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 * realtime. 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 struct tdq *tdq; 967 int cpu; 968 969 TDQ_LOCK_ASSERT(from, MA_OWNED); 970 TDQ_LOCK_ASSERT(to, MA_OWNED); 971 972 tdq = from; 973 cpu = TDQ_ID(to); 974 td = tdq_steal(tdq, cpu); 975 if (td == NULL) 976 return (NULL); 977 978 /* 979 * Although the run queue is locked the thread may be 980 * blocked. We can not set the lock until it is unblocked. 981 */ 982 thread_lock_block_wait(td); 983 sched_rem(td); 984 THREAD_LOCKPTR_ASSERT(td, TDQ_LOCKPTR(from)); 985 td->td_lock = TDQ_LOCKPTR(to); 986 td_get_sched(td)->ts_cpu = cpu; 987 tdq_add(to, td, SRQ_YIELDING); 988 989 return (td); 990 } 991 992 /* 993 * This tdq has idled. Try to steal a thread from another cpu and switch 994 * to it. 995 */ 996 static int 997 tdq_idled(struct tdq *tdq) 998 { 999 struct cpu_group *cg, *parent; 1000 struct tdq *steal; 1001 cpuset_t mask; 1002 int cpu, switchcnt, goup; 1003 1004 if (smp_started == 0 || steal_idle == 0 || tdq->tdq_cg == NULL) 1005 return (1); 1006 CPU_FILL(&mask); 1007 CPU_CLR(PCPU_GET(cpuid), &mask); 1008 restart: 1009 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt; 1010 for (cg = tdq->tdq_cg, goup = 0; ; ) { 1011 cpu = sched_highest(cg, &mask, steal_thresh, 1); 1012 /* 1013 * We were assigned a thread but not preempted. Returning 1014 * 0 here will cause our caller to switch to it. 1015 */ 1016 if (tdq->tdq_load) 1017 return (0); 1018 1019 /* 1020 * We found no CPU to steal from in this group. Escalate to 1021 * the parent and repeat. But if parent has only two children 1022 * groups we can avoid searching this group again by searching 1023 * the other one specifically and then escalating two levels. 1024 */ 1025 if (cpu == -1) { 1026 if (goup) { 1027 cg = cg->cg_parent; 1028 goup = 0; 1029 } 1030 parent = cg->cg_parent; 1031 if (parent == NULL) 1032 return (1); 1033 if (parent->cg_children == 2) { 1034 if (cg == &parent->cg_child[0]) 1035 cg = &parent->cg_child[1]; 1036 else 1037 cg = &parent->cg_child[0]; 1038 goup = 1; 1039 } else 1040 cg = parent; 1041 continue; 1042 } 1043 steal = TDQ_CPU(cpu); 1044 /* 1045 * The data returned by sched_highest() is stale and 1046 * the chosen CPU no longer has an eligible thread. 1047 * 1048 * Testing this ahead of tdq_lock_pair() only catches 1049 * this situation about 20% of the time on an 8 core 1050 * 16 thread Ryzen 7, but it still helps performance. 1051 */ 1052 if (steal->tdq_load < steal_thresh || 1053 steal->tdq_transferable == 0) 1054 goto restart; 1055 /* 1056 * Try to lock both queues. If we are assigned a thread while 1057 * waited for the lock, switch to it now instead of stealing. 1058 * If we can't get the lock, then somebody likely got there 1059 * first so continue searching. 1060 */ 1061 TDQ_LOCK(tdq); 1062 if (tdq->tdq_load > 0) { 1063 mi_switch(SW_VOL | SWT_IDLE); 1064 return (0); 1065 } 1066 if (TDQ_TRYLOCK_FLAGS(steal, MTX_DUPOK) == 0) { 1067 TDQ_UNLOCK(tdq); 1068 CPU_CLR(cpu, &mask); 1069 continue; 1070 } 1071 /* 1072 * The data returned by sched_highest() is stale and 1073 * the chosen CPU no longer has an eligible thread, or 1074 * we were preempted and the CPU loading info may be out 1075 * of date. The latter is rare. In either case restart 1076 * the search. 1077 */ 1078 if (steal->tdq_load < steal_thresh || 1079 steal->tdq_transferable == 0 || 1080 switchcnt != tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt) { 1081 tdq_unlock_pair(tdq, steal); 1082 goto restart; 1083 } 1084 /* 1085 * Steal the thread and switch to it. 1086 */ 1087 if (tdq_move(steal, tdq) != NULL) 1088 break; 1089 /* 1090 * We failed to acquire a thread even though it looked 1091 * like one was available. This could be due to affinity 1092 * restrictions or for other reasons. Loop again after 1093 * removing this CPU from the set. The restart logic 1094 * above does not restore this CPU to the set due to the 1095 * likelyhood of failing here again. 1096 */ 1097 CPU_CLR(cpu, &mask); 1098 tdq_unlock_pair(tdq, steal); 1099 } 1100 TDQ_UNLOCK(steal); 1101 mi_switch(SW_VOL | SWT_IDLE); 1102 return (0); 1103 } 1104 1105 /* 1106 * Notify a remote cpu of new work. Sends an IPI if criteria are met. 1107 */ 1108 static void 1109 tdq_notify(struct tdq *tdq, struct thread *td) 1110 { 1111 struct thread *ctd; 1112 int pri; 1113 int cpu; 1114 1115 if (tdq->tdq_owepreempt) 1116 return; 1117 cpu = td_get_sched(td)->ts_cpu; 1118 pri = td->td_priority; 1119 ctd = pcpu_find(cpu)->pc_curthread; 1120 if (!sched_shouldpreempt(pri, ctd->td_priority, 1)) 1121 return; 1122 1123 /* 1124 * Make sure that our caller's earlier update to tdq_load is 1125 * globally visible before we read tdq_cpu_idle. Idle thread 1126 * accesses both of them without locks, and the order is important. 1127 */ 1128 atomic_thread_fence_seq_cst(); 1129 1130 if (TD_IS_IDLETHREAD(ctd)) { 1131 /* 1132 * If the MD code has an idle wakeup routine try that before 1133 * falling back to IPI. 1134 */ 1135 if (!tdq->tdq_cpu_idle || cpu_idle_wakeup(cpu)) 1136 return; 1137 } 1138 1139 /* 1140 * The run queues have been updated, so any switch on the remote CPU 1141 * will satisfy the preemption request. 1142 */ 1143 tdq->tdq_owepreempt = 1; 1144 ipi_cpu(cpu, IPI_PREEMPT); 1145 } 1146 1147 /* 1148 * Steals load from a timeshare queue. Honors the rotating queue head 1149 * index. 1150 */ 1151 static struct thread * 1152 runq_steal_from(struct runq *rq, int cpu, u_char start) 1153 { 1154 struct rqbits *rqb; 1155 struct rqhead *rqh; 1156 struct thread *td, *first; 1157 int bit; 1158 int i; 1159 1160 rqb = &rq->rq_status; 1161 bit = start & (RQB_BPW -1); 1162 first = NULL; 1163 again: 1164 for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) { 1165 if (rqb->rqb_bits[i] == 0) 1166 continue; 1167 if (bit == 0) 1168 bit = RQB_FFS(rqb->rqb_bits[i]); 1169 for (; bit < RQB_BPW; bit++) { 1170 if ((rqb->rqb_bits[i] & (1ul << bit)) == 0) 1171 continue; 1172 rqh = &rq->rq_queues[bit + (i << RQB_L2BPW)]; 1173 TAILQ_FOREACH(td, rqh, td_runq) { 1174 if (first) { 1175 if (THREAD_CAN_MIGRATE(td) && 1176 THREAD_CAN_SCHED(td, cpu)) 1177 return (td); 1178 } else 1179 first = td; 1180 } 1181 } 1182 } 1183 if (start != 0) { 1184 start = 0; 1185 goto again; 1186 } 1187 1188 if (first && THREAD_CAN_MIGRATE(first) && 1189 THREAD_CAN_SCHED(first, cpu)) 1190 return (first); 1191 return (NULL); 1192 } 1193 1194 /* 1195 * Steals load from a standard linear queue. 1196 */ 1197 static struct thread * 1198 runq_steal(struct runq *rq, int cpu) 1199 { 1200 struct rqhead *rqh; 1201 struct rqbits *rqb; 1202 struct thread *td; 1203 int word; 1204 int bit; 1205 1206 rqb = &rq->rq_status; 1207 for (word = 0; word < RQB_LEN; word++) { 1208 if (rqb->rqb_bits[word] == 0) 1209 continue; 1210 for (bit = 0; bit < RQB_BPW; bit++) { 1211 if ((rqb->rqb_bits[word] & (1ul << bit)) == 0) 1212 continue; 1213 rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)]; 1214 TAILQ_FOREACH(td, rqh, td_runq) 1215 if (THREAD_CAN_MIGRATE(td) && 1216 THREAD_CAN_SCHED(td, cpu)) 1217 return (td); 1218 } 1219 } 1220 return (NULL); 1221 } 1222 1223 /* 1224 * Attempt to steal a thread in priority order from a thread queue. 1225 */ 1226 static struct thread * 1227 tdq_steal(struct tdq *tdq, int cpu) 1228 { 1229 struct thread *td; 1230 1231 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 1232 if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL) 1233 return (td); 1234 if ((td = runq_steal_from(&tdq->tdq_timeshare, 1235 cpu, tdq->tdq_ridx)) != NULL) 1236 return (td); 1237 return (runq_steal(&tdq->tdq_idle, cpu)); 1238 } 1239 1240 /* 1241 * Sets the thread lock and ts_cpu to match the requested cpu. Unlocks the 1242 * current lock and returns with the assigned queue locked. 1243 */ 1244 static inline struct tdq * 1245 sched_setcpu(struct thread *td, int cpu, int flags) 1246 { 1247 1248 struct tdq *tdq; 1249 struct mtx *mtx; 1250 1251 THREAD_LOCK_ASSERT(td, MA_OWNED); 1252 tdq = TDQ_CPU(cpu); 1253 td_get_sched(td)->ts_cpu = cpu; 1254 /* 1255 * If the lock matches just return the queue. 1256 */ 1257 if (td->td_lock == TDQ_LOCKPTR(tdq)) { 1258 KASSERT((flags & SRQ_HOLD) == 0, 1259 ("sched_setcpu: Invalid lock for SRQ_HOLD")); 1260 return (tdq); 1261 } 1262 1263 /* 1264 * The hard case, migration, we need to block the thread first to 1265 * prevent order reversals with other cpus locks. 