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. All advertising materials mentioning features or use of this software 19 * must display the following acknowledgement: 20 * This product includes software developed by the University of 21 * California, Berkeley and its contributors. 22 * 4. Neither the name of the University nor the names of its contributors 23 * may be used to endorse or promote products derived from this software 24 * without specific prior written permission. 25 * 26 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 27 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 28 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 29 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 30 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 31 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 32 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 33 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 34 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 35 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 36 * SUCH DAMAGE. 37 * 38 * @(#)kern_synch.c 8.9 (Berkeley) 5/19/95 39 * $FreeBSD: src/sys/kern/kern_synch.c,v 1.87.2.6 2002/10/13 07:29:53 kbyanc Exp $ 40 * $DragonFly: src/sys/kern/kern_synch.c,v 1.91 2008/09/09 04:06:13 dillon Exp $ 41 */ 42 43 #include "opt_ktrace.h" 44 45 #include <sys/param.h> 46 #include <sys/systm.h> 47 #include <sys/proc.h> 48 #include <sys/kernel.h> 49 #include <sys/signalvar.h> 50 #include <sys/resourcevar.h> 51 #include <sys/vmmeter.h> 52 #include <sys/sysctl.h> 53 #include <sys/lock.h> 54 #include <sys/uio.h> 55 #ifdef KTRACE 56 #include <sys/ktrace.h> 57 #endif 58 #include <sys/xwait.h> 59 #include <sys/ktr.h> 60 #include <sys/serialize.h> 61 62 #include <sys/signal2.h> 63 #include <sys/thread2.h> 64 #include <sys/spinlock2.h> 65 #include <sys/mutex2.h> 66 #include <sys/mplock2.h> 67 68 #include <machine/cpu.h> 69 #include <machine/smp.h> 70 71 TAILQ_HEAD(tslpque, thread); 72 73 static void sched_setup (void *dummy); 74 SYSINIT(sched_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, sched_setup, NULL) 75 76 int hogticks; 77 int lbolt; 78 int lbolt_syncer; 79 int sched_quantum; /* Roundrobin scheduling quantum in ticks. */ 80 int ncpus; 81 int ncpus2, ncpus2_shift, ncpus2_mask; /* note: mask not cpumask_t */ 82 int ncpus_fit, ncpus_fit_mask; /* note: mask not cpumask_t */ 83 int safepri; 84 int tsleep_now_works; 85 int tsleep_crypto_dump = 0; 86 87 static struct callout loadav_callout; 88 static struct callout schedcpu_callout; 89 MALLOC_DEFINE(M_TSLEEP, "tslpque", "tsleep queues"); 90 91 #define __DEALL(ident) __DEQUALIFY(void *, ident) 92 93 #if !defined(KTR_TSLEEP) 94 #define KTR_TSLEEP KTR_ALL 95 #endif 96 KTR_INFO_MASTER(tsleep); 97 KTR_INFO(KTR_TSLEEP, tsleep, tsleep_beg, 0, "tsleep enter %p", sizeof(void *)); 98 KTR_INFO(KTR_TSLEEP, tsleep, tsleep_end, 1, "tsleep exit", 0); 99 KTR_INFO(KTR_TSLEEP, tsleep, wakeup_beg, 2, "wakeup enter %p", sizeof(void *)); 100 KTR_INFO(KTR_TSLEEP, tsleep, wakeup_end, 3, "wakeup exit", 0); 101 KTR_INFO(KTR_TSLEEP, tsleep, ilockfail, 4, "interlock failed %p", sizeof(void *)); 102 103 #define logtsleep1(name) KTR_LOG(tsleep_ ## name) 104 #define logtsleep2(name, val) KTR_LOG(tsleep_ ## name, val) 105 106 struct loadavg averunnable = 107 { {0, 0, 0}, FSCALE }; /* load average, of runnable procs */ 108 /* 109 * Constants for averages over 1, 5, and 15 minutes 110 * when sampling at 5 second intervals. 111 */ 112 static fixpt_t cexp[3] = { 113 0.9200444146293232 * FSCALE, /* exp(-1/12) */ 114 0.9834714538216174 * FSCALE, /* exp(-1/60) */ 115 0.9944598480048967 * FSCALE, /* exp(-1/180) */ 116 }; 117 118 static void endtsleep (void *); 119 static void loadav (void *arg); 120 static void schedcpu (void *arg); 121 #ifdef SMP 122 static void tsleep_wakeup_remote(struct thread *td); 123 #endif 124 125 /* 126 * Adjust the scheduler quantum. The quantum is specified in microseconds. 