1 /*- 2 * Copyright (c) 1982, 1986, 1990, 1991, 1993 3 * The Regents of the University of California. All rights reserved. 4 * (c) UNIX System Laboratories, Inc. 5 * All or some portions of this file are derived from material licensed 6 * to the University of California by American Telephone and Telegraph 7 * Co. or Unix System Laboratories, Inc. and are reproduced herein with 8 * the permission of UNIX System Laboratories, Inc. 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 * 1. Redistributions of source code must retain the above copyright 14 * notice, this list of conditions and the following disclaimer. 15 * 2. Redistributions in binary form must reproduce the above copyright 16 * notice, this list of conditions and the following disclaimer in the 17 * documentation and/or other materials provided with the distribution. 18 * 3. Neither the name of the University nor the names of its contributors 19 * may be used to endorse or promote products derived from this software 20 * without specific prior written permission. 21 * 22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 * 34 * @(#)kern_synch.c 8.9 (Berkeley) 5/19/95 35 * $FreeBSD: src/sys/kern/kern_synch.c,v 1.87.2.6 2002/10/13 07:29:53 kbyanc Exp $ 36 */ 37 38 #include "opt_ktrace.h" 39 40 #include <sys/param.h> 41 #include <sys/systm.h> 42 #include <sys/proc.h> 43 #include <sys/kernel.h> 44 #include <sys/signalvar.h> 45 #include <sys/resourcevar.h> 46 #include <sys/vmmeter.h> 47 #include <sys/sysctl.h> 48 #include <sys/lock.h> 49 #include <sys/uio.h> 50 #ifdef KTRACE 51 #include <sys/ktrace.h> 52 #endif 53 #include <sys/xwait.h> 54 #include <sys/ktr.h> 55 #include <sys/serialize.h> 56 57 #include <sys/signal2.h> 58 #include <sys/thread2.h> 59 #include <sys/spinlock2.h> 60 #include <sys/mutex2.h> 61 62 #include <machine/cpu.h> 63 #include <machine/smp.h> 64 65 TAILQ_HEAD(tslpque, thread); 66 67 static void sched_setup (void *dummy); 68 SYSINIT(sched_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, sched_setup, NULL) 69 70 int hogticks; 71 int lbolt; 72 void *lbolt_syncer; 73 int sched_quantum; /* Roundrobin scheduling quantum in ticks. */ 74 int ncpus; 75 int ncpus2, ncpus2_shift, ncpus2_mask; /* note: mask not cpumask_t */ 76 int ncpus_fit, ncpus_fit_mask; /* note: mask not cpumask_t */ 77 int safepri; 78 int tsleep_now_works; 79 int tsleep_crypto_dump = 0; 80 81 static struct callout loadav_callout; 82 static struct callout schedcpu_callout; 83 MALLOC_DEFINE(M_TSLEEP, "tslpque", "tsleep queues"); 84 85 #define __DEALL(ident) __DEQUALIFY(void *, ident) 86 87 #if !defined(KTR_TSLEEP) 88 #define KTR_TSLEEP KTR_ALL 89 #endif 90 KTR_INFO_MASTER(tsleep); 91 KTR_INFO(KTR_TSLEEP, tsleep, tsleep_beg, 0, "tsleep enter %p", const volatile void *ident); 92 KTR_INFO(KTR_TSLEEP, tsleep, tsleep_end, 1, "tsleep exit"); 93 KTR_INFO(KTR_TSLEEP, tsleep, wakeup_beg, 2, "wakeup enter %p", const volatile void *ident); 94 KTR_INFO(KTR_TSLEEP, tsleep, wakeup_end, 3, "wakeup exit"); 95 KTR_INFO(KTR_TSLEEP, tsleep, ilockfail, 4, "interlock failed %p", const volatile void *ident); 96 97 #define logtsleep1(name) KTR_LOG(tsleep_ ## name) 98 #define logtsleep2(name, val) KTR_LOG(tsleep_ ## name, val) 99 100 struct loadavg averunnable = 101 { {0, 0, 0}, FSCALE }; /* load average, of runnable procs */ 102 /* 103 * Constants for averages over 1, 5, and 15 minutes 104 * when sampling at 5 second intervals. 105 */ 106 static fixpt_t cexp[3] = { 107 0.9200444146293232 * FSCALE, /* exp(-1/12) */ 108 0.9834714538216174 * FSCALE, /* exp(-1/60) */ 109 0.9944598480048967 * FSCALE, /* exp(-1/180) */ 110 }; 111 112 static void endtsleep (void *); 113 static void loadav (void *arg); 114 static void schedcpu (void *arg); 115 116 /* 117 * Adjust the scheduler quantum. The quantum is specified in microseconds. 118 * Note that 'tick' is in microseconds per tick. 119 */ 120 static int 121 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS) 122 { 123 int error, new_val; 124 125 new_val = sched_quantum * ustick; 126 error = sysctl_handle_int(oidp, &new_val, 0, req); 127 if (error != 0 || req->newptr == NULL) 128 return (error); 129 if (new_val < ustick) 130 return (EINVAL); 131 sched_quantum = new_val / ustick; 132 hogticks = 2 * sched_quantum; 133 return (0); 134 } 135 136 SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW, 137 0, sizeof sched_quantum, sysctl_kern_quantum, "I", ""); 138 139 static int pctcpu_decay = 10; 140 SYSCTL_INT(_kern, OID_AUTO, pctcpu_decay, CTLFLAG_RW, &pctcpu_decay, 0, ""); 141 142 /* 143 * kernel uses `FSCALE', userland (SHOULD) use kern.fscale 144 */ 145 int fscale __unused = FSCALE; /* exported to systat */ 146 SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, ""); 147 148 /* 149 * Recompute process priorities, once a second. 