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 987 * of 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 if ((lp->lwp_priority ^ newpriority) & ~PPQMASK) { 1079 if (lp->lwp_mpflags & LWP_MP_ONRUNQ) { 1080 dfly_remrunqueue_locked(rdd, lp); 1081 lp->lwp_priority = newpriority; 1082 lp->lwp_rqtype = newrqtype; 1083 lp->lwp_rqindex = (newpriority & PRIMASK) / PPQ; 1084 dfly_setrunqueue_locked(rdd, lp); 1085 checkpri = 1; 1086 } else { 1087 lp->lwp_priority = newpriority; 1088 lp->lwp_rqtype = newrqtype; 1089 lp->lwp_rqindex = (newpriority & PRIMASK) / PPQ; 1090 checkpri = 0; 1091 } 1092 } else { 1093 /* 1094 * In the same PPQ, uload cannot change. 1095 */ 1096 lp->lwp_priority = newpriority; 1097 checkpri = 1; 1098 rcpu = -1; 1099 } 1100 1101 /* 1102 * Adjust effective load. 1103 * 1104 * Calculate load then scale up or down geometrically based on p_nice. 1105 * Processes niced up (positive) are less important, and processes 1106 * niced downard (negative) are more important. The higher the uload, 1107 * the more important the thread. 1108 */ 1109 /* 0-511, 0-100% cpu */ 1110 delta_uload = lp->lwp_estcpu / NQS; 1111 delta_uload -= delta_uload * lp->lwp_proc->p_nice / (PRIO_MAX + 1); 1112 1113 1114 delta_uload -= lp->lwp_uload; 1115 lp->lwp_uload += delta_uload; 1116 if (lp->lwp_mpflags & LWP_MP_ULOAD) 1117 atomic_add_int(&dfly_pcpu[lp->lwp_qcpu].uload, delta_uload); 1118 1119 /* 1120 * Determine if we need to reschedule the target cpu. This only 1121 * occurs if the LWP is already on a scheduler queue, which means 1122 * that idle cpu notification has already occured. At most we 1123 * need only issue a need_user_resched() on the appropriate cpu. 1124 * 1125 * The LWP may be owned by a CPU different from the current one, 1126 * in which case dd->uschedcp may be modified without an MP lock 1127 * or a spinlock held. The worst that happens is that the code 1128 * below causes a spurious need_user_resched() on the target CPU 1129 * and dd->pri to be wrong for a short period of time, both of 1130 * which are harmless. 1131 * 1132 * If checkpri is 0 we are adjusting the priority of the current 1133 * process, possibly higher (less desireable), so ignore the upri 1134 * check which will fail in that case. 1135 */ 1136 if (rcpu >= 0) { 1137 if (CPUMASK_TESTBIT(dfly_rdyprocmask, rcpu) && 1138 (checkpri == 0 || 1139 (rdd->upri & ~PRIMASK) > 1140 (lp->lwp_priority & ~PRIMASK))) { 1141 if (rcpu == mycpu->gd_cpuid) { 1142 spin_unlock(&rdd->spin); 1143 need_user_resched(); 1144 } else { 1145 spin_unlock(&rdd->spin); 1146 lwkt_send_ipiq(globaldata_find(rcpu), 1147 dfly_need_user_resched_remote, 1148 NULL); 1149 } 1150 } else { 1151 spin_unlock(&rdd->spin); 1152 } 1153 } else { 1154 spin_unlock(&rdd->spin); 1155 } 1156 crit_exit(); 1157 } 1158 1159 static 1160 void 1161 dfly_yield(struct lwp *lp) 1162 { 1163 if (lp->lwp_qcpu != mycpu->gd_cpuid) 1164 return; 1165 KKASSERT(lp == curthread->td_lwp); 1166 1167 /* 1168 * Don't set need_user_resched() or mess with rrcount or anything. 1169 * the TDF flag will override everything as long as we release. 1170 */ 1171 atomic_set_int(&lp->lwp_thread->td_mpflags, TDF_MP_DIDYIELD); 1172 dfly_release_curproc(lp); 1173 } 1174 1175 /* 1176 * Thread was forcefully migrated to another cpu. Normally forced migrations 1177 * are used for iterations and the kernel returns to the original cpu before 1178 * returning and this is not needed. However, if the kernel migrates a 1179 * thread to another cpu and wants to leave it there, it has to call this 1180 * scheduler helper. 1181 * 1182 * Note that the lwkt_migratecpu() function also released the thread, so 1183 * we don't have to worry about that. 1184 */ 1185 static 1186 void 1187 dfly_changedcpu(struct lwp *lp) 1188 { 1189 dfly_pcpu_t dd = &dfly_pcpu[lp->lwp_qcpu]; 1190 dfly_pcpu_t rdd = &dfly_pcpu[mycpu->gd_cpuid]; 1191 1192 if (dd != rdd) { 1193 spin_lock(&dd->spin); 1194 dfly_changeqcpu_locked(lp, dd, rdd); 1195 spin_unlock(&dd->spin); 1196 } 1197 } 1198 1199 /* 1200 * Called from fork1() when a new child process is being created. 1201 * 1202 * Give the child process an initial estcpu that is more batch then 1203 * its parent and dock the parent for the fork (but do not 1204 * reschedule the parent). 1205 * 1206 * fast 1207 * 1208 * XXX lwp should be "spawning" instead of "forking" 1209 */ 1210 static void 1211 dfly_forking(struct lwp *plp, struct lwp *lp) 1212 { 1213 /* 1214 * Put the child 4 queue slots (out of 32) higher than the parent 1215 * (less desireable than the parent). 1216 */ 1217 lp->lwp_estcpu = ESTCPULIM(plp->lwp_estcpu + ESTCPUPPQ * 4); 1218 lp->lwp_forked = 1; 1219 lp->lwp_estfast = 0; 1220 1221 /* 1222 * Dock the parent a cost for the fork, protecting us from fork 1223 * bombs. If the parent is forking quickly make the child more 1224 * batchy. 