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_kdtrace.h" 43 #include "opt_sched.h" 44 45 #include <sys/param.h> 46 #include <sys/systm.h> 47 #include <sys/kdb.h> 48 #include <sys/kernel.h> 49 #include <sys/ktr.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/smp.h> 57 #include <sys/sx.h> 58 #include <sys/sysctl.h> 59 #include <sys/sysproto.h> 60 #include <sys/turnstile.h> 61 #include <sys/umtx.h> 62 #include <sys/vmmeter.h> 63 #include <sys/cpuset.h> 64 #include <sys/sbuf.h> 65 66 #ifdef HWPMC_HOOKS 67 #include <sys/pmckern.h> 68 #endif 69 70 #ifdef KDTRACE_HOOKS 71 #include <sys/dtrace_bsd.h> 72 int dtrace_vtime_active; 73 dtrace_vtime_switch_func_t dtrace_vtime_switch_func; 74 #endif 75 76 #include <machine/cpu.h> 77 #include <machine/smp.h> 78 79 #if defined(__sparc64__) 80 #error "This architecture is not currently compatible with ULE" 81 #endif 82 83 #define KTR_ULE 0 84 85 #define TS_NAME_LEN (MAXCOMLEN + sizeof(" td ") + sizeof(__XSTRING(UINT_MAX))) 86 #define TDQ_NAME_LEN (sizeof("sched lock ") + sizeof(__XSTRING(MAXCPU))) 87 #define TDQ_LOADNAME_LEN (PCPU_NAME_LEN + sizeof(" load")) 88 89 /* 90 * Thread scheduler specific section. All fields are protected 91 * by the thread lock. 92 */ 93 struct td_sched { 94 struct runq *ts_runq; /* Run-queue we're queued on. */ 95 short ts_flags; /* TSF_* flags. */ 96 u_char ts_cpu; /* CPU that we have affinity for. */ 97 int ts_rltick; /* Real last tick, for affinity. */ 98 int ts_slice; /* Ticks of slice remaining. */ 99 u_int ts_slptime; /* Number of ticks we vol. slept */ 100 u_int ts_runtime; /* Number of ticks we were running */ 101 int ts_ltick; /* Last tick that we were running on */ 102 int ts_incrtick; /* Last tick that we incremented on */ 103 int ts_ftick; /* First tick that we were running on */ 104 int ts_ticks; /* Tick count */ 105 #ifdef KTR 106 char ts_name[TS_NAME_LEN]; 107 #endif 108 }; 109 /* flags kept in ts_flags */ 110 #define TSF_BOUND 0x0001 /* Thread can not migrate. */ 111 #define TSF_XFERABLE 0x0002 /* Thread was added as transferable. */ 112 113 static struct td_sched td_sched0; 114 115 #define THREAD_CAN_MIGRATE(td) ((td)->td_pinned == 0) 116 #define THREAD_CAN_SCHED(td, cpu) \ 117 CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask) 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) 128 129 #define PRI_MIN_INTERACT PRI_MIN_TIMESHARE 130 #define PRI_MAX_INTERACT (PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE - 1) 131 #define PRI_MIN_BATCH (PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE) 132 #define PRI_MAX_BATCH PRI_MAX_TIMESHARE 133 134 /* 135 * Cpu percentage computation macros and defines. 136 * 137 * SCHED_TICK_SECS: Number of seconds to average the cpu usage across. 138 * SCHED_TICK_TARG: Number of hz ticks to average the cpu usage across. 139 * SCHED_TICK_MAX: Maximum number of ticks before scaling back. 140 * SCHED_TICK_SHIFT: Shift factor to avoid rounding away results. 141 * SCHED_TICK_HZ: Compute the number of hz ticks for a given ticks count. 142 * SCHED_TICK_TOTAL: Gives the amount of time we've been recording ticks. 143 */ 144 #define SCHED_TICK_SECS 10 145 #define SCHED_TICK_TARG (hz * SCHED_TICK_SECS) 146 #define SCHED_TICK_MAX (SCHED_TICK_TARG + hz) 147 #define SCHED_TICK_SHIFT 10 148 #define SCHED_TICK_HZ(ts) ((ts)->ts_ticks >> SCHED_TICK_SHIFT) 149 #define SCHED_TICK_TOTAL(ts) (max((ts)->ts_ltick - (ts)->ts_ftick, hz)) 150 151 /* 152 * These macros determine priorities for non-interactive threads. They are 153 * assigned a priority based on their recent cpu utilization as expressed 154 * by the ratio of ticks to the tick total. NHALF priorities at the start 155 * and end of the MIN to MAX timeshare range are only reachable with negative 156 * or positive nice respectively. 157 * 158 * PRI_RANGE: Priority range for utilization dependent priorities. 159 * PRI_NRESV: Number of nice values. 160 * PRI_TICKS: Compute a priority in PRI_RANGE from the ticks count and total. 161 * PRI_NICE: Determines the part of the priority inherited from nice. 162 */ 163 #define SCHED_PRI_NRESV (PRIO_MAX - PRIO_MIN) 164 #define SCHED_PRI_NHALF (SCHED_PRI_NRESV / 2) 165 #define SCHED_PRI_MIN (PRI_MIN_BATCH + SCHED_PRI_NHALF) 166 #define SCHED_PRI_MAX (PRI_MAX_BATCH - SCHED_PRI_NHALF) 167 #define SCHED_PRI_RANGE (SCHED_PRI_MAX - SCHED_PRI_MIN + 1) 168 #define SCHED_PRI_TICKS(ts) \ 169 (SCHED_TICK_HZ((ts)) / \ 170 (roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE)) 171 #define SCHED_PRI_NICE(nice) (nice) 172 173 /* 174 * These determine the interactivity of a process. Interactivity differs from 175 * cpu utilization in that it expresses the voluntary time slept vs time ran 176 * while cpu utilization includes all time not running. This more accurately 177 * models the intent of the thread. 178 * 179 * SLP_RUN_MAX: Maximum amount of sleep time + run time we'll accumulate 180 * before throttling back. 181 * SLP_RUN_FORK: Maximum slp+run time to inherit at fork time. 182 * INTERACT_MAX: Maximum interactivity value. Smaller is better. 183 * INTERACT_THRESH: Threshold for placement on the current runq. 184 */ 185 #define SCHED_SLP_RUN_MAX ((hz * 5) << SCHED_TICK_SHIFT) 186 #define SCHED_SLP_RUN_FORK ((hz / 2) << SCHED_TICK_SHIFT) 187 #define SCHED_INTERACT_MAX (100) 188 #define SCHED_INTERACT_HALF (SCHED_INTERACT_MAX / 2) 189 #define SCHED_INTERACT_THRESH (30) 190 191 /* 192 * tickincr: Converts a stathz tick into a hz domain scaled by 193 * the shift factor. Without the shift the error rate 194 * due to rounding would be unacceptably high. 195 * realstathz: stathz is sometimes 0 and run off of hz. 196 * sched_slice: Runtime of each thread before rescheduling. 197 * preempt_thresh: Priority threshold for preemption and remote IPIs. 198 */ 199 static int sched_interact = SCHED_INTERACT_THRESH; 200 static int realstathz; 201 static int tickincr; 202 static int sched_slice = 1; 203 #ifdef PREEMPTION 204 #ifdef FULL_PREEMPTION 205 static int preempt_thresh = PRI_MAX_IDLE; 206 #else 207 static int preempt_thresh = PRI_MIN_KERN; 208 #endif 209 #else 210 static int preempt_thresh = 0; 211 #endif 212 static int static_boost = PRI_MIN_BATCH; 213 static int sched_idlespins = 10000; 214 static int sched_idlespinthresh = 16; 215 216 /* 217 * tdq - per processor runqs and statistics. All fields are protected by the 218 * tdq_lock. The load and lowpri may be accessed without to avoid excess 219 * locking in sched_pickcpu(); 220 */ 221 struct tdq { 222 /* Ordered to improve efficiency of cpu_search() and switch(). */ 223 struct mtx tdq_lock; /* run queue lock. */ 224 struct cpu_group *tdq_cg; /* Pointer to cpu topology. */ 225 volatile int tdq_load; /* Aggregate load. */ 226 volatile int tdq_cpu_idle; /* cpu_idle() is active. */ 227 int tdq_sysload; /* For loadavg, !ITHD load. */ 228 int tdq_transferable; /* Transferable thread count. */ 229 short tdq_switchcnt; /* Switches this tick. */ 230 short tdq_oldswitchcnt; /* Switches last tick. */ 231 u_char tdq_lowpri; /* Lowest priority thread. */ 232 u_char tdq_ipipending; /* IPI pending. */ 233 u_char tdq_idx; /* Current insert index. */ 234 u_char tdq_ridx; /* Current removal index. */ 235 struct runq tdq_realtime; /* real-time run queue. */ 236 struct runq tdq_timeshare; /* timeshare run queue. */ 237 struct runq tdq_idle; /* Queue of IDLE threads. */ 238 char tdq_name[TDQ_NAME_LEN]; 239 #ifdef KTR 240 char tdq_loadname[TDQ_LOADNAME_LEN]; 241 #endif 242 } __aligned(64); 243 244 /* Idle thread states and config. */ 245 #define TDQ_RUNNING 1 246 #define TDQ_IDLE 2 247 248 #ifdef SMP 249 struct cpu_group *cpu_top; /* CPU topology */ 250 251 #define SCHED_AFFINITY_DEFAULT (max(1, hz / 1000)) 252 #define SCHED_AFFINITY(ts, t) ((ts)->ts_rltick > ticks - ((t) * affinity)) 253 254 /* 255 * Run-time tunables. 256 */ 257 static int rebalance = 1; 258 static int balance_interval = 128; /* Default set in sched_initticks(). */ 259 static int affinity; 260 static int steal_htt = 1; 261 static int steal_idle = 1; 262 static int steal_thresh = 2; 263 264 /* 265 * One thread queue per processor. 266 */ 267 static struct tdq tdq_cpu[MAXCPU]; 268 static struct tdq *balance_tdq; 269 static int balance_ticks; 270 271 #define TDQ_SELF() (&tdq_cpu[PCPU_GET(cpuid)]) 272 #define TDQ_CPU(x) (&tdq_cpu[(x)]) 273 #define TDQ_ID(x) ((int)((x) - tdq_cpu)) 274 #else /* !SMP */ 275 static struct tdq tdq_cpu; 276 277 #define TDQ_ID(x) (0) 278 #define TDQ_SELF() (&tdq_cpu) 279 #define TDQ_CPU(x) (&tdq_cpu) 280 #endif 281 282 #define TDQ_LOCK_ASSERT(t, type) mtx_assert(TDQ_LOCKPTR((t)), (type)) 283 #define TDQ_LOCK(t) mtx_lock_spin(TDQ_LOCKPTR((t))) 284 #define TDQ_LOCK_FLAGS(t, f) mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f)) 285 #define TDQ_UNLOCK(t) mtx_unlock_spin(TDQ_LOCKPTR((t))) 286 #define TDQ_LOCKPTR(t) (&(t)->tdq_lock) 287 288 static void sched_priority(struct thread *); 289 static void sched_thread_priority(struct thread *, u_char); 290 static int sched_interact_score(struct thread *); 291 static void sched_interact_update(struct thread *); 292 static void sched_interact_fork(struct thread *); 293 static void sched_pctcpu_update(struct td_sched *); 294 295 /* Operations on per processor queues */ 296 static struct thread *tdq_choose(struct tdq *); 297 static void tdq_setup(struct tdq *); 298 static void tdq_load_add(struct tdq *, struct thread *); 299 static void tdq_load_rem(struct tdq *, struct thread *); 300 static __inline void tdq_runq_add(struct tdq *, struct thread *, int); 301 static __inline void tdq_runq_rem(struct tdq *, struct thread *); 302 static inline int sched_shouldpreempt(int, int, int); 303 void tdq_print(int cpu); 304 static void runq_print(struct runq *rq); 305 static void tdq_add(struct tdq *, struct thread *, int); 306 #ifdef SMP 307 static int tdq_move(struct tdq *, struct tdq *); 308 static int tdq_idled(struct tdq *); 309 static void tdq_notify(struct tdq *, struct thread *); 310 static struct thread *tdq_steal(struct tdq *, int); 311 static struct thread *runq_steal(struct runq *, int); 312 static int sched_pickcpu(struct thread *, int); 313 static void sched_balance(void); 314 static int sched_balance_pair(struct tdq *, struct tdq *); 315 static inline struct tdq *sched_setcpu(struct thread *, int, int); 316 static inline void thread_unblock_switch(struct thread *, struct mtx *); 317 static struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int); 318 static int sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS); 319 static int sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, 320 struct cpu_group *cg, int indent); 321 #endif 322 323 static void sched_setup(void *dummy); 324 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL); 325 326 static void sched_initticks(void *dummy); 327 SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks, 328 NULL); 329 330 /* 331 * Print the threads waiting on a run-queue. 