1 /* 2 * Copyright (c) 1999 Peter Wemm <peter@FreeBSD.org> 3 * All rights reserved. 4 * 5 * Redistribution and use in source and binary forms, with or without 6 * modification, are permitted provided that the following conditions 7 * are met: 8 * 1. Redistributions of source code must retain the above copyright 9 * notice, this list of conditions and the following disclaimer. 10 * 2. Redistributions in binary form must reproduce the above copyright 11 * notice, this list of conditions and the following disclaimer in the 12 * documentation and/or other materials provided with the distribution. 13 * 14 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 15 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 16 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 17 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 18 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 19 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 20 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 21 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 22 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 23 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 24 * SUCH DAMAGE. 25 * 26 * $DragonFly: src/sys/kern/usched_bsd4.c,v 1.16 2006/07/11 01:01:50 dillon Exp $ 27 */ 28 29 #include <sys/param.h> 30 #include <sys/systm.h> 31 #include <sys/kernel.h> 32 #include <sys/lock.h> 33 #include <sys/queue.h> 34 #include <sys/proc.h> 35 #include <sys/rtprio.h> 36 #include <sys/uio.h> 37 #include <sys/sysctl.h> 38 #include <sys/resourcevar.h> 39 #include <sys/spinlock.h> 40 #include <machine/ipl.h> 41 #include <machine/cpu.h> 42 #include <machine/smp.h> 43 44 #include <sys/thread2.h> 45 #include <sys/spinlock2.h> 46 47 /* 48 * Priorities. Note that with 32 run queues per scheduler each queue 49 * represents four priority levels. 50 */ 51 52 #define MAXPRI 128 53 #define PRIMASK (MAXPRI - 1) 54 #define PRIBASE_REALTIME 0 55 #define PRIBASE_NORMAL MAXPRI 56 #define PRIBASE_IDLE (MAXPRI * 2) 57 #define PRIBASE_THREAD (MAXPRI * 3) 58 #define PRIBASE_NULL (MAXPRI * 4) 59 60 #define NQS 32 /* 32 run queues. */ 61 #define PPQ (MAXPRI / NQS) /* priorities per queue */ 62 #define PPQMASK (PPQ - 1) 63 64 /* 65 * NICEPPQ - number of nice units per priority queue 66 * ESTCPURAMP - number of scheduler ticks for estcpu to switch queues 67 * 68 * ESTCPUPPQ - number of estcpu units per priority queue 69 * ESTCPUMAX - number of estcpu units 70 * ESTCPUINCR - amount we have to increment p_estcpu per scheduling tick at 71 * 100% cpu. 72 */ 73 #define NICEPPQ 2 74 #define ESTCPURAMP 4 75 #define ESTCPUPPQ 512 76 #define ESTCPUMAX (ESTCPUPPQ * NQS) 77 #define ESTCPUINCR (ESTCPUPPQ / ESTCPURAMP) 78 #define PRIO_RANGE (PRIO_MAX - PRIO_MIN + 1) 79 80 #define ESTCPULIM(v) min((v), ESTCPUMAX) 81 82 TAILQ_HEAD(rq, lwp); 83 84 #define lwp_priority lwp_usdata.bsd4.priority 85 #define lwp_rqindex lwp_usdata.bsd4.rqindex 86 #define lwp_origcpu lwp_usdata.bsd4.origcpu 87 #define lwp_estcpu lwp_usdata.bsd4.estcpu 88 #define lwp_rqtype lwp_usdata.bsd4.rqtype 89 90 static void bsd4_acquire_curproc(struct lwp *lp); 91 static void bsd4_release_curproc(struct lwp *lp); 92 static void bsd4_select_curproc(globaldata_t gd); 93 static void bsd4_setrunqueue(struct lwp *lp); 94 static void bsd4_schedulerclock(struct lwp *lp, sysclock_t period, 95 sysclock_t cpstamp); 96 static void bsd4_recalculate_estcpu(struct lwp *lp); 97 static void bsd4_resetpriority(struct lwp *lp); 98 static void bsd4_forking(struct lwp *plp, struct lwp *lp); 99 static void bsd4_exiting(struct lwp *plp, struct lwp *lp); 100 101 #ifdef SMP 102 static void need_user_resched_remote(void *dummy); 103 #endif 104 static struct lwp *chooseproc_locked(struct lwp *chklp); 105 static void bsd4_remrunqueue_locked(struct lwp *lp); 106 static void bsd4_setrunqueue_locked(struct lwp *lp); 107 108 struct usched usched_bsd4 = { 109 { NULL }, 110 "bsd4", "Original DragonFly Scheduler", 111 NULL, /* default registration */ 112 NULL, /* default deregistration */ 113 bsd4_acquire_curproc, 114 bsd4_release_curproc, 115 bsd4_setrunqueue, 116 bsd4_schedulerclock, 117 bsd4_recalculate_estcpu, 118 bsd4_resetpriority, 119 bsd4_forking, 120 bsd4_exiting, 121 NULL /* setcpumask not supported */ 122 }; 123 124 struct usched_bsd4_pcpu { 125 struct thread helper_thread; 126 short rrcount; 127 short upri; 128 struct lwp *uschedcp; 129 }; 130 131 typedef struct usched_bsd4_pcpu *bsd4_pcpu_t; 132 133 /* 134 * We have NQS (32) run queues per scheduling class. For the normal 135 * class, there are 128 priorities scaled onto these 32 queues. New 136 * processes are added to the last entry in each queue, and processes 137 * are selected for running by taking them from the head and maintaining 138 * a simple FIFO arrangement. Realtime and Idle priority processes have 139 * and explicit 0-31 priority which maps directly onto their class queue 140 * index. When a queue has something in it, the corresponding bit is 141 * set in the queuebits variable, allowing a single read to determine 142 * the state of all 32 queues and then a ffs() to find the first busy 143 * queue. 