1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD 3 * 4 * Copyright (c) 2002-2007, Jeffrey Roberson <jeff@freebsd.org> 5 * All rights reserved. 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 1. Redistributions of source code must retain the above copyright 11 * notice unmodified, this list of conditions, and the following 12 * disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in the 15 * documentation and/or other materials provided with the distribution. 16 * 17 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR 18 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES 19 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. 20 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, 21 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT 22 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 23 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 24 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 25 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF 26 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 27 */ 28 29 /* 30 * This file implements the ULE scheduler. ULE supports independent CPU 31 * run queues and fine grain locking. It has superior interactive 32 * performance under load even on uni-processor systems. 33 * 34 * etymology: 35 * ULE is the last three letters in schedule. It owes its name to a 36 * generic user created for a scheduling system by Paul Mikesell at 37 * Isilon Systems and a general lack of creativity on the part of the author. 38 */ 39 40 #include <sys/cdefs.h> 41 __FBSDID("$FreeBSD$"); 42 43 #include "opt_hwpmc_hooks.h" 44 #include "opt_sched.h" 45 46 #include <sys/param.h> 47 #include <sys/systm.h> 48 #include <sys/kdb.h> 49 #include <sys/kernel.h> 50 #include <sys/ktr.h> 51 #include <sys/limits.h> 52 #include <sys/lock.h> 53 #include <sys/mutex.h> 54 #include <sys/proc.h> 55 #include <sys/resource.h> 56 #include <sys/resourcevar.h> 57 #include <sys/sched.h> 58 #include <sys/sdt.h> 59 #include <sys/smp.h> 60 #include <sys/sx.h> 61 #include <sys/sysctl.h> 62 #include <sys/sysproto.h> 63 #include <sys/turnstile.h> 64 #include <sys/umtxvar.h> 65 #include <sys/vmmeter.h> 66 #include <sys/cpuset.h> 67 #include <sys/sbuf.h> 68 69 #ifdef HWPMC_HOOKS 70 #include <sys/pmckern.h> 71 #endif 72 73 #ifdef KDTRACE_HOOKS 74 #include <sys/dtrace_bsd.h> 75 int __read_mostly dtrace_vtime_active; 76 dtrace_vtime_switch_func_t dtrace_vtime_switch_func; 77 #endif 78 79 #include <machine/cpu.h> 80 #include <machine/smp.h> 81 82 #define KTR_ULE 0 83 84 #define TS_NAME_LEN (MAXCOMLEN + sizeof(" td ") + sizeof(__XSTRING(UINT_MAX))) 85 #define TDQ_NAME_LEN (sizeof("sched lock ") + sizeof(__XSTRING(MAXCPU))) 86 #define TDQ_LOADNAME_LEN (sizeof("CPU ") + sizeof(__XSTRING(MAXCPU)) - 1 + sizeof(" load")) 87 88 /* 89 * Thread scheduler specific section. All fields are protected 90 * by the thread lock. 91 */ 92 struct td_sched { 93 struct runq *ts_runq; /* Run-queue we're queued on. */ 94 short ts_flags; /* TSF_* flags. */ 95 int ts_cpu; /* CPU that we have affinity for. */ 96 int ts_rltick; /* Real last tick, for affinity. */ 97 int ts_slice; /* Ticks of slice remaining. */ 98 u_int ts_slptime; /* Number of ticks we vol. slept */ 99 u_int ts_runtime; /* Number of ticks we were running */ 100 int ts_ltick; /* Last tick that we were running on */ 101 int ts_ftick; /* First tick that we were running on */ 102 int ts_ticks; /* Tick count */ 103 #ifdef KTR 104 char ts_name[TS_NAME_LEN]; 105 #endif 106 }; 107 /* flags kept in ts_flags */ 108 #define TSF_BOUND 0x0001 /* Thread can not migrate. */ 109 #define TSF_XFERABLE 0x0002 /* Thread was added as transferable. */ 110 111 #define THREAD_CAN_MIGRATE(td) ((td)->td_pinned == 0) 112 #define THREAD_CAN_SCHED(td, cpu) \ 113 CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask) 114 115 _Static_assert(sizeof(struct thread) + sizeof(struct td_sched) <= 116 sizeof(struct thread0_storage), 117 "increase struct thread0_storage.t0st_sched size"); 118 119 /* 120 * Priority ranges used for interactive and non-interactive timeshare 121 * threads. The timeshare priorities are split up into four ranges. 122 * The first range handles interactive threads. The last three ranges 123 * (NHALF, x, and NHALF) handle non-interactive threads with the outer 124 * ranges supporting nice values. 125 */ 126 #define PRI_TIMESHARE_RANGE (PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE + 1) 127 #define PRI_INTERACT_RANGE ((PRI_TIMESHARE_RANGE - SCHED_PRI_NRESV) / 2) 128 #define PRI_BATCH_RANGE (PRI_TIMESHARE_RANGE - PRI_INTERACT_RANGE) 129 130 #define PRI_MIN_INTERACT PRI_MIN_TIMESHARE 131 #define PRI_MAX_INTERACT (PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE - 1) 132 #define PRI_MIN_BATCH (PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE) 133 #define PRI_MAX_BATCH PRI_MAX_TIMESHARE 134 135 /* 136 * Cpu percentage computation macros and defines. 137 * 138 * SCHED_TICK_SECS: Number of seconds to average the cpu usage across. 139 * SCHED_TICK_TARG: Number of hz ticks to average the cpu usage across. 140 * SCHED_TICK_MAX: Maximum number of ticks before scaling back. 141 * SCHED_TICK_SHIFT: Shift factor to avoid rounding away results. 142 * SCHED_TICK_HZ: Compute the number of hz ticks for a given ticks count. 143 * SCHED_TICK_TOTAL: Gives the amount of time we've been recording ticks. 144 */ 145 #define SCHED_TICK_SECS 10 146 #define SCHED_TICK_TARG (hz * SCHED_TICK_SECS) 147 #define SCHED_TICK_MAX (SCHED_TICK_TARG + hz) 148 #define SCHED_TICK_SHIFT 10 149 #define SCHED_TICK_HZ(ts) ((ts)->ts_ticks >> SCHED_TICK_SHIFT) 150 #define SCHED_TICK_TOTAL(ts) (max((ts)->ts_ltick - (ts)->ts_ftick, hz)) 151 152 /* 153 * These macros determine priorities for non-interactive threads. They are 154 * assigned a priority based on their recent cpu utilization as expressed 155 * by the ratio of ticks to the tick total. NHALF priorities at the start 156 * and end of the MIN to MAX timeshare range are only reachable with negative 157 * or positive nice respectively. 158 * 159 * PRI_RANGE: Priority range for utilization dependent priorities. 160 * PRI_NRESV: Number of nice values. 161 * PRI_TICKS: Compute a priority in PRI_RANGE from the ticks count and total. 162 * PRI_NICE: Determines the part of the priority inherited from nice. 163 */ 164 #define SCHED_PRI_NRESV (PRIO_MAX - PRIO_MIN) 165 #define SCHED_PRI_NHALF (SCHED_PRI_NRESV / 2) 166 #define SCHED_PRI_MIN (PRI_MIN_BATCH + SCHED_PRI_NHALF) 167 #define SCHED_PRI_MAX (PRI_MAX_BATCH - SCHED_PRI_NHALF) 168 #define SCHED_PRI_RANGE (SCHED_PRI_MAX - SCHED_PRI_MIN + 1) 169 #define SCHED_PRI_TICKS(ts) \ 170 (SCHED_TICK_HZ((ts)) / \ 171 (roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE)) 172 #define SCHED_PRI_NICE(nice) (nice) 173 174 /* 175 * These determine the interactivity of a process. Interactivity differs from 176 * cpu utilization in that it expresses the voluntary time slept vs time ran 177 * while cpu utilization includes all time not running. This more accurately 178 * models the intent of the thread. 179 * 180 * SLP_RUN_MAX: Maximum amount of sleep time + run time we'll accumulate 181 * before throttling back. 182 * SLP_RUN_FORK: Maximum slp+run time to inherit at fork time. 183 * INTERACT_MAX: Maximum interactivity value. Smaller is better. 184 * INTERACT_THRESH: Threshold for placement on the current runq. 185 */ 186 #define SCHED_SLP_RUN_MAX ((hz * 5) << SCHED_TICK_SHIFT) 187 #define SCHED_SLP_RUN_FORK ((hz / 2) << SCHED_TICK_SHIFT) 188 #define SCHED_INTERACT_MAX (100) 189 #define SCHED_INTERACT_HALF (SCHED_INTERACT_MAX / 2) 190 #define SCHED_INTERACT_THRESH (30) 191 192 /* 193 * These parameters determine the slice behavior for batch work. 194 */ 195 #define SCHED_SLICE_DEFAULT_DIVISOR 10 /* ~94 ms, 12 stathz ticks. */ 196 #define SCHED_SLICE_MIN_DIVISOR 6 /* DEFAULT/MIN = ~16 ms. */ 197 198 /* Flags kept in td_flags. */ 199 #define TDF_PICKCPU TDF_SCHED0 /* Thread should pick new CPU. */ 200 #define TDF_SLICEEND TDF_SCHED2 /* Thread time slice is over. */ 201 202 /* 203 * tickincr: Converts a stathz tick into a hz domain scaled by 204 * the shift factor. Without the shift the error rate 205 * due to rounding would be unacceptably high. 206 * realstathz: stathz is sometimes 0 and run off of hz. 207 * sched_slice: Runtime of each thread before rescheduling. 208 * preempt_thresh: Priority threshold for preemption and remote IPIs. 209 */ 210 static u_int __read_mostly sched_interact = SCHED_INTERACT_THRESH; 211 static int __read_mostly tickincr = 8 << SCHED_TICK_SHIFT; 212 static int __read_mostly realstathz = 127; /* reset during boot. */ 213 static int __read_mostly sched_slice = 10; /* reset during boot. */ 214 static int __read_mostly sched_slice_min = 1; /* reset during boot. */ 215 #ifdef PREEMPTION 216 #ifdef FULL_PREEMPTION 217 static int __read_mostly preempt_thresh = PRI_MAX_IDLE; 218 #else 219 static int __read_mostly preempt_thresh = PRI_MIN_KERN; 220 #endif 221 #else 222 static int __read_mostly preempt_thresh = 0; 223 #endif 224 static int __read_mostly static_boost = PRI_MIN_BATCH; 225 static int __read_mostly sched_idlespins = 10000; 226 static int __read_mostly sched_idlespinthresh = -1; 227 228 /* 229 * tdq - per processor runqs and statistics. A mutex synchronizes access to 230 * most fields. Some fields are loaded or modified without the mutex. 231 * 232 * Locking protocols: 233 * (c) constant after initialization 234 * (f) flag, set with the tdq lock held, cleared on local CPU 235 * (l) all accesses are CPU-local 236 * (ls) stores are performed by the local CPU, loads may be lockless 237 * (t) all accesses are protected by the tdq mutex 238 * (ts) stores are serialized by the tdq mutex, loads may be lockless 239 */ 240 struct tdq { 241 /* 242 * Ordered to improve efficiency of cpu_search() and switch(). 243 * tdq_lock is padded to avoid false sharing with tdq_load and 244 * tdq_cpu_idle. 245 */ 246 struct mtx_padalign tdq_lock; /* run queue lock. */ 247 struct cpu_group *tdq_cg; /* (c) Pointer to cpu topology. */ 248 struct thread *tdq_curthread; /* (t) Current executing thread. */ 249 int tdq_load; /* (ts) Aggregate load. */ 250 int tdq_sysload; /* (ts) For loadavg, !ITHD load. */ 251 int tdq_cpu_idle; /* (ls) cpu_idle() is active. */ 252 int tdq_transferable; /* (ts) Transferable thread count. */ 253 short tdq_switchcnt; /* (l) Switches this tick. */ 254 short tdq_oldswitchcnt; /* (l) Switches last tick. */ 255 u_char tdq_lowpri; /* (ts) Lowest priority thread. */ 256 u_char tdq_owepreempt; /* (f) Remote preemption pending. */ 257 u_char tdq_idx; /* (t) Current insert index. */ 258 u_char tdq_ridx; /* (t) Current removal index. */ 259 int tdq_id; /* (c) cpuid. */ 260 struct runq tdq_realtime; /* (t) real-time run queue. */ 261 struct runq tdq_timeshare; /* (t) timeshare run queue. */ 262 struct runq tdq_idle; /* (t) Queue of IDLE threads. */ 263 char tdq_name[TDQ_NAME_LEN]; 264 #ifdef KTR 265 char tdq_loadname[TDQ_LOADNAME_LEN]; 266 #endif 267 }; 268 269 /* Idle thread states and config. */ 270 #define TDQ_RUNNING 1 271 #define TDQ_IDLE 2 272 273 /* Lockless accessors. */ 274 #define TDQ_LOAD(tdq) atomic_load_int(&(tdq)->tdq_load) 275 #define TDQ_TRANSFERABLE(tdq) atomic_load_int(&(tdq)->tdq_transferable) 276 #define TDQ_SWITCHCNT(tdq) (atomic_load_short(&(tdq)->tdq_switchcnt) + \ 277 atomic_load_short(&(tdq)->tdq_oldswitchcnt)) 278 #define TDQ_SWITCHCNT_INC(tdq) (atomic_store_short(&(tdq)->tdq_switchcnt, \ 279 atomic_load_short(&(tdq)->tdq_switchcnt) + 1)) 280 281 #ifdef SMP 282 struct cpu_group __read_mostly *cpu_top; /* CPU topology */ 283 284 #define SCHED_AFFINITY_DEFAULT (max(1, hz / 1000)) 285 #define SCHED_AFFINITY(ts, t) ((ts)->ts_rltick > ticks - ((t) * affinity)) 286 287 /* 288 * Run-time tunables. 289 */ 290 static int rebalance = 1; 291 static int balance_interval = 128; /* Default set in sched_initticks(). */ 292 static int __read_mostly affinity; 293 static int __read_mostly steal_idle = 1; 294 static int __read_mostly steal_thresh = 2; 295 static int __read_mostly always_steal = 0; 296 static int __read_mostly trysteal_limit = 2; 297 298 /* 299 * One thread queue per processor. 