1 /*- 2 * Copyright (c) 2002-2007, Jeffrey Roberson <jeff@freebsd.org> 3 * All rights reserved. 4 * 5 * Redistribution and use in source and binary forms, with or without 6 * modification, are permitted provided that the following conditions 7 * are met: 8 * 1. Redistributions of source code must retain the above copyright 9 * notice unmodified, this list of conditions, and the following 10 * disclaimer. 11 * 2. Redistributions in binary form must reproduce the above copyright 12 * notice, this list of conditions and the following disclaimer in the 13 * documentation and/or other materials provided with the distribution. 14 * 15 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR 16 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES 17 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. 18 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, 19 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT 20 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 21 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 22 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 23 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF 24 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 25 */ 26 27 /* 28 * This file implements the ULE scheduler. ULE supports independent CPU 29 * run queues and fine grain locking. It has superior interactive 30 * performance under load even on uni-processor systems. 31 * 32 * etymology: 33 * ULE is the last three letters in schedule. It owes its name to a 34 * generic user created for a scheduling system by Paul Mikesell at 35 * Isilon Systems and a general lack of creativity on the part of the author. 36 */ 37 38 #include <sys/cdefs.h> 39 __FBSDID("$FreeBSD$"); 40 41 #include "opt_hwpmc_hooks.h" 42 #include "opt_sched.h" 43 44 #include <sys/param.h> 45 #include <sys/systm.h> 46 #include <sys/kdb.h> 47 #include <sys/kernel.h> 48 #include <sys/ktr.h> 49 #include <sys/limits.h> 50 #include <sys/lock.h> 51 #include <sys/mutex.h> 52 #include <sys/proc.h> 53 #include <sys/resource.h> 54 #include <sys/resourcevar.h> 55 #include <sys/sched.h> 56 #include <sys/sdt.h> 57 #include <sys/smp.h> 58 #include <sys/sx.h> 59 #include <sys/sysctl.h> 60 #include <sys/sysproto.h> 61 #include <sys/turnstile.h> 62 #include <sys/umtx.h> 63 #include <sys/vmmeter.h> 64 #include <sys/cpuset.h> 65 #include <sys/sbuf.h> 66 67 #ifdef HWPMC_HOOKS 68 #include <sys/pmckern.h> 69 #endif 70 71 #ifdef KDTRACE_HOOKS 72 #include <sys/dtrace_bsd.h> 73 int dtrace_vtime_active; 74 dtrace_vtime_switch_func_t dtrace_vtime_switch_func; 75 #endif 76 77 #include <machine/cpu.h> 78 #include <machine/smp.h> 79 80 #define KTR_ULE 0 81 82 #define TS_NAME_LEN (MAXCOMLEN + sizeof(" td ") + sizeof(__XSTRING(UINT_MAX))) 83 #define TDQ_NAME_LEN (sizeof("sched lock ") + sizeof(__XSTRING(MAXCPU))) 84 #define TDQ_LOADNAME_LEN (sizeof("CPU ") + sizeof(__XSTRING(MAXCPU)) - 1 + sizeof(" load")) 85 86 /* 87 * Thread scheduler specific section. All fields are protected 88 * by the thread lock. 89 */ 90 struct td_sched { 91 struct runq *ts_runq; /* Run-queue we're queued on. */ 92 short ts_flags; /* TSF_* flags. */ 93 int ts_cpu; /* CPU that we have affinity for. */ 94 int ts_rltick; /* Real last tick, for affinity. */ 95 int ts_slice; /* Ticks of slice remaining. */ 96 u_int ts_slptime; /* Number of ticks we vol. slept */ 97 u_int ts_runtime; /* Number of ticks we were running */ 98 int ts_ltick; /* Last tick that we were running on */ 99 int ts_ftick; /* First tick that we were running on */ 100 int ts_ticks; /* Tick count */ 101 #ifdef KTR 102 char ts_name[TS_NAME_LEN]; 103 #endif 104 }; 105 /* flags kept in ts_flags */ 106 #define TSF_BOUND 0x0001 /* Thread can not migrate. */ 107 #define TSF_XFERABLE 0x0002 /* Thread was added as transferable. */ 108 109 #define THREAD_CAN_MIGRATE(td) ((td)->td_pinned == 0) 110 #define THREAD_CAN_SCHED(td, cpu) \ 111 CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask) 112 113 _Static_assert(sizeof(struct thread) + sizeof(struct td_sched) <= 114 sizeof(struct thread0_storage), 115 "increase struct thread0_storage.t0st_sched size"); 116 117 /* 118 * Priority ranges used for interactive and non-interactive timeshare 119 * threads. The timeshare priorities are split up into four ranges. 120 * The first range handles interactive threads. The last three ranges 121 * (NHALF, x, and NHALF) handle non-interactive threads with the outer 122 * ranges supporting nice values. 123 */ 124 #define PRI_TIMESHARE_RANGE (PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE + 1) 125 #define PRI_INTERACT_RANGE ((PRI_TIMESHARE_RANGE - SCHED_PRI_NRESV) / 2) 126 #define PRI_BATCH_RANGE (PRI_TIMESHARE_RANGE - PRI_INTERACT_RANGE) 127 128 #define PRI_MIN_INTERACT PRI_MIN_TIMESHARE 129 #define PRI_MAX_INTERACT (PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE - 1) 130 #define PRI_MIN_BATCH (PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE) 131 #define PRI_MAX_BATCH PRI_MAX_TIMESHARE 132 133 /* 134 * Cpu percentage computation macros and defines. 135 * 136 * SCHED_TICK_SECS: Number of seconds to average the cpu usage across. 137 * SCHED_TICK_TARG: Number of hz ticks to average the cpu usage across. 138 * SCHED_TICK_MAX: Maximum number of ticks before scaling back. 139 * SCHED_TICK_SHIFT: Shift factor to avoid rounding away results. 140 * SCHED_TICK_HZ: Compute the number of hz ticks for a given ticks count. 141 * SCHED_TICK_TOTAL: Gives the amount of time we've been recording ticks. 142 */ 143 #define SCHED_TICK_SECS 10 144 #define SCHED_TICK_TARG (hz * SCHED_TICK_SECS) 145 #define SCHED_TICK_MAX (SCHED_TICK_TARG + hz) 146 #define SCHED_TICK_SHIFT 10 147 #define SCHED_TICK_HZ(ts) ((ts)->ts_ticks >> SCHED_TICK_SHIFT) 148 #define SCHED_TICK_TOTAL(ts) (max((ts)->ts_ltick - (ts)->ts_ftick, hz)) 149 150 /* 151 * These macros determine priorities for non-interactive threads. They are 152 * assigned a priority based on their recent cpu utilization as expressed 153 * by the ratio of ticks to the tick total. NHALF priorities at the start 154 * and end of the MIN to MAX timeshare range are only reachable with negative 155 * or positive nice respectively. 156 * 157 * PRI_RANGE: Priority range for utilization dependent priorities. 158 * PRI_NRESV: Number of nice values. 159 * PRI_TICKS: Compute a priority in PRI_RANGE from the ticks count and total. 160 * PRI_NICE: Determines the part of the priority inherited from nice. 161 */ 162 #define SCHED_PRI_NRESV (PRIO_MAX - PRIO_MIN) 163 #define SCHED_PRI_NHALF (SCHED_PRI_NRESV / 2) 164 #define SCHED_PRI_MIN (PRI_MIN_BATCH + SCHED_PRI_NHALF) 165 #define SCHED_PRI_MAX (PRI_MAX_BATCH - SCHED_PRI_NHALF) 166 #define SCHED_PRI_RANGE (SCHED_PRI_MAX - SCHED_PRI_MIN + 1) 167 #define SCHED_PRI_TICKS(ts) \ 168 (SCHED_TICK_HZ((ts)) / \ 169 (roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE)) 170 #define SCHED_PRI_NICE(nice) (nice) 171 172 /* 173 * These determine the interactivity of a process. Interactivity differs from 174 * cpu utilization in that it expresses the voluntary time slept vs time ran 175 * while cpu utilization includes all time not running. This more accurately 176 * models the intent of the thread. 177 * 178 * SLP_RUN_MAX: Maximum amount of sleep time + run time we'll accumulate 179 * before throttling back. 180 * SLP_RUN_FORK: Maximum slp+run time to inherit at fork time. 181 * INTERACT_MAX: Maximum interactivity value. Smaller is better. 182 * INTERACT_THRESH: Threshold for placement on the current runq. 183 */ 184 #define SCHED_SLP_RUN_MAX ((hz * 5) << SCHED_TICK_SHIFT) 185 #define SCHED_SLP_RUN_FORK ((hz / 2) << SCHED_TICK_SHIFT) 186 #define SCHED_INTERACT_MAX (100) 187 #define SCHED_INTERACT_HALF (SCHED_INTERACT_MAX / 2) 188 #define SCHED_INTERACT_THRESH (30) 189 190 /* 191 * These parameters determine the slice behavior for batch work. 192 */ 193 #define SCHED_SLICE_DEFAULT_DIVISOR 10 /* ~94 ms, 12 stathz ticks. */ 194 #define SCHED_SLICE_MIN_DIVISOR 6 /* DEFAULT/MIN = ~16 ms. */ 195 196 /* Flags kept in td_flags. */ 197 #define TDF_SLICEEND TDF_SCHED2 /* Thread time slice is over. */ 198 199 /* 200 * tickincr: Converts a stathz tick into a hz domain scaled by 201 * the shift factor. Without the shift the error rate 202 * due to rounding would be unacceptably high. 203 * realstathz: stathz is sometimes 0 and run off of hz. 204 * sched_slice: Runtime of each thread before rescheduling. 205 * preempt_thresh: Priority threshold for preemption and remote IPIs. 206 */ 207 static int sched_interact = SCHED_INTERACT_THRESH; 208 static int tickincr = 8 << SCHED_TICK_SHIFT; 209 static int realstathz = 127; /* reset during boot. */ 210 static int sched_slice = 10; /* reset during boot. */ 211 static int sched_slice_min = 1; /* reset during boot. */ 212 #ifdef PREEMPTION 213 #ifdef FULL_PREEMPTION 214 static int preempt_thresh = PRI_MAX_IDLE; 215 #else 216 static int preempt_thresh = PRI_MIN_KERN; 217 #endif 218 #else 219 static int preempt_thresh = 0; 220 #endif 221 static int static_boost = PRI_MIN_BATCH; 222 static int sched_idlespins = 10000; 223 static int sched_idlespinthresh = -1; 224 225 /* 226 * tdq - per processor runqs and statistics. All fields are protected by the 227 * tdq_lock. The load and lowpri may be accessed without to avoid excess 228 * locking in sched_pickcpu(); 229 */ 230 struct tdq { 231 /* 232 * Ordered to improve efficiency of cpu_search() and switch(). 233 * tdq_lock is padded to avoid false sharing with tdq_load and 234 * tdq_cpu_idle. 235 */ 236 struct mtx_padalign tdq_lock; /* run queue lock. */ 237 struct cpu_group *tdq_cg; /* Pointer to cpu topology. */ 238 volatile int tdq_load; /* Aggregate load. */ 239 volatile int tdq_cpu_idle; /* cpu_idle() is active. */ 240 int tdq_sysload; /* For loadavg, !ITHD load. */ 241 int tdq_transferable; /* Transferable thread count. */ 242 short tdq_switchcnt; /* Switches this tick. */ 243 short tdq_oldswitchcnt; /* Switches last tick. */ 244 u_char tdq_lowpri; /* Lowest priority thread. */ 245 u_char tdq_ipipending; /* IPI pending. */ 246 u_char tdq_idx; /* Current insert index. */ 247 u_char tdq_ridx; /* Current removal index. */ 248 struct runq tdq_realtime; /* real-time run queue. */ 249 struct runq tdq_timeshare; /* timeshare run queue. */ 250 struct runq tdq_idle; /* Queue of IDLE threads. */ 251 char tdq_name[TDQ_NAME_LEN]; 252 #ifdef KTR 253 char tdq_loadname[TDQ_LOADNAME_LEN]; 254 #endif 255 } __aligned(64); 256 257 /* Idle thread states and config. */ 258 #define TDQ_RUNNING 1 259 #define TDQ_IDLE 2 260 261 #ifdef SMP 262 struct cpu_group *cpu_top; /* CPU topology */ 263 264 #define SCHED_AFFINITY_DEFAULT (max(1, hz / 1000)) 265 #define SCHED_AFFINITY(ts, t) ((ts)->ts_rltick > ticks - ((t) * affinity)) 266 267 /* 268 * Run-time tunables. 269 */ 270 static int rebalance = 1; 271 static int balance_interval = 128; /* Default set in sched_initticks(). */ 272 static int affinity; 273 static int steal_idle = 1; 274 static int steal_thresh = 2; 275 276 /* 277 * One thread queue per processor. 