1 /*- 2 * Copyright (c) 2002-2003, 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 #include <sys/cdefs.h> 28 __FBSDID("$FreeBSD$"); 29 30 #include <sys/param.h> 31 #include <sys/systm.h> 32 #include <sys/kernel.h> 33 #include <sys/ktr.h> 34 #include <sys/lock.h> 35 #include <sys/mutex.h> 36 #include <sys/proc.h> 37 #include <sys/resource.h> 38 #include <sys/sched.h> 39 #include <sys/smp.h> 40 #include <sys/sx.h> 41 #include <sys/sysctl.h> 42 #include <sys/sysproto.h> 43 #include <sys/vmmeter.h> 44 #ifdef DDB 45 #include <ddb/ddb.h> 46 #endif 47 #ifdef KTRACE 48 #include <sys/uio.h> 49 #include <sys/ktrace.h> 50 #endif 51 52 #include <machine/cpu.h> 53 54 #define KTR_ULE KTR_NFS 55 56 /* decay 95% of `p_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */ 57 /* XXX This is bogus compatability crap for ps */ 58 static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */ 59 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 60 61 static void sched_setup(void *dummy); 62 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL) 63 64 static SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "SCHED"); 65 66 static int sched_strict; 67 SYSCTL_INT(_kern_sched, OID_AUTO, strict, CTLFLAG_RD, &sched_strict, 0, ""); 68 69 static int slice_min = 1; 70 SYSCTL_INT(_kern_sched, OID_AUTO, slice_min, CTLFLAG_RW, &slice_min, 0, ""); 71 72 static int slice_max = 10; 73 SYSCTL_INT(_kern_sched, OID_AUTO, slice_max, CTLFLAG_RW, &slice_max, 0, ""); 74 75 int realstathz; 76 int tickincr = 1; 77 78 #ifdef SMP 79 /* Callout to handle load balancing SMP systems. */ 80 static struct callout kseq_lb_callout; 81 #endif 82 83 /* 84 * These datastructures are allocated within their parent datastructure but 85 * are scheduler specific. 86 */ 87 88 struct ke_sched { 89 int ske_slice; 90 struct runq *ske_runq; 91 /* The following variables are only used for pctcpu calculation */ 92 int ske_ltick; /* Last tick that we were running on */ 93 int ske_ftick; /* First tick that we were running on */ 94 int ske_ticks; /* Tick count */ 95 /* CPU that we have affinity for. */ 96 u_char ske_cpu; 97 }; 98 #define ke_slice ke_sched->ske_slice 99 #define ke_runq ke_sched->ske_runq 100 #define ke_ltick ke_sched->ske_ltick 101 #define ke_ftick ke_sched->ske_ftick 102 #define ke_ticks ke_sched->ske_ticks 103 #define ke_cpu ke_sched->ske_cpu 104 105 struct kg_sched { 106 int skg_slptime; /* Number of ticks we vol. slept */ 107 int skg_runtime; /* Number of ticks we were running */ 108 }; 109 #define kg_slptime kg_sched->skg_slptime 110 #define kg_runtime kg_sched->skg_runtime 111 112 struct td_sched { 113 int std_slptime; 114 }; 115 #define td_slptime td_sched->std_slptime 116 117 struct td_sched td_sched; 118 struct ke_sched ke_sched; 119 struct kg_sched kg_sched; 120 121 struct ke_sched *kse0_sched = &ke_sched; 122 struct kg_sched *ksegrp0_sched = &kg_sched; 123 struct p_sched *proc0_sched = NULL; 124 struct td_sched *thread0_sched = &td_sched; 125 126 /* 127 * The priority is primarily determined by the interactivity score. Thus, we 128 * give lower(better) priorities to kse groups that use less CPU. The nice 129 * value is then directly added to this to allow nice to have some effect 130 * on latency. 131 * 132 * PRI_RANGE: Total priority range for timeshare threads. 133 * PRI_NRESV: Number of nice values. 134 * PRI_BASE: The start of the dynamic range. 135 */ 136 #define SCHED_PRI_RANGE (PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE + 1) 137 #define SCHED_PRI_NRESV PRIO_TOTAL 138 #define SCHED_PRI_NHALF (PRIO_TOTAL / 2) 139 #define SCHED_PRI_NTHRESH (SCHED_PRI_NHALF - 1) 140 #define SCHED_PRI_BASE (PRI_MIN_TIMESHARE) 141 #define SCHED_PRI_INTERACT(score) \ 142 ((score) * SCHED_PRI_RANGE / SCHED_INTERACT_MAX) 143 144 /* 145 * These determine the interactivity of a process. 146 * 147 * SLP_RUN_MAX: Maximum amount of sleep time + run time we'll accumulate 148 * before throttling back. 149 * SLP_RUN_THROTTLE: Divisor for reducing slp/run time at fork time. 150 * INTERACT_MAX: Maximum interactivity value. Smaller is better. 151 * INTERACT_THRESH: Threshhold for placement on the current runq. 152 */ 153 #define SCHED_SLP_RUN_MAX ((hz * 2) << 10) 154 #define SCHED_SLP_RUN_THROTTLE (100) 155 #define SCHED_INTERACT_MAX (100) 156 #define SCHED_INTERACT_HALF (SCHED_INTERACT_MAX / 2) 157 #define SCHED_INTERACT_THRESH (20) 158 159 /* 160 * These parameters and macros determine the size of the time slice that is 161 * granted to each thread. 