127 * Note that 'tick' is in microseconds per tick. 128 */ 129 static int 130 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS) 131 { 132 int error, new_val; 133 134 new_val = sched_quantum * ustick; 135 error = sysctl_handle_int(oidp, &new_val, 0, req); 136 if (error != 0 || req->newptr == NULL) 137 return (error); 138 if (new_val < ustick) 139 return (EINVAL); 140 sched_quantum = new_val / ustick; 141 hogticks = 2 * sched_quantum; 142 return (0); 143 } 144 145 SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW, 146 0, sizeof sched_quantum, sysctl_kern_quantum, "I", ""); 147 148 /* 149 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the 150 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below 151 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT). 152 * 153 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used: 154 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits). 155 * 156 * If you don't want to bother with the faster/more-accurate formula, you 157 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate 158 * (more general) method of calculating the %age of CPU used by a process. 159 * 160 * decay 95% of `lwp_pctcpu' in 60 seconds; see CCPU_SHIFT before changing 161 */ 162 #define CCPU_SHIFT 11 163 164 static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */ 165 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 166 167 /* 168 * kernel uses `FSCALE', userland (SHOULD) use kern.fscale 169 */ 170 int fscale __unused = FSCALE; /* exported to systat */ 171 SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, ""); 172 173 /* 174 * Recompute process priorities, once a second. 175 * 176 * Since the userland schedulers are typically event oriented, if the 177 * estcpu calculation at wakeup() time is not sufficient to make a 178 * process runnable relative to other processes in the system we have 179 * a 1-second recalc to help out. 180 * 181 * This code also allows us to store sysclock_t data in the process structure 182 * without fear of an overrun, since sysclock_t are guarenteed to hold 183 * several seconds worth of count. 184 * 185 * WARNING! callouts can preempt normal threads. However, they will not 186 * preempt a thread holding a spinlock so we *can* safely use spinlocks. 187 */ 188 static int schedcpu_stats(struct proc *p, void *data __unused); 189 static int schedcpu_resource(struct proc *p, void *data __unused); 190 191 static void 192 schedcpu(void *arg) 193 { 194 allproc_scan(schedcpu_stats, NULL); 195 allproc_scan(schedcpu_resource, NULL); 196 wakeup((caddr_t)&lbolt); 197 wakeup((caddr_t)&lbolt_syncer); 198 callout_reset(&schedcpu_callout, hz, schedcpu, NULL); 199 } 200 201 /* 202 * General process statistics once a second 203 */ 204 static int 205 schedcpu_stats(struct proc *p, void *data __unused) 206 { 207 struct lwp *lp; 208 209 crit_enter(); 210 p->p_swtime++; 211 FOREACH_LWP_IN_PROC(lp, p) { 212 if (lp->lwp_stat == LSSLEEP) 213 lp->lwp_slptime++; 214 215 /* 216 * Only recalculate processes that are active or have slept 217 * less then 2 seconds. The schedulers understand this. 218 */ 219 if (lp->lwp_slptime <= 1) { 220 p->p_usched->recalculate(lp); 221 } else { 222 lp->lwp_pctcpu = (lp->lwp_pctcpu * ccpu) >> FSHIFT; 223 } 224 } 225 crit_exit(); 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 crit_enter(); 241 if (p->p_stat == SIDL || 242 p->p_stat == SZOMB || 243 p->p_limit == NULL 244 ) { 245 crit_exit(); 246 return(0); 247 } 248 249 ttime = 0; 250 FOREACH_LWP_IN_PROC(lp, p) { 251 /* 252 * We may have caught an lp in the middle of being 253 * created, lwp_thread can be NULL. 254 */ 255 if (lp->lwp_thread) { 256 ttime += lp->lwp_thread->td_sticks; 257 ttime += lp->lwp_thread->td_uticks; 258 } 259 } 260 261 switch(plimit_testcpulimit(p->p_limit, ttime)) { 262 case PLIMIT_TESTCPU_KILL: 263 killproc(p, "exceeded maximum CPU limit"); 264 break; 265 case PLIMIT_TESTCPU_XCPU: 266 if ((p->p_flag & P_XCPU) == 0) { 267 p->p_flag |= P_XCPU; 268 ksignal(p, SIGXCPU); 269 } 270 break; 271 default: 272 break; 273 } 274 crit_exit(); 275 return(0); 276 } 277 278 /* 279 * This is only used by ps. Generate a cpu percentage use over 280 * a period of one second. 