300 */ 301 static struct tdq __read_mostly *balance_tdq; 302 static int balance_ticks; 303 DPCPU_DEFINE_STATIC(struct tdq, tdq); 304 DPCPU_DEFINE_STATIC(uint32_t, randomval); 305 306 #define TDQ_SELF() ((struct tdq *)PCPU_GET(sched)) 307 #define TDQ_CPU(x) (DPCPU_ID_PTR((x), tdq)) 308 #define TDQ_ID(x) ((x)->tdq_id) 309 #else /* !SMP */ 310 static struct tdq tdq_cpu; 311 312 #define TDQ_ID(x) (0) 313 #define TDQ_SELF() (&tdq_cpu) 314 #define TDQ_CPU(x) (&tdq_cpu) 315 #endif 316 317 #define TDQ_LOCK_ASSERT(t, type) mtx_assert(TDQ_LOCKPTR((t)), (type)) 318 #define TDQ_LOCK(t) mtx_lock_spin(TDQ_LOCKPTR((t))) 319 #define TDQ_LOCK_FLAGS(t, f) mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f)) 320 #define TDQ_TRYLOCK(t) mtx_trylock_spin(TDQ_LOCKPTR((t))) 321 #define TDQ_TRYLOCK_FLAGS(t, f) mtx_trylock_spin_flags(TDQ_LOCKPTR((t)), (f)) 322 #define TDQ_UNLOCK(t) mtx_unlock_spin(TDQ_LOCKPTR((t))) 323 #define TDQ_LOCKPTR(t) ((struct mtx *)(&(t)->tdq_lock)) 324 325 static void sched_setpreempt(int); 326 static void sched_priority(struct thread *); 327 static void sched_thread_priority(struct thread *, u_char); 328 static int sched_interact_score(struct thread *); 329 static void sched_interact_update(struct thread *); 330 static void sched_interact_fork(struct thread *); 331 static void sched_pctcpu_update(struct td_sched *, int); 332 333 /* Operations on per processor queues */ 334 static struct thread *tdq_choose(struct tdq *); 335 static void tdq_setup(struct tdq *, int i); 336 static void tdq_load_add(struct tdq *, struct thread *); 337 static void tdq_load_rem(struct tdq *, struct thread *); 338 static __inline void tdq_runq_add(struct tdq *, struct thread *, int); 339 static __inline void tdq_runq_rem(struct tdq *, struct thread *); 340 static inline int sched_shouldpreempt(int, int, int); 341 static void tdq_print(int cpu); 342 static void runq_print(struct runq *rq); 343 static int tdq_add(struct tdq *, struct thread *, int); 344 #ifdef SMP 345 static int tdq_move(struct tdq *, struct tdq *); 346 static int tdq_idled(struct tdq *); 347 static void tdq_notify(struct tdq *, int lowpri); 348 static struct thread *tdq_steal(struct tdq *, int); 349 static struct thread *runq_steal(struct runq *, int); 350 static int sched_pickcpu(struct thread *, int); 351 static void sched_balance(void); 352 static bool sched_balance_pair(struct tdq *, struct tdq *); 353 static inline struct tdq *sched_setcpu(struct thread *, int, int); 354 static inline void thread_unblock_switch(struct thread *, struct mtx *); 355 static int sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS); 356 static int sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, 357 struct cpu_group *cg, int indent); 358 #endif 359 360 static void sched_setup(void *dummy); 361 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL); 362 363 static void sched_initticks(void *dummy); 364 SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks, 365 NULL); 366 367 SDT_PROVIDER_DEFINE(sched); 368 369 SDT_PROBE_DEFINE3(sched, , , change__pri, "struct thread *", 370 "struct proc *", "uint8_t"); 371 SDT_PROBE_DEFINE3(sched, , , dequeue, "struct thread *", 372 "struct proc *", "void *"); 373 SDT_PROBE_DEFINE4(sched, , , enqueue, "struct thread *", 374 "struct proc *", "void *", "int"); 375 SDT_PROBE_DEFINE4(sched, , , lend__pri, "struct thread *", 376 "struct proc *", "uint8_t", "struct thread *"); 377 SDT_PROBE_DEFINE2(sched, , , load__change, "int", "int"); 378 SDT_PROBE_DEFINE2(sched, , , off__cpu, "struct thread *", 379 "struct proc *"); 380 SDT_PROBE_DEFINE(sched, , , on__cpu); 381 SDT_PROBE_DEFINE(sched, , , remain__cpu); 382 SDT_PROBE_DEFINE2(sched, , , surrender, "struct thread *", 383 "struct proc *"); 384 385 /* 386 * Print the threads waiting on a run-queue. 387 */ 388 static void 389 runq_print(struct runq *rq) 390 { 391 struct rqhead *rqh; 392 struct thread *td; 393 int pri; 394 int j; 395 int i; 396 397 for (i = 0; i < RQB_LEN; i++) { 398 printf("\t\trunq bits %d 0x%zx\n", 399 i, rq->rq_status.rqb_bits[i]); 400 for (j = 0; j < RQB_BPW; j++) 401 if (rq->rq_status.rqb_bits[i] & (1ul << j)) { 402 pri = j + (i << RQB_L2BPW); 403 rqh = &rq->rq_queues[pri]; 404 TAILQ_FOREACH(td, rqh, td_runq) { 405 printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n", 406 td, td->td_name, td->td_priority, 407 td->td_rqindex, pri); 408 } 409 } 410 } 411 } 412 413 /* 414 * Print the status of a per-cpu thread queue. Should be a ddb show cmd. 415 */ 416 static void __unused 417 tdq_print(int cpu) 418 { 419 struct tdq *tdq; 420 421 tdq = TDQ_CPU(cpu); 422 423 printf("tdq %d:\n", TDQ_ID(tdq)); 424 printf("\tlock %p\n", TDQ_LOCKPTR(tdq)); 425 printf("\tLock name: %s\n", tdq->tdq_name); 426 printf("\tload: %d\n", tdq->tdq_load); 427 printf("\tswitch cnt: %d\n", tdq->tdq_switchcnt); 428 printf("\told switch cnt: %d\n", tdq->tdq_oldswitchcnt); 429 printf("\ttimeshare idx: %d\n", tdq->tdq_idx); 430 printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx); 431 printf("\tload transferable: %d\n", tdq->tdq_transferable); 432 printf("\tlowest priority: %d\n", tdq->tdq_lowpri); 433 printf("\trealtime runq:\n"); 434 runq_print(&tdq->tdq_realtime); 435 printf("\ttimeshare runq:\n"); 436 runq_print(&tdq->tdq_timeshare); 437 printf("\tidle runq:\n"); 438 runq_print(&tdq->tdq_idle); 439 } 440 441 static inline int 442 sched_shouldpreempt(int pri, int cpri, int remote) 443 { 444 /* 445 * If the new priority is not better than the current priority there is 446 * nothing to do. 447 */ 448 if (pri >= cpri) 449 return (0); 450 /* 451 * Always preempt idle. 452 */ 453 if (cpri >= PRI_MIN_IDLE) 454 return (1); 455 /* 456 * If preemption is disabled don't preempt others. 457 */ 458 if (preempt_thresh == 0) 459 return (0); 460 /* 461 * Preempt if we exceed the threshold. 462 */ 463 if (pri <= preempt_thresh) 464 return (1); 465 /* 466 * If we're interactive or better and there is non-interactive 467 * or worse running preempt only remote processors. 468 */ 469 if (remote && pri <= PRI_MAX_INTERACT && cpri > PRI_MAX_INTERACT) 470 return (1); 471 return (0); 472 } 473 474 /* 475 * Add a thread to the actual run-queue. Keeps transferable counts up to 476 * date with what is actually on the run-queue. Selects the correct 477 * queue position for timeshare threads. 478 */ 479 static __inline void 480 tdq_runq_add(struct tdq *tdq, struct thread *td, int flags) 481 { 482 struct td_sched *ts; 483 u_char pri; 484 485 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 486 THREAD_LOCK_BLOCKED_ASSERT(td, MA_OWNED); 487 488 pri = td->td_priority; 489 ts = td_get_sched(td); 490 TD_SET_RUNQ(td); 491 if (THREAD_CAN_MIGRATE(td)) { 492 tdq->tdq_transferable++; 493 ts->ts_flags |= TSF_XFERABLE; 494 } 495 if (pri < PRI_MIN_BATCH) { 496 ts->ts_runq = &tdq->tdq_realtime; 497 } else if (pri <= PRI_MAX_BATCH) { 498 ts->ts_runq = &tdq->tdq_timeshare; 499 KASSERT(pri <= PRI_MAX_BATCH && pri >= PRI_MIN_BATCH, 500 ("Invalid priority %d on timeshare runq", pri)); 501 /* 502 * This queue contains only priorities between MIN and MAX 503 * batch. Use the whole queue to represent these values. 504 */ 505 if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) { 506 pri = RQ_NQS * (pri - PRI_MIN_BATCH) / PRI_BATCH_RANGE; 507 pri = (pri + tdq->tdq_idx) % RQ_NQS; 508 /* 509 * This effectively shortens the queue by one so we 510 * can have a one slot difference between idx and 511 * ridx while we wait for threads to drain. 512 */ 513 if (tdq->tdq_ridx != tdq->tdq_idx && 514 pri == tdq->tdq_ridx) 515 pri = (unsigned char)(pri - 1) % RQ_NQS; 516 } else 517 pri = tdq->tdq_ridx; 518 runq_add_pri(ts->ts_runq, td, pri, flags); 519 return; 520 } else 521 ts->ts_runq = &tdq->tdq_idle; 522 runq_add(ts->ts_runq, td, flags); 523 } 524 525 /* 526 * Remove a thread from a run-queue. This typically happens when a thread 527 * is selected to run. Running threads are not on the queue and the 528 * transferable count does not reflect them. 529 */ 530 static __inline void 531 tdq_runq_rem(struct tdq *tdq, struct thread *td) 532 { 533 struct td_sched *ts; 534 535 ts = td_get_sched(td); 536 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 537 THREAD_LOCK_BLOCKED_ASSERT(td, MA_OWNED); 538 KASSERT(ts->ts_runq != NULL, 539 ("tdq_runq_remove: thread %p null ts_runq", td)); 540 if (ts->ts_flags & TSF_XFERABLE) { 541 tdq->tdq_transferable--; 542 ts->ts_flags &= ~TSF_XFERABLE; 543 } 544 if (ts->ts_runq == &tdq->tdq_timeshare) { 545 if (tdq->tdq_idx != tdq->tdq_ridx) 546 runq_remove_idx(ts->ts_runq, td, &tdq->tdq_ridx); 547 else 548 runq_remove_idx(ts->ts_runq, td, NULL); 549 } else 550 runq_remove(ts->ts_runq, td); 551 } 552 553 /* 554 * Load is maintained for all threads RUNNING and ON_RUNQ. Add the load 555 * for this thread to the referenced thread queue. 556 */ 557 static void 558 tdq_load_add(struct tdq *tdq, struct thread *td) 559 { 560 561 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 562 THREAD_LOCK_BLOCKED_ASSERT(td, MA_OWNED); 563 564 tdq->tdq_load++; 565 if ((td->td_flags & TDF_NOLOAD) == 0) 566 tdq->tdq_sysload++; 567 KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load); 568 SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load); 569 } 570 571 /* 572 * Remove the load from a thread that is transitioning to a sleep state or 573 * exiting. 574 */ 575 static void 576 tdq_load_rem(struct tdq *tdq, struct thread *td) 577 { 578 579 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 580 THREAD_LOCK_BLOCKED_ASSERT(td, MA_OWNED); 581 KASSERT(tdq->tdq_load != 0, 582 ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq))); 583 584 tdq->tdq_load--; 585 if ((td->td_flags & TDF_NOLOAD) == 0) 586 tdq->tdq_sysload--; 587 KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load); 588 SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load); 589 } 590 591 /* 592 * Bound timeshare latency by decreasing slice size as load increases. We 593 * consider the maximum latency as the sum of the threads waiting to run 594 * aside from curthread and target no more than sched_slice latency but 595 * no less than sched_slice_min runtime. 596 */ 597 static inline int 598 tdq_slice(struct tdq *tdq) 599 { 600 int load; 601 602 /* 603 * It is safe to use sys_load here because this is called from 604 * contexts where timeshare threads are running and so there 605 * cannot be higher priority load in the system. 606 */ 607 load = tdq->tdq_sysload - 1; 608 if (load >= SCHED_SLICE_MIN_DIVISOR) 609 return (sched_slice_min); 610 if (load <= 1) 611 return (sched_slice); 612 return (sched_slice / load); 613 } 614 615 /* 616 * Set lowpri to its exact value by searching the run-queue and 617 * evaluating curthread. curthread may be passed as an optimization. 618 */ 619 static void 620 tdq_setlowpri(struct tdq *tdq, struct thread *ctd) 621 { 622 struct thread *td; 623 624 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 625 if (ctd == NULL) 626 ctd = tdq->tdq_curthread; 627 td = tdq_choose(tdq); 628 if (td == NULL || td->td_priority > ctd->td_priority) 629 tdq->tdq_lowpri = ctd->td_priority; 630 else 631 tdq->tdq_lowpri = td->td_priority; 632 } 633 634 #ifdef SMP 635 /* 636 * We need some randomness. Implement a classic Linear Congruential 637 * Generator X_{n+1}=(aX_n+c) mod m. These values are optimized for 638 * m = 2^32, a = 69069 and c = 5. We only return the upper 16 bits 639 * of the random state (in the low bits of our answer) to keep 640 * the maximum randomness. 641 */ 642 static uint32_t 643 sched_random(void) 644 { 645 uint32_t *rndptr; 646 647 rndptr = DPCPU_PTR(randomval); 648 *rndptr = *rndptr * 69069 + 5; 649 650 return (*rndptr >> 16); 651 } 652 653 struct cpu_search { 654 cpuset_t *cs_mask; /* The mask of allowed CPUs to choose from. */ 655 int cs_prefer; /* Prefer this CPU and groups including it. */ 656 int cs_running; /* The thread is now running at cs_prefer. */ 657 int cs_pri; /* Min priority for low. */ 658 int cs_load; /* Max load for low, min load for high. */ 659 int cs_trans; /* Min transferable load for high. */ 660 }; 661 662 struct cpu_search_res { 663 int csr_cpu; /* The best CPU found. */ 664 int csr_load; /* The load of cs_cpu. */ 665 }; 666 667 /* 668 * Search the tree of cpu_groups for the lowest or highest loaded CPU. 669 * These routines actually compare the load on all paths through the tree 670 * and find the least loaded cpu on the least loaded path, which may differ 671 * from the least loaded cpu in the system. This balances work among caches 672 * and buses. 673 */ 674 static int 675 cpu_search_lowest(const struct cpu_group *cg, const struct cpu_search *s, 676 struct cpu_search_res *r) 677 { 678 struct cpu_search_res lr; 679 struct tdq *tdq; 680 int c, bload, l, load, p, total; 681 682 total = 0; 683 bload = INT_MAX; 684 r->csr_cpu = -1; 685 686 /* Loop through children CPU groups if there are any. */ 687 if (cg->cg_children > 0) { 688 for (c = cg->cg_children - 1; c >= 0; c--) { 689 load = cpu_search_lowest(&cg->cg_child[c], s, &lr); 690 total += load; 691 692 /* 693 * When balancing do not prefer SMT groups with load >1. 694 * It allows round-robin between SMT groups with equal 695 * load within parent group for more fair scheduling. 