1225 */ 1226 plp->lwp_estcpu = ESTCPULIM(plp->lwp_estcpu + ESTCPUPPQ / 16); 1227 } 1228 1229 /* 1230 * Called when a lwp is being removed from this scheduler, typically 1231 * during lwp_exit(). We have to clean out any ULOAD accounting before 1232 * we can let the lp go. The dd->spin lock is not needed for uload 1233 * updates. 1234 * 1235 * Scheduler dequeueing has already occurred, no further action in that 1236 * regard is needed. 1237 */ 1238 static void 1239 dfly_exiting(struct lwp *lp, struct proc *child_proc) 1240 { 1241 dfly_pcpu_t dd = &dfly_pcpu[lp->lwp_qcpu]; 1242 1243 if (lp->lwp_mpflags & LWP_MP_ULOAD) { 1244 atomic_clear_int(&lp->lwp_mpflags, LWP_MP_ULOAD); 1245 atomic_add_int(&dd->uload, -lp->lwp_uload); 1246 atomic_add_int(&dd->ucount, -1); 1247 atomic_add_int(&dfly_ucount, -1); 1248 } 1249 } 1250 1251 /* 1252 * This function cannot block in any way, but spinlocks are ok. 1253 * 1254 * Update the uload based on the state of the thread (whether it is going 1255 * to sleep or running again). The uload is meant to be a longer-term 1256 * load and not an instantanious load. 1257 */ 1258 static void 1259 dfly_uload_update(struct lwp *lp) 1260 { 1261 dfly_pcpu_t dd = &dfly_pcpu[lp->lwp_qcpu]; 1262 1263 if (lp->lwp_thread->td_flags & TDF_RUNQ) { 1264 if ((lp->lwp_mpflags & LWP_MP_ULOAD) == 0) { 1265 spin_lock(&dd->spin); 1266 if ((lp->lwp_mpflags & LWP_MP_ULOAD) == 0) { 1267 atomic_set_int(&lp->lwp_mpflags, 1268 LWP_MP_ULOAD); 1269 atomic_add_int(&dd->uload, lp->lwp_uload); 1270 atomic_add_int(&dd->ucount, 1); 1271 atomic_add_int(&dfly_ucount, 1); 1272 } 1273 spin_unlock(&dd->spin); 1274 } 1275 } else if (lp->lwp_slptime > 0) { 1276 if (lp->lwp_mpflags & LWP_MP_ULOAD) { 1277 spin_lock(&dd->spin); 1278 if (lp->lwp_mpflags & LWP_MP_ULOAD) { 1279 atomic_clear_int(&lp->lwp_mpflags, 1280 LWP_MP_ULOAD); 1281 atomic_add_int(&dd->uload, -lp->lwp_uload); 1282 atomic_add_int(&dd->ucount, -1); 1283 atomic_add_int(&dfly_ucount, -1); 1284 } 1285 spin_unlock(&dd->spin); 1286 } 1287 } 1288 } 1289 1290 /* 1291 * chooseproc() is called when a cpu needs a user process to LWKT schedule, 1292 * it selects a user process and returns it. If chklp is non-NULL and chklp 1293 * has a better or equal priority then the process that would otherwise be 1294 * chosen, NULL is returned. 1295 * 1296 * Until we fix the RUNQ code the chklp test has to be strict or we may 1297 * bounce between processes trying to acquire the current process designation. 1298 * 1299 * Must be called with rdd->spin locked. The spinlock is left intact through 1300 * the entire routine. dd->spin does not have to be locked. 1301 * 1302 * If worst is non-zero this function finds the worst thread instead of the 1303 * best thread (used by the schedulerclock-based rover). 1304 */ 1305 static 1306 struct lwp * 1307 dfly_chooseproc_locked(dfly_pcpu_t rdd, dfly_pcpu_t dd, 1308 struct lwp *chklp, int worst) 1309 { 1310 struct lwp *lp; 1311 struct rq *q; 1312 u_int32_t *which; 1313 u_int32_t pri; 1314 u_int32_t rtqbits; 1315 u_int32_t tsqbits; 1316 u_int32_t idqbits; 1317 1318 rtqbits = rdd->rtqueuebits; 1319 tsqbits = rdd->queuebits; 1320 idqbits = rdd->idqueuebits; 1321 1322 if (worst) { 1323 if (idqbits) { 1324 pri = bsrl(idqbits); 1325 q = &rdd->idqueues[pri]; 1326 which = &rdd->idqueuebits; 1327 } else if (tsqbits) { 1328 pri = bsrl(tsqbits); 1329 q = &rdd->queues[pri]; 1330 which = &rdd->queuebits; 1331 } else if (rtqbits) { 1332 pri = bsrl(rtqbits); 1333 q = &rdd->rtqueues[pri]; 1334 which = &rdd->rtqueuebits; 1335 } else { 1336 return (NULL); 1337 } 1338 lp = TAILQ_LAST(q, rq); 1339 } else { 1340 if (rtqbits) { 1341 pri = bsfl(rtqbits); 1342 q = &rdd->rtqueues[pri]; 1343 which = &rdd->rtqueuebits; 1344 } else if (tsqbits) { 1345 pri = bsfl(tsqbits); 1346 q = &rdd->queues[pri]; 1347 which = &rdd->queuebits; 1348 } else if (idqbits) { 1349 pri = bsfl(idqbits); 1350 q = &rdd->idqueues[pri]; 1351 which = &rdd->idqueuebits; 1352 } else { 1353 return (NULL); 1354 } 1355 lp = TAILQ_FIRST(q); 1356 } 1357 KASSERT(lp, ("chooseproc: no lwp on busy queue")); 1358 1359 /* 1360 * If the passed lwp <chklp> is reasonably close to the selected 1361 * lwp <lp>, return NULL (indicating that <chklp> should be kept). 