144 */ 145 static struct rq bsd4_queues[NQS]; 146 static struct rq bsd4_rtqueues[NQS]; 147 static struct rq bsd4_idqueues[NQS]; 148 static u_int32_t bsd4_queuebits; 149 static u_int32_t bsd4_rtqueuebits; 150 static u_int32_t bsd4_idqueuebits; 151 static cpumask_t bsd4_curprocmask = -1; /* currently running a user process */ 152 static cpumask_t bsd4_rdyprocmask; /* ready to accept a user process */ 153 static int bsd4_runqcount; 154 #ifdef SMP 155 static volatile int bsd4_scancpu; 156 #endif 157 static struct spinlock bsd4_spin; 158 static struct usched_bsd4_pcpu bsd4_pcpu[MAXCPU]; 159 160 SYSCTL_INT(_debug, OID_AUTO, bsd4_runqcount, CTLFLAG_RD, &bsd4_runqcount, 0, ""); 161 #ifdef INVARIANTS 162 static int usched_nonoptimal; 163 SYSCTL_INT(_debug, OID_AUTO, usched_nonoptimal, CTLFLAG_RW, 164 &usched_nonoptimal, 0, "acquire_curproc() was not optimal"); 165 static int usched_optimal; 166 SYSCTL_INT(_debug, OID_AUTO, usched_optimal, CTLFLAG_RW, 167 &usched_optimal, 0, "acquire_curproc() was optimal"); 168 #endif 169 static int usched_debug = -1; 170 SYSCTL_INT(_debug, OID_AUTO, scdebug, CTLFLAG_RW, &usched_debug, 0, ""); 171 #ifdef SMP 172 static int remote_resched_nonaffinity; 173 static int remote_resched_affinity; 174 static int choose_affinity; 175 SYSCTL_INT(_debug, OID_AUTO, remote_resched_nonaffinity, CTLFLAG_RD, 176 &remote_resched_nonaffinity, 0, "Number of remote rescheds"); 177 SYSCTL_INT(_debug, OID_AUTO, remote_resched_affinity, CTLFLAG_RD, 178 &remote_resched_affinity, 0, "Number of remote rescheds"); 179 SYSCTL_INT(_debug, OID_AUTO, choose_affinity, CTLFLAG_RD, 180 &choose_affinity, 0, "chooseproc() was smart"); 181 #endif 182 183 static int usched_bsd4_rrinterval = (ESTCPUFREQ + 9) / 10; 184 SYSCTL_INT(_kern, OID_AUTO, usched_bsd4_rrinterval, CTLFLAG_RW, 185 &usched_bsd4_rrinterval, 0, ""); 186 static int usched_bsd4_decay = ESTCPUINCR / 2; 187 SYSCTL_INT(_kern, OID_AUTO, usched_bsd4_decay, CTLFLAG_RW, 188 &usched_bsd4_decay, 0, ""); 189 190 /* 191 * Initialize the run queues at boot time. 192 */ 193 static void 194 rqinit(void *dummy) 195 { 196 int i; 197 198 spin_init(&bsd4_spin); 199 for (i = 0; i < NQS; i++) { 200 TAILQ_INIT(&bsd4_queues[i]); 201 TAILQ_INIT(&bsd4_rtqueues[i]); 202 TAILQ_INIT(&bsd4_idqueues[i]); 203 } 204 atomic_clear_int(&bsd4_curprocmask, 1); 205 } 206 SYSINIT(runqueue, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, rqinit, NULL) 207 208 /* 209 * BSD4_ACQUIRE_CURPROC 210 * 211 * This function is called when the kernel intends to return to userland. 212 * It is responsible for making the thread the current designated userland 213 * thread for this cpu, blocking if necessary. 214 * 215 * We are expected to handle userland reschedule requests here too. 216 * 217 * WARNING! THIS FUNCTION IS ALLOWED TO CAUSE THE CURRENT THREAD TO MIGRATE 218 * TO ANOTHER CPU! Because most of the kernel assumes that no migration will 219 * occur, this function is called only under very controlled circumstances. 220 * 221 * Basically we recalculate our estcpu to hopefully give us a more 222 * favorable disposition, setrunqueue, then wait for the curlwp 223 * designation to be handed to us (if the setrunqueue didn't do it). 224 * 225 * MPSAFE 226 */ 227 static void 228 bsd4_acquire_curproc(struct lwp *lp) 229 { 230 globaldata_t gd = mycpu; 231 bsd4_pcpu_t dd = &bsd4_pcpu[gd->gd_cpuid]; 232 233 /* 234 * Possibly select another thread, or keep the current thread. 235 */ 236 if (user_resched_wanted()) 237 bsd4_select_curproc(gd); 238 239 /* 240 * If uschedcp is still pointing to us, we're done 241 */ 242 if (dd->uschedcp == lp) 243 return; 244 245 /* 246 * If this cpu has no current thread, and the run queue is 247 * empty, we can safely select ourself. 248 */ 249 if (dd->uschedcp == NULL && bsd4_runqcount == 0) { 250 atomic_set_int(&bsd4_curprocmask, gd->gd_cpumask); 251 dd->uschedcp = lp; 252 dd->upri = lp->lwp_priority; 253 return; 254 } 255 256 /* 257 * Adjust estcpu and recalculate our priority, then put us back on 258 * the user process scheduler's runq. Only increment the involuntary 259 * context switch count if the setrunqueue call did not immediately 260 * schedule us. 261 * 262 * Loop until we become the currently scheduled process. Note that 263 * calling setrunqueue can cause us to be migrated to another cpu 264 * after we switch away. 265 */ 266 do { 267 crit_enter(); 268 bsd4_recalculate_estcpu(lp); 269 lwkt_deschedule_self(gd->gd_curthread); 270 bsd4_setrunqueue(lp); 271 if ((gd->gd_curthread->td_flags & TDF_RUNQ) == 0) 272 ++lp->lwp_stats->p_ru.ru_nivcsw; 273 lwkt_switch(); 274 crit_exit(); 275 gd = mycpu; 276 dd = &bsd4_pcpu[gd->gd_cpuid]; 277 } while (dd->uschedcp != lp); 278 KKASSERT((lp->lwp_proc->p_flag & P_ONRUNQ) == 0); 279 } 280 281 /* 282 * BSD4_RELEASE_CURPROC 283 * 284 * This routine detaches the current thread from the userland scheduler, 285 * usually because the thread needs to run in the kernel (at kernel priority) 286 * for a while. 