1 /* 2 * Copyright (c) 2012 The DragonFly Project. All rights reserved. 3 * Copyright (c) 1999 Peter Wemm <peter@FreeBSD.org>. All rights reserved. 4 * 5 * This code is derived from software contributed to The DragonFly Project 6 * by Matthew Dillon <dillon@backplane.com>, 7 * by Mihai Carabas <mihai.carabas@gmail.com> 8 * and many others. 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions 12 * are met: 13 * 14 * 1. Redistributions of source code must retain the above copyright 15 * notice, this list of conditions and the following disclaimer. 16 * 2. Redistributions in binary form must reproduce the above copyright 17 * notice, this list of conditions and the following disclaimer in 18 * the documentation and/or other materials provided with the 19 * distribution. 20 * 3. Neither the name of The DragonFly Project nor the names of its 21 * contributors may be used to endorse or promote products derived 22 * from this software without specific, prior written permission. 23 * 24 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 25 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 26 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 27 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 28 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 29 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 30 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 31 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 32 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 33 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 34 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 35 * SUCH DAMAGE. 36 */ 37 #include <sys/param.h> 38 #include <sys/systm.h> 39 #include <sys/kernel.h> 40 #include <sys/lock.h> 41 #include <sys/queue.h> 42 #include <sys/proc.h> 43 #include <sys/rtprio.h> 44 #include <sys/uio.h> 45 #include <sys/sysctl.h> 46 #include <sys/resourcevar.h> 47 #include <sys/spinlock.h> 48 #include <sys/cpu_topology.h> 49 #include <sys/thread2.h> 50 #include <sys/spinlock2.h> 51 #include <sys/mplock2.h> 52 53 #include <sys/ktr.h> 54 55 #include <machine/cpu.h> 56 #include <machine/smp.h> 57 58 /* 59 * Priorities. Note that with 32 run queues per scheduler each queue 60 * represents four priority levels. 61 */ 62 63 int dfly_rebalanced; 64 65 #define MAXPRI 128 66 #define PRIMASK (MAXPRI - 1) 67 #define PRIBASE_REALTIME 0 68 #define PRIBASE_NORMAL MAXPRI 69 #define PRIBASE_IDLE (MAXPRI * 2) 70 #define PRIBASE_THREAD (MAXPRI * 3) 71 #define PRIBASE_NULL (MAXPRI * 4) 72 73 #define NQS 32 /* 32 run queues. */ 74 #define PPQ (MAXPRI / NQS) /* priorities per queue */ 75 #define PPQMASK (PPQ - 1) 76 77 /* 78 * NICEPPQ - number of nice units per priority queue 79 * ESTCPUPPQ - number of estcpu units per priority queue 80 * ESTCPUMAX - number of estcpu units 81 */ 82 #define NICEPPQ 2 83 #define ESTCPUPPQ 512 84 #define ESTCPUMAX (ESTCPUPPQ * NQS) 85 #define BATCHMAX (ESTCPUFREQ * 30) 86 #define PRIO_RANGE (PRIO_MAX - PRIO_MIN + 1) 87 88 #define ESTCPULIM(v) min((v), ESTCPUMAX) 89 90 TAILQ_HEAD(rq, lwp); 91 92 #define lwp_priority lwp_usdata.dfly.priority 93 #define lwp_forked lwp_usdata.dfly.forked 94 #define lwp_rqindex lwp_usdata.dfly.rqindex 95 #define lwp_estcpu lwp_usdata.dfly.estcpu 96 #define lwp_estfast lwp_usdata.dfly.estfast 97 #define lwp_uload lwp_usdata.dfly.uload 98 #define lwp_rqtype lwp_usdata.dfly.rqtype 99 #define lwp_qcpu lwp_usdata.dfly.qcpu 100 #define lwp_rrcount lwp_usdata.dfly.rrcount 101 102 struct usched_dfly_pcpu { 103 struct spinlock spin; 104 struct thread helper_thread; 105 short unusde01; 106 short upri; 107 int uload; 108 int ucount; 109 struct lwp *uschedcp; 110 struct rq queues[NQS]; 111 struct rq rtqueues[NQS]; 112 struct rq idqueues[NQS]; 113 u_int32_t queuebits; 114 u_int32_t rtqueuebits; 115 u_int32_t idqueuebits; 116 int runqcount; 117 int cpuid; 118 cpumask_t cpumask; 119 cpu_node_t *cpunode; 120 }; 121 122 typedef struct usched_dfly_pcpu *dfly_pcpu_t; 123 124 static void dfly_acquire_curproc(struct lwp *lp); 125 static void dfly_release_curproc(struct lwp *lp); 126 static void dfly_select_curproc(globaldata_t gd); 127 static void dfly_setrunqueue(struct lwp *lp); 128 static void dfly_setrunqueue_dd(dfly_pcpu_t rdd, struct lwp *lp); 129 static void dfly_schedulerclock(struct lwp *lp, sysclock_t period, 130 sysclock_t cpstamp); 131 static void dfly_recalculate_estcpu(struct lwp *lp); 132 static void dfly_resetpriority(struct lwp *lp); 133 static void dfly_forking(struct lwp *plp, struct lwp *lp); 134 static void dfly_exiting(struct lwp *lp, struct proc *); 135 static void dfly_uload_update(struct lwp *lp); 136 static void dfly_yield(struct lwp *lp); 137 static void dfly_changeqcpu_locked(struct lwp *lp, 138 dfly_pcpu_t dd, dfly_pcpu_t rdd); 139 static dfly_pcpu_t dfly_choose_best_queue(struct lwp *lp); 140 static dfly_pcpu_t dfly_choose_worst_queue(dfly_pcpu_t dd); 141 static dfly_pcpu_t dfly_choose_queue_simple(dfly_pcpu_t dd, struct lwp *lp); 142 static void dfly_need_user_resched_remote(void *dummy); 143 static struct lwp *dfly_chooseproc_locked(dfly_pcpu_t rdd, dfly_pcpu_t dd, 144 struct lwp *chklp, int worst); 145 static void dfly_remrunqueue_locked(dfly_pcpu_t dd, struct lwp *lp); 146 static void dfly_setrunqueue_locked(dfly_pcpu_t dd, struct lwp *lp); 147 static void dfly_changedcpu(struct lwp *lp); 148 149 struct usched usched_dfly = { 150 { NULL }, 151 "dfly", "Original DragonFly Scheduler", 152 NULL, /* default registration */ 153 NULL, /* default deregistration */ 154 dfly_acquire_curproc, 155 dfly_release_curproc, 156 dfly_setrunqueue, 157 dfly_schedulerclock, 158 dfly_recalculate_estcpu, 159 dfly_resetpriority, 160 dfly_forking, 161 dfly_exiting, 162 dfly_uload_update, 163 NULL, /* setcpumask not supported */ 164 dfly_yield, 165 dfly_changedcpu 166 }; 167 168 /* 169 * We have NQS (32) run queues per scheduling class. For the normal 170 * class, there are 128 priorities scaled onto these 32 queues. New 171 * processes are added to the last entry in each queue, and processes 172 * are selected for running by taking them from the head and maintaining 173 * a simple FIFO arrangement. Realtime and Idle priority processes have 174 * and explicit 0-31 priority which maps directly onto their class queue 175 * index. When a queue has something in it, the corresponding bit is 176 * set in the queuebits variable, allowing a single read to determine 177 * the state of all 32 queues and then a ffs() to find the first busy 178 * queue. 179 */ 180 /* currently running a user process */ 181 static cpumask_t dfly_curprocmask = CPUMASK_INITIALIZER_ALLONES; 182 static cpumask_t dfly_rdyprocmask; /* ready to accept a user process */ 183 static volatile int dfly_scancpu; 184 static volatile int dfly_ucount; /* total running on whole system */ 185 static struct usched_dfly_pcpu dfly_pcpu[MAXCPU]; 186 static struct sysctl_ctx_list usched_dfly_sysctl_ctx; 187 static struct sysctl_oid *usched_dfly_sysctl_tree; 188 189 /* Debug info exposed through debug.* sysctl */ 190 191 static int usched_dfly_debug = -1; 192 SYSCTL_INT(_debug, OID_AUTO, dfly_scdebug, CTLFLAG_RW, 193 &usched_dfly_debug, 0, 194 "Print debug information for this pid"); 195 196 static int usched_dfly_pid_debug = -1; 197 SYSCTL_INT(_debug, OID_AUTO, dfly_pid_debug, CTLFLAG_RW, 198 &usched_dfly_pid_debug, 0, 199 "Print KTR debug information for this pid"); 200 201 static int usched_dfly_chooser = 0; 202 SYSCTL_INT(_debug, OID_AUTO, dfly_chooser, CTLFLAG_RW, 203 &usched_dfly_chooser, 0, 204 "Print KTR debug information for this pid"); 205 206 /* 207 * Tunning usched_dfly - configurable through kern.usched_dfly. 208 * 209 * weight1 - Tries to keep threads on their current cpu. If you 210 * make this value too large the scheduler will not be 211 * able to load-balance large loads. 212 * 213 * weight2 - If non-zero, detects thread pairs undergoing synchronous 214 * communications and tries to move them closer together. 215 * Behavior is adjusted by bit 4 of features (0x10). 216 * 217 * WARNING! Weight2 is a ridiculously sensitive parameter, 218 * a small value is recommended. 219 * 220 * weight3 - Weighting based on the number of recently runnable threads 221 * on the userland scheduling queue (ignoring their loads). 222 * A nominal value here prevents high-priority (low-load) 223 * threads from accumulating on one cpu core when other 224 * cores are available. 225 * 226 * This value should be left fairly small relative to weight1 227 * and weight4. 228 * 229 * weight4 - Weighting based on other cpu queues being available 230 * or running processes with higher lwp_priority's. 231 * 232 * This allows a thread to migrate to another nearby cpu if it 233 * is unable to run on the current cpu based on the other cpu 234 * being idle or running a lower priority (higher lwp_priority) 235 * thread. This value should be large enough to override weight1 236 * 237 * features - These flags can be set or cleared to enable or disable various 238 * features. 