1362 * 1363 * Note that we must error on the side of <chklp> to avoid bouncing 1364 * between threads in the acquire code. 1365 */ 1366 if (chklp) { 1367 if (chklp->lwp_priority < lp->lwp_priority + PPQ) 1368 return(NULL); 1369 } 1370 1371 KTR_COND_LOG(usched_chooseproc, 1372 lp->lwp_proc->p_pid == usched_dfly_pid_debug, 1373 lp->lwp_proc->p_pid, 1374 lp->lwp_thread->td_gd->gd_cpuid, 1375 mycpu->gd_cpuid); 1376 1377 KASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) != 0, ("not on runq6!")); 1378 atomic_clear_int(&lp->lwp_mpflags, LWP_MP_ONRUNQ); 1379 TAILQ_REMOVE(q, lp, lwp_procq); 1380 --rdd->runqcount; 1381 if (TAILQ_EMPTY(q)) 1382 *which &= ~(1 << pri); 1383 1384 /* 1385 * If we are choosing a process from rdd with the intent to 1386 * move it to dd, lwp_qcpu must be adjusted while rdd's spinlock 1387 * is still held. 1388 */ 1389 if (rdd != dd) { 1390 if (lp->lwp_mpflags & LWP_MP_ULOAD) { 1391 atomic_add_int(&rdd->uload, -lp->lwp_uload); 1392 atomic_add_int(&rdd->ucount, -1); 1393 atomic_add_int(&dfly_ucount, -1); 1394 } 1395 lp->lwp_qcpu = dd->cpuid; 1396 atomic_add_int(&dd->uload, lp->lwp_uload); 1397 atomic_add_int(&dd->ucount, 1); 1398 atomic_add_int(&dfly_ucount, 1); 1399 atomic_set_int(&lp->lwp_mpflags, LWP_MP_ULOAD); 1400 } 1401 return lp; 1402 } 1403 1404 /* 1405 * USED TO PUSH RUNNABLE LWPS TO THE LEAST LOADED CPU. 1406 * 1407 * Choose a cpu node to schedule lp on, hopefully nearby its current 1408 * node. 1409 * 1410 * We give the current node a modest advantage for obvious reasons. 1411 * 1412 * We also give the node the thread was woken up FROM a slight advantage 1413 * in order to try to schedule paired threads which synchronize/block waiting 1414 * for each other fairly close to each other. Similarly in a network setting 1415 * this feature will also attempt to place a user process near the kernel 1416 * protocol thread that is feeding it data. THIS IS A CRITICAL PART of the 1417 * algorithm as it heuristically groups synchronizing processes for locality 1418 * of reference in multi-socket systems. 1419 * 1420 * We check against running processes and give a big advantage if there 1421 * are none running. 1422 * 1423 * The caller will normally dfly_setrunqueue() lp on the returned queue. 1424 * 1425 * When the topology is known choose a cpu whos group has, in aggregate, 1426 * has the lowest weighted load. 1427 */ 1428 static 1429 dfly_pcpu_t 1430 dfly_choose_best_queue(struct lwp *lp) 1431 { 1432 cpumask_t wakemask; 1433 cpumask_t mask; 1434 cpu_node_t *cpup; 1435 cpu_node_t *cpun; 1436 cpu_node_t *cpub; 1437 dfly_pcpu_t dd = &dfly_pcpu[lp->lwp_qcpu]; 1438 dfly_pcpu_t rdd; 1439 int wakecpu; 1440 int cpuid; 1441 int n; 1442 int count; 1443 int load; 1444 int lowest_load; 1445 1446 /* 1447 * When the topology is unknown choose a random cpu that is hopefully 1448 * idle. 1449 */ 1450 if (dd->cpunode == NULL) 1451 return (dfly_choose_queue_simple(dd, lp)); 1452 1453 /* 1454 * Pairing mask 1455 */ 1456 if ((wakecpu = lp->lwp_thread->td_wakefromcpu) >= 0) 1457 wakemask = dfly_pcpu[wakecpu].cpumask; 1458 else 1459 CPUMASK_ASSZERO(wakemask); 1460 1461 /* 1462 * When the topology is known choose a cpu whos group has, in 1463 * aggregate, has the lowest weighted load. 1464 */ 1465 cpup = root_cpu_node; 1466 rdd = dd; 1467 1468 while (cpup) { 1469 /* 1470 * Degenerate case super-root 1471 */ 1472 if (cpup->child_no == 1) { 1473 cpup = cpup->child_node[0]; 1474 continue; 1475 } 1476 1477 /* 1478 * Terminal cpunode 1479 */ 1480 if (cpup->child_no == 0) { 1481 rdd = &dfly_pcpu[BSFCPUMASK(cpup->members)]; 1482 break; 1483 } 1484 1485 cpub = NULL; 1486 lowest_load = 0x7FFFFFFF; 1487 1488 for (n = 0; n < cpup->child_no; ++n) { 1489 /* 1490 * Accumulate load information for all cpus 1491 * which are members of this node. 1492 */ 1493 cpun = cpup->child_node[n]; 1494 mask = cpun->members; 1495 CPUMASK_ANDMASK(mask, usched_global_cpumask); 1496 CPUMASK_ANDMASK(mask, smp_active_mask); 1497 CPUMASK_ANDMASK(mask, lp->lwp_cpumask); 1498 if (CPUMASK_TESTZERO(mask)) 1499 continue; 1500 1501 count = 0; 1502 load = 0; 1503 1504 while (CPUMASK_TESTNZERO(mask)) { 1505 cpuid = BSFCPUMASK(mask); 1506 rdd = &dfly_pcpu[cpuid]; 1507 load += rdd->uload; 1508 load += rdd->ucount * usched_dfly_weight3; 1509 1510 if (rdd->uschedcp == NULL && 1511 rdd->runqcount == 0 && 1512 globaldata_find(cpuid)->gd_tdrunqcount == 0 1513 ) { 1514 load -= usched_dfly_weight4; 1515 } 1516 #if 0 1517 else if (rdd->upri > lp->lwp_priority + PPQ) { 1518 load -= usched_dfly_weight4 / 2; 1519 } 1520 #endif 1521 CPUMASK_NANDBIT(mask, cpuid); 1522 ++count; 1523 } 1524 1525 /* 1526 * Compensate if the lp is already accounted for in 1527 * the aggregate uload for this mask set. We want 1528 * to calculate the loads as if lp were not present, 1529 * otherwise the calculation is bogus. 1530 */ 1531 if ((lp->lwp_mpflags & LWP_MP_ULOAD) && 1532 CPUMASK_TESTMASK(dd->cpumask, cpun->members)) { 1533 load -= lp->lwp_uload; 1534 load -= usched_dfly_weight3; 1535 } 1536 1537 load /= count; 1538 1539 /* 1540 * Advantage the cpu group (lp) is already on. 1541 */ 1542 if (CPUMASK_TESTMASK(cpun->members, dd->cpumask)) 1543 load -= usched_dfly_weight1; 1544 1545 /* 1546 * Advantage the cpu group we want to pair (lp) to, 1547 * but don't let it go to the exact same cpu as 1548 * the wakecpu target. 