696 */ 697 if (__predict_false(s->cs_running) && 698 (cg->cg_child[c].cg_flags & CG_FLAG_THREAD) && 699 load >= 128 && (load & 128) != 0) 700 load += 128; 701 702 if (lr.csr_cpu >= 0 && (load < bload || 703 (load == bload && lr.csr_load < r->csr_load))) { 704 bload = load; 705 r->csr_cpu = lr.csr_cpu; 706 r->csr_load = lr.csr_load; 707 } 708 } 709 return (total); 710 } 711 712 /* Loop through children CPUs otherwise. */ 713 for (c = cg->cg_last; c >= cg->cg_first; c--) { 714 if (!CPU_ISSET(c, &cg->cg_mask)) 715 continue; 716 tdq = TDQ_CPU(c); 717 l = TDQ_LOAD(tdq); 718 if (c == s->cs_prefer) { 719 if (__predict_false(s->cs_running)) 720 l--; 721 p = 128; 722 } else 723 p = 0; 724 load = l * 256; 725 total += load - p; 726 727 /* 728 * Check this CPU is acceptable. 729 * If the threads is already on the CPU, don't look on the TDQ 730 * priority, since it can be the priority of the thread itself. 731 */ 732 if (l > s->cs_load || 733 (atomic_load_char(&tdq->tdq_lowpri) <= s->cs_pri && 734 (!s->cs_running || c != s->cs_prefer)) || 735 !CPU_ISSET(c, s->cs_mask)) 736 continue; 737 738 /* 739 * When balancing do not prefer CPUs with load > 1. 740 * It allows round-robin between CPUs with equal load 741 * within the CPU group for more fair scheduling. 742 */ 743 if (__predict_false(s->cs_running) && l > 0) 744 p = 0; 745 746 load -= sched_random() % 128; 747 if (bload > load - p) { 748 bload = load - p; 749 r->csr_cpu = c; 750 r->csr_load = load; 751 } 752 } 753 return (total); 754 } 755 756 static int 757 cpu_search_highest(const struct cpu_group *cg, const struct cpu_search *s, 758 struct cpu_search_res *r) 759 { 760 struct cpu_search_res lr; 761 struct tdq *tdq; 762 int c, bload, l, load, total; 763 764 total = 0; 765 bload = INT_MIN; 766 r->csr_cpu = -1; 767 768 /* Loop through children CPU groups if there are any. */ 769 if (cg->cg_children > 0) { 770 for (c = cg->cg_children - 1; c >= 0; c--) { 771 load = cpu_search_highest(&cg->cg_child[c], s, &lr); 772 total += load; 773 if (lr.csr_cpu >= 0 && (load > bload || 774 (load == bload && lr.csr_load > r->csr_load))) { 775 bload = load; 776 r->csr_cpu = lr.csr_cpu; 777 r->csr_load = lr.csr_load; 778 } 779 } 780 return (total); 781 } 782 783 /* Loop through children CPUs otherwise. */ 784 for (c = cg->cg_last; c >= cg->cg_first; c--) { 785 if (!CPU_ISSET(c, &cg->cg_mask)) 786 continue; 787 tdq = TDQ_CPU(c); 788 l = TDQ_LOAD(tdq); 789 load = l * 256; 790 total += load; 791 792 /* 793 * Check this CPU is acceptable. 794 */ 795 if (l < s->cs_load || TDQ_TRANSFERABLE(tdq) < s->cs_trans || 796 !CPU_ISSET(c, s->cs_mask)) 797 continue; 798 799 load -= sched_random() % 256; 800 if (load > bload) { 801 bload = load; 802 r->csr_cpu = c; 803 } 804 } 805 r->csr_load = bload; 806 return (total); 807 } 808 809 /* 810 * Find the cpu with the least load via the least loaded path that has a 811 * lowpri greater than pri pri. A pri of -1 indicates any priority is 812 * acceptable. 813 */ 814 static inline int 815 sched_lowest(const struct cpu_group *cg, cpuset_t *mask, int pri, int maxload, 816 int prefer, int running) 817 { 818 struct cpu_search s; 819 struct cpu_search_res r; 820 821 s.cs_prefer = prefer; 822 s.cs_running = running; 823 s.cs_mask = mask; 824 s.cs_pri = pri; 825 s.cs_load = maxload; 826 cpu_search_lowest(cg, &s, &r); 827 return (r.csr_cpu); 828 } 829 830 /* 831 * Find the cpu with the highest load via the highest loaded path. 832 */ 833 static inline int 834 sched_highest(const struct cpu_group *cg, cpuset_t *mask, int minload, 835 int mintrans) 836 { 837 struct cpu_search s; 838 struct cpu_search_res r; 839 840 s.cs_mask = mask; 841 s.cs_load = minload; 842 s.cs_trans = mintrans; 843 cpu_search_highest(cg, &s, &r); 844 return (r.csr_cpu); 845 } 846 847 static void 848 sched_balance_group(struct cpu_group *cg) 849 { 850 struct tdq *tdq; 851 struct thread *td; 852 cpuset_t hmask, lmask; 853 int high, low, anylow; 854 855 CPU_FILL(&hmask); 856 for (;;) { 857 high = sched_highest(cg, &hmask, 1, 0); 858 /* Stop if there is no more CPU with transferrable threads. */ 859 if (high == -1) 860 break; 861 CPU_CLR(high, &hmask); 862 CPU_COPY(&hmask, &lmask); 863 /* Stop if there is no more CPU left for low. */ 864 if (CPU_EMPTY(&lmask)) 865 break; 866 tdq = TDQ_CPU(high); 867 if (TDQ_LOAD(tdq) == 1) { 868 /* 869 * There is only one running thread. We can't move 870 * it from here, so tell it to pick new CPU by itself. 871 */ 872 TDQ_LOCK(tdq); 873 td = tdq->tdq_curthread; 874 if (td->td_lock == TDQ_LOCKPTR(tdq) && 875 (td->td_flags & TDF_IDLETD) == 0 && 876 THREAD_CAN_MIGRATE(td)) { 877 td->td_flags |= TDF_PICKCPU; 878 ast_sched_locked(td, TDA_SCHED); 879 if (high != curcpu) 880 ipi_cpu(high, IPI_AST); 881 } 882 TDQ_UNLOCK(tdq); 883 break; 884 } 885 anylow = 1; 886 nextlow: 887 if (TDQ_TRANSFERABLE(tdq) == 0) 888 continue; 889 low = sched_lowest(cg, &lmask, -1, TDQ_LOAD(tdq) - 1, high, 1); 890 /* Stop if we looked well and found no less loaded CPU. */ 891 if (anylow && low == -1) 892 break; 893 /* Go to next high if we found no less loaded CPU. */ 894 if (low == -1) 895 continue; 896 /* Transfer thread from high to low. */ 897 if (sched_balance_pair(tdq, TDQ_CPU(low))) { 898 /* CPU that got thread can no longer be a donor. */ 899 CPU_CLR(low, &hmask); 900 } else { 901 /* 902 * If failed, then there is no threads on high 903 * that can run on this low. Drop low from low 904 * mask and look for different one. 905 */ 906 CPU_CLR(low, &lmask); 907 anylow = 0; 908 goto nextlow; 909 } 910 } 911 } 912 913 static void 914 sched_balance(void) 915 { 916 struct tdq *tdq; 917 918 balance_ticks = max(balance_interval / 2, 1) + 919 (sched_random() % balance_interval); 920 tdq = TDQ_SELF(); 921 TDQ_UNLOCK(tdq); 922 sched_balance_group(cpu_top); 923 TDQ_LOCK(tdq); 924 } 925 926 /* 927 * Lock two thread queues using their address to maintain lock order. 928 */ 929 static void 930 tdq_lock_pair(struct tdq *one, struct tdq *two) 931 { 932 if (one < two) { 933 TDQ_LOCK(one); 934 TDQ_LOCK_FLAGS(two, MTX_DUPOK); 935 } else { 936 TDQ_LOCK(two); 937 TDQ_LOCK_FLAGS(one, MTX_DUPOK); 938 } 939 } 940 941 /* 942 * Unlock two thread queues. Order is not important here. 943 */ 944 static void 945 tdq_unlock_pair(struct tdq *one, struct tdq *two) 946 { 947 TDQ_UNLOCK(one); 948 TDQ_UNLOCK(two); 949 } 950 951 /* 952 * Transfer load between two imbalanced thread queues. Returns true if a thread 953 * was moved between the queues, and false otherwise. 954 */ 955 static bool 956 sched_balance_pair(struct tdq *high, struct tdq *low) 957 { 958 int cpu, lowpri; 959 bool ret; 960 961 ret = false; 962 tdq_lock_pair(high, low); 963 964 /* 965 * Transfer a thread from high to low. 966 */ 967 if (high->tdq_transferable != 0 && high->tdq_load > low->tdq_load) { 968 lowpri = tdq_move(high, low); 969 if (lowpri != -1) { 970 /* 971 * In case the target isn't the current CPU notify it of 972 * the new load, possibly sending an IPI to force it to 973 * reschedule. Otherwise maybe schedule a preemption. 974 */ 975 cpu = TDQ_ID(low); 976 if (cpu != PCPU_GET(cpuid)) 977 tdq_notify(low, lowpri); 978 else 979 sched_setpreempt(low->tdq_lowpri); 980 ret = true; 981 } 982 } 983 tdq_unlock_pair(high, low); 984 return (ret); 985 } 986 987 /* 988 * Move a thread from one thread queue to another. Returns -1 if the source 989 * queue was empty, else returns the maximum priority of all threads in 990 * the destination queue prior to the addition of the new thread. In the latter 991 * case, this priority can be used to determine whether an IPI needs to be 992 * delivered. 993 */ 994 static int 995 tdq_move(struct tdq *from, struct tdq *to) 996 { 997 struct thread *td; 998 int cpu; 999 1000 TDQ_LOCK_ASSERT(from, MA_OWNED); 1001 TDQ_LOCK_ASSERT(to, MA_OWNED); 1002 1003 cpu = TDQ_ID(to); 1004 td = tdq_steal(from, cpu); 1005 if (td == NULL) 1006 return (-1); 1007 1008 /* 1009 * Although the run queue is locked the thread may be 1010 * blocked. We can not set the lock until it is unblocked. 1011 */ 1012 thread_lock_block_wait(td); 1013 sched_rem(td); 1014 THREAD_LOCKPTR_ASSERT(td, TDQ_LOCKPTR(from)); 1015 td->td_lock = TDQ_LOCKPTR(to); 1016 td_get_sched(td)->ts_cpu = cpu; 1017 return (tdq_add(to, td, SRQ_YIELDING)); 1018 } 1019 1020 /* 1021 * This tdq has idled. Try to steal a thread from another cpu and switch 1022 * to it. 1023 */ 1024 static int 1025 tdq_idled(struct tdq *tdq) 1026 { 1027 struct cpu_group *cg, *parent; 1028 struct tdq *steal; 1029 cpuset_t mask; 1030 int cpu, switchcnt, goup; 1031 1032 if (smp_started == 0 || steal_idle == 0 || tdq->tdq_cg == NULL) 1033 return (1); 1034 CPU_FILL(&mask); 1035 CPU_CLR(PCPU_GET(cpuid), &mask); 1036 restart: 1037 switchcnt = TDQ_SWITCHCNT(tdq); 1038 for (cg = tdq->tdq_cg, goup = 0; ; ) { 1039 cpu = sched_highest(cg, &mask, steal_thresh, 1); 1040 /* 1041 * We were assigned a thread but not preempted. Returning 1042 * 0 here will cause our caller to switch to it. 1043 */ 1044 if (TDQ_LOAD(tdq)) 1045 return (0); 1046 1047 /* 1048 * We found no CPU to steal from in this group. Escalate to 1049 * the parent and repeat. But if parent has only two children 1050 * groups we can avoid searching this group again by searching 1051 * the other one specifically and then escalating two levels. 1052 */ 1053 if (cpu == -1) { 1054 if (goup) { 1055 cg = cg->cg_parent; 1056 goup = 0; 1057 } 1058 parent = cg->cg_parent; 1059 if (parent == NULL) 1060 return (1); 1061 if (parent->cg_children == 2) { 1062 if (cg == &parent->cg_child[0]) 1063 cg = &parent->cg_child[1]; 1064 else 1065 cg = &parent->cg_child[0]; 1066 goup = 1; 1067 } else 1068 cg = parent; 1069 continue; 1070 } 1071 steal = TDQ_CPU(cpu); 1072 /* 1073 * The data returned by sched_highest() is stale and 1074 * the chosen CPU no longer has an eligible thread. 1075 * 1076 * Testing this ahead of tdq_lock_pair() only catches 1077 * this situation about 20% of the time on an 8 core 1078 * 16 thread Ryzen 7, but it still helps performance. 1079 */ 1080 if (TDQ_LOAD(steal) < steal_thresh || 1081 TDQ_TRANSFERABLE(steal) == 0) 1082 goto restart; 1083 /* 1084 * Try to lock both queues. If we are assigned a thread while 1085 * waited for the lock, switch to it now instead of stealing. 1086 * If we can't get the lock, then somebody likely got there 1087 * first so continue searching. 1088 */ 1089 TDQ_LOCK(tdq); 1090 if (tdq->tdq_load > 0) { 1091 mi_switch(SW_VOL | SWT_IDLE); 1092 return (0); 1093 } 1094 if (TDQ_TRYLOCK_FLAGS(steal, MTX_DUPOK) == 0) { 1095 TDQ_UNLOCK(tdq); 1096 CPU_CLR(cpu, &mask); 1097 continue; 1098 } 1099 /* 1100 * The data returned by sched_highest() is stale and 1101 * the chosen CPU no longer has an eligible thread, or 1102 * we were preempted and the CPU loading info may be out 1103 * of date. The latter is rare. In either case restart 1104 * the search. 1105 */ 1106 if (TDQ_LOAD(steal) < steal_thresh || 1107 TDQ_TRANSFERABLE(steal) == 0 || 1108 switchcnt != TDQ_SWITCHCNT(tdq)) { 1109 tdq_unlock_pair(tdq, steal); 1110 goto restart; 1111 } 1112 /* 1113 * Steal the thread and switch to it. 1114 */ 1115 if (tdq_move(steal, tdq) != -1) 1116 break; 1117 /* 1118 * We failed to acquire a thread even though it looked 1119 * like one was available. This could be due to affinity 1120 * restrictions or for other reasons. Loop again after 1121 * removing this CPU from the set. The restart logic 1122 * above does not restore this CPU to the set due to the 1123 * likelyhood of failing here again. 1124 */ 1125 CPU_CLR(cpu, &mask); 1126 tdq_unlock_pair(tdq, steal); 1127 } 1128 TDQ_UNLOCK(steal); 1129 mi_switch(SW_VOL | SWT_IDLE); 1130 return (0); 1131 } 1132 1133 /* 1134 * Notify a remote cpu of new work. Sends an IPI if criteria are met. 1135 * 1136 * "lowpri" is the minimum scheduling priority among all threads on 1137 * the queue prior to the addition of the new thread. 1138 */ 1139 static void 1140 tdq_notify(struct tdq *tdq, int lowpri) 1141 { 1142 int cpu; 1143 1144 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 1145 KASSERT(tdq->tdq_lowpri <= lowpri, 1146 ("tdq_notify: lowpri %d > tdq_lowpri %d", lowpri, tdq->tdq_lowpri)); 1147 1148 if (tdq->tdq_owepreempt) 1149 return; 1150 1151 /* 1152 * Check to see if the newly added thread should preempt the one 1153 * currently running. 1154 */ 1155 if (!sched_shouldpreempt(tdq->tdq_lowpri, lowpri, 1)) 1156 return; 1157 1158 /* 1159 * Make sure that our caller's earlier update to tdq_load is 1160 * globally visible before we read tdq_cpu_idle. Idle thread 1161 * accesses both of them without locks, and the order is important. 1162 */ 1163 atomic_thread_fence_seq_cst(); 1164 1165 /* 1166 * Try to figure out if we can signal the idle thread instead of sending 1167 * an IPI. This check is racy; at worst, we will deliever an IPI 1168 * unnecessarily. 