1266 */ 1267 spinlock_enter(); 1268 mtx = thread_lock_block(td); 1269 if ((flags & SRQ_HOLD) == 0) 1270 mtx_unlock_spin(mtx); 1271 TDQ_LOCK(tdq); 1272 thread_lock_unblock(td, TDQ_LOCKPTR(tdq)); 1273 spinlock_exit(); 1274 return (tdq); 1275 } 1276 1277 SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding"); 1278 SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity"); 1279 SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity"); 1280 SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load"); 1281 SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu"); 1282 SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration"); 1283 1284 static int 1285 sched_pickcpu(struct thread *td, int flags) 1286 { 1287 struct cpu_group *cg, *ccg; 1288 struct td_sched *ts; 1289 struct tdq *tdq; 1290 cpuset_t *mask; 1291 int cpu, pri, r, self, intr; 1292 1293 self = PCPU_GET(cpuid); 1294 ts = td_get_sched(td); 1295 KASSERT(!CPU_ABSENT(ts->ts_cpu), ("sched_pickcpu: Start scheduler on " 1296 "absent CPU %d for thread %s.", ts->ts_cpu, td->td_name)); 1297 if (smp_started == 0) 1298 return (self); 1299 /* 1300 * Don't migrate a running thread from sched_switch(). 1301 */ 1302 if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td)) 1303 return (ts->ts_cpu); 1304 /* 1305 * Prefer to run interrupt threads on the processors that generate 1306 * the interrupt. 1307 */ 1308 if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) && 1309 curthread->td_intr_nesting_level) { 1310 tdq = TDQ_SELF(); 1311 if (tdq->tdq_lowpri >= PRI_MIN_IDLE) { 1312 SCHED_STAT_INC(pickcpu_idle_affinity); 1313 return (self); 1314 } 1315 ts->ts_cpu = self; 1316 intr = 1; 1317 cg = tdq->tdq_cg; 1318 goto llc; 1319 } else { 1320 intr = 0; 1321 tdq = TDQ_CPU(ts->ts_cpu); 1322 cg = tdq->tdq_cg; 1323 } 1324 /* 1325 * If the thread can run on the last cpu and the affinity has not 1326 * expired and it is idle, run it there. 1327 */ 1328 if (THREAD_CAN_SCHED(td, ts->ts_cpu) && 1329 tdq->tdq_lowpri >= PRI_MIN_IDLE && 1330 SCHED_AFFINITY(ts, CG_SHARE_L2)) { 1331 if (cg->cg_flags & CG_FLAG_THREAD) { 1332 /* Check all SMT threads for being idle. */ 1333 for (cpu = cg->cg_first; cpu <= cg->cg_last; cpu++) { 1334 if (CPU_ISSET(cpu, &cg->cg_mask) && 1335 TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE) 1336 break; 1337 } 1338 if (cpu > cg->cg_last) { 1339 SCHED_STAT_INC(pickcpu_idle_affinity); 1340 return (ts->ts_cpu); 1341 } 1342 } else { 1343 SCHED_STAT_INC(pickcpu_idle_affinity); 1344 return (ts->ts_cpu); 1345 } 1346 } 1347 llc: 1348 /* 1349 * Search for the last level cache CPU group in the tree. 1350 * Skip SMT, identical groups and caches with expired affinity. 1351 * Interrupt threads affinity is explicit and never expires. 1352 */ 1353 for (ccg = NULL; cg != NULL; cg = cg->cg_parent) { 1354 if (cg->cg_flags & CG_FLAG_THREAD) 1355 continue; 1356 if (cg->cg_children == 1 || cg->cg_count == 1) 1357 continue; 1358 if (cg->cg_level == CG_SHARE_NONE || 1359 (!intr && !SCHED_AFFINITY(ts, cg->cg_level))) 1360 continue; 1361 ccg = cg; 1362 } 1363 /* Found LLC shared by all CPUs, so do a global search. */ 1364 if (ccg == cpu_top) 1365 ccg = NULL; 1366 cpu = -1; 1367 mask = &td->td_cpuset->cs_mask; 1368 pri = td->td_priority; 1369 r = TD_IS_RUNNING(td); 1370 /* 1371 * Try hard to keep interrupts within found LLC. Search the LLC for 1372 * the least loaded CPU we can run now. For NUMA systems it should 1373 * be within target domain, and it also reduces scheduling overhead. 1374 */ 1375 if (ccg != NULL && intr) { 1376 cpu = sched_lowest(ccg, mask, pri, INT_MAX, ts->ts_cpu, r); 1377 if (cpu >= 0) 1378 SCHED_STAT_INC(pickcpu_intrbind); 1379 } else 1380 /* Search the LLC for the least loaded idle CPU we can run now. */ 1381 if (ccg != NULL) { 1382 cpu = sched_lowest(ccg, mask, max(pri, PRI_MAX_TIMESHARE), 1383 INT_MAX, ts->ts_cpu, r); 1384 if (cpu >= 0) 1385 SCHED_STAT_INC(pickcpu_affinity); 1386 } 1387 /* Search globally for the least loaded CPU we can run now. */ 1388 if (cpu < 0) { 1389 cpu = sched_lowest(cpu_top, mask, pri, INT_MAX, ts->ts_cpu, r); 1390 if (cpu >= 0) 1391 SCHED_STAT_INC(pickcpu_lowest); 1392 } 1393 /* Search globally for the least loaded CPU. */ 1394 if (cpu < 0) { 1395 cpu = sched_lowest(cpu_top, mask, -1, INT_MAX, ts->ts_cpu, r); 1396 if (cpu >= 0) 1397 SCHED_STAT_INC(pickcpu_lowest); 1398 } 1399 KASSERT(cpu >= 0, ("sched_pickcpu: Failed to find a cpu.")); 1400 KASSERT(!CPU_ABSENT(cpu), ("sched_pickcpu: Picked absent CPU %d.", cpu)); 1401 /* 1402 * Compare the lowest loaded cpu to current cpu. 1403 */ 1404 tdq = TDQ_CPU(cpu); 1405 if (THREAD_CAN_SCHED(td, self) && TDQ_SELF()->tdq_lowpri > pri && 1406 tdq->tdq_lowpri < PRI_MIN_IDLE && 1407 TDQ_SELF()->tdq_load <= tdq->tdq_load + 1) { 1408 SCHED_STAT_INC(pickcpu_local); 1409 cpu = self; 1410 } 1411 if (cpu != ts->ts_cpu) 1412 SCHED_STAT_INC(pickcpu_migration); 1413 return (cpu); 1414 } 1415 #endif 1416 1417 /* 1418 * Pick the highest priority task we have and return it. 1419 */ 1420 static struct thread * 1421 tdq_choose(struct tdq *tdq) 1422 { 1423 struct thread *td; 1424 1425 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 1426 td = runq_choose(&tdq->tdq_realtime); 1427 if (td != NULL) 1428 return (td); 1429 td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx); 1430 if (td != NULL) { 1431 KASSERT(td->td_priority >= PRI_MIN_BATCH, 1432 ("tdq_choose: Invalid priority on timeshare queue %d", 1433 td->td_priority)); 1434 return (td); 1435 } 1436 td = runq_choose(&tdq->tdq_idle); 1437 if (td != NULL) { 1438 KASSERT(td->td_priority >= PRI_MIN_IDLE, 1439 ("tdq_choose: Invalid priority on idle queue %d", 1440 td->td_priority)); 1441 return (td); 1442 } 1443 1444 return (NULL); 1445 } 1446 1447 /* 1448 * Initialize a thread queue. 1449 */ 1450 static void 1451 tdq_setup(struct tdq *tdq, int id) 1452 { 1453 1454 if (bootverbose) 1455 printf("ULE: setup cpu %d\n", id); 1456 runq_init(&tdq->tdq_realtime); 1457 runq_init(&tdq->tdq_timeshare); 1458 runq_init(&tdq->tdq_idle); 1459 tdq->tdq_id = id; 1460 snprintf(tdq->tdq_name, sizeof(tdq->tdq_name), 1461 "sched lock %d", (int)TDQ_ID(tdq)); 1462 mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock", MTX_SPIN); 1463 #ifdef KTR 1464 snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname), 1465 "CPU %d load", (int)TDQ_ID(tdq)); 1466 #endif 1467 } 1468 1469 #ifdef SMP 1470 static void 1471 sched_setup_smp(void) 1472 { 1473 struct tdq *tdq; 1474 int i; 1475 1476 cpu_top = smp_topo(); 1477 CPU_FOREACH(i) { 1478 tdq = DPCPU_ID_PTR(i, tdq); 1479 tdq_setup(tdq, i); 1480 tdq->tdq_cg = smp_topo_find(cpu_top, i); 1481 if (tdq->tdq_cg == NULL) 1482 panic("Can't find cpu group for %d\n", i); 1483 DPCPU_ID_SET(i, randomval, i * 69069 + 5); 1484 } 1485 PCPU_SET(sched, DPCPU_PTR(tdq)); 1486 balance_tdq = TDQ_SELF(); 1487 } 1488 #endif 1489 1490 /* 1491 * Setup the thread queues and initialize the topology based on MD 1492 * information. 1493 */ 1494 static void 1495 sched_setup(void *dummy) 1496 { 1497 struct tdq *tdq; 1498 1499 #ifdef SMP 1500 sched_setup_smp(); 1501 #else 1502 tdq_setup(TDQ_SELF(), 0); 1503 #endif 1504 tdq = TDQ_SELF(); 1505 1506 /* Add thread0's load since it's running. */ 1507 TDQ_LOCK(tdq); 1508 thread0.td_lock = TDQ_LOCKPTR(tdq); 1509 tdq_load_add(tdq, &thread0); 1510 tdq->tdq_lowpri = thread0.td_priority; 1511 TDQ_UNLOCK(tdq); 1512 } 1513 1514 /* 1515 * This routine determines time constants after stathz and hz are setup. 1516 */ 1517 /* ARGSUSED */ 1518 static void 1519 sched_initticks(void *dummy) 1520 { 1521 int incr; 1522 1523 realstathz = stathz ? stathz : hz; 1524 sched_slice = realstathz / SCHED_SLICE_DEFAULT_DIVISOR; 1525 sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR; 1526 hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) / 1527 realstathz); 1528 1529 /* 1530 * tickincr is shifted out by 10 to avoid rounding errors due to 1531 * hz not being evenly divisible by stathz on all platforms. 1532 */ 1533 incr = (hz << SCHED_TICK_SHIFT) / realstathz; 1534 /* 1535 * This does not work for values of stathz that are more than 1536 * 1 << SCHED_TICK_SHIFT * hz. In practice this does not happen. 