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