162 * 163 * SLICE_MIN: Minimum time slice granted, in units of ticks. 164 * SLICE_MAX: Maximum time slice granted. 165 * SLICE_RANGE: Range of available time slices scaled by hz. 166 * SLICE_SCALE: The number slices granted per val in the range of [0, max]. 167 * SLICE_NICE: Determine the amount of slice granted to a scaled nice. 168 */ 169 #define SCHED_SLICE_MIN (slice_min) 170 #define SCHED_SLICE_MAX (slice_max) 171 #define SCHED_SLICE_RANGE (SCHED_SLICE_MAX - SCHED_SLICE_MIN + 1) 172 #define SCHED_SLICE_SCALE(val, max) (((val) * SCHED_SLICE_RANGE) / (max)) 173 #define SCHED_SLICE_NICE(nice) \ 174 (SCHED_SLICE_MAX - SCHED_SLICE_SCALE((nice), SCHED_PRI_NTHRESH)) 175 176 /* 177 * This macro determines whether or not the kse belongs on the current or 178 * next run queue. 179 * 180 * XXX nice value should effect how interactive a kg is. 181 */ 182 #define SCHED_INTERACTIVE(kg) \ 183 (sched_interact_score(kg) < SCHED_INTERACT_THRESH) 184 #define SCHED_CURR(kg, ke) \ 185 (ke->ke_thread->td_priority < PRI_MIN_TIMESHARE || SCHED_INTERACTIVE(kg)) 186 187 /* 188 * Cpu percentage computation macros and defines. 189 * 190 * SCHED_CPU_TIME: Number of seconds to average the cpu usage across. 191 * SCHED_CPU_TICKS: Number of hz ticks to average the cpu usage across. 192 */ 193 194 #define SCHED_CPU_TIME 10 195 #define SCHED_CPU_TICKS (hz * SCHED_CPU_TIME) 196 197 /* 198 * kseq - per processor runqs and statistics. 199 */ 200 201 #define KSEQ_NCLASS (PRI_IDLE + 1) /* Number of run classes. */ 202 203 struct kseq { 204 struct runq ksq_idle; /* Queue of IDLE threads. */ 205 struct runq ksq_timeshare[2]; /* Run queues for !IDLE. */ 206 struct runq *ksq_next; /* Next timeshare queue. */ 207 struct runq *ksq_curr; /* Current queue. */ 208 int ksq_loads[KSEQ_NCLASS]; /* Load for each class */ 209 int ksq_load; /* Aggregate load. */ 210 short ksq_nice[PRIO_TOTAL + 1]; /* KSEs in each nice bin. */ 211 short ksq_nicemin; /* Least nice. */ 212 #ifdef SMP 213 int ksq_cpus; /* Count of CPUs in this kseq. */ 214 unsigned int ksq_rslices; /* Slices on run queue */ 215 #endif 216 }; 217 218 /* 219 * One kse queue per processor. 220 */ 221 #ifdef SMP 222 struct kseq kseq_cpu[MAXCPU]; 223 struct kseq *kseq_idmap[MAXCPU]; 224 #define KSEQ_SELF() (kseq_idmap[PCPU_GET(cpuid)]) 225 #define KSEQ_CPU(x) (kseq_idmap[(x)]) 226 #else 227 struct kseq kseq_cpu; 228 #define KSEQ_SELF() (&kseq_cpu) 229 #define KSEQ_CPU(x) (&kseq_cpu) 230 #endif 231 232 static void sched_slice(struct kse *ke); 233 static void sched_priority(struct ksegrp *kg); 234 static int sched_interact_score(struct ksegrp *kg); 235 static void sched_interact_update(struct ksegrp *kg); 236 void sched_pctcpu_update(struct kse *ke); 237 int sched_pickcpu(void); 238 239 /* Operations on per processor queues */ 240 static struct kse * kseq_choose(struct kseq *kseq); 241 static void kseq_setup(struct kseq *kseq); 242 static void kseq_add(struct kseq *kseq, struct kse *ke); 243 static void kseq_rem(struct kseq *kseq, struct kse *ke); 244 static void kseq_nice_add(struct kseq *kseq, int nice); 245 static void kseq_nice_rem(struct kseq *kseq, int nice); 246 void kseq_print(int cpu); 247 #ifdef SMP 248 struct kseq * kseq_load_highest(void); 249 void kseq_balance(void *arg); 250 void kseq_move(struct kseq *from, int cpu); 251 #endif 252 253 void 254 kseq_print(int cpu) 255 { 256 struct kseq *kseq; 257 int i; 258 259 kseq = KSEQ_CPU(cpu); 260 261 printf("kseq:\n"); 262 printf("\tload: %d\n", kseq->ksq_load); 263 printf("\tload ITHD: %d\n", kseq->ksq_loads[PRI_ITHD]); 264 printf("\tload REALTIME: %d\n", kseq->ksq_loads[PRI_REALTIME]); 265 printf("\tload TIMESHARE: %d\n", kseq->ksq_loads[PRI_TIMESHARE]); 266 printf("\tload IDLE: %d\n", kseq->ksq_loads[PRI_IDLE]); 267 printf("\tnicemin:\t%d\n", kseq->ksq_nicemin); 268 printf("\tnice counts:\n"); 269 for (i = 0; i < PRIO_TOTAL + 1; i++) 270 if (kseq->ksq_nice[i]) 271 printf("\t\t%d = %d\n", 272 i - SCHED_PRI_NHALF, kseq->ksq_nice[i]); 273 } 274 275 static void 276 kseq_add(struct kseq *kseq, struct kse *ke) 277 { 278 mtx_assert(&sched_lock, MA_OWNED); 279 kseq->ksq_loads[PRI_BASE(ke->ke_ksegrp->kg_pri_class)]++; 