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_clear_cpumask(&slpque_cpumasks[id], 370 gd->gd_cpumask); 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_set_cpumask(&slpque_cpumasks[id], gd->gd_cpumask); 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_clear_cpumask(&slpque_cpumasks[id], gd->gd_cpumask); 407 td->td_wchan = NULL; 408 td->td_wdomain = 0; 409 } 410 } 411 412 void 413 tsleep_remove(thread_t td) 414 { 415 _tsleep_remove(td); 416 } 417 418 /* 419 * General sleep call. Suspends the current process until a wakeup is 420 * performed on the specified identifier. The process will then be made 421 * runnable with the specified priority. Sleeps at most timo/hz seconds 422 * (0 means no timeout). If flags includes PCATCH flag, signals are checked 423 * before and after sleeping, else signals are not checked. Returns 0 if 424 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a 425 * signal needs to be delivered, ERESTART is returned if the current system 426 * call should be restarted if possible, and EINTR is returned if the system 427 * call should be interrupted by the signal (return EINTR). 428 * 429 * Note that if we are a process, we release_curproc() before messing with 430 * the LWKT scheduler. 431 * 432 * During autoconfiguration or after a panic, a sleep will simply 433 * lower the priority briefly to allow interrupts, then return. 434 * 435 * WARNING! This code can't block (short of switching away), or bad things 436 * will happen. No getting tokens, no blocking locks, etc. 437 */ 438 int 439 tsleep(const volatile void *ident, int flags, const char *wmesg, int timo) 440 { 441 struct thread *td = curthread; 442 struct lwp *lp = td->td_lwp; 443 struct proc *p = td->td_proc; /* may be NULL */ 444 globaldata_t gd; 445 int sig; 446 int catch; 447 int error; 448 int oldpri; 449 struct callout thandle; 450 451 /* 452 * Currently a severe hack. Make sure any delayed wakeups 453 * are flushed before we sleep or we might deadlock on whatever 454 * event we are sleeping on. 455 */ 456 if (td->td_flags & TDF_DELAYED_WAKEUP) 457 wakeup_end_delayed(); 458 459 /* 460 * NOTE: removed KTRPOINT, it could cause races due to blocking 461 * even in stable. Just scrap it for now. 462 */ 463 if (!tsleep_crypto_dump && (tsleep_now_works == 0 || panicstr)) { 464 /* 465 * After a panic, or before we actually have an operational 466 * softclock, just give interrupts a chance, then just return; 467 * 468 * don't run any other procs or panic below, 469 * in case this is the idle process and already asleep. 470 */ 471 splz(); 472 oldpri = td->td_pri; 473 lwkt_setpri_self(safepri); 474 lwkt_switch(); 475 lwkt_setpri_self(oldpri); 476 return (0); 477 } 478 logtsleep2(tsleep_beg, ident); 479 gd = td->td_gd; 480 KKASSERT(td != &gd->gd_idlethread); /* you must be kidding! */ 481 td->td_wakefromcpu = -1; /* overwritten by _wakeup */ 482 483 /* 484 * NOTE: all of this occurs on the current cpu, including any 485 * callout-based wakeups, so a critical section is a sufficient 486 * interlock. 487 * 488 * The entire sequence through to where we actually sleep must 489 * run without breaking the critical section. 490 */ 491 catch = flags & PCATCH; 492 error = 0; 493 sig = 0; 494 495 crit_enter_quick(td); 496 497 KASSERT(ident != NULL, ("tsleep: no ident")); 498 KASSERT(lp == NULL || 499 lp->lwp_stat == LSRUN || /* Obvious */ 500 lp->lwp_stat == LSSTOP, /* Set in tstop */ 501 ("tsleep %p %s %d", 502 ident, wmesg, lp->lwp_stat)); 503 504 /* 505 * We interlock the sleep queue if the caller has not already done 506 * it for us. This must be done before we potentially acquire any 507 * tokens or we can loose the wakeup. 