239 * 240 * 0x01 Enable idle-cpu pulling (default) 241 * 0x02 Enable proactive pushing (default) 242 * 0x04 Enable rebalancing rover (default) 243 * 0x08 Enable more proactive pushing (default) 244 * 0x10 (flip weight2 limit on same cpu) (default) 245 * 0x20 choose best cpu for forked process 246 * 0x40 choose current cpu for forked process 247 * 0x80 choose random cpu for forked process (default) 248 */ 249 static int usched_dfly_smt = 0; 250 static int usched_dfly_cache_coherent = 0; 251 static int usched_dfly_weight1 = 200; /* keep thread on current cpu */ 252 static int usched_dfly_weight2 = 180; /* synchronous peer's current cpu */ 253 static int usched_dfly_weight3 = 40; /* number of threads on queue */ 254 static int usched_dfly_weight4 = 160; /* availability of idle cores */ 255 static int usched_dfly_features = 0x8F; /* allow pulls */ 256 static int usched_dfly_fast_resched = 0;/* delta priority / resched */ 257 static int usched_dfly_swmask = ~PPQMASK; /* allow pulls */ 258 static int usched_dfly_rrinterval = (ESTCPUFREQ + 9) / 10; 259 static int usched_dfly_decay = 8; 260 261 /* KTR debug printings */ 262 263 KTR_INFO_MASTER(usched); 264 265 #if !defined(KTR_USCHED_DFLY) 266 #define KTR_USCHED_DFLY KTR_ALL 267 #endif 268 269 KTR_INFO(KTR_USCHED_DFLY, usched, chooseproc, 0, 270 "USCHED_DFLY(chooseproc: pid %d, old_cpuid %d, curr_cpuid %d)", 271 pid_t pid, int old_cpuid, int curr); 272 273 /* 274 * This function is called when the kernel intends to return to userland. 275 * It is responsible for making the thread the current designated userland 276 * thread for this cpu, blocking if necessary. 277 * 278 * The kernel will not depress our LWKT priority until after we return, 279 * in case we have to shove over to another cpu. 280 * 281 * We must determine our thread's disposition before we switch away. This 282 * is very sensitive code. 283 * 284 * WARNING! THIS FUNCTION IS ALLOWED TO CAUSE THE CURRENT THREAD TO MIGRATE 285 * TO ANOTHER CPU! Because most of the kernel assumes that no migration will 286 * occur, this function is called only under very controlled circumstances. 287 */ 288 static void 289 dfly_acquire_curproc(struct lwp *lp) 290 { 291 globaldata_t gd; 292 dfly_pcpu_t dd; 293 dfly_pcpu_t rdd; 294 thread_t td; 295 int force_resched; 296 297 /* 298 * Make sure we aren't sitting on a tsleep queue. 299 */ 300 td = lp->lwp_thread; 301 crit_enter_quick(td); 302 if (td->td_flags & TDF_TSLEEPQ) 303 tsleep_remove(td); 304 dfly_recalculate_estcpu(lp); 305 306 gd = mycpu; 307 dd = &dfly_pcpu[gd->gd_cpuid]; 308 309 /* 310 * Process any pending interrupts/ipi's, then handle reschedule 311 * requests. dfly_release_curproc() will try to assign a new 312 * uschedcp that isn't us and otherwise NULL it out. 313 */ 314 force_resched = 0; 315 if ((td->td_mpflags & TDF_MP_BATCH_DEMARC) && 316 lp->lwp_rrcount >= usched_dfly_rrinterval / 2) { 317 force_resched = 1; 318 } 319 320 if (user_resched_wanted()) { 321 if (dd->uschedcp == lp) 322 force_resched = 1; 323 clear_user_resched(); 324 dfly_release_curproc(lp); 325 } 326 327 /* 328 * Loop until we are the current user thread. 329 * 330 * NOTE: dd spinlock not held at top of loop. 331 */ 332 if (dd->uschedcp == lp) 333 lwkt_yield_quick(); 334 335 while (dd->uschedcp != lp) { 336 lwkt_yield_quick(); 337 338 spin_lock(&dd->spin); 339 340 if (force_resched && 341 (usched_dfly_features & 0x08) && 342 (rdd = dfly_choose_best_queue(lp)) != dd) { 343 /* 344 * We are not or are no longer the current lwp and a 345 * forced reschedule was requested. Figure out the 346 * best cpu to run on (our current cpu will be given 347 * significant weight). 348 * 349 * (if a reschedule was not requested we want to 350 * move this step after the uschedcp tests). 351 */ 352 dfly_changeqcpu_locked(lp, dd, rdd); 353 spin_unlock(&dd->spin); 354 lwkt_deschedule(lp->lwp_thread); 355 dfly_setrunqueue_dd(rdd, lp); 356 lwkt_switch(); 357 gd = mycpu; 358 dd = &dfly_pcpu[gd->gd_cpuid]; 359 continue; 360 } 361 362 /* 363 * Either no reschedule was requested or the best queue was 364 * dd, and no current process has been selected. We can 365 * trivially become the current lwp on the current cpu. 366 */ 367 if (dd->uschedcp == NULL) { 368 atomic_clear_int(&lp->lwp_thread->td_mpflags, 369 TDF_MP_DIDYIELD); 370 ATOMIC_CPUMASK_ORBIT(dfly_curprocmask, gd->gd_cpuid); 371 dd->uschedcp = lp; 372 dd->upri = lp->lwp_priority; 373 KKASSERT(lp->lwp_qcpu == dd->cpuid); 374 spin_unlock(&dd->spin); 375 break; 376 } 377 378 /* 379 * Put us back on the same run queue unconditionally. 380 * 381 * Set rrinterval to force placement at end of queue. 382 * Select the worst queue to ensure we round-robin, 383 * but do not change estcpu. 384 */ 385 if (lp->lwp_thread->td_mpflags & TDF_MP_DIDYIELD) { 386 u_int32_t tsqbits; 387 388 switch(lp->lwp_rqtype) { 389 case RTP_PRIO_NORMAL: 390 tsqbits = dd->queuebits; 391 spin_unlock(&dd->spin); 392 393 lp->lwp_rrcount = usched_dfly_rrinterval; 394 if (tsqbits) 395 lp->lwp_rqindex = bsrl(tsqbits); 396 break; 397 default: 398 spin_unlock(&dd->spin); 399 break; 400 } 401 lwkt_deschedule(lp->lwp_thread); 402 dfly_setrunqueue_dd(dd, lp); 403 atomic_clear_int(&lp->lwp_thread->td_mpflags, 404 TDF_MP_DIDYIELD); 405 lwkt_switch(); 406 gd = mycpu; 407 dd = &dfly_pcpu[gd->gd_cpuid]; 408 continue; 409 } 410 411 /* 412 * Can we steal the current designated user thread? 413 * 414 * If we do the other thread will stall when it tries to 415 * return to userland, possibly rescheduling elsewhere. 416 * 417 * It is important to do a masked test to avoid the edge 418 * case where two near-equal-priority threads are constantly 419 * interrupting each other. 420 * 421 * In the exact match case another thread has already gained 422 * uschedcp and lowered its priority, if we steal it the 423 * other thread will stay stuck on the LWKT runq and not 424 * push to another cpu. So don't steal on equal-priority even 425 * though it might appear to be more beneficial due to not 426 * having to switch back to the other thread's context. 427 * 428 * usched_dfly_fast_resched requires that two threads be 429 * significantly far apart in priority in order to interrupt. 430 * 431 * If better but not sufficiently far apart, the current 432 * uschedcp will be interrupted at the next scheduler clock. 433 */ 434 if (dd->uschedcp && 435 (dd->upri & ~PPQMASK) > 436 (lp->lwp_priority & ~PPQMASK) + usched_dfly_fast_resched) { 437 dd->uschedcp = lp; 438 dd->upri = lp->lwp_priority; 439 KKASSERT(lp->lwp_qcpu == dd->cpuid); 440 spin_unlock(&dd->spin); 441 break; 442 } 443 /* 444 * We are not the current lwp, figure out the best cpu 445 * to run on (our current cpu will be given significant 446 * weight). Loop on cpu change. 447 */ 448 if ((usched_dfly_features & 0x02) && 449 force_resched == 0 && 450 (rdd = dfly_choose_best_queue(lp)) != dd) { 451 dfly_changeqcpu_locked(lp, dd, rdd); 452 spin_unlock(&dd->spin); 453 lwkt_deschedule(lp->lwp_thread); 454 dfly_setrunqueue_dd(rdd, lp); 455 lwkt_switch(); 456 gd = mycpu; 457 dd = &dfly_pcpu[gd->gd_cpuid]; 458 continue; 459 } 460 461 /* 462 * We cannot become the current lwp, place the lp on the 463 * run-queue of this or another cpu and deschedule ourselves. 464 * 465 * When we are reactivated we will have another chance. 466 * 467 * Reload after a switch or setrunqueue/switch possibly 468 * moved us to another cpu. 469 */ 470 spin_unlock(&dd->spin); 471 lwkt_deschedule(lp->lwp_thread); 472 dfly_setrunqueue_dd(dd, lp); 473 lwkt_switch(); 474 gd = mycpu; 475 dd = &dfly_pcpu[gd->gd_cpuid]; 476 } 477 478 /* 479 * Make sure upri is synchronized, then yield to LWKT threads as 480 * needed before returning. This could result in another reschedule. 481 * XXX 482 */ 483 crit_exit_quick(td); 484 485 KKASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) == 0); 486 } 487 488 /* 489 * DFLY_RELEASE_CURPROC 490 * 491 * This routine detaches the current thread from the userland scheduler, 492 * usually because the thread needs to run or block in the kernel (at 493 * kernel priority) for a while. 494 * 495 * This routine is also responsible for selecting a new thread to 496 * make the current thread. 497 * 498 * NOTE: This implementation differs from the dummy example in that 499 * dfly_select_curproc() is able to select the current process, whereas 500 * dummy_select_curproc() is not able to select the current process. 501 * This means we have to NULL out uschedcp. 502 * 503 * Additionally, note that we may already be on a run queue if releasing 504 * via the lwkt_switch() in dfly_setrunqueue(). 505 */ 506 static void 507 dfly_release_curproc(struct lwp *lp) 508 { 509 globaldata_t gd = mycpu; 510 dfly_pcpu_t dd = &dfly_pcpu[gd->gd_cpuid]; 511 512 /* 513 * Make sure td_wakefromcpu is defaulted. This will be overwritten 514 * by wakeup(). 515 */ 516 if (dd->uschedcp == lp) { 517 KKASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) == 0); 518 spin_lock(&dd->spin); 519 if (dd->uschedcp == lp) { 520 dd->uschedcp = NULL; /* don't let lp be selected */ 521 dd->upri = PRIBASE_NULL; 522 ATOMIC_CPUMASK_NANDBIT(dfly_curprocmask, gd->gd_cpuid); 523 spin_unlock(&dd->spin); 524 dfly_select_curproc(gd); 525 } else { 526 spin_unlock(&dd->spin); 527 } 528 } 529 } 530 531 /* 532 * DFLY_SELECT_CURPROC 533 * 534 * Select a new current process for this cpu and clear any pending user 535 * reschedule request. The cpu currently has no current process. 