280 kseq->ksq_load++; 281 if (ke->ke_ksegrp->kg_pri_class == PRI_TIMESHARE) 282 CTR6(KTR_ULE, "Add kse %p to %p (slice: %d, pri: %d, nice: %d(%d))", 283 ke, ke->ke_runq, ke->ke_slice, ke->ke_thread->td_priority, 284 ke->ke_ksegrp->kg_nice, kseq->ksq_nicemin); 285 if (ke->ke_ksegrp->kg_pri_class == PRI_TIMESHARE) 286 kseq_nice_add(kseq, ke->ke_ksegrp->kg_nice); 287 #ifdef SMP 288 kseq->ksq_rslices += ke->ke_slice; 289 #endif 290 } 291 292 static void 293 kseq_rem(struct kseq *kseq, struct kse *ke) 294 { 295 mtx_assert(&sched_lock, MA_OWNED); 296 kseq->ksq_loads[PRI_BASE(ke->ke_ksegrp->kg_pri_class)]--; 297 kseq->ksq_load--; 298 ke->ke_runq = NULL; 299 if (ke->ke_ksegrp->kg_pri_class == PRI_TIMESHARE) 300 kseq_nice_rem(kseq, ke->ke_ksegrp->kg_nice); 301 #ifdef SMP 302 kseq->ksq_rslices -= ke->ke_slice; 303 #endif 304 } 305 306 static void 307 kseq_nice_add(struct kseq *kseq, int nice) 308 { 309 mtx_assert(&sched_lock, MA_OWNED); 310 /* Normalize to zero. */ 311 kseq->ksq_nice[nice + SCHED_PRI_NHALF]++; 312 if (nice < kseq->ksq_nicemin || kseq->ksq_loads[PRI_TIMESHARE] == 1) 313 kseq->ksq_nicemin = nice; 314 } 315 316 static void 317 kseq_nice_rem(struct kseq *kseq, int nice) 318 { 319 int n; 320 321 mtx_assert(&sched_lock, MA_OWNED); 322 /* Normalize to zero. */ 323 n = nice + SCHED_PRI_NHALF; 324 kseq->ksq_nice[n]--; 325 KASSERT(kseq->ksq_nice[n] >= 0, ("Negative nice count.")); 326 327 /* 328 * If this wasn't the smallest nice value or there are more in 329 * this bucket we can just return. Otherwise we have to recalculate 330 * the smallest nice. 331 */ 332 if (nice != kseq->ksq_nicemin || 333 kseq->ksq_nice[n] != 0 || 334 kseq->ksq_loads[PRI_TIMESHARE] == 0) 335 return; 336 337 for (; n < SCHED_PRI_NRESV + 1; n++) 338 if (kseq->ksq_nice[n]) { 339 kseq->ksq_nicemin = n - SCHED_PRI_NHALF; 340 return; 341 } 342 } 343 344 #ifdef SMP 345 /* 346 * kseq_balance is a simple CPU load balancing algorithm. It operates by 347 * finding the least loaded and most loaded cpu and equalizing their load 348 * by migrating some processes. 349 * 350 * Dealing only with two CPUs at a time has two advantages. Firstly, most 351 * installations will only have 2 cpus. Secondly, load balancing too much at 352 * once can have an unpleasant effect on the system. The scheduler rarely has 353 * enough information to make perfect decisions. So this algorithm chooses 354 * algorithm simplicity and more gradual effects on load in larger systems. 355 * 356 * It could be improved by considering the priorities and slices assigned to 357 * each task prior to balancing them. There are many pathological cases with 358 * any approach and so the semi random algorithm below may work as well as any. 359 * 360 */ 361 void 362 kseq_balance(void *arg) 363 { 364 struct kseq *kseq; 365 int high_load; 366 int low_load; 367 int high_cpu; 368 int low_cpu; 369 int move; 370 int diff; 371 int i; 372 373 high_cpu = 0; 374 low_cpu = 0; 375 high_load = 0; 376 low_load = -1; 377 378 mtx_lock_spin(&sched_lock); 379 if (smp_started == 0) 380 goto out; 381 382 for (i = 0; i < mp_maxid; i++) { 383 if (CPU_ABSENT(i) || (i & stopped_cpus) != 0) 384 continue; 385 kseq = KSEQ_CPU(i); 386 if (kseq->ksq_load > high_load) { 387 high_load = kseq->ksq_load; 388 high_cpu = i; 389 } 390 if (low_load == -1 || kseq->ksq_load < low_load) { 391 low_load = kseq->ksq_load; 392 low_cpu = i; 393 } 394 } 395 396 kseq = KSEQ_CPU(high_cpu); 397 398 /* 399 * Nothing to do. 400 */ 401 if (high_load < kseq->ksq_cpus + 1) 402 goto out; 403 404 high_load -= kseq->ksq_cpus; 405 406 if (low_load >= high_load) 407 goto out; 408 409 diff = high_load - low_load; 410 move = diff / 2; 411 if (diff & 0x1) 412 move++; 413 414 for (i = 0; i < move; i++) 415 kseq_move(kseq, low_cpu); 416 417 out: 418 mtx_unlock_spin(&sched_lock); 419 callout_reset(&kseq_lb_callout, hz, kseq_balance, NULL); 420 421 return; 422 } 423 424 struct kseq * 425 kseq_load_highest(void) 426 { 427 struct kseq *kseq; 428 int load; 429 int cpu; 430 int i; 431 432 mtx_assert(&sched_lock, MA_OWNED); 433 cpu = 0; 434 load = 0; 435 436 for (i = 0; i < mp_maxid; i++) { 437 if (CPU_ABSENT(i) || (i & stopped_cpus) != 0) 438 continue; 439 kseq = KSEQ_CPU(i); 440 if (kseq->ksq_load > load) { 441 load = kseq->ksq_load; 442 cpu = i; 443 } 444 } 445 kseq = KSEQ_CPU(cpu); 446 447 if (load > kseq->ksq_cpus) 448 return (kseq); 449 450 return (NULL); 451 } 452 453 void 454 kseq_move(struct kseq *from, int cpu) 455 { 456 struct kse *ke; 457 458 ke = kseq_choose(from); 459 runq_remove(ke->ke_runq, ke); 460 ke->ke_state = KES_THREAD; 461 kseq_rem(from, ke); 462 463 ke->ke_cpu = cpu; 464 sched_add(ke); 465 } 466 #endif 467 468 struct kse * 469 kseq_choose(struct kseq *kseq) 470 { 471 struct kse *ke; 472 struct runq *swap; 473 474 mtx_assert(&sched_lock, MA_OWNED); 475 swap = NULL; 476 477 for (;;) { 478 ke = runq_choose(kseq->ksq_curr); 479 if (ke == NULL) { 480 /* 481 * We already swaped once and didn't get anywhere. 482 */ 483 if (swap) 484 break; 485 swap = kseq->ksq_curr; 486 kseq->ksq_curr = kseq->ksq_next; 487 kseq->ksq_next = swap; 488 continue; 489 } 490 /* 491 * If we encounter a slice of 0 the kse is in a 492 * TIMESHARE kse group and its nice was too far out 493 * of the range that receives slices. 494 */ 495 if (ke->ke_slice == 0) { 496 runq_remove(ke->ke_runq, ke); 497 sched_slice(ke); 498 ke->ke_runq = kseq->ksq_next; 499 runq_add(ke->ke_runq, ke); 500 continue; 501 } 502 return (ke); 503 } 504 505 return (runq_choose(&kseq->ksq_idle)); 506 } 507 508 static void 509 kseq_setup(struct kseq *kseq) 510 { 511 runq_init(&kseq->ksq_timeshare[0]); 512 runq_init(&kseq->ksq_timeshare[1]); 513 runq_init(&kseq->ksq_idle); 514 515 kseq->ksq_curr = &kseq->ksq_timeshare[0]; 516 kseq->ksq_next = &kseq->ksq_timeshare[1]; 517 518 kseq->ksq_loads[PRI_ITHD] = 0; 519 kseq->ksq_loads[PRI_REALTIME] = 0; 520 kseq->ksq_loads[PRI_TIMESHARE] = 0; 521 kseq->ksq_loads[PRI_IDLE] = 0; 522 kseq->ksq_load = 0; 523 #ifdef SMP 524 kseq->ksq_rslices = 0; 525 #endif 526 } 527 528 static void 529 sched_setup(void *dummy) 530 { 531 #ifdef SMP 532 int i; 533 #endif 534 535 slice_min = (hz/100); /* 10ms */ 536 slice_max = (hz/7); /* ~140ms */ 537 538 #ifdef SMP 539 /* init kseqs */ 540 /* Create the idmap. */ 541 #ifdef ULE_HTT_EXPERIMENTAL 542 if (smp_topology == NULL) { 543 #else 544 if (1) { 545 #endif 546 for (i = 0; i < MAXCPU; i++) { 547 kseq_setup(&kseq_cpu[i]); 548 kseq_idmap[i] = &kseq_cpu[i]; 549 kseq_cpu[i].ksq_cpus = 1; 550 } 551 } else { 552 int j; 553 554 for (i = 0; i < smp_topology->ct_count; i++) { 555 struct cpu_group *cg; 556 557 cg = &smp_topology->ct_group[i]; 558 kseq_setup(&kseq_cpu[i]); 559 560 for (j = 0; j < MAXCPU; j++) 561 if ((cg->cg_mask & (1 << j)) != 0) 562 kseq_idmap[j] = &kseq_cpu[i]; 563 kseq_cpu[i].ksq_cpus = cg->cg_count; 564 } 565 } 566 callout_init(&kseq_lb_callout, 1); 567 kseq_balance(NULL); 568 #else 569 kseq_setup(KSEQ_SELF()); 570 #endif 571 mtx_lock_spin(&sched_lock); 572 kseq_add(KSEQ_SELF(), &kse0); 573 mtx_unlock_spin(&sched_lock); 574 } 575 576 /* 577 * Scale the scheduling priority according to the "interactivity" of this 578 * process. 579 */ 580 static void 581 sched_priority(struct ksegrp *kg) 582 { 583 int pri; 584 585 if (kg->kg_pri_class != PRI_TIMESHARE) 586 return; 587 588 pri = SCHED_PRI_INTERACT(sched_interact_score(kg)); 589 pri += SCHED_PRI_BASE; 590 pri += kg->kg_nice; 591 592 if (pri > PRI_MAX_TIMESHARE) 593 pri = PRI_MAX_TIMESHARE; 594 else if (pri < PRI_MIN_TIMESHARE) 595 pri = PRI_MIN_TIMESHARE; 596 597 kg->kg_user_pri = pri; 598 599 return; 600 } 601 602 /* 603 * Calculate a time slice based on the properties of the kseg and the runq 604 * that we're on. This is only for PRI_TIMESHARE ksegrps. 605 */ 606 static void 607 sched_slice(struct kse *ke) 608 { 609 struct kseq *kseq; 610 struct ksegrp *kg; 611 612 kg = ke->ke_ksegrp; 613 kseq = KSEQ_CPU(ke->ke_cpu); 614 615 /* 616 * Rationale: 617 * KSEs in interactive ksegs get the minimum slice so that we 618 * quickly notice if it abuses its advantage. 619 * 620 * KSEs in non-interactive ksegs are assigned a slice that is 621 * based on the ksegs nice value relative to the least nice kseg 622 * on the run queue for this cpu. 623 * 624 * If the KSE is less nice than all others it gets the maximum 625 * slice and other KSEs will adjust their slice relative to 626 * this when they first expire. 627 * 628 * There is 20 point window that starts relative to the least 629 * nice kse on the run queue. Slice size is determined by 630 * the kse distance from the last nice ksegrp. 