150 * 151 * Since the userland schedulers are typically event oriented, if the 152 * estcpu calculation at wakeup() time is not sufficient to make a 153 * process runnable relative to other processes in the system we have 154 * a 1-second recalc to help out. 155 * 156 * This code also allows us to store sysclock_t data in the process structure 157 * without fear of an overrun, since sysclock_t are guarenteed to hold 158 * several seconds worth of count. 159 * 160 * WARNING! callouts can preempt normal threads. However, they will not 161 * preempt a thread holding a spinlock so we *can* safely use spinlocks. 162 */ 163 static int schedcpu_stats(struct proc *p, void *data __unused); 164 static int schedcpu_resource(struct proc *p, void *data __unused); 165 166 static void 167 schedcpu(void *arg) 168 { 169 allproc_scan(schedcpu_stats, NULL); 170 allproc_scan(schedcpu_resource, NULL); 171 wakeup((caddr_t)&lbolt); 172 wakeup(lbolt_syncer); 173 callout_reset(&schedcpu_callout, hz, schedcpu, NULL); 174 } 175 176 /* 177 * General process statistics once a second 178 */ 179 static int 180 schedcpu_stats(struct proc *p, void *data __unused) 181 { 182 struct lwp *lp; 183 184 /* 185 * Threads may not be completely set up if process in SIDL state. 186 */ 187 if (p->p_stat == SIDL) 188 return(0); 189 190 PHOLD(p); 191 if (lwkt_trytoken(&p->p_token) == FALSE) { 192 PRELE(p); 193 return(0); 194 } 195 196 p->p_swtime++; 197 FOREACH_LWP_IN_PROC(lp, p) { 198 if (lp->lwp_stat == LSSLEEP) { 199 ++lp->lwp_slptime; 200 if (lp->lwp_slptime == 1) 201 p->p_usched->uload_update(lp); 202 } 203 204 /* 205 * Only recalculate processes that are active or have slept 206 * less then 2 seconds. The schedulers understand this. 207 * Otherwise decay by 50% per second. 208 */ 209 if (lp->lwp_slptime <= 1) { 210 p->p_usched->recalculate(lp); 211 } else { 212 int decay; 213 214 decay = pctcpu_decay; 215 cpu_ccfence(); 216 if (decay <= 1) 217 decay = 1; 218 if (decay > 100) 219 decay = 100; 220 lp->lwp_pctcpu = (lp->lwp_pctcpu * (decay - 1)) / decay; 221 } 222 } 223 lwkt_reltoken(&p->p_token); 224 lwkt_yield(); 225 PRELE(p); 226 return(0); 227 } 228 229 /* 230 * Resource checks. XXX break out since ksignal/killproc can block, 231 * limiting us to one process killed per second. There is probably 232 * a better way. 233 */ 234 static int 235 schedcpu_resource(struct proc *p, void *data __unused) 236 { 237 u_int64_t ttime; 238 struct lwp *lp; 239 240 if (p->p_stat == SIDL) 241 return(0); 242 243 PHOLD(p); 244 if (lwkt_trytoken(&p->p_token) == FALSE) { 245 PRELE(p); 246 return(0); 247 } 248 249 if (p->p_stat == SZOMB || p->p_limit == NULL) { 250 lwkt_reltoken(&p->p_token); 251 PRELE(p); 252 return(0); 253 } 254 255 ttime = 0; 256 FOREACH_LWP_IN_PROC(lp, p) { 257 /* 258 * We may have caught an lp in the middle of being 259 * created, lwp_thread can be NULL. 260 */ 261 if (lp->lwp_thread) { 262 ttime += lp->lwp_thread->td_sticks; 263 ttime += lp->lwp_thread->td_uticks; 264 } 265 } 266 267 switch(plimit_testcpulimit(p->p_limit, ttime)) { 268 case PLIMIT_TESTCPU_KILL: 269 killproc(p, "exceeded maximum CPU limit"); 270 break; 271 case PLIMIT_TESTCPU_XCPU: 272 if ((p->p_flags & P_XCPU) == 0) { 273 p->p_flags |= P_XCPU; 274 ksignal(p, SIGXCPU); 275 } 276 break; 277 default: 278 break; 279 } 280 lwkt_reltoken(&p->p_token); 281 lwkt_yield(); 282 PRELE(p); 283 return(0); 284 } 285 286 /* 287 * This is only used by ps. Generate a cpu percentage use over 288 * a period of one second. 289 */ 290 void 291 updatepcpu(struct lwp *lp, int cpticks, int ttlticks) 292 { 293 fixpt_t acc; 294 int remticks; 295 296 acc = (cpticks << FSHIFT) / ttlticks; 297 if (ttlticks >= ESTCPUFREQ) { 298 lp->lwp_pctcpu = acc; 299 } else { 300 remticks = ESTCPUFREQ - ttlticks; 301 lp->lwp_pctcpu = (acc * ttlticks + lp->lwp_pctcpu * remticks) / 302 ESTCPUFREQ; 303 } 304 } 305 306 /* 307 * tsleep/wakeup hash table parameters. Try to find the sweet spot for 308 * like addresses being slept on. 309 */ 310 #define TABLESIZE 4001 311 #define LOOKUP(x) (((u_int)(uintptr_t)(x)) % TABLESIZE) 312 313 static cpumask_t slpque_cpumasks[TABLESIZE]; 314 315 /* 316 * General scheduler initialization. We force a reschedule 25 times 317 * a second by default. Note that cpu0 is initialized in early boot and 318 * cannot make any high level calls. 