281 * 282 * MPSAFE 283 */ 284 void 285 updatepcpu(struct lwp *lp, int cpticks, int ttlticks) 286 { 287 fixpt_t acc; 288 int remticks; 289 290 acc = (cpticks << FSHIFT) / ttlticks; 291 if (ttlticks >= ESTCPUFREQ) { 292 lp->lwp_pctcpu = acc; 293 } else { 294 remticks = ESTCPUFREQ - ttlticks; 295 lp->lwp_pctcpu = (acc * ttlticks + lp->lwp_pctcpu * remticks) / 296 ESTCPUFREQ; 297 } 298 } 299 300 /* 301 * tsleep/wakeup hash table parameters. Try to find the sweet spot for 302 * like addresses being slept on. 303 */ 304 #define TABLESIZE 1024 305 #define LOOKUP(x) (((intptr_t)(x) >> 6) & (TABLESIZE - 1)) 306 307 static cpumask_t slpque_cpumasks[TABLESIZE]; 308 309 /* 310 * General scheduler initialization. We force a reschedule 25 times 311 * a second by default. Note that cpu0 is initialized in early boot and 312 * cannot make any high level calls. 313 * 314 * Each cpu has its own sleep queue. 315 */ 316 void 317 sleep_gdinit(globaldata_t gd) 318 { 319 static struct tslpque slpque_cpu0[TABLESIZE]; 320 int i; 321 322 if (gd->gd_cpuid == 0) { 323 sched_quantum = (hz + 24) / 25; 324 hogticks = 2 * sched_quantum; 325 326 gd->gd_tsleep_hash = slpque_cpu0; 327 } else { 328 gd->gd_tsleep_hash = kmalloc(sizeof(slpque_cpu0), 329 M_TSLEEP, M_WAITOK | M_ZERO); 330 } 331 for (i = 0; i < TABLESIZE; ++i) 332 TAILQ_INIT(&gd->gd_tsleep_hash[i]); 333 } 334 335 /* 336 * This is a dandy function that allows us to interlock tsleep/wakeup 337 * operations with unspecified upper level locks, such as lockmgr locks, 338 * simply by holding a critical section. The sequence is: 339 * 340 * (acquire upper level lock) 341 * tsleep_interlock(blah) 342 * (release upper level lock) 343 * tsleep(blah, ...) 344 * 345 * Basically this functions queues us on the tsleep queue without actually 346 * descheduling us. When tsleep() is later called with PINTERLOCK it 347 * assumes the thread was already queued, otherwise it queues it there. 348 * 349 * Thus it is possible to receive the wakeup prior to going to sleep and 350 * the race conditions are covered. 351 */ 352 static __inline void 353 _tsleep_interlock(globaldata_t gd, const volatile void *ident, int flags) 354 { 355 thread_t td = gd->gd_curthread; 356 int id; 357 358 crit_enter_quick(td); 359 if (td->td_flags & TDF_TSLEEPQ) { 360 id = LOOKUP(td->td_wchan); 361 TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_sleepq); 362 if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL) 363 atomic_clear_cpumask(&slpque_cpumasks[id], gd->gd_cpumask); 364 } else { 365 td->td_flags |= TDF_TSLEEPQ; 366 } 367 id = LOOKUP(ident); 368 TAILQ_INSERT_TAIL(&gd->gd_tsleep_hash[id], td, td_sleepq); 369 atomic_set_cpumask(&slpque_cpumasks[id], gd->gd_cpumask); 370 td->td_wchan = ident; 371 td->td_wdomain = flags & PDOMAIN_MASK; 372 crit_exit_quick(td); 373 } 374 375 void 376 tsleep_interlock(const volatile void *ident, int flags) 377 { 378 _tsleep_interlock(mycpu, ident, flags); 379 } 380 381 /* 382 * Remove thread from sleepq. Must be called with a critical section held. 383 */ 384 static __inline void 385 _tsleep_remove(thread_t td) 386 { 387 globaldata_t gd = mycpu; 388 int id; 389 390 KKASSERT(td->td_gd == gd); 391 if (td->td_flags & TDF_TSLEEPQ) { 392 td->td_flags &= ~TDF_TSLEEPQ; 393 id = LOOKUP(td->td_wchan); 394 TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_sleepq); 395 if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL) 396 atomic_clear_cpumask(&slpque_cpumasks[id], gd->gd_cpumask); 397 td->td_wchan = NULL; 398 td->td_wdomain = 0; 399 } 400 } 401 402 void 403 tsleep_remove(thread_t td) 404 { 405 _tsleep_remove(td); 406 } 407 408 /* 409 * This function removes a thread from the tsleep queue and schedules 410 * it. This function may act asynchronously. The target thread may be 411 * sleeping on a different cpu. 412 * 413 * This function mus be called while in a critical section but if the 414 * target thread is sleeping on a different cpu we cannot safely probe 415 * td_flags. 