1549 * 1550 * We do this by checking whether cpun is a 1551 * terminal node or not. All cpun's at the same 1552 * level will either all be terminal or all not 1553 * terminal. 1554 * 1555 * If it is and we match we disadvantage the load. 1556 * If it is and we don't match we advantage the load. 1557 * 1558 * Also note that we are effectively disadvantaging 1559 * all-but-one by the same amount, so it won't effect 1560 * the weight1 factor for the all-but-one nodes. 1561 */ 1562 if (CPUMASK_TESTMASK(cpun->members, wakemask)) { 1563 if (cpun->child_no != 0) { 1564 /* advantage */ 1565 load -= usched_dfly_weight2; 1566 } else { 1567 if (usched_dfly_features & 0x10) 1568 load += usched_dfly_weight2; 1569 else 1570 load -= usched_dfly_weight2; 1571 } 1572 } 1573 1574 /* 1575 * Calculate the best load 1576 */ 1577 if (cpub == NULL || lowest_load > load || 1578 (lowest_load == load && 1579 CPUMASK_TESTMASK(cpun->members, dd->cpumask)) 1580 ) { 1581 lowest_load = load; 1582 cpub = cpun; 1583 } 1584 } 1585 cpup = cpub; 1586 } 1587 if (usched_dfly_chooser) 1588 kprintf("lp %02d->%02d %s\n", 1589 lp->lwp_qcpu, rdd->cpuid, lp->lwp_proc->p_comm); 1590 return (rdd); 1591 } 1592 1593 /* 1594 * USED TO PULL RUNNABLE LWPS FROM THE MOST LOADED CPU. 1595 * 1596 * Choose the worst queue close to dd's cpu node with a non-empty runq 1597 * that is NOT dd. Also require that the moving of the highest-load thread 1598 * from rdd to dd does not cause the uload's to cross each other. 1599 * 1600 * This is used by the thread chooser when the current cpu's queues are 1601 * empty to steal a thread from another cpu's queue. We want to offload 1602 * the most heavily-loaded queue. 1603 */ 1604 static 1605 dfly_pcpu_t 1606 dfly_choose_worst_queue(dfly_pcpu_t dd) 1607 { 1608 cpumask_t mask; 1609 cpu_node_t *cpup; 1610 cpu_node_t *cpun; 1611 cpu_node_t *cpub; 1612 dfly_pcpu_t rdd; 1613 int cpuid; 1614 int n; 1615 int count; 1616 int load; 1617 #if 0 1618 int pri; 1619 int hpri; 1620 #endif 1621 int highest_load; 1622 1623 /* 1624 * When the topology is unknown choose a random cpu that is hopefully 1625 * idle. 1626 */ 1627 if (dd->cpunode == NULL) { 1628 return (NULL); 1629 } 1630 1631 /* 1632 * When the topology is known choose a cpu whos group has, in 1633 * aggregate, has the lowest weighted load. 1634 */ 1635 cpup = root_cpu_node; 1636 rdd = dd; 1637 while (cpup) { 1638 /* 1639 * Degenerate case super-root 1640 */ 1641 if (cpup->child_no == 1) { 1642 cpup = cpup->child_node[0]; 1643 continue; 1644 } 1645 1646 /* 1647 * Terminal cpunode 1648 */ 1649 if (cpup->child_no == 0) { 1650 rdd = &dfly_pcpu[BSFCPUMASK(cpup->members)]; 1651 break; 1652 } 1653 1654 cpub = NULL; 1655 highest_load = 0; 1656 1657 for (n = 0; n < cpup->child_no; ++n) { 1658 /* 1659 * Accumulate load information for all cpus 1660 * which are members of this node. 1661 */ 1662 cpun = cpup->child_node[n]; 1663 mask = cpun->members; 1664 CPUMASK_ANDMASK(mask, usched_global_cpumask); 1665 CPUMASK_ANDMASK(mask, smp_active_mask); 1666 if (CPUMASK_TESTZERO(mask)) 1667 continue; 1668 count = 0; 1669 load = 0; 1670 1671 while (CPUMASK_TESTNZERO(mask)) { 1672 cpuid = BSFCPUMASK(mask); 1673 rdd = &dfly_pcpu[cpuid]; 1674 load += rdd->uload; 1675 load += rdd->ucount * usched_dfly_weight3; 1676 if (rdd->uschedcp == NULL && 1677 rdd->runqcount == 0 && 1678 globaldata_find(cpuid)->gd_tdrunqcount == 0 1679 ) { 1680 load -= usched_dfly_weight4; 1681 } 1682 #if 0 1683 else if (rdd->upri > dd->upri + PPQ) { 1684 load -= usched_dfly_weight4 / 2; 1685 } 1686 #endif 1687 CPUMASK_NANDBIT(mask, cpuid); 1688 ++count; 1689 } 1690 load /= count; 1691 1692 /* 1693 * Prefer candidates which are somewhat closer to 1694 * our cpu. 1695 */ 1696 if (CPUMASK_TESTMASK(dd->cpumask, cpun->members)) 1697 load += usched_dfly_weight1; 1698 1699 /* 1700 * The best candidate is the one with the worst 1701 * (highest) load. 1702 */ 1703 if (cpub == NULL || highest_load < load) { 1704 highest_load = load; 1705 cpub = cpun; 1706 } 1707 } 1708 cpup = cpub; 1709 } 1710 1711 /* 1712 * We never return our own node (dd), and only return a remote 1713 * node if it's load is significantly worse than ours (i.e. where 1714 * stealing a thread would be considered reasonable). 1715 * 1716 * This also helps us avoid breaking paired threads apart which 1717 * can have disastrous effects on performance. 