278 */ 279 static struct tdq tdq_cpu[MAXCPU]; 280 static struct tdq *balance_tdq; 281 static int balance_ticks; 282 static DPCPU_DEFINE(uint32_t, randomval); 283 284 #define TDQ_SELF() (&tdq_cpu[PCPU_GET(cpuid)]) 285 #define TDQ_CPU(x) (&tdq_cpu[(x)]) 286 #define TDQ_ID(x) ((int)((x) - tdq_cpu)) 287 #else /* !SMP */ 288 static struct tdq tdq_cpu; 289 290 #define TDQ_ID(x) (0) 291 #define TDQ_SELF() (&tdq_cpu) 292 #define TDQ_CPU(x) (&tdq_cpu) 293 #endif 294 295 #define TDQ_LOCK_ASSERT(t, type) mtx_assert(TDQ_LOCKPTR((t)), (type)) 296 #define TDQ_LOCK(t) mtx_lock_spin(TDQ_LOCKPTR((t))) 297 #define TDQ_LOCK_FLAGS(t, f) mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f)) 298 #define TDQ_UNLOCK(t) mtx_unlock_spin(TDQ_LOCKPTR((t))) 299 #define TDQ_LOCKPTR(t) ((struct mtx *)(&(t)->tdq_lock)) 300 301 static void sched_priority(struct thread *); 302 static void sched_thread_priority(struct thread *, u_char); 303 static int sched_interact_score(struct thread *); 304 static void sched_interact_update(struct thread *); 305 static void sched_interact_fork(struct thread *); 306 static void sched_pctcpu_update(struct td_sched *, int); 307 308 /* Operations on per processor queues */ 309 static struct thread *tdq_choose(struct tdq *); 310 static void tdq_setup(struct tdq *); 311 static void tdq_load_add(struct tdq *, struct thread *); 312 static void tdq_load_rem(struct tdq *, struct thread *); 313 static __inline void tdq_runq_add(struct tdq *, struct thread *, int); 314 static __inline void tdq_runq_rem(struct tdq *, struct thread *); 315 static inline int sched_shouldpreempt(int, int, int); 316 void tdq_print(int cpu); 317 static void runq_print(struct runq *rq); 318 static void tdq_add(struct tdq *, struct thread *, int); 319 #ifdef SMP 320 static int tdq_move(struct tdq *, struct tdq *); 321 static int tdq_idled(struct tdq *); 322 static void tdq_notify(struct tdq *, int); 323 static struct thread *tdq_steal(struct tdq *, int); 324 static struct thread *runq_steal(struct runq *, int); 325 static int sched_pickcpu(struct thread *, int); 326 static void sched_balance(void); 327 static int sched_balance_pair(struct tdq *, struct tdq *); 328 static inline struct tdq *sched_setcpu(struct thread *, int, int); 329 static inline void thread_unblock_switch(struct thread *, struct mtx *); 330 static struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int); 331 static int sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS); 332 static int sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, 333 struct cpu_group *cg, int indent); 334 #endif 335 336 static void sched_setup(void *dummy); 337 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL); 338 339 static void sched_initticks(void *dummy); 340 SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks, 341 NULL); 342 343 SDT_PROVIDER_DEFINE(sched); 344 345 SDT_PROBE_DEFINE3(sched, , , change__pri, "struct thread *", 346 "struct proc *", "uint8_t"); 347 SDT_PROBE_DEFINE3(sched, , , dequeue, "struct thread *", 348 "struct proc *", "void *"); 349 SDT_PROBE_DEFINE4(sched, , , enqueue, "struct thread *", 350 "struct proc *", "void *", "int"); 351 SDT_PROBE_DEFINE4(sched, , , lend__pri, "struct thread *", 352 "struct proc *", "uint8_t", "struct thread *"); 353 SDT_PROBE_DEFINE2(sched, , , load__change, "int", "int"); 354 SDT_PROBE_DEFINE2(sched, , , off__cpu, "struct thread *", 355 "struct proc *"); 356 SDT_PROBE_DEFINE(sched, , , on__cpu); 357 SDT_PROBE_DEFINE(sched, , , remain__cpu); 358 SDT_PROBE_DEFINE2(sched, , , surrender, "struct thread *", 359 "struct proc *"); 360 361 /* 362 * Print the threads waiting on a run-queue. 363 */ 364 static void 365 runq_print(struct runq *rq) 366 { 367 struct rqhead *rqh; 368 struct thread *td; 369 int pri; 370 int j; 371 int i; 372 373 for (i = 0; i < RQB_LEN; i++) { 374 printf("\t\trunq bits %d 0x%zx\n", 375 i, rq->rq_status.rqb_bits[i]); 376 for (j = 0; j < RQB_BPW; j++) 377 if (rq->rq_status.rqb_bits[i] & (1ul << j)) { 378 pri = j + (i << RQB_L2BPW); 379 rqh = &rq->rq_queues[pri]; 380 TAILQ_FOREACH(td, rqh, td_runq) { 381 printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n", 382 td, td->td_name, td->td_priority, 383 td->td_rqindex, pri); 384 } 385 } 386 } 387 } 388 389 /* 390 * Print the status of a per-cpu thread queue. Should be a ddb show cmd. 391 */ 392 void 393 tdq_print(int cpu) 394 { 395 struct tdq *tdq; 396 397 tdq = TDQ_CPU(cpu); 398 399 printf("tdq %d:\n", TDQ_ID(tdq)); 400 printf("\tlock %p\n", TDQ_LOCKPTR(tdq)); 401 printf("\tLock name: %s\n", tdq->tdq_name); 402 printf("\tload: %d\n", tdq->tdq_load); 403 printf("\tswitch cnt: %d\n", tdq->tdq_switchcnt); 404 printf("\told switch cnt: %d\n", tdq->tdq_oldswitchcnt); 405 printf("\ttimeshare idx: %d\n", tdq->tdq_idx); 406 printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx); 407 printf("\tload transferable: %d\n", tdq->tdq_transferable); 408 printf("\tlowest priority: %d\n", tdq->tdq_lowpri); 409 printf("\trealtime runq:\n"); 410 runq_print(&tdq->tdq_realtime); 411 printf("\ttimeshare runq:\n"); 412 runq_print(&tdq->tdq_timeshare); 413 printf("\tidle runq:\n"); 414 runq_print(&tdq->tdq_idle); 415 } 416 417 static inline int 418 sched_shouldpreempt(int pri, int cpri, int remote) 419 { 420 /* 421 * If the new priority is not better than the current priority there is 422 * nothing to do. 423 */ 424 if (pri >= cpri) 425 return (0); 426 /* 427 * Always preempt idle. 428 */ 429 if (cpri >= PRI_MIN_IDLE) 430 return (1); 431 /* 432 * If preemption is disabled don't preempt others. 433 */ 434 if (preempt_thresh == 0) 435 return (0); 436 /* 437 * Preempt if we exceed the threshold. 438 */ 439 if (pri <= preempt_thresh) 440 return (1); 441 /* 442 * If we're interactive or better and there is non-interactive 443 * or worse running preempt only remote processors. 444 */ 445 if (remote && pri <= PRI_MAX_INTERACT && cpri > PRI_MAX_INTERACT) 446 return (1); 447 return (0); 448 } 449 450 /* 451 * Add a thread to the actual run-queue. Keeps transferable counts up to 452 * date with what is actually on the run-queue. Selects the correct 453 * queue position for timeshare threads. 454 */ 455 static __inline void 456 tdq_runq_add(struct tdq *tdq, struct thread *td, int flags) 457 { 458 struct td_sched *ts; 459 u_char pri; 460 461 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 462 THREAD_LOCK_ASSERT(td, MA_OWNED); 463 464 pri = td->td_priority; 465 ts = td_get_sched(td); 466 TD_SET_RUNQ(td); 467 if (THREAD_CAN_MIGRATE(td)) { 468 tdq->tdq_transferable++; 469 ts->ts_flags |= TSF_XFERABLE; 470 } 471 if (pri < PRI_MIN_BATCH) { 472 ts->ts_runq = &tdq->tdq_realtime; 473 } else if (pri <= PRI_MAX_BATCH) { 474 ts->ts_runq = &tdq->tdq_timeshare; 475 KASSERT(pri <= PRI_MAX_BATCH && pri >= PRI_MIN_BATCH, 476 ("Invalid priority %d on timeshare runq", pri)); 477 /* 478 * This queue contains only priorities between MIN and MAX 479 * realtime. Use the whole queue to represent these values. 480 */ 481 if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) { 482 pri = RQ_NQS * (pri - PRI_MIN_BATCH) / PRI_BATCH_RANGE; 483 pri = (pri + tdq->tdq_idx) % RQ_NQS; 484 /* 485 * This effectively shortens the queue by one so we 486 * can have a one slot difference between idx and 487 * ridx while we wait for threads to drain. 488 */ 489 if (tdq->tdq_ridx != tdq->tdq_idx && 490 pri == tdq->tdq_ridx) 491 pri = (unsigned char)(pri - 1) % RQ_NQS; 492 } else 493 pri = tdq->tdq_ridx; 494 runq_add_pri(ts->ts_runq, td, pri, flags); 495 return; 496 } else 497 ts->ts_runq = &tdq->tdq_idle; 498 runq_add(ts->ts_runq, td, flags); 499 } 500 501 /* 502 * Remove a thread from a run-queue. This typically happens when a thread 503 * is selected to run. Running threads are not on the queue and the 504 * transferable count does not reflect them. 505 */ 506 static __inline void 507 tdq_runq_rem(struct tdq *tdq, struct thread *td) 508 { 509 struct td_sched *ts; 510 511 ts = td_get_sched(td); 512 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 513 KASSERT(ts->ts_runq != NULL, 514 ("tdq_runq_remove: thread %p null ts_runq", td)); 515 if (ts->ts_flags & TSF_XFERABLE) { 516 tdq->tdq_transferable--; 517 ts->ts_flags &= ~TSF_XFERABLE; 518 } 519 if (ts->ts_runq == &tdq->tdq_timeshare) { 520 if (tdq->tdq_idx != tdq->tdq_ridx) 521 runq_remove_idx(ts->ts_runq, td, &tdq->tdq_ridx); 522 else 523 runq_remove_idx(ts->ts_runq, td, NULL); 524 } else 525 runq_remove(ts->ts_runq, td); 526 } 527 528 /* 529 * Load is maintained for all threads RUNNING and ON_RUNQ. Add the load 530 * for this thread to the referenced thread queue. 531 */ 532 static void 533 tdq_load_add(struct tdq *tdq, struct thread *td) 534 { 535 536 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 537 THREAD_LOCK_ASSERT(td, MA_OWNED); 538 539 tdq->tdq_load++; 540 if ((td->td_flags & TDF_NOLOAD) == 0) 541 tdq->tdq_sysload++; 542 KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load); 543 SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load); 544 } 545 546 /* 547 * Remove the load from a thread that is transitioning to a sleep state or 548 * exiting. 549 */ 550 static void 551 tdq_load_rem(struct tdq *tdq, struct thread *td) 552 { 553 554 THREAD_LOCK_ASSERT(td, MA_OWNED); 555 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 556 KASSERT(tdq->tdq_load != 0, 557 ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq))); 558 559 tdq->tdq_load--; 560 if ((td->td_flags & TDF_NOLOAD) == 0) 561 tdq->tdq_sysload--; 562 KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load); 563 SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load); 564 } 565 566 /* 567 * Bound timeshare latency by decreasing slice size as load increases. We 568 * consider the maximum latency as the sum of the threads waiting to run 569 * aside from curthread and target no more than sched_slice latency but 570 * no less than sched_slice_min runtime. 571 */ 572 static inline int 573 tdq_slice(struct tdq *tdq) 574 { 575 int load; 576 577 /* 578 * It is safe to use sys_load here because this is called from 579 * contexts where timeshare threads are running and so there 580 * cannot be higher priority load in the system. 581 */ 582 load = tdq->tdq_sysload - 1; 583 if (load >= SCHED_SLICE_MIN_DIVISOR) 584 return (sched_slice_min); 585 if (load <= 1) 586 return (sched_slice); 587 return (sched_slice / load); 588 } 589 590 /* 591 * Set lowpri to its exact value by searching the run-queue and 592 * evaluating curthread. curthread may be passed as an optimization. 593 */ 594 static void 595 tdq_setlowpri(struct tdq *tdq, struct thread *ctd) 596 { 597 struct thread *td; 598 599 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 600 if (ctd == NULL) 601 ctd = pcpu_find(TDQ_ID(tdq))->pc_curthread; 602 td = tdq_choose(tdq); 603 if (td == NULL || td->td_priority > ctd->td_priority) 604 tdq->tdq_lowpri = ctd->td_priority; 605 else 606 tdq->tdq_lowpri = td->td_priority; 607 } 608 609 #ifdef SMP 610 /* 611 * We need some randomness. Implement a classic Linear Congruential 612 * Generator X_{n+1}=(aX_n+c) mod m. These values are optimized for 613 * m = 2^32, a = 69069 and c = 5. We only return the upper 16 bits 614 * of the random state (in the low bits of our answer) to keep 615 * the maximum randomness. 