416 * 417 * This function is only called from a different cpu via setrunnable() 418 * when the thread is in a known sleep. However, multiple wakeups are 419 * possible and we must hold the td to prevent a race against the thread 420 * exiting. 421 */ 422 static __inline 423 void 424 _tsleep_wakeup(struct thread *td) 425 { 426 #ifdef SMP 427 globaldata_t gd = mycpu; 428 429 if (td->td_gd != gd) { 430 lwkt_hold(td); 431 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)tsleep_wakeup_remote, td); 432 return; 433 } 434 #endif 435 _tsleep_remove(td); 436 if (td->td_flags & TDF_TSLEEP_DESCHEDULED) { 437 td->td_flags &= ~TDF_TSLEEP_DESCHEDULED; 438 lwkt_schedule(td); 439 } 440 } 441 442 #ifdef SMP 443 static 444 void 445 tsleep_wakeup_remote(struct thread *td) 446 { 447 _tsleep_wakeup(td); 448 lwkt_rele(td); 449 } 450 #endif 451 452 453 /* 454 * General sleep call. Suspends the current process until a wakeup is 455 * performed on the specified identifier. The process will then be made 456 * runnable with the specified priority. Sleeps at most timo/hz seconds 457 * (0 means no timeout). If flags includes PCATCH flag, signals are checked 458 * before and after sleeping, else signals are not checked. Returns 0 if 459 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a 460 * signal needs to be delivered, ERESTART is returned if the current system 461 * call should be restarted if possible, and EINTR is returned if the system 462 * call should be interrupted by the signal (return EINTR). 463 * 464 * Note that if we are a process, we release_curproc() before messing with 465 * the LWKT scheduler. 466 * 467 * During autoconfiguration or after a panic, a sleep will simply 468 * lower the priority briefly to allow interrupts, then return. 469 */ 470 int 471 tsleep(const volatile void *ident, int flags, const char *wmesg, int timo) 472 { 473 struct thread *td = curthread; 474 struct lwp *lp = td->td_lwp; 475 struct proc *p = td->td_proc; /* may be NULL */ 476 globaldata_t gd; 477 int sig; 478 int catch; 479 int id; 480 int error; 481 int oldpri; 482 struct callout thandle; 483 484 /* 485 * NOTE: removed KTRPOINT, it could cause races due to blocking 486 * even in stable. Just scrap it for now. 487 */ 488 if (!tsleep_crypto_dump && (tsleep_now_works == 0 || panicstr)) { 489 /* 490 * After a panic, or before we actually have an operational 491 * softclock, just give interrupts a chance, then just return; 492 * 493 * don't run any other procs or panic below, 494 * in case this is the idle process and already asleep. 495 */ 496 splz(); 497 oldpri = td->td_pri; 498 lwkt_setpri_self(safepri); 499 lwkt_switch(); 500 lwkt_setpri_self(oldpri); 501 return (0); 502 } 503 logtsleep2(tsleep_beg, ident); 504 gd = td->td_gd; 505 KKASSERT(td != &gd->gd_idlethread); /* you must be kidding! */ 506 507 /* 508 * NOTE: all of this occurs on the current cpu, including any 509 * callout-based wakeups, so a critical section is a sufficient 510 * interlock. 511 * 512 * The entire sequence through to where we actually sleep must 513 * run without breaking the critical section. 514 */ 515 catch = flags & PCATCH; 516 error = 0; 517 sig = 0; 518 519 crit_enter_quick(td); 520 521 KASSERT(ident != NULL, ("tsleep: no ident")); 522 KASSERT(lp == NULL || 523 lp->lwp_stat == LSRUN || /* Obvious */ 524 lp->lwp_stat == LSSTOP, /* Set in tstop */ 525 ("tsleep %p %s %d", 526 ident, wmesg, lp->lwp_stat)); 527 528 /* 529 * Setup for the current process (if this is a process). 530 */ 531 if (lp) { 532 if (catch) { 533 /* 534 * Early termination if PCATCH was set and a 535 * signal is pending, interlocked with the 536 * critical section. 537 * 538 * Early termination only occurs when tsleep() is 539 * entered while in a normal LSRUN state. 