319 * 320 * Each cpu has its own sleep queue. 321 */ 322 void 323 sleep_gdinit(globaldata_t gd) 324 { 325 static struct tslpque slpque_cpu0[TABLESIZE]; 326 int i; 327 328 if (gd->gd_cpuid == 0) { 329 sched_quantum = (hz + 24) / 25; 330 hogticks = 2 * sched_quantum; 331 332 gd->gd_tsleep_hash = slpque_cpu0; 333 } else { 334 gd->gd_tsleep_hash = kmalloc(sizeof(slpque_cpu0), 335 M_TSLEEP, M_WAITOK | M_ZERO); 336 } 337 for (i = 0; i < TABLESIZE; ++i) 338 TAILQ_INIT(&gd->gd_tsleep_hash[i]); 339 } 340 341 /* 342 * This is a dandy function that allows us to interlock tsleep/wakeup 343 * operations with unspecified upper level locks, such as lockmgr locks, 344 * simply by holding a critical section. The sequence is: 345 * 346 * (acquire upper level lock) 347 * tsleep_interlock(blah) 348 * (release upper level lock) 349 * tsleep(blah, ...) 350 * 351 * Basically this functions queues us on the tsleep queue without actually 352 * descheduling us. When tsleep() is later called with PINTERLOCK it 353 * assumes the thread was already queued, otherwise it queues it there. 354 * 355 * Thus it is possible to receive the wakeup prior to going to sleep and 356 * the race conditions are covered. 357 */ 358 static __inline void 359 _tsleep_interlock(globaldata_t gd, const volatile void *ident, int flags) 360 { 361 thread_t td = gd->gd_curthread; 362 int id; 363 364 crit_enter_quick(td); 365 if (td->td_flags & TDF_TSLEEPQ) { 366 id = LOOKUP(td->td_wchan); 367 TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_sleepq); 368 if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL) { 369 ATOMIC_CPUMASK_NANDBIT(slpque_cpumasks[id], 370 gd->gd_cpuid); 371 } 372 } else { 373 td->td_flags |= TDF_TSLEEPQ; 374 } 375 id = LOOKUP(ident); 376 TAILQ_INSERT_TAIL(&gd->gd_tsleep_hash[id], td, td_sleepq); 377 ATOMIC_CPUMASK_ORBIT(slpque_cpumasks[id], gd->gd_cpuid); 378 td->td_wchan = ident; 379 td->td_wdomain = flags & PDOMAIN_MASK; 380 crit_exit_quick(td); 381 } 382 383 void 384 tsleep_interlock(const volatile void *ident, int flags) 385 { 386 _tsleep_interlock(mycpu, ident, flags); 387 } 388 389 /* 390 * Remove thread from sleepq. Must be called with a critical section held. 391 * The thread must not be migrating. 392 */ 393 static __inline void 394 _tsleep_remove(thread_t td) 395 { 396 globaldata_t gd = mycpu; 397 int id; 398 399 KKASSERT(td->td_gd == gd && IN_CRITICAL_SECT(td)); 400 KKASSERT((td->td_flags & TDF_MIGRATING) == 0); 401 if (td->td_flags & TDF_TSLEEPQ) { 402 td->td_flags &= ~TDF_TSLEEPQ; 403 id = LOOKUP(td->td_wchan); 404 TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_sleepq); 405 if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL) { 406 ATOMIC_CPUMASK_NANDBIT(slpque_cpumasks[id], 407 gd->gd_cpuid); 408 } 409 td->td_wchan = NULL; 410 td->td_wdomain = 0; 411 } 412 } 413 414 void 415 tsleep_remove(thread_t td) 416 { 417 _tsleep_remove(td); 418 } 419 420 /* 421 * General sleep call. Suspends the current process until a wakeup is 422 * performed on the specified identifier. The process will then be made 423 * runnable with the specified priority. Sleeps at most timo/hz seconds 424 * (0 means no timeout). If flags includes PCATCH flag, signals are checked 425 * before and after sleeping, else signals are not checked. Returns 0 if 426 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a 427 * signal needs to be delivered, ERESTART is returned if the current system 428 * call should be restarted if possible, and EINTR is returned if the system 429 * call should be interrupted by the signal (return EINTR). 430 * 431 * Note that if we are a process, we release_curproc() before messing with 432 * the LWKT scheduler. 433 * 434 * During autoconfiguration or after a panic, a sleep will simply 435 * lower the priority briefly to allow interrupts, then return. 