508 */ 509 if ((flags & PINTERLOCKED) == 0) { 510 _tsleep_interlock(gd, ident, flags); 511 } 512 513 /* 514 * Setup for the current process (if this is a process). We must 515 * interlock with lwp_token to avoid remote wakeup races via 516 * setrunnable() 517 */ 518 if (lp) { 519 lwkt_gettoken(&lp->lwp_token); 520 if (catch) { 521 /* 522 * Early termination if PCATCH was set and a 523 * signal is pending, interlocked with the 524 * critical section. 525 * 526 * Early termination only occurs when tsleep() is 527 * entered while in a normal LSRUN state. 528 */ 529 if ((sig = CURSIG(lp)) != 0) 530 goto resume; 531 532 /* 533 * Causes ksignal to wake us up if a signal is 534 * received (interlocked with p->p_token). 535 */ 536 lp->lwp_flags |= LWP_SINTR; 537 } 538 } else { 539 KKASSERT(p == NULL); 540 } 541 542 /* 543 * Make sure the current process has been untangled from 544 * the userland scheduler and initialize slptime to start 545 * counting. 546 * 547 * NOTE: td->td_wakefromcpu is pre-set by the release function 548 * for the dfly scheduler, and then adjusted by _wakeup() 549 */ 550 if (lp) { 551 p->p_usched->release_curproc(lp); 552 lp->lwp_slptime = 0; 553 } 554 555 /* 556 * If the interlocked flag is set but our cpu bit in the slpqueue 557 * is no longer set, then a wakeup was processed inbetween the 558 * tsleep_interlock() (ours or the callers), and here. This can 559 * occur under numerous circumstances including when we release the 560 * current process. 561 * 562 * Extreme loads can cause the sending of an IPI (e.g. wakeup()'s) 563 * to process incoming IPIs, thus draining incoming wakeups. 564 */ 565 if ((td->td_flags & TDF_TSLEEPQ) == 0) { 566 logtsleep2(ilockfail, ident); 567 goto resume; 568 } 569 570 /* 571 * scheduling is blocked while in a critical section. Coincide 572 * the descheduled-by-tsleep flag with the descheduling of the 573 * lwkt. 574 * 575 * The timer callout is localized on our cpu and interlocked by 576 * our critical section. 577 */ 578 lwkt_deschedule_self(td); 579 td->td_flags |= TDF_TSLEEP_DESCHEDULED; 580 td->td_wmesg = wmesg; 581 582 /* 583 * Setup the timeout, if any. The timeout is only operable while 584 * the thread is flagged descheduled. 585 */ 586 KKASSERT((td->td_flags & TDF_TIMEOUT) == 0); 587 if (timo) { 588 callout_init_mp(&thandle); 589 callout_reset(&thandle, timo, endtsleep, td); 590 } 591 592 /* 593 * Beddy bye bye. 594 */ 595 if (lp) { 596 /* 597 * Ok, we are sleeping. Place us in the SSLEEP state. 598 */ 599 KKASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) == 0); 600 601 /* 602 * tstop() sets LSSTOP, so don't fiddle with that. 603 */ 604 if (lp->lwp_stat != LSSTOP) 605 lp->lwp_stat = LSSLEEP; 606 lp->lwp_ru.ru_nvcsw++; 607 p->p_usched->uload_update(lp); 608 lwkt_switch(); 609 610 /* 611 * And when we are woken up, put us back in LSRUN. If we 612 * slept for over a second, recalculate our estcpu. 613 */ 614 lp->lwp_stat = LSRUN; 615 if (lp->lwp_slptime) { 616 p->p_usched->uload_update(lp); 617 p->p_usched->recalculate(lp); 618 } 619 lp->lwp_slptime = 0; 620 } else { 621 lwkt_switch(); 622 } 623 624 /* 625 * Make sure we haven't switched cpus while we were asleep. It's 626 * not supposed to happen. Cleanup our temporary flags. 