540 */ 541 if ((sig = CURSIG(lp)) != 0) 542 goto resume; 543 544 /* 545 * Early termination if PCATCH was set and a 546 * mailbox signal was possibly delivered prior to 547 * the system call even being made, in order to 548 * allow the user to interlock without having to 549 * make additional system calls. 550 */ 551 if (p->p_flag & P_MAILBOX) 552 goto resume; 553 554 /* 555 * Causes ksignal to wake us up when. 556 */ 557 lp->lwp_flag |= LWP_SINTR; 558 } 559 } 560 561 /* 562 * We interlock the sleep queue if the caller has not already done 563 * it for us. 564 */ 565 if ((flags & PINTERLOCKED) == 0) { 566 id = LOOKUP(ident); 567 _tsleep_interlock(gd, ident, flags); 568 } 569 570 /* 571 * 572 * If no interlock was set we do an integrated interlock here. 573 * Make sure the current process has been untangled from 574 * the userland scheduler and initialize slptime to start 575 * counting. We must interlock the sleep queue before doing 576 * this to avoid wakeup/process-ipi races which can occur under 577 * heavy loads. 578 */ 579 if (lp) { 580 p->p_usched->release_curproc(lp); 581 lp->lwp_slptime = 0; 582 } 583 584 /* 585 * If the interlocked flag is set but our cpu bit in the slpqueue 586 * is no longer set, then a wakeup was processed inbetween the 587 * tsleep_interlock() (ours or the callers), and here. This can 588 * occur under numerous circumstances including when we release the 589 * current process. 590 * 591 * Extreme loads can cause the sending of an IPI (e.g. wakeup()'s) 592 * to process incoming IPIs, thus draining incoming wakeups. 593 */ 594 if ((td->td_flags & TDF_TSLEEPQ) == 0) { 595 logtsleep2(ilockfail, ident); 596 goto resume; 597 } 598 599 /* 600 * scheduling is blocked while in a critical section. Coincide 601 * the descheduled-by-tsleep flag with the descheduling of the 602 * lwkt. 603 */ 604 lwkt_deschedule_self(td); 605 td->td_flags |= TDF_TSLEEP_DESCHEDULED; 606 td->td_wmesg = wmesg; 607 608 /* 609 * Setup the timeout, if any 610 */ 611 if (timo) { 612 callout_init(&thandle); 613 callout_reset(&thandle, timo, endtsleep, td); 614 } 615 616 /* 617 * Beddy bye bye. 618 */ 619 if (lp) { 620 /* 621 * Ok, we are sleeping. Place us in the SSLEEP state. 622 */ 623 KKASSERT((lp->lwp_flag & LWP_ONRUNQ) == 0); 624 /* 625 * tstop() sets LSSTOP, so don't fiddle with that. 626 */ 627 if (lp->lwp_stat != LSSTOP) 628 lp->lwp_stat = LSSLEEP; 629 lp->lwp_ru.ru_nvcsw++; 630 lwkt_switch(); 631 632 /* 633 * And when we are woken up, put us back in LSRUN. If we 634 * slept for over a second, recalculate our estcpu. 635 */ 636 lp->lwp_stat = LSRUN; 637 if (lp->lwp_slptime) 638 p->p_usched->recalculate(lp); 639 lp->lwp_slptime = 0; 640 } else { 641 lwkt_switch(); 642 } 643 644 /* 645 * Make sure we haven't switched cpus while we were asleep. It's 646 * not supposed to happen. Cleanup our temporary flags. 647 */ 648 KKASSERT(gd == td->td_gd); 649 650 /* 651 * Cleanup the timeout. 652 */ 653 if (timo) { 654 if (td->td_flags & TDF_TIMEOUT) { 655 td->td_flags &= ~TDF_TIMEOUT; 656 error = EWOULDBLOCK; 657 } else { 658 callout_stop(&thandle); 659 } 660 } 661 662 /* 663 * Make sure we have been removed from the sleepq. This should 664 * have been done for us already. 665 * 666 * However, it is possible for a scheduling IPI to be in flight 667 * from a previous tsleep/tsleep_interlock or due to a straight-out 668 * call to lwkt_schedule() (in the case of an interrupt thread). 669 * So don't complain if DESCHEDULED is still set. 670 */ 671 _tsleep_remove(td); 672 td->td_wmesg = NULL; 673 if (td->td_flags & TDF_TSLEEP_DESCHEDULED) { 674 td->td_flags &= ~TDF_TSLEEP_DESCHEDULED; 675 } 676 677 /* 678 * Figure out the correct error return. If interrupted by a 679 * signal we want to return EINTR or ERESTART. 680 * 681 * If P_MAILBOX is set no automatic system call restart occurs 682 * and we return EINTR. P_MAILBOX is meant to be used as an 683 * interlock, the user must poll it prior to any system call 684 * that it wishes to interlock a mailbox signal against since 685 * the flag is cleared on *any* system call that sleeps. 