1537 */ 1538 if (incr == 0) 1539 incr = 1; 1540 tickincr = incr; 1541 #ifdef SMP 1542 /* 1543 * Set the default balance interval now that we know 1544 * what realstathz is. 1545 */ 1546 balance_interval = realstathz; 1547 balance_ticks = balance_interval; 1548 affinity = SCHED_AFFINITY_DEFAULT; 1549 #endif 1550 if (sched_idlespinthresh < 0) 1551 sched_idlespinthresh = 2 * max(10000, 6 * hz) / realstathz; 1552 } 1553 1554 /* 1555 * This is the core of the interactivity algorithm. Determines a score based 1556 * on past behavior. It is the ratio of sleep time to run time scaled to 1557 * a [0, 100] integer. This is the voluntary sleep time of a process, which 1558 * differs from the cpu usage because it does not account for time spent 1559 * waiting on a run-queue. Would be prettier if we had floating point. 1560 * 1561 * When a thread's sleep time is greater than its run time the 1562 * calculation is: 1563 * 1564 * scaling factor 1565 * interactivity score = --------------------- 1566 * sleep time / run time 1567 * 1568 * 1569 * When a thread's run time is greater than its sleep time the 1570 * calculation is: 1571 * 1572 * scaling factor 1573 * interactivity score = 2 * scaling factor - --------------------- 1574 * run time / sleep time 1575 */ 1576 static int 1577 sched_interact_score(struct thread *td) 1578 { 1579 struct td_sched *ts; 1580 int div; 1581 1582 ts = td_get_sched(td); 1583 /* 1584 * The score is only needed if this is likely to be an interactive 1585 * task. Don't go through the expense of computing it if there's 1586 * no chance. 1587 */ 1588 if (sched_interact <= SCHED_INTERACT_HALF && 1589 ts->ts_runtime >= ts->ts_slptime) 1590 return (SCHED_INTERACT_HALF); 1591 1592 if (ts->ts_runtime > ts->ts_slptime) { 1593 div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF); 1594 return (SCHED_INTERACT_HALF + 1595 (SCHED_INTERACT_HALF - (ts->ts_slptime / div))); 1596 } 1597 if (ts->ts_slptime > ts->ts_runtime) { 1598 div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF); 1599 return (ts->ts_runtime / div); 1600 } 1601 /* runtime == slptime */ 1602 if (ts->ts_runtime) 1603 return (SCHED_INTERACT_HALF); 1604 1605 /* 1606 * This can happen if slptime and runtime are 0. 1607 */ 1608 return (0); 1609 1610 } 1611 1612 /* 1613 * Scale the scheduling priority according to the "interactivity" of this 1614 * process. 1615 */ 1616 static void 1617 sched_priority(struct thread *td) 1618 { 1619 u_int pri, score; 1620 1621 if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE) 1622 return; 1623 /* 1624 * If the score is interactive we place the thread in the realtime 1625 * queue with a priority that is less than kernel and interrupt 1626 * priorities. These threads are not subject to nice restrictions. 1627 * 1628 * Scores greater than this are placed on the normal timeshare queue 1629 * where the priority is partially decided by the most recent cpu 1630 * utilization and the rest is decided by nice value. 1631 * 1632 * The nice value of the process has a linear effect on the calculated 1633 * score. Negative nice values make it easier for a thread to be 1634 * considered interactive. 1635 */ 1636 score = imax(0, sched_interact_score(td) + td->td_proc->p_nice); 1637 if (score < sched_interact) { 1638 pri = PRI_MIN_INTERACT; 1639 pri += (PRI_MAX_INTERACT - PRI_MIN_INTERACT + 1) * score / 1640 sched_interact; 1641 KASSERT(pri >= PRI_MIN_INTERACT && pri <= PRI_MAX_INTERACT, 1642 ("sched_priority: invalid interactive priority %u score %u", 1643 pri, score)); 1644 } else { 1645 pri = SCHED_PRI_MIN; 1646 if (td_get_sched(td)->ts_ticks) 1647 pri += min(SCHED_PRI_TICKS(td_get_sched(td)), 1648 SCHED_PRI_RANGE - 1); 1649 pri += SCHED_PRI_NICE(td->td_proc->p_nice); 1650 KASSERT(pri >= PRI_MIN_BATCH && pri <= PRI_MAX_BATCH, 1651 ("sched_priority: invalid priority %u: nice %d, " 1652 "ticks %d ftick %d ltick %d tick pri %d", 1653 pri, td->td_proc->p_nice, td_get_sched(td)->ts_ticks, 1654 td_get_sched(td)->ts_ftick, td_get_sched(td)->ts_ltick, 1655 SCHED_PRI_TICKS(td_get_sched(td)))); 1656 } 1657 sched_user_prio(td, pri); 1658 1659 return; 1660 } 1661 1662 /* 1663 * This routine enforces a maximum limit on the amount of scheduling history 1664 * kept. It is called after either the slptime or runtime is adjusted. This 1665 * function is ugly due to integer math. 1666 */ 1667 static void 1668 sched_interact_update(struct thread *td) 1669 { 1670 struct td_sched *ts; 1671 u_int sum; 1672 1673 ts = td_get_sched(td); 1674 sum = ts->ts_runtime + ts->ts_slptime; 1675 if (sum < SCHED_SLP_RUN_MAX) 1676 return; 1677 /* 1678 * This only happens from two places: 1679 * 1) We have added an unusual amount of run time from fork_exit. 1680 * 2) We have added an unusual amount of sleep time from sched_sleep(). 1681 */ 1682 if (sum > SCHED_SLP_RUN_MAX * 2) { 1683 if (ts->ts_runtime > ts->ts_slptime) { 1684 ts->ts_runtime = SCHED_SLP_RUN_MAX; 1685 ts->ts_slptime = 1; 1686 } else { 1687 ts->ts_slptime = SCHED_SLP_RUN_MAX; 1688 ts->ts_runtime = 1; 1689 } 1690 return; 1691 } 1692 /* 1693 * If we have exceeded by more than 1/5th then the algorithm below 1694 * will not bring us back into range. Dividing by two here forces 1695 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX] 1696 */ 1697 if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) { 1698 ts->ts_runtime /= 2; 1699 ts->ts_slptime /= 2; 1700 return; 1701 } 1702 ts->ts_runtime = (ts->ts_runtime / 5) * 4; 1703 ts->ts_slptime = (ts->ts_slptime / 5) * 4; 1704 } 1705 1706 /* 1707 * Scale back the interactivity history when a child thread is created. The 1708 * history is inherited from the parent but the thread may behave totally 1709 * differently. For example, a shell spawning a compiler process. We want 1710 * to learn that the compiler is behaving badly very quickly. 1711 */ 1712 static void 1713 sched_interact_fork(struct thread *td) 1714 { 1715 struct td_sched *ts; 1716 int ratio; 1717 int sum; 1718 1719 ts = td_get_sched(td); 1720 sum = ts->ts_runtime + ts->ts_slptime; 1721 if (sum > SCHED_SLP_RUN_FORK) { 1722 ratio = sum / SCHED_SLP_RUN_FORK; 1723 ts->ts_runtime /= ratio; 1724 ts->ts_slptime /= ratio; 1725 } 1726 } 1727 1728 /* 1729 * Called from proc0_init() to setup the scheduler fields. 1730 */ 1731 void 1732 schedinit(void) 1733 { 1734 struct td_sched *ts0; 1735 1736 /* 1737 * Set up the scheduler specific parts of thread0. 1738 */ 1739 ts0 = td_get_sched(&thread0); 1740 ts0->ts_ltick = ticks; 1741 ts0->ts_ftick = ticks; 1742 ts0->ts_slice = 0; 1743 ts0->ts_cpu = curcpu; /* set valid CPU number */ 1744 } 1745 1746 /* 1747 * schedinit_ap() is needed prior to calling sched_throw(NULL) to ensure that 1748 * the pcpu requirements are met for any calls in the period between curthread 1749 * initialization and sched_throw(). One can safely add threads to the queue 1750 * before sched_throw(), for instance, as long as the thread lock is setup 1751 * correctly. 1752 * 1753 * TDQ_SELF() relies on the below sched pcpu setting; it may be used only 1754 * after schedinit_ap(). 1755 */ 1756 void 1757 schedinit_ap(void) 1758 { 1759 1760 #ifdef SMP 1761 PCPU_SET(sched, DPCPU_PTR(tdq)); 1762 #endif 1763 PCPU_GET(idlethread)->td_lock = TDQ_LOCKPTR(TDQ_SELF()); 1764 } 1765 1766 /* 1767 * This is only somewhat accurate since given many processes of the same 1768 * priority they will switch when their slices run out, which will be 1769 * at most sched_slice stathz ticks. 1770 */ 1771 int 1772 sched_rr_interval(void) 1773 { 1774 1775 /* Convert sched_slice from stathz to hz. */ 1776 return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz)); 1777 } 1778 1779 /* 1780 * Update the percent cpu tracking information when it is requested or 1781 * the total history exceeds the maximum. We keep a sliding history of 1782 * tick counts that slowly decays. This is less precise than the 4BSD 1783 * mechanism since it happens with less regular and frequent events. 1784 */ 1785 static void 1786 sched_pctcpu_update(struct td_sched *ts, int run) 1787 { 1788 int t = ticks; 1789 1790 /* 1791 * The signed difference may be negative if the thread hasn't run for 1792 * over half of the ticks rollover period. 