1169 */ 1170 cpu = TDQ_ID(tdq); 1171 if (TD_IS_IDLETHREAD(tdq->tdq_curthread) && 1172 (atomic_load_int(&tdq->tdq_cpu_idle) == 0 || cpu_idle_wakeup(cpu))) 1173 return; 1174 1175 /* 1176 * The run queues have been updated, so any switch on the remote CPU 1177 * will satisfy the preemption request. 1178 */ 1179 tdq->tdq_owepreempt = 1; 1180 ipi_cpu(cpu, IPI_PREEMPT); 1181 } 1182 1183 /* 1184 * Steals load from a timeshare queue. Honors the rotating queue head 1185 * index. 1186 */ 1187 static struct thread * 1188 runq_steal_from(struct runq *rq, int cpu, u_char start) 1189 { 1190 struct rqbits *rqb; 1191 struct rqhead *rqh; 1192 struct thread *td, *first; 1193 int bit; 1194 int i; 1195 1196 rqb = &rq->rq_status; 1197 bit = start & (RQB_BPW -1); 1198 first = NULL; 1199 again: 1200 for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) { 1201 if (rqb->rqb_bits[i] == 0) 1202 continue; 1203 if (bit == 0) 1204 bit = RQB_FFS(rqb->rqb_bits[i]); 1205 for (; bit < RQB_BPW; bit++) { 1206 if ((rqb->rqb_bits[i] & (1ul << bit)) == 0) 1207 continue; 1208 rqh = &rq->rq_queues[bit + (i << RQB_L2BPW)]; 1209 TAILQ_FOREACH(td, rqh, td_runq) { 1210 if (first) { 1211 if (THREAD_CAN_MIGRATE(td) && 1212 THREAD_CAN_SCHED(td, cpu)) 1213 return (td); 1214 } else 1215 first = td; 1216 } 1217 } 1218 } 1219 if (start != 0) { 1220 start = 0; 1221 goto again; 1222 } 1223 1224 if (first && THREAD_CAN_MIGRATE(first) && 1225 THREAD_CAN_SCHED(first, cpu)) 1226 return (first); 1227 return (NULL); 1228 } 1229 1230 /* 1231 * Steals load from a standard linear queue. 1232 */ 1233 static struct thread * 1234 runq_steal(struct runq *rq, int cpu) 1235 { 1236 struct rqhead *rqh; 1237 struct rqbits *rqb; 1238 struct thread *td; 1239 int word; 1240 int bit; 1241 1242 rqb = &rq->rq_status; 1243 for (word = 0; word < RQB_LEN; word++) { 1244 if (rqb->rqb_bits[word] == 0) 1245 continue; 1246 for (bit = 0; bit < RQB_BPW; bit++) { 1247 if ((rqb->rqb_bits[word] & (1ul << bit)) == 0) 1248 continue; 1249 rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)]; 1250 TAILQ_FOREACH(td, rqh, td_runq) 1251 if (THREAD_CAN_MIGRATE(td) && 1252 THREAD_CAN_SCHED(td, cpu)) 1253 return (td); 1254 } 1255 } 1256 return (NULL); 1257 } 1258 1259 /* 1260 * Attempt to steal a thread in priority order from a thread queue. 1261 */ 1262 static struct thread * 1263 tdq_steal(struct tdq *tdq, int cpu) 1264 { 1265 struct thread *td; 1266 1267 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 1268 if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL) 1269 return (td); 1270 if ((td = runq_steal_from(&tdq->tdq_timeshare, 1271 cpu, tdq->tdq_ridx)) != NULL) 1272 return (td); 1273 return (runq_steal(&tdq->tdq_idle, cpu)); 1274 } 1275 1276 /* 1277 * Sets the thread lock and ts_cpu to match the requested cpu. Unlocks the 1278 * current lock and returns with the assigned queue locked. 1279 */ 1280 static inline struct tdq * 1281 sched_setcpu(struct thread *td, int cpu, int flags) 1282 { 1283 1284 struct tdq *tdq; 1285 struct mtx *mtx; 1286 1287 THREAD_LOCK_ASSERT(td, MA_OWNED); 1288 tdq = TDQ_CPU(cpu); 1289 td_get_sched(td)->ts_cpu = cpu; 1290 /* 1291 * If the lock matches just return the queue. 1292 */ 1293 if (td->td_lock == TDQ_LOCKPTR(tdq)) { 1294 KASSERT((flags & SRQ_HOLD) == 0, 1295 ("sched_setcpu: Invalid lock for SRQ_HOLD")); 1296 return (tdq); 1297 } 1298 1299 /* 1300 * The hard case, migration, we need to block the thread first to 1301 * prevent order reversals with other cpus locks. 1302 */ 1303 spinlock_enter(); 1304 mtx = thread_lock_block(td); 1305 if ((flags & SRQ_HOLD) == 0) 1306 mtx_unlock_spin(mtx); 1307 TDQ_LOCK(tdq); 1308 thread_lock_unblock(td, TDQ_LOCKPTR(tdq)); 1309 spinlock_exit(); 1310 return (tdq); 1311 } 1312 1313 SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding"); 1314 SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity"); 1315 SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity"); 1316 SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load"); 1317 SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu"); 1318 SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration"); 1319 1320 static int 1321 sched_pickcpu(struct thread *td, int flags) 1322 { 1323 struct cpu_group *cg, *ccg; 1324 struct td_sched *ts; 1325 struct tdq *tdq; 1326 cpuset_t *mask; 1327 int cpu, pri, r, self, intr; 1328 1329 self = PCPU_GET(cpuid); 1330 ts = td_get_sched(td); 1331 KASSERT(!CPU_ABSENT(ts->ts_cpu), ("sched_pickcpu: Start scheduler on " 1332 "absent CPU %d for thread %s.", ts->ts_cpu, td->td_name)); 1333 if (smp_started == 0) 1334 return (self); 1335 /* 1336 * Don't migrate a running thread from sched_switch(). 1337 */ 1338 if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td)) 1339 return (ts->ts_cpu); 1340 /* 1341 * Prefer to run interrupt threads on the processors that generate 1342 * the interrupt. 1343 */ 1344 if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) && 1345 curthread->td_intr_nesting_level) { 1346 tdq = TDQ_SELF(); 1347 if (tdq->tdq_lowpri >= PRI_MIN_IDLE) { 1348 SCHED_STAT_INC(pickcpu_idle_affinity); 1349 return (self); 1350 } 1351 ts->ts_cpu = self; 1352 intr = 1; 1353 cg = tdq->tdq_cg; 1354 goto llc; 1355 } else { 1356 intr = 0; 1357 tdq = TDQ_CPU(ts->ts_cpu); 1358 cg = tdq->tdq_cg; 1359 } 1360 /* 1361 * If the thread can run on the last cpu and the affinity has not 1362 * expired and it is idle, run it there. 1363 */ 1364 if (THREAD_CAN_SCHED(td, ts->ts_cpu) && 1365 atomic_load_char(&tdq->tdq_lowpri) >= PRI_MIN_IDLE && 1366 SCHED_AFFINITY(ts, CG_SHARE_L2)) { 1367 if (cg->cg_flags & CG_FLAG_THREAD) { 1368 /* Check all SMT threads for being idle. */ 1369 for (cpu = cg->cg_first; cpu <= cg->cg_last; cpu++) { 1370 pri = 1371 atomic_load_char(&TDQ_CPU(cpu)->tdq_lowpri); 1372 if (CPU_ISSET(cpu, &cg->cg_mask) && 1373 pri < PRI_MIN_IDLE) 1374 break; 1375 } 1376 if (cpu > cg->cg_last) { 1377 SCHED_STAT_INC(pickcpu_idle_affinity); 1378 return (ts->ts_cpu); 1379 } 1380 } else { 1381 SCHED_STAT_INC(pickcpu_idle_affinity); 1382 return (ts->ts_cpu); 1383 } 1384 } 1385 llc: 1386 /* 1387 * Search for the last level cache CPU group in the tree. 1388 * Skip SMT, identical groups and caches with expired affinity. 1389 * Interrupt threads affinity is explicit and never expires. 1390 */ 1391 for (ccg = NULL; cg != NULL; cg = cg->cg_parent) { 1392 if (cg->cg_flags & CG_FLAG_THREAD) 1393 continue; 1394 if (cg->cg_children == 1 || cg->cg_count == 1) 1395 continue; 1396 if (cg->cg_level == CG_SHARE_NONE || 1397 (!intr && !SCHED_AFFINITY(ts, cg->cg_level))) 1398 continue; 1399 ccg = cg; 1400 } 1401 /* Found LLC shared by all CPUs, so do a global search. */ 1402 if (ccg == cpu_top) 1403 ccg = NULL; 1404 cpu = -1; 1405 mask = &td->td_cpuset->cs_mask; 1406 pri = td->td_priority; 1407 r = TD_IS_RUNNING(td); 1408 /* 1409 * Try hard to keep interrupts within found LLC. Search the LLC for 1410 * the least loaded CPU we can run now. For NUMA systems it should 1411 * be within target domain, and it also reduces scheduling overhead. 1412 */ 1413 if (ccg != NULL && intr) { 1414 cpu = sched_lowest(ccg, mask, pri, INT_MAX, ts->ts_cpu, r); 1415 if (cpu >= 0) 1416 SCHED_STAT_INC(pickcpu_intrbind); 1417 } else 1418 /* Search the LLC for the least loaded idle CPU we can run now. */ 1419 if (ccg != NULL) { 1420 cpu = sched_lowest(ccg, mask, max(pri, PRI_MAX_TIMESHARE), 1421 INT_MAX, ts->ts_cpu, r); 1422 if (cpu >= 0) 1423 SCHED_STAT_INC(pickcpu_affinity); 1424 } 1425 /* Search globally for the least loaded CPU we can run now. */ 1426 if (cpu < 0) { 1427 cpu = sched_lowest(cpu_top, mask, pri, INT_MAX, ts->ts_cpu, r); 1428 if (cpu >= 0) 1429 SCHED_STAT_INC(pickcpu_lowest); 1430 } 1431 /* Search globally for the least loaded CPU. */ 1432 if (cpu < 0) { 1433 cpu = sched_lowest(cpu_top, mask, -1, INT_MAX, ts->ts_cpu, r); 1434 if (cpu >= 0) 1435 SCHED_STAT_INC(pickcpu_lowest); 1436 } 1437 KASSERT(cpu >= 0, ("sched_pickcpu: Failed to find a cpu.")); 1438 KASSERT(!CPU_ABSENT(cpu), ("sched_pickcpu: Picked absent CPU %d.", cpu)); 1439 /* 1440 * Compare the lowest loaded cpu to current cpu. 1441 */ 1442 tdq = TDQ_CPU(cpu); 1443 if (THREAD_CAN_SCHED(td, self) && TDQ_SELF()->tdq_lowpri > pri && 1444 atomic_load_char(&tdq->tdq_lowpri) < PRI_MIN_IDLE && 1445 TDQ_LOAD(TDQ_SELF()) <= TDQ_LOAD(tdq) + 1) { 1446 SCHED_STAT_INC(pickcpu_local); 1447 cpu = self; 1448 } 1449 if (cpu != ts->ts_cpu) 1450 SCHED_STAT_INC(pickcpu_migration); 1451 return (cpu); 1452 } 1453 #endif 1454 1455 /* 1456 * Pick the highest priority task we have and return it. 1457 */ 1458 static struct thread * 1459 tdq_choose(struct tdq *tdq) 1460 { 1461 struct thread *td; 1462 1463 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 1464 td = runq_choose(&tdq->tdq_realtime); 1465 if (td != NULL) 1466 return (td); 1467 td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx); 1468 if (td != NULL) { 1469 KASSERT(td->td_priority >= PRI_MIN_BATCH, 1470 ("tdq_choose: Invalid priority on timeshare queue %d", 1471 td->td_priority)); 1472 return (td); 1473 } 1474 td = runq_choose(&tdq->tdq_idle); 1475 if (td != NULL) { 1476 KASSERT(td->td_priority >= PRI_MIN_IDLE, 1477 ("tdq_choose: Invalid priority on idle queue %d", 1478 td->td_priority)); 1479 return (td); 1480 } 1481 1482 return (NULL); 1483 } 1484 1485 /* 1486 * Initialize a thread queue. 1487 */ 1488 static void 1489 tdq_setup(struct tdq *tdq, int id) 1490 { 1491 1492 if (bootverbose) 1493 printf("ULE: setup cpu %d\n", id); 1494 runq_init(&tdq->tdq_realtime); 1495 runq_init(&tdq->tdq_timeshare); 1496 runq_init(&tdq->tdq_idle); 1497 tdq->tdq_id = id; 1498 snprintf(tdq->tdq_name, sizeof(tdq->tdq_name), 1499 "sched lock %d", (int)TDQ_ID(tdq)); 1500 mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock", MTX_SPIN); 1501 #ifdef KTR 1502 snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname), 1503 "CPU %d load", (int)TDQ_ID(tdq)); 1504 #endif 1505 } 1506 1507 #ifdef SMP 1508 static void 1509 sched_setup_smp(void) 1510 { 1511 struct tdq *tdq; 1512 int i; 1513 1514 cpu_top = smp_topo(); 1515 CPU_FOREACH(i) { 1516 tdq = DPCPU_ID_PTR(i, tdq); 1517 tdq_setup(tdq, i); 1518 tdq->tdq_cg = smp_topo_find(cpu_top, i); 1519 if (tdq->tdq_cg == NULL) 1520 panic("Can't find cpu group for %d\n", i); 1521 DPCPU_ID_SET(i, randomval, i * 69069 + 5); 1522 } 1523 PCPU_SET(sched, DPCPU_PTR(tdq)); 1524 balance_tdq = TDQ_SELF(); 1525 } 1526 #endif 1527 1528 /* 1529 * Setup the thread queues and initialize the topology based on MD 1530 * information. 1531 */ 1532 static void 1533 sched_setup(void *dummy) 1534 { 1535 struct tdq *tdq; 1536 1537 #ifdef SMP 1538 sched_setup_smp(); 1539 #else 1540 tdq_setup(TDQ_SELF(), 0); 1541 #endif 1542 tdq = TDQ_SELF(); 1543 1544 /* Add thread0's load since it's running. */ 1545 TDQ_LOCK(tdq); 1546 thread0.td_lock = TDQ_LOCKPTR(tdq); 1547 tdq_load_add(tdq, &thread0); 1548 tdq->tdq_curthread = &thread0; 1549 tdq->tdq_lowpri = thread0.td_priority; 1550 TDQ_UNLOCK(tdq); 1551 } 1552 1553 /* 1554 * This routine determines time constants after stathz and hz are setup. 1555 */ 1556 /* ARGSUSED */ 1557 static void 1558 sched_initticks(void *dummy) 1559 { 1560 int incr; 1561 1562 realstathz = stathz ? stathz : hz; 1563 sched_slice = realstathz / SCHED_SLICE_DEFAULT_DIVISOR; 1564 sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR; 1565 hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) / 1566 realstathz); 1567 1568 /* 1569 * tickincr is shifted out by 10 to avoid rounding errors due to 1570 * hz not being evenly divisible by stathz on all platforms. 1571 */ 1572 incr = (hz << SCHED_TICK_SHIFT) / realstathz; 1573 /* 1574 * This does not work for values of stathz that are more than 1575 * 1 << SCHED_TICK_SHIFT * hz. In practice this does not happen. 1576 */ 1577 if (incr == 0) 1578 incr = 1; 1579 tickincr = incr; 1580 #ifdef SMP 1581 /* 1582 * Set the default balance interval now that we know 1583 * what realstathz is. 1584 */ 1585 balance_interval = realstathz; 1586 balance_ticks = balance_interval; 1587 affinity = SCHED_AFFINITY_DEFAULT; 1588 #endif 1589 if (sched_idlespinthresh < 0) 1590 sched_idlespinthresh = 2 * max(10000, 6 * hz) / realstathz; 1591 } 1592 1593 /* 1594 * This is the core of the interactivity algorithm. Determines a score based 1595 * on past behavior. It is the ratio of sleep time to run time scaled to 1596 * a [0, 100] integer. This is the voluntary sleep time of a process, which 1597 * differs from the cpu usage because it does not account for time spent 1598 * waiting on a run-queue. Would be prettier if we had floating point. 