332 */ 333 static void 334 runq_print(struct runq *rq) 335 { 336 struct rqhead *rqh; 337 struct thread *td; 338 int pri; 339 int j; 340 int i; 341 342 for (i = 0; i < RQB_LEN; i++) { 343 printf("\t\trunq bits %d 0x%zx\n", 344 i, rq->rq_status.rqb_bits[i]); 345 for (j = 0; j < RQB_BPW; j++) 346 if (rq->rq_status.rqb_bits[i] & (1ul << j)) { 347 pri = j + (i << RQB_L2BPW); 348 rqh = &rq->rq_queues[pri]; 349 TAILQ_FOREACH(td, rqh, td_runq) { 350 printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n", 351 td, td->td_name, td->td_priority, 352 td->td_rqindex, pri); 353 } 354 } 355 } 356 } 357 358 /* 359 * Print the status of a per-cpu thread queue. Should be a ddb show cmd. 360 */ 361 void 362 tdq_print(int cpu) 363 { 364 struct tdq *tdq; 365 366 tdq = TDQ_CPU(cpu); 367 368 printf("tdq %d:\n", TDQ_ID(tdq)); 369 printf("\tlock %p\n", TDQ_LOCKPTR(tdq)); 370 printf("\tLock name: %s\n", tdq->tdq_name); 371 printf("\tload: %d\n", tdq->tdq_load); 372 printf("\tswitch cnt: %d\n", tdq->tdq_switchcnt); 373 printf("\told switch cnt: %d\n", tdq->tdq_oldswitchcnt); 374 printf("\ttimeshare idx: %d\n", tdq->tdq_idx); 375 printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx); 376 printf("\tload transferable: %d\n", tdq->tdq_transferable); 377 printf("\tlowest priority: %d\n", tdq->tdq_lowpri); 378 printf("\trealtime runq:\n"); 379 runq_print(&tdq->tdq_realtime); 380 printf("\ttimeshare runq:\n"); 381 runq_print(&tdq->tdq_timeshare); 382 printf("\tidle runq:\n"); 383 runq_print(&tdq->tdq_idle); 384 } 385 386 static inline int 387 sched_shouldpreempt(int pri, int cpri, int remote) 388 { 389 /* 390 * If the new priority is not better than the current priority there is 391 * nothing to do. 392 */ 393 if (pri >= cpri) 394 return (0); 395 /* 396 * Always preempt idle. 397 */ 398 if (cpri >= PRI_MIN_IDLE) 399 return (1); 400 /* 401 * If preemption is disabled don't preempt others. 402 */ 403 if (preempt_thresh == 0) 404 return (0); 405 /* 406 * Preempt if we exceed the threshold. 407 */ 408 if (pri <= preempt_thresh) 409 return (1); 410 /* 411 * If we're interactive or better and there is non-interactive 412 * or worse running preempt only remote processors. 413 */ 414 if (remote && pri <= PRI_MAX_INTERACT && cpri > PRI_MAX_INTERACT) 415 return (1); 416 return (0); 417 } 418 419 #define TS_RQ_PPQ (((PRI_MAX_BATCH - PRI_MIN_BATCH) + 1) / RQ_NQS) 420 /* 421 * Add a thread to the actual run-queue. Keeps transferable counts up to 422 * date with what is actually on the run-queue. Selects the correct 423 * queue position for timeshare threads. 424 */ 425 static __inline void 426 tdq_runq_add(struct tdq *tdq, struct thread *td, int flags) 427 { 428 struct td_sched *ts; 429 u_char pri; 430 431 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 432 THREAD_LOCK_ASSERT(td, MA_OWNED); 433 434 pri = td->td_priority; 435 ts = td->td_sched; 436 TD_SET_RUNQ(td); 437 if (THREAD_CAN_MIGRATE(td)) { 438 tdq->tdq_transferable++; 439 ts->ts_flags |= TSF_XFERABLE; 440 } 441 if (pri < PRI_MIN_BATCH) { 442 ts->ts_runq = &tdq->tdq_realtime; 443 } else if (pri <= PRI_MAX_BATCH) { 444 ts->ts_runq = &tdq->tdq_timeshare; 445 KASSERT(pri <= PRI_MAX_BATCH && pri >= PRI_MIN_BATCH, 446 ("Invalid priority %d on timeshare runq", pri)); 447 /* 448 * This queue contains only priorities between MIN and MAX 449 * realtime. Use the whole queue to represent these values. 450 */ 451 if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) { 452 pri = (pri - PRI_MIN_BATCH) / TS_RQ_PPQ; 453 pri = (pri + tdq->tdq_idx) % RQ_NQS; 454 /* 455 * This effectively shortens the queue by one so we 456 * can have a one slot difference between idx and 457 * ridx while we wait for threads to drain. 458 */ 459 if (tdq->tdq_ridx != tdq->tdq_idx && 460 pri == tdq->tdq_ridx) 461 pri = (unsigned char)(pri - 1) % RQ_NQS; 462 } else 463 pri = tdq->tdq_ridx; 464 runq_add_pri(ts->ts_runq, td, pri, flags); 465 return; 466 } else 467 ts->ts_runq = &tdq->tdq_idle; 468 runq_add(ts->ts_runq, td, flags); 469 } 470 471 /* 472 * Remove a thread from a run-queue. This typically happens when a thread 473 * is selected to run. Running threads are not on the queue and the 474 * transferable count does not reflect them. 475 */ 476 static __inline void 477 tdq_runq_rem(struct tdq *tdq, struct thread *td) 478 { 479 struct td_sched *ts; 480 481 ts = td->td_sched; 482 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 483 KASSERT(ts->ts_runq != NULL, 484 ("tdq_runq_remove: thread %p null ts_runq", td)); 485 if (ts->ts_flags & TSF_XFERABLE) { 486 tdq->tdq_transferable--; 487 ts->ts_flags &= ~TSF_XFERABLE; 488 } 489 if (ts->ts_runq == &tdq->tdq_timeshare) { 490 if (tdq->tdq_idx != tdq->tdq_ridx) 491 runq_remove_idx(ts->ts_runq, td, &tdq->tdq_ridx); 492 else 493 runq_remove_idx(ts->ts_runq, td, NULL); 494 } else 495 runq_remove(ts->ts_runq, td); 496 } 497 498 /* 499 * Load is maintained for all threads RUNNING and ON_RUNQ. Add the load 500 * for this thread to the referenced thread queue. 501 */ 502 static void 503 tdq_load_add(struct tdq *tdq, struct thread *td) 504 { 505 506 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 507 THREAD_LOCK_ASSERT(td, MA_OWNED); 508 509 tdq->tdq_load++; 510 if ((td->td_flags & TDF_NOLOAD) == 0) 511 tdq->tdq_sysload++; 512 KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load); 513 } 514 515 /* 516 * Remove the load from a thread that is transitioning to a sleep state or 517 * exiting. 518 */ 519 static void 520 tdq_load_rem(struct tdq *tdq, struct thread *td) 521 { 522 523 THREAD_LOCK_ASSERT(td, MA_OWNED); 524 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 525 KASSERT(tdq->tdq_load != 0, 526 ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq))); 527 528 tdq->tdq_load--; 529 if ((td->td_flags & TDF_NOLOAD) == 0) 530 tdq->tdq_sysload--; 531 KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load); 532 } 533 534 /* 535 * Set lowpri to its exact value by searching the run-queue and 536 * evaluating curthread. curthread may be passed as an optimization. 537 */ 538 static void 539 tdq_setlowpri(struct tdq *tdq, struct thread *ctd) 540 { 541 struct thread *td; 542 543 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 544 if (ctd == NULL) 545 ctd = pcpu_find(TDQ_ID(tdq))->pc_curthread; 546 td = tdq_choose(tdq); 547 if (td == NULL || td->td_priority > ctd->td_priority) 548 tdq->tdq_lowpri = ctd->td_priority; 549 else 550 tdq->tdq_lowpri = td->td_priority; 551 } 552 553 #ifdef SMP 554 struct cpu_search { 555 cpuset_t cs_mask; 556 u_int cs_load; 557 u_int cs_cpu; 558 int cs_limit; /* Min priority for low min load for high. */ 559 }; 560 561 #define CPU_SEARCH_LOWEST 0x1 562 #define CPU_SEARCH_HIGHEST 0x2 563 #define CPU_SEARCH_BOTH (CPU_SEARCH_LOWEST|CPU_SEARCH_HIGHEST) 564 565 #define CPUSET_FOREACH(cpu, mask) \ 566 for ((cpu) = 0; (cpu) <= mp_maxid; (cpu)++) \ 567 if ((mask) & 1 << (cpu)) 568 569 static __inline int cpu_search(struct cpu_group *cg, struct cpu_search *low, 570 struct cpu_search *high, const int match); 571 int cpu_search_lowest(struct cpu_group *cg, struct cpu_search *low); 572 int cpu_search_highest(struct cpu_group *cg, struct cpu_search *high); 573 int cpu_search_both(struct cpu_group *cg, struct cpu_search *low, 574 struct cpu_search *high); 575 576 /* 577 * This routine compares according to the match argument and should be 578 * reduced in actual instantiations via constant propagation and dead code 579 * elimination. 580 */ 581 static __inline int 582 cpu_compare(int cpu, struct cpu_search *low, struct cpu_search *high, 583 const int match) 584 { 585 struct tdq *tdq; 586 587 tdq = TDQ_CPU(cpu); 588 if (match & CPU_SEARCH_LOWEST) 589 if (CPU_ISSET(cpu, &low->cs_mask) && 590 tdq->tdq_load < low->cs_load && 591 tdq->tdq_lowpri > low->cs_limit) { 592 low->cs_cpu = cpu; 593 low->cs_load = tdq->tdq_load; 594 } 595 if (match & CPU_SEARCH_HIGHEST) 596 if (CPU_ISSET(cpu, &high->cs_mask) && 597 tdq->tdq_load >= high->cs_limit && 598 tdq->tdq_load > high->cs_load && 599 tdq->tdq_transferable) { 600 high->cs_cpu = cpu; 601 high->cs_load = tdq->tdq_load; 602 } 603 return (tdq->tdq_load); 604 } 605 606 /* 607 * Search the tree of cpu_groups for the lowest or highest loaded cpu 608 * according to the match argument. This routine actually compares the 609 * load on all paths through the tree and finds the least loaded cpu on 610 * the least loaded path, which may differ from the least loaded cpu in 611 * the system. This balances work among caches and busses. 