536 * 537 * This routine is also responsible for equal-priority round-robining, 538 * typically triggered from dfly_schedulerclock(). In our dummy example 539 * all the 'user' threads are LWKT scheduled all at once and we just 540 * call lwkt_switch(). 541 * 542 * The calling process is not on the queue and cannot be selected. 543 */ 544 static 545 void 546 dfly_select_curproc(globaldata_t gd) 547 { 548 dfly_pcpu_t dd = &dfly_pcpu[gd->gd_cpuid]; 549 struct lwp *nlp; 550 int cpuid = gd->gd_cpuid; 551 552 crit_enter_gd(gd); 553 554 spin_lock(&dd->spin); 555 nlp = dfly_chooseproc_locked(dd, dd, dd->uschedcp, 0); 556 557 if (nlp) { 558 ATOMIC_CPUMASK_ORBIT(dfly_curprocmask, cpuid); 559 dd->upri = nlp->lwp_priority; 560 dd->uschedcp = nlp; 561 #if 0 562 dd->rrcount = 0; /* reset round robin */ 563 #endif 564 spin_unlock(&dd->spin); 565 lwkt_acquire(nlp->lwp_thread); 566 lwkt_schedule(nlp->lwp_thread); 567 } else { 568 spin_unlock(&dd->spin); 569 } 570 crit_exit_gd(gd); 571 } 572 573 /* 574 * Place the specified lwp on the user scheduler's run queue. This routine 575 * must be called with the thread descheduled. The lwp must be runnable. 576 * It must not be possible for anyone else to explicitly schedule this thread. 577 * 578 * The thread may be the current thread as a special case. 579 */ 580 static void 581 dfly_setrunqueue(struct lwp *lp) 582 { 583 dfly_pcpu_t dd; 584 dfly_pcpu_t rdd; 585 586 /* 587 * First validate the process LWKT state. 588 */ 589 KASSERT(lp->lwp_stat == LSRUN, ("setrunqueue: lwp not LSRUN")); 590 KASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) == 0, 591 ("lwp %d/%d already on runq! flag %08x/%08x", lp->lwp_proc->p_pid, 592 lp->lwp_tid, lp->lwp_proc->p_flags, lp->lwp_flags)); 593 KKASSERT((lp->lwp_thread->td_flags & TDF_RUNQ) == 0); 594 595 /* 596 * NOTE: dd/rdd do not necessarily represent the current cpu. 597 * Instead they may represent the cpu the thread was last 598 * scheduled on or inherited by its parent. 599 */ 600 dd = &dfly_pcpu[lp->lwp_qcpu]; 601 rdd = dd; 602 603 /* 604 * This process is not supposed to be scheduled anywhere or assigned 605 * as the current process anywhere. Assert the condition. 606 */ 607 KKASSERT(rdd->uschedcp != lp); 608 609 /* 610 * Ok, we have to setrunqueue some target cpu and request a reschedule 611 * if necessary. 612 * 613 * We have to choose the best target cpu. It might not be the current 614 * target even if the current cpu has no running user thread (for 615 * example, because the current cpu might be a hyperthread and its 616 * sibling has a thread assigned). 617 * 618 * If we just forked it is most optimal to run the child on the same 619 * cpu just in case the parent decides to wait for it (thus getting 620 * off that cpu). As long as there is nothing else runnable on the 621 * cpu, that is. If we did this unconditionally a parent forking 622 * multiple children before waiting (e.g. make -j N) leaves other 623 * cpus idle that could be working. 624 */ 625 if (lp->lwp_forked) { 626 lp->lwp_forked = 0; 627 if (usched_dfly_features & 0x20) 628 rdd = dfly_choose_best_queue(lp); 629 else if (usched_dfly_features & 0x40) 630 rdd = &dfly_pcpu[lp->lwp_qcpu]; 631 else if (usched_dfly_features & 0x80) 632 rdd = dfly_choose_queue_simple(rdd, lp); 633 else if (dfly_pcpu[lp->lwp_qcpu].runqcount) 634 rdd = dfly_choose_best_queue(lp); 635 else 636 rdd = &dfly_pcpu[lp->lwp_qcpu]; 637 } else { 638 rdd = dfly_choose_best_queue(lp); 639 /* rdd = &dfly_pcpu[lp->lwp_qcpu]; */ 640 } 641 if (lp->lwp_qcpu != rdd->cpuid) { 642 spin_lock(&dd->spin); 643 dfly_changeqcpu_locked(lp, dd, rdd); 644 spin_unlock(&dd->spin); 645 } 646 dfly_setrunqueue_dd(rdd, lp); 647 } 648 649 /* 650 * Change qcpu to rdd->cpuid. The dd the lp is CURRENTLY on must be 651 * spin-locked on-call. rdd does not have to be. 652 */ 653 static void 654 dfly_changeqcpu_locked(struct lwp *lp, dfly_pcpu_t dd, dfly_pcpu_t rdd) 655 { 656 if (lp->lwp_qcpu != rdd->cpuid) { 657 if (lp->lwp_mpflags & LWP_MP_ULOAD) { 658 atomic_clear_int(&lp->lwp_mpflags, LWP_MP_ULOAD); 659 atomic_add_int(&dd->uload, -lp->lwp_uload); 660 atomic_add_int(&dd->ucount, -1); 661 atomic_add_int(&dfly_ucount, -1); 662 } 663 lp->lwp_qcpu = rdd->cpuid; 664 } 665 } 666 667 /* 668 * Place lp on rdd's runqueue. Nothing is locked on call. This function 669 * also performs all necessary ancillary notification actions. 670 */ 671 static void 672 dfly_setrunqueue_dd(dfly_pcpu_t rdd, struct lwp *lp) 673 { 674 globaldata_t rgd; 675 676 /* 677 * We might be moving the lp to another cpu's run queue, and once 678 * on the runqueue (even if it is our cpu's), another cpu can rip 679 * it away from us. 680 * 681 * TDF_MIGRATING might already be set if this is part of a 682 * remrunqueue+setrunqueue sequence. 683 */ 684 if ((lp->lwp_thread->td_flags & TDF_MIGRATING) == 0) 685 lwkt_giveaway(lp->lwp_thread); 686 687 rgd = globaldata_find(rdd->cpuid); 688 689 /* 690 * We lose control of the lp the moment we release the spinlock 691 * after having placed it on the queue. i.e. another cpu could pick 692 * it up, or it could exit, or its priority could be further 693 * adjusted, or something like that. 694 * 695 * WARNING! rdd can point to a foreign cpu! 696 */ 697 spin_lock(&rdd->spin); 698 dfly_setrunqueue_locked(rdd, lp); 699 700 /* 701 * Potentially interrupt the currently-running thread 702 */ 703 if ((rdd->upri & ~PPQMASK) <= (lp->lwp_priority & ~PPQMASK)) { 704 /* 705 * Currently running thread is better or same, do not 706 * interrupt. 707 */ 708 spin_unlock(&rdd->spin); 709 } else if ((rdd->upri & ~PPQMASK) <= (lp->lwp_priority & ~PPQMASK) + 710 usched_dfly_fast_resched) { 711 /* 712 * Currently running thread is not better, but not so bad 713 * that we need to interrupt it. Let it run for one more 714 * scheduler tick. 715 */ 716 if (rdd->uschedcp && 717 rdd->uschedcp->lwp_rrcount < usched_dfly_rrinterval) { 718 rdd->uschedcp->lwp_rrcount = usched_dfly_rrinterval - 1; 719 } 720 spin_unlock(&rdd->spin); 721 } else if (rgd == mycpu) { 722 /* 723 * We should interrupt the currently running thread, which 724 * is on the current cpu. However, if DIDYIELD is set we 725 * round-robin unconditionally and do not interrupt it. 726 */ 727 spin_unlock(&rdd->spin); 728 if (rdd->uschedcp == NULL) 729 wakeup_mycpu(&rdd->helper_thread); /* XXX */ 730 if ((lp->lwp_thread->td_mpflags & TDF_MP_DIDYIELD) == 0) 731 need_user_resched(); 732 } else { 733 /* 734 * We should interrupt the currently running thread, which 735 * is on a different cpu. 736 */ 737 spin_unlock(&rdd->spin); 738 lwkt_send_ipiq(rgd, dfly_need_user_resched_remote, NULL); 739 } 740 } 741 742 /* 743 * This routine is called from a systimer IPI. It MUST be MP-safe and 744 * the BGL IS NOT HELD ON ENTRY. This routine is called at ESTCPUFREQ on 745 * each cpu. 746 */ 747 static 748 void 749 dfly_schedulerclock(struct lwp *lp, sysclock_t period, sysclock_t cpstamp) 750 { 751 globaldata_t gd = mycpu; 752 dfly_pcpu_t dd = &dfly_pcpu[gd->gd_cpuid]; 753 754 /* 755 * Spinlocks also hold a critical section so there should not be 756 * any active. 757 */ 758 KKASSERT(gd->gd_spinlocks == 0); 759 760 if (lp == NULL) 761 return; 762 763 /* 764 * Do we need to round-robin? We round-robin 10 times a second. 765 * This should only occur for cpu-bound batch processes. 766 */ 767 if (++lp->lwp_rrcount >= usched_dfly_rrinterval) { 768 lp->lwp_thread->td_wakefromcpu = -1; 769 need_user_resched(); 770 } 771 772 /* 773 * Adjust estcpu upward using a real time equivalent calculation, 774 * and recalculate lp's priority. 775 */ 776 lp->lwp_estcpu = ESTCPULIM(lp->lwp_estcpu + ESTCPUMAX / ESTCPUFREQ + 1); 777 dfly_resetpriority(lp); 778 779 /* 780 * Rebalance two cpus every 8 ticks, pulling the worst thread 781 * from the worst cpu's queue into a rotating cpu number. 782 * 783 * This mechanic is needed because the push algorithms can 784 * steady-state in an non-optimal configuration. We need to mix it 785 * up a little, even if it means breaking up a paired thread, so 786 * the push algorithms can rebalance the degenerate conditions. 787 * This portion of the algorithm exists to ensure stability at the 788 * selected weightings. 789 * 790 * Because we might be breaking up optimal conditions we do not want 791 * to execute this too quickly, hence we only rebalance approximately 792 * ~7-8 times per second. The push's, on the otherhand, are capable 793 * moving threads to other cpus at a much higher rate. 794 * 795 * We choose the most heavily loaded thread from the worst queue 796 * in order to ensure that multiple heavy-weight threads on the same 797 * queue get broken up, and also because these threads are the most 798 * likely to be able to remain in place. Hopefully then any pairings, 799 * if applicable, migrate to where these threads are. 800 */ 801 if ((usched_dfly_features & 0x04) && 802 ((u_int)sched_ticks & 7) == 0 && 803 (u_int)sched_ticks / 8 % ncpus == gd->gd_cpuid) { 804 /* 805 * Our cpu is up. 806 */ 807 struct lwp *nlp; 808 dfly_pcpu_t rdd; 809 810 rdd = dfly_choose_worst_queue(dd); 811 if (rdd) { 812 spin_lock(&dd->spin); 813 if (spin_trylock(&rdd->spin)) { 814 nlp = dfly_chooseproc_locked(rdd, dd, NULL, 1); 815 spin_unlock(&rdd->spin); 816 if (nlp == NULL) 817 spin_unlock(&dd->spin); 818 } else { 819 spin_unlock(&dd->spin); 820 nlp = NULL; 821 } 822 } else { 823 nlp = NULL; 824 } 825 /* dd->spin held if nlp != NULL */ 826 827 /* 828 * Either schedule it or add it to our queue. 