1599 * 1600 * When a thread's sleep time is greater than its run time the 1601 * calculation is: 1602 * 1603 * scaling factor 1604 * interactivity score = --------------------- 1605 * sleep time / run time 1606 * 1607 * 1608 * When a thread's run time is greater than its sleep time the 1609 * calculation is: 1610 * 1611 * scaling factor 1612 * interactivity score = 2 * scaling factor - --------------------- 1613 * run time / sleep time 1614 */ 1615 static int 1616 sched_interact_score(struct thread *td) 1617 { 1618 struct td_sched *ts; 1619 int div; 1620 1621 ts = td_get_sched(td); 1622 /* 1623 * The score is only needed if this is likely to be an interactive 1624 * task. Don't go through the expense of computing it if there's 1625 * no chance. 1626 */ 1627 if (sched_interact <= SCHED_INTERACT_HALF && 1628 ts->ts_runtime >= ts->ts_slptime) 1629 return (SCHED_INTERACT_HALF); 1630 1631 if (ts->ts_runtime > ts->ts_slptime) { 1632 div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF); 1633 return (SCHED_INTERACT_HALF + 1634 (SCHED_INTERACT_HALF - (ts->ts_slptime / div))); 1635 } 1636 if (ts->ts_slptime > ts->ts_runtime) { 1637 div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF); 1638 return (ts->ts_runtime / div); 1639 } 1640 /* runtime == slptime */ 1641 if (ts->ts_runtime) 1642 return (SCHED_INTERACT_HALF); 1643 1644 /* 1645 * This can happen if slptime and runtime are 0. 1646 */ 1647 return (0); 1648 1649 } 1650 1651 /* 1652 * Scale the scheduling priority according to the "interactivity" of this 1653 * process. 1654 */ 1655 static void 1656 sched_priority(struct thread *td) 1657 { 1658 u_int pri, score; 1659 1660 if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE) 1661 return; 1662 /* 1663 * If the score is interactive we place the thread in the realtime 1664 * queue with a priority that is less than kernel and interrupt 1665 * priorities. These threads are not subject to nice restrictions. 1666 * 1667 * Scores greater than this are placed on the normal timeshare queue 1668 * where the priority is partially decided by the most recent cpu 1669 * utilization and the rest is decided by nice value. 1670 * 1671 * The nice value of the process has a linear effect on the calculated 1672 * score. Negative nice values make it easier for a thread to be 1673 * considered interactive. 1674 */ 1675 score = imax(0, sched_interact_score(td) + td->td_proc->p_nice); 1676 if (score < sched_interact) { 1677 pri = PRI_MIN_INTERACT; 1678 pri += (PRI_MAX_INTERACT - PRI_MIN_INTERACT + 1) * score / 1679 sched_interact; 1680 KASSERT(pri >= PRI_MIN_INTERACT && pri <= PRI_MAX_INTERACT, 1681 ("sched_priority: invalid interactive priority %u score %u", 1682 pri, score)); 1683 } else { 1684 pri = SCHED_PRI_MIN; 1685 if (td_get_sched(td)->ts_ticks) 1686 pri += min(SCHED_PRI_TICKS(td_get_sched(td)), 1687 SCHED_PRI_RANGE - 1); 1688 pri += SCHED_PRI_NICE(td->td_proc->p_nice); 1689 KASSERT(pri >= PRI_MIN_BATCH && pri <= PRI_MAX_BATCH, 1690 ("sched_priority: invalid priority %u: nice %d, " 1691 "ticks %d ftick %d ltick %d tick pri %d", 1692 pri, td->td_proc->p_nice, td_get_sched(td)->ts_ticks, 1693 td_get_sched(td)->ts_ftick, td_get_sched(td)->ts_ltick, 1694 SCHED_PRI_TICKS(td_get_sched(td)))); 1695 } 1696 sched_user_prio(td, pri); 1697 1698 return; 1699 } 1700 1701 /* 1702 * This routine enforces a maximum limit on the amount of scheduling history 1703 * kept. It is called after either the slptime or runtime is adjusted. This 1704 * function is ugly due to integer math. 1705 */ 1706 static void 1707 sched_interact_update(struct thread *td) 1708 { 1709 struct td_sched *ts; 1710 u_int sum; 1711 1712 ts = td_get_sched(td); 1713 sum = ts->ts_runtime + ts->ts_slptime; 1714 if (sum < SCHED_SLP_RUN_MAX) 1715 return; 1716 /* 1717 * This only happens from two places: 1718 * 1) We have added an unusual amount of run time from fork_exit. 1719 * 2) We have added an unusual amount of sleep time from sched_sleep(). 1720 */ 1721 if (sum > SCHED_SLP_RUN_MAX * 2) { 1722 if (ts->ts_runtime > ts->ts_slptime) { 1723 ts->ts_runtime = SCHED_SLP_RUN_MAX; 1724 ts->ts_slptime = 1; 1725 } else { 1726 ts->ts_slptime = SCHED_SLP_RUN_MAX; 1727 ts->ts_runtime = 1; 1728 } 1729 return; 1730 } 1731 /* 1732 * If we have exceeded by more than 1/5th then the algorithm below 1733 * will not bring us back into range. Dividing by two here forces 1734 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX] 1735 */ 1736 if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) { 1737 ts->ts_runtime /= 2; 1738 ts->ts_slptime /= 2; 1739 return; 1740 } 1741 ts->ts_runtime = (ts->ts_runtime / 5) * 4; 1742 ts->ts_slptime = (ts->ts_slptime / 5) * 4; 1743 } 1744 1745 /* 1746 * Scale back the interactivity history when a child thread is created. The 1747 * history is inherited from the parent but the thread may behave totally 1748 * differently. For example, a shell spawning a compiler process. We want 1749 * to learn that the compiler is behaving badly very quickly. 1750 */ 1751 static void 1752 sched_interact_fork(struct thread *td) 1753 { 1754 struct td_sched *ts; 1755 int ratio; 1756 int sum; 1757 1758 ts = td_get_sched(td); 1759 sum = ts->ts_runtime + ts->ts_slptime; 1760 if (sum > SCHED_SLP_RUN_FORK) { 1761 ratio = sum / SCHED_SLP_RUN_FORK; 1762 ts->ts_runtime /= ratio; 1763 ts->ts_slptime /= ratio; 1764 } 1765 } 1766 1767 /* 1768 * Called from proc0_init() to setup the scheduler fields. 1769 */ 1770 void 1771 schedinit(void) 1772 { 1773 struct td_sched *ts0; 1774 1775 /* 1776 * Set up the scheduler specific parts of thread0. 1777 */ 1778 ts0 = td_get_sched(&thread0); 1779 ts0->ts_ltick = ticks; 1780 ts0->ts_ftick = ticks; 1781 ts0->ts_slice = 0; 1782 ts0->ts_cpu = curcpu; /* set valid CPU number */ 1783 } 1784 1785 /* 1786 * schedinit_ap() is needed prior to calling sched_throw(NULL) to ensure that 1787 * the pcpu requirements are met for any calls in the period between curthread 1788 * initialization and sched_throw(). One can safely add threads to the queue 1789 * before sched_throw(), for instance, as long as the thread lock is setup 1790 * correctly. 1791 * 1792 * TDQ_SELF() relies on the below sched pcpu setting; it may be used only 1793 * after schedinit_ap(). 1794 */ 1795 void 1796 schedinit_ap(void) 1797 { 1798 1799 #ifdef SMP 1800 PCPU_SET(sched, DPCPU_PTR(tdq)); 1801 #endif 1802 PCPU_GET(idlethread)->td_lock = TDQ_LOCKPTR(TDQ_SELF()); 1803 } 1804 1805 /* 1806 * This is only somewhat accurate since given many processes of the same 1807 * priority they will switch when their slices run out, which will be 1808 * at most sched_slice stathz ticks. 1809 */ 1810 int 1811 sched_rr_interval(void) 1812 { 1813 1814 /* Convert sched_slice from stathz to hz. */ 1815 return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz)); 1816 } 1817 1818 /* 1819 * Update the percent cpu tracking information when it is requested or 1820 * the total history exceeds the maximum. We keep a sliding history of 1821 * tick counts that slowly decays. This is less precise than the 4BSD 1822 * mechanism since it happens with less regular and frequent events. 1823 */ 1824 static void 1825 sched_pctcpu_update(struct td_sched *ts, int run) 1826 { 1827 int t = ticks; 1828 1829 /* 1830 * The signed difference may be negative if the thread hasn't run for 1831 * over half of the ticks rollover period. 1832 */ 1833 if ((u_int)(t - ts->ts_ltick) >= SCHED_TICK_TARG) { 1834 ts->ts_ticks = 0; 1835 ts->ts_ftick = t - SCHED_TICK_TARG; 1836 } else if (t - ts->ts_ftick >= SCHED_TICK_MAX) { 1837 ts->ts_ticks = (ts->ts_ticks / (ts->ts_ltick - ts->ts_ftick)) * 1838 (ts->ts_ltick - (t - SCHED_TICK_TARG)); 1839 ts->ts_ftick = t - SCHED_TICK_TARG; 1840 } 1841 if (run) 1842 ts->ts_ticks += (t - ts->ts_ltick) << SCHED_TICK_SHIFT; 1843 ts->ts_ltick = t; 1844 } 1845 1846 /* 1847 * Adjust the priority of a thread. Move it to the appropriate run-queue 1848 * if necessary. This is the back-end for several priority related 1849 * functions. 1850 */ 1851 static void 1852 sched_thread_priority(struct thread *td, u_char prio) 1853 { 1854 struct tdq *tdq; 1855 int oldpri; 1856 1857 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio", 1858 "prio:%d", td->td_priority, "new prio:%d", prio, 1859 KTR_ATTR_LINKED, sched_tdname(curthread)); 1860 SDT_PROBE3(sched, , , change__pri, td, td->td_proc, prio); 1861 if (td != curthread && prio < td->td_priority) { 1862 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread), 1863 "lend prio", "prio:%d", td->td_priority, "new prio:%d", 1864 prio, KTR_ATTR_LINKED, sched_tdname(td)); 1865 SDT_PROBE4(sched, , , lend__pri, td, td->td_proc, prio, 1866 curthread); 1867 } 1868 THREAD_LOCK_ASSERT(td, MA_OWNED); 1869 if (td->td_priority == prio) 1870 return; 1871 /* 1872 * If the priority has been elevated due to priority 1873 * propagation, we may have to move ourselves to a new 1874 * queue. This could be optimized to not re-add in some 1875 * cases. 1876 */ 1877 if (TD_ON_RUNQ(td) && prio < td->td_priority) { 1878 sched_rem(td); 1879 td->td_priority = prio; 1880 sched_add(td, SRQ_BORROWING | SRQ_HOLDTD); 1881 return; 1882 } 1883 /* 1884 * If the thread is currently running we may have to adjust the lowpri 1885 * information so other cpus are aware of our current priority. 1886 */ 1887 if (TD_IS_RUNNING(td)) { 1888 tdq = TDQ_CPU(td_get_sched(td)->ts_cpu); 1889 oldpri = td->td_priority; 1890 td->td_priority = prio; 1891 if (prio < tdq->tdq_lowpri) 1892 tdq->tdq_lowpri = prio; 1893 else if (tdq->tdq_lowpri == oldpri) 1894 tdq_setlowpri(tdq, td); 1895 return; 1896 } 1897 td->td_priority = prio; 1898 } 1899 1900 /* 1901 * Update a thread's priority when it is lent another thread's 1902 * priority. 1903 */ 1904 void 1905 sched_lend_prio(struct thread *td, u_char prio) 1906 { 1907 1908 td->td_flags |= TDF_BORROWING; 1909 sched_thread_priority(td, prio); 1910 } 1911 1912 /* 1913 * Restore a thread's priority when priority propagation is 1914 * over. The prio argument is the minimum priority the thread 1915 * needs to have to satisfy other possible priority lending 1916 * requests. If the thread's regular priority is less 1917 * important than prio, the thread will keep a priority boost 1918 * of prio. 1919 */ 1920 void 1921 sched_unlend_prio(struct thread *td, u_char prio) 1922 { 1923 u_char base_pri; 1924 1925 if (td->td_base_pri >= PRI_MIN_TIMESHARE && 1926 td->td_base_pri <= PRI_MAX_TIMESHARE) 1927 base_pri = td->td_user_pri; 1928 else 1929 base_pri = td->td_base_pri; 1930 if (prio >= base_pri) { 1931 td->td_flags &= ~TDF_BORROWING; 1932 sched_thread_priority(td, base_pri); 1933 } else 1934 sched_lend_prio(td, prio); 1935 } 1936 1937 /* 1938 * Standard entry for setting the priority to an absolute value. 1939 */ 1940 void 1941 sched_prio(struct thread *td, u_char prio) 1942 { 1943 u_char oldprio; 1944 1945 /* First, update the base priority. */ 1946 td->td_base_pri = prio; 1947 1948 /* 1949 * If the thread is borrowing another thread's priority, don't 1950 * ever lower the priority. 1951 */ 1952 if (td->td_flags & TDF_BORROWING && td->td_priority < prio) 1953 return; 1954 1955 /* Change the real priority. */ 1956 oldprio = td->td_priority; 1957 sched_thread_priority(td, prio); 1958 1959 /* 1960 * If the thread is on a turnstile, then let the turnstile update 1961 * its state. 1962 */ 1963 if (TD_ON_LOCK(td) && oldprio != prio) 1964 turnstile_adjust(td, oldprio); 1965 } 1966 1967 /* 1968 * Set the base interrupt thread priority. 1969 */ 1970 void 1971 sched_ithread_prio(struct thread *td, u_char prio) 1972 { 1973 THREAD_LOCK_ASSERT(td, MA_OWNED); 1974 MPASS(td->td_pri_class == PRI_ITHD); 1975 td->td_base_ithread_pri = prio; 1976 sched_prio(td, prio); 1977 } 1978 1979 /* 1980 * Set the base user priority, does not effect current running priority. 1981 */ 1982 void 1983 sched_user_prio(struct thread *td, u_char prio) 1984 { 1985 1986 td->td_base_user_pri = prio; 1987 if (td->td_lend_user_pri <= prio) 1988 return; 1989 td->td_user_pri = prio; 1990 } 1991 1992 void 1993 sched_lend_user_prio(struct thread *td, u_char prio) 1994 { 1995 1996 THREAD_LOCK_ASSERT(td, MA_OWNED); 1997 td->td_lend_user_pri = prio; 1998 td->td_user_pri = min(prio, td->td_base_user_pri); 1999 if (td->td_priority > td->td_user_pri) 2000 sched_prio(td, td->td_user_pri); 2001 else if (td->td_priority != td->td_user_pri) 2002 ast_sched_locked(td, TDA_SCHED); 2003 } 2004 2005 /* 2006 * Like the above but first check if there is anything to do. 2007 */ 2008 void 2009 sched_lend_user_prio_cond(struct thread *td, u_char prio) 2010 { 2011 2012 if (td->td_lend_user_pri != prio) 2013 goto lend; 2014 if (td->td_user_pri != min(prio, td->td_base_user_pri)) 2015 goto lend; 2016 if (td->td_priority != td->td_user_pri) 2017 goto lend; 2018 return; 2019 2020 lend: 2021 thread_lock(td); 2022 sched_lend_user_prio(td, prio); 2023 thread_unlock(td); 2024 } 2025 2026 #ifdef SMP 2027 /* 2028 * This tdq is about to idle. Try to steal a thread from another CPU before 2029 * choosing the idle thread. 