627 */ 628 KKASSERT(gd == td->td_gd); 629 630 /* 631 * Cleanup the timeout. If the timeout has already occured thandle 632 * has already been stopped, otherwise stop thandle. If the timeout 633 * is running (the callout thread must be blocked trying to get 634 * lwp_token) then wait for us to get scheduled. 635 */ 636 if (timo) { 637 while (td->td_flags & TDF_TIMEOUT_RUNNING) { 638 lwkt_deschedule_self(td); 639 td->td_wmesg = "tsrace"; 640 lwkt_switch(); 641 kprintf("td %p %s: timeout race\n", td, td->td_comm); 642 } 643 if (td->td_flags & TDF_TIMEOUT) { 644 td->td_flags &= ~TDF_TIMEOUT; 645 error = EWOULDBLOCK; 646 } else { 647 /* does not block when on same cpu */ 648 callout_stop(&thandle); 649 } 650 } 651 td->td_flags &= ~TDF_TSLEEP_DESCHEDULED; 652 653 /* 654 * Make sure we have been removed from the sleepq. In most 655 * cases this will have been done for us already but it is 656 * possible for a scheduling IPI to be in-flight from a 657 * previous tsleep/tsleep_interlock() or due to a straight-out 658 * call to lwkt_schedule() (in the case of an interrupt thread), 659 * causing a spurious wakeup. 660 */ 661 _tsleep_remove(td); 662 td->td_wmesg = NULL; 663 664 /* 665 * Figure out the correct error return. If interrupted by a 666 * signal we want to return EINTR or ERESTART. 667 */ 668 resume: 669 if (lp) { 670 if (catch && error == 0) { 671 if (sig != 0 || (sig = CURSIG(lp))) { 672 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 673 error = EINTR; 674 else 675 error = ERESTART; 676 } 677 } 678 lp->lwp_flags &= ~LWP_SINTR; 679 lwkt_reltoken(&lp->lwp_token); 680 } 681 logtsleep1(tsleep_end); 682 crit_exit_quick(td); 683 return (error); 684 } 685 686 /* 687 * Interlocked spinlock sleep. An exclusively held spinlock must 688 * be passed to ssleep(). The function will atomically release the 689 * spinlock and tsleep on the ident, then reacquire the spinlock and 690 * return. 691 * 692 * This routine is fairly important along the critical path, so optimize it 693 * heavily. 694 */ 695 int 696 ssleep(const volatile void *ident, struct spinlock *spin, int flags, 697 const char *wmesg, int timo) 698 { 699 globaldata_t gd = mycpu; 700 int error; 701 702 _tsleep_interlock(gd, ident, flags); 703 spin_unlock_quick(gd, spin); 704 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 705 _spin_lock_quick(gd, spin, wmesg); 706 707 return (error); 708 } 709 710 int 711 lksleep(const volatile void *ident, struct lock *lock, int flags, 712 const char *wmesg, int timo) 713 { 714 globaldata_t gd = mycpu; 715 int error; 716 717 _tsleep_interlock(gd, ident, flags); 718 lockmgr(lock, LK_RELEASE); 719 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 720 lockmgr(lock, LK_EXCLUSIVE); 721 722 return (error); 723 } 724 725 /* 726 * Interlocked mutex sleep. An exclusively held mutex must be passed 727 * to mtxsleep(). The function will atomically release the mutex 728 * and tsleep on the ident, then reacquire the mutex and return. 729 */ 730 int 731 mtxsleep(const volatile void *ident, struct mtx *mtx, int flags, 732 const char *wmesg, int timo) 733 { 734 globaldata_t gd = mycpu; 735 int error; 736 737 _tsleep_interlock(gd, ident, flags); 738 mtx_unlock(mtx); 739 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 740 mtx_lock_ex_quick(mtx, wmesg); 741 742 return (error); 743 } 744 745 /* 746 * Interlocked serializer sleep. An exclusively held serializer must 747 * be passed to zsleep(). The function will atomically release 748 * the serializer and tsleep on the ident, then reacquire the serializer 749 * and return. 