(struct thread *td) 2201 { 2202 struct thread *ctd; 2203 int cpri; 2204 int pri; 2205 2206 THREAD_LOCK_ASSERT(curthread, MA_OWNED); 2207 2208 ctd = curthread; 2209 pri = td->td_priority; 2210 cpri = ctd->td_priority; 2211 if (pri < cpri) 2212 ctd->td_flags |= TDF_NEEDRESCHED; 2213 if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd)) 2214 return; 2215 if (!sched_shouldpreempt(pri, cpri, 0)) 2216 return; 2217 ctd->td_owepreempt = 1; 2218 } 2219 2220 /* 2221 * Add a thread to a thread queue. Initializes priority, slice, runq, and 2222 * add it to the appropriate queue. This is the internal function called 2223 * when the tdq is predetermined. 2224 */ 2225 void 2226 tdq_add(struct tdq *tdq, struct thread *td, int flags) 2227 { 2228 struct td_sched *ts; 2229 int class; 2230 2231 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2232 KASSERT((td->td_inhibitors == 0), 2233 ("sched_add: trying to run inhibited thread")); 2234 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)), 2235 ("sched_add: bad thread state")); 2236 KASSERT(td->td_flags & TDF_INMEM, 2237 ("sched_add: thread swapped out")); 2238 2239 ts = td->td_sched; 2240 class = PRI_BASE(td->td_pri_class); 2241 TD_SET_RUNQ(td); 2242 if (ts->ts_slice == 0) 2243 ts->ts_slice = sched_slice; 2244 /* 2245 * Pick the run queue based on priority. 2246 */ 2247 if (td->td_priority <= PRI_MAX_REALTIME) 2248 ts->ts_runq = &tdq->tdq_realtime; 2249 else if (td->td_priority <= PRI_MAX_TIMESHARE) 2250 ts->ts_runq = &tdq->tdq_timeshare; 2251 else 2252 ts->ts_runq = &tdq->tdq_idle; 2253 if (td->td_priority < tdq->tdq_lowpri) 2254 tdq->tdq_lowpri = td->td_priority; 2255 tdq_runq_add(tdq, ts, flags); 2256 tdq_load_add(tdq, ts); 2257 } 2258 2259 /* 2260 * Select the target thread queue and add a thread to it. Request 2261 * preemption or IPI a remote processor if required. 2262 */ 2263 void 2264 sched_add(struct thread *td, int flags) 2265 { 2266 struct td_sched *ts; 2267 struct tdq *tdq; 2268 #ifdef SMP 2269 int cpuid; 2270 int cpu; 2271 #endif 2272 CTR5(KTR_SCHED, "sched_add: %p(%s) prio %d by %p(%s)", 2273 td, td->td_name, td->td_priority, curthread, 2274 curthread->td_name); 2275 THREAD_LOCK_ASSERT(td, MA_OWNED); 2276 ts = td->td_sched; 2277 /* 2278 * Recalculate the priority before we select the target cpu or 2279 * run-queue. 2280 */ 2281 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) 2282 sched_priority(td); 2283 #ifdef SMP 2284 cpuid = PCPU_GET(cpuid); 2285 /* 2286 * Pick the destination cpu and if it isn't ours transfer to the 2287 * target cpu. 2288 */ 2289 cpu = sched_pickcpu(ts, flags); 2290 tdq = sched_setcpu(ts, cpu, flags); 2291 tdq_add(tdq, td, flags); 2292 if (cpu != cpuid) { 2293 tdq_notify(tdq, ts); 2294 return; 2295 } 2296 #else 2297 tdq = TDQ_SELF(); 2298 TDQ_LOCK(tdq); 2299 /* 2300 * Now that the thread is moving to the run-queue, set the lock 2301 * to the scheduler's lock. 2302 */ 2303 thread_lock_set(td, TDQ_LOCKPTR(tdq)); 2304 tdq_add(tdq, td, flags); 2305 #endif 2306 if (!(flags & SRQ_YIELDING)) 2307 sched_setpreempt(td); 2308 } 2309 2310 /* 2311 * Remove a thread from a run-queue without running it. This is used 2312 * when we're stealing a thread from a remote queue. Otherwise all threads 2313 * exit by calling sched_exit_thread() and sched_throw() themselves. 