631 * 632 * If you are outside of the window you will get no slice and 633 * you will be reevaluated each time you are selected on the 634 * run queue. 635 * 636 */ 637 638 if (!SCHED_INTERACTIVE(kg)) { 639 int nice; 640 641 nice = kg->kg_nice + (0 - kseq->ksq_nicemin); 642 if (kseq->ksq_loads[PRI_TIMESHARE] == 0 || 643 kg->kg_nice < kseq->ksq_nicemin) 644 ke->ke_slice = SCHED_SLICE_MAX; 645 else if (nice <= SCHED_PRI_NTHRESH) 646 ke->ke_slice = SCHED_SLICE_NICE(nice); 647 else 648 ke->ke_slice = 0; 649 } else 650 ke->ke_slice = SCHED_SLICE_MIN; 651 652 CTR6(KTR_ULE, 653 "Sliced %p(%d) (nice: %d, nicemin: %d, load: %d, interactive: %d)", 654 ke, ke->ke_slice, kg->kg_nice, kseq->ksq_nicemin, 655 kseq->ksq_loads[PRI_TIMESHARE], SCHED_INTERACTIVE(kg)); 656 657 /* 658 * Check to see if we need to scale back the slp and run time 659 * in the kg. This will cause us to forget old interactivity 660 * while maintaining the current ratio. 661 */ 662 sched_interact_update(kg); 663 664 return; 665 } 666 667 static void 668 sched_interact_update(struct ksegrp *kg) 669 { 670 /* XXX Fixme, use a linear algorithm and not a while loop. */ 671 while ((kg->kg_runtime + kg->kg_slptime) > SCHED_SLP_RUN_MAX) { 672 kg->kg_runtime = (kg->kg_runtime / 5) * 4; 673 kg->kg_slptime = (kg->kg_slptime / 5) * 4; 674 } 675 } 676 677 static int 678 sched_interact_score(struct ksegrp *kg) 679 { 680 int div; 681 682 if (kg->kg_runtime > kg->kg_slptime) { 683 div = max(1, kg->kg_runtime / SCHED_INTERACT_HALF); 684 return (SCHED_INTERACT_HALF + 685 (SCHED_INTERACT_HALF - (kg->kg_slptime / div))); 686 } if (kg->kg_slptime > kg->kg_runtime) { 687 div = max(1, kg->kg_slptime / SCHED_INTERACT_HALF); 688 return (kg->kg_runtime / div); 689 } 690 691 /* 692 * This can happen if slptime and runtime are 0. 693 */ 694 return (0); 695 696 } 697 698 /* 699 * This is only somewhat accurate since given many processes of the same 700 * priority they will switch when their slices run out, which will be 701 * at most SCHED_SLICE_MAX. 702 */ 703 int 704 sched_rr_interval(void) 705 { 706 return (SCHED_SLICE_MAX); 707 } 708 709 void 710 sched_pctcpu_update(struct kse *ke) 711 { 712 /* 713 * Adjust counters and watermark for pctcpu calc. 714 */ 715 716 /* 717 * Shift the tick count out so that the divide doesn't round away 718 * our results. 719 */ 720 ke->ke_ticks <<= 10; 721 ke->ke_ticks = (ke->ke_ticks / (ke->ke_ltick - ke->ke_ftick)) * 722 SCHED_CPU_TICKS; 723 ke->ke_ticks >>= 10; 724 ke->ke_ltick = ticks; 725 ke->ke_ftick = ke->ke_ltick - SCHED_CPU_TICKS; 726 } 727 728 #ifdef SMP 729 /* XXX Should be changed to kseq_load_lowest() */ 730 int 731 sched_pickcpu(void) 732 { 733 struct kseq *kseq; 734 int load; 735 int cpu; 736 int i; 737 738 mtx_assert(&sched_lock, MA_OWNED); 739 if (!smp_started) 740 return (0); 741 742 load = 0; 743 cpu = 0; 744 745 for (i = 0; i < mp_maxid; i++) { 746 if (CPU_ABSENT(i) || (i & stopped_cpus) != 0) 747 continue; 748 kseq = KSEQ_CPU(i); 749 if (kseq->ksq_load < load) { 750 cpu = i; 751 load = kseq->ksq_load; 752 } 753 } 754 755 CTR1(KTR_RUNQ, "sched_pickcpu: %d", cpu); 756 return (cpu); 757 } 758 #else 759 int 760 sched_pickcpu(void) 761 { 762 return (0); 763 } 764 #endif 765 766 void 767 sched_prio(struct thread *td, u_char prio) 768 { 769 struct kse *ke; 770 struct runq *rq; 771 772 mtx_assert(&sched_lock, MA_OWNED); 773 ke = td->td_kse; 774 td->td_priority = prio; 775 776 if (TD_ON_RUNQ(td)) { 777 rq = ke->ke_runq; 778 779 runq_remove(rq, ke); 780 runq_add(rq, ke); 781 } 782 } 783 784 void 785 sched_switchout(struct thread *td) 786 { 787 struct kse *ke; 788 789 mtx_assert(&sched_lock, MA_OWNED); 790 791 ke = td->td_kse; 792 793 td->td_last_kse = ke; 794 td->td_lastcpu = td->td_oncpu; 795 td->td_oncpu = NOCPU; 796 td->td_flags &= ~TDF_NEEDRESCHED; 797 798 if (TD_IS_RUNNING(td)) { 799 /* 800 * This queue is always correct except for idle threads which 801 * have a higher priority due to priority propagation. 802 */ 803 if (ke->ke_ksegrp->kg_pri_class == PRI_IDLE && 804 ke->ke_thread->td_priority > PRI_MIN_IDLE) 805 ke->ke_runq = KSEQ_SELF()->ksq_curr; 806 runq_add(ke->ke_runq, ke); 807 /* setrunqueue(td); */ 808 return; 809 } 810 if (ke->ke_runq) 811 kseq_rem(KSEQ_CPU(ke->ke_cpu), ke); 812 /* 813 * We will not be on the run queue. So we must be 814 * sleeping or similar. 