616 */ 617 static uint32_t 618 sched_random(void) 619 { 620 uint32_t *rndptr; 621 622 rndptr = DPCPU_PTR(randomval); 623 *rndptr = *rndptr * 69069 + 5; 624 625 return (*rndptr >> 16); 626 } 627 628 struct cpu_search { 629 cpuset_t cs_mask; 630 u_int cs_prefer; 631 int cs_pri; /* Min priority for low. */ 632 int cs_limit; /* Max load for low, min load for high. */ 633 int cs_cpu; 634 int cs_load; 635 }; 636 637 #define CPU_SEARCH_LOWEST 0x1 638 #define CPU_SEARCH_HIGHEST 0x2 639 #define CPU_SEARCH_BOTH (CPU_SEARCH_LOWEST|CPU_SEARCH_HIGHEST) 640 641 #define CPUSET_FOREACH(cpu, mask) \ 642 for ((cpu) = 0; (cpu) <= mp_maxid; (cpu)++) \ 643 if (CPU_ISSET(cpu, &mask)) 644 645 static __always_inline int cpu_search(const struct cpu_group *cg, 646 struct cpu_search *low, struct cpu_search *high, const int match); 647 int __noinline cpu_search_lowest(const struct cpu_group *cg, 648 struct cpu_search *low); 649 int __noinline cpu_search_highest(const struct cpu_group *cg, 650 struct cpu_search *high); 651 int __noinline cpu_search_both(const struct cpu_group *cg, 652 struct cpu_search *low, struct cpu_search *high); 653 654 /* 655 * Search the tree of cpu_groups for the lowest or highest loaded cpu 656 * according to the match argument. This routine actually compares the 657 * load on all paths through the tree and finds the least loaded cpu on 658 * the least loaded path, which may differ from the least loaded cpu in 659 * the system. This balances work among caches and buses. 660 * 661 * This inline is instantiated in three forms below using constants for the 662 * match argument. It is reduced to the minimum set for each case. It is 663 * also recursive to the depth of the tree. 664 */ 665 static __always_inline int 666 cpu_search(const struct cpu_group *cg, struct cpu_search *low, 667 struct cpu_search *high, const int match) 668 { 669 struct cpu_search lgroup; 670 struct cpu_search hgroup; 671 cpuset_t cpumask; 672 struct cpu_group *child; 673 struct tdq *tdq; 674 int cpu, i, hload, lload, load, total, rnd; 675 676 total = 0; 677 cpumask = cg->cg_mask; 678 if (match & CPU_SEARCH_LOWEST) { 679 lload = INT_MAX; 680 lgroup = *low; 681 } 682 if (match & CPU_SEARCH_HIGHEST) { 683 hload = INT_MIN; 684 hgroup = *high; 685 } 686 687 /* Iterate through the child CPU groups and then remaining CPUs. */ 688 for (i = cg->cg_children, cpu = mp_maxid; ; ) { 689 if (i == 0) { 690 #ifdef HAVE_INLINE_FFSL 691 cpu = CPU_FFS(&cpumask) - 1; 692 #else 693 while (cpu >= 0 && !CPU_ISSET(cpu, &cpumask)) 694 cpu--; 695 #endif 696 if (cpu < 0) 697 break; 698 child = NULL; 699 } else 700 child = &cg->cg_child[i - 1]; 701 702 if (match & CPU_SEARCH_LOWEST) 703 lgroup.cs_cpu = -1; 704 if (match & CPU_SEARCH_HIGHEST) 705 hgroup.cs_cpu = -1; 706 if (child) { /* Handle child CPU group. */ 707 CPU_NAND(&cpumask, &child->cg_mask); 708 switch (match) { 709 case CPU_SEARCH_LOWEST: 710 load = cpu_search_lowest(child, &lgroup); 711 break; 712 case CPU_SEARCH_HIGHEST: 713 load = cpu_search_highest(child, &hgroup); 714 break; 715 case CPU_SEARCH_BOTH: 716 load = cpu_search_both(child, &lgroup, &hgroup); 717 break; 718 } 719 } else { /* Handle child CPU. */ 720 CPU_CLR(cpu, &cpumask); 721 tdq = TDQ_CPU(cpu); 722 load = tdq->tdq_load * 256; 723 rnd = sched_random() % 32; 724 if (match & CPU_SEARCH_LOWEST) { 725 if (cpu == low->cs_prefer) 726 load -= 64; 727 /* If that CPU is allowed and get data. */ 728 if (tdq->tdq_lowpri > lgroup.cs_pri && 729 tdq->tdq_load <= lgroup.cs_limit && 730 CPU_ISSET(cpu, &lgroup.cs_mask)) { 731 lgroup.cs_cpu = cpu; 732 lgroup.cs_load = load - rnd; 733 } 734 } 735 if (match & CPU_SEARCH_HIGHEST) 736 if (tdq->tdq_load >= hgroup.cs_limit && 737 tdq->tdq_transferable && 738 CPU_ISSET(cpu, &hgroup.cs_mask)) { 739 hgroup.cs_cpu = cpu; 740 hgroup.cs_load = load - rnd; 741 } 742 } 743 total += load; 744 745 /* We have info about child item. Compare it. */ 746 if (match & CPU_SEARCH_LOWEST) { 747 if (lgroup.cs_cpu >= 0 && 748 (load < lload || 749 (load == lload && lgroup.cs_load < low->cs_load))) { 750 lload = load; 751 low->cs_cpu = lgroup.cs_cpu; 752 low->cs_load = lgroup.cs_load; 753 } 754 } 755 if (match & CPU_SEARCH_HIGHEST) 756 if (hgroup.cs_cpu >= 0 && 757 (load > hload || 758 (load == hload && hgroup.cs_load > high->cs_load))) { 759 hload = load; 760 high->cs_cpu = hgroup.cs_cpu; 761 high->cs_load = hgroup.cs_load; 762 } 763 if (child) { 764 i--; 765 if (i == 0 && CPU_EMPTY(&cpumask)) 766 break; 767 } 768 #ifndef HAVE_INLINE_FFSL 769 else 770 cpu--; 771 #endif 772 } 773 return (total); 774 } 775 776 /* 777 * cpu_search instantiations must pass constants to maintain the inline 778 * optimization. 779 */ 780 int 781 cpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low) 782 { 783 return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST); 784 } 785 786 int 787 cpu_search_highest(const struct cpu_group *cg, struct cpu_search *high) 788 { 789 return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST); 790 } 791 792 int 793 cpu_search_both(const struct cpu_group *cg, struct cpu_search *low, 794 struct cpu_search *high) 795 { 796 return cpu_search(cg, low, high, CPU_SEARCH_BOTH); 797 } 798 799 /* 800 * Find the cpu with the least load via the least loaded path that has a 801 * lowpri greater than pri pri. A pri of -1 indicates any priority is 802 * acceptable. 803 */ 804 static inline int 805 sched_lowest(const struct cpu_group *cg, cpuset_t mask, int pri, int maxload, 806 int prefer) 807 { 808 struct cpu_search low; 809 810 low.cs_cpu = -1; 811 low.cs_prefer = prefer; 812 low.cs_mask = mask; 813 low.cs_pri = pri; 814 low.cs_limit = maxload; 815 cpu_search_lowest(cg, &low); 816 return low.cs_cpu; 817 } 818 819 /* 820 * Find the cpu with the highest load via the highest loaded path. 821 */ 822 static inline int 823 sched_highest(const struct cpu_group *cg, cpuset_t mask, int minload) 824 { 825 struct cpu_search high; 826 827 high.cs_cpu = -1; 828 high.cs_mask = mask; 829 high.cs_limit = minload; 830 cpu_search_highest(cg, &high); 831 return high.cs_cpu; 832 } 833 834 static void 835 sched_balance_group(struct cpu_group *cg) 836 { 837 cpuset_t hmask, lmask; 838 int high, low, anylow; 839 840 CPU_FILL(&hmask); 841 for (;;) { 842 high = sched_highest(cg, hmask, 1); 843 /* Stop if there is no more CPU with transferrable threads. */ 844 if (high == -1) 845 break; 846 CPU_CLR(high, &hmask); 847 CPU_COPY(&hmask, &lmask); 848 /* Stop if there is no more CPU left for low. */ 849 if (CPU_EMPTY(&lmask)) 850 break; 851 anylow = 1; 852 nextlow: 853 low = sched_lowest(cg, lmask, -1, 854 TDQ_CPU(high)->tdq_load - 1, high); 855 /* Stop if we looked well and found no less loaded CPU. */ 856 if (anylow && low == -1) 857 break; 858 /* Go to next high if we found no less loaded CPU. */ 859 if (low == -1) 860 continue; 861 /* Transfer thread from high to low. */ 862 if (sched_balance_pair(TDQ_CPU(high), TDQ_CPU(low))) { 863 /* CPU that got thread can no longer be a donor. */ 864 CPU_CLR(low, &hmask); 865 } else { 866 /* 867 * If failed, then there is no threads on high 868 * that can run on this low. Drop low from low 869 * mask and look for different one. 870 */ 871 CPU_CLR(low, &lmask); 872 anylow = 0; 873 goto nextlow; 874 } 875 } 876 } 877 878 static void 879 sched_balance(void) 880 { 881 struct tdq *tdq; 882 883 if (smp_started == 0 || rebalance == 0) 884 return; 885 886 balance_ticks = max(balance_interval / 2, 1) + 887 (sched_random() % balance_interval); 888 tdq = TDQ_SELF(); 889 TDQ_UNLOCK(tdq); 890 sched_balance_group(cpu_top); 891 TDQ_LOCK(tdq); 892 } 893 894 /* 895 * Lock two thread queues using their address to maintain lock order. 896 */ 897 static void 898 tdq_lock_pair(struct tdq *one, struct tdq *two) 899 { 900 if (one < two) { 901 TDQ_LOCK(one); 902 TDQ_LOCK_FLAGS(two, MTX_DUPOK); 903 } else { 904 TDQ_LOCK(two); 905 TDQ_LOCK_FLAGS(one, MTX_DUPOK); 906 } 907 } 908 909 /* 910 * Unlock two thread queues. Order is not important here. 911 */ 912 static void 913 tdq_unlock_pair(struct tdq *one, struct tdq *two) 914 { 915 TDQ_UNLOCK(one); 916 TDQ_UNLOCK(two); 917 } 918 919 /* 920 * Transfer load between two imbalanced thread queues. 921 */ 922 static int 923 sched_balance_pair(struct tdq *high, struct tdq *low) 924 { 925 int moved; 926 int cpu; 927 928 tdq_lock_pair(high, low); 929 moved = 0; 930 /* 931 * Determine what the imbalance is and then adjust that to how many 932 * threads we actually have to give up (transferable). 933 */ 934 if (high->tdq_transferable != 0 && high->tdq_load > low->tdq_load && 935 (moved = tdq_move(high, low)) > 0) { 936 /* 937 * In case the target isn't the current cpu IPI it to force a 938 * reschedule with the new workload. 939 */ 940 cpu = TDQ_ID(low); 941 if (cpu != PCPU_GET(cpuid)) 942 ipi_cpu(cpu, IPI_PREEMPT); 943 } 944 tdq_unlock_pair(high, low); 945 return (moved); 946 } 947 948 /* 949 * Move a thread from one thread queue to another. 950 */ 951 static int 952 tdq_move(struct tdq *from, struct tdq *to) 953 { 954 struct td_sched *ts; 955 struct thread *td; 956 struct tdq *tdq; 957 int cpu; 958 959 TDQ_LOCK_ASSERT(from, MA_OWNED); 960 TDQ_LOCK_ASSERT(to, MA_OWNED); 961 962 tdq = from; 963 cpu = TDQ_ID(to); 964 td = tdq_steal(tdq, cpu); 965 if (td == NULL) 966 return (0); 967 ts = td_get_sched(td); 968 /* 969 * Although the run queue is locked the thread may be blocked. Lock 970 * it to clear this and acquire the run-queue lock. 971 */ 972 thread_lock(td); 973 /* Drop recursive lock on from acquired via thread_lock(). */ 974 TDQ_UNLOCK(from); 975 sched_rem(td); 976 ts->ts_cpu = cpu; 977 td->td_lock = TDQ_LOCKPTR(to); 978 tdq_add(to, td, SRQ_YIELDING); 979 return (1); 980 } 981 982 /* 983 * This tdq has idled. Try to steal a thread from another cpu and switch 984 * to it. 985 */ 986 static int 987 tdq_idled(struct tdq *tdq) 988 { 989 struct cpu_group *cg; 990 struct tdq *steal; 991 cpuset_t mask; 992 int thresh; 993 int cpu; 994 995 if (smp_started == 0 || steal_idle == 0) 996 return (1); 997 CPU_FILL(&mask); 998 CPU_CLR(PCPU_GET(cpuid), &mask); 999 /* We don't want to be preempted while we're iterating. */ 1000 spinlock_enter(); 1001 for (cg = tdq->tdq_cg; cg != NULL; ) { 1002 if ((cg->cg_flags & CG_FLAG_THREAD) == 0) 1003 thresh = steal_thresh; 1004 else 1005 thresh = 1; 1006 cpu = sched_highest(cg, mask, thresh); 1007 if (cpu == -1) { 1008 cg = cg->cg_parent; 1009 continue; 1010 } 1011 steal = TDQ_CPU(cpu); 1012 CPU_CLR(cpu, &mask); 1013 tdq_lock_pair(tdq, steal); 1014 if (steal->tdq_load < thresh || steal->tdq_transferable == 0) { 1015 tdq_unlock_pair(tdq, steal); 1016 continue; 1017 } 1018 /* 1019 * If a thread was added while interrupts were disabled don't 1020 * steal one here. If we fail to acquire one due to affinity 1021 * restrictions loop again with this cpu removed from the 1022 * set. 1023 */ 1024 if (tdq->tdq_load == 0 && tdq_move(steal, tdq) == 0) { 1025 tdq_unlock_pair(tdq, steal); 1026 continue; 1027 } 1028 spinlock_exit(); 1029 TDQ_UNLOCK(steal); 1030 mi_switch(SW_VOL | SWT_IDLE, NULL); 1031 thread_unlock(curthread); 1032 1033 return (0); 1034 } 1035 spinlock_exit(); 1036 return (1); 1037 } 1038 1039 /* 1040 * Notify a remote cpu of new work. Sends an IPI if criteria are met. 1041 */ 1042 static void 1043 tdq_notify(struct tdq *tdq, int pri) 1044 { 1045 struct thread *ctd; 1046 int cpu; 1047 1048 if (tdq->tdq_ipipending) 1049 return; 1050 cpu = TD_ID(tdq); 1051 ctd = pcpu_find(cpu)->pc_curthread; 1052 if (!sched_shouldpreempt(pri, ctd->td_priority, 1)) 1053 return; 1054 1055 /* 1056 * Make sure that our caller's earlier update to tdq_load is 1057 * globally visible before we read tdq_cpu_idle. Idle thread 1058 * accesses both of them without locks, and the order is important. 