612 * 613 * This inline is instantiated in three forms below using constants for the 614 * match argument. It is reduced to the minimum set for each case. It is 615 * also recursive to the depth of the tree. 616 */ 617 static __inline int 618 cpu_search(struct cpu_group *cg, struct cpu_search *low, 619 struct cpu_search *high, const int match) 620 { 621 int total; 622 623 total = 0; 624 if (cg->cg_children) { 625 struct cpu_search lgroup; 626 struct cpu_search hgroup; 627 struct cpu_group *child; 628 u_int lload; 629 int hload; 630 int load; 631 int i; 632 633 lload = -1; 634 hload = -1; 635 for (i = 0; i < cg->cg_children; i++) { 636 child = &cg->cg_child[i]; 637 if (match & CPU_SEARCH_LOWEST) { 638 lgroup = *low; 639 lgroup.cs_load = -1; 640 } 641 if (match & CPU_SEARCH_HIGHEST) { 642 hgroup = *high; 643 lgroup.cs_load = 0; 644 } 645 switch (match) { 646 case CPU_SEARCH_LOWEST: 647 load = cpu_search_lowest(child, &lgroup); 648 break; 649 case CPU_SEARCH_HIGHEST: 650 load = cpu_search_highest(child, &hgroup); 651 break; 652 case CPU_SEARCH_BOTH: 653 load = cpu_search_both(child, &lgroup, &hgroup); 654 break; 655 } 656 total += load; 657 if (match & CPU_SEARCH_LOWEST) 658 if (load < lload || low->cs_cpu == -1) { 659 *low = lgroup; 660 lload = load; 661 } 662 if (match & CPU_SEARCH_HIGHEST) 663 if (load > hload || high->cs_cpu == -1) { 664 hload = load; 665 *high = hgroup; 666 } 667 } 668 } else { 669 int cpu; 670 671 CPUSET_FOREACH(cpu, cg->cg_mask) 672 total += cpu_compare(cpu, low, high, match); 673 } 674 return (total); 675 } 676 677 /* 678 * cpu_search instantiations must pass constants to maintain the inline 679 * optimization. 680 */ 681 int 682 cpu_search_lowest(struct cpu_group *cg, struct cpu_search *low) 683 { 684 return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST); 685 } 686 687 int 688 cpu_search_highest(struct cpu_group *cg, struct cpu_search *high) 689 { 690 return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST); 691 } 692 693 int 694 cpu_search_both(struct cpu_group *cg, struct cpu_search *low, 695 struct cpu_search *high) 696 { 697 return cpu_search(cg, low, high, CPU_SEARCH_BOTH); 698 } 699 700 /* 701 * Find the cpu with the least load via the least loaded path that has a 702 * lowpri greater than pri pri. A pri of -1 indicates any priority is 703 * acceptable. 704 */ 705 static inline int 706 sched_lowest(struct cpu_group *cg, cpuset_t mask, int pri) 707 { 708 struct cpu_search low; 709 710 low.cs_cpu = -1; 711 low.cs_load = -1; 712 low.cs_mask = mask; 713 low.cs_limit = pri; 714 cpu_search_lowest(cg, &low); 715 return low.cs_cpu; 716 } 717 718 /* 719 * Find the cpu with the highest load via the highest loaded path. 720 */ 721 static inline int 722 sched_highest(struct cpu_group *cg, cpuset_t mask, int minload) 723 { 724 struct cpu_search high; 725 726 high.cs_cpu = -1; 727 high.cs_load = 0; 728 high.cs_mask = mask; 729 high.cs_limit = minload; 730 cpu_search_highest(cg, &high); 731 return high.cs_cpu; 732 } 733 734 /* 735 * Simultaneously find the highest and lowest loaded cpu reachable via 736 * cg. 737 */ 738 static inline void 739 sched_both(struct cpu_group *cg, cpuset_t mask, int *lowcpu, int *highcpu) 740 { 741 struct cpu_search high; 742 struct cpu_search low; 743 744 low.cs_cpu = -1; 745 low.cs_limit = -1; 746 low.cs_load = -1; 747 low.cs_mask = mask; 748 high.cs_load = 0; 749 high.cs_cpu = -1; 750 high.cs_limit = -1; 751 high.cs_mask = mask; 752 cpu_search_both(cg, &low, &high); 753 *lowcpu = low.cs_cpu; 754 *highcpu = high.cs_cpu; 755 return; 756 } 757 758 static void 759 sched_balance_group(struct cpu_group *cg) 760 { 761 cpuset_t mask; 762 int high; 763 int low; 764 int i; 765 766 CPU_FILL(&mask); 767 for (;;) { 768 sched_both(cg, mask, &low, &high); 769 if (low == high || low == -1 || high == -1) 770 break; 771 if (sched_balance_pair(TDQ_CPU(high), TDQ_CPU(low))) 772 break; 773 /* 774 * If we failed to move any threads determine which cpu 775 * to kick out of the set and try again. 776 */ 777 if (TDQ_CPU(high)->tdq_transferable == 0) 778 CPU_CLR(high, &mask); 779 else 780 CPU_CLR(low, &mask); 781 } 782 783 for (i = 0; i < cg->cg_children; i++) 784 sched_balance_group(&cg->cg_child[i]); 785 } 786 787 static void 788 sched_balance(void) 789 { 790 struct tdq *tdq; 791 792 /* 793 * Select a random time between .5 * balance_interval and 794 * 1.5 * balance_interval. 795 */ 796 balance_ticks = max(balance_interval / 2, 1); 797 balance_ticks += random() % balance_interval; 798 if (smp_started == 0 || rebalance == 0) 799 return; 800 tdq = TDQ_SELF(); 801 TDQ_UNLOCK(tdq); 802 sched_balance_group(cpu_top); 803 TDQ_LOCK(tdq); 804 } 805 806 /* 807 * Lock two thread queues using their address to maintain lock order. 808 */ 809 static void 810 tdq_lock_pair(struct tdq *one, struct tdq *two) 811 { 812 if (one < two) { 813 TDQ_LOCK(one); 814 TDQ_LOCK_FLAGS(two, MTX_DUPOK); 815 } else { 816 TDQ_LOCK(two); 817 TDQ_LOCK_FLAGS(one, MTX_DUPOK); 818 } 819 } 820 821 /* 822 * Unlock two thread queues. Order is not important here. 823 */ 824 static void 825 tdq_unlock_pair(struct tdq *one, struct tdq *two) 826 { 827 TDQ_UNLOCK(one); 828 TDQ_UNLOCK(two); 829 } 830 831 /* 832 * Transfer load between two imbalanced thread queues. 833 */ 834 static int 835 sched_balance_pair(struct tdq *high, struct tdq *low) 836 { 837 int transferable; 838 int high_load; 839 int low_load; 840 int moved; 841 int move; 842 int diff; 843 int i; 844 845 tdq_lock_pair(high, low); 846 transferable = high->tdq_transferable; 847 high_load = high->tdq_load; 848 low_load = low->tdq_load; 849 moved = 0; 850 /* 851 * Determine what the imbalance is and then adjust that to how many 852 * threads we actually have to give up (transferable). 853 */ 854 if (transferable != 0) { 855 diff = high_load - low_load; 856 move = diff / 2; 857 if (diff & 0x1) 858 move++; 859 move = min(move, transferable); 860 for (i = 0; i < move; i++) 861 moved += tdq_move(high, low); 862 /* 863 * IPI the target cpu to force it to reschedule with the new 864 * workload. 865 */ 866 ipi_cpu(TDQ_ID(low), IPI_PREEMPT); 867 } 868 tdq_unlock_pair(high, low); 869 return (moved); 870 } 871 872 /* 873 * Move a thread from one thread queue to another. 874 */ 875 static int 876 tdq_move(struct tdq *from, struct tdq *to) 877 { 878 struct td_sched *ts; 879 struct thread *td; 880 struct tdq *tdq; 881 int cpu; 882 883 TDQ_LOCK_ASSERT(from, MA_OWNED); 884 TDQ_LOCK_ASSERT(to, MA_OWNED); 885 886 tdq = from; 887 cpu = TDQ_ID(to); 888 td = tdq_steal(tdq, cpu); 889 if (td == NULL) 890 return (0); 891 ts = td->td_sched; 892 /* 893 * Although the run queue is locked the thread may be blocked. Lock 894 * it to clear this and acquire the run-queue lock. 895 */ 896 thread_lock(td); 897 /* Drop recursive lock on from acquired via thread_lock(). */ 898 TDQ_UNLOCK(from); 899 sched_rem(td); 900 ts->ts_cpu = cpu; 901 td->td_lock = TDQ_LOCKPTR(to); 902 tdq_add(to, td, SRQ_YIELDING); 903 return (1); 904 } 905 906 /* 907 * This tdq has idled. Try to steal a thread from another cpu and switch 908 * to it. 909 */ 910 static int 911 tdq_idled(struct tdq *tdq) 912 { 913 struct cpu_group *cg; 914 struct tdq *steal; 915 cpuset_t mask; 916 int thresh; 917 int cpu; 918 919 if (smp_started == 0 || steal_idle == 0) 920 return (1); 921 CPU_FILL(&mask); 922 CPU_CLR(PCPU_GET(cpuid), &mask); 923 /* We don't want to be preempted while we're iterating. */ 924 spinlock_enter(); 925 for (cg = tdq->tdq_cg; cg != NULL; ) { 926 if ((cg->cg_flags & CG_FLAG_THREAD) == 0) 927 thresh = steal_thresh; 928 else 929 thresh = 1; 930 cpu = sched_highest(cg, mask, thresh); 931 if (cpu == -1) { 932 cg = cg->cg_parent; 933 continue; 934 } 935 steal = TDQ_CPU(cpu); 936 CPU_CLR(cpu, &mask); 937 tdq_lock_pair(tdq, steal); 938 if (steal->tdq_load < thresh || steal->tdq_transferable == 0) { 939 tdq_unlock_pair(tdq, steal); 940 continue; 941 } 942 /* 943 * If a thread was added while interrupts were disabled don't 944 * steal one here. If we fail to acquire one due to affinity 945 * restrictions loop again with this cpu removed from the 946 * set. 947 */ 948 if (tdq->tdq_load == 0 && tdq_move(steal, tdq) == 0) { 949 tdq_unlock_pair(tdq, steal); 950 continue; 951 } 952 spinlock_exit(); 953 TDQ_UNLOCK(steal); 954 mi_switch(SW_VOL | SWT_IDLE, NULL); 955 thread_unlock(curthread); 956 957 return (0); 958 } 959 spinlock_exit(); 960 return (1); 961 } 962 963 /* 964 * Notify a remote cpu of new work. Sends an IPI if criteria are met. 965 */ 966 static void 967 tdq_notify(struct tdq *tdq, struct thread *td) 968 { 969 struct thread *ctd; 970 int pri; 971 int cpu; 972 973 if (tdq->tdq_ipipending) 974 return; 975 cpu = td->td_sched->ts_cpu; 976 pri = td->td_priority; 977 ctd = pcpu_find(cpu)->pc_curthread; 978 if (!sched_shouldpreempt(pri, ctd->td_priority, 1)) 979 return; 980 if (TD_IS_IDLETHREAD(ctd)) { 981 /* 982 * If the MD code has an idle wakeup routine try that before 983 * falling back to IPI. 984 */ 985 if (!tdq->tdq_cpu_idle || cpu_idle_wakeup(cpu)) 986 return; 987 } 988 tdq->tdq_ipipending = 1; 989 ipi_cpu(cpu, IPI_PREEMPT); 990 } 991 992 /* 993 * Steals load from a timeshare queue. Honors the rotating queue head 994 * index. 995 */ 996 static struct thread * 997 runq_steal_from(struct runq *rq, int cpu, u_char start) 998 { 999 struct rqbits *rqb; 1000 struct rqhead *rqh; 1001 struct thread *td; 1002 int first; 1003 int bit; 1004 int pri; 1005 int i; 1006 1007 rqb = &rq->rq_status; 1008 bit = start & (RQB_BPW -1); 1009 pri = 0; 1010 first = 0; 1011 again: 1012 for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) { 1013 if (rqb->rqb_bits[i] == 0) 1014 continue; 1015 if (bit != 0) { 1016 for (pri = bit; pri < RQB_BPW; pri++) 1017 if (rqb->rqb_bits[i] & (1ul << pri)) 1018 break; 1019 if (pri >= RQB_BPW) 1020 continue; 1021 } else 1022 pri = RQB_FFS(rqb->rqb_bits[i]); 1023 pri += (i << RQB_L2BPW); 1024 rqh = &rq->rq_queues[pri]; 1025 TAILQ_FOREACH(td, rqh, td_runq) { 1026 if (first && THREAD_CAN_MIGRATE(td) && 1027 THREAD_CAN_SCHED(td, cpu)) 1028 return (td); 1029 first = 1; 1030 } 1031 } 1032 if (start != 0) { 1033 start = 0; 1034 goto again; 1035 } 1036 1037 return (NULL); 1038 } 1039 1040 /* 1041 * Steals load from a standard linear queue. 