829 */ 830 if (nlp && 831 (nlp->lwp_priority & ~PPQMASK) < (dd->upri & ~PPQMASK)) { 832 ATOMIC_CPUMASK_ORMASK(dfly_curprocmask, dd->cpumask); 833 dd->upri = nlp->lwp_priority; 834 dd->uschedcp = nlp; 835 #if 0 836 dd->rrcount = 0; /* reset round robin */ 837 #endif 838 spin_unlock(&dd->spin); 839 lwkt_acquire(nlp->lwp_thread); 840 lwkt_schedule(nlp->lwp_thread); 841 } else if (nlp) { 842 dfly_setrunqueue_locked(dd, nlp); 843 spin_unlock(&dd->spin); 844 } 845 } 846 } 847 848 /* 849 * Called from acquire and from kern_synch's one-second timer (one of the 850 * callout helper threads) with a critical section held. 851 * 852 * Adjust p_estcpu based on our single-cpu load, p_nice, and compensate for 853 * overall system load. 854 * 855 * Note that no recalculation occurs for a process which sleeps and wakes 856 * up in the same tick. That is, a system doing thousands of context 857 * switches per second will still only do serious estcpu calculations 858 * ESTCPUFREQ times per second. 859 */ 860 static 861 void 862 dfly_recalculate_estcpu(struct lwp *lp) 863 { 864 globaldata_t gd = mycpu; 865 sysclock_t cpbase; 866 sysclock_t ttlticks; 867 int estcpu; 868 int decay_factor; 869 int ucount; 870 871 /* 872 * We have to subtract periodic to get the last schedclock 873 * timeout time, otherwise we would get the upcoming timeout. 874 * Keep in mind that a process can migrate between cpus and 875 * while the scheduler clock should be very close, boundary 876 * conditions could lead to a small negative delta. 877 */ 878 cpbase = gd->gd_schedclock.time - gd->gd_schedclock.periodic; 879 880 if (lp->lwp_slptime > 1) { 881 /* 882 * Too much time has passed, do a coarse correction. 883 */ 884 lp->lwp_estcpu = lp->lwp_estcpu >> 1; 885 dfly_resetpriority(lp); 886 lp->lwp_cpbase = cpbase; 887 lp->lwp_cpticks = 0; 888 lp->lwp_estfast = 0; 889 } else if (lp->lwp_cpbase != cpbase) { 890 /* 891 * Adjust estcpu if we are in a different tick. Don't waste 892 * time if we are in the same tick. 893 * 894 * First calculate the number of ticks in the measurement 895 * interval. The ttlticks calculation can wind up 0 due to 896 * a bug in the handling of lwp_slptime (as yet not found), 897 * so make sure we do not get a divide by 0 panic. 898 */ 899 ttlticks = (cpbase - lp->lwp_cpbase) / 900 gd->gd_schedclock.periodic; 901 if ((ssysclock_t)ttlticks < 0) { 902 ttlticks = 0; 903 lp->lwp_cpbase = cpbase; 904 } 905 if (ttlticks == 0) 906 return; 907 updatepcpu(lp, lp->lwp_cpticks, ttlticks); 908 909 /* 910 * Calculate the percentage of one cpu being used then 911 * compensate for any system load in excess of ncpus. 912 * 913 * For example, if we have 8 cores and 16 running cpu-bound 914 * processes then all things being equal each process will 915 * get 50% of one cpu. We need to pump this value back 916 * up to 100% so the estcpu calculation properly adjusts 917 * the process's dynamic priority. 918 * 919 * estcpu is scaled by ESTCPUMAX, pctcpu is scaled by FSCALE. 920 */ 921 estcpu = (lp->lwp_pctcpu * ESTCPUMAX) >> FSHIFT; 922 ucount = dfly_ucount; 923 if (ucount > ncpus) { 924 estcpu += estcpu * (ucount - ncpus) / ncpus; 925 } 926 927 if (usched_dfly_debug == lp->lwp_proc->p_pid) { 928 kprintf("pid %d lwp %p estcpu %3d %3d cp %d/%d", 929 lp->lwp_proc->p_pid, lp, 930 estcpu, lp->lwp_estcpu, 931 lp->lwp_cpticks, ttlticks); 932 } 933 934 /* 935 * Adjust lp->lwp_esetcpu. The decay factor determines how 936 * quickly lwp_estcpu collapses to its realtime calculation. 937 * A slower collapse gives us a more accurate number over 938 * the long term but can create problems with bursty threads 939 * or threads which become cpu hogs. 940 * 941 * To solve this problem, newly started lwps and lwps which 942 * are restarting after having been asleep for a while are 943 * given a much, much faster decay in order to quickly 944 * detect whether they become cpu-bound. 945 * 946 * NOTE: p_nice is accounted for in dfly_resetpriority(), 947 * and not here, but we must still ensure that a 948 * cpu-bound nice -20 process does not completely 949 * override a cpu-bound nice +20 process. 950 * 951 * NOTE: We must use ESTCPULIM() here to deal with any 952 * overshoot. 953 */ 954 decay_factor = usched_dfly_decay; 955 if (decay_factor < 1) 956 decay_factor = 1; 957 if (decay_factor > 1024) 958 decay_factor = 1024; 959 960 if (lp->lwp_estfast < usched_dfly_decay) { 961 ++lp->lwp_estfast; 962 lp->lwp_estcpu = ESTCPULIM( 963 (lp->lwp_estcpu * lp->lwp_estfast + estcpu) / 964 (lp->lwp_estfast + 1)); 965 } else { 966 lp->lwp_estcpu = ESTCPULIM( 967 (lp->lwp_estcpu * decay_factor + estcpu) / 968 (decay_factor + 1)); 969 } 970 971 if (usched_dfly_debug == lp->lwp_proc->p_pid) 972 kprintf(" finalestcpu %d\n", lp->lwp_estcpu); 973 dfly_resetpriority(lp); 974 lp->lwp_cpbase += ttlticks * gd->gd_schedclock.periodic; 975 lp->lwp_cpticks = 0; 976 } 977 } 978 979 /* 980 * Compute the priority of a process when running in user mode. 981 * Arrange to reschedule if the resulting priority is better 982 * than that of the current process. 983 * 984 * This routine may be called with any process. 985 * 986 * This routine is called by fork1() for initial setup with the process of 987 * the run queue, and also may be called normally with the process on or 988 * off the run queue. 989 */ 990 static void 991 dfly_resetpriority(struct lwp *lp) 992 { 993 dfly_pcpu_t rdd; 994 int newpriority; 995 u_short newrqtype; 996 int rcpu; 997 int checkpri; 998 int estcpu; 999 int delta_uload; 1000 1001 crit_enter(); 1002 1003 /* 1004 * Lock the scheduler (lp) belongs to. This can be on a different 1005 * cpu. Handle races. This loop breaks out with the appropriate 1006 * rdd locked. 1007 */ 1008 for (;;) { 1009 rcpu = lp->lwp_qcpu; 1010 cpu_ccfence(); 1011 rdd = &dfly_pcpu[rcpu]; 1012 spin_lock(&rdd->spin); 1013 if (rcpu == lp->lwp_qcpu) 1014 break; 1015 spin_unlock(&rdd->spin); 1016 } 1017 1018 /* 1019 * Calculate the new priority and queue type 1020 */ 1021 newrqtype = lp->lwp_rtprio.type; 1022 1023 switch(newrqtype) { 1024 case RTP_PRIO_REALTIME: 1025 case RTP_PRIO_FIFO: 1026 newpriority = PRIBASE_REALTIME + 1027 (lp->lwp_rtprio.prio & PRIMASK); 1028 break; 1029 case RTP_PRIO_NORMAL: 1030 /* 1031 * 1032 */ 1033 estcpu = lp->lwp_estcpu; 1034 1035 /* 1036 * p_nice piece Adds (0-40) * 2 0-80 1037 * estcpu Adds 16384 * 4 / 512 0-128 1038 */ 1039 newpriority = (lp->lwp_proc->p_nice - PRIO_MIN) * PPQ / NICEPPQ; 1040 newpriority += estcpu * PPQ / ESTCPUPPQ; 1041 newpriority = newpriority * MAXPRI / (PRIO_RANGE * PPQ / 1042 NICEPPQ + ESTCPUMAX * PPQ / ESTCPUPPQ); 1043 newpriority = PRIBASE_NORMAL + (newpriority & PRIMASK); 1044 break; 1045 case RTP_PRIO_IDLE: 1046 newpriority = PRIBASE_IDLE + (lp->lwp_rtprio.prio & PRIMASK); 1047 break; 1048 case RTP_PRIO_THREAD: 1049 newpriority = PRIBASE_THREAD + (lp->lwp_rtprio.prio & PRIMASK); 1050 break; 1051 default: 1052 panic("Bad RTP_PRIO %d", newrqtype); 1053 /* NOT REACHED */ 1054 } 1055 1056 /* 1057 * The LWKT scheduler doesn't dive usched structures, give it a hint 1058 * on the relative priority of user threads running in the kernel. 1059 * The LWKT scheduler will always ensure that a user thread running 1060 * in the kernel will get cpu some time, regardless of its upri, 1061 * but can decide not to instantly switch from one kernel or user 1062 * mode user thread to a kernel-mode user thread when it has a less 1063 * desireable user priority. 1064 * 1065 * td_upri has normal sense (higher values are more desireable), so 1066 * negate it. 1067 */ 1068 lp->lwp_thread->td_upri = -(newpriority & usched_dfly_swmask); 1069 1070 /* 1071 * The newpriority incorporates the queue type so do a simple masked 1072 * check to determine if the process has moved to another queue. If 1073 * it has, and it is currently on a run queue, then move it. 1074 * 1075 * Since uload is ~PPQMASK masked, no modifications are necessary if 1076 * we end up in the same run queue. 1077 * 1078 * Reset rrcount if moving to a higher-priority queue, otherwise 1079 * retain rrcount. 1080 */ 1081 if ((lp->lwp_priority ^ newpriority) & ~PPQMASK) { 1082 if (lp->lwp_priority < newpriority) 1083 lp->lwp_rrcount = 0; 1084 if (lp->lwp_mpflags & LWP_MP_ONRUNQ) { 1085 dfly_remrunqueue_locked(rdd, lp); 1086 lp->lwp_priority = newpriority; 1087 lp->lwp_rqtype = newrqtype; 1088 lp->lwp_rqindex = (newpriority & PRIMASK) / PPQ; 1089 dfly_setrunqueue_locked(rdd, lp); 1090 checkpri = 1; 1091 } else { 1092 lp->lwp_priority = newpriority; 1093 lp->lwp_rqtype = newrqtype; 1094 lp->lwp_rqindex = (newpriority & PRIMASK) / PPQ; 1095 checkpri = 0; 1096 } 1097 } else { 1098 /* 1099 * In the same PPQ, uload cannot change. 1100 */ 1101 lp->lwp_priority = newpriority; 1102 checkpri = 1; 1103 rcpu = -1; 1104 } 1105 1106 /* 1107 * Adjust effective load. 1108 * 1109 * Calculate load then scale up or down geometrically based on p_nice. 1110 * Processes niced up (positive) are less important, and processes 1111 * niced downard (negative) are more important. The higher the uload, 1112 * the more important the thread. 