2030 */ 2031 static void 2032 tdq_trysteal(struct tdq *tdq) 2033 { 2034 struct cpu_group *cg, *parent; 2035 struct tdq *steal; 2036 cpuset_t mask; 2037 int cpu, i, goup; 2038 2039 if (smp_started == 0 || steal_idle == 0 || trysteal_limit == 0 || 2040 tdq->tdq_cg == NULL) 2041 return; 2042 CPU_FILL(&mask); 2043 CPU_CLR(PCPU_GET(cpuid), &mask); 2044 /* We don't want to be preempted while we're iterating. */ 2045 spinlock_enter(); 2046 TDQ_UNLOCK(tdq); 2047 for (i = 1, cg = tdq->tdq_cg, goup = 0; ; ) { 2048 cpu = sched_highest(cg, &mask, steal_thresh, 1); 2049 /* 2050 * If a thread was added while interrupts were disabled don't 2051 * steal one here. 2052 */ 2053 if (TDQ_LOAD(tdq) > 0) { 2054 TDQ_LOCK(tdq); 2055 break; 2056 } 2057 2058 /* 2059 * We found no CPU to steal from in this group. Escalate to 2060 * the parent and repeat. But if parent has only two children 2061 * groups we can avoid searching this group again by searching 2062 * the other one specifically and then escalating two levels. 2063 */ 2064 if (cpu == -1) { 2065 if (goup) { 2066 cg = cg->cg_parent; 2067 goup = 0; 2068 } 2069 if (++i > trysteal_limit) { 2070 TDQ_LOCK(tdq); 2071 break; 2072 } 2073 parent = cg->cg_parent; 2074 if (parent == NULL) { 2075 TDQ_LOCK(tdq); 2076 break; 2077 } 2078 if (parent->cg_children == 2) { 2079 if (cg == &parent->cg_child[0]) 2080 cg = &parent->cg_child[1]; 2081 else 2082 cg = &parent->cg_child[0]; 2083 goup = 1; 2084 } else 2085 cg = parent; 2086 continue; 2087 } 2088 steal = TDQ_CPU(cpu); 2089 /* 2090 * The data returned by sched_highest() is stale and 2091 * the chosen CPU no longer has an eligible thread. 2092 * At this point unconditionally exit the loop to bound 2093 * the time spent in the critcal section. 2094 */ 2095 if (TDQ_LOAD(steal) < steal_thresh || 2096 TDQ_TRANSFERABLE(steal) == 0) 2097 continue; 2098 /* 2099 * Try to lock both queues. If we are assigned a thread while 2100 * waited for the lock, switch to it now instead of stealing. 2101 * If we can't get the lock, then somebody likely got there 2102 * first. 2103 */ 2104 TDQ_LOCK(tdq); 2105 if (tdq->tdq_load > 0) 2106 break; 2107 if (TDQ_TRYLOCK_FLAGS(steal, MTX_DUPOK) == 0) 2108 break; 2109 /* 2110 * The data returned by sched_highest() is stale and 2111 * the chosen CPU no longer has an eligible thread. 2112 */ 2113 if (TDQ_LOAD(steal) < steal_thresh || 2114 TDQ_TRANSFERABLE(steal) == 0) { 2115 TDQ_UNLOCK(steal); 2116 break; 2117 } 2118 /* 2119 * If we fail to acquire one due to affinity restrictions, 2120 * bail out and let the idle thread to a more complete search 2121 * outside of a critical section. 2122 */ 2123 if (tdq_move(steal, tdq) == -1) { 2124 TDQ_UNLOCK(steal); 2125 break; 2126 } 2127 TDQ_UNLOCK(steal); 2128 break; 2129 } 2130 spinlock_exit(); 2131 } 2132 #endif 2133 2134 /* 2135 * Handle migration from sched_switch(). This happens only for 2136 * cpu binding. 2137 */ 2138 static struct mtx * 2139 sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags) 2140 { 2141 struct tdq *tdn; 2142 #ifdef SMP 2143 int lowpri; 2144 #endif 2145 2146 KASSERT(THREAD_CAN_MIGRATE(td) || 2147 (td_get_sched(td)->ts_flags & TSF_BOUND) != 0, 2148 ("Thread %p shouldn't migrate", td)); 2149 KASSERT(!CPU_ABSENT(td_get_sched(td)->ts_cpu), ("sched_switch_migrate: " 2150 "thread %s queued on absent CPU %d.", td->td_name, 2151 td_get_sched(td)->ts_cpu)); 2152 tdn = TDQ_CPU(td_get_sched(td)->ts_cpu); 2153 #ifdef SMP 2154 tdq_load_rem(tdq, td); 2155 /* 2156 * Do the lock dance required to avoid LOR. We have an 2157 * extra spinlock nesting from sched_switch() which will 2158 * prevent preemption while we're holding neither run-queue lock. 2159 */ 2160 TDQ_UNLOCK(tdq); 2161 TDQ_LOCK(tdn); 2162 lowpri = tdq_add(tdn, td, flags); 2163 tdq_notify(tdn, lowpri); 2164 TDQ_UNLOCK(tdn); 2165 TDQ_LOCK(tdq); 2166 #endif 2167 return (TDQ_LOCKPTR(tdn)); 2168 } 2169 2170 /* 2171 * thread_lock_unblock() that does not assume td_lock is blocked. 2172 */ 2173 static inline void 2174 thread_unblock_switch(struct thread *td, struct mtx *mtx) 2175 { 2176 atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock, 2177 (uintptr_t)mtx); 2178 } 2179 2180 /* 2181 * Switch threads. This function has to handle threads coming in while 2182 * blocked for some reason, running, or idle. It also must deal with 2183 * migrating a thread from one queue to another as running threads may 2184 * be assigned elsewhere via binding. 2185 */ 2186 void 2187 sched_switch(struct thread *td, int flags) 2188 { 2189 struct thread *newtd; 2190 struct tdq *tdq; 2191 struct td_sched *ts; 2192 struct mtx *mtx; 2193 int srqflag; 2194 int cpuid, preempted; 2195 #ifdef SMP 2196 int pickcpu; 2197 #endif 2198 2199 THREAD_LOCK_ASSERT(td, MA_OWNED); 2200 2201 cpuid = PCPU_GET(cpuid); 2202 tdq = TDQ_SELF(); 2203 ts = td_get_sched(td); 2204 sched_pctcpu_update(ts, 1); 2205 #ifdef SMP 2206 pickcpu = (td->td_flags & TDF_PICKCPU) != 0; 2207 if (pickcpu) 2208 ts->ts_rltick = ticks - affinity * MAX_CACHE_LEVELS; 2209 else 2210 ts->ts_rltick = ticks; 2211 #endif 2212 td->td_lastcpu = td->td_oncpu; 2213 preempted = (td->td_flags & TDF_SLICEEND) == 0 && 2214 (flags & SW_PREEMPT) != 0; 2215 td->td_flags &= ~(TDF_PICKCPU | TDF_SLICEEND); 2216 ast_unsched_locked(td, TDA_SCHED); 2217 td->td_owepreempt = 0; 2218 atomic_store_char(&tdq->tdq_owepreempt, 0); 2219 if (!TD_IS_IDLETHREAD(td)) 2220 TDQ_SWITCHCNT_INC(tdq); 2221 2222 /* 2223 * Always block the thread lock so we can drop the tdq lock early. 2224 */ 2225 mtx = thread_lock_block(td); 2226 spinlock_enter(); 2227 if (TD_IS_IDLETHREAD(td)) { 2228 MPASS(mtx == TDQ_LOCKPTR(tdq)); 2229 TD_SET_CAN_RUN(td); 2230 } else if (TD_IS_RUNNING(td)) { 2231 MPASS(mtx == TDQ_LOCKPTR(tdq)); 2232 srqflag = preempted ? 2233 SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED : 2234 SRQ_OURSELF|SRQ_YIELDING; 2235 #ifdef SMP 2236 if (THREAD_CAN_MIGRATE(td) && (!THREAD_CAN_SCHED(td, ts->ts_cpu) 2237 || pickcpu)) 2238 ts->ts_cpu = sched_pickcpu(td, 0); 2239 #endif 2240 if (ts->ts_cpu == cpuid) 2241 tdq_runq_add(tdq, td, srqflag); 2242 else 2243 mtx = sched_switch_migrate(tdq, td, srqflag); 2244 } else { 2245 /* This thread must be going to sleep. */ 2246 if (mtx != TDQ_LOCKPTR(tdq)) { 2247 mtx_unlock_spin(mtx); 2248 TDQ_LOCK(tdq); 2249 } 2250 tdq_load_rem(tdq, td); 2251 #ifdef SMP 2252 if (tdq->tdq_load == 0) 2253 tdq_trysteal(tdq); 2254 #endif 2255 } 2256 2257 #if (KTR_COMPILE & KTR_SCHED) != 0 2258 if (TD_IS_IDLETHREAD(td)) 2259 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "idle", 2260 "prio:%d", td->td_priority); 2261 else 2262 KTR_STATE3(KTR_SCHED, "thread", sched_tdname(td), KTDSTATE(td), 2263 "prio:%d", td->td_priority, "wmesg:\"%s\"", td->td_wmesg, 2264 "lockname:\"%s\"", td->td_lockname); 2265 #endif 2266 2267 /* 2268 * We enter here with the thread blocked and assigned to the 2269 * appropriate cpu run-queue or sleep-queue and with the current 2270 * thread-queue locked. 2271 */ 2272 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED); 2273 MPASS(td == tdq->tdq_curthread); 2274 newtd = choosethread(); 2275 sched_pctcpu_update(td_get_sched(newtd), 0); 2276 TDQ_UNLOCK(tdq); 2277 2278 /* 2279 * Call the MD code to switch contexts if necessary. 2280 */ 2281 if (td != newtd) { 2282 #ifdef HWPMC_HOOKS 2283 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 2284 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT); 2285 #endif 2286 SDT_PROBE2(sched, , , off__cpu, newtd, newtd->td_proc); 2287 2288 #ifdef KDTRACE_HOOKS 2289 /* 2290 * If DTrace has set the active vtime enum to anything 2291 * other than INACTIVE (0), then it should have set the 2292 * function to call. 2293 */ 2294 if (dtrace_vtime_active) 2295 (*dtrace_vtime_switch_func)(newtd); 2296 #endif 2297 td->td_oncpu = NOCPU; 2298 cpu_switch(td, newtd, mtx); 2299 cpuid = td->td_oncpu = PCPU_GET(cpuid); 2300 2301 SDT_PROBE0(sched, , , on__cpu); 2302 #ifdef HWPMC_HOOKS 2303 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 2304 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN); 2305 #endif 2306 } else { 2307 thread_unblock_switch(td, mtx); 2308 SDT_PROBE0(sched, , , remain__cpu); 2309 } 2310 KASSERT(curthread->td_md.md_spinlock_count == 1, 2311 ("invalid count %d", curthread->td_md.md_spinlock_count)); 2312 2313 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running", 2314 "prio:%d", td->td_priority); 2315 } 2316 2317 /* 2318 * Adjust thread priorities as a result of a nice request. 2319 */ 2320 void 2321 sched_nice(struct proc *p, int nice) 2322 { 2323 struct thread *td; 2324 2325 PROC_LOCK_ASSERT(p, MA_OWNED); 2326 2327 p->p_nice = nice; 2328 FOREACH_THREAD_IN_PROC(p, td) { 2329 thread_lock(td); 2330 sched_priority(td); 2331 sched_prio(td, td->td_base_user_pri); 2332 thread_unlock(td); 2333 } 2334 } 2335 2336 /* 2337 * Record the sleep time for the interactivity scorer. 2338 */ 2339 void 2340 sched_sleep(struct thread *td, int prio) 2341 { 2342 2343 THREAD_LOCK_ASSERT(td, MA_OWNED); 2344 2345 td->td_slptick = ticks; 2346 if (TD_IS_SUSPENDED(td) || prio >= PSOCK) 2347 td->td_flags |= TDF_CANSWAP; 2348 if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE) 2349 return; 2350 if (static_boost == 1 && prio) 2351 sched_prio(td, prio); 2352 else if (static_boost && td->td_priority > static_boost) 2353 sched_prio(td, static_boost); 2354 } 2355 2356 /* 2357 * Schedule a thread to resume execution and record how long it voluntarily 2358 * slept. We also update the pctcpu, interactivity, and priority. 2359 * 2360 * Requires the thread lock on entry, drops on exit. 2361 */ 2362 void 2363 sched_wakeup(struct thread *td, int srqflags) 2364 { 2365 struct td_sched *ts; 2366 int slptick; 2367 2368 THREAD_LOCK_ASSERT(td, MA_OWNED); 2369 ts = td_get_sched(td); 2370 td->td_flags &= ~TDF_CANSWAP; 2371 2372 /* 2373 * If we slept for more than a tick update our interactivity and 2374 * priority. 2375 */ 2376 slptick = td->td_slptick; 2377 td->td_slptick = 0; 2378 if (slptick && slptick != ticks) { 2379 ts->ts_slptime += (ticks - slptick) << SCHED_TICK_SHIFT; 2380 sched_interact_update(td); 2381 sched_pctcpu_update(ts, 0); 2382 } 2383 2384 /* 2385 * When resuming an idle ithread, restore its base ithread 2386 * priority. 2387 */ 2388 if (PRI_BASE(td->td_pri_class) == PRI_ITHD && 2389 td->td_priority != td->td_base_ithread_pri) 2390 sched_prio(td, td->td_base_ithread_pri); 2391 2392 /* 2393 * Reset the slice value since we slept and advanced the round-robin. 2394 */ 2395 ts->ts_slice = 0; 2396 sched_add(td, SRQ_BORING | srqflags); 2397 } 2398 2399 /* 2400 * Penalize the parent for creating a new child and initialize the child's 2401 * priority. 2402 */ 2403 void 2404 sched_fork(struct thread *td, struct thread *child) 2405 { 2406 THREAD_LOCK_ASSERT(td, MA_OWNED); 2407 sched_pctcpu_update(td_get_sched(td), 1); 2408 sched_fork_thread(td, child); 2409 /* 2410 * Penalize the parent and child for forking. 2411 */ 2412 sched_interact_fork(child); 2413 sched_priority(child); 2414 td_get_sched(td)->ts_runtime += tickincr; 2415 sched_interact_update(td); 2416 sched_priority(td); 2417 } 2418 2419 /* 2420 * Fork a new thread, may be within the same process. 2421 */ 2422 void 2423 sched_fork_thread(struct thread *td, struct thread *child) 2424 { 2425 struct td_sched *ts; 2426 struct td_sched *ts2; 2427 struct tdq *tdq; 2428 2429 tdq = TDQ_SELF(); 2430 THREAD_LOCK_ASSERT(td, MA_OWNED); 2431 /* 2432 * Initialize child. 2433 */ 2434 ts = td_get_sched(td); 2435 ts2 = td_get_sched(child); 2436 child->td_oncpu = NOCPU; 2437 child->td_lastcpu = NOCPU; 2438 child->td_lock = TDQ_LOCKPTR(tdq); 2439 child->td_cpuset = cpuset_ref(td->td_cpuset); 2440 child->td_domain.dr_policy = td->td_cpuset->cs_domain; 2441 ts2->ts_cpu = ts->ts_cpu; 2442 ts2->ts_flags = 0; 2443 /* 2444 * Grab our parents cpu estimation information. 2445 */ 2446 ts2->ts_ticks = ts->ts_ticks; 2447 ts2->ts_ltick = ts->ts_ltick; 2448 ts2->ts_ftick = ts->ts_ftick; 2449 /* 2450 * Do not inherit any borrowed priority from the parent. 2451 */ 2452 child->td_priority = child->td_base_pri; 2453 /* 2454 * And update interactivity score. 2455 */ 2456 ts2->ts_slptime = ts->ts_slptime; 2457 ts2->ts_runtime = ts->ts_runtime; 2458 /* Attempt to quickly learn interactivity. */ 2459 ts2->ts_slice = tdq_slice(tdq) - sched_slice_min; 2460 #ifdef KTR 2461 bzero(ts2->ts_name, sizeof(ts2->ts_name)); 2462 #endif 2463 } 2464 2465 /* 2466 * Adjust the priority class of a thread. 