815 */ 816 if (td->td_proc->p_flag & P_SA) 817 kse_reassign(ke); 818 } 819 820 void 821 sched_switchin(struct thread *td) 822 { 823 /* struct kse *ke = td->td_kse; */ 824 mtx_assert(&sched_lock, MA_OWNED); 825 826 td->td_oncpu = PCPU_GET(cpuid); 827 } 828 829 void 830 sched_nice(struct ksegrp *kg, int nice) 831 { 832 struct kse *ke; 833 struct thread *td; 834 struct kseq *kseq; 835 836 PROC_LOCK_ASSERT(kg->kg_proc, MA_OWNED); 837 mtx_assert(&sched_lock, MA_OWNED); 838 /* 839 * We need to adjust the nice counts for running KSEs. 840 */ 841 if (kg->kg_pri_class == PRI_TIMESHARE) 842 FOREACH_KSE_IN_GROUP(kg, ke) { 843 if (ke->ke_runq == NULL) 844 continue; 845 kseq = KSEQ_CPU(ke->ke_cpu); 846 kseq_nice_rem(kseq, kg->kg_nice); 847 kseq_nice_add(kseq, nice); 848 } 849 kg->kg_nice = nice; 850 sched_priority(kg); 851 FOREACH_THREAD_IN_GROUP(kg, td) 852 td->td_flags |= TDF_NEEDRESCHED; 853 } 854 855 void 856 sched_sleep(struct thread *td, u_char prio) 857 { 858 mtx_assert(&sched_lock, MA_OWNED); 859 860 td->td_slptime = ticks; 861 td->td_priority = prio; 862 863 CTR2(KTR_ULE, "sleep kse %p (tick: %d)", 864 td->td_kse, td->td_slptime); 865 } 866 867 void 868 sched_wakeup(struct thread *td) 869 { 870 mtx_assert(&sched_lock, MA_OWNED); 871 872 /* 873 * Let the kseg know how long we slept for. This is because process 874 * interactivity behavior is modeled in the kseg. 875 */ 876 if (td->td_slptime) { 877 struct ksegrp *kg; 878 int hzticks; 879 880 kg = td->td_ksegrp; 881 hzticks = ticks - td->td_slptime; 882 kg->kg_slptime += hzticks << 10; 883 sched_interact_update(kg); 884 sched_priority(kg); 885 if (td->td_kse) 886 sched_slice(td->td_kse); 887 CTR2(KTR_ULE, "wakeup kse %p (%d ticks)", 888 td->td_kse, hzticks); 889 td->td_slptime = 0; 890 } 891 setrunqueue(td); 892 if (td->td_priority < curthread->td_priority) 893 curthread->td_flags |= TDF_NEEDRESCHED; 894 } 895 896 /* 897 * Penalize the parent for creating a new child and initialize the child's 898 * priority. 899 */ 900 void 901 sched_fork(struct proc *p, struct proc *p1) 902 { 903 904 mtx_assert(&sched_lock, MA_OWNED); 905 906 sched_fork_ksegrp(FIRST_KSEGRP_IN_PROC(p), FIRST_KSEGRP_IN_PROC(p1)); 907 sched_fork_kse(FIRST_KSE_IN_PROC(p), FIRST_KSE_IN_PROC(p1)); 908 sched_fork_thread(FIRST_THREAD_IN_PROC(p), FIRST_THREAD_IN_PROC(p1)); 909 } 910 911 void 912 sched_fork_kse(struct kse *ke, struct kse *child) 913 { 914 915 child->ke_slice = 1; /* Attempt to quickly learn interactivity. */ 916 child->ke_cpu = ke->ke_cpu; /* sched_pickcpu(); */ 917 child->ke_runq = NULL; 918 919 /* 920 * Claim that we've been running for one second for statistical 921 * purposes. 922 */ 923 child->ke_ticks = 0; 924 child->ke_ltick = ticks; 925 child->ke_ftick = ticks - hz; 926 } 927 928 void 929 sched_fork_ksegrp(struct ksegrp *kg, struct ksegrp *child) 930 { 931 932 PROC_LOCK_ASSERT(child->kg_proc, MA_OWNED); 933 /* XXX Need something better here */ 934 935 child->kg_slptime = kg->kg_slptime / SCHED_SLP_RUN_THROTTLE; 936 child->kg_runtime = kg->kg_runtime / SCHED_SLP_RUN_THROTTLE; 937 kg->kg_runtime += tickincr << 10; 938 sched_interact_update(kg); 939 940 child->kg_user_pri = kg->kg_user_pri; 941 child->kg_nice = kg->kg_nice; 942 } 943 944 void 945 sched_fork_thread(struct thread *td, struct thread *child) 946 { 947 } 948 949 void 950 sched_class(struct ksegrp *kg, int class) 951 { 952 struct kseq *kseq; 953 struct kse *ke; 954 955 mtx_assert(&sched_lock, MA_OWNED); 956 if (kg->kg_pri_class == class) 957 return; 958 959 FOREACH_KSE_IN_GROUP(kg, ke) { 960 if (ke->ke_state != KES_ONRUNQ && 961 ke->ke_state != KES_THREAD) 962 continue; 963 kseq = KSEQ_CPU(ke->ke_cpu); 964 965 kseq->ksq_loads[PRI_BASE(kg->kg_pri_class)]--; 966 kseq->ksq_loads[PRI_BASE(class)]++; 967 968 if (kg->kg_pri_class == PRI_TIMESHARE) 969 kseq_nice_rem(kseq, kg->kg_nice); 970 else if (class == PRI_TIMESHARE) 971 kseq_nice_add(kseq, kg->kg_nice); 972 } 973 974 kg->kg_pri_class = class; 975 } 976 977 /* 978 * Return some of the child's priority and interactivity to the parent. 