1042 */ 1043 static struct thread * 1044 runq_steal(struct runq *rq, int cpu) 1045 { 1046 struct rqhead *rqh; 1047 struct rqbits *rqb; 1048 struct thread *td; 1049 int word; 1050 int bit; 1051 1052 rqb = &rq->rq_status; 1053 for (word = 0; word < RQB_LEN; word++) { 1054 if (rqb->rqb_bits[word] == 0) 1055 continue; 1056 for (bit = 0; bit < RQB_BPW; bit++) { 1057 if ((rqb->rqb_bits[word] & (1ul << bit)) == 0) 1058 continue; 1059 rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)]; 1060 TAILQ_FOREACH(td, rqh, td_runq) 1061 if (THREAD_CAN_MIGRATE(td) && 1062 THREAD_CAN_SCHED(td, cpu)) 1063 return (td); 1064 } 1065 } 1066 return (NULL); 1067 } 1068 1069 /* 1070 * Attempt to steal a thread in priority order from a thread queue. 1071 */ 1072 static struct thread * 1073 tdq_steal(struct tdq *tdq, int cpu) 1074 { 1075 struct thread *td; 1076 1077 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 1078 if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL) 1079 return (td); 1080 if ((td = runq_steal_from(&tdq->tdq_timeshare, 1081 cpu, tdq->tdq_ridx)) != NULL) 1082 return (td); 1083 return (runq_steal(&tdq->tdq_idle, cpu)); 1084 } 1085 1086 /* 1087 * Sets the thread lock and ts_cpu to match the requested cpu. Unlocks the 1088 * current lock and returns with the assigned queue locked. 1089 */ 1090 static inline struct tdq * 1091 sched_setcpu(struct thread *td, int cpu, int flags) 1092 { 1093 1094 struct tdq *tdq; 1095 1096 THREAD_LOCK_ASSERT(td, MA_OWNED); 1097 tdq = TDQ_CPU(cpu); 1098 td->td_sched->ts_cpu = cpu; 1099 /* 1100 * If the lock matches just return the queue. 1101 */ 1102 if (td->td_lock == TDQ_LOCKPTR(tdq)) 1103 return (tdq); 1104 #ifdef notyet 1105 /* 1106 * If the thread isn't running its lockptr is a 1107 * turnstile or a sleepqueue. We can just lock_set without 1108 * blocking. 1109 */ 1110 if (TD_CAN_RUN(td)) { 1111 TDQ_LOCK(tdq); 1112 thread_lock_set(td, TDQ_LOCKPTR(tdq)); 1113 return (tdq); 1114 } 1115 #endif 1116 /* 1117 * The hard case, migration, we need to block the thread first to 1118 * prevent order reversals with other cpus locks. 1119 */ 1120 spinlock_enter(); 1121 thread_lock_block(td); 1122 TDQ_LOCK(tdq); 1123 thread_lock_unblock(td, TDQ_LOCKPTR(tdq)); 1124 spinlock_exit(); 1125 return (tdq); 1126 } 1127 1128 SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding"); 1129 SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity"); 1130 SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity"); 1131 SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load"); 1132 SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu"); 1133 SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration"); 1134 1135 static int 1136 sched_pickcpu(struct thread *td, int flags) 1137 { 1138 struct cpu_group *cg; 1139 struct td_sched *ts; 1140 struct tdq *tdq; 1141 cpuset_t mask; 1142 int self; 1143 int pri; 1144 int cpu; 1145 1146 self = PCPU_GET(cpuid); 1147 ts = td->td_sched; 1148 if (smp_started == 0) 1149 return (self); 1150 /* 1151 * Don't migrate a running thread from sched_switch(). 1152 */ 1153 if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td)) 1154 return (ts->ts_cpu); 1155 /* 1156 * Prefer to run interrupt threads on the processors that generate 1157 * the interrupt. 1158 */ 1159 if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) && 1160 curthread->td_intr_nesting_level && ts->ts_cpu != self) { 1161 SCHED_STAT_INC(pickcpu_intrbind); 1162 ts->ts_cpu = self; 1163 } 1164 /* 1165 * If the thread can run on the last cpu and the affinity has not 1166 * expired or it is idle run it there. 1167 */ 1168 pri = td->td_priority; 1169 tdq = TDQ_CPU(ts->ts_cpu); 1170 if (THREAD_CAN_SCHED(td, ts->ts_cpu)) { 1171 if (tdq->tdq_lowpri > PRI_MIN_IDLE) { 1172 SCHED_STAT_INC(pickcpu_idle_affinity); 1173 return (ts->ts_cpu); 1174 } 1175 if (SCHED_AFFINITY(ts, CG_SHARE_L2) && tdq->tdq_lowpri > pri) { 1176 SCHED_STAT_INC(pickcpu_affinity); 1177 return (ts->ts_cpu); 1178 } 1179 } 1180 /* 1181 * Search for the highest level in the tree that still has affinity. 1182 */ 1183 cg = NULL; 1184 for (cg = tdq->tdq_cg; cg != NULL; cg = cg->cg_parent) 1185 if (SCHED_AFFINITY(ts, cg->cg_level)) 1186 break; 1187 cpu = -1; 1188 mask = td->td_cpuset->cs_mask; 1189 if (cg) 1190 cpu = sched_lowest(cg, mask, pri); 1191 if (cpu == -1) 1192 cpu = sched_lowest(cpu_top, mask, -1); 1193 /* 1194 * Compare the lowest loaded cpu to current cpu. 1195 */ 1196 if (THREAD_CAN_SCHED(td, self) && TDQ_CPU(self)->tdq_lowpri > pri && 1197 TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE) { 1198 SCHED_STAT_INC(pickcpu_local); 1199 cpu = self; 1200 } else 1201 SCHED_STAT_INC(pickcpu_lowest); 1202 if (cpu != ts->ts_cpu) 1203 SCHED_STAT_INC(pickcpu_migration); 1204 KASSERT(cpu != -1, ("sched_pickcpu: Failed to find a cpu.")); 1205 return (cpu); 1206 } 1207 #endif 1208 1209 /* 1210 * Pick the highest priority task we have and return it. 1211 */ 1212 static struct thread * 1213 tdq_choose(struct tdq *tdq) 1214 { 1215 struct thread *td; 1216 1217 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 1218 td = runq_choose(&tdq->tdq_realtime); 1219 if (td != NULL) 1220 return (td); 1221 td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx); 1222 if (td != NULL) { 1223 KASSERT(td->td_priority >= PRI_MIN_BATCH, 1224 ("tdq_choose: Invalid priority on timeshare queue %d", 1225 td->td_priority)); 1226 return (td); 1227 } 1228 td = runq_choose(&tdq->tdq_idle); 1229 if (td != NULL) { 1230 KASSERT(td->td_priority >= PRI_MIN_IDLE, 1231 ("tdq_choose: Invalid priority on idle queue %d", 1232 td->td_priority)); 1233 return (td); 1234 } 1235 1236 return (NULL); 1237 } 1238 1239 /* 1240 * Initialize a thread queue. 1241 */ 1242 static void 1243 tdq_setup(struct tdq *tdq) 1244 { 1245 1246 if (bootverbose) 1247 printf("ULE: setup cpu %d\n", TDQ_ID(tdq)); 1248 runq_init(&tdq->tdq_realtime); 1249 runq_init(&tdq->tdq_timeshare); 1250 runq_init(&tdq->tdq_idle); 1251 snprintf(tdq->tdq_name, sizeof(tdq->tdq_name), 1252 "sched lock %d", (int)TDQ_ID(tdq)); 1253 mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock", 1254 MTX_SPIN | MTX_RECURSE); 1255 #ifdef KTR 1256 snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname), 1257 "CPU %d load", (int)TDQ_ID(tdq)); 1258 #endif 1259 } 1260 1261 #ifdef SMP 1262 static void 1263 sched_setup_smp(void) 1264 { 1265 struct tdq *tdq; 1266 int i; 1267 1268 cpu_top = smp_topo(); 1269 CPU_FOREACH(i) { 1270 tdq = TDQ_CPU(i); 1271 tdq_setup(tdq); 1272 tdq->tdq_cg = smp_topo_find(cpu_top, i); 1273 if (tdq->tdq_cg == NULL) 1274 panic("Can't find cpu group for %d\n", i); 1275 } 1276 balance_tdq = TDQ_SELF(); 1277 sched_balance(); 1278 } 1279 #endif 1280 1281 /* 1282 * Setup the thread queues and initialize the topology based on MD 1283 * information. 1284 */ 1285 static void 1286 sched_setup(void *dummy) 1287 { 1288 struct tdq *tdq; 1289 1290 tdq = TDQ_SELF(); 1291 #ifdef SMP 1292 sched_setup_smp(); 1293 #else 1294 tdq_setup(tdq); 1295 #endif 1296 /* 1297 * To avoid divide-by-zero, we set realstathz a dummy value 1298 * in case which sched_clock() called before sched_initticks(). 1299 */ 1300 realstathz = hz; 1301 sched_slice = (realstathz/10); /* ~100ms */ 1302 tickincr = 1 << SCHED_TICK_SHIFT; 1303 1304 /* Add thread0's load since it's running. */ 1305 TDQ_LOCK(tdq); 1306 thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF()); 1307 tdq_load_add(tdq, &thread0); 1308 tdq->tdq_lowpri = thread0.td_priority; 1309 TDQ_UNLOCK(tdq); 1310 } 1311 1312 /* 1313 * This routine determines the tickincr after stathz and hz are setup. 1314 */ 1315 /* ARGSUSED */ 1316 static void 1317 sched_initticks(void *dummy) 1318 { 1319 int incr; 1320 1321 realstathz = stathz ? stathz : hz; 1322 sched_slice = (realstathz/10); /* ~100ms */ 1323 1324 /* 1325 * tickincr is shifted out by 10 to avoid rounding errors due to 1326 * hz not being evenly divisible by stathz on all platforms. 1327 */ 1328 incr = (hz << SCHED_TICK_SHIFT) / realstathz; 1329 /* 1330 * This does not work for values of stathz that are more than 1331 * 1 << SCHED_TICK_SHIFT * hz. In practice this does not happen. 1332 */ 1333 if (incr == 0) 1334 incr = 1; 1335 tickincr = incr; 1336 #ifdef SMP 1337 /* 1338 * Set the default balance interval now that we know 1339 * what realstathz is. 1340 */ 1341 balance_interval = realstathz; 1342 /* 1343 * Set steal thresh to roughly log2(mp_ncpu) but no greater than 4. 1344 * This prevents excess thrashing on large machines and excess idle 1345 * on smaller machines. 1346 */ 1347 steal_thresh = min(fls(mp_ncpus) - 1, 3); 1348 affinity = SCHED_AFFINITY_DEFAULT; 1349 #endif 1350 } 1351 1352 1353 /* 1354 * This is the core of the interactivity algorithm. Determines a score based 1355 * on past behavior. It is the ratio of sleep time to run time scaled to 1356 * a [0, 100] integer. This is the voluntary sleep time of a process, which 1357 * differs from the cpu usage because it does not account for time spent 1358 * waiting on a run-queue. Would be prettier if we had floating point. 1359 */ 1360 static int 1361 sched_interact_score(struct thread *td) 1362 { 1363 struct td_sched *ts; 1364 int div; 1365 1366 ts = td->td_sched; 1367 /* 1368 * The score is only needed if this is likely to be an interactive 1369 * task. Don't go through the expense of computing it if there's 1370 * no chance. 1371 */ 1372 if (sched_interact <= SCHED_INTERACT_HALF && 1373 ts->ts_runtime >= ts->ts_slptime) 1374 return (SCHED_INTERACT_HALF); 1375 1376 if (ts->ts_runtime > ts->ts_slptime) { 1377 div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF); 1378 return (SCHED_INTERACT_HALF + 1379 (SCHED_INTERACT_HALF - (ts->ts_slptime / div))); 1380 } 1381 if (ts->ts_slptime > ts->ts_runtime) { 1382 div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF); 1383 return (ts->ts_runtime / div); 1384 } 1385 /* runtime == slptime */ 1386 if (ts->ts_runtime) 1387 return (SCHED_INTERACT_HALF); 1388 1389 /* 1390 * This can happen if slptime and runtime are 0. 1391 */ 1392 return (0); 1393 1394 } 1395 1396 /* 1397 * Scale the scheduling priority according to the "interactivity" of this 1398 * process. 