2467 */ 2468 void 2469 sched_class(struct thread *td, int class) 2470 { 2471 2472 THREAD_LOCK_ASSERT(td, MA_OWNED); 2473 if (td->td_pri_class == class) 2474 return; 2475 td->td_pri_class = class; 2476 } 2477 2478 /* 2479 * Return some of the child's priority and interactivity to the parent. 2480 */ 2481 void 2482 sched_exit(struct proc *p, struct thread *child) 2483 { 2484 struct thread *td; 2485 2486 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit", 2487 "prio:%d", child->td_priority); 2488 PROC_LOCK_ASSERT(p, MA_OWNED); 2489 td = FIRST_THREAD_IN_PROC(p); 2490 sched_exit_thread(td, child); 2491 } 2492 2493 /* 2494 * Penalize another thread for the time spent on this one. This helps to 2495 * worsen the priority and interactivity of processes which schedule batch 2496 * jobs such as make. This has little effect on the make process itself but 2497 * causes new processes spawned by it to receive worse scores immediately. 2498 */ 2499 void 2500 sched_exit_thread(struct thread *td, struct thread *child) 2501 { 2502 2503 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit", 2504 "prio:%d", child->td_priority); 2505 /* 2506 * Give the child's runtime to the parent without returning the 2507 * sleep time as a penalty to the parent. This causes shells that 2508 * launch expensive things to mark their children as expensive. 2509 */ 2510 thread_lock(td); 2511 td_get_sched(td)->ts_runtime += td_get_sched(child)->ts_runtime; 2512 sched_interact_update(td); 2513 sched_priority(td); 2514 thread_unlock(td); 2515 } 2516 2517 void 2518 sched_preempt(struct thread *td) 2519 { 2520 struct tdq *tdq; 2521 int flags; 2522 2523 SDT_PROBE2(sched, , , surrender, td, td->td_proc); 2524 2525 thread_lock(td); 2526 tdq = TDQ_SELF(); 2527 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2528 if (td->td_priority > tdq->tdq_lowpri) { 2529 if (td->td_critnest == 1) { 2530 flags = SW_INVOL | SW_PREEMPT; 2531 flags |= TD_IS_IDLETHREAD(td) ? SWT_REMOTEWAKEIDLE : 2532 SWT_REMOTEPREEMPT; 2533 mi_switch(flags); 2534 /* Switch dropped thread lock. */ 2535 return; 2536 } 2537 td->td_owepreempt = 1; 2538 } else { 2539 tdq->tdq_owepreempt = 0; 2540 } 2541 thread_unlock(td); 2542 } 2543 2544 /* 2545 * Fix priorities on return to user-space. Priorities may be elevated due 2546 * to static priorities in msleep() or similar. 2547 */ 2548 void 2549 sched_userret_slowpath(struct thread *td) 2550 { 2551 2552 thread_lock(td); 2553 td->td_priority = td->td_user_pri; 2554 td->td_base_pri = td->td_user_pri; 2555 tdq_setlowpri(TDQ_SELF(), td); 2556 thread_unlock(td); 2557 } 2558 2559 SCHED_STAT_DEFINE(ithread_demotions, "Interrupt thread priority demotions"); 2560 SCHED_STAT_DEFINE(ithread_preemptions, 2561 "Interrupt thread preemptions due to time-sharing"); 2562 2563 /* 2564 * Return time slice for a given thread. For ithreads this is 2565 * sched_slice. For other threads it is tdq_slice(tdq). 2566 */ 2567 static inline int 2568 td_slice(struct thread *td, struct tdq *tdq) 2569 { 2570 if (PRI_BASE(td->td_pri_class) == PRI_ITHD) 2571 return (sched_slice); 2572 return (tdq_slice(tdq)); 2573 } 2574 2575 /* 2576 * Handle a stathz tick. This is really only relevant for timeshare 2577 * and interrupt threads. 2578 */ 2579 void 2580 sched_clock(struct thread *td, int cnt) 2581 { 2582 struct tdq *tdq; 2583 struct td_sched *ts; 2584 2585 THREAD_LOCK_ASSERT(td, MA_OWNED); 2586 tdq = TDQ_SELF(); 2587 #ifdef SMP 2588 /* 2589 * We run the long term load balancer infrequently on the first cpu. 2590 */ 2591 if (balance_tdq == tdq && smp_started != 0 && rebalance != 0 && 2592 balance_ticks != 0) { 2593 balance_ticks -= cnt; 2594 if (balance_ticks <= 0) 2595 sched_balance(); 2596 } 2597 #endif 2598 /* 2599 * Save the old switch count so we have a record of the last ticks 2600 * activity. Initialize the new switch count based on our load. 2601 * If there is some activity seed it to reflect that. 2602 */ 2603 tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt; 2604 tdq->tdq_switchcnt = tdq->tdq_load; 2605 2606 /* 2607 * Advance the insert index once for each tick to ensure that all 2608 * threads get a chance to run. 2609 */ 2610 if (tdq->tdq_idx == tdq->tdq_ridx) { 2611 tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS; 2612 if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx])) 2613 tdq->tdq_ridx = tdq->tdq_idx; 2614 } 2615 ts = td_get_sched(td); 2616 sched_pctcpu_update(ts, 1); 2617 if ((td->td_pri_class & PRI_FIFO_BIT) || TD_IS_IDLETHREAD(td)) 2618 return; 2619 2620 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) { 2621 /* 2622 * We used a tick; charge it to the thread so 2623 * that we can compute our interactivity. 2624 */ 2625 td_get_sched(td)->ts_runtime += tickincr * cnt; 2626 sched_interact_update(td); 2627 sched_priority(td); 2628 } 2629 2630 /* 2631 * Force a context switch if the current thread has used up a full 2632 * time slice (default is 100ms). 2633 */ 2634 ts->ts_slice += cnt; 2635 if (ts->ts_slice >= td_slice(td, tdq)) { 2636 ts->ts_slice = 0; 2637 2638 /* 2639 * If an ithread uses a full quantum, demote its 2640 * priority and preempt it. 2641 */ 2642 if (PRI_BASE(td->td_pri_class) == PRI_ITHD) { 2643 SCHED_STAT_INC(ithread_preemptions); 2644 td->td_owepreempt = 1; 2645 if (td->td_base_pri + RQ_PPQ < PRI_MAX_ITHD) { 2646 SCHED_STAT_INC(ithread_demotions); 2647 sched_prio(td, td->td_base_pri + RQ_PPQ); 2648 } 2649 } else { 2650 ast_sched_locked(td, TDA_SCHED); 2651 td->td_flags |= TDF_SLICEEND; 2652 } 2653 } 2654 } 2655 2656 u_int 2657 sched_estcpu(struct thread *td __unused) 2658 { 2659 2660 return (0); 2661 } 2662 2663 /* 2664 * Return whether the current CPU has runnable tasks. Used for in-kernel 2665 * cooperative idle threads. 2666 */ 2667 int 2668 sched_runnable(void) 2669 { 2670 struct tdq *tdq; 2671 int load; 2672 2673 load = 1; 2674 2675 tdq = TDQ_SELF(); 2676 if ((curthread->td_flags & TDF_IDLETD) != 0) { 2677 if (TDQ_LOAD(tdq) > 0) 2678 goto out; 2679 } else 2680 if (TDQ_LOAD(tdq) - 1 > 0) 2681 goto out; 2682 load = 0; 2683 out: 2684 return (load); 2685 } 2686 2687 /* 2688 * Choose the highest priority thread to run. The thread is removed from 2689 * the run-queue while running however the load remains. 2690 */ 2691 struct thread * 2692 sched_choose(void) 2693 { 2694 struct thread *td; 2695 struct tdq *tdq; 2696 2697 tdq = TDQ_SELF(); 2698 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2699 td = tdq_choose(tdq); 2700 if (td != NULL) { 2701 tdq_runq_rem(tdq, td); 2702 tdq->tdq_lowpri = td->td_priority; 2703 } else { 2704 tdq->tdq_lowpri = PRI_MAX_IDLE; 2705 td = PCPU_GET(idlethread); 2706 } 2707 tdq->tdq_curthread = td; 2708 return (td); 2709 } 2710 2711 /* 2712 * Set owepreempt if the currently running thread has lower priority than "pri". 2713 * Preemption never happens directly in ULE, we always request it once we exit a 2714 * critical section. 2715 */ 2716 static void 2717 sched_setpreempt(int pri) 2718 { 2719 struct thread *ctd; 2720 int cpri; 2721 2722 ctd = curthread; 2723 THREAD_LOCK_ASSERT(ctd, MA_OWNED); 2724 2725 cpri = ctd->td_priority; 2726 if (pri < cpri) 2727 ast_sched_locked(ctd, TDA_SCHED); 2728 if (KERNEL_PANICKED() || pri >= cpri || cold || TD_IS_INHIBITED(ctd)) 2729 return; 2730 if (!sched_shouldpreempt(pri, cpri, 0)) 2731 return; 2732 ctd->td_owepreempt = 1; 2733 } 2734 2735 /* 2736 * Add a thread to a thread queue. Select the appropriate runq and add the 2737 * thread to it. This is the internal function called when the tdq is 2738 * predetermined. 2739 */ 2740 static int 2741 tdq_add(struct tdq *tdq, struct thread *td, int flags) 2742 { 2743 int lowpri; 2744 2745 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2746 THREAD_LOCK_BLOCKED_ASSERT(td, MA_OWNED); 2747 KASSERT((td->td_inhibitors == 0), 2748 ("sched_add: trying to run inhibited thread")); 2749 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)), 2750 ("sched_add: bad thread state")); 2751 KASSERT(td->td_flags & TDF_INMEM, 2752 ("sched_add: thread swapped out")); 2753 2754 lowpri = tdq->tdq_lowpri; 2755 if (td->td_priority < lowpri) 2756 tdq->tdq_lowpri = td->td_priority; 2757 tdq_runq_add(tdq, td, flags); 2758 tdq_load_add(tdq, td); 2759 return (lowpri); 2760 } 2761 2762 /* 2763 * Select the target thread queue and add a thread to it. Request 2764 * preemption or IPI a remote processor if required. 2765 * 2766 * Requires the thread lock on entry, drops on exit. 2767 */ 2768 void 2769 sched_add(struct thread *td, int flags) 2770 { 2771 struct tdq *tdq; 2772 #ifdef SMP 2773 int cpu, lowpri; 2774 #endif 2775 2776 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add", 2777 "prio:%d", td->td_priority, KTR_ATTR_LINKED, 2778 sched_tdname(curthread)); 2779 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup", 2780 KTR_ATTR_LINKED, sched_tdname(td)); 2781 SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL, 2782 flags & SRQ_PREEMPTED); 2783 THREAD_LOCK_ASSERT(td, MA_OWNED); 2784 /* 2785 * Recalculate the priority before we select the target cpu or 2786 * run-queue. 2787 */ 2788 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) 2789 sched_priority(td); 2790 #ifdef SMP 2791 /* 2792 * Pick the destination cpu and if it isn't ours transfer to the 2793 * target cpu. 2794 */ 2795 cpu = sched_pickcpu(td, flags); 2796 tdq = sched_setcpu(td, cpu, flags); 2797 lowpri = tdq_add(tdq, td, flags); 2798 if (cpu != PCPU_GET(cpuid)) 2799 tdq_notify(tdq, lowpri); 2800 else if (!(flags & SRQ_YIELDING)) 2801 sched_setpreempt(td->td_priority); 2802 #else 2803 tdq = TDQ_SELF(); 2804 /* 2805 * Now that the thread is moving to the run-queue, set the lock 2806 * to the scheduler's lock. 2807 */ 2808 if (td->td_lock != TDQ_LOCKPTR(tdq)) { 2809 TDQ_LOCK(tdq); 2810 if ((flags & SRQ_HOLD) != 0) 2811 td->td_lock = TDQ_LOCKPTR(tdq); 2812 else 2813 thread_lock_set(td, TDQ_LOCKPTR(tdq)); 2814 } 2815 (void)tdq_add(tdq, td, flags); 2816 if (!(flags & SRQ_YIELDING)) 2817 sched_setpreempt(td->td_priority); 2818 #endif 2819 if (!(flags & SRQ_HOLDTD)) 2820 thread_unlock(td); 2821 } 2822 2823 /* 2824 * Remove a thread from a run-queue without running it. This is used 2825 * when we're stealing a thread from a remote queue. Otherwise all threads 2826 * exit by calling sched_exit_thread() and sched_throw() themselves. 2827 */ 2828 void 2829 sched_rem(struct thread *td) 2830 { 2831 struct tdq *tdq; 2832 2833 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem", 2834 "prio:%d", td->td_priority); 2835 SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL); 2836 tdq = TDQ_CPU(td_get_sched(td)->ts_cpu); 2837 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2838 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2839 KASSERT(TD_ON_RUNQ(td), 2840 ("sched_rem: thread not on run queue")); 2841 tdq_runq_rem(tdq, td); 2842 tdq_load_rem(tdq, td); 2843 TD_SET_CAN_RUN(td); 2844 if (td->td_priority == tdq->tdq_lowpri) 2845 tdq_setlowpri(tdq, NULL); 2846 } 2847 2848 /* 2849 * Fetch cpu utilization information. Updates on demand. 2850 */ 2851 fixpt_t 2852 sched_pctcpu(struct thread *td) 2853 { 2854 fixpt_t pctcpu; 2855 struct td_sched *ts; 2856 2857 pctcpu = 0; 2858 ts = td_get_sched(td); 2859 2860 THREAD_LOCK_ASSERT(td, MA_OWNED); 2861 sched_pctcpu_update(ts, TD_IS_RUNNING(td)); 2862 if (ts->ts_ticks) { 2863 int rtick; 2864 2865 /* How many rtick per second ? */ 2866 rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz); 2867 pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT; 2868 } 2869 2870 return (pctcpu); 2871 } 2872 2873 /* 2874 * Enforce affinity settings for a thread. Called after adjustments to 2875 * cpumask. 2876 */ 2877 void 2878 sched_affinity(struct thread *td) 2879 { 2880 #ifdef SMP 2881 struct td_sched *ts; 2882 2883 THREAD_LOCK_ASSERT(td, MA_OWNED); 2884 ts = td_get_sched(td); 2885 if (THREAD_CAN_SCHED(td, ts->ts_cpu)) 2886 return; 2887 if (TD_ON_RUNQ(td)) { 2888 sched_rem(td); 2889 sched_add(td, SRQ_BORING | SRQ_HOLDTD); 2890 return; 2891 } 2892 if (!TD_IS_RUNNING(td)) 2893 return; 2894 /* 2895 * Force a switch before returning to userspace. If the 2896 * target thread is not running locally send an ipi to force 2897 * the issue. 2898 */ 2899 ast_sched_locked(td, TDA_SCHED); 2900 if (td != curthread) 2901 ipi_cpu(ts->ts_cpu, IPI_PREEMPT); 2902 #endif 2903 } 2904 2905 /* 2906 * Bind a thread to a target cpu. 2907 */ 2908 void 2909 sched_bind(struct thread *td, int cpu) 2910 { 2911 struct td_sched *ts; 2912 2913 THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED); 2914 KASSERT(td == curthread, ("sched_bind: can only bind curthread")); 2915 ts = td_get_sched(td); 2916 if (ts->ts_flags & TSF_BOUND) 2917 sched_unbind(td); 2918 KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td)); 2919 ts->ts_flags |= TSF_BOUND; 2920 sched_pin(); 2921 if (PCPU_GET(cpuid) == cpu) 2922 return; 2923 ts->ts_cpu = cpu; 2924 /* When we return from mi_switch we'll be on the correct cpu. */ 2925 mi_switch(SW_VOL | SWT_BIND); 2926 thread_lock(td); 2927 } 2928 2929 /* 2930 * Release a bound thread. 