287 * 288 * This routine is also responsible for selecting a new thread to 289 * make the current thread. 290 * 291 * NOTE: This implementation differs from the dummy example in that 292 * bsd4_select_curproc() is able to select the current process, whereas 293 * dummy_select_curproc() is not able to select the current process. 294 * This means we have to NULL out uschedcp. 295 * 296 * Additionally, note that we may already be on a run queue if releasing 297 * via the lwkt_switch() in bsd4_setrunqueue(). 298 * 299 * WARNING! The MP lock may be in an unsynchronized state due to the 300 * way get_mplock() works and the fact that this function may be called 301 * from a passive release during a lwkt_switch(). try_mplock() will deal 302 * with this for us but you should be aware that td_mpcount may not be 303 * useable. 304 * 305 * MPSAFE 306 */ 307 static void 308 bsd4_release_curproc(struct lwp *lp) 309 { 310 globaldata_t gd = mycpu; 311 bsd4_pcpu_t dd = &bsd4_pcpu[gd->gd_cpuid]; 312 313 if (dd->uschedcp == lp) { 314 /* 315 * Note: we leave ou curprocmask bit set to prevent 316 * unnecessary scheduler helper wakeups. 317 * bsd4_select_curproc() will clean it up. 318 */ 319 KKASSERT((lp->lwp_proc->p_flag & P_ONRUNQ) == 0); 320 dd->uschedcp = NULL; /* don't let lp be selected */ 321 bsd4_select_curproc(gd); 322 } 323 } 324 325 /* 326 * BSD4_SELECT_CURPROC 327 * 328 * Select a new current process for this cpu. This satisfies a user 329 * scheduler reschedule request so clear that too. 330 * 331 * This routine is also responsible for equal-priority round-robining, 332 * typically triggered from bsd4_schedulerclock(). In our dummy example 333 * all the 'user' threads are LWKT scheduled all at once and we just 334 * call lwkt_switch(). 335 * 336 * MPSAFE 337 */ 338 static 339 void 340 bsd4_select_curproc(globaldata_t gd) 341 { 342 bsd4_pcpu_t dd = &bsd4_pcpu[gd->gd_cpuid]; 343 struct lwp *nlp; 344 int cpuid = gd->gd_cpuid; 345 346 crit_enter_gd(gd); 347 clear_user_resched(); /* This satisfied the reschedule request */ 348 dd->rrcount = 0; /* Reset the round-robin counter */ 349 350 spin_lock_wr(&bsd4_spin); 351 if ((nlp = chooseproc_locked(dd->uschedcp)) != NULL) { 352 atomic_set_int(&bsd4_curprocmask, 1 << cpuid); 353 dd->upri = nlp->lwp_priority; 354 dd->uschedcp = nlp; 355 spin_unlock_wr(&bsd4_spin); 356 #ifdef SMP 357 lwkt_acquire(nlp->lwp_thread); 358 #endif 359 lwkt_schedule(nlp->lwp_thread); 360 } else if (dd->uschedcp) { 361 dd->upri = dd->uschedcp->lwp_priority; 362 spin_unlock_wr(&bsd4_spin); 363 KKASSERT(bsd4_curprocmask & (1 << cpuid)); 364 } else if (bsd4_runqcount && (bsd4_rdyprocmask & (1 << cpuid))) { 365 atomic_clear_int(&bsd4_curprocmask, 1 << cpuid); 366 atomic_clear_int(&bsd4_rdyprocmask, 1 << cpuid); 367 dd->uschedcp = NULL; 368 dd->upri = PRIBASE_NULL; 369 spin_unlock_wr(&bsd4_spin); 370 lwkt_schedule(&dd->helper_thread); 371 } else { 372 dd->uschedcp = NULL; 373 dd->upri = PRIBASE_NULL; 374 atomic_clear_int(&bsd4_curprocmask, 1 << cpuid); 375 spin_unlock_wr(&bsd4_spin); 376 } 377 crit_exit_gd(gd); 378 } 379 380 /* 381 * BSD4_SETRUNQUEUE 382 * 383 * This routine is called to schedule a new user process after a fork. 384 * 385 * The caller may set P_PASSIVE_ACQ in p_flag to indicate that we should 386 * attempt to leave the thread on the current cpu. 387 * 388 * If P_PASSIVE_ACQ is set setrunqueue() will not wakeup potential target 389 * cpus in an attempt to keep the process on the current cpu at least for 390 * a little while to take advantage of locality of reference (e.g. fork/exec 391 * or short fork/exit, and uio_yield()). 392 * 393 * CPU AFFINITY: cpu affinity is handled by attempting to either schedule 394 * or (user level) preempt on the same cpu that a process was previously 395 * scheduled to. If we cannot do this but we are at enough of a higher 396 * priority then the processes running on other cpus, we will allow the 397 * process to be stolen by another cpu. 398 * 399 * WARNING! This routine cannot block. bsd4_acquire_curproc() does 400 * a deschedule/switch interlock and we can be moved to another cpu 401 * the moment we are switched out. Our LWKT run state is the only 402 * thing preventing the transfer. 