1059 */ 1060 atomic_thread_fence_seq_cst(); 1061 1062 if (TD_IS_IDLETHREAD(ctd)) { 1063 /* 1064 * If the MD code has an idle wakeup routine try that before 1065 * falling back to IPI. 1066 */ 1067 if (!tdq->tdq_cpu_idle || cpu_idle_wakeup(cpu)) 1068 return; 1069 } 1070 tdq->tdq_ipipending = 1; 1071 ipi_cpu(cpu, IPI_PREEMPT); 1072 } 1073 1074 /* 1075 * Steals load from a timeshare queue. Honors the rotating queue head 1076 * index. 1077 */ 1078 static struct thread * 1079 runq_steal_from(struct runq *rq, int cpu, u_char start) 1080 { 1081 struct rqbits *rqb; 1082 struct rqhead *rqh; 1083 struct thread *td, *first; 1084 int bit; 1085 int i; 1086 1087 rqb = &rq->rq_status; 1088 bit = start & (RQB_BPW -1); 1089 first = NULL; 1090 again: 1091 for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) { 1092 if (rqb->rqb_bits[i] == 0) 1093 continue; 1094 if (bit == 0) 1095 bit = RQB_FFS(rqb->rqb_bits[i]); 1096 for (; bit < RQB_BPW; bit++) { 1097 if ((rqb->rqb_bits[i] & (1ul << bit)) == 0) 1098 continue; 1099 rqh = &rq->rq_queues[bit + (i << RQB_L2BPW)]; 1100 TAILQ_FOREACH(td, rqh, td_runq) { 1101 if (first && THREAD_CAN_MIGRATE(td) && 1102 THREAD_CAN_SCHED(td, cpu)) 1103 return (td); 1104 first = td; 1105 } 1106 } 1107 } 1108 if (start != 0) { 1109 start = 0; 1110 goto again; 1111 } 1112 1113 if (first && THREAD_CAN_MIGRATE(first) && 1114 THREAD_CAN_SCHED(first, cpu)) 1115 return (first); 1116 return (NULL); 1117 } 1118 1119 /* 1120 * Steals load from a standard linear queue. 1121 */ 1122 static struct thread * 1123 runq_steal(struct runq *rq, int cpu) 1124 { 1125 struct rqhead *rqh; 1126 struct rqbits *rqb; 1127 struct thread *td; 1128 int word; 1129 int bit; 1130 1131 rqb = &rq->rq_status; 1132 for (word = 0; word < RQB_LEN; word++) { 1133 if (rqb->rqb_bits[word] == 0) 1134 continue; 1135 for (bit = 0; bit < RQB_BPW; bit++) { 1136 if ((rqb->rqb_bits[word] & (1ul << bit)) == 0) 1137 continue; 1138 rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)]; 1139 TAILQ_FOREACH(td, rqh, td_runq) 1140 if (THREAD_CAN_MIGRATE(td) && 1141 THREAD_CAN_SCHED(td, cpu)) 1142 return (td); 1143 } 1144 } 1145 return (NULL); 1146 } 1147 1148 /* 1149 * Attempt to steal a thread in priority order from a thread queue. 1150 */ 1151 static struct thread * 1152 tdq_steal(struct tdq *tdq, int cpu) 1153 { 1154 struct thread *td; 1155 1156 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 1157 if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL) 1158 return (td); 1159 if ((td = runq_steal_from(&tdq->tdq_timeshare, 1160 cpu, tdq->tdq_ridx)) != NULL) 1161 return (td); 1162 return (runq_steal(&tdq->tdq_idle, cpu)); 1163 } 1164 1165 /* 1166 * Sets the thread lock and ts_cpu to match the requested cpu. Unlocks the 1167 * current lock and returns with the assigned queue locked. 1168 */ 1169 static inline struct tdq * 1170 sched_setcpu(struct thread *td, int cpu, int flags) 1171 { 1172 1173 struct tdq *tdq; 1174 1175 THREAD_LOCK_ASSERT(td, MA_OWNED); 1176 tdq = TDQ_CPU(cpu); 1177 td_get_sched(td)->ts_cpu = cpu; 1178 /* 1179 * If the lock matches just return the queue. 1180 */ 1181 if (td->td_lock == TDQ_LOCKPTR(tdq)) 1182 return (tdq); 1183 #ifdef notyet 1184 /* 1185 * If the thread isn't running its lockptr is a 1186 * turnstile or a sleepqueue. We can just lock_set without 1187 * blocking. 1188 */ 1189 if (TD_CAN_RUN(td)) { 1190 TDQ_LOCK(tdq); 1191 thread_lock_set(td, TDQ_LOCKPTR(tdq)); 1192 return (tdq); 1193 } 1194 #endif 1195 /* 1196 * The hard case, migration, we need to block the thread first to 1197 * prevent order reversals with other cpus locks. 1198 */ 1199 spinlock_enter(); 1200 thread_lock_block(td); 1201 TDQ_LOCK(tdq); 1202 thread_lock_unblock(td, TDQ_LOCKPTR(tdq)); 1203 spinlock_exit(); 1204 return (tdq); 1205 } 1206 1207 SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding"); 1208 SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity"); 1209 SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity"); 1210 SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load"); 1211 SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu"); 1212 SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration"); 1213 1214 static int 1215 sched_pickcpu(struct thread *td, int flags) 1216 { 1217 struct cpu_group *cg, *ccg; 1218 struct td_sched *ts; 1219 struct tdq *tdq; 1220 cpuset_t mask; 1221 int cpu, pri, self; 1222 1223 self = PCPU_GET(cpuid); 1224 ts = td_get_sched(td); 1225 if (smp_started == 0) 1226 return (self); 1227 /* 1228 * Don't migrate a running thread from sched_switch(). 1229 */ 1230 if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td)) 1231 return (ts->ts_cpu); 1232 /* 1233 * Prefer to run interrupt threads on the processors that generate 1234 * the interrupt. 1235 */ 1236 pri = td->td_priority; 1237 if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) && 1238 curthread->td_intr_nesting_level && ts->ts_cpu != self) { 1239 SCHED_STAT_INC(pickcpu_intrbind); 1240 ts->ts_cpu = self; 1241 if (TDQ_CPU(self)->tdq_lowpri > pri) { 1242 SCHED_STAT_INC(pickcpu_affinity); 1243 return (ts->ts_cpu); 1244 } 1245 } 1246 /* 1247 * If the thread can run on the last cpu and the affinity has not 1248 * expired or it is idle run it there. 1249 */ 1250 tdq = TDQ_CPU(ts->ts_cpu); 1251 cg = tdq->tdq_cg; 1252 if (THREAD_CAN_SCHED(td, ts->ts_cpu) && 1253 tdq->tdq_lowpri >= PRI_MIN_IDLE && 1254 SCHED_AFFINITY(ts, CG_SHARE_L2)) { 1255 if (cg->cg_flags & CG_FLAG_THREAD) { 1256 CPUSET_FOREACH(cpu, cg->cg_mask) { 1257 if (TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE) 1258 break; 1259 } 1260 } else 1261 cpu = INT_MAX; 1262 if (cpu > mp_maxid) { 1263 SCHED_STAT_INC(pickcpu_idle_affinity); 1264 return (ts->ts_cpu); 1265 } 1266 } 1267 /* 1268 * Search for the last level cache CPU group in the tree. 1269 * Skip caches with expired affinity time and SMT groups. 1270 * Affinity to higher level caches will be handled less aggressively. 1271 */ 1272 for (ccg = NULL; cg != NULL; cg = cg->cg_parent) { 1273 if (cg->cg_flags & CG_FLAG_THREAD) 1274 continue; 1275 if (!SCHED_AFFINITY(ts, cg->cg_level)) 1276 continue; 1277 ccg = cg; 1278 } 1279 if (ccg != NULL) 1280 cg = ccg; 1281 cpu = -1; 1282 /* Search the group for the less loaded idle CPU we can run now. */ 1283 mask = td->td_cpuset->cs_mask; 1284 if (cg != NULL && cg != cpu_top && 1285 CPU_CMP(&cg->cg_mask, &cpu_top->cg_mask) != 0) 1286 cpu = sched_lowest(cg, mask, max(pri, PRI_MAX_TIMESHARE), 1287 INT_MAX, ts->ts_cpu); 1288 /* Search globally for the less loaded CPU we can run now. */ 1289 if (cpu == -1) 1290 cpu = sched_lowest(cpu_top, mask, pri, INT_MAX, ts->ts_cpu); 1291 /* Search globally for the less loaded CPU. */ 1292 if (cpu == -1) 1293 cpu = sched_lowest(cpu_top, mask, -1, INT_MAX, ts->ts_cpu); 1294 KASSERT(cpu != -1, ("sched_pickcpu: Failed to find a cpu.")); 1295 /* 1296 * Compare the lowest loaded cpu to current cpu. 1297 */ 1298 if (THREAD_CAN_SCHED(td, self) && TDQ_CPU(self)->tdq_lowpri > pri && 1299 TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE && 1300 TDQ_CPU(self)->tdq_load <= TDQ_CPU(cpu)->tdq_load + 1) { 1301 SCHED_STAT_INC(pickcpu_local); 1302 cpu = self; 1303 } else 1304 SCHED_STAT_INC(pickcpu_lowest); 1305 if (cpu != ts->ts_cpu) 1306 SCHED_STAT_INC(pickcpu_migration); 1307 return (cpu); 1308 } 1309 #endif 1310 1311 /* 1312 * Pick the highest priority task we have and return it. 1313 */ 1314 static struct thread * 1315 tdq_choose(struct tdq *tdq) 1316 { 1317 struct thread *td; 1318 1319 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 1320 td = runq_choose(&tdq->tdq_realtime); 1321 if (td != NULL) 1322 return (td); 1323 td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx); 1324 if (td != NULL) { 1325 KASSERT(td->td_priority >= PRI_MIN_BATCH, 1326 ("tdq_choose: Invalid priority on timeshare queue %d", 1327 td->td_priority)); 1328 return (td); 1329 } 1330 td = runq_choose(&tdq->tdq_idle); 1331 if (td != NULL) { 1332 KASSERT(td->td_priority >= PRI_MIN_IDLE, 1333 ("tdq_choose: Invalid priority on idle queue %d", 1334 td->td_priority)); 1335 return (td); 1336 } 1337 1338 return (NULL); 1339 } 1340 1341 /* 1342 * Initialize a thread queue. 1343 */ 1344 static void 1345 tdq_setup(struct tdq *tdq) 1346 { 1347 1348 if (bootverbose) 1349 printf("ULE: setup cpu %d\n", TDQ_ID(tdq)); 1350 runq_init(&tdq->tdq_realtime); 1351 runq_init(&tdq->tdq_timeshare); 1352 runq_init(&tdq->tdq_idle); 1353 snprintf(tdq->tdq_name, sizeof(tdq->tdq_name), 1354 "sched lock %d", (int)TDQ_ID(tdq)); 1355 mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock", 1356 MTX_SPIN | MTX_RECURSE); 1357 #ifdef KTR 1358 snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname), 1359 "CPU %d load", (int)TDQ_ID(tdq)); 1360 #endif 1361 } 1362 1363 #ifdef SMP 1364 static void 1365 sched_setup_smp(void) 1366 { 1367 struct tdq *tdq; 1368 int i; 1369 1370 cpu_top = smp_topo(); 1371 CPU_FOREACH(i) { 1372 tdq = TDQ_CPU(i); 1373 tdq_setup(tdq); 1374 tdq->tdq_cg = smp_topo_find(cpu_top, i); 1375 if (tdq->tdq_cg == NULL) 1376 panic("Can't find cpu group for %d\n", i); 1377 } 1378 balance_tdq = TDQ_SELF(); 1379 sched_balance(); 1380 } 1381 #endif 1382 1383 /* 1384 * Setup the thread queues and initialize the topology based on MD 1385 * information. 1386 */ 1387 static void 1388 sched_setup(void *dummy) 1389 { 1390 struct tdq *tdq; 1391 1392 tdq = TDQ_SELF(); 1393 #ifdef SMP 1394 sched_setup_smp(); 1395 #else 1396 tdq_setup(tdq); 1397 #endif 1398 1399 /* Add thread0's load since it's running. */ 1400 TDQ_LOCK(tdq); 1401 thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF()); 1402 tdq_load_add(tdq, &thread0); 1403 tdq->tdq_lowpri = thread0.td_priority; 1404 TDQ_UNLOCK(tdq); 1405 } 1406 1407 /* 1408 * This routine determines time constants after stathz and hz are setup. 1409 */ 1410 /* ARGSUSED */ 1411 static void 1412 sched_initticks(void *dummy) 1413 { 1414 int incr; 1415 1416 realstathz = stathz ? stathz : hz; 1417 sched_slice = realstathz / SCHED_SLICE_DEFAULT_DIVISOR; 1418 sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR; 1419 hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) / 1420 realstathz); 1421 1422 /* 1423 * tickincr is shifted out by 10 to avoid rounding errors due to 1424 * hz not being evenly divisible by stathz on all platforms. 1425 */ 1426 incr = (hz << SCHED_TICK_SHIFT) / realstathz; 1427 /* 1428 * This does not work for values of stathz that are more than 1429 * 1 << SCHED_TICK_SHIFT * hz. In practice this does not happen. 