1113 */ 1114 /* 0-511, 0-100% cpu */ 1115 delta_uload = lp->lwp_estcpu / NQS; 1116 delta_uload -= delta_uload * lp->lwp_proc->p_nice / (PRIO_MAX + 1); 1117 1118 1119 delta_uload -= lp->lwp_uload; 1120 lp->lwp_uload += delta_uload; 1121 if (lp->lwp_mpflags & LWP_MP_ULOAD) 1122 atomic_add_int(&dfly_pcpu[lp->lwp_qcpu].uload, delta_uload); 1123 1124 /* 1125 * Determine if we need to reschedule the target cpu. This only 1126 * occurs if the LWP is already on a scheduler queue, which means 1127 * that idle cpu notification has already occured. At most we 1128 * need only issue a need_user_resched() on the appropriate cpu. 1129 * 1130 * The LWP may be owned by a CPU different from the current one, 1131 * in which case dd->uschedcp may be modified without an MP lock 1132 * or a spinlock held. The worst that happens is that the code 1133 * below causes a spurious need_user_resched() on the target CPU 1134 * and dd->pri to be wrong for a short period of time, both of 1135 * which are harmless. 1136 * 1137 * If checkpri is 0 we are adjusting the priority of the current 1138 * process, possibly higher (less desireable), so ignore the upri 1139 * check which will fail in that case. 1140 */ 1141 if (rcpu >= 0) { 1142 if (CPUMASK_TESTBIT(dfly_rdyprocmask, rcpu) && 1143 (checkpri == 0 || 1144 (rdd->upri & ~PRIMASK) > 1145 (lp->lwp_priority & ~PRIMASK))) { 1146 if (rcpu == mycpu->gd_cpuid) { 1147 spin_unlock(&rdd->spin); 1148 need_user_resched(); 1149 } else { 1150 spin_unlock(&rdd->spin); 1151 lwkt_send_ipiq(globaldata_find(rcpu), 1152 dfly_need_user_resched_remote, 1153 NULL); 1154 } 1155 } else { 1156 spin_unlock(&rdd->spin); 1157 } 1158 } else { 1159 spin_unlock(&rdd->spin); 1160 } 1161 crit_exit(); 1162 } 1163 1164 static 1165 void 1166 dfly_yield(struct lwp *lp) 1167 { 1168 if (lp->lwp_qcpu != mycpu->gd_cpuid) 1169 return; 1170 KKASSERT(lp == curthread->td_lwp); 1171 1172 /* 1173 * Don't set need_user_resched() or mess with rrcount or anything. 1174 * the TDF flag will override everything as long as we release. 1175 */ 1176 atomic_set_int(&lp->lwp_thread->td_mpflags, TDF_MP_DIDYIELD); 1177 dfly_release_curproc(lp); 1178 } 1179 1180 /* 1181 * Thread was forcefully migrated to another cpu. Normally forced migrations 1182 * are used for iterations and the kernel returns to the original cpu before 1183 * returning and this is not needed. However, if the kernel migrates a 1184 * thread to another cpu and wants to leave it there, it has to call this 1185 * scheduler helper. 1186 * 1187 * Note that the lwkt_migratecpu() function also released the thread, so 1188 * we don't have to worry about that. 1189 */ 1190 static 1191 void 1192 dfly_changedcpu(struct lwp *lp) 1193 { 1194 dfly_pcpu_t dd = &dfly_pcpu[lp->lwp_qcpu]; 1195 dfly_pcpu_t rdd = &dfly_pcpu[mycpu->gd_cpuid]; 1196 1197 if (dd != rdd) { 1198 spin_lock(&dd->spin); 1199 dfly_changeqcpu_locked(lp, dd, rdd); 1200 spin_unlock(&dd->spin); 1201 } 1202 } 1203 1204 /* 1205 * Called from fork1() when a new child process is being created. 1206 * 1207 * Give the child process an initial estcpu that is more batch then 1208 * its parent and dock the parent for the fork (but do not 1209 * reschedule the parent). 1210 * 1211 * fast 1212 * 1213 * XXX lwp should be "spawning" instead of "forking" 1214 */ 1215 static void 1216 dfly_forking(struct lwp *plp, struct lwp *lp) 1217 { 1218 /* 1219 * Put the child 4 queue slots (out of 32) higher than the parent 1220 * (less desireable than the parent). 1221 */ 1222 lp->lwp_estcpu = ESTCPULIM(plp->lwp_estcpu + ESTCPUPPQ * 4); 1223 lp->lwp_forked = 1; 1224 lp->lwp_estfast = 0; 1225 1226 /* 1227 * Dock the parent a cost for the fork, protecting us from fork 1228 * bombs. If the parent is forking quickly make the child more 1229 * batchy. 1230 */ 1231 plp->lwp_estcpu = ESTCPULIM(plp->lwp_estcpu + ESTCPUPPQ / 16); 1232 } 1233 1234 /* 1235 * Called when a lwp is being removed from this scheduler, typically 1236 * during lwp_exit(). We have to clean out any ULOAD accounting before 1237 * we can let the lp go. The dd->spin lock is not needed for uload 1238 * updates. 1239 * 1240 * Scheduler dequeueing has already occurred, no further action in that 1241 * regard is needed. 1242 */ 1243 static void 1244 dfly_exiting(struct lwp *lp, struct proc *child_proc) 1245 { 1246 dfly_pcpu_t dd = &dfly_pcpu[lp->lwp_qcpu]; 1247 1248 if (lp->lwp_mpflags & LWP_MP_ULOAD) { 1249 atomic_clear_int(&lp->lwp_mpflags, LWP_MP_ULOAD); 1250 atomic_add_int(&dd->uload, -lp->lwp_uload); 1251 atomic_add_int(&dd->ucount, -1); 1252 atomic_add_int(&dfly_ucount, -1); 1253 } 1254 } 1255 1256 /* 1257 * This function cannot block in any way, but spinlocks are ok. 1258 * 1259 * Update the uload based on the state of the thread (whether it is going 1260 * to sleep or running again). The uload is meant to be a longer-term 1261 * load and not an instantanious load. 1262 */ 1263 static void 1264 dfly_uload_update(struct lwp *lp) 1265 { 1266 dfly_pcpu_t dd = &dfly_pcpu[lp->lwp_qcpu]; 1267 1268 if (lp->lwp_thread->td_flags & TDF_RUNQ) { 1269 if ((lp->lwp_mpflags & LWP_MP_ULOAD) == 0) { 1270 spin_lock(&dd->spin); 1271 if ((lp->lwp_mpflags & LWP_MP_ULOAD) == 0) { 1272 atomic_set_int(&lp->lwp_mpflags, 1273 LWP_MP_ULOAD); 1274 atomic_add_int(&dd->uload, lp->lwp_uload); 1275 atomic_add_int(&dd->ucount, 1); 1276 atomic_add_int(&dfly_ucount, 1); 1277 } 1278 spin_unlock(&dd->spin); 1279 } 1280 } else if (lp->lwp_slptime > 0) { 1281 if (lp->lwp_mpflags & LWP_MP_ULOAD) { 1282 spin_lock(&dd->spin); 1283 if (lp->lwp_mpflags & LWP_MP_ULOAD) { 1284 atomic_clear_int(&lp->lwp_mpflags, 1285 LWP_MP_ULOAD); 1286 atomic_add_int(&dd->uload, -lp->lwp_uload); 1287 atomic_add_int(&dd->ucount, -1); 1288 atomic_add_int(&dfly_ucount, -1); 1289 } 1290 spin_unlock(&dd->spin); 1291 } 1292 } 1293 } 1294 1295 /* 1296 * chooseproc() is called when a cpu needs a user process to LWKT schedule, 1297 * it selects a user process and returns it. If chklp is non-NULL and chklp 1298 * has a better or equal priority then the process that would otherwise be 1299 * chosen, NULL is returned. 1300 * 1301 * Until we fix the RUNQ code the chklp test has to be strict or we may 1302 * bounce between processes trying to acquire the current process designation. 1303 * 1304 * Must be called with rdd->spin locked. The spinlock is left intact through 1305 * the entire routine. dd->spin does not have to be locked. 1306 * 1307 * If worst is non-zero this function finds the worst thread instead of the 1308 * best thread (used by the schedulerclock-based rover). 1309 */ 1310 static 1311 struct lwp * 1312 dfly_chooseproc_locked(dfly_pcpu_t rdd, dfly_pcpu_t dd, 1313 struct lwp *chklp, int worst) 1314 { 1315 struct lwp *lp; 1316 struct rq *q; 1317 u_int32_t *which; 1318 u_int32_t pri; 1319 u_int32_t rtqbits; 1320 u_int32_t tsqbits; 1321 u_int32_t idqbits; 1322 1323 rtqbits = rdd->rtqueuebits; 1324 tsqbits = rdd->queuebits; 1325 idqbits = rdd->idqueuebits; 1326 1327 if (worst) { 1328 if (idqbits) { 1329 pri = bsrl(idqbits); 1330 q = &rdd->idqueues[pri]; 1331 which = &rdd->idqueuebits; 1332 } else if (tsqbits) { 1333 pri = bsrl(tsqbits); 1334 q = &rdd->queues[pri]; 1335 which = &rdd->queuebits; 1336 } else if (rtqbits) { 1337 pri = bsrl(rtqbits); 1338 q = &rdd->rtqueues[pri]; 1339 which = &rdd->rtqueuebits; 1340 } else { 1341 return (NULL); 1342 } 1343 lp = TAILQ_LAST(q, rq); 1344 } else { 1345 if (rtqbits) { 1346 pri = bsfl(rtqbits); 1347 q = &rdd->rtqueues[pri]; 1348 which = &rdd->rtqueuebits; 1349 } else if (tsqbits) { 1350 pri = bsfl(tsqbits); 1351 q = &rdd->queues[pri]; 1352 which = &rdd->queuebits; 1353 } else if (idqbits) { 1354 pri = bsfl(idqbits); 1355 q = &rdd->idqueues[pri]; 1356 which = &rdd->idqueuebits; 1357 } else { 1358 return (NULL); 1359 } 1360 lp = TAILQ_FIRST(q); 1361 } 1362 KASSERT(lp, ("chooseproc: no lwp on busy queue")); 1363 1364 /* 1365 * If the passed lwp <chklp> is reasonably close to the selected 1366 * lwp <lp>, return NULL (indicating that <chklp> should be kept). 1367 * 1368 * Note that we must error on the side of <chklp> to avoid bouncing 1369 * between threads in the acquire code. 1370 */ 1371 if (chklp) { 1372 if (chklp->lwp_priority < lp->lwp_priority + PPQ) 1373 return(NULL); 1374 } 1375 1376 KTR_COND_LOG(usched_chooseproc, 1377 lp->lwp_proc->p_pid == usched_dfly_pid_debug, 1378 lp->lwp_proc->p_pid, 1379 lp->lwp_thread->td_gd->gd_cpuid, 1380 mycpu->gd_cpuid); 1381 1382 KASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) != 0, ("not on runq6!")); 1383 atomic_clear_int(&lp->lwp_mpflags, LWP_MP_ONRUNQ); 1384 TAILQ_REMOVE(q, lp, lwp_procq); 1385 --rdd->runqcount; 1386 if (TAILQ_EMPTY(q)) 1387 *which &= ~(1 << pri); 1388 1389 /* 1390 * If we are choosing a process from rdd with the intent to 1391 * move it to dd, lwp_qcpu must be adjusted while rdd's spinlock 1392 * is still held. 1393 */ 1394 if (rdd != dd) { 1395 if (lp->lwp_mpflags & LWP_MP_ULOAD) { 1396 atomic_add_int(&rdd->uload, -lp->lwp_uload); 1397 atomic_add_int(&rdd->ucount, -1); 1398 atomic_add_int(&dfly_ucount, -1); 1399 } 1400 lp->lwp_qcpu = dd->cpuid; 1401 atomic_add_int(&dd->uload, lp->lwp_uload); 1402 atomic_add_int(&dd->ucount, 1); 1403 atomic_add_int(&dfly_ucount, 1); 1404 atomic_set_int(&lp->lwp_mpflags, LWP_MP_ULOAD); 1405 } 1406 return lp; 1407 } 1408 1409 /* 1410 * USED TO PUSH RUNNABLE LWPS TO THE LEAST LOADED CPU. 1411 * 1412 * Choose a cpu node to schedule lp on, hopefully nearby its current 1413 * node. 1414 * 1415 * We give the current node a modest advantage for obvious reasons. 