1718 */ 1719 if (rdd == dd) 1720 return(NULL); 1721 1722 #if 0 1723 hpri = 0; 1724 if (rdd->rtqueuebits && hpri < (pri = bsrl(rdd->rtqueuebits))) 1725 hpri = pri; 1726 if (rdd->queuebits && hpri < (pri = bsrl(rdd->queuebits))) 1727 hpri = pri; 1728 if (rdd->idqueuebits && hpri < (pri = bsrl(rdd->idqueuebits))) 1729 hpri = pri; 1730 hpri *= PPQ; 1731 if (rdd->uload - hpri < dd->uload + hpri) 1732 return(NULL); 1733 #endif 1734 return (rdd); 1735 } 1736 1737 static 1738 dfly_pcpu_t 1739 dfly_choose_queue_simple(dfly_pcpu_t dd, struct lwp *lp) 1740 { 1741 dfly_pcpu_t rdd; 1742 cpumask_t tmpmask; 1743 cpumask_t mask; 1744 int cpuid; 1745 1746 /* 1747 * Fallback to the original heuristic, select random cpu, 1748 * first checking cpus not currently running a user thread. 1749 */ 1750 ++dfly_scancpu; 1751 cpuid = (dfly_scancpu & 0xFFFF) % ncpus; 1752 mask = dfly_rdyprocmask; 1753 CPUMASK_NANDMASK(mask, dfly_curprocmask); 1754 CPUMASK_ANDMASK(mask, lp->lwp_cpumask); 1755 CPUMASK_ANDMASK(mask, smp_active_mask); 1756 CPUMASK_ANDMASK(mask, usched_global_cpumask); 1757 1758 while (CPUMASK_TESTNZERO(mask)) { 1759 CPUMASK_ASSNBMASK(tmpmask, cpuid); 1760 if (CPUMASK_TESTMASK(tmpmask, mask)) { 1761 CPUMASK_ANDMASK(tmpmask, mask); 1762 cpuid = BSFCPUMASK(tmpmask); 1763 } else { 1764 cpuid = BSFCPUMASK(mask); 1765 } 1766 rdd = &dfly_pcpu[cpuid]; 1767 1768 if ((rdd->upri & ~PPQMASK) >= (lp->lwp_priority & ~PPQMASK)) 1769 goto found; 1770 CPUMASK_NANDBIT(mask, cpuid); 1771 } 1772 1773 /* 1774 * Then cpus which might have a currently running lp 1775 */ 1776 cpuid = (dfly_scancpu & 0xFFFF) % ncpus; 1777 mask = dfly_rdyprocmask; 1778 CPUMASK_ANDMASK(mask, dfly_curprocmask); 1779 CPUMASK_ANDMASK(mask, lp->lwp_cpumask); 1780 CPUMASK_ANDMASK(mask, smp_active_mask); 1781 CPUMASK_ANDMASK(mask, usched_global_cpumask); 1782 1783 while (CPUMASK_TESTNZERO(mask)) { 1784 CPUMASK_ASSNBMASK(tmpmask, cpuid); 1785 if (CPUMASK_TESTMASK(tmpmask, mask)) { 1786 CPUMASK_ANDMASK(tmpmask, mask); 1787 cpuid = BSFCPUMASK(tmpmask); 1788 } else { 1789 cpuid = BSFCPUMASK(mask); 1790 } 1791 rdd = &dfly_pcpu[cpuid]; 1792 1793 if ((rdd->upri & ~PPQMASK) > (lp->lwp_priority & ~PPQMASK)) 1794 goto found; 1795 CPUMASK_NANDBIT(mask, cpuid); 1796 } 1797 1798 /* 1799 * If we cannot find a suitable cpu we reload from dfly_scancpu 1800 * and round-robin. Other cpus will pickup as they release their 1801 * current lwps or become ready. 1802 * 1803 * Avoid a degenerate system lockup case if usched_global_cpumask 1804 * is set to 0 or otherwise does not cover lwp_cpumask. 1805 * 1806 * We only kick the target helper thread in this case, we do not 1807 * set the user resched flag because 1808 */ 1809 cpuid = (dfly_scancpu & 0xFFFF) % ncpus; 1810 if (CPUMASK_TESTBIT(usched_global_cpumask, cpuid) == 0) 1811 cpuid = 0; 1812 rdd = &dfly_pcpu[cpuid]; 1813 found: 1814 return (rdd); 1815 } 1816 1817 static 1818 void 1819 dfly_need_user_resched_remote(void *dummy) 1820 { 1821 globaldata_t gd = mycpu; 1822 dfly_pcpu_t dd = &dfly_pcpu[gd->gd_cpuid]; 1823 1824 /* 1825 * Flag reschedule needed 1826 */ 1827 need_user_resched(); 1828 1829 /* 1830 * If no user thread is currently running we need to kick the helper 1831 * on our cpu to recover. Otherwise the cpu will never schedule 1832 * anything again. 1833 * 1834 * We cannot schedule the process ourselves because this is an 1835 * IPI callback and we cannot acquire spinlocks in an IPI callback. 1836 * 1837 * Call wakeup_mycpu to avoid sending IPIs to other CPUs 1838 */ 1839 if (dd->uschedcp == NULL && 1840 CPUMASK_TESTBIT(dfly_rdyprocmask, gd->gd_cpuid)) { 1841 ATOMIC_CPUMASK_NANDBIT(dfly_rdyprocmask, gd->gd_cpuid); 1842 wakeup_mycpu(&dd->helper_thread); 1843 } 1844 } 1845 1846 /* 1847 * dfly_remrunqueue_locked() removes a given process from the run queue 1848 * that it is on, clearing the queue busy bit if it becomes empty. 1849 * 1850 * Note that user process scheduler is different from the LWKT schedule. 1851 * The user process scheduler only manages user processes but it uses LWKT 1852 * underneath, and a user process operating in the kernel will often be 1853 * 'released' from our management. 1854 * 1855 * uload is NOT adjusted here. It is only adjusted if the lwkt_thread goes 1856 * to sleep or the lwp is moved to a different runq. 1857 */ 1858 static void 1859 dfly_remrunqueue_locked(dfly_pcpu_t rdd, struct lwp *lp) 1860 { 1861 struct rq *q; 1862 u_int32_t *which; 1863 u_int8_t pri; 1864 1865 KKASSERT(rdd->runqcount >= 0); 1866 1867 pri = lp->lwp_rqindex; 1868 1869 switch(lp->lwp_rqtype) { 1870 case RTP_PRIO_NORMAL: 1871 q = &rdd->queues[pri]; 1872 which = &rdd->queuebits; 1873 break; 1874 case RTP_PRIO_REALTIME: 1875 case RTP_PRIO_FIFO: 1876 q = &rdd->rtqueues[pri]; 1877 which = &rdd->rtqueuebits; 1878 break; 1879 case RTP_PRIO_IDLE: 1880 q = &rdd->idqueues[pri]; 1881 which = &rdd->idqueuebits; 1882 break; 1883 default: 1884 panic("remrunqueue: invalid rtprio type"); 1885 /* NOT REACHED */ 1886 } 1887 KKASSERT(lp->lwp_mpflags & LWP_MP_ONRUNQ); 1888 atomic_clear_int(&lp->lwp_mpflags, LWP_MP_ONRUNQ); 1889 TAILQ_REMOVE(q, lp, lwp_procq); 1890 --rdd->runqcount; 1891 if (TAILQ_EMPTY(q)) { 1892 KASSERT((*which & (1 << pri)) != 0, 1893 ("remrunqueue: remove from empty queue")); 1894 *which &= ~(1 << pri); 1895 } 1896 } 1897 1898 /* 1899 * dfly_setrunqueue_locked() 1900 * 1901 * Add a process whos rqtype and rqindex had previously been calculated 1902 * onto the appropriate run queue. Determine if the addition requires 1903 * a reschedule on a cpu and return the cpuid or -1. 