403 * 404 * The associated thread must NOT currently be scheduled (but can be the 405 * current process after it has been LWKT descheduled). It must NOT be on 406 * a bsd4 scheduler queue either. The purpose of this routine is to put 407 * it on a scheduler queue or make it the current user process and LWKT 408 * schedule it. It is possible that the thread is in the middle of a LWKT 409 * switchout on another cpu, lwkt_acquire() deals with that case. 410 * 411 * The process must be runnable. 412 * 413 * MPSAFE 414 */ 415 static void 416 bsd4_setrunqueue(struct lwp *lp) 417 { 418 globaldata_t gd; 419 bsd4_pcpu_t dd; 420 int cpuid; 421 #ifdef SMP 422 cpumask_t mask; 423 cpumask_t tmpmask; 424 #endif 425 426 /* 427 * First validate the process state relative to the current cpu. 428 * We don't need the spinlock for this, just a critical section. 429 * We are in control of the process. 430 */ 431 crit_enter(); 432 KASSERT(lp->lwp_proc->p_stat == SRUN, ("setrunqueue: proc not SRUN")); 433 KASSERT((lp->lwp_proc->p_flag & P_ONRUNQ) == 0, 434 ("lwp %d/%d already on runq! flag %08x", lp->lwp_proc->p_pid, 435 lp->lwp_tid, lp->lwp_proc->p_flag)); 436 KKASSERT((lp->lwp_thread->td_flags & TDF_RUNQ) == 0); 437 438 /* 439 * Note: gd and dd are relative to the target thread's last cpu, 440 * NOT our current cpu. 441 */ 442 gd = lp->lwp_thread->td_gd; 443 dd = &bsd4_pcpu[gd->gd_cpuid]; 444 445 /* 446 * This process is not supposed to be scheduled anywhere or assigned 447 * as the current process anywhere. Assert the condition. 448 */ 449 KKASSERT(dd->uschedcp != lp); 450 451 /* 452 * Check local cpu affinity. The associated thread is stable at 453 * the moment. Note that we may be checking another cpu here so we 454 * have to be careful. We can only assign uschedcp on OUR cpu. 455 * 456 * This allows us to avoid actually queueing the process. 457 * acquire_curproc() will handle any threads we mistakenly schedule. 458 */ 459 cpuid = gd->gd_cpuid; 460 if (gd == mycpu && (bsd4_curprocmask & (1 << cpuid)) == 0) { 461 atomic_set_int(&bsd4_curprocmask, 1 << cpuid); 462 dd->uschedcp = lp; 463 dd->upri = lp->lwp_priority; 464 lwkt_schedule(lp->lwp_thread); 465 crit_exit(); 466 return; 467 } 468 469 /* 470 * gd and cpuid may still 'hint' at another cpu. Even so we have 471 * to place this process on the userland scheduler's run queue for 472 * action by the target cpu. 473 */ 474 #ifdef SMP 475 /* 476 * XXX fixme. Could be part of a remrunqueue/setrunqueue 477 * operation when the priority is recalculated, so TDF_MIGRATING 478 * may already be set. 479 */ 480 if ((lp->lwp_thread->td_flags & TDF_MIGRATING) == 0) 481 lwkt_giveaway(lp->lwp_thread); 482 #endif 483 484 /* 485 * We lose control of lp the moment we release the spinlock after 486 * having placed lp on the queue. i.e. another cpu could pick it 487 * up and it could exit, or its priority could be further adjusted, 488 * or something like that. 489 */ 490 spin_lock_wr(&bsd4_spin); 491 bsd4_setrunqueue_locked(lp); 492 493 /* 494 * gd, dd, and cpuid are still our target cpu 'hint', not our current 495 * cpu info. 496 * 497 * We always try to schedule a LWP to its original cpu first. It 498 * is possible for the scheduler helper or setrunqueue to assign 499 * the LWP to a different cpu before the one we asked for wakes 500 * up. 501 * 502 * If the LWP has higher priority (lower lwp_priority value) on 503 * its target cpu, reschedule on that cpu. 504 */ 505 if ((lp->lwp_thread->td_flags & TDF_NORESCHED) == 0) { 506 if ((dd->upri & ~PRIMASK) > (lp->lwp_priority & ~PRIMASK)) { 507 dd->upri = lp->lwp_priority; 508 spin_unlock_wr(&bsd4_spin); 509 #ifdef SMP 510 if (gd == mycpu) { 511 need_user_resched(); 512 } else { 513 lwkt_send_ipiq(gd, need_user_resched_remote, 514 NULL); 515 } 516 #else 517 need_user_resched(); 518 #endif 519 crit_exit(); 520 return; 521 } 522 } 523 spin_unlock_wr(&bsd4_spin); 524 525 #ifdef SMP 526 /* 527 * Otherwise the LWP has a lower priority or we were asked not 528 * to reschedule. Look for an idle cpu whos scheduler helper 529 * is ready to accept more work. 530 * 531 * Look for an idle cpu starting at our rotator (bsd4_scancpu). 532 * 533 * If no cpus are ready to accept work, just return. 