979 */ 980 void 981 sched_exit(struct proc *p, struct proc *child) 982 { 983 /* XXX Need something better here */ 984 mtx_assert(&sched_lock, MA_OWNED); 985 sched_exit_kse(FIRST_KSE_IN_PROC(p), FIRST_KSE_IN_PROC(child)); 986 sched_exit_ksegrp(FIRST_KSEGRP_IN_PROC(p), FIRST_KSEGRP_IN_PROC(child)); 987 } 988 989 void 990 sched_exit_kse(struct kse *ke, struct kse *child) 991 { 992 kseq_rem(KSEQ_CPU(child->ke_cpu), child); 993 } 994 995 void 996 sched_exit_ksegrp(struct ksegrp *kg, struct ksegrp *child) 997 { 998 /* kg->kg_slptime += child->kg_slptime; */ 999 kg->kg_runtime += child->kg_runtime; 1000 sched_interact_update(kg); 1001 } 1002 1003 void 1004 sched_exit_thread(struct thread *td, struct thread *child) 1005 { 1006 } 1007 1008 void 1009 sched_clock(struct kse *ke) 1010 { 1011 struct kseq *kseq; 1012 struct ksegrp *kg; 1013 struct thread *td; 1014 #if 0 1015 struct kse *nke; 1016 #endif 1017 1018 /* 1019 * sched_setup() apparently happens prior to stathz being set. We 1020 * need to resolve the timers earlier in the boot so we can avoid 1021 * calculating this here. 1022 */ 1023 if (realstathz == 0) { 1024 realstathz = stathz ? stathz : hz; 1025 tickincr = hz / realstathz; 1026 /* 1027 * XXX This does not work for values of stathz that are much 1028 * larger than hz. 1029 */ 1030 if (tickincr == 0) 1031 tickincr = 1; 1032 } 1033 1034 td = ke->ke_thread; 1035 kg = ke->ke_ksegrp; 1036 1037 mtx_assert(&sched_lock, MA_OWNED); 1038 KASSERT((td != NULL), ("schedclock: null thread pointer")); 1039 1040 /* Adjust ticks for pctcpu */ 1041 ke->ke_ticks++; 1042 ke->ke_ltick = ticks; 1043 1044 /* Go up to one second beyond our max and then trim back down */ 1045 if (ke->ke_ftick + SCHED_CPU_TICKS + hz < ke->ke_ltick) 1046 sched_pctcpu_update(ke); 1047 1048 if (td->td_flags & TDF_IDLETD) 1049 return; 1050 1051 CTR4(KTR_ULE, "Tick kse %p (slice: %d, slptime: %d, runtime: %d)", 1052 ke, ke->ke_slice, kg->kg_slptime >> 10, kg->kg_runtime >> 10); 1053 1054 /* 1055 * We only do slicing code for TIMESHARE ksegrps. 1056 */ 1057 if (kg->kg_pri_class != PRI_TIMESHARE) 1058 return; 1059 /* 1060 * Check for a higher priority task on the run queue. This can happen 1061 * on SMP if another processor woke up a process on our runq. 1062 */ 1063 kseq = KSEQ_SELF(); 1064 #if 0 1065 if (kseq->ksq_load > 1 && (nke = kseq_choose(kseq)) != NULL) { 1066 if (sched_strict && 1067 nke->ke_thread->td_priority < td->td_priority) 1068 td->td_flags |= TDF_NEEDRESCHED; 1069 else if (nke->ke_thread->td_priority < 1070 td->td_priority SCHED_PRIO_SLOP) 1071 1072 if (nke->ke_thread->td_priority < td->td_priority) 1073 td->td_flags |= TDF_NEEDRESCHED; 1074 } 1075 #endif 1076 /* 1077 * We used a tick charge it to the ksegrp so that we can compute our 1078 * interactivity. 1079 */ 1080 kg->kg_runtime += tickincr << 10; 1081 sched_interact_update(kg); 1082 1083 /* 1084 * We used up one time slice. 1085 */ 1086 ke->ke_slice--; 1087 #ifdef SMP 1088 kseq->ksq_rslices--; 1089 #endif 1090 1091 if (ke->ke_slice > 0) 1092 return; 1093 /* 1094 * We're out of time, recompute priorities and requeue. 1095 */ 1096 kseq_rem(kseq, ke); 1097 sched_priority(kg); 1098 sched_slice(ke); 1099 if (SCHED_CURR(kg, ke)) 1100 ke->ke_runq = kseq->ksq_curr; 1101 else 1102 ke->ke_runq = kseq->ksq_next; 1103 kseq_add(kseq, ke); 1104 td->td_flags |= TDF_NEEDRESCHED; 1105 } 1106 1107 int 1108 sched_runnable(void) 1109 { 1110 struct kseq *kseq; 1111 int load; 1112 1113 load = 1; 1114 1115 mtx_lock_spin(&sched_lock); 1116 kseq = KSEQ_SELF(); 1117 1118 if (kseq->ksq_load) 1119 goto out; 1120 #ifdef SMP 1121 /* 1122 * For SMP we may steal other processor's KSEs. Just search until we 1123 * verify that at least on other cpu has a runnable task. 1124 */ 1125 if (smp_started) { 1126 int i; 1127 1128 for (i = 0; i < mp_maxid; i++) { 1129 if (CPU_ABSENT(i) || (i & stopped_cpus) != 0) 1130 continue; 1131 kseq = KSEQ_CPU(i); 1132 if (kseq->ksq_load > kseq->ksq_cpus) 1133 goto out; 1134 } 1135 } 1136 #endif 1137 load = 0; 1138 out: 1139 mtx_unlock_spin(&sched_lock); 1140 return (load); 1141 } 1142 1143 void 1144 sched_userret(struct thread *td) 1145 { 1146 struct ksegrp *kg; 1147 struct kseq *kseq; 1148 struct kse *ke; 1149 1150 kg = td->td_ksegrp; 1151 1152 if (td->td_priority != kg->kg_user_pri) { 1153 mtx_lock_spin(&sched_lock); 1154 td->td_priority = kg->kg_user_pri; 1155 kseq = KSEQ_SELF(); 1156 if (td->td_ksegrp->kg_pri_class == PRI_TIMESHARE && 1157 #ifdef SMP 1158 kseq->ksq_load > kseq->ksq_cpus && 1159 #else 1160 kseq->ksq_load > 1 && 1161 #endif 1162 (ke = kseq_choose(kseq)) != NULL && 1163 ke->ke_thread->td_priority < td->td_priority) 1164 curthread->td_flags |= TDF_NEEDRESCHED; 1165 mtx_unlock_spin(&sched_lock); 1166 } 1167 } 1168 1169 struct kse * 1170 sched_choose(void) 1171 { 1172 struct kseq *kseq; 1173 struct kse *ke; 1174 1175 mtx_assert(&sched_lock, MA_OWNED); 1176 #ifdef SMP 1177 retry: 1178 #endif 1179 kseq = KSEQ_SELF(); 1180 ke = kseq_choose(kseq); 1181 if (ke) { 1182 runq_remove(ke->ke_runq, ke); 1183 ke->ke_state = KES_THREAD; 1184 1185 if (ke->ke_ksegrp->kg_pri_class == PRI_TIMESHARE) { 1186 CTR4(KTR_ULE, "Run kse %p from %p (slice: %d, pri: %d)", 1187 ke, ke->ke_runq, ke->ke_slice, 1188 ke->ke_thread->td_priority); 1189 } 1190 return (ke); 1191 } 1192 1193 #ifdef SMP 1194 if (smp_started) { 1195 /* 1196 * Find the cpu with the highest load and steal one proc. 1197 */ 1198 if ((kseq = kseq_load_highest()) == NULL) 1199 return (NULL); 1200 1201 /* 1202 * Remove this kse from this kseq and runq and then requeue 1203 * on the current processor. Then we will dequeue it 1204 * normally above. 1205 */ 1206 kseq_move(kseq, PCPU_GET(cpuid)); 1207 goto retry; 1208 } 1209 #endif 1210 1211 return (NULL); 1212 } 1213 1214 void 1215 sched_add(struct kse *ke) 1216 { 1217 struct kseq *kseq; 1218 struct ksegrp *kg; 1219 1220 mtx_assert(&sched_lock, MA_OWNED); 1221 KASSERT((ke->ke_thread != NULL), ("sched_add: No thread on KSE")); 1222 KASSERT((ke->ke_thread->td_kse != NULL), 1223 ("sched_add: No KSE on thread")); 1224 KASSERT(ke->ke_state != KES_ONRUNQ, 1225 ("sched_add: kse %p (%s) already in run queue", ke, 1226 ke->ke_proc->p_comm)); 1227 KASSERT(ke->ke_proc->p_sflag & PS_INMEM, 1228 ("sched_add: process swapped out")); 1229 KASSERT(ke->ke_runq == NULL, 1230 ("sched_add: KSE %p is still assigned to a run queue", ke)); 1231 1232 kg = ke->ke_ksegrp; 1233 1234 switch (PRI_BASE(kg->kg_pri_class)) { 1235 case PRI_ITHD: 1236 case PRI_REALTIME: 1237 kseq = KSEQ_SELF(); 1238 ke->ke_runq = kseq->ksq_curr; 1239 ke->ke_slice = SCHED_SLICE_MAX; 1240 ke->ke_cpu = PCPU_GET(cpuid); 1241 break; 1242 case PRI_TIMESHARE: 1243 kseq = KSEQ_CPU(ke->ke_cpu); 1244 if (SCHED_CURR(kg, ke)) 1245 ke->ke_runq = kseq->ksq_curr; 1246 else 1247 ke->ke_runq = kseq->ksq_next; 1248 break; 1249 case PRI_IDLE: 1250 kseq = KSEQ_CPU(ke->ke_cpu); 1251 /* 1252 * This is for priority prop. 1253 */ 1254 if (ke->ke_thread->td_priority > PRI_MIN_IDLE) 1255 ke->ke_runq = kseq->ksq_curr; 1256 else 1257 ke->ke_runq = &kseq->ksq_idle; 1258 ke->ke_slice = SCHED_SLICE_MIN; 1259 break; 1260 default: 1261 panic("Unknown pri class.\n"); 1262 break; 1263 } 1264 1265 ke->ke_ksegrp->kg_runq_kses++; 1266 ke->ke_state = KES_ONRUNQ; 1267 1268 runq_add(ke->ke_runq, ke); 1269 kseq_add(kseq, ke); 1270 } 1271 1272 void 1273 sched_rem(struct kse *ke) 1274 { 1275 struct kseq *kseq; 1276 1277 mtx_assert(&sched_lock, MA_OWNED); 1278 KASSERT((ke->ke_state == KES_ONRUNQ), ("KSE not on run queue")); 1279 1280 ke->ke_state = KES_THREAD; 1281 ke->ke_ksegrp->kg_runq_kses--; 1282 kseq = KSEQ_CPU(ke->ke_cpu); 1283 runq_remove(ke->ke_runq, ke); 1284 kseq_rem(kseq, ke); 1285 } 1286 1287 fixpt_t 1288 sched_pctcpu(struct kse *ke) 1289 { 1290 fixpt_t pctcpu; 1291 1292 pctcpu = 0; 1293 1294 mtx_lock_spin(&sched_lock); 1295 if (ke->ke_ticks) { 1296 int rtick; 1297 1298 /* 1299 * Don't update more frequently than twice a second. Allowing 1300 * this causes the cpu usage to decay away too quickly due to 1301 * rounding errors. 1302 */ 1303 if (ke->ke_ltick < (ticks - (hz / 2))) 1304 sched_pctcpu_update(ke); 1305 1306 /* How many rtick per second ? */ 1307 rtick = min(ke->ke_ticks / SCHED_CPU_TIME, SCHED_CPU_TICKS); 1308 pctcpu = (FSCALE * ((FSCALE * rtick)/realstathz)) >> FSHIFT; 1309 } 1310 1311 ke->ke_proc->p_swtime = ke->ke_ltick - ke->ke_ftick; 1312 mtx_unlock_spin(&sched_lock); 1313 1314 return (pctcpu); 1315 } 1316 1317 int 1318 sched_sizeof_kse(void) 1319 { 1320 return (sizeof(struct kse) + sizeof(struct ke_sched)); 1321 } 1322 1323 int 1324 sched_sizeof_ksegrp(void) 1325 { 1326 return (sizeof(struct ksegrp) + sizeof(struct kg_sched)); 1327 } 1328 1329 int 1330 sched_sizeof_proc(void) 1331 { 1332 return (sizeof(struct proc)); 1333 } 1334 1335 int 1336 sched_sizeof_thread(void) 1337 { 1338 return (sizeof(struct thread) + sizeof(struct td_sched)); 1339 } 1340