1430 */ 1431 if (incr == 0) 1432 incr = 1; 1433 tickincr = incr; 1434 #ifdef SMP 1435 /* 1436 * Set the default balance interval now that we know 1437 * what realstathz is. 1438 */ 1439 balance_interval = realstathz; 1440 affinity = SCHED_AFFINITY_DEFAULT; 1441 #endif 1442 if (sched_idlespinthresh < 0) 1443 sched_idlespinthresh = 2 * max(10000, 6 * hz) / realstathz; 1444 } 1445 1446 1447 /* 1448 * This is the core of the interactivity algorithm. Determines a score based 1449 * on past behavior. It is the ratio of sleep time to run time scaled to 1450 * a [0, 100] integer. This is the voluntary sleep time of a process, which 1451 * differs from the cpu usage because it does not account for time spent 1452 * waiting on a run-queue. Would be prettier if we had floating point. 1453 * 1454 * When a thread's sleep time is greater than its run time the 1455 * calculation is: 1456 * 1457 * scaling factor 1458 * interactivity score = --------------------- 1459 * sleep time / run time 1460 * 1461 * 1462 * When a thread's run time is greater than its sleep time the 1463 * calculation is: 1464 * 1465 * scaling factor 1466 * interactivity score = --------------------- + scaling factor 1467 * run time / sleep time 1468 */ 1469 static int 1470 sched_interact_score(struct thread *td) 1471 { 1472 struct td_sched *ts; 1473 int div; 1474 1475 ts = td_get_sched(td); 1476 /* 1477 * The score is only needed if this is likely to be an interactive 1478 * task. Don't go through the expense of computing it if there's 1479 * no chance. 1480 */ 1481 if (sched_interact <= SCHED_INTERACT_HALF && 1482 ts->ts_runtime >= ts->ts_slptime) 1483 return (SCHED_INTERACT_HALF); 1484 1485 if (ts->ts_runtime > ts->ts_slptime) { 1486 div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF); 1487 return (SCHED_INTERACT_HALF + 1488 (SCHED_INTERACT_HALF - (ts->ts_slptime / div))); 1489 } 1490 if (ts->ts_slptime > ts->ts_runtime) { 1491 div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF); 1492 return (ts->ts_runtime / div); 1493 } 1494 /* runtime == slptime */ 1495 if (ts->ts_runtime) 1496 return (SCHED_INTERACT_HALF); 1497 1498 /* 1499 * This can happen if slptime and runtime are 0. 1500 */ 1501 return (0); 1502 1503 } 1504 1505 /* 1506 * Scale the scheduling priority according to the "interactivity" of this 1507 * process. 1508 */ 1509 static void 1510 sched_priority(struct thread *td) 1511 { 1512 int score; 1513 int pri; 1514 1515 if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE) 1516 return; 1517 /* 1518 * If the score is interactive we place the thread in the realtime 1519 * queue with a priority that is less than kernel and interrupt 1520 * priorities. These threads are not subject to nice restrictions. 1521 * 1522 * Scores greater than this are placed on the normal timeshare queue 1523 * where the priority is partially decided by the most recent cpu 1524 * utilization and the rest is decided by nice value. 1525 * 1526 * The nice value of the process has a linear effect on the calculated 1527 * score. Negative nice values make it easier for a thread to be 1528 * considered interactive. 1529 */ 1530 score = imax(0, sched_interact_score(td) + td->td_proc->p_nice); 1531 if (score < sched_interact) { 1532 pri = PRI_MIN_INTERACT; 1533 pri += ((PRI_MAX_INTERACT - PRI_MIN_INTERACT + 1) / 1534 sched_interact) * score; 1535 KASSERT(pri >= PRI_MIN_INTERACT && pri <= PRI_MAX_INTERACT, 1536 ("sched_priority: invalid interactive priority %d score %d", 1537 pri, score)); 1538 } else { 1539 pri = SCHED_PRI_MIN; 1540 if (td_get_sched(td)->ts_ticks) 1541 pri += min(SCHED_PRI_TICKS(td_get_sched(td)), 1542 SCHED_PRI_RANGE - 1); 1543 pri += SCHED_PRI_NICE(td->td_proc->p_nice); 1544 KASSERT(pri >= PRI_MIN_BATCH && pri <= PRI_MAX_BATCH, 1545 ("sched_priority: invalid priority %d: nice %d, " 1546 "ticks %d ftick %d ltick %d tick pri %d", 1547 pri, td->td_proc->p_nice, td_get_sched(td)->ts_ticks, 1548 td_get_sched(td)->ts_ftick, td_get_sched(td)->ts_ltick, 1549 SCHED_PRI_TICKS(td_get_sched(td)))); 1550 } 1551 sched_user_prio(td, pri); 1552 1553 return; 1554 } 1555 1556 /* 1557 * This routine enforces a maximum limit on the amount of scheduling history 1558 * kept. It is called after either the slptime or runtime is adjusted. This 1559 * function is ugly due to integer math. 1560 */ 1561 static void 1562 sched_interact_update(struct thread *td) 1563 { 1564 struct td_sched *ts; 1565 u_int sum; 1566 1567 ts = td_get_sched(td); 1568 sum = ts->ts_runtime + ts->ts_slptime; 1569 if (sum < SCHED_SLP_RUN_MAX) 1570 return; 1571 /* 1572 * This only happens from two places: 1573 * 1) We have added an unusual amount of run time from fork_exit. 1574 * 2) We have added an unusual amount of sleep time from sched_sleep(). 1575 */ 1576 if (sum > SCHED_SLP_RUN_MAX * 2) { 1577 if (ts->ts_runtime > ts->ts_slptime) { 1578 ts->ts_runtime = SCHED_SLP_RUN_MAX; 1579 ts->ts_slptime = 1; 1580 } else { 1581 ts->ts_slptime = SCHED_SLP_RUN_MAX; 1582 ts->ts_runtime = 1; 1583 } 1584 return; 1585 } 1586 /* 1587 * If we have exceeded by more than 1/5th then the algorithm below 1588 * will not bring us back into range. Dividing by two here forces 1589 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX] 1590 */ 1591 if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) { 1592 ts->ts_runtime /= 2; 1593 ts->ts_slptime /= 2; 1594 return; 1595 } 1596 ts->ts_runtime = (ts->ts_runtime / 5) * 4; 1597 ts->ts_slptime = (ts->ts_slptime / 5) * 4; 1598 } 1599 1600 /* 1601 * Scale back the interactivity history when a child thread is created. The 1602 * history is inherited from the parent but the thread may behave totally 1603 * differently. For example, a shell spawning a compiler process. We want 1604 * to learn that the compiler is behaving badly very quickly. 1605 */ 1606 static void 1607 sched_interact_fork(struct thread *td) 1608 { 1609 struct td_sched *ts; 1610 int ratio; 1611 int sum; 1612 1613 ts = td_get_sched(td); 1614 sum = ts->ts_runtime + ts->ts_slptime; 1615 if (sum > SCHED_SLP_RUN_FORK) { 1616 ratio = sum / SCHED_SLP_RUN_FORK; 1617 ts->ts_runtime /= ratio; 1618 ts->ts_slptime /= ratio; 1619 } 1620 } 1621 1622 /* 1623 * Called from proc0_init() to setup the scheduler fields. 1624 */ 1625 void 1626 schedinit(void) 1627 { 1628 struct td_sched *ts0; 1629 1630 /* 1631 * Set up the scheduler specific parts of thread0. 1632 */ 1633 ts0 = td_get_sched(&thread0); 1634 ts0->ts_ltick = ticks; 1635 ts0->ts_ftick = ticks; 1636 ts0->ts_slice = 0; 1637 } 1638 1639 /* 1640 * This is only somewhat accurate since given many processes of the same 1641 * priority they will switch when their slices run out, which will be 1642 * at most sched_slice stathz ticks. 1643 */ 1644 int 1645 sched_rr_interval(void) 1646 { 1647 1648 /* Convert sched_slice from stathz to hz. */ 1649 return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz)); 1650 } 1651 1652 /* 1653 * Update the percent cpu tracking information when it is requested or 1654 * the total history exceeds the maximum. We keep a sliding history of 1655 * tick counts that slowly decays. This is less precise than the 4BSD 1656 * mechanism since it happens with less regular and frequent events. 1657 */ 1658 static void 1659 sched_pctcpu_update(struct td_sched *ts, int run) 1660 { 1661 int t = ticks; 1662 1663 if (t - ts->ts_ltick >= SCHED_TICK_TARG) { 1664 ts->ts_ticks = 0; 1665 ts->ts_ftick = t - SCHED_TICK_TARG; 1666 } else if (t - ts->ts_ftick >= SCHED_TICK_MAX) { 1667 ts->ts_ticks = (ts->ts_ticks / (ts->ts_ltick - ts->ts_ftick)) * 1668 (ts->ts_ltick - (t - SCHED_TICK_TARG)); 1669 ts->ts_ftick = t - SCHED_TICK_TARG; 1670 } 1671 if (run) 1672 ts->ts_ticks += (t - ts->ts_ltick) << SCHED_TICK_SHIFT; 1673 ts->ts_ltick = t; 1674 } 1675 1676 static void 1677 sched_check_preempt(struct tdq *tdq, struct thread *td) 1678 { 1679 1680 KASSERT(TD_IS_RUNNING(td), ("thread is not running")); 1681 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 1682 KASSERT(tdq == TDQ_CPU(td->td_sched->ts_cpu), 1683 ("tdq does not contain td")); 1684 1685 if (tdq == TDQ_SELF()) { 1686 if (sched_shouldpreempt(tdq->tdq_lowpri, td->td_priority, 0)) 1687 td->td_owepreempt = 1; 1688 } else 1689 tdq_notify(tdq, tdq->tdq_lowpri); 1690 } 1691 1692 /* 1693 * Adjust the priority of a thread. Move it to the appropriate run-queue 1694 * if necessary. This is the back-end for several priority related 1695 * functions. 1696 */ 1697 static void 1698 sched_thread_priority(struct thread *td, u_char prio) 1699 { 1700 struct td_sched *ts; 1701 struct tdq *tdq; 1702 int oldpri; 1703 1704 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio", 1705 "prio:%d", td->td_priority, "new prio:%d", prio, 1706 KTR_ATTR_LINKED, sched_tdname(curthread)); 1707 SDT_PROBE3(sched, , , change__pri, td, td->td_proc, prio); 1708 if (td != curthread && prio < td->td_priority) { 1709 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread), 1710 "lend prio", "prio:%d", td->td_priority, "new prio:%d", 1711 prio, KTR_ATTR_LINKED, sched_tdname(td)); 1712 SDT_PROBE4(sched, , , lend__pri, td, td->td_proc, prio, 1713 curthread); 1714 } 1715 ts = td_get_sched(td); 1716 THREAD_LOCK_ASSERT(td, MA_OWNED); 1717 if (td->td_priority == prio) 1718 return; 1719 /* 1720 * If the priority has been elevated due to priority 1721 * propagation, we may have to move ourselves to a new 1722 * queue. This could be optimized to not re-add in some 1723 * cases. 1724 */ 1725 if (TD_ON_RUNQ(td) && prio < td->td_priority) { 1726 sched_rem(td); 1727 td->td_priority = prio; 1728 sched_add(td, SRQ_BORROWING); 1729 return; 1730 } 1731 /* 1732 * If the thread is currently running we may have to adjust the lowpri 1733 * information so other cpus are aware of our current priority. 1734 */ 1735 if (TD_IS_RUNNING(td)) { 1736 tdq = TDQ_CPU(ts->ts_cpu); 1737 oldpri = td->td_priority; 1738 td->td_priority = prio; 1739 if (prio < tdq->tdq_lowpri) 1740 tdq->tdq_lowpri = prio; 1741 else if (tdq->tdq_lowpri == oldpri) 1742 tdq_setlowpri(tdq, td); 1743 1744 if (oldpri < prio) 1745 sched_check_preempt(tdq, td); 1746 return; 1747 } 1748 td->td_priority = prio; 1749 } 1750 1751 /* 1752 * Update a thread's priority when it is lent another thread's 1753 * priority. 1754 */ 1755 void 1756 sched_lend_prio(struct thread *td, u_char prio) 1757 { 1758 1759 td->td_flags |= TDF_BORROWING; 1760 sched_thread_priority(td, prio); 1761 } 1762 1763 /* 1764 * Restore a thread's priority when priority propagation is 1765 * over. The prio argument is the minimum priority the thread 1766 * needs to have to satisfy other possible priority lending 1767 * requests. If the thread's regular priority is less 1768 * important than prio, the thread will keep a priority boost 1769 * of prio. 1770 */ 1771 void 1772 sched_unlend_prio(struct thread *td, u_char prio) 1773 { 1774 u_char base_pri; 1775 1776 if (td->td_base_pri >= PRI_MIN_TIMESHARE && 1777 td->td_base_pri <= PRI_MAX_TIMESHARE) 1778 base_pri = td->td_user_pri; 1779 else 1780 base_pri = td->td_base_pri; 1781 if (prio >= base_pri) { 1782 td->td_flags &= ~TDF_BORROWING; 1783 sched_thread_priority(td, base_pri); 1784 } else 1785 sched_lend_prio(td, prio); 1786 } 1787 1788 /* 1789 * Standard entry for setting the priority to an absolute value. 