1399 */ 1400 static void 1401 sched_priority(struct thread *td) 1402 { 1403 int score; 1404 int pri; 1405 1406 if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE) 1407 return; 1408 /* 1409 * If the score is interactive we place the thread in the realtime 1410 * queue with a priority that is less than kernel and interrupt 1411 * priorities. These threads are not subject to nice restrictions. 1412 * 1413 * Scores greater than this are placed on the normal timeshare queue 1414 * where the priority is partially decided by the most recent cpu 1415 * utilization and the rest is decided by nice value. 1416 * 1417 * The nice value of the process has a linear effect on the calculated 1418 * score. Negative nice values make it easier for a thread to be 1419 * considered interactive. 1420 */ 1421 score = imax(0, sched_interact_score(td) + td->td_proc->p_nice); 1422 if (score < sched_interact) { 1423 pri = PRI_MIN_INTERACT; 1424 pri += ((PRI_MAX_INTERACT - PRI_MIN_INTERACT + 1) / 1425 sched_interact) * score; 1426 KASSERT(pri >= PRI_MIN_INTERACT && pri <= PRI_MAX_INTERACT, 1427 ("sched_priority: invalid interactive priority %d score %d", 1428 pri, score)); 1429 } else { 1430 pri = SCHED_PRI_MIN; 1431 if (td->td_sched->ts_ticks) 1432 pri += SCHED_PRI_TICKS(td->td_sched); 1433 pri += SCHED_PRI_NICE(td->td_proc->p_nice); 1434 KASSERT(pri >= PRI_MIN_BATCH && pri <= PRI_MAX_BATCH, 1435 ("sched_priority: invalid priority %d: nice %d, " 1436 "ticks %d ftick %d ltick %d tick pri %d", 1437 pri, td->td_proc->p_nice, td->td_sched->ts_ticks, 1438 td->td_sched->ts_ftick, td->td_sched->ts_ltick, 1439 SCHED_PRI_TICKS(td->td_sched))); 1440 } 1441 sched_user_prio(td, pri); 1442 1443 return; 1444 } 1445 1446 /* 1447 * This routine enforces a maximum limit on the amount of scheduling history 1448 * kept. It is called after either the slptime or runtime is adjusted. This 1449 * function is ugly due to integer math. 1450 */ 1451 static void 1452 sched_interact_update(struct thread *td) 1453 { 1454 struct td_sched *ts; 1455 u_int sum; 1456 1457 ts = td->td_sched; 1458 sum = ts->ts_runtime + ts->ts_slptime; 1459 if (sum < SCHED_SLP_RUN_MAX) 1460 return; 1461 /* 1462 * This only happens from two places: 1463 * 1) We have added an unusual amount of run time from fork_exit. 1464 * 2) We have added an unusual amount of sleep time from sched_sleep(). 1465 */ 1466 if (sum > SCHED_SLP_RUN_MAX * 2) { 1467 if (ts->ts_runtime > ts->ts_slptime) { 1468 ts->ts_runtime = SCHED_SLP_RUN_MAX; 1469 ts->ts_slptime = 1; 1470 } else { 1471 ts->ts_slptime = SCHED_SLP_RUN_MAX; 1472 ts->ts_runtime = 1; 1473 } 1474 return; 1475 } 1476 /* 1477 * If we have exceeded by more than 1/5th then the algorithm below 1478 * will not bring us back into range. Dividing by two here forces 1479 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX] 1480 */ 1481 if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) { 1482 ts->ts_runtime /= 2; 1483 ts->ts_slptime /= 2; 1484 return; 1485 } 1486 ts->ts_runtime = (ts->ts_runtime / 5) * 4; 1487 ts->ts_slptime = (ts->ts_slptime / 5) * 4; 1488 } 1489 1490 /* 1491 * Scale back the interactivity history when a child thread is created. The 1492 * history is inherited from the parent but the thread may behave totally 1493 * differently. For example, a shell spawning a compiler process. We want 1494 * to learn that the compiler is behaving badly very quickly. 1495 */ 1496 static void 1497 sched_interact_fork(struct thread *td) 1498 { 1499 int ratio; 1500 int sum; 1501 1502 sum = td->td_sched->ts_runtime + td->td_sched->ts_slptime; 1503 if (sum > SCHED_SLP_RUN_FORK) { 1504 ratio = sum / SCHED_SLP_RUN_FORK; 1505 td->td_sched->ts_runtime /= ratio; 1506 td->td_sched->ts_slptime /= ratio; 1507 } 1508 } 1509 1510 /* 1511 * Called from proc0_init() to setup the scheduler fields. 1512 */ 1513 void 1514 schedinit(void) 1515 { 1516 1517 /* 1518 * Set up the scheduler specific parts of proc0. 1519 */ 1520 proc0.p_sched = NULL; /* XXX */ 1521 thread0.td_sched = &td_sched0; 1522 td_sched0.ts_ltick = ticks; 1523 td_sched0.ts_ftick = ticks; 1524 td_sched0.ts_slice = sched_slice; 1525 } 1526 1527 /* 1528 * This is only somewhat accurate since given many processes of the same 1529 * priority they will switch when their slices run out, which will be 1530 * at most sched_slice stathz ticks. 1531 */ 1532 int 1533 sched_rr_interval(void) 1534 { 1535 1536 /* Convert sched_slice to hz */ 1537 return (hz/(realstathz/sched_slice)); 1538 } 1539 1540 /* 1541 * Update the percent cpu tracking information when it is requested or 1542 * the total history exceeds the maximum. We keep a sliding history of 1543 * tick counts that slowly decays. This is less precise than the 4BSD 1544 * mechanism since it happens with less regular and frequent events. 1545 */ 1546 static void 1547 sched_pctcpu_update(struct td_sched *ts) 1548 { 1549 1550 if (ts->ts_ticks == 0) 1551 return; 1552 if (ticks - (hz / 10) < ts->ts_ltick && 1553 SCHED_TICK_TOTAL(ts) < SCHED_TICK_MAX) 1554 return; 1555 /* 1556 * Adjust counters and watermark for pctcpu calc. 1557 */ 1558 if (ts->ts_ltick > ticks - SCHED_TICK_TARG) 1559 ts->ts_ticks = (ts->ts_ticks / (ticks - ts->ts_ftick)) * 1560 SCHED_TICK_TARG; 1561 else 1562 ts->ts_ticks = 0; 1563 ts->ts_ltick = ticks; 1564 ts->ts_ftick = ts->ts_ltick - SCHED_TICK_TARG; 1565 } 1566 1567 /* 1568 * Adjust the priority of a thread. Move it to the appropriate run-queue 1569 * if necessary. This is the back-end for several priority related 1570 * functions. 1571 */ 1572 static void 1573 sched_thread_priority(struct thread *td, u_char prio) 1574 { 1575 struct td_sched *ts; 1576 struct tdq *tdq; 1577 int oldpri; 1578 1579 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio", 1580 "prio:%d", td->td_priority, "new prio:%d", prio, 1581 KTR_ATTR_LINKED, sched_tdname(curthread)); 1582 if (td != curthread && prio > td->td_priority) { 1583 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread), 1584 "lend prio", "prio:%d", td->td_priority, "new prio:%d", 1585 prio, KTR_ATTR_LINKED, sched_tdname(td)); 1586 } 1587 ts = td->td_sched; 1588 THREAD_LOCK_ASSERT(td, MA_OWNED); 1589 if (td->td_priority == prio) 1590 return; 1591 /* 1592 * If the priority has been elevated due to priority 1593 * propagation, we may have to move ourselves to a new 1594 * queue. This could be optimized to not re-add in some 1595 * cases. 1596 */ 1597 if (TD_ON_RUNQ(td) && prio < td->td_priority) { 1598 sched_rem(td); 1599 td->td_priority = prio; 1600 sched_add(td, SRQ_BORROWING); 1601 return; 1602 } 1603 /* 1604 * If the thread is currently running we may have to adjust the lowpri 1605 * information so other cpus are aware of our current priority. 1606 */ 1607 if (TD_IS_RUNNING(td)) { 1608 tdq = TDQ_CPU(ts->ts_cpu); 1609 oldpri = td->td_priority; 1610 td->td_priority = prio; 1611 if (prio < tdq->tdq_lowpri) 1612 tdq->tdq_lowpri = prio; 1613 else if (tdq->tdq_lowpri == oldpri) 1614 tdq_setlowpri(tdq, td); 1615 return; 1616 } 1617 td->td_priority = prio; 1618 } 1619 1620 /* 1621 * Update a thread's priority when it is lent another thread's 1622 * priority. 1623 */ 1624 void 1625 sched_lend_prio(struct thread *td, u_char prio) 1626 { 1627 1628 td->td_flags |= TDF_BORROWING; 1629 sched_thread_priority(td, prio); 1630 } 1631 1632 /* 1633 * Restore a thread's priority when priority propagation is 1634 * over. The prio argument is the minimum priority the thread 1635 * needs to have to satisfy other possible priority lending 1636 * requests. If the thread's regular priority is less 1637 * important than prio, the thread will keep a priority boost 1638 * of prio. 1639 */ 1640 void 1641 sched_unlend_prio(struct thread *td, u_char prio) 1642 { 1643 u_char base_pri; 1644 1645 if (td->td_base_pri >= PRI_MIN_TIMESHARE && 1646 td->td_base_pri <= PRI_MAX_TIMESHARE) 1647 base_pri = td->td_user_pri; 1648 else 1649 base_pri = td->td_base_pri; 1650 if (prio >= base_pri) { 1651 td->td_flags &= ~TDF_BORROWING; 1652 sched_thread_priority(td, base_pri); 1653 } else 1654 sched_lend_prio(td, prio); 1655 } 1656 1657 /* 1658 * Standard entry for setting the priority to an absolute value. 1659 */ 1660 void 1661 sched_prio(struct thread *td, u_char prio) 1662 { 1663 u_char oldprio; 1664 1665 /* First, update the base priority. */ 1666 td->td_base_pri = prio; 1667 1668 /* 1669 * If the thread is borrowing another thread's priority, don't 1670 * ever lower the priority. 1671 */ 1672 if (td->td_flags & TDF_BORROWING && td->td_priority < prio) 1673 return; 1674 1675 /* Change the real priority. */ 1676 oldprio = td->td_priority; 1677 sched_thread_priority(td, prio); 1678 1679 /* 1680 * If the thread is on a turnstile, then let the turnstile update 1681 * its state. 1682 */ 1683 if (TD_ON_LOCK(td) && oldprio != prio) 1684 turnstile_adjust(td, oldprio); 1685 } 1686 1687 /* 1688 * Set the base user priority, does not effect current running priority. 1689 */ 1690 void 1691 sched_user_prio(struct thread *td, u_char prio) 1692 { 1693 1694 td->td_base_user_pri = prio; 1695 if (td->td_lend_user_pri <= prio) 1696 return; 1697 td->td_user_pri = prio; 1698 } 1699 1700 void 1701 sched_lend_user_prio(struct thread *td, u_char prio) 1702 { 1703 1704 THREAD_LOCK_ASSERT(td, MA_OWNED); 1705 td->td_lend_user_pri = prio; 1706 td->td_user_pri = min(prio, td->td_base_user_pri); 1707 if (td->td_priority > td->td_user_pri) 1708 sched_prio(td, td->td_user_pri); 1709 else if (td->td_priority != td->td_user_pri) 1710 td->td_flags |= TDF_NEEDRESCHED; 1711 } 1712 1713 /* 1714 * Handle migration from sched_switch(). This happens only for 1715 * cpu binding. 