686 */ 687 resume: 688 if (p) { 689 if (catch && error == 0) { 690 if ((p->p_flag & P_MAILBOX) && sig == 0) { 691 error = EINTR; 692 } else if (sig != 0 || (sig = CURSIG(lp))) { 693 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 694 error = EINTR; 695 else 696 error = ERESTART; 697 } 698 } 699 lp->lwp_flag &= ~(LWP_BREAKTSLEEP | LWP_SINTR); 700 p->p_flag &= ~P_MAILBOX; 701 } 702 logtsleep1(tsleep_end); 703 crit_exit_quick(td); 704 return (error); 705 } 706 707 /* 708 * Interlocked spinlock sleep. An exclusively held spinlock must 709 * be passed to ssleep(). The function will atomically release the 710 * spinlock and tsleep on the ident, then reacquire the spinlock and 711 * return. 712 * 713 * This routine is fairly important along the critical path, so optimize it 714 * heavily. 715 */ 716 int 717 ssleep(const volatile void *ident, struct spinlock *spin, int flags, 718 const char *wmesg, int timo) 719 { 720 globaldata_t gd = mycpu; 721 int error; 722 723 _tsleep_interlock(gd, ident, flags); 724 spin_unlock_quick(gd, spin); 725 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 726 spin_lock_quick(gd, spin); 727 728 return (error); 729 } 730 731 int 732 lksleep(const volatile void *ident, struct lock *lock, int flags, 733 const char *wmesg, int timo) 734 { 735 globaldata_t gd = mycpu; 736 int error; 737 738 _tsleep_interlock(gd, ident, flags); 739 lockmgr(lock, LK_RELEASE); 740 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 741 lockmgr(lock, LK_EXCLUSIVE); 742 743 return (error); 744 } 745 746 /* 747 * Interlocked mutex sleep. An exclusively held mutex must be passed 748 * to mtxsleep(). The function will atomically release the mutex 749 * and tsleep on the ident, then reacquire the mutex and return. 750 */ 751 int 752 mtxsleep(const volatile void *ident, struct mtx *mtx, int flags, 753 const char *wmesg, int timo) 754 { 755 globaldata_t gd = mycpu; 756 int error; 757 758 _tsleep_interlock(gd, ident, flags); 759 mtx_unlock(mtx); 760 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 761 mtx_lock_ex_quick(mtx, wmesg); 762 763 return (error); 764 } 765 766 /* 767 * Interlocked serializer sleep. An exclusively held serializer must 768 * be passed to zsleep(). The function will atomically release 769 * the serializer and tsleep on the ident, then reacquire the serializer 770 * and return. 771 */ 772 int 773 zsleep(const volatile void *ident, struct lwkt_serialize *slz, int flags, 774 const char *wmesg, int timo) 775 { 776 globaldata_t gd = mycpu; 777 int ret; 778 779 ASSERT_SERIALIZED(slz); 780 781 _tsleep_interlock(gd, ident, flags); 782 lwkt_serialize_exit(slz); 783 ret = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 784 lwkt_serialize_enter(slz); 785 786 return ret; 787 } 788 789 /* 790 * Directly block on the LWKT thread by descheduling it. This 791 * is much faster then tsleep(), but the only legal way to wake 792 * us up is to directly schedule the thread. 793 * 794 * Setting TDF_SINTR will cause new signals to directly schedule us. 795 * 796 * This routine must be called while in a critical section. 797 */ 798 int 799 lwkt_sleep(const char *wmesg, int flags) 800 { 801 thread_t td = curthread; 802 int sig; 803 804 if ((flags & PCATCH) == 0 || td->td_lwp == NULL) { 805 td->td_flags |= TDF_BLOCKED; 806 td->td_wmesg = wmesg; 807 lwkt_deschedule_self(td); 808 lwkt_switch(); 809 td->td_wmesg = NULL; 810 td->td_flags &= ~TDF_BLOCKED; 811 return(0); 812 } 813 if ((sig = CURSIG(td->td_lwp)) != 0) { 814 if (SIGISMEMBER(td->td_proc->p_sigacts->ps_sigintr, sig)) 815 return(EINTR); 816 else 817 return(ERESTART); 818 819 } 820 td->td_flags |= TDF_BLOCKED | TDF_SINTR; 821 td->td_wmesg = wmesg; 822 lwkt_deschedule_self(td); 823 lwkt_switch(); 824 td->td_flags &= ~(TDF_BLOCKED | TDF_SINTR); 825 td->td_wmesg = NULL; 826 return(0); 827 } 828 829 /* 830 * Implement the timeout for tsleep. 