1793 */ 1794 if ((u_int)(t - ts->ts_ltick) >= SCHED_TICK_TARG) { 1795 ts->ts_ticks = 0; 1796 ts->ts_ftick = t - SCHED_TICK_TARG; 1797 } else if (t - ts->ts_ftick >= SCHED_TICK_MAX) { 1798 ts->ts_ticks = (ts->ts_ticks / (ts->ts_ltick - ts->ts_ftick)) * 1799 (ts->ts_ltick - (t - SCHED_TICK_TARG)); 1800 ts->ts_ftick = t - SCHED_TICK_TARG; 1801 } 1802 if (run) 1803 ts->ts_ticks += (t - ts->ts_ltick) << SCHED_TICK_SHIFT; 1804 ts->ts_ltick = t; 1805 } 1806 1807 /* 1808 * Adjust the priority of a thread. Move it to the appropriate run-queue 1809 * if necessary. This is the back-end for several priority related 1810 * functions. 1811 */ 1812 static void 1813 sched_thread_priority(struct thread *td, u_char prio) 1814 { 1815 struct td_sched *ts; 1816 struct tdq *tdq; 1817 int oldpri; 1818 1819 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio", 1820 "prio:%d", td->td_priority, "new prio:%d", prio, 1821 KTR_ATTR_LINKED, sched_tdname(curthread)); 1822 SDT_PROBE3(sched, , , change__pri, td, td->td_proc, prio); 1823 if (td != curthread && prio < td->td_priority) { 1824 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread), 1825 "lend prio", "prio:%d", td->td_priority, "new prio:%d", 1826 prio, KTR_ATTR_LINKED, sched_tdname(td)); 1827 SDT_PROBE4(sched, , , lend__pri, td, td->td_proc, prio, 1828 curthread); 1829 } 1830 ts = td_get_sched(td); 1831 THREAD_LOCK_ASSERT(td, MA_OWNED); 1832 if (td->td_priority == prio) 1833 return; 1834 /* 1835 * If the priority has been elevated due to priority 1836 * propagation, we may have to move ourselves to a new 1837 * queue. This could be optimized to not re-add in some 1838 * cases. 1839 */ 1840 if (TD_ON_RUNQ(td) && prio < td->td_priority) { 1841 sched_rem(td); 1842 td->td_priority = prio; 1843 sched_add(td, SRQ_BORROWING | SRQ_HOLDTD); 1844 return; 1845 } 1846 /* 1847 * If the thread is currently running we may have to adjust the lowpri 1848 * information so other cpus are aware of our current priority. 1849 */ 1850 if (TD_IS_RUNNING(td)) { 1851 tdq = TDQ_CPU(ts->ts_cpu); 1852 oldpri = td->td_priority; 1853 td->td_priority = prio; 1854 if (prio < tdq->tdq_lowpri) 1855 tdq->tdq_lowpri = prio; 1856 else if (tdq->tdq_lowpri == oldpri) 1857 tdq_setlowpri(tdq, td); 1858 return; 1859 } 1860 td->td_priority = prio; 1861 } 1862 1863 /* 1864 * Update a thread's priority when it is lent another thread's 1865 * priority. 1866 */ 1867 void 1868 sched_lend_prio(struct thread *td, u_char prio) 1869 { 1870 1871 td->td_flags |= TDF_BORROWING; 1872 sched_thread_priority(td, prio); 1873 } 1874 1875 /* 1876 * Restore a thread's priority when priority propagation is 1877 * over. The prio argument is the minimum priority the thread 1878 * needs to have to satisfy other possible priority lending 1879 * requests. If the thread's regular priority is less 1880 * important than prio, the thread will keep a priority boost 1881 * of prio. 1882 */ 1883 void 1884 sched_unlend_prio(struct thread *td, u_char prio) 1885 { 1886 u_char base_pri; 1887 1888 if (td->td_base_pri >= PRI_MIN_TIMESHARE && 1889 td->td_base_pri <= PRI_MAX_TIMESHARE) 1890 base_pri = td->td_user_pri; 1891 else 1892 base_pri = td->td_base_pri; 1893 if (prio >= base_pri) { 1894 td->td_flags &= ~TDF_BORROWING; 1895 sched_thread_priority(td, base_pri); 1896 } else 1897 sched_lend_prio(td, prio); 1898 } 1899 1900 /* 1901 * Standard entry for setting the priority to an absolute value. 1902 */ 1903 void 1904 sched_prio(struct thread *td, u_char prio) 1905 { 1906 u_char oldprio; 1907 1908 /* First, update the base priority. */ 1909 td->td_base_pri = prio; 1910 1911 /* 1912 * If the thread is borrowing another thread's priority, don't 1913 * ever lower the priority. 1914 */ 1915 if (td->td_flags & TDF_BORROWING && td->td_priority < prio) 1916 return; 1917 1918 /* Change the real priority. */ 1919 oldprio = td->td_priority; 1920 sched_thread_priority(td, prio); 1921 1922 /* 1923 * If the thread is on a turnstile, then let the turnstile update 1924 * its state. 1925 */ 1926 if (TD_ON_LOCK(td) && oldprio != prio) 1927 turnstile_adjust(td, oldprio); 1928 } 1929 1930 /* 1931 * Set the base user priority, does not effect current running priority. 1932 */ 1933 void 1934 sched_user_prio(struct thread *td, u_char prio) 1935 { 1936 1937 td->td_base_user_pri = prio; 1938 if (td->td_lend_user_pri <= prio) 1939 return; 1940 td->td_user_pri = prio; 1941 } 1942 1943 void 1944 sched_lend_user_prio(struct thread *td, u_char prio) 1945 { 1946 1947 THREAD_LOCK_ASSERT(td, MA_OWNED); 1948 td->td_lend_user_pri = prio; 1949 td->td_user_pri = min(prio, td->td_base_user_pri); 1950 if (td->td_priority > td->td_user_pri) 1951 sched_prio(td, td->td_user_pri); 1952 else if (td->td_priority != td->td_user_pri) 1953 td->td_flags |= TDF_NEEDRESCHED; 1954 } 1955 1956 /* 1957 * Like the above but first check if there is anything to do. 1958 */ 1959 void 1960 sched_lend_user_prio_cond(struct thread *td, u_char prio) 1961 { 1962 1963 if (td->td_lend_user_pri != prio) 1964 goto lend; 1965 if (td->td_user_pri != min(prio, td->td_base_user_pri)) 1966 goto lend; 1967 if (td->td_priority != td->td_user_pri) 1968 goto lend; 1969 return; 1970 1971 lend: 1972 thread_lock(td); 1973 sched_lend_user_prio(td, prio); 1974 thread_unlock(td); 1975 } 1976 1977 #ifdef SMP 1978 /* 1979 * This tdq is about to idle. Try to steal a thread from another CPU before 1980 * choosing the idle thread. 1981 */ 1982 static void 1983 tdq_trysteal(struct tdq *tdq) 1984 { 1985 struct cpu_group *cg, *parent; 1986 struct tdq *steal; 1987 cpuset_t mask; 1988 int cpu, i, goup; 1989 1990 if (smp_started == 0 || steal_idle == 0 || trysteal_limit == 0 || 1991 tdq->tdq_cg == NULL) 1992 return; 1993 CPU_FILL(&mask); 1994 CPU_CLR(PCPU_GET(cpuid), &mask); 1995 /* We don't want to be preempted while we're iterating. */ 1996 spinlock_enter(); 1997 TDQ_UNLOCK(tdq); 1998 for (i = 1, cg = tdq->tdq_cg, goup = 0; ; ) { 1999 cpu = sched_highest(cg, &mask, steal_thresh, 1); 2000 /* 2001 * If a thread was added while interrupts were disabled don't 2002 * steal one here. 2003 */ 2004 if (tdq->tdq_load > 0) { 2005 TDQ_LOCK(tdq); 2006 break; 2007 } 2008 2009 /* 2010 * We found no CPU to steal from in this group. Escalate to 2011 * the parent and repeat. But if parent has only two children 2012 * groups we can avoid searching this group again by searching 2013 * the other one specifically and then escalating two levels. 2014 */ 2015 if (cpu == -1) { 2016 if (goup) { 2017 cg = cg->cg_parent; 2018 goup = 0; 2019 } 2020 if (++i > trysteal_limit) { 2021 TDQ_LOCK(tdq); 2022 break; 2023 } 2024 parent = cg->cg_parent; 2025 if (parent == NULL) { 2026 TDQ_LOCK(tdq); 2027 break; 2028 } 2029 if (parent->cg_children == 2) { 2030 if (cg == &parent->cg_child[0]) 2031 cg = &parent->cg_child[1]; 2032 else 2033 cg = &parent->cg_child[0]; 2034 goup = 1; 2035 } else 2036 cg = parent; 2037 continue; 2038 } 2039 steal = TDQ_CPU(cpu); 2040 /* 2041 * The data returned by sched_highest() is stale and 2042 * the chosen CPU no longer has an eligible thread. 2043 * At this point unconditonally exit the loop to bound 2044 * the time spent in the critcal section. 2045 */ 2046 if (steal->tdq_load < steal_thresh || 2047 steal->tdq_transferable == 0) 2048 continue; 2049 /* 2050 * Try to lock both queues. If we are assigned a thread while 2051 * waited for the lock, switch to it now instead of stealing. 2052 * If we can't get the lock, then somebody likely got there 2053 * first. 2054 */ 2055 TDQ_LOCK(tdq); 2056 if (tdq->tdq_load > 0) 2057 break; 2058 if (TDQ_TRYLOCK_FLAGS(steal, MTX_DUPOK) == 0) 2059 break; 2060 /* 2061 * The data returned by sched_highest() is stale and 2062 * the chosen CPU no longer has an eligible thread. 2063 */ 2064 if (steal->tdq_load < steal_thresh || 2065 steal->tdq_transferable == 0) { 2066 TDQ_UNLOCK(steal); 2067 break; 2068 } 2069 /* 2070 * If we fail to acquire one due to affinity restrictions, 2071 * bail out and let the idle thread to a more complete search 2072 * outside of a critical section. 2073 */ 2074 if (tdq_move(steal, tdq) == NULL) { 2075 TDQ_UNLOCK(steal); 2076 break; 2077 } 2078 TDQ_UNLOCK(steal); 2079 break; 2080 } 2081 spinlock_exit(); 2082 } 2083 #endif 2084 2085 /* 2086 * Handle migration from sched_switch(). This happens only for 2087 * cpu binding. 