436 * 437 * WARNING! This code can't block (short of switching away), or bad things 438 * will happen. No getting tokens, no blocking locks, etc. 439 */ 440 int 441 tsleep(const volatile void *ident, int flags, const char *wmesg, int timo) 442 { 443 struct thread *td = curthread; 444 struct lwp *lp = td->td_lwp; 445 struct proc *p = td->td_proc; /* may be NULL */ 446 globaldata_t gd; 447 int sig; 448 int catch; 449 int error; 450 int oldpri; 451 struct callout thandle; 452 453 /* 454 * Currently a severe hack. Make sure any delayed wakeups 455 * are flushed before we sleep or we might deadlock on whatever 456 * event we are sleeping on. 457 */ 458 if (td->td_flags & TDF_DELAYED_WAKEUP) 459 wakeup_end_delayed(); 460 461 /* 462 * NOTE: removed KTRPOINT, it could cause races due to blocking 463 * even in stable. Just scrap it for now. 464 */ 465 if (!tsleep_crypto_dump && (tsleep_now_works == 0 || panicstr)) { 466 /* 467 * After a panic, or before we actually have an operational 468 * softclock, just give interrupts a chance, then just return; 469 * 470 * don't run any other procs or panic below, 471 * in case this is the idle process and already asleep. 472 */ 473 splz(); 474 oldpri = td->td_pri; 475 lwkt_setpri_self(safepri); 476 lwkt_switch(); 477 lwkt_setpri_self(oldpri); 478 return (0); 479 } 480 logtsleep2(tsleep_beg, ident); 481 gd = td->td_gd; 482 KKASSERT(td != &gd->gd_idlethread); /* you must be kidding! */ 483 td->td_wakefromcpu = -1; /* overwritten by _wakeup */ 484 485 /* 486 * NOTE: all of this occurs on the current cpu, including any 487 * callout-based wakeups, so a critical section is a sufficient 488 * interlock. 489 * 490 * The entire sequence through to where we actually sleep must 491 * run without breaking the critical section. 492 */ 493 catch = flags & PCATCH; 494 error = 0; 495 sig = 0; 496 497 crit_enter_quick(td); 498 499 KASSERT(ident != NULL, ("tsleep: no ident")); 500 KASSERT(lp == NULL || 501 lp->lwp_stat == LSRUN || /* Obvious */ 502 lp->lwp_stat == LSSTOP, /* Set in tstop */ 503 ("tsleep %p %s %d", 504 ident, wmesg, lp->lwp_stat)); 505 506 /* 507 * We interlock the sleep queue if the caller has not already done 508 * it for us. This must be done before we potentially acquire any 509 * tokens or we can loose the wakeup. 510 */ 511 if ((flags & PINTERLOCKED) == 0) { 512 _tsleep_interlock(gd, ident, flags); 513 } 514 515 /* 516 * Setup for the current process (if this is a process). We must 517 * interlock with lwp_token to avoid remote wakeup races via 518 * setrunnable() 519 */ 520 if (lp) { 521 lwkt_gettoken(&lp->lwp_token); 522 if (catch) { 523 /* 524 * Early termination if PCATCH was set and a 525 * signal is pending, interlocked with the 526 * critical section. 527 * 528 * Early termination only occurs when tsleep() is 529 * entered while in a normal LSRUN state. 530 */ 531 if ((sig = CURSIG(lp)) != 0) 532 goto resume; 533 534 /* 535 * Causes ksignal to wake us up if a signal is 536 * received (interlocked with p->p_token). 537 */ 538 lp->lwp_flags |= LWP_SINTR; 539 } 540 } else { 541 KKASSERT(p == NULL); 542 } 543 544 /* 545 * Make sure the current process has been untangled from 546 * the userland scheduler and initialize slptime to start 547 * counting. 548 * 549 * NOTE: td->td_wakefromcpu is pre-set by the release function 550 * for the dfly scheduler, and then adjusted by _wakeup() 551 */ 552 if (lp) { 553 p->p_usched->release_curproc(lp); 554 lp->lwp_slptime = 0; 555 } 556 557 /* 558 * If the interlocked flag is set but our cpu bit in the slpqueue 559 * is no longer set, then a wakeup was processed inbetween the 560 * tsleep_interlock() (ours or the callers), and here. This can 561 * occur under numerous circumstances including when we release the 562 * current process. 563 * 564 * Extreme loads can cause the sending of an IPI (e.g. wakeup()'s) 565 * to process incoming IPIs, thus draining incoming wakeups. 566 */ 567 if ((td->td_flags & TDF_TSLEEPQ) == 0) { 568 logtsleep2(ilockfail, ident); 569 goto resume; 570 } 571 572 /* 573 * scheduling is blocked while in a critical section. Coincide 574 * the descheduled-by-tsleep flag with the descheduling of the 575 * lwkt. 576 * 577 * The timer callout is localized on our cpu and interlocked by 578 * our critical section. 579 */ 580 lwkt_deschedule_self(td); 581 td->td_flags |= TDF_TSLEEP_DESCHEDULED; 582 td->td_wmesg = wmesg; 583 584 /* 585 * Setup the timeout, if any. The timeout is only operable while 586 * the thread is flagged descheduled. 587 */ 588 KKASSERT((td->td_flags & TDF_TIMEOUT) == 0); 589 if (timo) { 590 callout_init_mp(&thandle); 591 callout_reset(&thandle, timo, endtsleep, td); 592 } 593 594 /* 595 * Beddy bye bye. 596 */ 597 if (lp) { 598 /* 599 * Ok, we are sleeping. Place us in the SSLEEP state. 