1790 */ 1791 void 1792 sched_prio(struct thread *td, u_char prio) 1793 { 1794 u_char oldprio; 1795 1796 /* First, update the base priority. */ 1797 td->td_base_pri = prio; 1798 1799 /* 1800 * If the thread is borrowing another thread's priority, don't 1801 * ever lower the priority. 1802 */ 1803 if (td->td_flags & TDF_BORROWING && td->td_priority < prio) 1804 return; 1805 1806 /* Change the real priority. */ 1807 oldprio = td->td_priority; 1808 sched_thread_priority(td, prio); 1809 1810 /* 1811 * If the thread is on a turnstile, then let the turnstile update 1812 * its state. 1813 */ 1814 if (TD_ON_LOCK(td) && oldprio != prio) 1815 turnstile_adjust(td, oldprio); 1816 } 1817 1818 /* 1819 * Set the base user priority, does not effect current running priority. 1820 */ 1821 void 1822 sched_user_prio(struct thread *td, u_char prio) 1823 { 1824 1825 td->td_base_user_pri = prio; 1826 if (td->td_lend_user_pri <= prio) 1827 return; 1828 td->td_user_pri = prio; 1829 } 1830 1831 void 1832 sched_lend_user_prio(struct thread *td, u_char prio) 1833 { 1834 1835 THREAD_LOCK_ASSERT(td, MA_OWNED); 1836 td->td_lend_user_pri = prio; 1837 td->td_user_pri = min(prio, td->td_base_user_pri); 1838 if (td->td_priority > td->td_user_pri) 1839 sched_prio(td, td->td_user_pri); 1840 else if (td->td_priority != td->td_user_pri) 1841 td->td_flags |= TDF_NEEDRESCHED; 1842 } 1843 1844 /* 1845 * Handle migration from sched_switch(). This happens only for 1846 * cpu binding. 1847 */ 1848 static struct mtx * 1849 sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags) 1850 { 1851 struct tdq *tdn; 1852 1853 tdn = TDQ_CPU(td_get_sched(td)->ts_cpu); 1854 #ifdef SMP 1855 tdq_load_rem(tdq, td); 1856 /* 1857 * Do the lock dance required to avoid LOR. We grab an extra 1858 * spinlock nesting to prevent preemption while we're 1859 * not holding either run-queue lock. 1860 */ 1861 spinlock_enter(); 1862 thread_lock_block(td); /* This releases the lock on tdq. */ 1863 1864 /* 1865 * Acquire both run-queue locks before placing the thread on the new 1866 * run-queue to avoid deadlocks created by placing a thread with a 1867 * blocked lock on the run-queue of a remote processor. The deadlock 1868 * occurs when a third processor attempts to lock the two queues in 1869 * question while the target processor is spinning with its own 1870 * run-queue lock held while waiting for the blocked lock to clear. 1871 */ 1872 tdq_lock_pair(tdn, tdq); 1873 tdq_add(tdn, td, flags); 1874 tdq_notify(tdn, td->td_priority); 1875 TDQ_UNLOCK(tdn); 1876 spinlock_exit(); 1877 #endif 1878 return (TDQ_LOCKPTR(tdn)); 1879 } 1880 1881 /* 1882 * Variadic version of thread_lock_unblock() that does not assume td_lock 1883 * is blocked. 1884 */ 1885 static inline void 1886 thread_unblock_switch(struct thread *td, struct mtx *mtx) 1887 { 1888 atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock, 1889 (uintptr_t)mtx); 1890 } 1891 1892 /* 1893 * Switch threads. This function has to handle threads coming in while 1894 * blocked for some reason, running, or idle. It also must deal with 1895 * migrating a thread from one queue to another as running threads may 1896 * be assigned elsewhere via binding. 1897 */ 1898 void 1899 sched_switch(struct thread *td, struct thread *newtd, int flags) 1900 { 1901 struct tdq *tdq; 1902 struct td_sched *ts; 1903 struct mtx *mtx; 1904 int srqflag; 1905 int cpuid, preempted; 1906 1907 THREAD_LOCK_ASSERT(td, MA_OWNED); 1908 KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument")); 1909 1910 cpuid = PCPU_GET(cpuid); 1911 tdq = TDQ_CPU(cpuid); 1912 ts = td_get_sched(td); 1913 mtx = td->td_lock; 1914 sched_pctcpu_update(ts, 1); 1915 ts->ts_rltick = ticks; 1916 td->td_lastcpu = td->td_oncpu; 1917 td->td_oncpu = NOCPU; 1918 preempted = (td->td_flags & TDF_SLICEEND) == 0 && 1919 (flags & SW_PREEMPT) != 0; 1920 td->td_flags &= ~(TDF_NEEDRESCHED | TDF_SLICEEND); 1921 td->td_owepreempt = 0; 1922 if (!TD_IS_IDLETHREAD(td)) 1923 tdq->tdq_switchcnt++; 1924 /* 1925 * The lock pointer in an idle thread should never change. Reset it 1926 * to CAN_RUN as well. 1927 */ 1928 if (TD_IS_IDLETHREAD(td)) { 1929 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1930 TD_SET_CAN_RUN(td); 1931 } else if (TD_IS_RUNNING(td)) { 1932 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1933 srqflag = preempted ? 1934 SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED : 1935 SRQ_OURSELF|SRQ_YIELDING; 1936 #ifdef SMP 1937 if (THREAD_CAN_MIGRATE(td) && !THREAD_CAN_SCHED(td, ts->ts_cpu)) 1938 ts->ts_cpu = sched_pickcpu(td, 0); 1939 #endif 1940 if (ts->ts_cpu == cpuid) 1941 tdq_runq_add(tdq, td, srqflag); 1942 else { 1943 KASSERT(THREAD_CAN_MIGRATE(td) || 1944 (ts->ts_flags & TSF_BOUND) != 0, 1945 ("Thread %p shouldn't migrate", td)); 1946 mtx = sched_switch_migrate(tdq, td, srqflag); 1947 } 1948 } else { 1949 /* This thread must be going to sleep. */ 1950 TDQ_LOCK(tdq); 1951 mtx = thread_lock_block(td); 1952 tdq_load_rem(tdq, td); 1953 } 1954 /* 1955 * We enter here with the thread blocked and assigned to the 1956 * appropriate cpu run-queue or sleep-queue and with the current 1957 * thread-queue locked. 1958 */ 1959 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED); 1960 newtd = choosethread(); 1961 /* 1962 * Call the MD code to switch contexts if necessary. 1963 */ 1964 if (td != newtd) { 1965 #ifdef HWPMC_HOOKS 1966 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1967 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT); 1968 #endif 1969 SDT_PROBE2(sched, , , off__cpu, newtd, newtd->td_proc); 1970 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object); 1971 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd; 1972 sched_pctcpu_update(td_get_sched(newtd), 0); 1973 1974 #ifdef KDTRACE_HOOKS 1975 /* 1976 * If DTrace has set the active vtime enum to anything 1977 * other than INACTIVE (0), then it should have set the 1978 * function to call. 1979 */ 1980 if (dtrace_vtime_active) 1981 (*dtrace_vtime_switch_func)(newtd); 1982 #endif 1983 1984 cpu_switch(td, newtd, mtx); 1985 /* 1986 * We may return from cpu_switch on a different cpu. However, 1987 * we always return with td_lock pointing to the current cpu's 1988 * run queue lock. 1989 */ 1990 cpuid = PCPU_GET(cpuid); 1991 tdq = TDQ_CPU(cpuid); 1992 lock_profile_obtain_lock_success( 1993 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__); 1994 1995 SDT_PROBE0(sched, , , on__cpu); 1996 #ifdef HWPMC_HOOKS 1997 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1998 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN); 1999 #endif 2000 } else { 2001 thread_unblock_switch(td, mtx); 2002 SDT_PROBE0(sched, , , remain__cpu); 2003 } 2004 /* 2005 * Assert that all went well and return. 2006 */ 2007 TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED); 2008 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2009 td->td_oncpu = cpuid; 2010 } 2011 2012 /* 2013 * Adjust thread priorities as a result of a nice request. 2014 */ 2015 void 2016 sched_nice(struct proc *p, int nice) 2017 { 2018 struct thread *td; 2019 2020 PROC_LOCK_ASSERT(p, MA_OWNED); 2021 2022 p->p_nice = nice; 2023 FOREACH_THREAD_IN_PROC(p, td) { 2024 thread_lock(td); 2025 sched_priority(td); 2026 sched_prio(td, td->td_base_user_pri); 2027 thread_unlock(td); 2028 } 2029 } 2030 2031 /* 2032 * Record the sleep time for the interactivity scorer. 2033 */ 2034 void 2035 sched_sleep(struct thread *td, int prio) 2036 { 2037 2038 THREAD_LOCK_ASSERT(td, MA_OWNED); 2039 2040 td->td_slptick = ticks; 2041 if (TD_IS_SUSPENDED(td) || prio >= PSOCK) 2042 td->td_flags |= TDF_CANSWAP; 2043 if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE) 2044 return; 2045 if (static_boost == 1 && prio) 2046 sched_prio(td, prio); 2047 else if (static_boost && td->td_priority > static_boost) 2048 sched_prio(td, static_boost); 2049 } 2050 2051 /* 2052 * Schedule a thread to resume execution and record how long it voluntarily 2053 * slept. We also update the pctcpu, interactivity, and priority. 2054 */ 2055 void 2056 sched_wakeup(struct thread *td) 2057 { 2058 struct td_sched *ts; 2059 int slptick; 2060 2061 THREAD_LOCK_ASSERT(td, MA_OWNED); 2062 ts = td_get_sched(td); 2063 td->td_flags &= ~TDF_CANSWAP; 2064 /* 2065 * If we slept for more than a tick update our interactivity and 2066 * priority. 2067 */ 2068 slptick = td->td_slptick; 2069 td->td_slptick = 0; 2070 if (slptick && slptick != ticks) { 2071 ts->ts_slptime += (ticks - slptick) << SCHED_TICK_SHIFT; 2072 sched_interact_update(td); 2073 sched_pctcpu_update(ts, 0); 2074 } 2075 /* 2076 * Reset the slice value since we slept and advanced the round-robin. 2077 */ 2078 ts->ts_slice = 0; 2079 sched_add(td, SRQ_BORING); 2080 } 2081 2082 /* 2083 * Penalize the parent for creating a new child and initialize the child's 2084 * priority. 2085 */ 2086 void 2087 sched_fork(struct thread *td, struct thread *child) 2088 { 2089 THREAD_LOCK_ASSERT(td, MA_OWNED); 2090 sched_pctcpu_update(td_get_sched(td), 1); 2091 sched_fork_thread(td, child); 2092 /* 2093 * Penalize the parent and child for forking. 2094 */ 2095 sched_interact_fork(child); 2096 sched_priority(child); 2097 td_get_sched(td)->ts_runtime += tickincr; 2098 sched_interact_update(td); 2099 sched_priority(td); 2100 } 2101 2102 /* 2103 * Fork a new thread, may be within the same process. 2104 */ 2105 void 2106 sched_fork_thread(struct thread *td, struct thread *child) 2107 { 2108 struct td_sched *ts; 2109 struct td_sched *ts2; 2110 struct tdq *tdq; 2111 2112 tdq = TDQ_SELF(); 2113 THREAD_LOCK_ASSERT(td, MA_OWNED); 2114 /* 2115 * Initialize child. 2116 */ 2117 ts = td_get_sched(td); 2118 ts2 = td_get_sched(child); 2119 child->td_oncpu = NOCPU; 2120 child->td_lastcpu = NOCPU; 2121 child->td_lock = TDQ_LOCKPTR(tdq); 2122 child->td_cpuset = cpuset_ref(td->td_cpuset); 2123 ts2->ts_cpu = ts->ts_cpu; 2124 ts2->ts_flags = 0; 2125 /* 2126 * Grab our parents cpu estimation information. 2127 */ 2128 ts2->ts_ticks = ts->ts_ticks; 2129 ts2->ts_ltick = ts->ts_ltick; 2130 ts2->ts_ftick = ts->ts_ftick; 2131 /* 2132 * Do not inherit any borrowed priority from the parent. 2133 */ 2134 child->td_priority = child->td_base_pri; 2135 /* 2136 * And update interactivity score. 2137 */ 2138 ts2->ts_slptime = ts->ts_slptime; 2139 ts2->ts_runtime = ts->ts_runtime; 2140 /* Attempt to quickly learn interactivity. */ 2141 ts2->ts_slice = tdq_slice(tdq) - sched_slice_min; 2142 #ifdef KTR 2143 bzero(ts2->ts_name, sizeof(ts2->ts_name)); 2144 #endif 2145 } 2146 2147 /* 2148 * Adjust the priority class of a thread. 2149 */ 2150 void 2151 sched_class(struct thread *td, int class) 2152 { 2153 2154 THREAD_LOCK_ASSERT(td, MA_OWNED); 2155 if (td->td_pri_class == class) 2156 return; 2157 td->td_pri_class = class; 2158 } 2159 2160 /* 2161 * Return some of the child's priority and interactivity to the parent. 