2088 */ 2089 static struct mtx * 2090 sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags) 2091 { 2092 struct tdq *tdn; 2093 2094 KASSERT(THREAD_CAN_MIGRATE(td) || 2095 (td_get_sched(td)->ts_flags & TSF_BOUND) != 0, 2096 ("Thread %p shouldn't migrate", td)); 2097 KASSERT(!CPU_ABSENT(td_get_sched(td)->ts_cpu), ("sched_switch_migrate: " 2098 "thread %s queued on absent CPU %d.", td->td_name, 2099 td_get_sched(td)->ts_cpu)); 2100 tdn = TDQ_CPU(td_get_sched(td)->ts_cpu); 2101 #ifdef SMP 2102 tdq_load_rem(tdq, td); 2103 /* 2104 * Do the lock dance required to avoid LOR. We have an 2105 * extra spinlock nesting from sched_switch() which will 2106 * prevent preemption while we're holding neither run-queue lock. 2107 */ 2108 TDQ_UNLOCK(tdq); 2109 TDQ_LOCK(tdn); 2110 tdq_add(tdn, td, flags); 2111 tdq_notify(tdn, td); 2112 TDQ_UNLOCK(tdn); 2113 TDQ_LOCK(tdq); 2114 #endif 2115 return (TDQ_LOCKPTR(tdn)); 2116 } 2117 2118 /* 2119 * thread_lock_unblock() that does not assume td_lock is blocked. 2120 */ 2121 static inline void 2122 thread_unblock_switch(struct thread *td, struct mtx *mtx) 2123 { 2124 atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock, 2125 (uintptr_t)mtx); 2126 } 2127 2128 /* 2129 * Switch threads. This function has to handle threads coming in while 2130 * blocked for some reason, running, or idle. It also must deal with 2131 * migrating a thread from one queue to another as running threads may 2132 * be assigned elsewhere via binding. 2133 */ 2134 void 2135 sched_switch(struct thread *td, int flags) 2136 { 2137 struct thread *newtd; 2138 struct tdq *tdq; 2139 struct td_sched *ts; 2140 struct mtx *mtx; 2141 int srqflag; 2142 int cpuid, preempted; 2143 #ifdef SMP 2144 int pickcpu; 2145 #endif 2146 2147 THREAD_LOCK_ASSERT(td, MA_OWNED); 2148 2149 cpuid = PCPU_GET(cpuid); 2150 tdq = TDQ_SELF(); 2151 ts = td_get_sched(td); 2152 sched_pctcpu_update(ts, 1); 2153 #ifdef SMP 2154 pickcpu = (td->td_flags & TDF_PICKCPU) != 0; 2155 if (pickcpu) 2156 ts->ts_rltick = ticks - affinity * MAX_CACHE_LEVELS; 2157 else 2158 ts->ts_rltick = ticks; 2159 #endif 2160 td->td_lastcpu = td->td_oncpu; 2161 preempted = (td->td_flags & TDF_SLICEEND) == 0 && 2162 (flags & SW_PREEMPT) != 0; 2163 td->td_flags &= ~(TDF_NEEDRESCHED | TDF_PICKCPU | TDF_SLICEEND); 2164 td->td_owepreempt = 0; 2165 tdq->tdq_owepreempt = 0; 2166 if (!TD_IS_IDLETHREAD(td)) 2167 tdq->tdq_switchcnt++; 2168 2169 /* 2170 * Always block the thread lock so we can drop the tdq lock early. 2171 */ 2172 mtx = thread_lock_block(td); 2173 spinlock_enter(); 2174 if (TD_IS_IDLETHREAD(td)) { 2175 MPASS(mtx == TDQ_LOCKPTR(tdq)); 2176 TD_SET_CAN_RUN(td); 2177 } else if (TD_IS_RUNNING(td)) { 2178 MPASS(mtx == TDQ_LOCKPTR(tdq)); 2179 srqflag = preempted ? 2180 SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED : 2181 SRQ_OURSELF|SRQ_YIELDING; 2182 #ifdef SMP 2183 if (THREAD_CAN_MIGRATE(td) && (!THREAD_CAN_SCHED(td, ts->ts_cpu) 2184 || pickcpu)) 2185 ts->ts_cpu = sched_pickcpu(td, 0); 2186 #endif 2187 if (ts->ts_cpu == cpuid) 2188 tdq_runq_add(tdq, td, srqflag); 2189 else 2190 mtx = sched_switch_migrate(tdq, td, srqflag); 2191 } else { 2192 /* This thread must be going to sleep. */ 2193 if (mtx != TDQ_LOCKPTR(tdq)) { 2194 mtx_unlock_spin(mtx); 2195 TDQ_LOCK(tdq); 2196 } 2197 tdq_load_rem(tdq, td); 2198 #ifdef SMP 2199 if (tdq->tdq_load == 0) 2200 tdq_trysteal(tdq); 2201 #endif 2202 } 2203 2204 #if (KTR_COMPILE & KTR_SCHED) != 0 2205 if (TD_IS_IDLETHREAD(td)) 2206 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "idle", 2207 "prio:%d", td->td_priority); 2208 else 2209 KTR_STATE3(KTR_SCHED, "thread", sched_tdname(td), KTDSTATE(td), 2210 "prio:%d", td->td_priority, "wmesg:\"%s\"", td->td_wmesg, 2211 "lockname:\"%s\"", td->td_lockname); 2212 #endif 2213 2214 /* 2215 * We enter here with the thread blocked and assigned to the 2216 * appropriate cpu run-queue or sleep-queue and with the current 2217 * thread-queue locked. 2218 */ 2219 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED); 2220 newtd = choosethread(); 2221 sched_pctcpu_update(td_get_sched(newtd), 0); 2222 TDQ_UNLOCK(tdq); 2223 2224 /* 2225 * Call the MD code to switch contexts if necessary. 2226 */ 2227 if (td != newtd) { 2228 #ifdef HWPMC_HOOKS 2229 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 2230 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT); 2231 #endif 2232 SDT_PROBE2(sched, , , off__cpu, newtd, newtd->td_proc); 2233 2234 #ifdef KDTRACE_HOOKS 2235 /* 2236 * If DTrace has set the active vtime enum to anything 2237 * other than INACTIVE (0), then it should have set the 2238 * function to call. 2239 */ 2240 if (dtrace_vtime_active) 2241 (*dtrace_vtime_switch_func)(newtd); 2242 #endif 2243 td->td_oncpu = NOCPU; 2244 cpu_switch(td, newtd, mtx); 2245 cpuid = td->td_oncpu = PCPU_GET(cpuid); 2246 2247 SDT_PROBE0(sched, , , on__cpu); 2248 #ifdef HWPMC_HOOKS 2249 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 2250 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN); 2251 #endif 2252 } else { 2253 thread_unblock_switch(td, mtx); 2254 SDT_PROBE0(sched, , , remain__cpu); 2255 } 2256 KASSERT(curthread->td_md.md_spinlock_count == 1, 2257 ("invalid count %d", curthread->td_md.md_spinlock_count)); 2258 2259 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running", 2260 "prio:%d", td->td_priority); 2261 } 2262 2263 /* 2264 * Adjust thread priorities as a result of a nice request. 2265 */ 2266 void 2267 sched_nice(struct proc *p, int nice) 2268 { 2269 struct thread *td; 2270 2271 PROC_LOCK_ASSERT(p, MA_OWNED); 2272 2273 p->p_nice = nice; 2274 FOREACH_THREAD_IN_PROC(p, td) { 2275 thread_lock(td); 2276 sched_priority(td); 2277 sched_prio(td, td->td_base_user_pri); 2278 thread_unlock(td); 2279 } 2280 } 2281 2282 /* 2283 * Record the sleep time for the interactivity scorer. 2284 */ 2285 void 2286 sched_sleep(struct thread *td, int prio) 2287 { 2288 2289 THREAD_LOCK_ASSERT(td, MA_OWNED); 2290 2291 td->td_slptick = ticks; 2292 if (TD_IS_SUSPENDED(td) || prio >= PSOCK) 2293 td->td_flags |= TDF_CANSWAP; 2294 if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE) 2295 return; 2296 if (static_boost == 1 && prio) 2297 sched_prio(td, prio); 2298 else if (static_boost && td->td_priority > static_boost) 2299 sched_prio(td, static_boost); 2300 } 2301 2302 /* 2303 * Schedule a thread to resume execution and record how long it voluntarily 2304 * slept. We also update the pctcpu, interactivity, and priority. 2305 * 2306 * Requires the thread lock on entry, drops on exit. 2307 */ 2308 void 2309 sched_wakeup(struct thread *td, int srqflags) 2310 { 2311 struct td_sched *ts; 2312 int slptick; 2313 2314 THREAD_LOCK_ASSERT(td, MA_OWNED); 2315 ts = td_get_sched(td); 2316 td->td_flags &= ~TDF_CANSWAP; 2317 2318 /* 2319 * If we slept for more than a tick update our interactivity and 2320 * priority. 2321 */ 2322 slptick = td->td_slptick; 2323 td->td_slptick = 0; 2324 if (slptick && slptick != ticks) { 2325 ts->ts_slptime += (ticks - slptick) << SCHED_TICK_SHIFT; 2326 sched_interact_update(td); 2327 sched_pctcpu_update(ts, 0); 2328 } 2329 /* 2330 * Reset the slice value since we slept and advanced the round-robin. 2331 */ 2332 ts->ts_slice = 0; 2333 sched_add(td, SRQ_BORING | srqflags); 2334 } 2335 2336 /* 2337 * Penalize the parent for creating a new child and initialize the child's 2338 * priority. 2339 */ 2340 void 2341 sched_fork(struct thread *td, struct thread *child) 2342 { 2343 THREAD_LOCK_ASSERT(td, MA_OWNED); 2344 sched_pctcpu_update(td_get_sched(td), 1); 2345 sched_fork_thread(td, child); 2346 /* 2347 * Penalize the parent and child for forking. 2348 */ 2349 sched_interact_fork(child); 2350 sched_priority(child); 2351 td_get_sched(td)->ts_runtime += tickincr; 2352 sched_interact_update(td); 2353 sched_priority(td); 2354 } 2355 2356 /* 2357 * Fork a new thread, may be within the same process. 2358 */ 2359 void 2360 sched_fork_thread(struct thread *td, struct thread *child) 2361 { 2362 struct td_sched *ts; 2363 struct td_sched *ts2; 2364 struct tdq *tdq; 2365 2366 tdq = TDQ_SELF(); 2367 THREAD_LOCK_ASSERT(td, MA_OWNED); 2368 /* 2369 * Initialize child. 2370 */ 2371 ts = td_get_sched(td); 2372 ts2 = td_get_sched(child); 2373 child->td_oncpu = NOCPU; 2374 child->td_lastcpu = NOCPU; 2375 child->td_lock = TDQ_LOCKPTR(tdq); 2376 child->td_cpuset = cpuset_ref(td->td_cpuset); 2377 child->td_domain.dr_policy = td->td_cpuset->cs_domain; 2378 ts2->ts_cpu = ts->ts_cpu; 2379 ts2->ts_flags = 0; 2380 /* 2381 * Grab our parents cpu estimation information. 