600 */ 601 KKASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) == 0); 602 603 /* 604 * tstop() sets LSSTOP, so don't fiddle with that. 605 */ 606 if (lp->lwp_stat != LSSTOP) 607 lp->lwp_stat = LSSLEEP; 608 lp->lwp_ru.ru_nvcsw++; 609 p->p_usched->uload_update(lp); 610 lwkt_switch(); 611 612 /* 613 * And when we are woken up, put us back in LSRUN. If we 614 * slept for over a second, recalculate our estcpu. 615 */ 616 lp->lwp_stat = LSRUN; 617 if (lp->lwp_slptime) { 618 p->p_usched->uload_update(lp); 619 p->p_usched->recalculate(lp); 620 } 621 lp->lwp_slptime = 0; 622 } else { 623 lwkt_switch(); 624 } 625 626 /* 627 * Make sure we haven't switched cpus while we were asleep. It's 628 * not supposed to happen. Cleanup our temporary flags. 629 */ 630 KKASSERT(gd == td->td_gd); 631 632 /* 633 * Cleanup the timeout. If the timeout has already occured thandle 634 * has already been stopped, otherwise stop thandle. If the timeout 635 * is running (the callout thread must be blocked trying to get 636 * lwp_token) then wait for us to get scheduled. 637 */ 638 if (timo) { 639 while (td->td_flags & TDF_TIMEOUT_RUNNING) { 640 lwkt_deschedule_self(td); 641 td->td_wmesg = "tsrace"; 642 lwkt_switch(); 643 kprintf("td %p %s: timeout race\n", td, td->td_comm); 644 } 645 if (td->td_flags & TDF_TIMEOUT) { 646 td->td_flags &= ~TDF_TIMEOUT; 647 error = EWOULDBLOCK; 648 } else { 649 /* does not block when on same cpu */ 650 callout_stop(&thandle); 651 } 652 } 653 td->td_flags &= ~TDF_TSLEEP_DESCHEDULED; 654 655 /* 656 * Make sure we have been removed from the sleepq. In most 657 * cases this will have been done for us already but it is 658 * possible for a scheduling IPI to be in-flight from a 659 * previous tsleep/tsleep_interlock() or due to a straight-out 660 * call to lwkt_schedule() (in the case of an interrupt thread), 661 * causing a spurious wakeup. 662 */ 663 _tsleep_remove(td); 664 td->td_wmesg = NULL; 665 666 /* 667 * Figure out the correct error return. If interrupted by a 668 * signal we want to return EINTR or ERESTART. 669 */ 670 resume: 671 if (lp) { 672 if (catch && error == 0) { 673 if (sig != 0 || (sig = CURSIG(lp))) { 674 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 675 error = EINTR; 676 else 677 error = ERESTART; 678 } 679 } 680 lp->lwp_flags &= ~LWP_SINTR; 681 lwkt_reltoken(&lp->lwp_token); 682 } 683 logtsleep1(tsleep_end); 684 crit_exit_quick(td); 685 return (error); 686 } 687 688 /* 689 * Interlocked spinlock sleep. An exclusively held spinlock must 690 * be passed to ssleep(). The function will atomically release the 691 * spinlock and tsleep on the ident, then reacquire the spinlock and 692 * return. 693 * 694 * This routine is fairly important along the critical path, so optimize it 695 * heavily. 696 */ 697 int 698 ssleep(const volatile void *ident, struct spinlock *spin, int flags, 699 const char *wmesg, int timo) 700 { 701 globaldata_t gd = mycpu; 702 int error; 703 704 _tsleep_interlock(gd, ident, flags); 705 spin_unlock_quick(gd, spin); 706 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 707 _spin_lock_quick(gd, spin, wmesg); 708 709 return (error); 710 } 711 712 int 713 lksleep(const volatile void *ident, struct lock *lock, int flags, 714 const char *wmesg, int timo) 715 { 716 globaldata_t gd = mycpu; 717 int error; 718 719 _tsleep_interlock(gd, ident, flags); 720 lockmgr(lock, LK_RELEASE); 721 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 722 lockmgr(lock, LK_EXCLUSIVE); 723 724 return (error); 725 } 726 727 /* 728 * Interlocked mutex sleep. An exclusively held mutex must be passed 729 * to mtxsleep(). The function will atomically release the mutex 730 * and tsleep on the ident, then reacquire the mutex and return. 731 */ 732 int 733 mtxsleep(const volatile void *ident, struct mtx *mtx, int flags, 734 const char *wmesg, int timo) 735 { 736 globaldata_t gd = mycpu; 737 int error; 738 739 _tsleep_interlock(gd, ident, flags); 740 mtx_unlock(mtx); 741 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 742 mtx_lock_ex_quick(mtx, wmesg); 743 744 return (error); 745 } 746 747 /* 748 * Interlocked serializer sleep. An exclusively held serializer must 749 * be passed to zsleep(). The function will atomically release 750 * the serializer and tsleep on the ident, then reacquire the serializer 751 * and return. 