1416 * 1417 * We also give the node the thread was woken up FROM a slight advantage 1418 * in order to try to schedule paired threads which synchronize/block waiting 1419 * for each other fairly close to each other. Similarly in a network setting 1420 * this feature will also attempt to place a user process near the kernel 1421 * protocol thread that is feeding it data. THIS IS A CRITICAL PART of the 1422 * algorithm as it heuristically groups synchronizing processes for locality 1423 * of reference in multi-socket systems. 1424 * 1425 * We check against running processes and give a big advantage if there 1426 * are none running. 1427 * 1428 * The caller will normally dfly_setrunqueue() lp on the returned queue. 1429 * 1430 * When the topology is known choose a cpu whos group has, in aggregate, 1431 * has the lowest weighted load. 1432 */ 1433 static 1434 dfly_pcpu_t 1435 dfly_choose_best_queue(struct lwp *lp) 1436 { 1437 cpumask_t wakemask; 1438 cpumask_t mask; 1439 cpu_node_t *cpup; 1440 cpu_node_t *cpun; 1441 cpu_node_t *cpub; 1442 dfly_pcpu_t dd = &dfly_pcpu[lp->lwp_qcpu]; 1443 dfly_pcpu_t rdd; 1444 int wakecpu; 1445 int cpuid; 1446 int n; 1447 int count; 1448 int load; 1449 int lowest_load; 1450 1451 /* 1452 * When the topology is unknown choose a random cpu that is hopefully 1453 * idle. 1454 */ 1455 if (dd->cpunode == NULL) 1456 return (dfly_choose_queue_simple(dd, lp)); 1457 1458 /* 1459 * Pairing mask 1460 */ 1461 if ((wakecpu = lp->lwp_thread->td_wakefromcpu) >= 0) 1462 wakemask = dfly_pcpu[wakecpu].cpumask; 1463 else 1464 CPUMASK_ASSZERO(wakemask); 1465 1466 /* 1467 * When the topology is known choose a cpu whos group has, in 1468 * aggregate, has the lowest weighted load. 1469 */ 1470 cpup = root_cpu_node; 1471 rdd = dd; 1472 1473 while (cpup) { 1474 /* 1475 * Degenerate case super-root 1476 */ 1477 if (cpup->child_no == 1) { 1478 cpup = cpup->child_node[0]; 1479 continue; 1480 } 1481 1482 /* 1483 * Terminal cpunode 1484 */ 1485 if (cpup->child_no == 0) { 1486 rdd = &dfly_pcpu[BSFCPUMASK(cpup->members)]; 1487 break; 1488 } 1489 1490 cpub = NULL; 1491 lowest_load = 0x7FFFFFFF; 1492 1493 for (n = 0; n < cpup->child_no; ++n) { 1494 /* 1495 * Accumulate load information for all cpus 1496 * which are members of this node. 1497 */ 1498 cpun = cpup->child_node[n]; 1499 mask = cpun->members; 1500 CPUMASK_ANDMASK(mask, usched_global_cpumask); 1501 CPUMASK_ANDMASK(mask, smp_active_mask); 1502 CPUMASK_ANDMASK(mask, lp->lwp_cpumask); 1503 if (CPUMASK_TESTZERO(mask)) 1504 continue; 1505 1506 count = 0; 1507 load = 0; 1508 1509 while (CPUMASK_TESTNZERO(mask)) { 1510 cpuid = BSFCPUMASK(mask); 1511 rdd = &dfly_pcpu[cpuid]; 1512 load += rdd->uload; 1513 load += rdd->ucount * usched_dfly_weight3; 1514 1515 if (rdd->uschedcp == NULL && 1516 rdd->runqcount == 0 && 1517 globaldata_find(cpuid)->gd_tdrunqcount == 0 1518 ) { 1519 load -= usched_dfly_weight4; 1520 } 1521 #if 0 1522 else if (rdd->upri > lp->lwp_priority + PPQ) { 1523 load -= usched_dfly_weight4 / 2; 1524 } 1525 #endif 1526 CPUMASK_NANDBIT(mask, cpuid); 1527 ++count; 1528 } 1529 1530 /* 1531 * Compensate if the lp is already accounted for in 1532 * the aggregate uload for this mask set. We want 1533 * to calculate the loads as if lp were not present, 1534 * otherwise the calculation is bogus. 1535 */ 1536 if ((lp->lwp_mpflags & LWP_MP_ULOAD) && 1537 CPUMASK_TESTMASK(dd->cpumask, cpun->members)) { 1538 load -= lp->lwp_uload; 1539 load -= usched_dfly_weight3; 1540 } 1541 1542 load /= count; 1543 1544 /* 1545 * Advantage the cpu group (lp) is already on. 1546 */ 1547 if (CPUMASK_TESTMASK(cpun->members, dd->cpumask)) 1548 load -= usched_dfly_weight1; 1549 1550 /* 1551 * Advantage the cpu group we want to pair (lp) to, 1552 * but don't let it go to the exact same cpu as 1553 * the wakecpu target. 1554 * 1555 * We do this by checking whether cpun is a 1556 * terminal node or not. All cpun's at the same 1557 * level will either all be terminal or all not 1558 * terminal. 1559 * 1560 * If it is and we match we disadvantage the load. 1561 * If it is and we don't match we advantage the load. 1562 * 1563 * Also note that we are effectively disadvantaging 1564 * all-but-one by the same amount, so it won't effect 1565 * the weight1 factor for the all-but-one nodes. 1566 */ 1567 if (CPUMASK_TESTMASK(cpun->members, wakemask)) { 1568 if (cpun->child_no != 0) { 1569 /* advantage */ 1570 load -= usched_dfly_weight2; 1571 } else { 1572 if (usched_dfly_features & 0x10) 1573 load += usched_dfly_weight2; 1574 else 1575 load -= usched_dfly_weight2; 1576 } 1577 } 1578 1579 /* 1580 * Calculate the best load 1581 */ 1582 if (cpub == NULL || lowest_load > load || 1583 (lowest_load == load && 1584 CPUMASK_TESTMASK(cpun->members, dd->cpumask)) 1585 ) { 1586 lowest_load = load; 1587 cpub = cpun; 1588 } 1589 } 1590 cpup = cpub; 1591 } 1592 if (usched_dfly_chooser) 1593 kprintf("lp %02d->%02d %s\n", 1594 lp->lwp_qcpu, rdd->cpuid, lp->lwp_proc->p_comm); 1595 return (rdd); 1596 } 1597 1598 /* 1599 * USED TO PULL RUNNABLE LWPS FROM THE MOST LOADED CPU. 1600 * 1601 * Choose the worst queue close to dd's cpu node with a non-empty runq 1602 * that is NOT dd. Also require that the moving of the highest-load thread 1603 * from rdd to dd does not cause the uload's to cross each other. 1604 * 1605 * This is used by the thread chooser when the current cpu's queues are 1606 * empty to steal a thread from another cpu's queue. We want to offload 1607 * the most heavily-loaded queue. 1608 */ 1609 static 1610 dfly_pcpu_t 1611 dfly_choose_worst_queue(dfly_pcpu_t dd) 1612 { 1613 cpumask_t mask; 1614 cpu_node_t *cpup; 1615 cpu_node_t *cpun; 1616 cpu_node_t *cpub; 1617 dfly_pcpu_t rdd; 1618 int cpuid; 1619 int n; 1620 int count; 1621 int load; 1622 #if 0 1623 int pri; 1624 int hpri; 1625 #endif 1626 int highest_load; 1627 1628 /* 1629 * When the topology is unknown choose a random cpu that is hopefully 1630 * idle. 1631 */ 1632 if (dd->cpunode == NULL) { 1633 return (NULL); 1634 } 1635 1636 /* 1637 * When the topology is known choose a cpu whos group has, in 1638 * aggregate, has the lowest weighted load. 1639 */ 1640 cpup = root_cpu_node; 1641 rdd = dd; 1642 while (cpup) { 1643 /* 1644 * Degenerate case super-root 1645 */ 1646 if (cpup->child_no == 1) { 1647 cpup = cpup->child_node[0]; 1648 continue; 1649 } 1650 1651 /* 1652 * Terminal cpunode 1653 */ 1654 if (cpup->child_no == 0) { 1655 rdd = &dfly_pcpu[BSFCPUMASK(cpup->members)]; 1656 break; 1657 } 1658 1659 cpub = NULL; 1660 highest_load = 0; 1661 1662 for (n = 0; n < cpup->child_no; ++n) { 1663 /* 1664 * Accumulate load information for all cpus 1665 * which are members of this node. 1666 */ 1667 cpun = cpup->child_node[n]; 1668 mask = cpun->members; 1669 CPUMASK_ANDMASK(mask, usched_global_cpumask); 1670 CPUMASK_ANDMASK(mask, smp_active_mask); 1671 if (CPUMASK_TESTZERO(mask)) 1672 continue; 1673 count = 0; 1674 load = 0; 1675 1676 while (CPUMASK_TESTNZERO(mask)) { 1677 cpuid = BSFCPUMASK(mask); 1678 rdd = &dfly_pcpu[cpuid]; 1679 load += rdd->uload; 1680 load += rdd->ucount * usched_dfly_weight3; 1681 if (rdd->uschedcp == NULL && 1682 rdd->runqcount == 0 && 1683 globaldata_find(cpuid)->gd_tdrunqcount == 0 1684 ) { 1685 load -= usched_dfly_weight4; 1686 } 1687 #if 0 1688 else if (rdd->upri > dd->upri + PPQ) { 1689 load -= usched_dfly_weight4 / 2; 1690 } 1691 #endif 1692 CPUMASK_NANDBIT(mask, cpuid); 1693 ++count; 1694 } 1695 load /= count; 1696 1697 /* 1698 * Prefer candidates which are somewhat closer to 1699 * our cpu. 1700 */ 1701 if (CPUMASK_TESTMASK(dd->cpumask, cpun->members)) 1702 load += usched_dfly_weight1; 1703 1704 /* 1705 * The best candidate is the one with the worst 1706 * (highest) load. 1707 */ 1708 if (cpub == NULL || highest_load < load) { 1709 highest_load = load; 1710 cpub = cpun; 1711 } 1712 } 1713 cpup = cpub; 1714 } 1715 1716 /* 1717 * We never return our own node (dd), and only return a remote 1718 * node if it's load is significantly worse than ours (i.e. where 1719 * stealing a thread would be considered reasonable). 1720 * 1721 * This also helps us avoid breaking paired threads apart which 1722 * can have disastrous effects on performance. 1723 */ 1724 if (rdd == dd) 1725 return(NULL); 1726 1727 #if 0 1728 hpri = 0; 1729 if (rdd->rtqueuebits && hpri < (pri = bsrl(rdd->rtqueuebits))) 1730 hpri = pri; 1731 if (rdd->queuebits && hpri < (pri = bsrl(rdd->queuebits))) 1732 hpri = pri; 1733 if (rdd->idqueuebits && hpri < (pri = bsrl(rdd->idqueuebits))) 1734 hpri = pri; 1735 hpri *= PPQ; 1736 if (rdd->uload - hpri < dd->uload + hpri) 1737 return(NULL); 1738 #endif 1739 return (rdd); 1740 } 1741 1742 static 1743 dfly_pcpu_t 1744 dfly_choose_queue_simple(dfly_pcpu_t dd, struct lwp *lp) 1745 { 1746 dfly_pcpu_t rdd; 1747 cpumask_t tmpmask; 1748 cpumask_t mask; 1749 int cpuid; 1750 1751 /* 1752 * Fallback to the original heuristic, select random cpu, 1753 * first checking cpus not currently running a user thread. 