1716 */ 1717 static struct mtx * 1718 sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags) 1719 { 1720 struct tdq *tdn; 1721 1722 tdn = TDQ_CPU(td->td_sched->ts_cpu); 1723 #ifdef SMP 1724 tdq_load_rem(tdq, td); 1725 /* 1726 * Do the lock dance required to avoid LOR. We grab an extra 1727 * spinlock nesting to prevent preemption while we're 1728 * not holding either run-queue lock. 1729 */ 1730 spinlock_enter(); 1731 thread_lock_block(td); /* This releases the lock on tdq. */ 1732 1733 /* 1734 * Acquire both run-queue locks before placing the thread on the new 1735 * run-queue to avoid deadlocks created by placing a thread with a 1736 * blocked lock on the run-queue of a remote processor. The deadlock 1737 * occurs when a third processor attempts to lock the two queues in 1738 * question while the target processor is spinning with its own 1739 * run-queue lock held while waiting for the blocked lock to clear. 1740 */ 1741 tdq_lock_pair(tdn, tdq); 1742 tdq_add(tdn, td, flags); 1743 tdq_notify(tdn, td); 1744 TDQ_UNLOCK(tdn); 1745 spinlock_exit(); 1746 #endif 1747 return (TDQ_LOCKPTR(tdn)); 1748 } 1749 1750 /* 1751 * Variadic version of thread_lock_unblock() that does not assume td_lock 1752 * is blocked. 1753 */ 1754 static inline void 1755 thread_unblock_switch(struct thread *td, struct mtx *mtx) 1756 { 1757 atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock, 1758 (uintptr_t)mtx); 1759 } 1760 1761 /* 1762 * Switch threads. This function has to handle threads coming in while 1763 * blocked for some reason, running, or idle. It also must deal with 1764 * migrating a thread from one queue to another as running threads may 1765 * be assigned elsewhere via binding. 1766 */ 1767 void 1768 sched_switch(struct thread *td, struct thread *newtd, int flags) 1769 { 1770 struct tdq *tdq; 1771 struct td_sched *ts; 1772 struct mtx *mtx; 1773 int srqflag; 1774 int cpuid; 1775 1776 THREAD_LOCK_ASSERT(td, MA_OWNED); 1777 KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument")); 1778 1779 cpuid = PCPU_GET(cpuid); 1780 tdq = TDQ_CPU(cpuid); 1781 ts = td->td_sched; 1782 mtx = td->td_lock; 1783 ts->ts_rltick = ticks; 1784 td->td_lastcpu = td->td_oncpu; 1785 td->td_oncpu = NOCPU; 1786 td->td_flags &= ~TDF_NEEDRESCHED; 1787 td->td_owepreempt = 0; 1788 tdq->tdq_switchcnt++; 1789 /* 1790 * The lock pointer in an idle thread should never change. Reset it 1791 * to CAN_RUN as well. 1792 */ 1793 if (TD_IS_IDLETHREAD(td)) { 1794 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1795 TD_SET_CAN_RUN(td); 1796 } else if (TD_IS_RUNNING(td)) { 1797 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1798 srqflag = (flags & SW_PREEMPT) ? 1799 SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED : 1800 SRQ_OURSELF|SRQ_YIELDING; 1801 #ifdef SMP 1802 if (THREAD_CAN_MIGRATE(td) && !THREAD_CAN_SCHED(td, ts->ts_cpu)) 1803 ts->ts_cpu = sched_pickcpu(td, 0); 1804 #endif 1805 if (ts->ts_cpu == cpuid) 1806 tdq_runq_add(tdq, td, srqflag); 1807 else { 1808 KASSERT(THREAD_CAN_MIGRATE(td) || 1809 (ts->ts_flags & TSF_BOUND) != 0, 1810 ("Thread %p shouldn't migrate", td)); 1811 mtx = sched_switch_migrate(tdq, td, srqflag); 1812 } 1813 } else { 1814 /* This thread must be going to sleep. */ 1815 TDQ_LOCK(tdq); 1816 mtx = thread_lock_block(td); 1817 tdq_load_rem(tdq, td); 1818 } 1819 /* 1820 * We enter here with the thread blocked and assigned to the 1821 * appropriate cpu run-queue or sleep-queue and with the current 1822 * thread-queue locked. 1823 */ 1824 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED); 1825 newtd = choosethread(); 1826 /* 1827 * Call the MD code to switch contexts if necessary. 1828 */ 1829 if (td != newtd) { 1830 #ifdef HWPMC_HOOKS 1831 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1832 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT); 1833 #endif 1834 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object); 1835 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd; 1836 1837 #ifdef KDTRACE_HOOKS 1838 /* 1839 * If DTrace has set the active vtime enum to anything 1840 * other than INACTIVE (0), then it should have set the 1841 * function to call. 1842 */ 1843 if (dtrace_vtime_active) 1844 (*dtrace_vtime_switch_func)(newtd); 1845 #endif 1846 1847 cpu_switch(td, newtd, mtx); 1848 /* 1849 * We may return from cpu_switch on a different cpu. However, 1850 * we always return with td_lock pointing to the current cpu's 1851 * run queue lock. 1852 */ 1853 cpuid = PCPU_GET(cpuid); 1854 tdq = TDQ_CPU(cpuid); 1855 lock_profile_obtain_lock_success( 1856 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__); 1857 #ifdef HWPMC_HOOKS 1858 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1859 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN); 1860 #endif 1861 } else 1862 thread_unblock_switch(td, mtx); 1863 /* 1864 * Assert that all went well and return. 1865 */ 1866 TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED); 1867 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1868 td->td_oncpu = cpuid; 1869 } 1870 1871 /* 1872 * Adjust thread priorities as a result of a nice request. 1873 */ 1874 void 1875 sched_nice(struct proc *p, int nice) 1876 { 1877 struct thread *td; 1878 1879 PROC_LOCK_ASSERT(p, MA_OWNED); 1880 1881 p->p_nice = nice; 1882 FOREACH_THREAD_IN_PROC(p, td) { 1883 thread_lock(td); 1884 sched_priority(td); 1885 sched_prio(td, td->td_base_user_pri); 1886 thread_unlock(td); 1887 } 1888 } 1889 1890 /* 1891 * Record the sleep time for the interactivity scorer. 1892 */ 1893 void 1894 sched_sleep(struct thread *td, int prio) 1895 { 1896 1897 THREAD_LOCK_ASSERT(td, MA_OWNED); 1898 1899 td->td_slptick = ticks; 1900 if (TD_IS_SUSPENDED(td) || prio >= PSOCK) 1901 td->td_flags |= TDF_CANSWAP; 1902 if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE) 1903 return; 1904 if (static_boost == 1 && prio) 1905 sched_prio(td, prio); 1906 else if (static_boost && td->td_priority > static_boost) 1907 sched_prio(td, static_boost); 1908 } 1909 1910 /* 1911 * Schedule a thread to resume execution and record how long it voluntarily 1912 * slept. We also update the pctcpu, interactivity, and priority. 1913 */ 1914 void 1915 sched_wakeup(struct thread *td) 1916 { 1917 struct td_sched *ts; 1918 int slptick; 1919 1920 THREAD_LOCK_ASSERT(td, MA_OWNED); 1921 ts = td->td_sched; 1922 td->td_flags &= ~TDF_CANSWAP; 1923 /* 1924 * If we slept for more than a tick update our interactivity and 1925 * priority. 1926 */ 1927 slptick = td->td_slptick; 1928 td->td_slptick = 0; 1929 if (slptick && slptick != ticks) { 1930 u_int hzticks; 1931 1932 hzticks = (ticks - slptick) << SCHED_TICK_SHIFT; 1933 ts->ts_slptime += hzticks; 1934 sched_interact_update(td); 1935 sched_pctcpu_update(ts); 1936 } 1937 /* Reset the slice value after we sleep. */ 1938 ts->ts_slice = sched_slice; 1939 sched_add(td, SRQ_BORING); 1940 } 1941 1942 /* 1943 * Penalize the parent for creating a new child and initialize the child's 1944 * priority. 1945 */ 1946 void 1947 sched_fork(struct thread *td, struct thread *child) 1948 { 1949 THREAD_LOCK_ASSERT(td, MA_OWNED); 1950 sched_fork_thread(td, child); 1951 /* 1952 * Penalize the parent and child for forking. 1953 */ 1954 sched_interact_fork(child); 1955 sched_priority(child); 1956 td->td_sched->ts_runtime += tickincr; 1957 sched_interact_update(td); 1958 sched_priority(td); 1959 } 1960 1961 /* 1962 * Fork a new thread, may be within the same process. 1963 */ 1964 void 1965 sched_fork_thread(struct thread *td, struct thread *child) 1966 { 1967 struct td_sched *ts; 1968 struct td_sched *ts2; 1969 1970 THREAD_LOCK_ASSERT(td, MA_OWNED); 1971 /* 1972 * Initialize child. 1973 */ 1974 ts = td->td_sched; 1975 ts2 = child->td_sched; 1976 child->td_lock = TDQ_LOCKPTR(TDQ_SELF()); 1977 child->td_cpuset = cpuset_ref(td->td_cpuset); 1978 ts2->ts_cpu = ts->ts_cpu; 1979 ts2->ts_flags = 0; 1980 /* 1981 * Grab our parents cpu estimation information. 1982 */ 1983 ts2->ts_ticks = ts->ts_ticks; 1984 ts2->ts_ltick = ts->ts_ltick; 1985 ts2->ts_incrtick = ts->ts_incrtick; 1986 ts2->ts_ftick = ts->ts_ftick; 1987 /* 1988 * Do not inherit any borrowed priority from the parent. 1989 */ 1990 child->td_priority = child->td_base_pri; 1991 /* 1992 * And update interactivity score. 1993 */ 1994 ts2->ts_slptime = ts->ts_slptime; 1995 ts2->ts_runtime = ts->ts_runtime; 1996 ts2->ts_slice = 1; /* Attempt to quickly learn interactivity. */ 1997 #ifdef KTR 1998 bzero(ts2->ts_name, sizeof(ts2->ts_name)); 1999 #endif 2000 } 2001 2002 /* 2003 * Adjust the priority class of a thread. 2004 */ 2005 void 2006 sched_class(struct thread *td, int class) 2007 { 2008 2009 THREAD_LOCK_ASSERT(td, MA_OWNED); 2010 if (td->td_pri_class == class) 2011 return; 2012 td->td_pri_class = class; 2013 } 2014 2015 /* 2016 * Return some of the child's priority and interactivity to the parent. 2017 */ 2018 void 2019 sched_exit(struct proc *p, struct thread *child) 2020 { 2021 struct thread *td; 2022 2023 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit", 2024 "prio:td", child->td_priority); 2025 PROC_LOCK_ASSERT(p, MA_OWNED); 2026 td = FIRST_THREAD_IN_PROC(p); 2027 sched_exit_thread(td, child); 2028 } 2029 2030 /* 2031 * Penalize another thread for the time spent on this one. This helps to 2032 * worsen the priority and interactivity of processes which schedule batch 2033 * jobs such as make. This has little effect on the make process itself but 2034 * causes new processes spawned by it to receive worse scores immediately. 2035 */ 2036 void 2037 sched_exit_thread(struct thread *td, struct thread *child) 2038 { 2039 2040 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit", 2041 "prio:td", child->td_priority); 2042 /* 2043 * Give the child's runtime to the parent without returning the 2044 * sleep time as a penalty to the parent. This causes shells that 2045 * launch expensive things to mark their children as expensive. 