534 * 535 * XXX P_PASSIVE_ACQ 536 */ 537 mask = ~bsd4_curprocmask & bsd4_rdyprocmask & mycpu->gd_other_cpus & 538 lp->lwp_cpumask; 539 if (mask) { 540 cpuid = bsd4_scancpu; 541 if (++cpuid == ncpus) 542 cpuid = 0; 543 tmpmask = ~((1 << cpuid) - 1); 544 if (mask & tmpmask) 545 cpuid = bsfl(mask & tmpmask); 546 else 547 cpuid = bsfl(mask); 548 atomic_clear_int(&bsd4_rdyprocmask, 1 << cpuid); 549 bsd4_scancpu = cpuid; 550 lwkt_schedule(&bsd4_pcpu[cpuid].helper_thread); 551 } 552 #endif 553 crit_exit(); 554 } 555 556 /* 557 * This routine is called from a systimer IPI. It MUST be MP-safe and 558 * the BGL IS NOT HELD ON ENTRY. This routine is called at ESTCPUFREQ on 559 * each cpu. 560 * 561 * Because this is effectively a 'fast' interrupt, we cannot safely 562 * use spinlocks unless gd_spinlock_rd is NULL and gd_spinlocks_wr is 0, 563 * even if the spinlocks are 'non conflicting'. This is due to the way 564 * spinlock conflicts against cached read locks are handled. 565 * 566 * MPSAFE 567 */ 568 static 569 void 570 bsd4_schedulerclock(struct lwp *lp, sysclock_t period, sysclock_t cpstamp) 571 { 572 globaldata_t gd = mycpu; 573 bsd4_pcpu_t dd = &bsd4_pcpu[gd->gd_cpuid]; 574 575 /* 576 * Do we need to round-robin? We round-robin 10 times a second. 577 * This should only occur for cpu-bound batch processes. 578 */ 579 if (++dd->rrcount >= usched_bsd4_rrinterval) { 580 dd->rrcount = 0; 581 need_user_resched(); 582 } 583 584 /* 585 * As the process accumulates cpu time p_estcpu is bumped and may 586 * push the process into another scheduling queue. It typically 587 * takes 4 ticks to bump the queue. 588 */ 589 lp->lwp_estcpu = ESTCPULIM(lp->lwp_estcpu + ESTCPUINCR); 590 591 /* 592 * Reducing p_origcpu over time causes more of our estcpu to be 593 * returned to the parent when we exit. This is a small tweak 594 * for the batch detection heuristic. 595 */ 596 if (lp->lwp_origcpu) 597 --lp->lwp_origcpu; 598 599 /* 600 * We can only safely call bsd4_resetpriority(), which uses spinlocks, 601 * if we aren't interrupting a thread that is using spinlocks. 602 * Otherwise we can deadlock with another cpu waiting for our read 603 * spinlocks to clear. 604 */ 605 if (gd->gd_spinlock_rd == NULL && gd->gd_spinlocks_wr == 0) 606 bsd4_resetpriority(lp); 607 else 608 need_user_resched(); 609 } 610 611 /* 612 * Called from acquire and from kern_synch's one-second timer (one of the 613 * callout helper threads) with a critical section held. 614 * 615 * Decay p_estcpu based on the number of ticks we haven't been running 616 * and our p_nice. As the load increases each process observes a larger 617 * number of idle ticks (because other processes are running in them). 618 * This observation leads to a larger correction which tends to make the 619 * system more 'batchy'. 620 * 621 * Note that no recalculation occurs for a process which sleeps and wakes 622 * up in the same tick. That is, a system doing thousands of context 623 * switches per second will still only do serious estcpu calculations 624 * ESTCPUFREQ times per second. 625 * 626 * MPSAFE 627 */ 628 static 629 void 630 bsd4_recalculate_estcpu(struct lwp *lp) 631 { 632 globaldata_t gd = mycpu; 633 sysclock_t cpbase; 634 int loadfac; 635 int ndecay; 636 int nticks; 637 int nleft; 638 639 /* 640 * We have to subtract periodic to get the last schedclock 641 * timeout time, otherwise we would get the upcoming timeout. 642 * Keep in mind that a process can migrate between cpus and 643 * while the scheduler clock should be very close, boundary 644 * conditions could lead to a small negative delta. 645 */ 646 cpbase = gd->gd_schedclock.time - gd->gd_schedclock.periodic; 647 648 if (lp->lwp_slptime > 1) { 649 /* 650 * Too much time has passed, do a coarse correction. 651 */ 652 lp->lwp_estcpu = lp->lwp_estcpu >> 1; 653 bsd4_resetpriority(lp); 654 lp->lwp_cpbase = cpbase; 655 lp->lwp_cpticks = 0; 656 } else if (lp->lwp_cpbase != cpbase) { 657 /* 658 * Adjust estcpu if we are in a different tick. Don't waste 659 * time if we are in the same tick. 660 * 661 * First calculate the number of ticks in the measurement 662 * interval. The nticks calculation can wind up 0 due to 663 * a bug in the handling of lwp_slptime (as yet not found), 664 * so make sure we do not get a divide by 0 panic. 665 */ 666 nticks = (cpbase - lp->lwp_cpbase) / gd->gd_schedclock.periodic; 667 if (nticks <= 0) 668 nticks = 1; 669 updatepcpu(lp, lp->lwp_cpticks, nticks); 670 671 if ((nleft = nticks - lp->lwp_cpticks) < 0) 672 nleft = 0; 673 if (usched_debug == lp->lwp_proc->p_pid) { 674 printf("pid %d tid %d estcpu %d cpticks %d nticks %d nleft %d", 675 lp->lwp_proc->p_pid, lp->lwp_tid, lp->lwp_estcpu, 676 lp->lwp_cpticks, nticks, nleft); 677 } 678 679 /* 680 * Calculate a decay value based on ticks remaining scaled 681 * down by the instantanious load and p_nice. 682 */ 683 if ((loadfac = bsd4_runqcount) < 2) 684 loadfac = 2; 685 ndecay = nleft * usched_bsd4_decay * 2 * 686 (PRIO_MAX * 2 - lp->lwp_proc->p_nice) / (loadfac * PRIO_MAX * 2); 687 688 /* 689 * Adjust p_estcpu. Handle a border case where batch jobs 690 * can get stalled long enough to decay to zero when they 691 * shouldn't. 692 */ 693 if (lp->lwp_estcpu > ndecay * 2) 694 lp->lwp_estcpu -= ndecay; 695 else 696 lp->lwp_estcpu >>= 1; 697 698 if (usched_debug == lp->lwp_proc->p_pid) 699 printf(" ndecay %d estcpu %d\n", ndecay, lp->lwp_estcpu); 700 bsd4_resetpriority(lp); 701 lp->lwp_cpbase = cpbase; 702 lp->lwp_cpticks = 0; 703 } 704 } 705 706 /* 707 * Compute the priority of a process when running in user mode. 708 * Arrange to reschedule if the resulting priority is better 709 * than that of the current process. 