1754 */ 1755 ++dfly_scancpu; 1756 cpuid = (dfly_scancpu & 0xFFFF) % ncpus; 1757 mask = dfly_rdyprocmask; 1758 CPUMASK_NANDMASK(mask, dfly_curprocmask); 1759 CPUMASK_ANDMASK(mask, lp->lwp_cpumask); 1760 CPUMASK_ANDMASK(mask, smp_active_mask); 1761 CPUMASK_ANDMASK(mask, usched_global_cpumask); 1762 1763 while (CPUMASK_TESTNZERO(mask)) { 1764 CPUMASK_ASSNBMASK(tmpmask, cpuid); 1765 if (CPUMASK_TESTMASK(tmpmask, mask)) { 1766 CPUMASK_ANDMASK(tmpmask, mask); 1767 cpuid = BSFCPUMASK(tmpmask); 1768 } else { 1769 cpuid = BSFCPUMASK(mask); 1770 } 1771 rdd = &dfly_pcpu[cpuid]; 1772 1773 if ((rdd->upri & ~PPQMASK) >= (lp->lwp_priority & ~PPQMASK)) 1774 goto found; 1775 CPUMASK_NANDBIT(mask, cpuid); 1776 } 1777 1778 /* 1779 * Then cpus which might have a currently running lp 1780 */ 1781 cpuid = (dfly_scancpu & 0xFFFF) % ncpus; 1782 mask = dfly_rdyprocmask; 1783 CPUMASK_ANDMASK(mask, dfly_curprocmask); 1784 CPUMASK_ANDMASK(mask, lp->lwp_cpumask); 1785 CPUMASK_ANDMASK(mask, smp_active_mask); 1786 CPUMASK_ANDMASK(mask, usched_global_cpumask); 1787 1788 while (CPUMASK_TESTNZERO(mask)) { 1789 CPUMASK_ASSNBMASK(tmpmask, cpuid); 1790 if (CPUMASK_TESTMASK(tmpmask, mask)) { 1791 CPUMASK_ANDMASK(tmpmask, mask); 1792 cpuid = BSFCPUMASK(tmpmask); 1793 } else { 1794 cpuid = BSFCPUMASK(mask); 1795 } 1796 rdd = &dfly_pcpu[cpuid]; 1797 1798 if ((rdd->upri & ~PPQMASK) > (lp->lwp_priority & ~PPQMASK)) 1799 goto found; 1800 CPUMASK_NANDBIT(mask, cpuid); 1801 } 1802 1803 /* 1804 * If we cannot find a suitable cpu we reload from dfly_scancpu 1805 * and round-robin. Other cpus will pickup as they release their 1806 * current lwps or become ready. 1807 * 1808 * Avoid a degenerate system lockup case if usched_global_cpumask 1809 * is set to 0 or otherwise does not cover lwp_cpumask. 1810 * 1811 * We only kick the target helper thread in this case, we do not 1812 * set the user resched flag because 1813 */ 1814 cpuid = (dfly_scancpu & 0xFFFF) % ncpus; 1815 if (CPUMASK_TESTBIT(usched_global_cpumask, cpuid) == 0) 1816 cpuid = 0; 1817 rdd = &dfly_pcpu[cpuid]; 1818 found: 1819 return (rdd); 1820 } 1821 1822 static 1823 void 1824 dfly_need_user_resched_remote(void *dummy) 1825 { 1826 globaldata_t gd = mycpu; 1827 dfly_pcpu_t dd = &dfly_pcpu[gd->gd_cpuid]; 1828 1829 /* 1830 * Flag reschedule needed 1831 */ 1832 need_user_resched(); 1833 1834 /* 1835 * If no user thread is currently running we need to kick the helper 1836 * on our cpu to recover. Otherwise the cpu will never schedule 1837 * anything again. 1838 * 1839 * We cannot schedule the process ourselves because this is an 1840 * IPI callback and we cannot acquire spinlocks in an IPI callback. 1841 * 1842 * Call wakeup_mycpu to avoid sending IPIs to other CPUs 1843 */ 1844 if (dd->uschedcp == NULL && 1845 CPUMASK_TESTBIT(dfly_rdyprocmask, gd->gd_cpuid)) { 1846 ATOMIC_CPUMASK_NANDBIT(dfly_rdyprocmask, gd->gd_cpuid); 1847 wakeup_mycpu(&dd->helper_thread); 1848 } 1849 } 1850 1851 /* 1852 * dfly_remrunqueue_locked() removes a given process from the run queue 1853 * that it is on, clearing the queue busy bit if it becomes empty. 1854 * 1855 * Note that user process scheduler is different from the LWKT schedule. 1856 * The user process scheduler only manages user processes but it uses LWKT 1857 * underneath, and a user process operating in the kernel will often be 1858 * 'released' from our management. 1859 * 1860 * uload is NOT adjusted here. It is only adjusted if the lwkt_thread goes 1861 * to sleep or the lwp is moved to a different runq. 1862 */ 1863 static void 1864 dfly_remrunqueue_locked(dfly_pcpu_t rdd, struct lwp *lp) 1865 { 1866 struct rq *q; 1867 u_int32_t *which; 1868 u_int8_t pri; 1869 1870 KKASSERT(rdd->runqcount >= 0); 1871 1872 pri = lp->lwp_rqindex; 1873 1874 switch(lp->lwp_rqtype) { 1875 case RTP_PRIO_NORMAL: 1876 q = &rdd->queues[pri]; 1877 which = &rdd->queuebits; 1878 break; 1879 case RTP_PRIO_REALTIME: 1880 case RTP_PRIO_FIFO: 1881 q = &rdd->rtqueues[pri]; 1882 which = &rdd->rtqueuebits; 1883 break; 1884 case RTP_PRIO_IDLE: 1885 q = &rdd->idqueues[pri]; 1886 which = &rdd->idqueuebits; 1887 break; 1888 default: 1889 panic("remrunqueue: invalid rtprio type"); 1890 /* NOT REACHED */ 1891 } 1892 KKASSERT(lp->lwp_mpflags & LWP_MP_ONRUNQ); 1893 atomic_clear_int(&lp->lwp_mpflags, LWP_MP_ONRUNQ); 1894 TAILQ_REMOVE(q, lp, lwp_procq); 1895 --rdd->runqcount; 1896 if (TAILQ_EMPTY(q)) { 1897 KASSERT((*which & (1 << pri)) != 0, 1898 ("remrunqueue: remove from empty queue")); 1899 *which &= ~(1 << pri); 1900 } 1901 } 1902 1903 /* 1904 * dfly_setrunqueue_locked() 1905 * 1906 * Add a process whos rqtype and rqindex had previously been calculated 1907 * onto the appropriate run queue. Determine if the addition requires 1908 * a reschedule on a cpu and return the cpuid or -1. 1909 * 1910 * NOTE: Lower priorities are better priorities. 1911 * 1912 * NOTE ON ULOAD: This variable specifies the aggregate load on a cpu, the 1913 * sum of the rough lwp_priority for all running and runnable 1914 * processes. Lower priority processes (higher lwp_priority 1915 * values) actually DO count as more load, not less, because 1916 * these are the programs which require the most care with 1917 * regards to cpu selection. 1918 */ 1919 static void 1920 dfly_setrunqueue_locked(dfly_pcpu_t rdd, struct lwp *lp) 1921 { 1922 u_int32_t *which; 1923 struct rq *q; 1924 int pri; 1925 1926 KKASSERT(lp->lwp_qcpu == rdd->cpuid); 1927 1928 if ((lp->lwp_mpflags & LWP_MP_ULOAD) == 0) { 1929 atomic_set_int(&lp->lwp_mpflags, LWP_MP_ULOAD); 1930 atomic_add_int(&dfly_pcpu[lp->lwp_qcpu].uload, lp->lwp_uload); 1931 atomic_add_int(&dfly_pcpu[lp->lwp_qcpu].ucount, 1); 1932 atomic_add_int(&dfly_ucount, 1); 1933 } 1934 1935 pri = lp->lwp_rqindex; 1936 1937 switch(lp->lwp_rqtype) { 1938 case RTP_PRIO_NORMAL: 1939 q = &rdd->queues[pri]; 1940 which = &rdd->queuebits; 1941 break; 1942 case RTP_PRIO_REALTIME: 1943 case RTP_PRIO_FIFO: 1944 q = &rdd->rtqueues[pri]; 1945 which = &rdd->rtqueuebits; 1946 break; 1947 case RTP_PRIO_IDLE: 1948 q = &rdd->idqueues[pri]; 1949 which = &rdd->idqueuebits; 1950 break; 1951 default: 1952 panic("remrunqueue: invalid rtprio type"); 1953 /* NOT REACHED */ 1954 } 1955 1956 /* 1957 * Place us on the selected queue. Determine if we should be 1958 * placed at the head of the queue or at the end. 1959 * 1960 * We are placed at the tail if our round-robin count has expired, 1961 * or is about to expire and the system thinks its a good place to 1962 * round-robin, or there is already a next thread on the queue 1963 * (it might be trying to pick up where it left off and we don't 1964 * want to interfere). 1965 */ 1966 KKASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) == 0); 1967 atomic_set_int(&lp->lwp_mpflags, LWP_MP_ONRUNQ); 1968 ++rdd->runqcount; 1969 1970 if (lp->lwp_rrcount >= usched_dfly_rrinterval || 1971 (lp->lwp_rrcount >= usched_dfly_rrinterval / 2 && 1972 (lp->lwp_thread->td_mpflags & TDF_MP_BATCH_DEMARC)) 1973 ) { 1974 /* 1975 * Place on tail 1976 */ 1977 atomic_clear_int(&lp->lwp_thread->td_mpflags, 1978 TDF_MP_BATCH_DEMARC); 1979 lp->lwp_rrcount = 0; 1980 TAILQ_INSERT_TAIL(q, lp, lwp_procq); 1981 } else { 1982 /* 1983 * Retain rrcount and place on head. Count is retained 1984 * even if the queue is empty. 1985 */ 1986 TAILQ_INSERT_HEAD(q, lp, lwp_procq); 1987 } 1988 *which |= 1 << pri; 1989 } 1990 1991 /* 1992 * For SMP systems a user scheduler helper thread is created for each 1993 * cpu and is used to allow one cpu to wakeup another for the purposes of 1994 * scheduling userland threads from setrunqueue(). 1995 * 1996 * UP systems do not need the helper since there is only one cpu. 1997 * 1998 * We can't use the idle thread for this because we might block. 1999 * Additionally, doing things this way allows us to HLT idle cpus 2000 * on MP systems. 2001 */ 2002 static void 2003 dfly_helper_thread(void *dummy) 2004 { 2005 globaldata_t gd; 2006 dfly_pcpu_t dd; 2007 dfly_pcpu_t rdd; 2008 struct lwp *nlp; 2009 cpumask_t mask; 2010 int cpuid; 2011 2012 gd = mycpu; 2013 cpuid = gd->gd_cpuid; /* doesn't change */ 2014 mask = gd->gd_cpumask; /* doesn't change */ 2015 dd = &dfly_pcpu[cpuid]; 2016 2017 /* 2018 * Since we only want to be woken up only when no user processes 2019 * are scheduled on a cpu, run at an ultra low priority. 2020 */ 2021 lwkt_setpri_self(TDPRI_USER_SCHEDULER); 2022 2023 tsleep(&dd->helper_thread, 0, "schslp", 0); 2024 2025 for (;;) { 2026 /* 2027 * We use the LWKT deschedule-interlock trick to avoid racing 2028 * dfly_rdyprocmask. This means we cannot block through to the 2029 * manual lwkt_switch() call we make below. 2030 */ 2031 crit_enter_gd(gd); 2032 tsleep_interlock(&dd->helper_thread, 0); 2033 2034 spin_lock(&dd->spin); 2035 2036 ATOMIC_CPUMASK_ORMASK(dfly_rdyprocmask, mask); 2037 clear_user_resched(); /* This satisfied the reschedule request */ 2038 #if 0 2039 dd->rrcount = 0; /* Reset the round-robin counter */ 2040 #endif 2041 2042 if (dd->runqcount || dd->uschedcp != NULL) { 2043 /* 2044 * Threads are available. A thread may or may not be 2045 * currently scheduled. Get the best thread already queued 2046 * to this cpu. 2047 */ 2048 nlp = dfly_chooseproc_locked(dd, dd, dd->uschedcp, 0); 2049 if (nlp) { 2050 ATOMIC_CPUMASK_ORMASK(dfly_curprocmask, mask); 2051 dd->upri = nlp->lwp_priority; 2052 dd->uschedcp = nlp; 2053 #if 0 2054 dd->rrcount = 0; /* reset round robin */ 2055 #endif 2056 spin_unlock(&dd->spin); 2057 lwkt_acquire(nlp->lwp_thread); 2058 lwkt_schedule(nlp->lwp_thread); 2059 } else { 2060 /* 2061 * This situation should not occur because we had 2062 * at least one thread available. 