750 */ 751 int 752 zsleep(const volatile void *ident, struct lwkt_serialize *slz, int flags, 753 const char *wmesg, int timo) 754 { 755 globaldata_t gd = mycpu; 756 int ret; 757 758 ASSERT_SERIALIZED(slz); 759 760 _tsleep_interlock(gd, ident, flags); 761 lwkt_serialize_exit(slz); 762 ret = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 763 lwkt_serialize_enter(slz); 764 765 return ret; 766 } 767 768 /* 769 * Directly block on the LWKT thread by descheduling it. This 770 * is much faster then tsleep(), but the only legal way to wake 771 * us up is to directly schedule the thread. 772 * 773 * Setting TDF_SINTR will cause new signals to directly schedule us. 774 * 775 * This routine must be called while in a critical section. 776 */ 777 int 778 lwkt_sleep(const char *wmesg, int flags) 779 { 780 thread_t td = curthread; 781 int sig; 782 783 if ((flags & PCATCH) == 0 || td->td_lwp == NULL) { 784 td->td_flags |= TDF_BLOCKED; 785 td->td_wmesg = wmesg; 786 lwkt_deschedule_self(td); 787 lwkt_switch(); 788 td->td_wmesg = NULL; 789 td->td_flags &= ~TDF_BLOCKED; 790 return(0); 791 } 792 if ((sig = CURSIG(td->td_lwp)) != 0) { 793 if (SIGISMEMBER(td->td_proc->p_sigacts->ps_sigintr, sig)) 794 return(EINTR); 795 else 796 return(ERESTART); 797 798 } 799 td->td_flags |= TDF_BLOCKED | TDF_SINTR; 800 td->td_wmesg = wmesg; 801 lwkt_deschedule_self(td); 802 lwkt_switch(); 803 td->td_flags &= ~(TDF_BLOCKED | TDF_SINTR); 804 td->td_wmesg = NULL; 805 return(0); 806 } 807 808 /* 809 * Implement the timeout for tsleep. 810 * 811 * This type of callout timeout is scheduled on the same cpu the process 812 * is sleeping on. Also, at the moment, the MP lock is held. 813 */ 814 static void 815 endtsleep(void *arg) 816 { 817 thread_t td = arg; 818 struct lwp *lp; 819 820 /* 821 * We are going to have to get the lwp_token, which means we might 822 * block. This can race a tsleep getting woken up by other means 823 * so set TDF_TIMEOUT_RUNNING to force the tsleep to wait for our 824 * processing to complete (sorry tsleep!). 825 * 826 * We can safely set td_flags because td MUST be on the same cpu 827 * as we are. 828 */ 829 KKASSERT(td->td_gd == mycpu); 830 crit_enter(); 831 td->td_flags |= TDF_TIMEOUT_RUNNING | TDF_TIMEOUT; 832 833 /* 834 * This can block but TDF_TIMEOUT_RUNNING will prevent the thread 835 * from exiting the tsleep on us. The flag is interlocked by virtue 836 * of lp being on the same cpu as we are. 837 */ 838 if ((lp = td->td_lwp) != NULL) 839 lwkt_gettoken(&lp->lwp_token); 840 841 KKASSERT(td->td_flags & TDF_TSLEEP_DESCHEDULED); 842 843 if (lp) { 844 /* 845 * callout timer should never be set in tstop() because 846 * it passes a timeout of 0. 847 */ 848 KKASSERT(lp->lwp_stat != LSSTOP); 849 setrunnable(lp); 850 lwkt_reltoken(&lp->lwp_token); 851 } else { 852 _tsleep_remove(td); 853 lwkt_schedule(td); 854 } 855 KKASSERT(td->td_gd == mycpu); 856 td->td_flags &= ~TDF_TIMEOUT_RUNNING; 857 crit_exit(); 858 } 859 860 /* 861 * Make all processes sleeping on the specified identifier runnable. 