2046 */ 2047 thread_lock(td); 2048 td->td_sched->ts_runtime += child->td_sched->ts_runtime; 2049 sched_interact_update(td); 2050 sched_priority(td); 2051 thread_unlock(td); 2052 } 2053 2054 void 2055 sched_preempt(struct thread *td) 2056 { 2057 struct tdq *tdq; 2058 2059 thread_lock(td); 2060 tdq = TDQ_SELF(); 2061 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2062 tdq->tdq_ipipending = 0; 2063 if (td->td_priority > tdq->tdq_lowpri) { 2064 int flags; 2065 2066 flags = SW_INVOL | SW_PREEMPT; 2067 if (td->td_critnest > 1) 2068 td->td_owepreempt = 1; 2069 else if (TD_IS_IDLETHREAD(td)) 2070 mi_switch(flags | SWT_REMOTEWAKEIDLE, NULL); 2071 else 2072 mi_switch(flags | SWT_REMOTEPREEMPT, NULL); 2073 } 2074 thread_unlock(td); 2075 } 2076 2077 /* 2078 * Fix priorities on return to user-space. Priorities may be elevated due 2079 * to static priorities in msleep() or similar. 2080 */ 2081 void 2082 sched_userret(struct thread *td) 2083 { 2084 /* 2085 * XXX we cheat slightly on the locking here to avoid locking in 2086 * the usual case. Setting td_priority here is essentially an 2087 * incomplete workaround for not setting it properly elsewhere. 2088 * Now that some interrupt handlers are threads, not setting it 2089 * properly elsewhere can clobber it in the window between setting 2090 * it here and returning to user mode, so don't waste time setting 2091 * it perfectly here. 2092 */ 2093 KASSERT((td->td_flags & TDF_BORROWING) == 0, 2094 ("thread with borrowed priority returning to userland")); 2095 if (td->td_priority != td->td_user_pri) { 2096 thread_lock(td); 2097 td->td_priority = td->td_user_pri; 2098 td->td_base_pri = td->td_user_pri; 2099 tdq_setlowpri(TDQ_SELF(), td); 2100 thread_unlock(td); 2101 } 2102 } 2103 2104 /* 2105 * Handle a stathz tick. This is really only relevant for timeshare 2106 * threads. 2107 */ 2108 void 2109 sched_clock(struct thread *td) 2110 { 2111 struct tdq *tdq; 2112 struct td_sched *ts; 2113 2114 THREAD_LOCK_ASSERT(td, MA_OWNED); 2115 tdq = TDQ_SELF(); 2116 #ifdef SMP 2117 /* 2118 * We run the long term load balancer infrequently on the first cpu. 2119 */ 2120 if (balance_tdq == tdq) { 2121 if (balance_ticks && --balance_ticks == 0) 2122 sched_balance(); 2123 } 2124 #endif 2125 /* 2126 * Save the old switch count so we have a record of the last ticks 2127 * activity. Initialize the new switch count based on our load. 2128 * If there is some activity seed it to reflect that. 2129 */ 2130 tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt; 2131 tdq->tdq_switchcnt = tdq->tdq_load; 2132 /* 2133 * Advance the insert index once for each tick to ensure that all 2134 * threads get a chance to run. 2135 */ 2136 if (tdq->tdq_idx == tdq->tdq_ridx) { 2137 tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS; 2138 if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx])) 2139 tdq->tdq_ridx = tdq->tdq_idx; 2140 } 2141 ts = td->td_sched; 2142 if (td->td_pri_class & PRI_FIFO_BIT) 2143 return; 2144 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) { 2145 /* 2146 * We used a tick; charge it to the thread so 2147 * that we can compute our interactivity. 2148 */ 2149 td->td_sched->ts_runtime += tickincr; 2150 sched_interact_update(td); 2151 sched_priority(td); 2152 } 2153 /* 2154 * We used up one time slice. 2155 */ 2156 if (--ts->ts_slice > 0) 2157 return; 2158 /* 2159 * We're out of time, force a requeue at userret(). 2160 */ 2161 ts->ts_slice = sched_slice; 2162 td->td_flags |= TDF_NEEDRESCHED; 2163 } 2164 2165 /* 2166 * Called once per hz tick. Used for cpu utilization information. This 2167 * is easier than trying to scale based on stathz. 2168 */ 2169 void 2170 sched_tick(int cnt) 2171 { 2172 struct td_sched *ts; 2173 2174 ts = curthread->td_sched; 2175 /* 2176 * Ticks is updated asynchronously on a single cpu. Check here to 2177 * avoid incrementing ts_ticks multiple times in a single tick. 2178 */ 2179 if (ts->ts_incrtick == ticks) 2180 return; 2181 /* Adjust ticks for pctcpu */ 2182 ts->ts_ticks += cnt << SCHED_TICK_SHIFT; 2183 ts->ts_ltick = ticks; 2184 ts->ts_incrtick = ticks; 2185 /* 2186 * Update if we've exceeded our desired tick threshold by over one 2187 * second. 2188 */ 2189 if (ts->ts_ftick + SCHED_TICK_MAX < ts->ts_ltick) 2190 sched_pctcpu_update(ts); 2191 } 2192 2193 /* 2194 * Return whether the current CPU has runnable tasks. Used for in-kernel 2195 * cooperative idle threads. 2196 */ 2197 int 2198 sched_runnable(void) 2199 { 2200 struct tdq *tdq; 2201 int load; 2202 2203 load = 1; 2204 2205 tdq = TDQ_SELF(); 2206 if ((curthread->td_flags & TDF_IDLETD) != 0) { 2207 if (tdq->tdq_load > 0) 2208 goto out; 2209 } else 2210 if (tdq->tdq_load - 1 > 0) 2211 goto out; 2212 load = 0; 2213 out: 2214 return (load); 2215 } 2216 2217 /* 2218 * Choose the highest priority thread to run. The thread is removed from 2219 * the run-queue while running however the load remains. For SMP we set 2220 * the tdq in the global idle bitmask if it idles here. 2221 */ 2222 struct thread * 2223 sched_choose(void) 2224 { 2225 struct thread *td; 2226 struct tdq *tdq; 2227 2228 tdq = TDQ_SELF(); 2229 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2230 td = tdq_choose(tdq); 2231 if (td) { 2232 td->td_sched->ts_ltick = ticks; 2233 tdq_runq_rem(tdq, td); 2234 tdq->tdq_lowpri = td->td_priority; 2235 return (td); 2236 } 2237 tdq->tdq_lowpri = PRI_MAX_IDLE; 2238 return (PCPU_GET(idlethread)); 2239 } 2240 2241 /* 2242 * Set owepreempt if necessary. Preemption never happens directly in ULE, 2243 * we always request it once we exit a critical section. 2244 */ 2245 static inline void 2246 sched_setpreempt(struct thread *td) 2247 { 2248 struct thread *ctd; 2249 int cpri; 2250 int pri; 2251 2252 THREAD_LOCK_ASSERT(curthread, MA_OWNED); 2253 2254 ctd = curthread; 2255 pri = td->td_priority; 2256 cpri = ctd->td_priority; 2257 if (pri < cpri) 2258 ctd->td_flags |= TDF_NEEDRESCHED; 2259 if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd)) 2260 return; 2261 if (!sched_shouldpreempt(pri, cpri, 0)) 2262 return; 2263 ctd->td_owepreempt = 1; 2264 } 2265 2266 /* 2267 * Add a thread to a thread queue. Select the appropriate runq and add the 2268 * thread to it. This is the internal function called when the tdq is 2269 * predetermined. 2270 */ 2271 void 2272 tdq_add(struct tdq *tdq, struct thread *td, int flags) 2273 { 2274 2275 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2276 KASSERT((td->td_inhibitors == 0), 2277 ("sched_add: trying to run inhibited thread")); 2278 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)), 2279 ("sched_add: bad thread state")); 2280 KASSERT(td->td_flags & TDF_INMEM, 2281 ("sched_add: thread swapped out")); 2282 2283 if (td->td_priority < tdq->tdq_lowpri) 2284 tdq->tdq_lowpri = td->td_priority; 2285 tdq_runq_add(tdq, td, flags); 2286 tdq_load_add(tdq, td); 2287 } 2288 2289 /* 2290 * Select the target thread queue and add a thread to it. Request 2291 * preemption or IPI a remote processor if required. 2292 */ 2293 void 2294 sched_add(struct thread *td, int flags) 2295 { 2296 struct tdq *tdq; 2297 #ifdef SMP 2298 int cpu; 2299 #endif 2300 2301 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add", 2302 "prio:%d", td->td_priority, KTR_ATTR_LINKED, 2303 sched_tdname(curthread)); 2304 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup", 2305 KTR_ATTR_LINKED, sched_tdname(td)); 2306 THREAD_LOCK_ASSERT(td, MA_OWNED); 2307 /* 2308 * Recalculate the priority before we select the target cpu or 2309 * run-queue. 2310 */ 2311 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) 2312 sched_priority(td); 2313 #ifdef SMP 2314 /* 2315 * Pick the destination cpu and if it isn't ours transfer to the 2316 * target cpu. 2317 */ 2318 cpu = sched_pickcpu(td, flags); 2319 tdq = sched_setcpu(td, cpu, flags); 2320 tdq_add(tdq, td, flags); 2321 if (cpu != PCPU_GET(cpuid)) { 2322 tdq_notify(tdq, td); 2323 return; 2324 } 2325 #else 2326 tdq = TDQ_SELF(); 2327 TDQ_LOCK(tdq); 2328 /* 2329 * Now that the thread is moving to the run-queue, set the lock 2330 * to the scheduler's lock. 2331 */ 2332 thread_lock_set(td, TDQ_LOCKPTR(tdq)); 2333 tdq_add(tdq, td, flags); 2334 #endif 2335 if (!(flags & SRQ_YIELDING)) 2336 sched_setpreempt(td); 2337 } 2338 2339 /* 2340 * Remove a thread from a run-queue without running it. This is used 2341 * when we're stealing a thread from a remote queue. Otherwise all threads 2342 * exit by calling sched_exit_thread() and sched_throw() themselves. 2343 */ 2344 void 2345 sched_rem(struct thread *td) 2346 { 2347 struct tdq *tdq; 2348 2349 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem", 2350 "prio:%d", td->td_priority); 2351 tdq = TDQ_CPU(td->td_sched->ts_cpu); 2352 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2353 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2354 KASSERT(TD_ON_RUNQ(td), 2355 ("sched_rem: thread not on run queue")); 2356 tdq_runq_rem(tdq, td); 2357 tdq_load_rem(tdq, td); 2358 TD_SET_CAN_RUN(td); 2359 if (td->td_priority == tdq->tdq_lowpri) 2360 tdq_setlowpri(tdq, NULL); 2361 } 2362 2363 /* 2364 * Fetch cpu utilization information. Updates on demand. 2365 */ 2366 fixpt_t 2367 sched_pctcpu(struct thread *td) 2368 { 2369 fixpt_t pctcpu; 2370 struct td_sched *ts; 2371 2372 pctcpu = 0; 2373 ts = td->td_sched; 2374 if (ts == NULL) 2375 return (0); 2376 2377 THREAD_LOCK_ASSERT(td, MA_OWNED); 2378 if (ts->ts_ticks) { 2379 int rtick; 2380 2381 sched_pctcpu_update(ts); 2382 /* How many rtick per second ? */ 2383 rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz); 2384 pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT; 2385 } 2386 2387 return (pctcpu); 2388 } 2389 2390 /* 2391 * Enforce affinity settings for a thread. Called after adjustments to 2392 * cpumask. 2393 */ 2394 void 2395 sched_affinity(struct thread *td) 2396 { 2397 #ifdef SMP 2398 struct td_sched *ts; 2399 2400 THREAD_LOCK_ASSERT(td, MA_OWNED); 2401 ts = td->td_sched; 2402 if (THREAD_CAN_SCHED(td, ts->ts_cpu)) 2403 return; 2404 if (TD_ON_RUNQ(td)) { 2405 sched_rem(td); 2406 sched_add(td, SRQ_BORING); 2407 return; 2408 } 2409 if (!TD_IS_RUNNING(td)) 2410 return; 2411 /* 2412 * Force a switch before returning to userspace. If the 2413 * target thread is not running locally send an ipi to force 2414 * the issue. 