2382 */ 2383 ts2->ts_ticks = ts->ts_ticks; 2384 ts2->ts_ltick = ts->ts_ltick; 2385 ts2->ts_ftick = ts->ts_ftick; 2386 /* 2387 * Do not inherit any borrowed priority from the parent. 2388 */ 2389 child->td_priority = child->td_base_pri; 2390 /* 2391 * And update interactivity score. 2392 */ 2393 ts2->ts_slptime = ts->ts_slptime; 2394 ts2->ts_runtime = ts->ts_runtime; 2395 /* Attempt to quickly learn interactivity. */ 2396 ts2->ts_slice = tdq_slice(tdq) - sched_slice_min; 2397 #ifdef KTR 2398 bzero(ts2->ts_name, sizeof(ts2->ts_name)); 2399 #endif 2400 } 2401 2402 /* 2403 * Adjust the priority class of a thread. 2404 */ 2405 void 2406 sched_class(struct thread *td, int class) 2407 { 2408 2409 THREAD_LOCK_ASSERT(td, MA_OWNED); 2410 if (td->td_pri_class == class) 2411 return; 2412 td->td_pri_class = class; 2413 } 2414 2415 /* 2416 * Return some of the child's priority and interactivity to the parent. 2417 */ 2418 void 2419 sched_exit(struct proc *p, struct thread *child) 2420 { 2421 struct thread *td; 2422 2423 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit", 2424 "prio:%d", child->td_priority); 2425 PROC_LOCK_ASSERT(p, MA_OWNED); 2426 td = FIRST_THREAD_IN_PROC(p); 2427 sched_exit_thread(td, child); 2428 } 2429 2430 /* 2431 * Penalize another thread for the time spent on this one. This helps to 2432 * worsen the priority and interactivity of processes which schedule batch 2433 * jobs such as make. This has little effect on the make process itself but 2434 * causes new processes spawned by it to receive worse scores immediately. 2435 */ 2436 void 2437 sched_exit_thread(struct thread *td, struct thread *child) 2438 { 2439 2440 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit", 2441 "prio:%d", child->td_priority); 2442 /* 2443 * Give the child's runtime to the parent without returning the 2444 * sleep time as a penalty to the parent. This causes shells that 2445 * launch expensive things to mark their children as expensive. 2446 */ 2447 thread_lock(td); 2448 td_get_sched(td)->ts_runtime += td_get_sched(child)->ts_runtime; 2449 sched_interact_update(td); 2450 sched_priority(td); 2451 thread_unlock(td); 2452 } 2453 2454 void 2455 sched_preempt(struct thread *td) 2456 { 2457 struct tdq *tdq; 2458 int flags; 2459 2460 SDT_PROBE2(sched, , , surrender, td, td->td_proc); 2461 2462 thread_lock(td); 2463 tdq = TDQ_SELF(); 2464 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2465 if (td->td_priority > tdq->tdq_lowpri) { 2466 if (td->td_critnest == 1) { 2467 flags = SW_INVOL | SW_PREEMPT; 2468 flags |= TD_IS_IDLETHREAD(td) ? SWT_REMOTEWAKEIDLE : 2469 SWT_REMOTEPREEMPT; 2470 mi_switch(flags); 2471 /* Switch dropped thread lock. */ 2472 return; 2473 } 2474 td->td_owepreempt = 1; 2475 } else { 2476 tdq->tdq_owepreempt = 0; 2477 } 2478 thread_unlock(td); 2479 } 2480 2481 /* 2482 * Fix priorities on return to user-space. Priorities may be elevated due 2483 * to static priorities in msleep() or similar. 2484 */ 2485 void 2486 sched_userret_slowpath(struct thread *td) 2487 { 2488 2489 thread_lock(td); 2490 td->td_priority = td->td_user_pri; 2491 td->td_base_pri = td->td_user_pri; 2492 tdq_setlowpri(TDQ_SELF(), td); 2493 thread_unlock(td); 2494 } 2495 2496 /* 2497 * Handle a stathz tick. This is really only relevant for timeshare 2498 * threads. 2499 */ 2500 void 2501 sched_clock(struct thread *td, int cnt) 2502 { 2503 struct tdq *tdq; 2504 struct td_sched *ts; 2505 2506 THREAD_LOCK_ASSERT(td, MA_OWNED); 2507 tdq = TDQ_SELF(); 2508 #ifdef SMP 2509 /* 2510 * We run the long term load balancer infrequently on the first cpu. 2511 */ 2512 if (balance_tdq == tdq && smp_started != 0 && rebalance != 0 && 2513 balance_ticks != 0) { 2514 balance_ticks -= cnt; 2515 if (balance_ticks <= 0) 2516 sched_balance(); 2517 } 2518 #endif 2519 /* 2520 * Save the old switch count so we have a record of the last ticks 2521 * activity. Initialize the new switch count based on our load. 2522 * If there is some activity seed it to reflect that. 2523 */ 2524 tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt; 2525 tdq->tdq_switchcnt = tdq->tdq_load; 2526 /* 2527 * Advance the insert index once for each tick to ensure that all 2528 * threads get a chance to run. 2529 */ 2530 if (tdq->tdq_idx == tdq->tdq_ridx) { 2531 tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS; 2532 if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx])) 2533 tdq->tdq_ridx = tdq->tdq_idx; 2534 } 2535 ts = td_get_sched(td); 2536 sched_pctcpu_update(ts, 1); 2537 if ((td->td_pri_class & PRI_FIFO_BIT) || TD_IS_IDLETHREAD(td)) 2538 return; 2539 2540 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) { 2541 /* 2542 * We used a tick; charge it to the thread so 2543 * that we can compute our interactivity. 2544 */ 2545 td_get_sched(td)->ts_runtime += tickincr * cnt; 2546 sched_interact_update(td); 2547 sched_priority(td); 2548 } 2549 2550 /* 2551 * Force a context switch if the current thread has used up a full 2552 * time slice (default is 100ms). 2553 */ 2554 ts->ts_slice += cnt; 2555 if (ts->ts_slice >= tdq_slice(tdq)) { 2556 ts->ts_slice = 0; 2557 td->td_flags |= TDF_NEEDRESCHED | TDF_SLICEEND; 2558 } 2559 } 2560 2561 u_int 2562 sched_estcpu(struct thread *td __unused) 2563 { 2564 2565 return (0); 2566 } 2567 2568 /* 2569 * Return whether the current CPU has runnable tasks. Used for in-kernel 2570 * cooperative idle threads. 2571 */ 2572 int 2573 sched_runnable(void) 2574 { 2575 struct tdq *tdq; 2576 int load; 2577 2578 load = 1; 2579 2580 tdq = TDQ_SELF(); 2581 if ((curthread->td_flags & TDF_IDLETD) != 0) { 2582 if (tdq->tdq_load > 0) 2583 goto out; 2584 } else 2585 if (tdq->tdq_load - 1 > 0) 2586 goto out; 2587 load = 0; 2588 out: 2589 return (load); 2590 } 2591 2592 /* 2593 * Choose the highest priority thread to run. The thread is removed from 2594 * the run-queue while running however the load remains. For SMP we set 2595 * the tdq in the global idle bitmask if it idles here. 2596 */ 2597 struct thread * 2598 sched_choose(void) 2599 { 2600 struct thread *td; 2601 struct tdq *tdq; 2602 2603 tdq = TDQ_SELF(); 2604 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2605 td = tdq_choose(tdq); 2606 if (td) { 2607 tdq_runq_rem(tdq, td); 2608 tdq->tdq_lowpri = td->td_priority; 2609 return (td); 2610 } 2611 tdq->tdq_lowpri = PRI_MAX_IDLE; 2612 return (PCPU_GET(idlethread)); 2613 } 2614 2615 /* 2616 * Set owepreempt if necessary. Preemption never happens directly in ULE, 2617 * we always request it once we exit a critical section. 2618 */ 2619 static inline void 2620 sched_setpreempt(struct thread *td) 2621 { 2622 struct thread *ctd; 2623 int cpri; 2624 int pri; 2625 2626 THREAD_LOCK_ASSERT(curthread, MA_OWNED); 2627 2628 ctd = curthread; 2629 pri = td->td_priority; 2630 cpri = ctd->td_priority; 2631 if (pri < cpri) 2632 ctd->td_flags |= TDF_NEEDRESCHED; 2633 if (KERNEL_PANICKED() || pri >= cpri || cold || TD_IS_INHIBITED(ctd)) 2634 return; 2635 if (!sched_shouldpreempt(pri, cpri, 0)) 2636 return; 2637 ctd->td_owepreempt = 1; 2638 } 2639 2640 /* 2641 * Add a thread to a thread queue. Select the appropriate runq and add the 2642 * thread to it. This is the internal function called when the tdq is 2643 * predetermined. 2644 */ 2645 void 2646 tdq_add(struct tdq *tdq, struct thread *td, int flags) 2647 { 2648 2649 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2650 THREAD_LOCK_BLOCKED_ASSERT(td, MA_OWNED); 2651 KASSERT((td->td_inhibitors == 0), 2652 ("sched_add: trying to run inhibited thread")); 2653 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)), 2654 ("sched_add: bad thread state")); 2655 KASSERT(td->td_flags & TDF_INMEM, 2656 ("sched_add: thread swapped out")); 2657 2658 if (td->td_priority < tdq->tdq_lowpri) 2659 tdq->tdq_lowpri = td->td_priority; 2660 tdq_runq_add(tdq, td, flags); 2661 tdq_load_add(tdq, td); 2662 } 2663 2664 /* 2665 * Select the target thread queue and add a thread to it. Request 2666 * preemption or IPI a remote processor if required. 2667 * 2668 * Requires the thread lock on entry, drops on exit. 2669 */ 2670 void 2671 sched_add(struct thread *td, int flags) 2672 { 2673 struct tdq *tdq; 2674 #ifdef SMP 2675 int cpu; 2676 #endif 2677 2678 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add", 2679 "prio:%d", td->td_priority, KTR_ATTR_LINKED, 2680 sched_tdname(curthread)); 2681 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup", 2682 KTR_ATTR_LINKED, sched_tdname(td)); 2683 SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL, 2684 flags & SRQ_PREEMPTED); 2685 THREAD_LOCK_ASSERT(td, MA_OWNED); 2686 /* 2687 * Recalculate the priority before we select the target cpu or 2688 * run-queue. 