862 * count may be zero or one only. 863 * 864 * The domain encodes the sleep/wakeup domain, flags, plus the originating 865 * cpu. 866 * 867 * This call may run without the MP lock held. We can only manipulate thread 868 * state on the cpu owning the thread. We CANNOT manipulate process state 869 * at all. 870 * 871 * _wakeup() can be passed to an IPI so we can't use (const volatile 872 * void *ident). 873 */ 874 static void 875 _wakeup(void *ident, int domain) 876 { 877 struct tslpque *qp; 878 struct thread *td; 879 struct thread *ntd; 880 globaldata_t gd; 881 cpumask_t mask; 882 int id; 883 884 crit_enter(); 885 logtsleep2(wakeup_beg, ident); 886 gd = mycpu; 887 id = LOOKUP(ident); 888 qp = &gd->gd_tsleep_hash[id]; 889 restart: 890 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) { 891 ntd = TAILQ_NEXT(td, td_sleepq); 892 if (td->td_wchan == ident && 893 td->td_wdomain == (domain & PDOMAIN_MASK) 894 ) { 895 KKASSERT(td->td_gd == gd); 896 _tsleep_remove(td); 897 td->td_wakefromcpu = PWAKEUP_DECODE(domain); 898 if (td->td_flags & TDF_TSLEEP_DESCHEDULED) { 899 lwkt_schedule(td); 900 if (domain & PWAKEUP_ONE) 901 goto done; 902 } 903 goto restart; 904 } 905 } 906 907 /* 908 * We finished checking the current cpu but there still may be 909 * more work to do. Either wakeup_one was requested and no matching 910 * thread was found, or a normal wakeup was requested and we have 911 * to continue checking cpus. 912 * 913 * It should be noted that this scheme is actually less expensive then 914 * the old scheme when waking up multiple threads, since we send 915 * only one IPI message per target candidate which may then schedule 916 * multiple threads. Before we could have wound up sending an IPI 917 * message for each thread on the target cpu (!= current cpu) that 918 * needed to be woken up. 919 * 920 * NOTE: Wakeups occuring on remote cpus are asynchronous. This 921 * should be ok since we are passing idents in the IPI rather then 922 * thread pointers. 923 */ 924 if ((domain & PWAKEUP_MYCPU) == 0 && 925 (mask = slpque_cpumasks[id] & gd->gd_other_cpus) != 0) { 926 lwkt_send_ipiq2_mask(mask, _wakeup, ident, 927 domain | PWAKEUP_MYCPU); 928 } 929 done: 930 logtsleep1(wakeup_end); 931 crit_exit(); 932 } 933 934 /* 935 * Wakeup all threads tsleep()ing on the specified ident, on all cpus 936 */ 937 void 938 wakeup(const volatile void *ident) 939 { 940 globaldata_t gd = mycpu; 941 thread_t td = gd->gd_curthread; 942 943 if (td && (td->td_flags & TDF_DELAYED_WAKEUP)) { 944 /* 945 * If we are in a delayed wakeup section, record up to two wakeups in 946 * a per-CPU queue and issue them when we block or exit the delayed 947 * wakeup section. 948 */ 949 if (atomic_cmpset_ptr(&gd->gd_delayed_wakeup[0], NULL, ident)) 950 return; 951 if (atomic_cmpset_ptr(&gd->gd_delayed_wakeup[1], NULL, ident)) 952 return; 953 954 ident = atomic_swap_ptr(__DEQUALIFY(volatile void **, &gd->gd_delayed_wakeup[1]), 955 __DEALL(ident)); 956 ident = atomic_swap_ptr(__DEQUALIFY(volatile void **, &gd->gd_delayed_wakeup[0]), 957 __DEALL(ident)); 958 } 959 960 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, gd->gd_cpuid)); 961 } 962 963 /* 964 * Wakeup one thread tsleep()ing on the specified ident, on any cpu. 965 */ 966 void 967 wakeup_one(const volatile void *ident) 968 { 969 /* XXX potentially round-robin the first responding cpu */ 970 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mycpu->gd_cpuid) | 971 PWAKEUP_ONE); 972 } 973 974 /* 975 * Wakeup threads tsleep()ing on the specified ident on the current cpu 976 * only. 977 */ 978 void 979 wakeup_mycpu(const volatile void *ident) 980 { 981 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mycpu->gd_cpuid) | 982 PWAKEUP_MYCPU); 983 } 984 985 /* 986 * Wakeup one thread tsleep()ing on the specified ident on the current cpu 987 * only. 988 */ 989 void 990 wakeup_mycpu_one(const volatile void *ident) 991 { 992 /* XXX potentially round-robin the first responding cpu */ 993 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mycpu->gd_cpuid) | 994 PWAKEUP_MYCPU | PWAKEUP_ONE); 995 } 996 997 /* 998 * Wakeup all thread tsleep()ing on the specified ident on the specified cpu 999 * only. 1000 */ 1001 void 1002 wakeup_oncpu(globaldata_t gd, const volatile void *ident) 1003 { 1004 globaldata_t mygd = mycpu; 1005 if (gd == mycpu) { 1006 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mygd->gd_cpuid) | 1007 PWAKEUP_MYCPU); 1008 } else { 1009 lwkt_send_ipiq2(gd, _wakeup, __DEALL(ident), 1010 PWAKEUP_ENCODE(0, mygd->gd_cpuid) | 1011 PWAKEUP_MYCPU); 1012 } 1013 } 1014 1015 /* 1016 * Wakeup one thread tsleep()ing on the specified ident on the specified cpu 1017 * only. 1018 */ 1019 void 1020 wakeup_oncpu_one(globaldata_t gd, const volatile void *ident) 1021 { 1022 globaldata_t mygd = mycpu; 1023 if (gd == mygd) { 1024 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mygd->gd_cpuid) | 1025 PWAKEUP_MYCPU | PWAKEUP_ONE); 1026 } else { 1027 lwkt_send_ipiq2(gd, _wakeup, __DEALL(ident), 1028 PWAKEUP_ENCODE(0, mygd->gd_cpuid) | 1029 PWAKEUP_MYCPU | PWAKEUP_ONE); 1030 } 1031 } 1032 1033 /* 1034 * Wakeup all threads waiting on the specified ident that slept using 1035 * the specified domain, on all cpus. 1036 */ 1037 void 1038 wakeup_domain(const volatile void *ident, int domain) 1039 { 1040 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(domain, mycpu->gd_cpuid)); 1041 } 1042 1043 /* 1044 * Wakeup one thread waiting on the specified ident that slept using 1045 * the specified domain, on any cpu. 1046 */ 1047 void 1048 wakeup_domain_one(const volatile void *ident, int domain) 1049 { 1050 /* XXX potentially round-robin the first responding cpu */ 1051 _wakeup(__DEALL(ident), 1052 PWAKEUP_ENCODE(domain, mycpu->gd_cpuid) | PWAKEUP_ONE); 1053 } 1054 1055 void 1056 wakeup_start_delayed(void) 1057 { 1058 globaldata_t gd = mycpu; 1059 1060 crit_enter(); 1061 gd->gd_curthread->td_flags |= TDF_DELAYED_WAKEUP; 1062 crit_exit(); 1063 } 1064 1065 void 1066 wakeup_end_delayed(void) 1067 { 1068 globaldata_t gd = mycpu; 1069 1070 if (gd->gd_curthread->td_flags & TDF_DELAYED_WAKEUP) { 1071 crit_enter(); 1072 gd->gd_curthread->td_flags &= ~TDF_DELAYED_WAKEUP; 1073 if (gd->gd_delayed_wakeup[0] || gd->gd_delayed_wakeup[1]) { 1074 if (gd->gd_delayed_wakeup[0]) { 1075 wakeup(gd->gd_delayed_wakeup[0]); 1076 gd->gd_delayed_wakeup[0] = NULL; 1077 } 1078 if (gd->gd_delayed_wakeup[1]) { 1079 wakeup(gd->gd_delayed_wakeup[1]); 1080 gd->gd_delayed_wakeup[1] = NULL; 1081 } 1082 } 1083 crit_exit(); 1084 } 1085 } 1086 1087 /* 1088 * setrunnable() 1089 * 1090 * Make a process runnable. lp->lwp_token must be held on call and this 1091 * function must be called from the cpu owning lp. 1092 * 1093 * This only has an effect if we are in LSSTOP or LSSLEEP. 