1904 * 1905 * NOTE: Lower priorities are better priorities. 1906 * 1907 * NOTE ON ULOAD: This variable specifies the aggregate load on a cpu, the 1908 * sum of the rough lwp_priority for all running and runnable 1909 * processes. Lower priority processes (higher lwp_priority 1910 * values) actually DO count as more load, not less, because 1911 * these are the programs which require the most care with 1912 * regards to cpu selection. 1913 */ 1914 static void 1915 dfly_setrunqueue_locked(dfly_pcpu_t rdd, struct lwp *lp) 1916 { 1917 struct rq *q; 1918 u_int32_t *which; 1919 int pri; 1920 1921 KKASSERT(lp->lwp_qcpu == rdd->cpuid); 1922 1923 if ((lp->lwp_mpflags & LWP_MP_ULOAD) == 0) { 1924 atomic_set_int(&lp->lwp_mpflags, LWP_MP_ULOAD); 1925 atomic_add_int(&dfly_pcpu[lp->lwp_qcpu].uload, lp->lwp_uload); 1926 atomic_add_int(&dfly_pcpu[lp->lwp_qcpu].ucount, 1); 1927 atomic_add_int(&dfly_ucount, 1); 1928 } 1929 1930 pri = lp->lwp_rqindex; 1931 1932 switch(lp->lwp_rqtype) { 1933 case RTP_PRIO_NORMAL: 1934 q = &rdd->queues[pri]; 1935 which = &rdd->queuebits; 1936 break; 1937 case RTP_PRIO_REALTIME: 1938 case RTP_PRIO_FIFO: 1939 q = &rdd->rtqueues[pri]; 1940 which = &rdd->rtqueuebits; 1941 break; 1942 case RTP_PRIO_IDLE: 1943 q = &rdd->idqueues[pri]; 1944 which = &rdd->idqueuebits; 1945 break; 1946 default: 1947 panic("remrunqueue: invalid rtprio type"); 1948 /* NOT REACHED */ 1949 } 1950 1951 /* 1952 * Place us on the selected queue. Determine if we should be 1953 * placed at the head of the queue or at the end. 1954 * 1955 * We are placed at the tail if our round-robin count has expired, 1956 * or is about to expire and the system thinks its a good place to 1957 * round-robin, or there is already a next thread on the queue 1958 * (it might be trying to pick up where it left off and we don't 1959 * want to interfere). 1960 */ 1961 KKASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) == 0); 1962 atomic_set_int(&lp->lwp_mpflags, LWP_MP_ONRUNQ); 1963 ++rdd->runqcount; 1964 1965 if (lp->lwp_rrcount >= usched_dfly_rrinterval || 1966 (lp->lwp_rrcount >= usched_dfly_rrinterval / 2 && 1967 (lp->lwp_thread->td_mpflags & TDF_MP_BATCH_DEMARC)) || 1968 !TAILQ_EMPTY(q) 1969 ) { 1970 atomic_clear_int(&lp->lwp_thread->td_mpflags, 1971 TDF_MP_BATCH_DEMARC); 1972 lp->lwp_rrcount = 0; 1973 TAILQ_INSERT_TAIL(q, lp, lwp_procq); 1974 } else { 1975 if (TAILQ_EMPTY(q)) 1976 lp->lwp_rrcount = 0; 1977 TAILQ_INSERT_HEAD(q, lp, lwp_procq); 1978 } 1979 *which |= 1 << pri; 1980 } 1981 1982 /* 1983 * For SMP systems a user scheduler helper thread is created for each 1984 * cpu and is used to allow one cpu to wakeup another for the purposes of 1985 * scheduling userland threads from setrunqueue(). 1986 * 1987 * UP systems do not need the helper since there is only one cpu. 1988 * 1989 * We can't use the idle thread for this because we might block. 1990 * Additionally, doing things this way allows us to HLT idle cpus 1991 * on MP systems. 1992 */ 1993 static void 1994 dfly_helper_thread(void *dummy) 1995 { 1996 globaldata_t gd; 1997 dfly_pcpu_t dd; 1998 dfly_pcpu_t rdd; 1999 struct lwp *nlp; 2000 cpumask_t mask; 2001 int cpuid; 2002 2003 gd = mycpu; 2004 cpuid = gd->gd_cpuid; /* doesn't change */ 2005 mask = gd->gd_cpumask; /* doesn't change */ 2006 dd = &dfly_pcpu[cpuid]; 2007 2008 /* 2009 * Since we only want to be woken up only when no user processes 2010 * are scheduled on a cpu, run at an ultra low priority. 2011 */ 2012 lwkt_setpri_self(TDPRI_USER_SCHEDULER); 2013 2014 tsleep(&dd->helper_thread, 0, "schslp", 0); 2015 2016 for (;;) { 2017 /* 2018 * We use the LWKT deschedule-interlock trick to avoid racing 2019 * dfly_rdyprocmask. This means we cannot block through to the 2020 * manual lwkt_switch() call we make below. 