710 * 711 * This routine may be called with any process. 712 * 713 * This routine is called by fork1() for initial setup with the process 714 * of the run queue, and also may be called normally with the process on or 715 * off the run queue. 716 * 717 * MPSAFE 718 */ 719 static void 720 bsd4_resetpriority(struct lwp *lp) 721 { 722 bsd4_pcpu_t dd; 723 int newpriority; 724 u_short newrqtype; 725 int reschedcpu; 726 727 /* 728 * Calculate the new priority and queue type 729 */ 730 crit_enter(); 731 spin_lock_wr(&bsd4_spin); 732 733 newrqtype = lp->lwp_rtprio.type; 734 735 switch(newrqtype) { 736 case RTP_PRIO_REALTIME: 737 case RTP_PRIO_FIFO: 738 newpriority = PRIBASE_REALTIME + 739 (lp->lwp_rtprio.prio & PRIMASK); 740 break; 741 case RTP_PRIO_NORMAL: 742 newpriority = (lp->lwp_proc->p_nice - PRIO_MIN) * PPQ / NICEPPQ; 743 newpriority += lp->lwp_estcpu * PPQ / ESTCPUPPQ; 744 newpriority = newpriority * MAXPRI / (PRIO_RANGE * PPQ / 745 NICEPPQ + ESTCPUMAX * PPQ / ESTCPUPPQ); 746 newpriority = PRIBASE_NORMAL + (newpriority & PRIMASK); 747 break; 748 case RTP_PRIO_IDLE: 749 newpriority = PRIBASE_IDLE + (lp->lwp_rtprio.prio & PRIMASK); 750 break; 751 case RTP_PRIO_THREAD: 752 newpriority = PRIBASE_THREAD + (lp->lwp_rtprio.prio & PRIMASK); 753 break; 754 default: 755 panic("Bad RTP_PRIO %d", newrqtype); 756 /* NOT REACHED */ 757 } 758 759 /* 760 * The newpriority incorporates the queue type so do a simple masked 761 * check to determine if the process has moved to another queue. If 762 * it has, and it is currently on a run queue, then move it. 763 */ 764 if ((lp->lwp_priority ^ newpriority) & ~PPQMASK) { 765 lp->lwp_priority = newpriority; 766 if (lp->lwp_proc->p_flag & P_ONRUNQ) { 767 bsd4_remrunqueue_locked(lp); 768 lp->lwp_rqtype = newrqtype; 769 lp->lwp_rqindex = (newpriority & PRIMASK) / PPQ; 770 bsd4_setrunqueue_locked(lp); 771 reschedcpu = lp->lwp_thread->td_gd->gd_cpuid; 772 } else { 773 lp->lwp_rqtype = newrqtype; 774 lp->lwp_rqindex = (newpriority & PRIMASK) / PPQ; 775 reschedcpu = -1; 776 } 777 } else { 778 lp->lwp_priority = newpriority; 779 reschedcpu = -1; 780 } 781 spin_unlock_wr(&bsd4_spin); 782 783 /* 784 * Determine if we need to reschedule the target cpu. This only 785 * occurs if the LWP is already on a scheduler queue, which means 786 * that idle cpu notification has already occured. At most we 787 * need only issue a need_user_resched() on the appropriate cpu. 788 */ 789 if (reschedcpu >= 0) { 790 dd = &bsd4_pcpu[reschedcpu]; 791 KKASSERT(dd->uschedcp != lp); 792 if ((dd->upri & ~PRIMASK) > (lp->lwp_priority & ~PRIMASK)) { 793 dd->upri = lp->lwp_priority; 794 #ifdef SMP 795 if (reschedcpu == mycpu->gd_cpuid) { 796 need_user_resched(); 797 } else { 798 lwkt_send_ipiq(lp->lwp_thread->td_gd, 799 need_user_resched_remote, NULL); 800 } 801 #else 802 need_user_resched(); 803 #endif 804 } 805 } 806 crit_exit(); 807 } 808 809 /* 810 * Called from fork1() when a new child process is being created. 811 * 812 * Give the child process an initial estcpu that is more batch then 813 * its parent and dock the parent for the fork (but do not 814 * reschedule the parent). This comprises the main part of our batch 815 * detection heuristic for both parallel forking and sequential execs. 816 * 817 * Interactive processes will decay the boosted estcpu quickly while batch 818 * processes will tend to compound it. 819 * XXX lwp should be "spawning" instead of "forking" 820 * 821 * MPSAFE 822 */ 823 static void 824 bsd4_forking(struct lwp *plp, struct lwp *lp) 825 { 826 lp->lwp_estcpu = ESTCPULIM(plp->lwp_estcpu + ESTCPUPPQ); 827 lp->lwp_origcpu = lp->lwp_estcpu; 828 plp->lwp_estcpu = ESTCPULIM(plp->lwp_estcpu + ESTCPUPPQ); 829 } 830 831 /* 832 * Called when the parent reaps a child. Propogate cpu use by the child 833 * back to the parent. 834 * 835 * MPSAFE 836 */ 837 static void 838 bsd4_exiting(struct lwp *plp, struct lwp *lp) 839 { 840 int delta; 841 842 if (plp->lwp_proc->p_pid != 1) { 843 delta = lp->lwp_estcpu - lp->lwp_origcpu; 844 if (delta > 0) 845 plp->lwp_estcpu = ESTCPULIM(plp->lwp_estcpu + delta); 846 } 847 } 848 849 850 /* 851 * chooseproc() is called when a cpu needs a user process to LWKT schedule, 852 * it selects a user process and returns it. If chklp is non-NULL and chklp 853 * has a better or equal priority then the process that would otherwise be 854 * chosen, NULL is returned. 