2689 */ 2690 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) 2691 sched_priority(td); 2692 #ifdef SMP 2693 /* 2694 * Pick the destination cpu and if it isn't ours transfer to the 2695 * target cpu. 2696 */ 2697 cpu = sched_pickcpu(td, flags); 2698 tdq = sched_setcpu(td, cpu, flags); 2699 tdq_add(tdq, td, flags); 2700 if (cpu != PCPU_GET(cpuid)) 2701 tdq_notify(tdq, td); 2702 else if (!(flags & SRQ_YIELDING)) 2703 sched_setpreempt(td); 2704 #else 2705 tdq = TDQ_SELF(); 2706 /* 2707 * Now that the thread is moving to the run-queue, set the lock 2708 * to the scheduler's lock. 2709 */ 2710 if (td->td_lock != TDQ_LOCKPTR(tdq)) { 2711 TDQ_LOCK(tdq); 2712 if ((flags & SRQ_HOLD) != 0) 2713 td->td_lock = TDQ_LOCKPTR(tdq); 2714 else 2715 thread_lock_set(td, TDQ_LOCKPTR(tdq)); 2716 } 2717 tdq_add(tdq, td, flags); 2718 if (!(flags & SRQ_YIELDING)) 2719 sched_setpreempt(td); 2720 #endif 2721 if (!(flags & SRQ_HOLDTD)) 2722 thread_unlock(td); 2723 } 2724 2725 /* 2726 * Remove a thread from a run-queue without running it. This is used 2727 * when we're stealing a thread from a remote queue. Otherwise all threads 2728 * exit by calling sched_exit_thread() and sched_throw() themselves. 2729 */ 2730 void 2731 sched_rem(struct thread *td) 2732 { 2733 struct tdq *tdq; 2734 2735 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem", 2736 "prio:%d", td->td_priority); 2737 SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL); 2738 tdq = TDQ_CPU(td_get_sched(td)->ts_cpu); 2739 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2740 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2741 KASSERT(TD_ON_RUNQ(td), 2742 ("sched_rem: thread not on run queue")); 2743 tdq_runq_rem(tdq, td); 2744 tdq_load_rem(tdq, td); 2745 TD_SET_CAN_RUN(td); 2746 if (td->td_priority == tdq->tdq_lowpri) 2747 tdq_setlowpri(tdq, NULL); 2748 } 2749 2750 /* 2751 * Fetch cpu utilization information. Updates on demand. 2752 */ 2753 fixpt_t 2754 sched_pctcpu(struct thread *td) 2755 { 2756 fixpt_t pctcpu; 2757 struct td_sched *ts; 2758 2759 pctcpu = 0; 2760 ts = td_get_sched(td); 2761 2762 THREAD_LOCK_ASSERT(td, MA_OWNED); 2763 sched_pctcpu_update(ts, TD_IS_RUNNING(td)); 2764 if (ts->ts_ticks) { 2765 int rtick; 2766 2767 /* How many rtick per second ? */ 2768 rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz); 2769 pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT; 2770 } 2771 2772 return (pctcpu); 2773 } 2774 2775 /* 2776 * Enforce affinity settings for a thread. Called after adjustments to 2777 * cpumask. 2778 */ 2779 void 2780 sched_affinity(struct thread *td) 2781 { 2782 #ifdef SMP 2783 struct td_sched *ts; 2784 2785 THREAD_LOCK_ASSERT(td, MA_OWNED); 2786 ts = td_get_sched(td); 2787 if (THREAD_CAN_SCHED(td, ts->ts_cpu)) 2788 return; 2789 if (TD_ON_RUNQ(td)) { 2790 sched_rem(td); 2791 sched_add(td, SRQ_BORING | SRQ_HOLDTD); 2792 return; 2793 } 2794 if (!TD_IS_RUNNING(td)) 2795 return; 2796 /* 2797 * Force a switch before returning to userspace. If the 2798 * target thread is not running locally send an ipi to force 2799 * the issue. 2800 */ 2801 td->td_flags |= TDF_NEEDRESCHED; 2802 if (td != curthread) 2803 ipi_cpu(ts->ts_cpu, IPI_PREEMPT); 2804 #endif 2805 } 2806 2807 /* 2808 * Bind a thread to a target cpu. 2809 */ 2810 void 2811 sched_bind(struct thread *td, int cpu) 2812 { 2813 struct td_sched *ts; 2814 2815 THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED); 2816 KASSERT(td == curthread, ("sched_bind: can only bind curthread")); 2817 ts = td_get_sched(td); 2818 if (ts->ts_flags & TSF_BOUND) 2819 sched_unbind(td); 2820 KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td)); 2821 ts->ts_flags |= TSF_BOUND; 2822 sched_pin(); 2823 if (PCPU_GET(cpuid) == cpu) 2824 return; 2825 ts->ts_cpu = cpu; 2826 /* When we return from mi_switch we'll be on the correct cpu. */ 2827 mi_switch(SW_VOL); 2828 thread_lock(td); 2829 } 2830 2831 /* 2832 * Release a bound thread. 2833 */ 2834 void 2835 sched_unbind(struct thread *td) 2836 { 2837 struct td_sched *ts; 2838 2839 THREAD_LOCK_ASSERT(td, MA_OWNED); 2840 KASSERT(td == curthread, ("sched_unbind: can only bind curthread")); 2841 ts = td_get_sched(td); 2842 if ((ts->ts_flags & TSF_BOUND) == 0) 2843 return; 2844 ts->ts_flags &= ~TSF_BOUND; 2845 sched_unpin(); 2846 } 2847 2848 int 2849 sched_is_bound(struct thread *td) 2850 { 2851 THREAD_LOCK_ASSERT(td, MA_OWNED); 2852 return (td_get_sched(td)->ts_flags & TSF_BOUND); 2853 } 2854 2855 /* 2856 * Basic yield call. 2857 */ 2858 void 2859 sched_relinquish(struct thread *td) 2860 { 2861 thread_lock(td); 2862 mi_switch(SW_VOL | SWT_RELINQUISH); 2863 } 2864 2865 /* 2866 * Return the total system load. 2867 */ 2868 int 2869 sched_load(void) 2870 { 2871 #ifdef SMP 2872 int total; 2873 int i; 2874 2875 total = 0; 2876 CPU_FOREACH(i) 2877 total += TDQ_CPU(i)->tdq_sysload; 2878 return (total); 2879 #else 2880 return (TDQ_SELF()->tdq_sysload); 2881 #endif 2882 } 2883 2884 int 2885 sched_sizeof_proc(void) 2886 { 2887 return (sizeof(struct proc)); 2888 } 2889 2890 int 2891 sched_sizeof_thread(void) 2892 { 2893 return (sizeof(struct thread) + sizeof(struct td_sched)); 2894 } 2895 2896 #ifdef SMP 2897 #define TDQ_IDLESPIN(tdq) \ 2898 ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0) 2899 #else 2900 #define TDQ_IDLESPIN(tdq) 1 2901 #endif 2902 2903 /* 2904 * The actual idle process. 2905 */ 2906 void 2907 sched_idletd(void *dummy) 2908 { 2909 struct thread *td; 2910 struct tdq *tdq; 2911 int oldswitchcnt, switchcnt; 2912 int i; 2913 2914 mtx_assert(&Giant, MA_NOTOWNED); 2915 td = curthread; 2916 tdq = TDQ_SELF(); 2917 THREAD_NO_SLEEPING(); 2918 oldswitchcnt = -1; 2919 for (;;) { 2920 if (tdq->tdq_load) { 2921 thread_lock(td); 2922 mi_switch(SW_VOL | SWT_IDLE); 2923 } 2924 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt; 2925 #ifdef SMP 2926 if (always_steal || switchcnt != oldswitchcnt) { 2927 oldswitchcnt = switchcnt; 2928 if (tdq_idled(tdq) == 0) 2929 continue; 2930 } 2931 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt; 2932 #else 2933 oldswitchcnt = switchcnt; 2934 #endif 2935 /* 2936 * If we're switching very frequently, spin while checking 2937 * for load rather than entering a low power state that 2938 * may require an IPI. However, don't do any busy 2939 * loops while on SMT machines as this simply steals 2940 * cycles from cores doing useful work. 2941 */ 2942 if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) { 2943 for (i = 0; i < sched_idlespins; i++) { 2944 if (tdq->tdq_load) 2945 break; 2946 cpu_spinwait(); 2947 } 2948 } 2949 2950 /* If there was context switch during spin, restart it. */ 2951 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt; 2952 if (tdq->tdq_load != 0 || switchcnt != oldswitchcnt) 2953 continue; 2954 2955 /* Run main MD idle handler. */ 2956 tdq->tdq_cpu_idle = 1; 2957 /* 2958 * Make sure that tdq_cpu_idle update is globally visible 2959 * before cpu_idle() read tdq_load. The order is important 2960 * to avoid race with tdq_notify. 2961 */ 2962 atomic_thread_fence_seq_cst(); 2963 /* 2964 * Checking for again after the fence picks up assigned 2965 * threads often enough to make it worthwhile to do so in 2966 * order to avoid calling cpu_idle(). 2967 */ 2968 if (tdq->tdq_load != 0) { 2969 tdq->tdq_cpu_idle = 0; 2970 continue; 2971 } 2972 cpu_idle(switchcnt * 4 > sched_idlespinthresh); 2973 tdq->tdq_cpu_idle = 0; 2974 2975 /* 2976 * Account thread-less hardware interrupts and 2977 * other wakeup reasons equal to context switches. 2978 */ 2979 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt; 2980 if (switchcnt != oldswitchcnt) 2981 continue; 2982 tdq->tdq_switchcnt++; 2983 oldswitchcnt++; 2984 } 2985 } 2986 2987 /* 2988 * sched_throw_grab() chooses a thread from the queue to switch to 2989 * next. It returns with the tdq lock dropped in a spinlock section to 2990 * keep interrupts disabled until the CPU is running in a proper threaded 2991 * context. 