752 */ 753 int 754 zsleep(const volatile void *ident, struct lwkt_serialize *slz, int flags, 755 const char *wmesg, int timo) 756 { 757 globaldata_t gd = mycpu; 758 int ret; 759 760 ASSERT_SERIALIZED(slz); 761 762 _tsleep_interlock(gd, ident, flags); 763 lwkt_serialize_exit(slz); 764 ret = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 765 lwkt_serialize_enter(slz); 766 767 return ret; 768 } 769 770 /* 771 * Directly block on the LWKT thread by descheduling it. This 772 * is much faster then tsleep(), but the only legal way to wake 773 * us up is to directly schedule the thread. 774 * 775 * Setting TDF_SINTR will cause new signals to directly schedule us. 776 * 777 * This routine must be called while in a critical section. 778 */ 779 int 780 lwkt_sleep(const char *wmesg, int flags) 781 { 782 thread_t td = curthread; 783 int sig; 784 785 if ((flags & PCATCH) == 0 || td->td_lwp == NULL) { 786 td->td_flags |= TDF_BLOCKED; 787 td->td_wmesg = wmesg; 788 lwkt_deschedule_self(td); 789 lwkt_switch(); 790 td->td_wmesg = NULL; 791 td->td_flags &= ~TDF_BLOCKED; 792 return(0); 793 } 794 if ((sig = CURSIG(td->td_lwp)) != 0) { 795 if (SIGISMEMBER(td->td_proc->p_sigacts->ps_sigintr, sig)) 796 return(EINTR); 797 else 798 return(ERESTART); 799 800 } 801 td->td_flags |= TDF_BLOCKED | TDF_SINTR; 802 td->td_wmesg = wmesg; 803 lwkt_deschedule_self(td); 804 lwkt_switch(); 805 td->td_flags &= ~(TDF_BLOCKED | TDF_SINTR); 806 td->td_wmesg = NULL; 807 return(0); 808 } 809 810 /* 811 * Implement the timeout for tsleep. 812 * 813 * This type of callout timeout is scheduled on the same cpu the process 814 * is sleeping on. Also, at the moment, the MP lock is held. 815 */ 816 static void 817 endtsleep(void *arg) 818 { 819 thread_t td = arg; 820 struct lwp *lp; 821 822 /* 823 * We are going to have to get the lwp_token, which means we might 824 * block. This can race a tsleep getting woken up by other means 825 * so set TDF_TIMEOUT_RUNNING to force the tsleep to wait for our 826 * processing to complete (sorry tsleep!). 827 * 828 * We can safely set td_flags because td MUST be on the same cpu 829 * as we are. 830 */ 831 KKASSERT(td->td_gd == mycpu); 832 crit_enter(); 833 td->td_flags |= TDF_TIMEOUT_RUNNING | TDF_TIMEOUT; 834 835 /* 836 * This can block but TDF_TIMEOUT_RUNNING will prevent the thread 837 * from exiting the tsleep on us. The flag is interlocked by virtue 838 * of lp being on the same cpu as we are. 839 */ 840 if ((lp = td->td_lwp) != NULL) 841 lwkt_gettoken(&lp->lwp_token); 842 843 KKASSERT(td->td_flags & TDF_TSLEEP_DESCHEDULED); 844 845 if (lp) { 846 /* 847 * callout timer should never be set in tstop() because 848 * it passes a timeout of 0. 849 */ 850 KKASSERT(lp->lwp_stat != LSSTOP); 851 setrunnable(lp); 852 lwkt_reltoken(&lp->lwp_token); 853 } else { 854 _tsleep_remove(td); 855 lwkt_schedule(td); 856 } 857 KKASSERT(td->td_gd == mycpu); 858 td->td_flags &= ~TDF_TIMEOUT_RUNNING; 859 crit_exit(); 860 } 861 862 /* 863 * Make all processes sleeping on the specified identifier runnable. 864 * count may be zero or one only. 865 * 866 * The domain encodes the sleep/wakeup domain, flags, plus the originating 867 * cpu. 868 * 869 * This call may run without the MP lock held. We can only manipulate thread 870 * state on the cpu owning the thread. We CANNOT manipulate process state 871 * at all. 872 * 873 * _wakeup() can be passed to an IPI so we can't use (const volatile 874 * void *ident). 875 */ 876 static void 877 _wakeup(void *ident, int domain) 878 { 879 struct tslpque *qp; 880 struct thread *td; 881 struct thread *ntd; 882 globaldata_t gd; 883 cpumask_t mask; 884 int id; 885 886 crit_enter(); 887 logtsleep2(wakeup_beg, ident); 888 gd = mycpu; 889 id = LOOKUP(ident); 890 qp = &gd->gd_tsleep_hash[id]; 891 restart: 892 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) { 893 ntd = TAILQ_NEXT(td, td_sleepq); 894 if (td->td_wchan == ident && 895 td->td_wdomain == (domain & PDOMAIN_MASK) 896 ) { 897 KKASSERT(td->td_gd == gd); 898 _tsleep_remove(td); 899 td->td_wakefromcpu = PWAKEUP_DECODE(domain); 900 if (td->td_flags & TDF_TSLEEP_DESCHEDULED) { 901 lwkt_schedule(td); 902 if (domain & PWAKEUP_ONE) 903 goto done; 904 } 905 goto restart; 906 } 907 } 908 909 /* 910 * We finished checking the current cpu but there still may be 911 * more work to do. Either wakeup_one was requested and no matching 912 * thread was found, or a normal wakeup was requested and we have 913 * to continue checking cpus. 914 * 915 * It should be noted that this scheme is actually less expensive then 916 * the old scheme when waking up multiple threads, since we send 917 * only one IPI message per target candidate which may then schedule 918 * multiple threads. Before we could have wound up sending an IPI 919 * message for each thread on the target cpu (!