855 * 856 * Until we fix the RUNQ code the chklp test has to be strict or we may 857 * bounce between processes trying to acquire the current process designation. 858 * 859 * MPSAFE - must be called with bsd4_spin exclusive held. The spinlock is 860 * left intact through the entire routine. 861 */ 862 static 863 struct lwp * 864 chooseproc_locked(struct lwp *chklp) 865 { 866 struct lwp *lp; 867 struct rq *q; 868 u_int32_t *which, *which2; 869 u_int32_t pri; 870 u_int32_t rtqbits; 871 u_int32_t tsqbits; 872 u_int32_t idqbits; 873 cpumask_t cpumask; 874 875 rtqbits = bsd4_rtqueuebits; 876 tsqbits = bsd4_queuebits; 877 idqbits = bsd4_idqueuebits; 878 cpumask = mycpu->gd_cpumask; 879 880 #ifdef SMP 881 again: 882 #endif 883 if (rtqbits) { 884 pri = bsfl(rtqbits); 885 q = &bsd4_rtqueues[pri]; 886 which = &bsd4_rtqueuebits; 887 which2 = &rtqbits; 888 } else if (tsqbits) { 889 pri = bsfl(tsqbits); 890 q = &bsd4_queues[pri]; 891 which = &bsd4_queuebits; 892 which2 = &tsqbits; 893 } else if (idqbits) { 894 pri = bsfl(idqbits); 895 q = &bsd4_idqueues[pri]; 896 which = &bsd4_idqueuebits; 897 which2 = &idqbits; 898 } else { 899 return NULL; 900 } 901 lp = TAILQ_FIRST(q); 902 KASSERT(lp, ("chooseproc: no lwp on busy queue")); 903 904 #ifdef SMP 905 while ((lp->lwp_cpumask & cpumask) == 0) { 906 lp = TAILQ_NEXT(lp, lwp_procq); 907 if (lp == NULL) { 908 *which2 &= ~(1 << pri); 909 goto again; 910 } 911 } 912 #endif 913 914 /* 915 * If the passed lwp <chklp> is reasonably close to the selected 916 * lwp <lp>, return NULL (indicating that <chklp> should be kept). 917 * 918 * Note that we must error on the side of <chklp> to avoid bouncing 919 * between threads in the acquire code. 920 */ 921 if (chklp) { 922 if (chklp->lwp_priority < lp->lwp_priority + PPQ) 923 return(NULL); 924 } 925 926 #ifdef SMP 927 /* 928 * If the chosen lwp does not reside on this cpu spend a few 929 * cycles looking for a better candidate at the same priority level. 930 * This is a fallback check, setrunqueue() tries to wakeup the 931 * correct cpu and is our front-line affinity. 932 */ 933 if (lp->lwp_thread->td_gd != mycpu && 934 (chklp = TAILQ_NEXT(lp, lwp_procq)) != NULL 935 ) { 936 if (chklp->lwp_thread->td_gd == mycpu) { 937 ++choose_affinity; 938 lp = chklp; 939 } 940 } 941 #endif 942 943 TAILQ_REMOVE(q, lp, lwp_procq); 944 --bsd4_runqcount; 945 if (TAILQ_EMPTY(q)) 946 *which &= ~(1 << pri); 947 KASSERT((lp->lwp_proc->p_flag & P_ONRUNQ) != 0, ("not on runq6!")); 948 lp->lwp_proc->p_flag &= ~P_ONRUNQ; 949 return lp; 950 } 951 952 #ifdef SMP 953 /* 954 * Called via an ipi message to reschedule on another cpu. 955 * 956 * MPSAFE 957 */ 958 static 959 void 960 need_user_resched_remote(void *dummy) 961 { 962 need_user_resched(); 963 } 964 965 #endif 966 967 968 /* 969 * bsd4_remrunqueue_locked() removes a given process from the run queue 970 * that it is on, clearing the queue busy bit if it becomes empty. 971 * 972 * Note that user process scheduler is different from the LWKT schedule. 973 * The user process scheduler only manages user processes but it uses LWKT 974 * underneath, and a user process operating in the kernel will often be 975 * 'released' from our management. 976 * 977 * MPSAFE - bsd4_spin must be held exclusively on call 978 */ 979 static void 980 bsd4_remrunqueue_locked(struct lwp *lp) 981 { 982 struct rq *q; 983 u_int32_t *which; 984 u_int8_t pri; 985 986 KKASSERT(lp->lwp_proc->p_flag & P_ONRUNQ); 987 lp->lwp_proc->p_flag &= ~P_ONRUNQ; 988 --bsd4_runqcount; 989 KKASSERT(bsd4_runqcount >= 0); 990 991 pri = lp->lwp_rqindex; 992 switch(lp->lwp_rqtype) { 993 case RTP_PRIO_NORMAL: 994 q = &bsd4_queues[pri]; 995 which = &bsd4_queuebits; 996 break; 997 case RTP_PRIO_REALTIME: 998 case RTP_PRIO_FIFO: 999 q = &bsd4_rtqueues[pri]; 1000 which = &bsd4_rtqueuebits; 1001 break; 1002 case RTP_PRIO_IDLE: 1003 q = &bsd4_idqueues[pri]; 1004 which = &bsd4_idqueuebits; 1005 break; 1006 default: 1007 panic("remrunqueue: invalid rtprio type"); 1008 /* NOT REACHED */ 1009 } 1010 TAILQ_REMOVE(q, lp, lwp_procq); 1011 if (TAILQ_EMPTY(q)) { 1012 KASSERT((*which & (1 << pri)) != 0, 1013 ("remrunqueue: remove from empty queue")); 1014 *which &= ~(1 << pri); 1015 } 1016 } 1017 1018 /* 1019 * bsd4_setrunqueue_locked() 1020 * 1021 * Add a process whos rqtype and rqindex had previously been calculated 1022 * onto the appropriate run queue. Determine if the addition requires 1023 * a reschedule on a cpu and return the cpuid or -1. 1024 * 1025 * NOTE: Lower priorities are better priorities. 