2415 */ 2416 td->td_flags |= TDF_NEEDRESCHED; 2417 if (td != curthread) 2418 ipi_cpu(ts->ts_cpu, IPI_PREEMPT); 2419 #endif 2420 } 2421 2422 /* 2423 * Bind a thread to a target cpu. 2424 */ 2425 void 2426 sched_bind(struct thread *td, int cpu) 2427 { 2428 struct td_sched *ts; 2429 2430 THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED); 2431 KASSERT(td == curthread, ("sched_bind: can only bind curthread")); 2432 ts = td->td_sched; 2433 if (ts->ts_flags & TSF_BOUND) 2434 sched_unbind(td); 2435 KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td)); 2436 ts->ts_flags |= TSF_BOUND; 2437 sched_pin(); 2438 if (PCPU_GET(cpuid) == cpu) 2439 return; 2440 ts->ts_cpu = cpu; 2441 /* When we return from mi_switch we'll be on the correct cpu. */ 2442 mi_switch(SW_VOL, NULL); 2443 } 2444 2445 /* 2446 * Release a bound thread. 2447 */ 2448 void 2449 sched_unbind(struct thread *td) 2450 { 2451 struct td_sched *ts; 2452 2453 THREAD_LOCK_ASSERT(td, MA_OWNED); 2454 KASSERT(td == curthread, ("sched_unbind: can only bind curthread")); 2455 ts = td->td_sched; 2456 if ((ts->ts_flags & TSF_BOUND) == 0) 2457 return; 2458 ts->ts_flags &= ~TSF_BOUND; 2459 sched_unpin(); 2460 } 2461 2462 int 2463 sched_is_bound(struct thread *td) 2464 { 2465 THREAD_LOCK_ASSERT(td, MA_OWNED); 2466 return (td->td_sched->ts_flags & TSF_BOUND); 2467 } 2468 2469 /* 2470 * Basic yield call. 2471 */ 2472 void 2473 sched_relinquish(struct thread *td) 2474 { 2475 thread_lock(td); 2476 mi_switch(SW_VOL | SWT_RELINQUISH, NULL); 2477 thread_unlock(td); 2478 } 2479 2480 /* 2481 * Return the total system load. 2482 */ 2483 int 2484 sched_load(void) 2485 { 2486 #ifdef SMP 2487 int total; 2488 int i; 2489 2490 total = 0; 2491 CPU_FOREACH(i) 2492 total += TDQ_CPU(i)->tdq_sysload; 2493 return (total); 2494 #else 2495 return (TDQ_SELF()->tdq_sysload); 2496 #endif 2497 } 2498 2499 int 2500 sched_sizeof_proc(void) 2501 { 2502 return (sizeof(struct proc)); 2503 } 2504 2505 int 2506 sched_sizeof_thread(void) 2507 { 2508 return (sizeof(struct thread) + sizeof(struct td_sched)); 2509 } 2510 2511 #ifdef SMP 2512 #define TDQ_IDLESPIN(tdq) \ 2513 ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0) 2514 #else 2515 #define TDQ_IDLESPIN(tdq) 1 2516 #endif 2517 2518 /* 2519 * The actual idle process. 2520 */ 2521 void 2522 sched_idletd(void *dummy) 2523 { 2524 struct thread *td; 2525 struct tdq *tdq; 2526 int switchcnt; 2527 int i; 2528 2529 mtx_assert(&Giant, MA_NOTOWNED); 2530 td = curthread; 2531 tdq = TDQ_SELF(); 2532 for (;;) { 2533 #ifdef SMP 2534 if (tdq_idled(tdq) == 0) 2535 continue; 2536 #endif 2537 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt; 2538 /* 2539 * If we're switching very frequently, spin while checking 2540 * for load rather than entering a low power state that 2541 * may require an IPI. However, don't do any busy 2542 * loops while on SMT machines as this simply steals 2543 * cycles from cores doing useful work. 2544 */ 2545 if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) { 2546 for (i = 0; i < sched_idlespins; i++) { 2547 if (tdq->tdq_load) 2548 break; 2549 cpu_spinwait(); 2550 } 2551 } 2552 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt; 2553 if (tdq->tdq_load == 0) { 2554 tdq->tdq_cpu_idle = 1; 2555 if (tdq->tdq_load == 0) { 2556 cpu_idle(switchcnt > sched_idlespinthresh * 4); 2557 tdq->tdq_switchcnt++; 2558 } 2559 tdq->tdq_cpu_idle = 0; 2560 } 2561 if (tdq->tdq_load) { 2562 thread_lock(td); 2563 mi_switch(SW_VOL | SWT_IDLE, NULL); 2564 thread_unlock(td); 2565 } 2566 } 2567 } 2568 2569 /* 2570 * A CPU is entering for the first time or a thread is exiting. 2571 */ 2572 void 2573 sched_throw(struct thread *td) 2574 { 2575 struct thread *newtd; 2576 struct tdq *tdq; 2577 2578 tdq = TDQ_SELF(); 2579 if (td == NULL) { 2580 /* Correct spinlock nesting and acquire the correct lock. */ 2581 TDQ_LOCK(tdq); 2582 spinlock_exit(); 2583 } else { 2584 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2585 tdq_load_rem(tdq, td); 2586 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object); 2587 } 2588 KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count")); 2589 newtd = choosethread(); 2590 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd; 2591 PCPU_SET(switchtime, cpu_ticks()); 2592 PCPU_SET(switchticks, ticks); 2593 cpu_throw(td, newtd); /* doesn't return */ 2594 } 2595 2596 /* 2597 * This is called from fork_exit(). Just acquire the correct locks and 2598 * let fork do the rest of the work. 2599 */ 2600 void 2601 sched_fork_exit(struct thread *td) 2602 { 2603 struct td_sched *ts; 2604 struct tdq *tdq; 2605 int cpuid; 2606 2607 /* 2608 * Finish setting up thread glue so that it begins execution in a 2609 * non-nested critical section with the scheduler lock held. 2610 */ 2611 cpuid = PCPU_GET(cpuid); 2612 tdq = TDQ_CPU(cpuid); 2613 ts = td->td_sched; 2614 if (TD_IS_IDLETHREAD(td)) 2615 td->td_lock = TDQ_LOCKPTR(tdq); 2616 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2617 td->td_oncpu = cpuid; 2618 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED); 2619 lock_profile_obtain_lock_success( 2620 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__); 2621 } 2622 2623 /* 2624 * Create on first use to catch odd startup conditons. 2625 */ 2626 char * 2627 sched_tdname(struct thread *td) 2628 { 2629 #ifdef KTR 2630 struct td_sched *ts; 2631 2632 ts = td->td_sched; 2633 if (ts->ts_name[0] == '\0') 2634 snprintf(ts->ts_name, sizeof(ts->ts_name), 2635 "%s tid %d", td->td_name, td->td_tid); 2636 return (ts->ts_name); 2637 #else 2638 return (td->td_name); 2639 #endif 2640 } 2641 2642 #ifdef SMP 2643 2644 /* 2645 * Build the CPU topology dump string. Is recursively called to collect 2646 * the topology tree. 2647 */ 2648 static int 2649 sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg, 2650 int indent) 2651 { 2652 int i, first; 2653 2654 sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent, 2655 "", 1 + indent / 2, cg->cg_level); 2656 sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"0x%x\">", indent, "", 2657 cg->cg_count, cg->cg_mask); 2658 first = TRUE; 2659 for (i = 0; i < MAXCPU; i++) { 2660 if ((cg->cg_mask & (1 << i)) != 0) { 2661 if (!first) 2662 sbuf_printf(sb, ", "); 2663 else 2664 first = FALSE; 2665 sbuf_printf(sb, "%d", i); 2666 } 2667 } 2668 sbuf_printf(sb, "</cpu>\n"); 2669 2670 if (cg->cg_flags != 0) { 2671 sbuf_printf(sb, "%*s <flags>", indent, ""); 2672 if ((cg->cg_flags & CG_FLAG_HTT) != 0) 2673 sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>"); 2674 if ((cg->cg_flags & CG_FLAG_THREAD) != 0) 2675 sbuf_printf(sb, "<flag name=\"THREAD\">THREAD group</flag>"); 2676 if ((cg->cg_flags & CG_FLAG_SMT) != 0) 2677 sbuf_printf(sb, "<flag name=\"SMT\">SMT group</flag>"); 2678 sbuf_printf(sb, "</flags>\n"); 2679 } 2680 2681 if (cg->cg_children > 0) { 2682 sbuf_printf(sb, "%*s <children>\n", indent, ""); 2683 for (i = 0; i < cg->cg_children; i++) 2684 sysctl_kern_sched_topology_spec_internal(sb, 2685 &cg->cg_child[i], indent+2); 2686 sbuf_printf(sb, "%*s </children>\n", indent, ""); 2687 } 2688 sbuf_printf(sb, "%*s</group>\n", indent, ""); 2689 return (0); 2690 } 2691 2692 /* 2693 * Sysctl handler for retrieving topology dump. It's a wrapper for 2694 * the recursive sysctl_kern_smp_topology_spec_internal(). 2695 */ 2696 static int 2697 sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS) 2698 { 2699 struct sbuf *topo; 2700 int err; 2701 2702 KASSERT(cpu_top != NULL, ("cpu_top isn't initialized")); 2703 2704 topo = sbuf_new(NULL, NULL, 500, SBUF_AUTOEXTEND); 2705 if (topo == NULL) 2706 return (ENOMEM); 2707 2708 sbuf_printf(topo, "<groups>\n"); 2709 err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1); 2710 sbuf_printf(topo, "</groups>\n"); 2711 2712 if (err == 0) { 2713 sbuf_finish(topo); 2714 err = SYSCTL_OUT(req, sbuf_data(topo), sbuf_len(topo)); 2715 } 2716 sbuf_delete(topo); 2717 return (err); 2718 } 2719 2720 #endif 2721 2722 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler"); 2723 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0, 2724 "Scheduler name"); 2725 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0, 2726 "Slice size for timeshare threads"); 2727 SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0, 2728 "Interactivity score threshold"); 2729 SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW, &preempt_thresh, 2730 0,"Min priority for preemption, lower priorities have greater precedence"); 2731 SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost, 2732 0,"Controls whether static kernel priorities are assigned to sleeping threads."); 2733 SYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins, 2734 0,"Number of times idle will spin waiting for new work."); 2735 SYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW, &sched_idlespinthresh, 2736 0,"Threshold before we will permit idle spinning."); 2737 #ifdef SMP 2738 SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0, 2739 "Number of hz ticks to keep thread affinity for"); 2740 SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0, 2741 "Enables the long-term load balancer"); 2742 SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW, 2743 &balance_interval, 0, 2744 "Average frequency in stathz ticks to run the long-term balancer"); 2745 SYSCTL_INT(_kern_sched, OID_AUTO, steal_htt, CTLFLAG_RW, &steal_htt, 0, 2746 "Steals work from another hyper-threaded core on idle"); 2747 SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0, 2748 "Attempts to steal work from other cores before idling"); 2749 SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0, 2750 "Minimum load on remote cpu before we'll steal"); 2751 2752 /* Retrieve SMP topology */ 2753 SYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING | 2754 CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A", 2755 "XML dump of detected CPU topology"); 2756 2757 #endif 2758 2759 /* ps compat. All cpu percentages from ULE are weighted. */ 2760 static int ccpu = 0; 2761 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 2762