2162 */ 2163 void 2164 sched_exit(struct proc *p, struct thread *child) 2165 { 2166 struct thread *td; 2167 2168 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit", 2169 "prio:%d", child->td_priority); 2170 PROC_LOCK_ASSERT(p, MA_OWNED); 2171 td = FIRST_THREAD_IN_PROC(p); 2172 sched_exit_thread(td, child); 2173 } 2174 2175 /* 2176 * Penalize another thread for the time spent on this one. This helps to 2177 * worsen the priority and interactivity of processes which schedule batch 2178 * jobs such as make. This has little effect on the make process itself but 2179 * causes new processes spawned by it to receive worse scores immediately. 2180 */ 2181 void 2182 sched_exit_thread(struct thread *td, struct thread *child) 2183 { 2184 2185 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit", 2186 "prio:%d", child->td_priority); 2187 /* 2188 * Give the child's runtime to the parent without returning the 2189 * sleep time as a penalty to the parent. This causes shells that 2190 * launch expensive things to mark their children as expensive. 2191 */ 2192 thread_lock(td); 2193 td_get_sched(td)->ts_runtime += td_get_sched(child)->ts_runtime; 2194 sched_interact_update(td); 2195 sched_priority(td); 2196 thread_unlock(td); 2197 } 2198 2199 void 2200 sched_preempt(struct thread *td) 2201 { 2202 struct tdq *tdq; 2203 2204 SDT_PROBE2(sched, , , surrender, td, td->td_proc); 2205 2206 thread_lock(td); 2207 tdq = TDQ_SELF(); 2208 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2209 tdq->tdq_ipipending = 0; 2210 if (td->td_priority > tdq->tdq_lowpri) { 2211 int flags; 2212 2213 flags = SW_INVOL | SW_PREEMPT; 2214 if (td->td_critnest > 1) 2215 td->td_owepreempt = 1; 2216 else if (TD_IS_IDLETHREAD(td)) 2217 mi_switch(flags | SWT_REMOTEWAKEIDLE, NULL); 2218 else 2219 mi_switch(flags | SWT_REMOTEPREEMPT, NULL); 2220 } 2221 thread_unlock(td); 2222 } 2223 2224 /* 2225 * Fix priorities on return to user-space. Priorities may be elevated due 2226 * to static priorities in msleep() or similar. 2227 */ 2228 void 2229 sched_userret(struct thread *td) 2230 { 2231 /* 2232 * XXX we cheat slightly on the locking here to avoid locking in 2233 * the usual case. Setting td_priority here is essentially an 2234 * incomplete workaround for not setting it properly elsewhere. 2235 * Now that some interrupt handlers are threads, not setting it 2236 * properly elsewhere can clobber it in the window between setting 2237 * it here and returning to user mode, so don't waste time setting 2238 * it perfectly here. 2239 */ 2240 KASSERT((td->td_flags & TDF_BORROWING) == 0, 2241 ("thread with borrowed priority returning to userland")); 2242 if (td->td_priority != td->td_user_pri) { 2243 thread_lock(td); 2244 td->td_priority = td->td_user_pri; 2245 td->td_base_pri = td->td_user_pri; 2246 tdq_setlowpri(TDQ_SELF(), td); 2247 thread_unlock(td); 2248 } 2249 } 2250 2251 /* 2252 * Handle a stathz tick. This is really only relevant for timeshare 2253 * threads. 2254 */ 2255 void 2256 sched_clock(struct thread *td) 2257 { 2258 struct tdq *tdq; 2259 struct td_sched *ts; 2260 2261 THREAD_LOCK_ASSERT(td, MA_OWNED); 2262 tdq = TDQ_SELF(); 2263 #ifdef SMP 2264 /* 2265 * We run the long term load balancer infrequently on the first cpu. 2266 */ 2267 if (balance_tdq == tdq) { 2268 if (balance_ticks && --balance_ticks == 0) 2269 sched_balance(); 2270 } 2271 #endif 2272 /* 2273 * Save the old switch count so we have a record of the last ticks 2274 * activity. Initialize the new switch count based on our load. 2275 * If there is some activity seed it to reflect that. 2276 */ 2277 tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt; 2278 tdq->tdq_switchcnt = tdq->tdq_load; 2279 /* 2280 * Advance the insert index once for each tick to ensure that all 2281 * threads get a chance to run. 2282 */ 2283 if (tdq->tdq_idx == tdq->tdq_ridx) { 2284 tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS; 2285 if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx])) 2286 tdq->tdq_ridx = tdq->tdq_idx; 2287 } 2288 ts = td_get_sched(td); 2289 sched_pctcpu_update(ts, 1); 2290 if (td->td_pri_class & PRI_FIFO_BIT) 2291 return; 2292 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) { 2293 /* 2294 * We used a tick; charge it to the thread so 2295 * that we can compute our interactivity. 2296 */ 2297 td_get_sched(td)->ts_runtime += tickincr; 2298 sched_interact_update(td); 2299 sched_priority(td); 2300 } 2301 2302 /* 2303 * Force a context switch if the current thread has used up a full 2304 * time slice (default is 100ms). 2305 */ 2306 if (!TD_IS_IDLETHREAD(td) && ++ts->ts_slice >= tdq_slice(tdq)) { 2307 ts->ts_slice = 0; 2308 td->td_flags |= TDF_NEEDRESCHED | TDF_SLICEEND; 2309 } 2310 } 2311 2312 u_int 2313 sched_estcpu(struct thread *td __unused) 2314 { 2315 2316 return (0); 2317 } 2318 2319 /* 2320 * Return whether the current CPU has runnable tasks. Used for in-kernel 2321 * cooperative idle threads. 2322 */ 2323 int 2324 sched_runnable(void) 2325 { 2326 struct tdq *tdq; 2327 int load; 2328 2329 load = 1; 2330 2331 tdq = TDQ_SELF(); 2332 if ((curthread->td_flags & TDF_IDLETD) != 0) { 2333 if (tdq->tdq_load > 0) 2334 goto out; 2335 } else 2336 if (tdq->tdq_load - 1 > 0) 2337 goto out; 2338 load = 0; 2339 out: 2340 return (load); 2341 } 2342 2343 /* 2344 * Choose the highest priority thread to run. The thread is removed from 2345 * the run-queue while running however the load remains. For SMP we set 2346 * the tdq in the global idle bitmask if it idles here. 2347 */ 2348 struct thread * 2349 sched_choose(void) 2350 { 2351 struct thread *td; 2352 struct tdq *tdq; 2353 2354 tdq = TDQ_SELF(); 2355 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2356 td = tdq_choose(tdq); 2357 if (td) { 2358 tdq_runq_rem(tdq, td); 2359 tdq->tdq_lowpri = td->td_priority; 2360 return (td); 2361 } 2362 tdq->tdq_lowpri = PRI_MAX_IDLE; 2363 return (PCPU_GET(idlethread)); 2364 } 2365 2366 /* 2367 * Set owepreempt if necessary. Preemption never happens directly in ULE, 2368 * we always request it once we exit a critical section. 2369 */ 2370 static inline void 2371 sched_setpreempt(struct thread *td) 2372 { 2373 struct thread *ctd; 2374 int cpri; 2375 int pri; 2376 2377 THREAD_LOCK_ASSERT(curthread, MA_OWNED); 2378 2379 ctd = curthread; 2380 pri = td->td_priority; 2381 cpri = ctd->td_priority; 2382 if (pri < cpri) 2383 ctd->td_flags |= TDF_NEEDRESCHED; 2384 if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd)) 2385 return; 2386 if (!sched_shouldpreempt(pri, cpri, 0)) 2387 return; 2388 ctd->td_owepreempt = 1; 2389 } 2390 2391 /* 2392 * Add a thread to a thread queue. Select the appropriate runq and add the 2393 * thread to it. This is the internal function called when the tdq is 2394 * predetermined. 2395 */ 2396 void 2397 tdq_add(struct tdq *tdq, struct thread *td, int flags) 2398 { 2399 2400 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2401 KASSERT((td->td_inhibitors == 0), 2402 ("sched_add: trying to run inhibited thread")); 2403 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)), 2404 ("sched_add: bad thread state")); 2405 KASSERT(td->td_flags & TDF_INMEM, 2406 ("sched_add: thread swapped out")); 2407 2408 if (td->td_priority < tdq->tdq_lowpri) 2409 tdq->tdq_lowpri = td->td_priority; 2410 tdq_runq_add(tdq, td, flags); 2411 tdq_load_add(tdq, td); 2412 } 2413 2414 /* 2415 * Select the target thread queue and add a thread to it. Request 2416 * preemption or IPI a remote processor if required. 2417 */ 2418 void 2419 sched_add(struct thread *td, int flags) 2420 { 2421 struct tdq *tdq; 2422 #ifdef SMP 2423 int cpu; 2424 #endif 2425 2426 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add", 2427 "prio:%d", td->td_priority, KTR_ATTR_LINKED, 2428 sched_tdname(curthread)); 2429 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup", 2430 KTR_ATTR_LINKED, sched_tdname(td)); 2431 SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL, 2432 flags & SRQ_PREEMPTED); 2433 THREAD_LOCK_ASSERT(td, MA_OWNED); 2434 /* 2435 * Recalculate the priority before we select the target cpu or 2436 * run-queue. 2437 */ 2438 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) 2439 sched_priority(td); 2440 #ifdef SMP 2441 /* 2442 * Pick the destination cpu and if it isn't ours transfer to the 2443 * target cpu. 2444 */ 2445 cpu = sched_pickcpu(td, flags); 2446 tdq = sched_setcpu(td, cpu, flags); 2447 tdq_add(tdq, td, flags); 2448 if (cpu != PCPU_GET(cpuid)) { 2449 tdq_notify(tdq, td->td_priority); 2450 return; 2451 } 2452 #else 2453 tdq = TDQ_SELF(); 2454 TDQ_LOCK(tdq); 2455 /* 2456 * Now that the thread is moving to the run-queue, set the lock 2457 * to the scheduler's lock. 2458 */ 2459 thread_lock_set(td, TDQ_LOCKPTR(tdq)); 2460 tdq_add(tdq, td, flags); 2461 #endif 2462 if (!(flags & SRQ_YIELDING)) 2463 sched_setpreempt(td); 2464 } 2465 2466 /* 2467 * Remove a thread from a run-queue without running it. This is used 2468 * when we're stealing a thread from a remote queue. Otherwise all threads 2469 * exit by calling sched_exit_thread() and sched_throw() themselves. 2470 */ 2471 void 2472 sched_rem(struct thread *td) 2473 { 2474 struct tdq *tdq; 2475 2476 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem", 2477 "prio:%d", td->td_priority); 2478 SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL); 2479 tdq = TDQ_CPU(td_get_sched(td)->ts_cpu); 2480 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2481 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2482 KASSERT(TD_ON_RUNQ(td), 2483 ("sched_rem: thread not on run queue")); 2484 tdq_runq_rem(tdq, td); 2485 tdq_load_rem(tdq, td); 2486 TD_SET_CAN_RUN(td); 2487 if (td->td_priority == tdq->tdq_lowpri) 2488 tdq_setlowpri(tdq, NULL); 2489 } 2490 2491 /* 2492 * Fetch cpu utilization information. Updates on demand. 2493 */ 2494 fixpt_t 2495 sched_pctcpu(struct thread *td) 2496 { 2497 fixpt_t pctcpu; 2498 struct td_sched *ts; 2499 2500 pctcpu = 0; 2501 ts = td_get_sched(td); 2502 2503 THREAD_LOCK_ASSERT(td, MA_OWNED); 2504 sched_pctcpu_update(ts, TD_IS_RUNNING(td)); 2505 if (ts->ts_ticks) { 2506 int rtick; 2507 2508 /* How many rtick per second ? */ 2509 rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz); 2510 pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT; 2511 } 2512 2513 return (pctcpu); 2514 } 2515 2516 /* 2517 * Enforce affinity settings for a thread. Called after adjustments to 2518 * cpumask. 2519 */ 2520 void 2521 sched_affinity(struct thread *td) 2522 { 2523 #ifdef SMP 2524 struct td_sched *ts; 2525 2526 THREAD_LOCK_ASSERT(td, MA_OWNED); 2527 ts = td_get_sched(td); 2528 if (THREAD_CAN_SCHED(td, ts->ts_cpu)) 2529 return; 2530 if (TD_ON_RUNQ(td)) { 2531 sched_rem(td); 2532 sched_add(td, SRQ_BORING); 2533 return; 2534 } 2535 if (!TD_IS_RUNNING(td)) 2536 return; 2537 /* 2538 * Force a switch before returning to userspace. If the 2539 * target thread is not running locally send an ipi to force 2540 * the issue. 2541 */ 2542 td->td_flags |= TDF_NEEDRESCHED; 2543 if (td != curthread) 2544 ipi_cpu(ts->ts_cpu, IPI_PREEMPT); 2545 #endif 2546 } 2547 2548 /* 2549 * Bind a thread to a target cpu. 2550 */ 2551 void 2552 sched_bind(struct thread *td, int cpu) 2553 { 2554 struct td_sched *ts; 2555 2556 THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED); 2557 KASSERT(td == curthread, ("sched_bind: can only bind curthread")); 2558 ts = td_get_sched(td); 2559 if (ts->ts_flags & TSF_BOUND) 2560 sched_unbind(td); 2561 KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td)); 2562 ts->ts_flags |= TSF_BOUND; 2563 sched_pin(); 2564 if (PCPU_GET(cpuid) == cpu) 2565 return; 2566 ts->ts_cpu = cpu; 2567 /* When we return from mi_switch we'll be on the correct cpu. */ 2568 mi_switch(SW_VOL, NULL); 2569 } 2570 2571 /* 2572 * Release a bound thread. 