2314 */ 2315 void 2316 sched_rem(struct thread *td) 2317 { 2318 struct tdq *tdq; 2319 struct td_sched *ts; 2320 2321 CTR5(KTR_SCHED, "sched_rem: %p(%s) prio %d by %p(%s)", 2322 td, td->td_name, td->td_priority, curthread, 2323 curthread->td_name); 2324 ts = td->td_sched; 2325 tdq = TDQ_CPU(ts->ts_cpu); 2326 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2327 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2328 KASSERT(TD_ON_RUNQ(td), 2329 ("sched_rem: thread not on run queue")); 2330 tdq_runq_rem(tdq, ts); 2331 tdq_load_rem(tdq, ts); 2332 TD_SET_CAN_RUN(td); 2333 if (td->td_priority == tdq->tdq_lowpri) 2334 tdq_setlowpri(tdq, NULL); 2335 } 2336 2337 /* 2338 * Fetch cpu utilization information. Updates on demand. 2339 */ 2340 fixpt_t 2341 sched_pctcpu(struct thread *td) 2342 { 2343 fixpt_t pctcpu; 2344 struct td_sched *ts; 2345 2346 pctcpu = 0; 2347 ts = td->td_sched; 2348 if (ts == NULL) 2349 return (0); 2350 2351 thread_lock(td); 2352 if (ts->ts_ticks) { 2353 int rtick; 2354 2355 sched_pctcpu_update(ts); 2356 /* How many rtick per second ? */ 2357 rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz); 2358 pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT; 2359 } 2360 thread_unlock(td); 2361 2362 return (pctcpu); 2363 } 2364 2365 /* 2366 * Enforce affinity settings for a thread. Called after adjustments to 2367 * cpumask. 2368 */ 2369 void 2370 sched_affinity(struct thread *td) 2371 { 2372 #ifdef SMP 2373 struct td_sched *ts; 2374 int cpu; 2375 2376 THREAD_LOCK_ASSERT(td, MA_OWNED); 2377 ts = td->td_sched; 2378 if (THREAD_CAN_SCHED(td, ts->ts_cpu)) 2379 return; 2380 if (!TD_IS_RUNNING(td)) 2381 return; 2382 td->td_flags |= TDF_NEEDRESCHED; 2383 if (!THREAD_CAN_MIGRATE(td)) 2384 return; 2385 /* 2386 * Assign the new cpu and force a switch before returning to 2387 * userspace. If the target thread is not running locally send 2388 * an ipi to force the issue. 2389 */ 2390 cpu = ts->ts_cpu; 2391 ts->ts_cpu = sched_pickcpu(ts, 0); 2392 if (cpu != PCPU_GET(cpuid)) 2393 ipi_selected(1 << cpu, IPI_PREEMPT); 2394 #endif 2395 } 2396 2397 /* 2398 * Bind a thread to a target cpu. 2399 */ 2400 void 2401 sched_bind(struct thread *td, int cpu) 2402 { 2403 struct td_sched *ts; 2404 2405 THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED); 2406 ts = td->td_sched; 2407 if (ts->ts_flags & TSF_BOUND) 2408 sched_unbind(td); 2409 ts->ts_flags |= TSF_BOUND; 2410 sched_pin(); 2411 if (PCPU_GET(cpuid) == cpu) 2412 return; 2413 ts->ts_cpu = cpu; 2414 /* When we return from mi_switch we'll be on the correct cpu. */ 2415 mi_switch(SW_VOL, NULL); 2416 } 2417 2418 /* 2419 * Release a bound thread. 2420 */ 2421 void 2422 sched_unbind(struct thread *td) 2423 { 2424 struct td_sched *ts; 2425 2426 THREAD_LOCK_ASSERT(td, MA_OWNED); 2427 ts = td->td_sched; 2428 if ((ts->ts_flags & TSF_BOUND) == 0) 2429 return; 2430 ts->ts_flags &= ~TSF_BOUND; 2431 sched_unpin(); 2432 } 2433 2434 int 2435 sched_is_bound(struct thread *td) 2436 { 2437 THREAD_LOCK_ASSERT(td, MA_OWNED); 2438 return (td->td_sched->ts_flags & TSF_BOUND); 2439 } 2440 2441 /* 2442 * Basic yield call. 2443 */ 2444 void 2445 sched_relinquish(struct thread *td) 2446 { 2447 thread_lock(td); 2448 SCHED_STAT_INC(switch_relinquish); 2449 mi_switch(SW_VOL, NULL); 2450 thread_unlock(td); 2451 } 2452 2453 /* 2454 * Return the total system load. 2455 */ 2456 int 2457 sched_load(void) 2458 { 2459 #ifdef SMP 2460 int total; 2461 int i; 2462 2463 total = 0; 2464 for (i = 0; i <= mp_maxid; i++) 2465 total += TDQ_CPU(i)->tdq_sysload; 2466 return (total); 2467 #else 2468 return (TDQ_SELF()->tdq_sysload); 2469 #endif 2470 } 2471 2472 int 2473 sched_sizeof_proc(void) 2474 { 2475 return (sizeof(struct proc)); 2476 } 2477 2478 int 2479 sched_sizeof_thread(void) 2480 { 2481 return (sizeof(struct thread) + sizeof(struct td_sched)); 2482 } 2483 2484 /* 2485 * The actual idle process. 