= current cpu) that 920 * needed to be woken up. 921 * 922 * NOTE: Wakeups occuring on remote cpus are asynchronous. This 923 * should be ok since we are passing idents in the IPI rather then 924 * thread pointers. 925 */ 926 if ((domain & PWAKEUP_MYCPU) == 0) { 927 mask = slpque_cpumasks[id]; 928 CPUMASK_ANDMASK(mask, gd->gd_other_cpus); 929 if (CPUMASK_TESTNZERO(mask)) { 930 lwkt_send_ipiq2_mask(mask, _wakeup, ident, 931 domain | PWAKEUP_MYCPU); 932 } 933 } 934 done: 935 logtsleep1(wakeup_end); 936 crit_exit(); 937 } 938 939 /* 940 * Wakeup all threads tsleep()ing on the specified ident, on all cpus 941 */ 942 void 943 wakeup(const volatile void *ident) 944 { 945 globaldata_t gd = mycpu; 946 thread_t td = gd->gd_curthread; 947 948 if (td && (td->td_flags & TDF_DELAYED_WAKEUP)) { 949 /* 950 * If we are in a delayed wakeup section, record up to two wakeups in 951 * a per-CPU queue and issue them when we block or exit the delayed 952 * wakeup section. 953 */ 954 if (atomic_cmpset_ptr(&gd->gd_delayed_wakeup[0], NULL, ident)) 955 return; 956 if (atomic_cmpset_ptr(&gd->gd_delayed_wakeup[1], NULL, ident)) 957 return; 958 959 ident = atomic_swap_ptr(__DEQUALIFY(volatile void **, &gd->gd_delayed_wakeup[1]), 960 __DEALL(ident)); 961 ident = atomic_swap_ptr(__DEQUALIFY(volatile void **, &gd->gd_delayed_wakeup[0]), 962 __DEALL(ident)); 963 } 964 965 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, gd->gd_cpuid)); 966 } 967 968 /* 969 * Wakeup one thread tsleep()ing on the specified ident, on any cpu. 970 */ 971 void 972 wakeup_one(const volatile void *ident) 973 { 974 /* XXX potentially round-robin the first responding cpu */ 975 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mycpu->gd_cpuid) | 976 PWAKEUP_ONE); 977 } 978 979 /* 980 * Wakeup threads tsleep()ing on the specified ident on the current cpu 981 * only. 982 */ 983 void 984 wakeup_mycpu(const volatile void *ident) 985 { 986 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mycpu->gd_cpuid) | 987 PWAKEUP_MYCPU); 988 } 989 990 /* 991 * Wakeup one thread tsleep()ing on the specified ident on the current cpu 992 * only. 993 */ 994 void 995 wakeup_mycpu_one(const volatile void *ident) 996 { 997 /* XXX potentially round-robin the first responding cpu */ 998 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mycpu->gd_cpuid) | 999 PWAKEUP_MYCPU | PWAKEUP_ONE); 1000 } 1001 1002 /* 1003 * Wakeup all thread tsleep()ing on the specified ident on the specified cpu 1004 * only. 1005 */ 1006 void 1007 wakeup_oncpu(globaldata_t gd, const volatile void *ident) 1008 { 1009 globaldata_t mygd = mycpu; 1010 if (gd == mycpu) { 1011 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mygd->gd_cpuid) | 1012 PWAKEUP_MYCPU); 1013 } else { 1014 lwkt_send_ipiq2(gd, _wakeup, __DEALL(ident), 1015 PWAKEUP_ENCODE(0, mygd->gd_cpuid) | 1016 PWAKEUP_MYCPU); 1017 } 1018 } 1019 1020 /* 1021 * Wakeup one thread tsleep()ing on the specified ident on the specified cpu 1022 * only. 1023 */ 1024 void 1025 wakeup_oncpu_one(globaldata_t gd, const volatile void *ident) 1026 { 1027 globaldata_t mygd = mycpu; 1028 if (gd == mygd) { 1029 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mygd->gd_cpuid) | 1030 PWAKEUP_MYCPU | PWAKEUP_ONE); 1031 } else { 1032 lwkt_send_ipiq2(gd, _wakeup, __DEALL(ident), 1033 PWAKEUP_ENCODE(0, mygd->gd_cpuid) | 1034 PWAKEUP_MYCPU | PWAKEUP_ONE); 1035 } 1036 } 1037 1038 /* 1039 * Wakeup all threads waiting on the specified ident that slept using 1040 * the specified domain, on all cpus. 1041 */ 1042 void 1043 wakeup_domain(const volatile void *ident, int domain) 1044 { 1045 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(domain, mycpu->gd_cpuid)); 1046 } 1047 1048 /* 1049 * Wakeup one thread waiting on the specified ident that slept using 1050 * the specified domain, on any cpu. 1051 */ 1052 void 1053 wakeup_domain_one(const volatile void *ident, int domain) 1054 { 1055 /* XXX potentially round-robin the first responding cpu */ 1056 _wakeup(__DEALL(ident), 1057 PWAKEUP_ENCODE(domain, mycpu->gd_cpuid) | PWAKEUP_ONE); 1058 } 1059 1060 void 1061 wakeup_start_delayed(void) 1062 { 1063 globaldata_t gd = mycpu; 1064 1065 crit_enter(); 1066 gd->gd_curthread->td_flags |= TDF_DELAYED_WAKEUP; 1067 crit_exit(); 1068 } 1069 1070 void 1071 wakeup_end_delayed(void) 1072 { 1073 globaldata_t gd = mycpu; 