2573 */ 2574 void 2575 sched_unbind(struct thread *td) 2576 { 2577 struct td_sched *ts; 2578 2579 THREAD_LOCK_ASSERT(td, MA_OWNED); 2580 KASSERT(td == curthread, ("sched_unbind: can only bind curthread")); 2581 ts = td_get_sched(td); 2582 if ((ts->ts_flags & TSF_BOUND) == 0) 2583 return; 2584 ts->ts_flags &= ~TSF_BOUND; 2585 sched_unpin(); 2586 } 2587 2588 int 2589 sched_is_bound(struct thread *td) 2590 { 2591 THREAD_LOCK_ASSERT(td, MA_OWNED); 2592 return (td_get_sched(td)->ts_flags & TSF_BOUND); 2593 } 2594 2595 /* 2596 * Basic yield call. 2597 */ 2598 void 2599 sched_relinquish(struct thread *td) 2600 { 2601 thread_lock(td); 2602 mi_switch(SW_VOL | SWT_RELINQUISH, NULL); 2603 thread_unlock(td); 2604 } 2605 2606 /* 2607 * Return the total system load. 2608 */ 2609 int 2610 sched_load(void) 2611 { 2612 #ifdef SMP 2613 int total; 2614 int i; 2615 2616 total = 0; 2617 CPU_FOREACH(i) 2618 total += TDQ_CPU(i)->tdq_sysload; 2619 return (total); 2620 #else 2621 return (TDQ_SELF()->tdq_sysload); 2622 #endif 2623 } 2624 2625 int 2626 sched_sizeof_proc(void) 2627 { 2628 return (sizeof(struct proc)); 2629 } 2630 2631 int 2632 sched_sizeof_thread(void) 2633 { 2634 return (sizeof(struct thread) + sizeof(struct td_sched)); 2635 } 2636 2637 #ifdef SMP 2638 #define TDQ_IDLESPIN(tdq) \ 2639 ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0) 2640 #else 2641 #define TDQ_IDLESPIN(tdq) 1 2642 #endif 2643 2644 /* 2645 * The actual idle process. 2646 */ 2647 void 2648 sched_idletd(void *dummy) 2649 { 2650 struct thread *td; 2651 struct tdq *tdq; 2652 int oldswitchcnt, switchcnt; 2653 int i; 2654 2655 mtx_assert(&Giant, MA_NOTOWNED); 2656 td = curthread; 2657 tdq = TDQ_SELF(); 2658 THREAD_NO_SLEEPING(); 2659 oldswitchcnt = -1; 2660 for (;;) { 2661 if (tdq->tdq_load) { 2662 thread_lock(td); 2663 mi_switch(SW_VOL | SWT_IDLE, NULL); 2664 thread_unlock(td); 2665 } 2666 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt; 2667 #ifdef SMP 2668 if (switchcnt != oldswitchcnt) { 2669 oldswitchcnt = switchcnt; 2670 if (tdq_idled(tdq) == 0) 2671 continue; 2672 } 2673 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt; 2674 #else 2675 oldswitchcnt = switchcnt; 2676 #endif 2677 /* 2678 * If we're switching very frequently, spin while checking 2679 * for load rather than entering a low power state that 2680 * may require an IPI. However, don't do any busy 2681 * loops while on SMT machines as this simply steals 2682 * cycles from cores doing useful work. 2683 */ 2684 if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) { 2685 for (i = 0; i < sched_idlespins; i++) { 2686 if (tdq->tdq_load) 2687 break; 2688 cpu_spinwait(); 2689 } 2690 } 2691 2692 /* If there was context switch during spin, restart it. */ 2693 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt; 2694 if (tdq->tdq_load != 0 || switchcnt != oldswitchcnt) 2695 continue; 2696 2697 /* Run main MD idle handler. */ 2698 tdq->tdq_cpu_idle = 1; 2699 /* 2700 * Make sure that tdq_cpu_idle update is globally visible 2701 * before cpu_idle() read tdq_load. The order is important 2702 * to avoid race with tdq_notify. 2703 */ 2704 atomic_thread_fence_seq_cst(); 2705 cpu_idle(switchcnt * 4 > sched_idlespinthresh); 2706 tdq->tdq_cpu_idle = 0; 2707 2708 /* 2709 * Account thread-less hardware interrupts and 2710 * other wakeup reasons equal to context switches. 2711 */ 2712 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt; 2713 if (switchcnt != oldswitchcnt) 2714 continue; 2715 tdq->tdq_switchcnt++; 2716 oldswitchcnt++; 2717 } 2718 } 2719 2720 /* 2721 * A CPU is entering for the first time or a thread is exiting. 2722 */ 2723 void 2724 sched_throw(struct thread *td) 2725 { 2726 struct thread *newtd; 2727 struct tdq *tdq; 2728 2729 tdq = TDQ_SELF(); 2730 if (td == NULL) { 2731 /* Correct spinlock nesting and acquire the correct lock. */ 2732 TDQ_LOCK(tdq); 2733 spinlock_exit(); 2734 PCPU_SET(switchtime, cpu_ticks()); 2735 PCPU_SET(switchticks, ticks); 2736 } else { 2737 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2738 tdq_load_rem(tdq, td); 2739 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object); 2740 td->td_lastcpu = td->td_oncpu; 2741 td->td_oncpu = NOCPU; 2742 } 2743 KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count")); 2744 newtd = choosethread(); 2745 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd; 2746 cpu_throw(td, newtd); /* doesn't return */ 2747 } 2748 2749 /* 2750 * This is called from fork_exit(). Just acquire the correct locks and 2751 * let fork do the rest of the work. 2752 */ 2753 void 2754 sched_fork_exit(struct thread *td) 2755 { 2756 struct tdq *tdq; 2757 int cpuid; 2758 2759 /* 2760 * Finish setting up thread glue so that it begins execution in a 2761 * non-nested critical section with the scheduler lock held. 2762 */ 2763 cpuid = PCPU_GET(cpuid); 2764 tdq = TDQ_CPU(cpuid); 2765 if (TD_IS_IDLETHREAD(td)) 2766 td->td_lock = TDQ_LOCKPTR(tdq); 2767 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2768 td->td_oncpu = cpuid; 2769 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED); 2770 lock_profile_obtain_lock_success( 2771 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__); 2772 } 2773 2774 /* 2775 * Create on first use to catch odd startup conditons. 2776 */ 2777 char * 2778 sched_tdname(struct thread *td) 2779 { 2780 #ifdef KTR 2781 struct td_sched *ts; 2782 2783 ts = td_get_sched(td); 2784 if (ts->ts_name[0] == '\0') 2785 snprintf(ts->ts_name, sizeof(ts->ts_name), 2786 "%s tid %d", td->td_name, td->td_tid); 2787 return (ts->ts_name); 2788 #else 2789 return (td->td_name); 2790 #endif 2791 } 2792 2793 #ifdef KTR 2794 void 2795 sched_clear_tdname(struct thread *td) 2796 { 2797 struct td_sched *ts; 2798 2799 ts = td_get_sched(td); 2800 ts->ts_name[0] = '\0'; 2801 } 2802 #endif 2803 2804 #ifdef SMP 2805 2806 /* 2807 * Build the CPU topology dump string. Is recursively called to collect 2808 * the topology tree. 2809 */ 2810 static int 2811 sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg, 2812 int indent) 2813 { 2814 char cpusetbuf[CPUSETBUFSIZ]; 2815 int i, first; 2816 2817 sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent, 2818 "", 1 + indent / 2, cg->cg_level); 2819 sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"%s\">", indent, "", 2820 cg->cg_count, cpusetobj_strprint(cpusetbuf, &cg->cg_mask)); 2821 first = TRUE; 2822 for (i = 0; i < MAXCPU; i++) { 2823 if (CPU_ISSET(i, &cg->cg_mask)) { 2824 if (!first) 2825 sbuf_printf(sb, ", "); 2826 else 2827 first = FALSE; 2828 sbuf_printf(sb, "%d", i); 2829 } 2830 } 2831 sbuf_printf(sb, "</cpu>\n"); 2832 2833 if (cg->cg_flags != 0) { 2834 sbuf_printf(sb, "%*s <flags>", indent, ""); 2835 if ((cg->cg_flags & CG_FLAG_HTT) != 0) 2836 sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>"); 2837 if ((cg->cg_flags & CG_FLAG_THREAD) != 0) 2838 sbuf_printf(sb, "<flag name=\"THREAD\">THREAD group</flag>"); 2839 if ((cg->cg_flags & CG_FLAG_SMT) != 0) 2840 sbuf_printf(sb, "<flag name=\"SMT\">SMT group</flag>"); 2841 sbuf_printf(sb, "</flags>\n"); 2842 } 2843 2844 if (cg->cg_children > 0) { 2845 sbuf_printf(sb, "%*s <children>\n", indent, ""); 2846 for (i = 0; i < cg->cg_children; i++) 2847 sysctl_kern_sched_topology_spec_internal(sb, 2848 &cg->cg_child[i], indent+2); 2849 sbuf_printf(sb, "%*s </children>\n", indent, ""); 2850 } 2851 sbuf_printf(sb, "%*s</group>\n", indent, ""); 2852 return (0); 2853 } 2854 2855 /* 2856 * Sysctl handler for retrieving topology dump. It's a wrapper for 2857 * the recursive sysctl_kern_smp_topology_spec_internal(). 2858 */ 2859 static int 2860 sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS) 2861 { 2862 struct sbuf *topo; 2863 int err; 2864 2865 KASSERT(cpu_top != NULL, ("cpu_top isn't initialized")); 2866 2867 topo = sbuf_new_for_sysctl(NULL, NULL, 512, req); 2868 if (topo == NULL) 2869 return (ENOMEM); 2870 2871 sbuf_printf(topo, "<groups>\n"); 2872 err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1); 2873 sbuf_printf(topo, "</groups>\n"); 2874 2875 if (err == 0) { 2876 err = sbuf_finish(topo); 2877 } 2878 sbuf_delete(topo); 2879 return (err); 2880 } 2881 2882 #endif 2883 2884 static int 2885 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS) 2886 { 2887 int error, new_val, period; 2888 2889 period = 1000000 / realstathz; 2890 new_val = period * sched_slice; 2891 error = sysctl_handle_int(oidp, &new_val, 0, req); 2892 if (error != 0 || req->newptr == NULL) 2893 return (error); 2894 if (new_val <= 0) 2895 return (EINVAL); 2896 sched_slice = imax(1, (new_val + period / 2) / period); 2897 sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR; 2898 hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) / 2899 realstathz); 2900 return (0); 2901 } 2902 2903 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler"); 2904 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0, 2905 "Scheduler name"); 2906 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW, 2907 NULL, 0, sysctl_kern_quantum, "I", 2908 "Quantum for timeshare threads in microseconds"); 2909 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0, 2910 "Quantum for timeshare threads in stathz ticks"); 2911 SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0, 2912 "Interactivity score threshold"); 2913 SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW, 2914 &preempt_thresh, 0, 2915 "Maximal (lowest) priority for preemption"); 2916 SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost, 0, 2917 "Assign static kernel priorities to sleeping threads"); 2918 SYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins, 0, 2919 "Number of times idle thread will spin waiting for new work"); 2920 SYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW, 2921 &sched_idlespinthresh, 0, 2922 "Threshold before we will permit idle thread spinning"); 2923 #ifdef SMP 2924 SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0, 2925 "Number of hz ticks to keep thread affinity for"); 2926 SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0, 2927 "Enables the long-term load balancer"); 2928 SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW, 2929 &balance_interval, 0, 2930 "Average period in stathz ticks to run the long-term balancer"); 2931 SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0, 2932 "Attempts to steal work from other cores before idling"); 2933 SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0, 2934 "Minimum load on remote CPU before we'll steal"); 2935 SYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING | 2936 CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A", 2937 "XML dump of detected CPU topology"); 2938 #endif 2939 2940 /* ps compat. All cpu percentages from ULE are weighted. */ 2941 static int ccpu = 0; 2942 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 2943