831 * 832 * We set LWP_BREAKTSLEEP to indicate that an event has occured, but 833 * we only call setrunnable if the process is not stopped. 834 * 835 * This type of callout timeout is scheduled on the same cpu the process 836 * is sleeping on. Also, at the moment, the MP lock is held. 837 */ 838 static void 839 endtsleep(void *arg) 840 { 841 thread_t td = arg; 842 struct lwp *lp; 843 844 crit_enter(); 845 lwkt_gettoken(&proc_token); 846 847 /* 848 * cpu interlock. Thread flags are only manipulated on 849 * the cpu owning the thread. proc flags are only manipulated 850 * by the older of the MP lock. We have both. 851 */ 852 if (td->td_flags & TDF_TSLEEP_DESCHEDULED) { 853 td->td_flags |= TDF_TIMEOUT; 854 855 if ((lp = td->td_lwp) != NULL) { 856 lp->lwp_flag |= LWP_BREAKTSLEEP; 857 if (lp->lwp_proc->p_stat != SSTOP) 858 setrunnable(lp); 859 } else { 860 _tsleep_wakeup(td); 861 } 862 } 863 lwkt_reltoken(&proc_token); 864 crit_exit(); 865 } 866 867 /* 868 * Make all processes sleeping on the specified identifier runnable. 869 * count may be zero or one only. 870 * 871 * The domain encodes the sleep/wakeup domain AND the first cpu to check 872 * (which is always the current cpu). As we iterate across cpus 873 * 874 * This call may run without the MP lock held. We can only manipulate thread 875 * state on the cpu owning the thread. We CANNOT manipulate process state 876 * at all. 877 * 878 * _wakeup() can be passed to an IPI so we can't use (const volatile 879 * void *ident). 880 */ 881 static void 882 _wakeup(void *ident, int domain) 883 { 884 struct tslpque *qp; 885 struct thread *td; 886 struct thread *ntd; 887 globaldata_t gd; 888 #ifdef SMP 889 cpumask_t mask; 890 #endif 891 int id; 892 893 crit_enter(); 894 logtsleep2(wakeup_beg, ident); 895 gd = mycpu; 896 id = LOOKUP(ident); 897 qp = &gd->gd_tsleep_hash[id]; 898 restart: 899 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) { 900 ntd = TAILQ_NEXT(td, td_sleepq); 901 if (td->td_wchan == ident && 902 td->td_wdomain == (domain & PDOMAIN_MASK) 903 ) { 904 KKASSERT(td->td_gd == gd); 905 _tsleep_remove(td); 906 if (td->td_flags & TDF_TSLEEP_DESCHEDULED) { 907 td->td_flags &= ~TDF_TSLEEP_DESCHEDULED; 908 lwkt_schedule(td); 909 if (domain & PWAKEUP_ONE) 910 goto done; 911 } 912 goto restart; 913 } 914 } 915 916 #ifdef SMP 917 /* 918 * We finished checking the current cpu but there still may be 919 * more work to do. Either wakeup_one was requested and no matching 920 * thread was found, or a normal wakeup was requested and we have 921 * to continue checking cpus. 922 * 923 * It should be noted that this scheme is actually less expensive then 924 * the old scheme when waking up multiple threads, since we send 925 * only one IPI message per target candidate which may then schedule 926 * multiple threads. Before we could have wound up sending an IPI 927 * message for each thread on the target cpu (!= current cpu) that 928 * needed to be woken up. 929 * 930 * NOTE: Wakeups occuring on remote cpus are asynchronous. This 931 * should be ok since we are passing idents in the IPI rather then 932 * thread pointers. 933 */ 934 if ((domain & PWAKEUP_MYCPU) == 0 && 935 (mask = slpque_cpumasks[id] & gd->gd_other_cpus) != 0) { 936 lwkt_send_ipiq2_mask(mask, _wakeup, ident, 937 domain | PWAKEUP_MYCPU); 938 } 939 #endif 940 done: 941 logtsleep1(wakeup_end); 942 crit_exit(); 943 } 944 945 /* 946 * Wakeup all threads tsleep()ing on the specified ident, on all cpus 947 */ 948 void 949 wakeup(const volatile void *ident) 950 { 951 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mycpu->gd_cpuid)); 952 } 953 954 /* 955 * Wakeup one thread tsleep()ing on the specified ident, on any cpu. 956 */ 957 void 958 wakeup_one(const volatile void *ident) 959 { 960 /* XXX potentially round-robin the first responding cpu */ 961 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mycpu->gd_cpuid) | PWAKEUP_ONE); 962 } 963 964 /* 965 * Wakeup threads