1074 1075 if (gd->gd_curthread->td_flags & TDF_DELAYED_WAKEUP) { 1076 crit_enter(); 1077 gd->gd_curthread->td_flags &= ~TDF_DELAYED_WAKEUP; 1078 if (gd->gd_delayed_wakeup[0] || gd->gd_delayed_wakeup[1]) { 1079 if (gd->gd_delayed_wakeup[0]) { 1080 wakeup(gd->gd_delayed_wakeup[0]); 1081 gd->gd_delayed_wakeup[0] = NULL; 1082 } 1083 if (gd->gd_delayed_wakeup[1]) { 1084 wakeup(gd->gd_delayed_wakeup[1]); 1085 gd->gd_delayed_wakeup[1] = NULL; 1086 } 1087 } 1088 crit_exit(); 1089 } 1090 } 1091 1092 /* 1093 * setrunnable() 1094 * 1095 * Make a process runnable. lp->lwp_token must be held on call and this 1096 * function must be called from the cpu owning lp. 1097 * 1098 * This only has an effect if we are in LSSTOP or LSSLEEP. 1099 */ 1100 void 1101 setrunnable(struct lwp *lp) 1102 { 1103 thread_t td = lp->lwp_thread; 1104 1105 ASSERT_LWKT_TOKEN_HELD(&lp->lwp_token); 1106 KKASSERT(td->td_gd == mycpu); 1107 crit_enter(); 1108 if (lp->lwp_stat == LSSTOP) 1109 lp->lwp_stat = LSSLEEP; 1110 if (lp->lwp_stat == LSSLEEP) { 1111 _tsleep_remove(td); 1112 lwkt_schedule(td); 1113 } else if (td->td_flags & TDF_SINTR) { 1114 lwkt_schedule(td); 1115 } 1116 crit_exit(); 1117 } 1118 1119 /* 1120 * The process is stopped due to some condition, usually because p_stat is 1121 * set to SSTOP, but also possibly due to being traced. 1122 * 1123 * Caller must hold p->p_token 1124 * 1125 * NOTE! If the caller sets SSTOP, the caller must also clear P_WAITED 1126 * because the parent may check the child's status before the child actually 1127 * gets to this routine. 1128 * 1129 * This routine is called with the current lwp only, typically just 1130 * before returning to userland if the process state is detected as 1131 * possibly being in a stopped state. 1132 */ 1133 void 1134 tstop(void) 1135 { 1136 struct lwp *lp = curthread->td_lwp; 1137 struct proc *p = lp->lwp_proc; 1138 struct proc *q; 1139 1140 lwkt_gettoken(&lp->lwp_token); 1141 crit_enter(); 1142 1143 /* 1144 * If LWP_MP_WSTOP is set, we were sleeping 1145 * while our process was stopped. At this point 1146 * we were already counted as stopped. 1147 */ 1148 if ((lp->lwp_mpflags & LWP_MP_WSTOP) == 0) { 1149 /* 1150 * If we're the last thread to stop, signal 1151 * our parent. 1152 */ 1153 p->p_nstopped++; 1154 atomic_set_int(&lp->lwp_mpflags, LWP_MP_WSTOP); 1155 wakeup(&p->p_nstopped); 1156 if (p->p_nstopped == p->p_nthreads) { 1157 /* 1158 * Token required to interlock kern_wait() 1159 */ 1160 q = p->p_pptr; 1161 PHOLD(q); 1162 lwkt_gettoken(&q->p_token); 1163 p->p_flags &= ~P_WAITED; 1164 wakeup(p->p_pptr); 1165 if ((q->p_sigacts->ps_flag & PS_NOCLDSTOP) == 0) 1166 ksignal(q, SIGCHLD); 1167 lwkt_reltoken(&q->p_token); 1168 PRELE(q); 1169 } 1170 } 1171 while (p->p_stat == SSTOP) { 1172 lp->lwp_stat = LSSTOP; 1173 tsleep(p, 0, "stop", 0); 1174 } 1175 p->p_nstopped--; 1176 atomic_clear_int(&lp->lwp_mpflags, LWP_MP_WSTOP); 1177 crit_exit(); 1178 lwkt_reltoken(&lp->lwp_token); 1179 } 1180 1181 /* 1182 * Compute a tenex style load average of a quantity on 1183 * 1, 5 and 15 minute intervals. 1184 */ 1185 static int loadav_count_runnable(struct lwp *p, void *data); 1186 1187 static void 1188 loadav(void *arg) 1189 { 1190 struct loadavg *avg; 1191 int i, nrun; 1192 1193 nrun = 0; 1194 alllwp_scan(loadav_count_runnable, &nrun); 1195 avg = &averunnable; 1196 for (i = 0; i < 3; i++) { 1197 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] + 1198 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT; 1199 } 1200 1201 /* 1202 * Schedule the next update to occur after 5 seconds, but add a 1203 * random variation to avoid synchronisation with processes that 1204 * run at regular intervals. 1205 */ 1206 callout_reset(&loadav_callout, hz * 4 + (int)(krandom() % (hz * 2 + 1)), 1207 loadav, NULL); 1208 } 1209 1210 static int 1211 loadav_count_runnable(struct lwp *lp, void *data) 1212 { 1213 int *nrunp = data; 1214 thread_t td; 1215 1216 switch (lp->lwp_stat) { 1217 case LSRUN: 1218 if ((td = lp->lwp_thread) == NULL) 1219 break; 1220 if (td->td_flags & TDF_BLOCKED) 1221 break; 1222 ++*nrunp; 1223 break; 1224 default: 1225 break; 1226 } 1227 lwkt_yield(); 1228 return(0); 1229 } 1230 1231 /* ARGSUSED */ 1232 static void 1233 sched_setup(void *dummy) 1234 { 1235 callout_init_mp(&loadav_callout); 1236 callout_init_mp(&schedcpu_callout); 1237 1238 /* Kick off timeout driven events by calling first time. */ 1239 schedcpu(NULL); 1240 loadav(NULL); 1241 } 1242 1243