2931 */ 2932 void 2933 sched_unbind(struct thread *td) 2934 { 2935 struct td_sched *ts; 2936 2937 THREAD_LOCK_ASSERT(td, MA_OWNED); 2938 KASSERT(td == curthread, ("sched_unbind: can only bind curthread")); 2939 ts = td_get_sched(td); 2940 if ((ts->ts_flags & TSF_BOUND) == 0) 2941 return; 2942 ts->ts_flags &= ~TSF_BOUND; 2943 sched_unpin(); 2944 } 2945 2946 int 2947 sched_is_bound(struct thread *td) 2948 { 2949 THREAD_LOCK_ASSERT(td, MA_OWNED); 2950 return (td_get_sched(td)->ts_flags & TSF_BOUND); 2951 } 2952 2953 /* 2954 * Basic yield call. 2955 */ 2956 void 2957 sched_relinquish(struct thread *td) 2958 { 2959 thread_lock(td); 2960 mi_switch(SW_VOL | SWT_RELINQUISH); 2961 } 2962 2963 /* 2964 * Return the total system load. 2965 */ 2966 int 2967 sched_load(void) 2968 { 2969 #ifdef SMP 2970 int total; 2971 int i; 2972 2973 total = 0; 2974 CPU_FOREACH(i) 2975 total += atomic_load_int(&TDQ_CPU(i)->tdq_sysload); 2976 return (total); 2977 #else 2978 return (atomic_load_int(&TDQ_SELF()->tdq_sysload)); 2979 #endif 2980 } 2981 2982 int 2983 sched_sizeof_proc(void) 2984 { 2985 return (sizeof(struct proc)); 2986 } 2987 2988 int 2989 sched_sizeof_thread(void) 2990 { 2991 return (sizeof(struct thread) + sizeof(struct td_sched)); 2992 } 2993 2994 #ifdef SMP 2995 #define TDQ_IDLESPIN(tdq) \ 2996 ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0) 2997 #else 2998 #define TDQ_IDLESPIN(tdq) 1 2999 #endif 3000 3001 /* 3002 * The actual idle process. 3003 */ 3004 void 3005 sched_idletd(void *dummy) 3006 { 3007 struct thread *td; 3008 struct tdq *tdq; 3009 int oldswitchcnt, switchcnt; 3010 int i; 3011 3012 mtx_assert(&Giant, MA_NOTOWNED); 3013 td = curthread; 3014 tdq = TDQ_SELF(); 3015 THREAD_NO_SLEEPING(); 3016 oldswitchcnt = -1; 3017 for (;;) { 3018 if (TDQ_LOAD(tdq)) { 3019 thread_lock(td); 3020 mi_switch(SW_VOL | SWT_IDLE); 3021 } 3022 switchcnt = TDQ_SWITCHCNT(tdq); 3023 #ifdef SMP 3024 if (always_steal || switchcnt != oldswitchcnt) { 3025 oldswitchcnt = switchcnt; 3026 if (tdq_idled(tdq) == 0) 3027 continue; 3028 } 3029 switchcnt = TDQ_SWITCHCNT(tdq); 3030 #else 3031 oldswitchcnt = switchcnt; 3032 #endif 3033 /* 3034 * If we're switching very frequently, spin while checking 3035 * for load rather than entering a low power state that 3036 * may require an IPI. However, don't do any busy 3037 * loops while on SMT machines as this simply steals 3038 * cycles from cores doing useful work. 3039 */ 3040 if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) { 3041 for (i = 0; i < sched_idlespins; i++) { 3042 if (TDQ_LOAD(tdq)) 3043 break; 3044 cpu_spinwait(); 3045 } 3046 } 3047 3048 /* If there was context switch during spin, restart it. */ 3049 switchcnt = TDQ_SWITCHCNT(tdq); 3050 if (TDQ_LOAD(tdq) != 0 || switchcnt != oldswitchcnt) 3051 continue; 3052 3053 /* Run main MD idle handler. */ 3054 atomic_store_int(&tdq->tdq_cpu_idle, 1); 3055 /* 3056 * Make sure that the tdq_cpu_idle update is globally visible 3057 * before cpu_idle() reads tdq_load. The order is important 3058 * to avoid races with tdq_notify(). 3059 */ 3060 atomic_thread_fence_seq_cst(); 3061 /* 3062 * Checking for again after the fence picks up assigned 3063 * threads often enough to make it worthwhile to do so in 3064 * order to avoid calling cpu_idle(). 3065 */ 3066 if (TDQ_LOAD(tdq) != 0) { 3067 atomic_store_int(&tdq->tdq_cpu_idle, 0); 3068 continue; 3069 } 3070 cpu_idle(switchcnt * 4 > sched_idlespinthresh); 3071 atomic_store_int(&tdq->tdq_cpu_idle, 0); 3072 3073 /* 3074 * Account thread-less hardware interrupts and 3075 * other wakeup reasons equal to context switches. 3076 */ 3077 switchcnt = TDQ_SWITCHCNT(tdq); 3078 if (switchcnt != oldswitchcnt) 3079 continue; 3080 TDQ_SWITCHCNT_INC(tdq); 3081 oldswitchcnt++; 3082 } 3083 } 3084 3085 /* 3086 * sched_throw_grab() chooses a thread from the queue to switch to 3087 * next. It returns with the tdq lock dropped in a spinlock section to 3088 * keep interrupts disabled until the CPU is running in a proper threaded 3089 * context. 3090 */ 3091 static struct thread * 3092 sched_throw_grab(struct tdq *tdq) 3093 { 3094 struct thread *newtd; 3095 3096 newtd = choosethread(); 3097 spinlock_enter(); 3098 TDQ_UNLOCK(tdq); 3099 KASSERT(curthread->td_md.md_spinlock_count == 1, 3100 ("invalid count %d", curthread->td_md.md_spinlock_count)); 3101 return (newtd); 3102 } 3103 3104 /* 3105 * A CPU is entering for the first time. 3106 */ 3107 void 3108 sched_ap_entry(void) 3109 { 3110 struct thread *newtd; 3111 struct tdq *tdq; 3112 3113 tdq = TDQ_SELF(); 3114 3115 /* This should have been setup in schedinit_ap(). */ 3116 THREAD_LOCKPTR_ASSERT(curthread, TDQ_LOCKPTR(tdq)); 3117 3118 TDQ_LOCK(tdq); 3119 /* Correct spinlock nesting. */ 3120 spinlock_exit(); 3121 PCPU_SET(switchtime, cpu_ticks()); 3122 PCPU_SET(switchticks, ticks); 3123 3124 newtd = sched_throw_grab(tdq); 3125 3126 /* doesn't return */ 3127 cpu_throw(NULL, newtd); 3128 } 3129 3130 /* 3131 * A thread is exiting. 3132 */ 3133 void 3134 sched_throw(struct thread *td) 3135 { 3136 struct thread *newtd; 3137 struct tdq *tdq; 3138 3139 tdq = TDQ_SELF(); 3140 3141 MPASS(td != NULL); 3142 THREAD_LOCK_ASSERT(td, MA_OWNED); 3143 THREAD_LOCKPTR_ASSERT(td, TDQ_LOCKPTR(tdq)); 3144 3145 tdq_load_rem(tdq, td); 3146 td->td_lastcpu = td->td_oncpu; 3147 td->td_oncpu = NOCPU; 3148 thread_lock_block(td); 3149 3150 newtd = sched_throw_grab(tdq); 3151 3152 /* doesn't return */ 3153 cpu_switch(td, newtd, TDQ_LOCKPTR(tdq)); 3154 } 3155 3156 /* 3157 * This is called from fork_exit(). Just acquire the correct locks and 3158 * let fork do the rest of the work. 3159 */ 3160 void 3161 sched_fork_exit(struct thread *td) 3162 { 3163 struct tdq *tdq; 3164 int cpuid; 3165 3166 /* 3167 * Finish setting up thread glue so that it begins execution in a 3168 * non-nested critical section with the scheduler lock held. 3169 */ 3170 KASSERT(curthread->td_md.md_spinlock_count == 1, 3171 ("invalid count %d", curthread->td_md.md_spinlock_count)); 3172 cpuid = PCPU_GET(cpuid); 3173 tdq = TDQ_SELF(); 3174 TDQ_LOCK(tdq); 3175 spinlock_exit(); 3176 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 3177 td->td_oncpu = cpuid; 3178 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running", 3179 "prio:%d", td->td_priority); 3180 SDT_PROBE0(sched, , , on__cpu); 3181 } 3182 3183 /* 3184 * Create on first use to catch odd startup conditions. 3185 */ 3186 char * 3187 sched_tdname(struct thread *td) 3188 { 3189 #ifdef KTR 3190 struct td_sched *ts; 3191 3192 ts = td_get_sched(td); 3193 if (ts->ts_name[0] == '\0') 3194 snprintf(ts->ts_name, sizeof(ts->ts_name), 3195 "%s tid %d", td->td_name, td->td_tid); 3196 return (ts->ts_name); 3197 #else 3198 return (td->td_name); 3199 #endif 3200 } 3201 3202 #ifdef KTR 3203 void 3204 sched_clear_tdname(struct thread *td) 3205 { 3206 struct td_sched *ts; 3207 3208 ts = td_get_sched(td); 3209 ts->ts_name[0] = '\0'; 3210 } 3211 #endif 3212 3213 #ifdef SMP 3214 3215 /* 3216 * Build the CPU topology dump string. Is recursively called to collect 3217 * the topology tree. 3218 */ 3219 static int 3220 sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg, 3221 int indent) 3222 { 3223 char cpusetbuf[CPUSETBUFSIZ]; 3224 int i, first; 3225 3226 sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent, 3227 "", 1 + indent / 2, cg->cg_level); 3228 sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"%s\">", indent, "", 3229 cg->cg_count, cpusetobj_strprint(cpusetbuf, &cg->cg_mask)); 3230 first = TRUE; 3231 for (i = cg->cg_first; i <= cg->cg_last; i++) { 3232 if (CPU_ISSET(i, &cg->cg_mask)) { 3233 if (!first) 3234 sbuf_printf(sb, ", "); 3235 else 3236 first = FALSE; 3237 sbuf_printf(sb, "%d", i); 3238 } 3239 } 3240 sbuf_printf(sb, "</cpu>\n"); 3241 3242 if (cg->cg_flags != 0) { 3243 sbuf_printf(sb, "%*s <flags>", indent, ""); 3244 if ((cg->cg_flags & CG_FLAG_HTT) != 0) 3245 sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>"); 3246 if ((cg->cg_flags & CG_FLAG_THREAD) != 0) 3247 sbuf_printf(sb, "<flag name=\"THREAD\">THREAD group</flag>"); 3248 if ((cg->cg_flags & CG_FLAG_SMT) != 0) 3249 sbuf_printf(sb, "<flag name=\"SMT\">SMT group</flag>"); 3250 if ((cg->cg_flags & CG_FLAG_NODE) != 0) 3251 sbuf_printf(sb, "<flag name=\"NODE\">NUMA node</flag>"); 3252 sbuf_printf(sb, "</flags>\n"); 3253 } 3254 3255 if (cg->cg_children > 0) { 3256 sbuf_printf(sb, "%*s <children>\n", indent, ""); 3257 for (i = 0; i < cg->cg_children; i++) 3258 sysctl_kern_sched_topology_spec_internal(sb, 3259 &cg->cg_child[i], indent+2); 3260 sbuf_printf(sb, "%*s </children>\n", indent, ""); 3261 } 3262 sbuf_printf(sb, "%*s</group>\n", indent, ""); 3263 return (0); 3264 } 3265 3266 /* 3267 * Sysctl handler for retrieving topology dump. It's a wrapper for 3268 * the recursive sysctl_kern_smp_topology_spec_internal(). 3269 */ 3270 static int 3271 sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS) 3272 { 3273 struct sbuf *topo; 3274 int err; 3275 3276 KASSERT(cpu_top != NULL, ("cpu_top isn't initialized")); 3277 3278 topo = sbuf_new_for_sysctl(NULL, NULL, 512, req); 3279 if (topo == NULL) 3280 return (ENOMEM); 3281 3282 sbuf_printf(topo, "<groups>\n"); 3283 err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1); 3284 sbuf_printf(topo, "</groups>\n"); 3285 3286 if (err == 0) { 3287 err = sbuf_finish(topo); 3288 } 3289 sbuf_delete(topo); 3290 return (err); 3291 } 3292 3293 #endif 3294 3295 static int 3296 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS) 3297 { 3298 int error, new_val, period; 3299 3300 period = 1000000 / realstathz; 3301 new_val = period * sched_slice; 3302 error = sysctl_handle_int(oidp, &new_val, 0, req); 3303 if (error != 0 || req->newptr == NULL) 3304 return (error); 3305 if (new_val <= 0) 3306 return (EINVAL); 3307 sched_slice = imax(1, (new_val + period / 2) / period); 3308 sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR; 3309 hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) / 3310 realstathz); 3311 return (0); 3312 } 3313 3314 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 3315 "Scheduler"); 3316 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0, 3317 "Scheduler name"); 3318 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, 3319 CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, 0, 3320 sysctl_kern_quantum, "I", 3321 "Quantum for timeshare threads in microseconds"); 3322 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0, 3323 "Quantum for timeshare threads in stathz ticks"); 3324 SYSCTL_UINT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0, 3325 "Interactivity score threshold"); 3326 SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW, 3327 &preempt_thresh, 0, 3328 "Maximal (lowest) priority for preemption"); 3329 SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost, 0, 3330 "Assign static kernel priorities to sleeping threads"); 3331 SYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins, 0, 3332 "Number of times idle thread will spin waiting for new work"); 3333 SYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW, 3334 &sched_idlespinthresh, 0, 3335 "Threshold before we will permit idle thread spinning"); 3336 #ifdef SMP 3337 SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0, 3338 "Number of hz ticks to keep thread affinity for"); 3339 SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0, 3340 "Enables the long-term load balancer"); 3341 SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW, 3342 &balance_interval, 0, 3343 "Average period in stathz ticks to run the long-term balancer"); 3344 SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0, 3345 "Attempts to steal work from other cores before idling"); 3346 SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0, 3347 "Minimum load on remote CPU before we'll steal"); 3348 SYSCTL_INT(_kern_sched, OID_AUTO, trysteal_limit, CTLFLAG_RW, &trysteal_limit, 3349 0, "Topological distance limit for stealing threads in sched_switch()"); 3350 SYSCTL_INT(_kern_sched, OID_AUTO, always_steal, CTLFLAG_RW, &always_steal, 0, 3351 "Always run the stealer from the idle thread"); 3352 SYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING | 3353 CTLFLAG_MPSAFE | CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A", 3354 "XML dump of detected CPU topology"); 3355 #endif 3356 3357 /* ps compat. All cpu percentages from ULE are weighted. */ 3358 static int ccpu = 0; 3359 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, 3360 "Decay factor used for updating %CPU in 4BSD scheduler"); 3361