2486 */ 2487 void 2488 sched_idletd(void *dummy) 2489 { 2490 struct thread *td; 2491 struct tdq *tdq; 2492 2493 td = curthread; 2494 tdq = TDQ_SELF(); 2495 mtx_assert(&Giant, MA_NOTOWNED); 2496 /* ULE relies on preemption for idle interruption. */ 2497 for (;;) { 2498 #ifdef SMP 2499 if (tdq_idled(tdq)) 2500 cpu_idle(); 2501 #else 2502 cpu_idle(); 2503 #endif 2504 } 2505 } 2506 2507 /* 2508 * A CPU is entering for the first time or a thread is exiting. 2509 */ 2510 void 2511 sched_throw(struct thread *td) 2512 { 2513 struct thread *newtd; 2514 struct tdq *tdq; 2515 2516 tdq = TDQ_SELF(); 2517 if (td == NULL) { 2518 /* Correct spinlock nesting and acquire the correct lock. */ 2519 TDQ_LOCK(tdq); 2520 spinlock_exit(); 2521 } else { 2522 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2523 tdq_load_rem(tdq, td->td_sched); 2524 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object); 2525 } 2526 KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count")); 2527 newtd = choosethread(); 2528 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd; 2529 PCPU_SET(switchtime, cpu_ticks()); 2530 PCPU_SET(switchticks, ticks); 2531 cpu_throw(td, newtd); /* doesn't return */ 2532 } 2533 2534 /* 2535 * This is called from fork_exit(). Just acquire the correct locks and 2536 * let fork do the rest of the work. 2537 */ 2538 void 2539 sched_fork_exit(struct thread *td) 2540 { 2541 struct td_sched *ts; 2542 struct tdq *tdq; 2543 int cpuid; 2544 2545 /* 2546 * Finish setting up thread glue so that it begins execution in a 2547 * non-nested critical section with the scheduler lock held. 2548 */ 2549 cpuid = PCPU_GET(cpuid); 2550 tdq = TDQ_CPU(cpuid); 2551 ts = td->td_sched; 2552 if (TD_IS_IDLETHREAD(td)) 2553 td->td_lock = TDQ_LOCKPTR(tdq); 2554 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2555 td->td_oncpu = cpuid; 2556 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED); 2557 lock_profile_obtain_lock_success( 2558 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__); 2559 tdq->tdq_lowpri = td->td_priority; 2560 } 2561 2562 static SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, 2563 "Scheduler"); 2564 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0, 2565 "Scheduler name"); 2566 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0, 2567 "Slice size for timeshare threads"); 2568 SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0, 2569 "Interactivity score threshold"); 2570 SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW, &preempt_thresh, 2571 0,"Min priority for preemption, lower priorities have greater precedence"); 2572 #ifdef SMP 2573 SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0, 2574 "Number of hz ticks to keep thread affinity for"); 2575 SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0, 2576 "Enables the long-term load balancer"); 2577 SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW, 2578 &balance_interval, 0, 2579 "Average frequency in stathz ticks to run the long-term balancer"); 2580 SYSCTL_INT(_kern_sched, OID_AUTO, steal_htt, CTLFLAG_RW, &steal_htt, 0, 2581 "Steals work from another hyper-threaded core on idle"); 2582 SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0, 2583 "Attempts to steal work from other cores before idling"); 2584 SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0, 2585 "Minimum load on remote cpu before we'll steal"); 2586 #endif 2587 2588 /* ps compat. All cpu percentages from ULE are weighted. */ 2589 static int ccpu = 0; 2590 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 2591 2592 2593 #define KERN_SWITCH_INCLUDE 1 2594 #include "kern/kern_switch.c" 2595