1026 * 1027 * MPSAFE - bsd4_spin must be held exclusively on call 1028 */ 1029 static void 1030 bsd4_setrunqueue_locked(struct lwp *lp) 1031 { 1032 struct rq *q; 1033 u_int32_t *which; 1034 int pri; 1035 1036 KKASSERT((lp->lwp_proc->p_flag & P_ONRUNQ) == 0); 1037 lp->lwp_proc->p_flag |= P_ONRUNQ; 1038 ++bsd4_runqcount; 1039 1040 pri = lp->lwp_rqindex; 1041 1042 switch(lp->lwp_rqtype) { 1043 case RTP_PRIO_NORMAL: 1044 q = &bsd4_queues[pri]; 1045 which = &bsd4_queuebits; 1046 break; 1047 case RTP_PRIO_REALTIME: 1048 case RTP_PRIO_FIFO: 1049 q = &bsd4_rtqueues[pri]; 1050 which = &bsd4_rtqueuebits; 1051 break; 1052 case RTP_PRIO_IDLE: 1053 q = &bsd4_idqueues[pri]; 1054 which = &bsd4_idqueuebits; 1055 break; 1056 default: 1057 panic("remrunqueue: invalid rtprio type"); 1058 /* NOT REACHED */ 1059 } 1060 1061 /* 1062 * Add to the correct queue and set the appropriate bit. If no 1063 * lower priority (i.e. better) processes are in the queue then 1064 * we want a reschedule, calculate the best cpu for the job. 1065 * 1066 * Always run reschedules on the LWPs original cpu. 1067 */ 1068 TAILQ_INSERT_TAIL(q, lp, lwp_procq); 1069 *which |= 1 << pri; 1070 } 1071 1072 #ifdef SMP 1073 1074 /* 1075 * For SMP systems a user scheduler helper thread is created for each 1076 * cpu and is used to allow one cpu to wakeup another for the purposes of 1077 * scheduling userland threads from setrunqueue(). UP systems do not 1078 * need the helper since there is only one cpu. We can't use the idle 1079 * thread for this because we need to hold the MP lock. Additionally, 1080 * doing things this way allows us to HLT idle cpus on MP systems. 1081 * 1082 * MPSAFE 1083 */ 1084 static void 1085 sched_thread(void *dummy) 1086 { 1087 globaldata_t gd; 1088 bsd4_pcpu_t dd; 1089 struct lwp *nlp; 1090 cpumask_t cpumask; 1091 cpumask_t tmpmask; 1092 int cpuid; 1093 int tmpid; 1094 1095 gd = mycpu; 1096 cpuid = gd->gd_cpuid; /* doesn't change */ 1097 cpumask = 1 << cpuid; /* doesn't change */ 1098 dd = &bsd4_pcpu[cpuid]; 1099 1100 /* 1101 * The scheduler thread does not need to hold the MP lock. Since we 1102 * are woken up only when no user processes are scheduled on a cpu, we 1103 * can run at an ultra low priority. 1104 */ 1105 rel_mplock(); 1106 lwkt_setpri_self(TDPRI_USER_SCHEDULER); 1107 1108 for (;;) { 1109 /* 1110 * We use the LWKT deschedule-interlock trick to avoid racing 1111 * bsd4_rdyprocmask. This means we cannot block through to the 1112 * manual lwkt_switch() call we make below. 1113 */ 1114 crit_enter_gd(gd); 1115 lwkt_deschedule_self(gd->gd_curthread); 1116 spin_lock_wr(&bsd4_spin); 1117 atomic_set_int(&bsd4_rdyprocmask, cpumask); 1118 if ((bsd4_curprocmask & cpumask) == 0) { 1119 if ((nlp = chooseproc_locked(NULL)) != NULL) { 1120 atomic_set_int(&bsd4_curprocmask, cpumask); 1121 dd->upri = nlp->lwp_priority; 1122 dd->uschedcp = nlp; 1123 spin_unlock_wr(&bsd4_spin); 1124 lwkt_acquire(nlp->lwp_thread); 1125 lwkt_schedule(nlp->lwp_thread); 1126 } else { 1127 spin_unlock_wr(&bsd4_spin); 1128 } 1129 } else { 1130 /* 1131 * Someone scheduled us but raced. In order to not lose 1132 * track of the fact that there may be a LWP ready to go, 1133 * forward the request to another cpu if available. 1134 * 1135 * Rotate through cpus starting with cpuid + 1. Since cpuid 1136 * is already masked out by gd_other_cpus, just use ~cpumask. 1137 */ 1138 tmpmask = ~bsd4_curprocmask & bsd4_rdyprocmask & 1139 mycpu->gd_other_cpus; 1140 if (tmpmask) { 1141 if (tmpmask & ~(cpumask - 1)) 1142 tmpid = bsfl(tmpmask & ~(cpumask - 1)); 1143 else 1144 tmpid = bsfl(tmpmask); 1145 bsd4_scancpu = tmpid; 1146 atomic_clear_int(&bsd4_rdyprocmask, 1 << tmpid); 1147 spin_unlock_wr(&bsd4_spin); 1148 lwkt_schedule(&bsd4_pcpu[tmpid].helper_thread); 1149 } else { 1150 spin_unlock_wr(&bsd4_spin); 1151 } 1152 } 1153 crit_exit_gd(gd); 1154 lwkt_switch(); 1155 } 1156 } 1157 1158 /* 1159 * Setup our scheduler helpers. Note that curprocmask bit 0 has already 1160 * been cleared by rqinit() and we should not mess with it further. 1161 */ 1162 static void 1163 sched_thread_cpu_init(void) 1164 { 1165 int i; 1166 1167 if (bootverbose) 1168 printf("start scheduler helpers on cpus:"); 1169 1170 for (i = 0; i < ncpus; ++i) { 1171 bsd4_pcpu_t dd = &bsd4_pcpu[i]; 1172 cpumask_t mask = 1 << i; 1173 1174 if ((mask & smp_active_mask) == 0) 1175 continue; 1176 1177 if (bootverbose) 1178 printf(" %d", i); 1179 1180 lwkt_create(sched_thread, NULL, NULL, &dd->helper_thread, 1181 TDF_STOPREQ, i, "usched %d", i); 1182 1183 /* 1184 * Allow user scheduling on the target cpu. cpu #0 has already 1185 * been enabled in rqinit(). 1186 */ 1187 if (i) 1188 atomic_clear_int(&bsd4_curprocmask, mask); 1189 atomic_set_int(&bsd4_rdyprocmask, mask); 1190 } 1191 if (bootverbose) 1192 printf("\n"); 1193 } 1194 SYSINIT(uschedtd, SI_SUB_FINISH_SMP, SI_ORDER_ANY, sched_thread_cpu_init, NULL) 1195 1196 #endif 1197 1198