2992 */ 2993 static struct thread * 2994 sched_throw_grab(struct tdq *tdq) 2995 { 2996 struct thread *newtd; 2997 2998 newtd = choosethread(); 2999 spinlock_enter(); 3000 TDQ_UNLOCK(tdq); 3001 KASSERT(curthread->td_md.md_spinlock_count == 1, 3002 ("invalid count %d", curthread->td_md.md_spinlock_count)); 3003 return (newtd); 3004 } 3005 3006 /* 3007 * A CPU is entering for the first time. 3008 */ 3009 void 3010 sched_ap_entry(void) 3011 { 3012 struct thread *newtd; 3013 struct tdq *tdq; 3014 3015 tdq = TDQ_SELF(); 3016 3017 /* This should have been setup in schedinit_ap(). */ 3018 THREAD_LOCKPTR_ASSERT(curthread, TDQ_LOCKPTR(tdq)); 3019 3020 TDQ_LOCK(tdq); 3021 /* Correct spinlock nesting. */ 3022 spinlock_exit(); 3023 PCPU_SET(switchtime, cpu_ticks()); 3024 PCPU_SET(switchticks, ticks); 3025 3026 newtd = sched_throw_grab(tdq); 3027 3028 /* doesn't return */ 3029 cpu_throw(NULL, newtd); 3030 } 3031 3032 /* 3033 * A thread is exiting. 3034 */ 3035 void 3036 sched_throw(struct thread *td) 3037 { 3038 struct thread *newtd; 3039 struct tdq *tdq; 3040 3041 tdq = TDQ_SELF(); 3042 3043 MPASS(td != NULL); 3044 THREAD_LOCK_ASSERT(td, MA_OWNED); 3045 THREAD_LOCKPTR_ASSERT(td, TDQ_LOCKPTR(tdq)); 3046 3047 tdq_load_rem(tdq, td); 3048 td->td_lastcpu = td->td_oncpu; 3049 td->td_oncpu = NOCPU; 3050 thread_lock_block(td); 3051 3052 newtd = sched_throw_grab(tdq); 3053 3054 /* doesn't return */ 3055 cpu_switch(td, newtd, TDQ_LOCKPTR(tdq)); 3056 } 3057 3058 /* 3059 * This is called from fork_exit(). Just acquire the correct locks and 3060 * let fork do the rest of the work. 3061 */ 3062 void 3063 sched_fork_exit(struct thread *td) 3064 { 3065 struct tdq *tdq; 3066 int cpuid; 3067 3068 /* 3069 * Finish setting up thread glue so that it begins execution in a 3070 * non-nested critical section with the scheduler lock held. 3071 */ 3072 KASSERT(curthread->td_md.md_spinlock_count == 1, 3073 ("invalid count %d", curthread->td_md.md_spinlock_count)); 3074 cpuid = PCPU_GET(cpuid); 3075 tdq = TDQ_SELF(); 3076 TDQ_LOCK(tdq); 3077 spinlock_exit(); 3078 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 3079 td->td_oncpu = cpuid; 3080 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running", 3081 "prio:%d", td->td_priority); 3082 SDT_PROBE0(sched, , , on__cpu); 3083 } 3084 3085 /* 3086 * Create on first use to catch odd startup conditons. 3087 */ 3088 char * 3089 sched_tdname(struct thread *td) 3090 { 3091 #ifdef KTR 3092 struct td_sched *ts; 3093 3094 ts = td_get_sched(td); 3095 if (ts->ts_name[0] == '\0') 3096 snprintf(ts->ts_name, sizeof(ts->ts_name), 3097 "%s tid %d", td->td_name, td->td_tid); 3098 return (ts->ts_name); 3099 #else 3100 return (td->td_name); 3101 #endif 3102 } 3103 3104 #ifdef KTR 3105 void 3106 sched_clear_tdname(struct thread *td) 3107 { 3108 struct td_sched *ts; 3109 3110 ts = td_get_sched(td); 3111 ts->ts_name[0] = '\0'; 3112 } 3113 #endif 3114 3115 #ifdef SMP 3116 3117 /* 3118 * Build the CPU topology dump string. Is recursively called to collect 3119 * the topology tree. 3120 */ 3121 static int 3122 sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg, 3123 int indent) 3124 { 3125 char cpusetbuf[CPUSETBUFSIZ]; 3126 int i, first; 3127 3128 sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent, 3129 "", 1 + indent / 2, cg->cg_level); 3130 sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"%s\">", indent, "", 3131 cg->cg_count, cpusetobj_strprint(cpusetbuf, &cg->cg_mask)); 3132 first = TRUE; 3133 for (i = cg->cg_first; i <= cg->cg_last; i++) { 3134 if (CPU_ISSET(i, &cg->cg_mask)) { 3135 if (!first) 3136 sbuf_printf(sb, ", "); 3137 else 3138 first = FALSE; 3139 sbuf_printf(sb, "%d", i); 3140 } 3141 } 3142 sbuf_printf(sb, "</cpu>\n"); 3143 3144 if (cg->cg_flags != 0) { 3145 sbuf_printf(sb, "%*s <flags>", indent, ""); 3146 if ((cg->cg_flags & CG_FLAG_HTT) != 0) 3147 sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>"); 3148 if ((cg->cg_flags & CG_FLAG_THREAD) != 0) 3149 sbuf_printf(sb, "<flag name=\"THREAD\">THREAD group</flag>"); 3150 if ((cg->cg_flags & CG_FLAG_SMT) != 0) 3151 sbuf_printf(sb, "<flag name=\"SMT\">SMT group</flag>"); 3152 if ((cg->cg_flags & CG_FLAG_NODE) != 0) 3153 sbuf_printf(sb, "<flag name=\"NODE\">NUMA node</flag>"); 3154 sbuf_printf(sb, "</flags>\n"); 3155 } 3156 3157 if (cg->cg_children > 0) { 3158 sbuf_printf(sb, "%*s <children>\n", indent, ""); 3159 for (i = 0; i < cg->cg_children; i++) 3160 sysctl_kern_sched_topology_spec_internal(sb, 3161 &cg->cg_child[i], indent+2); 3162 sbuf_printf(sb, "%*s </children>\n", indent, ""); 3163 } 3164 sbuf_printf(sb, "%*s</group>\n", indent, ""); 3165 return (0); 3166 } 3167 3168 /* 3169 * Sysctl handler for retrieving topology dump. It's a wrapper for 3170 * the recursive sysctl_kern_smp_topology_spec_internal(). 3171 */ 3172 static int 3173 sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS) 3174 { 3175 struct sbuf *topo; 3176 int err; 3177 3178 KASSERT(cpu_top != NULL, ("cpu_top isn't initialized")); 3179 3180 topo = sbuf_new_for_sysctl(NULL, NULL, 512, req); 3181 if (topo == NULL) 3182 return (ENOMEM); 3183 3184 sbuf_printf(topo, "<groups>\n"); 3185 err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1); 3186 sbuf_printf(topo, "</groups>\n"); 3187 3188 if (err == 0) { 3189 err = sbuf_finish(topo); 3190 } 3191 sbuf_delete(topo); 3192 return (err); 3193 } 3194 3195 #endif 3196 3197 static int 3198 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS) 3199 { 3200 int error, new_val, period; 3201 3202 period = 1000000 / realstathz; 3203 new_val = period * sched_slice; 3204 error = sysctl_handle_int(oidp, &new_val, 0, req); 3205 if (error != 0 || req->newptr == NULL) 3206 return (error); 3207 if (new_val <= 0) 3208 return (EINVAL); 3209 sched_slice = imax(1, (new_val + period / 2) / period); 3210 sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR; 3211 hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) / 3212 realstathz); 3213 return (0); 3214 } 3215 3216 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 3217 "Scheduler"); 3218 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0, 3219 "Scheduler name"); 3220 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, 3221 CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, 0, 3222 sysctl_kern_quantum, "I", 3223 "Quantum for timeshare threads in microseconds"); 3224 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0, 3225 "Quantum for timeshare threads in stathz ticks"); 3226 SYSCTL_UINT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0, 3227 "Interactivity score threshold"); 3228 SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW, 3229 &preempt_thresh, 0, 3230 "Maximal (lowest) priority for preemption"); 3231 SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost, 0, 3232 "Assign static kernel priorities to sleeping threads"); 3233 SYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins, 0, 3234 "Number of times idle thread will spin waiting for new work"); 3235 SYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW, 3236 &sched_idlespinthresh, 0, 3237 "Threshold before we will permit idle thread spinning"); 3238 #ifdef SMP 3239 SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0, 3240 "Number of hz ticks to keep thread affinity for"); 3241 SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0, 3242 "Enables the long-term load balancer"); 3243 SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW, 3244 &balance_interval, 0, 3245 "Average period in stathz ticks to run the long-term balancer"); 3246 SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0, 3247 "Attempts to steal work from other cores before idling"); 3248 SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0, 3249 "Minimum load on remote CPU before we'll steal"); 3250 SYSCTL_INT(_kern_sched, OID_AUTO, trysteal_limit, CTLFLAG_RW, &trysteal_limit, 3251 0, "Topological distance limit for stealing threads in sched_switch()"); 3252 SYSCTL_INT(_kern_sched, OID_AUTO, always_steal, CTLFLAG_RW, &always_steal, 0, 3253 "Always run the stealer from the idle thread"); 3254 SYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING | 3255 CTLFLAG_MPSAFE | CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A", 3256 "XML dump of detected CPU topology"); 3257 #endif 3258 3259 /* ps compat. All cpu percentages from ULE are weighted. */ 3260 static int ccpu = 0; 3261 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, 3262 "Decay factor used for updating %CPU in 4BSD scheduler"); 3263