2063 */ 2064 spin_unlock(&dd->spin); 2065 } 2066 } else if (usched_dfly_features & 0x01) { 2067 /* 2068 * This cpu is devoid of runnable threads, steal a thread 2069 * from another cpu. Since we're stealing, might as well 2070 * load balance at the same time. 2071 * 2072 * We choose the highest-loaded thread from the worst queue. 2073 * 2074 * NOTE! This function only returns a non-NULL rdd when 2075 * another cpu's queue is obviously overloaded. We 2076 * do not want to perform the type of rebalancing 2077 * the schedclock does here because it would result 2078 * in insane process pulling when 'steady' state is 2079 * partially unbalanced (e.g. 6 runnables and only 2080 * 4 cores). 2081 */ 2082 rdd = dfly_choose_worst_queue(dd); 2083 if (rdd && spin_trylock(&rdd->spin)) { 2084 nlp = dfly_chooseproc_locked(rdd, dd, NULL, 1); 2085 spin_unlock(&rdd->spin); 2086 } else { 2087 nlp = NULL; 2088 } 2089 if (nlp) { 2090 ATOMIC_CPUMASK_ORMASK(dfly_curprocmask, mask); 2091 dd->upri = nlp->lwp_priority; 2092 dd->uschedcp = nlp; 2093 #if 0 2094 dd->rrcount = 0; /* reset round robin */ 2095 #endif 2096 spin_unlock(&dd->spin); 2097 lwkt_acquire(nlp->lwp_thread); 2098 lwkt_schedule(nlp->lwp_thread); 2099 } else { 2100 /* 2101 * Leave the thread on our run queue. Another 2102 * scheduler will try to pull it later. 2103 */ 2104 spin_unlock(&dd->spin); 2105 } 2106 } else { 2107 /* 2108 * devoid of runnable threads and not allowed to steal 2109 * any. 2110 */ 2111 spin_unlock(&dd->spin); 2112 } 2113 2114 /* 2115 * We're descheduled unless someone scheduled us. Switch away. 2116 * Exiting the critical section will cause splz() to be called 2117 * for us if interrupts and such are pending. 2118 */ 2119 crit_exit_gd(gd); 2120 tsleep(&dd->helper_thread, PINTERLOCKED, "schslp", 0); 2121 } 2122 } 2123 2124 #if 0 2125 static int 2126 sysctl_usched_dfly_stick_to_level(SYSCTL_HANDLER_ARGS) 2127 { 2128 int error, new_val; 2129 2130 new_val = usched_dfly_stick_to_level; 2131 2132 error = sysctl_handle_int(oidp, &new_val, 0, req); 2133 if (error != 0 || req->newptr == NULL) 2134 return (error); 2135 if (new_val > cpu_topology_levels_number - 1 || new_val < 0) 2136 return (EINVAL); 2137 usched_dfly_stick_to_level = new_val; 2138 return (0); 2139 } 2140 #endif 2141 2142 /* 2143 * Setup the queues and scheduler helpers (scheduler helpers are SMP only). 2144 * Note that curprocmask bit 0 has already been cleared by rqinit() and 2145 * we should not mess with it further. 2146 */ 2147 static void 2148 usched_dfly_cpu_init(void) 2149 { 2150 int i; 2151 int j; 2152 int smt_not_supported = 0; 2153 int cache_coherent_not_supported = 0; 2154 2155 if (bootverbose) 2156 kprintf("Start usched_dfly helpers on cpus:\n"); 2157 2158 sysctl_ctx_init(&usched_dfly_sysctl_ctx); 2159 usched_dfly_sysctl_tree = 2160 SYSCTL_ADD_NODE(&usched_dfly_sysctl_ctx, 2161 SYSCTL_STATIC_CHILDREN(_kern), OID_AUTO, 2162 "usched_dfly", CTLFLAG_RD, 0, ""); 2163 2164 for (i = 0; i < ncpus; ++i) { 2165 dfly_pcpu_t dd = &dfly_pcpu[i]; 2166 cpumask_t mask; 2167 2168 CPUMASK_ASSBIT(mask, i); 2169 if (CPUMASK_TESTMASK(mask, smp_active_mask) == 0) 2170 continue; 2171 2172 spin_init(&dd->spin, "uschedcpuinit"); 2173 dd->cpunode = get_cpu_node_by_cpuid(i); 2174 dd->cpuid = i; 2175 CPUMASK_ASSBIT(dd->cpumask, i); 2176 for (j = 0; j < NQS; j++) { 2177 TAILQ_INIT(&dd->queues[j]); 2178 TAILQ_INIT(&dd->rtqueues[j]); 2179 TAILQ_INIT(&dd->idqueues[j]); 2180 } 2181 ATOMIC_CPUMASK_NANDBIT(dfly_curprocmask, 0); 2182 2183 if (dd->cpunode == NULL) { 2184 smt_not_supported = 1; 2185 cache_coherent_not_supported = 1; 2186 if (bootverbose) 2187 kprintf (" cpu%d - WARNING: No CPU NODE " 2188 "found for cpu\n", i); 2189 } else { 2190 switch (dd->cpunode->type) { 2191 case THREAD_LEVEL: 2192 if (bootverbose) 2193 kprintf (" cpu%d - HyperThreading " 2194 "available. Core siblings: ", 2195 i); 2196 break; 2197 case CORE_LEVEL: 2198 smt_not_supported = 1; 2199 2200 if (bootverbose) 2201 kprintf (" cpu%d - No HT available, " 2202 "multi-core/physical " 2203 "cpu. Physical siblings: ", 2204 i); 2205 break; 2206 case CHIP_LEVEL: 2207 smt_not_supported = 1; 2208 2209 if (bootverbose) 2210 kprintf (" cpu%d - No HT available, " 2211 "single-core/physical cpu. " 2212 "Package siblings: ", 2213 i); 2214 break; 2215 default: 2216 /* Let's go for safe defaults here */ 2217 smt_not_supported = 1; 2218 cache_coherent_not_supported = 1; 2219 if (bootverbose) 2220 kprintf (" cpu%d - Unknown cpunode->" 2221 "type=%u. siblings: ", 2222 i, 2223 (u_int)dd->cpunode->type); 2224 break; 2225 } 2226 2227 if (bootverbose) { 2228 if (dd->cpunode->parent_node != NULL) { 2229 kprint_cpuset(&dd->cpunode-> 2230 parent_node->members); 2231 kprintf("\n"); 2232 } else { 2233 kprintf(" no siblings\n"); 2234 } 2235 } 2236 } 2237 2238 lwkt_create(dfly_helper_thread, NULL, NULL, &dd->helper_thread, 2239 0, i, "usched %d", i); 2240 2241 /* 2242 * Allow user scheduling on the target cpu. cpu #0 has already 2243 * been enabled in rqinit(). 2244 */ 2245 if (i) 2246 ATOMIC_CPUMASK_NANDMASK(dfly_curprocmask, mask); 2247 ATOMIC_CPUMASK_ORMASK(dfly_rdyprocmask, mask); 2248 dd->upri = PRIBASE_NULL; 2249 2250 } 2251 2252 /* usched_dfly sysctl configurable parameters */ 2253 2254 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx, 2255 SYSCTL_CHILDREN(usched_dfly_sysctl_tree), 2256 OID_AUTO, "rrinterval", CTLFLAG_RW, 2257 &usched_dfly_rrinterval, 0, ""); 2258 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx, 2259 SYSCTL_CHILDREN(usched_dfly_sysctl_tree), 2260 OID_AUTO, "decay", CTLFLAG_RW, 2261 &usched_dfly_decay, 0, "Extra decay when not running"); 2262 2263 /* Add enable/disable option for SMT scheduling if supported */ 2264 if (smt_not_supported) { 2265 usched_dfly_smt = 0; 2266 SYSCTL_ADD_STRING(&usched_dfly_sysctl_ctx, 2267 SYSCTL_CHILDREN(usched_dfly_sysctl_tree), 2268 OID_AUTO, "smt", CTLFLAG_RD, 2269 "NOT SUPPORTED", 0, "SMT NOT SUPPORTED"); 2270 } else { 2271 usched_dfly_smt = 1; 2272 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx, 2273 SYSCTL_CHILDREN(usched_dfly_sysctl_tree), 2274 OID_AUTO, "smt", CTLFLAG_RW, 2275 &usched_dfly_smt, 0, "Enable SMT scheduling"); 2276 } 2277 2278 /* 2279 * Add enable/disable option for cache coherent scheduling 2280 * if supported 2281 */ 2282 if (cache_coherent_not_supported) { 2283 usched_dfly_cache_coherent = 0; 2284 SYSCTL_ADD_STRING(&usched_dfly_sysctl_ctx, 2285 SYSCTL_CHILDREN(usched_dfly_sysctl_tree), 2286 OID_AUTO, "cache_coherent", CTLFLAG_RD, 2287 "NOT SUPPORTED", 0, 2288 "Cache coherence NOT SUPPORTED"); 2289 } else { 2290 usched_dfly_cache_coherent = 1; 2291 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx, 2292 SYSCTL_CHILDREN(usched_dfly_sysctl_tree), 2293 OID_AUTO, "cache_coherent", CTLFLAG_RW, 2294 &usched_dfly_cache_coherent, 0, 2295 "Enable/Disable cache coherent scheduling"); 2296 2297 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx, 2298 SYSCTL_CHILDREN(usched_dfly_sysctl_tree), 2299 OID_AUTO, "weight1", CTLFLAG_RW, 2300 &usched_dfly_weight1, 200, 2301 "Weight selection for current cpu"); 2302 2303 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx, 2304 SYSCTL_CHILDREN(usched_dfly_sysctl_tree), 2305 OID_AUTO, "weight2", CTLFLAG_RW, 2306 &usched_dfly_weight2, 180, 2307 "Weight selection for wakefrom cpu"); 2308 2309 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx, 2310 SYSCTL_CHILDREN(usched_dfly_sysctl_tree), 2311 OID_AUTO, "weight3", CTLFLAG_RW, 2312 &usched_dfly_weight3, 40, 2313 "Weight selection for num threads on queue"); 2314 2315 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx, 2316 SYSCTL_CHILDREN(usched_dfly_sysctl_tree), 2317 OID_AUTO, "weight4", CTLFLAG_RW, 2318 &usched_dfly_weight4, 160, 2319 "Availability of other idle cpus"); 2320 2321 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx, 2322 SYSCTL_CHILDREN(usched_dfly_sysctl_tree), 2323 OID_AUTO, "fast_resched", CTLFLAG_RW, 2324 &usched_dfly_fast_resched, 0, 2325 "Availability of other idle cpus"); 2326 2327 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx, 2328 SYSCTL_CHILDREN(usched_dfly_sysctl_tree), 2329 OID_AUTO, "features", CTLFLAG_RW, 2330 &usched_dfly_features, 0x8F, 2331 "Allow pulls into empty queues"); 2332 2333 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx, 2334 SYSCTL_CHILDREN(usched_dfly_sysctl_tree), 2335 OID_AUTO, "swmask", CTLFLAG_RW, 2336 &usched_dfly_swmask, ~PPQMASK, 2337 "Queue mask to force thread switch"); 2338 2339 #if 0 2340 SYSCTL_ADD_PROC(&usched_dfly_sysctl_ctx, 2341 SYSCTL_CHILDREN(usched_dfly_sysctl_tree), 2342 OID_AUTO, "stick_to_level", 2343 CTLTYPE_INT | CTLFLAG_RW, 2344 NULL, sizeof usched_dfly_stick_to_level, 2345 sysctl_usched_dfly_stick_to_level, "I", 2346 "Stick a process to this level. See sysctl" 2347 "paremter hw.cpu_topology.level_description"); 2348 #endif 2349 } 2350 } 2351 SYSINIT(uschedtd, SI_BOOT2_USCHED, SI_ORDER_SECOND, 2352 usched_dfly_cpu_init, NULL); 2353