2021 */ 2022 crit_enter_gd(gd); 2023 tsleep_interlock(&dd->helper_thread, 0); 2024 2025 spin_lock(&dd->spin); 2026 2027 ATOMIC_CPUMASK_ORMASK(dfly_rdyprocmask, mask); 2028 clear_user_resched(); /* This satisfied the reschedule request */ 2029 #if 0 2030 dd->rrcount = 0; /* Reset the round-robin counter */ 2031 #endif 2032 2033 if (dd->runqcount || dd->uschedcp != NULL) { 2034 /* 2035 * Threads are available. A thread may or may not be 2036 * currently scheduled. Get the best thread already queued 2037 * to this cpu. 2038 */ 2039 nlp = dfly_chooseproc_locked(dd, dd, dd->uschedcp, 0); 2040 if (nlp) { 2041 ATOMIC_CPUMASK_ORMASK(dfly_curprocmask, mask); 2042 dd->upri = nlp->lwp_priority; 2043 dd->uschedcp = nlp; 2044 #if 0 2045 dd->rrcount = 0; /* reset round robin */ 2046 #endif 2047 spin_unlock(&dd->spin); 2048 lwkt_acquire(nlp->lwp_thread); 2049 lwkt_schedule(nlp->lwp_thread); 2050 } else { 2051 /* 2052 * This situation should not occur because we had 2053 * at least one thread available. 2054 */ 2055 spin_unlock(&dd->spin); 2056 } 2057 } else if (usched_dfly_features & 0x01) { 2058 /* 2059 * This cpu is devoid of runnable threads, steal a thread 2060 * from another cpu. Since we're stealing, might as well 2061 * load balance at the same time. 2062 * 2063 * We choose the highest-loaded thread from the worst queue. 2064 * 2065 * NOTE! This function only returns a non-NULL rdd when 2066 * another cpu's queue is obviously overloaded. We 2067 * do not want to perform the type of rebalancing 2068 * the schedclock does here because it would result 2069 * in insane process pulling when 'steady' state is 2070 * partially unbalanced (e.g. 6 runnables and only 2071 * 4 cores). 2072 */ 2073 rdd = dfly_choose_worst_queue(dd); 2074 if (rdd && spin_trylock(&rdd->spin)) { 2075 nlp = dfly_chooseproc_locked(rdd, dd, NULL, 1); 2076 spin_unlock(&rdd->spin); 2077 } else { 2078 nlp = NULL; 2079 } 2080 if (nlp) { 2081 ATOMIC_CPUMASK_ORMASK(dfly_curprocmask, mask); 2082 dd->upri = nlp->lwp_priority; 2083 dd->uschedcp = nlp; 2084 #if 0 2085 dd->rrcount = 0; /* reset round robin */ 2086 #endif 2087 spin_unlock(&dd->spin); 2088 lwkt_acquire(nlp->lwp_thread); 2089 lwkt_schedule(nlp->lwp_thread); 2090 } else { 2091 /* 2092 * Leave the thread on our run queue. Another 2093 * scheduler will try to pull it later. 2094 */ 2095 spin_unlock(&dd->spin); 2096 } 2097 } else { 2098 /* 2099 * devoid of runnable threads and not allowed to steal 2100 * any. 2101 */ 2102 spin_unlock(&dd->spin); 2103 } 2104 2105 /* 2106 * We're descheduled unless someone scheduled us. Switch away. 2107 * Exiting the critical section will cause splz() to be called 2108 * for us if interrupts and such are pending. 2109 */ 2110 crit_exit_gd(gd); 2111 tsleep(&dd->helper_thread, PINTERLOCKED, "schslp", 0); 2112 } 2113 } 2114 2115 #if 0 2116 static int 2117 sysctl_usched_dfly_stick_to_level(SYSCTL_HANDLER_ARGS) 2118 { 2119 int error, new_val; 2120 2121 new_val = usched_dfly_stick_to_level; 2122 2123 error = sysctl_handle_int(oidp, &new_val, 0, req); 2124 if (error != 0 || req->newptr == NULL) 2125 return (error); 2126 if (new_val > cpu_topology_levels_number - 1 || new_val < 0) 2127 return (EINVAL); 2128 usched_dfly_stick_to_level = new_val; 2129 return (0); 2130 } 2131 #endif 2132 2133 /* 2134 * Setup the queues and scheduler helpers (scheduler helpers are SMP only). 2135 * Note that curprocmask bit 0 has already been cleared by rqinit() and 2136 * we should not mess with it further. 2137 */ 2138 static void 2139 usched_dfly_cpu_init(void) 2140 { 2141 int i; 2142 int j; 2143 int smt_not_supported = 0; 2144 int cache_coherent_not_supported = 0; 2145 2146 if (bootverbose) 2147 kprintf("Start usched_dfly helpers on cpus:\n"); 2148 2149 sysctl_ctx_init(&usched_dfly_sysctl_ctx); 2150 usched_dfly_sysctl_tree = 2151 SYSCTL_ADD_NODE(&usched_dfly_sysctl_ctx, 2152 SYSCTL_STATIC_CHILDREN(_kern), OID_AUTO, 2153 "usched_dfly", CTLFLAG_RD, 0, ""); 2154 2155 for (i = 0; i < ncpus; ++i) { 2156 dfly_pcpu_t dd = &dfly_pcpu[i]; 2157 cpumask_t mask; 2158 2159 CPUMASK_ASSBIT(mask, i); 2160 if (CPUMASK_TESTMASK(mask, smp_active_mask) == 0) 2161 continue; 2162 2163 spin_init(&dd->spin, "uschedcpuinit"); 2164 dd->cpunode = get_cpu_node_by_cpuid(i); 2165 dd->cpuid = i; 2166 CPUMASK_ASSBIT(dd->cpumask, i); 2167 for (j = 0; j < NQS; j++) { 2168 TAILQ_INIT(&dd->queues[j]); 2169 TAILQ_INIT(&dd->rtqueues[j]); 2170 TAILQ_INIT(&dd->idqueues[j]); 2171 } 2172 ATOMIC_CPUMASK_NANDBIT(dfly_curprocmask, 0); 2173 2174 if (dd->cpunode == NULL) { 2175 smt_not_supported = 1; 2176 cache_coherent_not_supported = 1; 2177 if (bootverbose) 2178 kprintf (" cpu%d - WARNING: No CPU NODE " 2179 "found for cpu\n", i); 2180 } else { 2181 switch (dd->cpunode->type) { 2182 case THREAD_LEVEL: 2183 if (bootverbose) 2184 kprintf (" cpu%d - HyperThreading " 2185 "available. Core siblings: ", 2186 i); 2187 break; 2188 case CORE_LEVEL: 2189 smt_not_supported = 1; 2190 2191 if (bootverbose) 2192 kprintf (" cpu%d - No HT available, " 2193 "multi-core/physical " 2194 "cpu. Physical siblings: ", 2195 i); 2196 break; 2197 case CHIP_LEVEL: 2198 smt_not_supported = 1; 2199 2200 if (bootverbose) 2201 kprintf (" cpu%d - No HT available, " 2202 "single-core/physical cpu. " 2203 "Package siblings: ", 2204 i); 2205 break; 2206 default: 2207 /* Let's go for safe defaults here */ 2208 smt_not_supported = 1; 2209 cache_coherent_not_supported = 1; 2210 if (bootverbose) 2211 kprintf (" cpu%d - Unknown cpunode->" 2212 "type=%u. siblings: ", 2213 i, 2214 (u_int)dd->cpunode->type); 2215 break; 2216 } 2217 2218 if (bootverbose) { 2219 if (dd->cpunode->parent_node != NULL) { 2220 kprint_cpuset(&dd->cpunode-> 2221 parent_node->members); 2222 kprintf("\n"); 2223 } else { 2224 kprintf(" no siblings\n"); 2225 } 2226 } 2227 } 2228 2229 lwkt_create(dfly_helper_thread, NULL, NULL, &dd->helper_thread, 2230 0, i, "usched %d", i); 2231 2232 /* 2233 * Allow user scheduling on the target cpu. cpu #0 has already 2234 * been enabled in rqinit(). 2235 */ 2236 if (i) 2237 ATOMIC_CPUMASK_NANDMASK(dfly_curprocmask, mask); 2238 ATOMIC_CPUMASK_ORMASK(dfly_rdyprocmask, mask); 2239 dd->upri = PRIBASE_NULL; 2240 2241 } 2242 2243 /* usched_dfly sysctl configurable parameters */ 2244 2245 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx, 2246 SYSCTL_CHILDREN(usched_dfly_sysctl_tree), 2247 OID_AUTO, "rrinterval", CTLFLAG_RW, 2248 &usched_dfly_rrinterval, 0, ""); 2249 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx, 2250 SYSCTL_CHILDREN(usched_dfly_sysctl_tree), 2251 OID_AUTO, "decay", CTLFLAG_RW, 2252 &usched_dfly_decay, 0, "Extra decay when not running"); 2253 2254 /* Add enable/disable option for SMT scheduling if supported */ 2255 if (smt_not_supported) { 2256 usched_dfly_smt = 0; 2257 SYSCTL_ADD_STRING(&usched_dfly_sysctl_ctx, 2258 SYSCTL_CHILDREN(usched_dfly_sysctl_tree), 2259 OID_AUTO, "smt", CTLFLAG_RD, 2260 "NOT SUPPORTED", 0, "SMT NOT SUPPORTED"); 2261 } else { 2262 usched_dfly_smt = 1; 2263 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx, 2264 SYSCTL_CHILDREN(usched_dfly_sysctl_tree), 2265 OID_AUTO, "smt", CTLFLAG_RW, 2266 &usched_dfly_smt, 0, "Enable SMT scheduling"); 2267 } 2268 2269 /* 2270 * Add enable/disable option for cache coherent scheduling 2271 * if supported 2272 */ 2273 if (cache_coherent_not_supported) { 2274 usched_dfly_cache_coherent = 0; 2275 SYSCTL_ADD_STRING(&usched_dfly_sysctl_ctx, 2276 SYSCTL_CHILDREN(usched_dfly_sysctl_tree), 2277 OID_AUTO, "cache_coherent", CTLFLAG_RD, 2278 "NOT SUPPORTED", 0, 2279 "Cache coherence NOT SUPPORTED"); 2280 } else { 2281 usched_dfly_cache_coherent = 1; 2282 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx, 2283 SYSCTL_CHILDREN(usched_dfly_sysctl_tree), 2284 OID_AUTO, "cache_coherent", CTLFLAG_RW, 2285 &usched_dfly_cache_coherent, 0, 2286 "Enable/Disable cache coherent scheduling"); 2287 2288 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx, 2289 SYSCTL_CHILDREN(usched_dfly_sysctl_tree), 2290 OID_AUTO, "weight1", CTLFLAG_RW, 2291 &usched_dfly_weight1, 200, 2292 "Weight selection for current cpu"); 2293 2294 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx, 2295 SYSCTL_CHILDREN(usched_dfly_sysctl_tree), 2296 OID_AUTO, "weight2", CTLFLAG_RW, 2297 &usched_dfly_weight2, 180, 2298 "Weight selection for wakefrom cpu"); 2299 2300 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx, 2301 SYSCTL_CHILDREN(usched_dfly_sysctl_tree), 2302 OID_AUTO, "weight3", CTLFLAG_RW, 2303 &usched_dfly_weight3, 40, 2304 "Weight selection for num threads on queue"); 2305 2306 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx, 2307 SYSCTL_CHILDREN(usched_dfly_sysctl_tree), 2308 OID_AUTO, "weight4", CTLFLAG_RW, 2309 &usched_dfly_weight4, 160, 2310 "Availability of other idle cpus"); 2311 2312 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx, 2313 SYSCTL_CHILDREN(usched_dfly_sysctl_tree), 2314 OID_AUTO, "fast_resched", CTLFLAG_RW, 2315 &usched_dfly_fast_resched, 0, 2316 "Availability of other idle cpus"); 2317 2318 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx, 2319 SYSCTL_CHILDREN(usched_dfly_sysctl_tree), 2320 OID_AUTO, "features", CTLFLAG_RW, 2321 &usched_dfly_features, 0x8F, 2322 "Allow pulls into empty queues"); 2323 2324 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx, 2325 SYSCTL_CHILDREN(usched_dfly_sysctl_tree), 2326 OID_AUTO, "swmask", CTLFLAG_RW, 2327 &usched_dfly_swmask, ~PPQMASK, 2328 "Queue mask to force thread switch"); 2329 2330 #if 0 2331 SYSCTL_ADD_PROC(&usched_dfly_sysctl_ctx, 2332 SYSCTL_CHILDREN(usched_dfly_sysctl_tree), 2333 OID_AUTO, "stick_to_level", 2334 CTLTYPE_INT | CTLFLAG_RW, 2335 NULL, sizeof usched_dfly_stick_to_level, 2336 sysctl_usched_dfly_stick_to_level, "I", 2337 "Stick a process to this level. See sysctl" 2338 "paremter hw.cpu_topology.level_description"); 2339 #endif 2340 } 2341 } 2342 SYSINIT(uschedtd, SI_BOOT2_USCHED, SI_ORDER_SECOND, 2343 usched_dfly_cpu_init, NULL); 2344