1094 */ 1095 void 1096 setrunnable(struct lwp *lp) 1097 { 1098 thread_t td = lp->lwp_thread; 1099 1100 ASSERT_LWKT_TOKEN_HELD(&lp->lwp_token); 1101 KKASSERT(td->td_gd == mycpu); 1102 crit_enter(); 1103 if (lp->lwp_stat == LSSTOP) 1104 lp->lwp_stat = LSSLEEP; 1105 if (lp->lwp_stat == LSSLEEP) { 1106 _tsleep_remove(td); 1107 lwkt_schedule(td); 1108 } else if (td->td_flags & TDF_SINTR) { 1109 lwkt_schedule(td); 1110 } 1111 crit_exit(); 1112 } 1113 1114 /* 1115 * The process is stopped due to some condition, usually because p_stat is 1116 * set to SSTOP, but also possibly due to being traced. 1117 * 1118 * Caller must hold p->p_token 1119 * 1120 * NOTE! If the caller sets SSTOP, the caller must also clear P_WAITED 1121 * because the parent may check the child's status before the child actually 1122 * gets to this routine. 1123 * 1124 * This routine is called with the current lwp only, typically just 1125 * before returning to userland if the process state is detected as 1126 * possibly being in a stopped state. 1127 */ 1128 void 1129 tstop(void) 1130 { 1131 struct lwp *lp = curthread->td_lwp; 1132 struct proc *p = lp->lwp_proc; 1133 struct proc *q; 1134 1135 lwkt_gettoken(&lp->lwp_token); 1136 crit_enter(); 1137 1138 /* 1139 * If LWP_MP_WSTOP is set, we were sleeping 1140 * while our process was stopped. At this point 1141 * we were already counted as stopped. 1142 */ 1143 if ((lp->lwp_mpflags & LWP_MP_WSTOP) == 0) { 1144 /* 1145 * If we're the last thread to stop, signal 1146 * our parent. 1147 */ 1148 p->p_nstopped++; 1149 atomic_set_int(&lp->lwp_mpflags, LWP_MP_WSTOP); 1150 wakeup(&p->p_nstopped); 1151 if (p->p_nstopped == p->p_nthreads) { 1152 /* 1153 * Token required to interlock kern_wait() 1154 */ 1155 q = p->p_pptr; 1156 PHOLD(q); 1157 lwkt_gettoken(&q->p_token); 1158 p->p_flags &= ~P_WAITED; 1159 wakeup(p->p_pptr); 1160 if ((q->p_sigacts->ps_flag & PS_NOCLDSTOP) == 0) 1161 ksignal(q, SIGCHLD); 1162 lwkt_reltoken(&q->p_token); 1163 PRELE(q); 1164 } 1165 } 1166 while (p->p_stat == SSTOP) { 1167 lp->lwp_stat = LSSTOP; 1168 tsleep(p, 0, "stop", 0); 1169 } 1170 p->p_nstopped--; 1171 atomic_clear_int(&lp->lwp_mpflags, LWP_MP_WSTOP); 1172 crit_exit(); 1173 lwkt_reltoken(&lp->lwp_token); 1174 } 1175 1176 /* 1177 * Compute a tenex style load average of a quantity on 1178 * 1, 5 and 15 minute intervals. 1179 */ 1180 static int loadav_count_runnable(struct lwp *p, void *data); 1181 1182 static void 1183 loadav(void *arg) 1184 { 1185 struct loadavg *avg; 1186 int i, nrun; 1187 1188 nrun = 0; 1189 alllwp_scan(loadav_count_runnable, &nrun); 1190 avg = &averunnable; 1191 for (i = 0; i < 3; i++) { 1192 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] + 1193 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT; 1194 } 1195 1196 /* 1197 * Schedule the next update to occur after 5 seconds, but add a 1198 * random variation to avoid synchronisation with processes that 1199 * run at regular intervals. 1200 */ 1201 callout_reset(&loadav_callout, hz * 4 + (int)(krandom() % (hz * 2 + 1)), 1202 loadav, NULL); 1203 } 1204 1205 static int 1206 loadav_count_runnable(struct lwp *lp, void *data) 1207 { 1208 int *nrunp = data; 1209 thread_t td; 1210 1211 switch (lp->lwp_stat) { 1212 case LSRUN: 1213 if ((td = lp->lwp_thread) == NULL) 1214 break; 1215 if (td->td_flags & TDF_BLOCKED) 1216 break; 1217 ++*nrunp; 1218 break; 1219 default: 1220 break; 1221 } 1222 lwkt_yield(); 1223 return(0); 1224 } 1225 1226 /* ARGSUSED */ 1227 static void 1228 sched_setup(void *dummy) 1229 { 1230 callout_init_mp(&loadav_callout); 1231 callout_init_mp(&schedcpu_callout); 1232 1233 /* Kick off timeout driven events by calling first time. */ 1234 schedcpu(NULL); 1235 loadav(NULL); 1236 } 1237 1238