tsleep()ing on the specified ident on the current cpu 966 * only. 967 */ 968 void 969 wakeup_mycpu(const volatile void *ident) 970 { 971 _wakeup(__DEALL(ident), PWAKEUP_MYCPU); 972 } 973 974 /* 975 * Wakeup one thread tsleep()ing on the specified ident on the current cpu 976 * only. 977 */ 978 void 979 wakeup_mycpu_one(const volatile void *ident) 980 { 981 /* XXX potentially round-robin the first responding cpu */ 982 _wakeup(__DEALL(ident), PWAKEUP_MYCPU|PWAKEUP_ONE); 983 } 984 985 /* 986 * Wakeup all thread tsleep()ing on the specified ident on the specified cpu 987 * only. 988 */ 989 void 990 wakeup_oncpu(globaldata_t gd, const volatile void *ident) 991 { 992 #ifdef SMP 993 if (gd == mycpu) { 994 _wakeup(__DEALL(ident), PWAKEUP_MYCPU); 995 } else { 996 lwkt_send_ipiq2(gd, _wakeup, __DEALL(ident), PWAKEUP_MYCPU); 997 } 998 #else 999 _wakeup(__DEALL(ident), PWAKEUP_MYCPU); 1000 #endif 1001 } 1002 1003 /* 1004 * Wakeup one thread tsleep()ing on the specified ident on the specified cpu 1005 * only. 1006 */ 1007 void 1008 wakeup_oncpu_one(globaldata_t gd, const volatile void *ident) 1009 { 1010 #ifdef SMP 1011 if (gd == mycpu) { 1012 _wakeup(__DEALL(ident), PWAKEUP_MYCPU | PWAKEUP_ONE); 1013 } else { 1014 lwkt_send_ipiq2(gd, _wakeup, __DEALL(ident), 1015 PWAKEUP_MYCPU | PWAKEUP_ONE); 1016 } 1017 #else 1018 _wakeup(__DEALL(ident), PWAKEUP_MYCPU | PWAKEUP_ONE); 1019 #endif 1020 } 1021 1022 /* 1023 * Wakeup all threads waiting on the specified ident that slept using 1024 * the specified domain, on all cpus. 1025 */ 1026 void 1027 wakeup_domain(const volatile void *ident, int domain) 1028 { 1029 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(domain, mycpu->gd_cpuid)); 1030 } 1031 1032 /* 1033 * Wakeup one thread waiting on the specified ident that slept using 1034 * the specified domain, on any cpu. 1035 */ 1036 void 1037 wakeup_domain_one(const volatile void *ident, int domain) 1038 { 1039 /* XXX potentially round-robin the first responding cpu */ 1040 _wakeup(__DEALL(ident), 1041 PWAKEUP_ENCODE(domain, mycpu->gd_cpuid) | PWAKEUP_ONE); 1042 } 1043 1044 /* 1045 * setrunnable() 1046 * 1047 * Make a process runnable. The proc_token must be held on call. This only 1048 * has an effect if we are in SSLEEP. We only break out of the 1049 * tsleep if LWP_BREAKTSLEEP is set, otherwise we just fix-up the state. 1050 * 1051 * NOTE: With proc_token held we can only safely manipulate the process 1052 * structure and the lp's lwp_stat. 1053 */ 1054 void 1055 setrunnable(struct lwp *lp) 1056 { 1057 ASSERT_LWKT_TOKEN_HELD(&proc_token); 1058 crit_enter(); 1059 if (lp->lwp_stat == LSSTOP) 1060 lp->lwp_stat = LSSLEEP; 1061 if (lp->lwp_stat == LSSLEEP && (lp->lwp_flag & LWP_BREAKTSLEEP)) 1062 _tsleep_wakeup(lp->lwp_thread); 1063 crit_exit(); 1064 } 1065 1066 /* 1067 * The process is stopped due to some condition, usually because p_stat is 1068 * set to SSTOP, but also possibly due to being traced. 1069 * 1070 * NOTE! If the caller sets SSTOP, the caller must also clear P_WAITED 1071 * because the parent may check the child's status before the child actually 1072 * gets to this routine. 1073 * 1074 * This routine is called with the current lwp only, typically just 1075 * before returning to userland. 1076 * 1077 * Setting LWP_BREAKTSLEEP before entering the tsleep will cause a passive 1078 * SIGCONT to break out of the tsleep. 1079 */ 1080 void 1081 tstop(void) 1082 { 1083 struct lwp *lp = curthread->td_lwp; 1084 struct proc *p = lp->lwp_proc; 1085 1086 crit_enter(); 1087 /* 1088 * If LWP_WSTOP is set, we were sleeping 1089 * while our process was stopped. At this point 1090 * we were already counted as stopped. 1091 */ 1092 if ((lp->lwp_flag & LWP_WSTOP) == 0) { 1093 /* 1094 * If we're the last thread to stop, signal 1095 * our parent. 1096 */ 1097 p->p_nstopped++; 1098 lp->lwp_flag |= LWP_WSTOP; 1099 wakeup(&p->p_nstopped); 1100 if (p->p_nstopped == p->p_nthreads) { 1101 p->p_flag &= ~P_WAITED; 1102 wakeup(p->p_pptr); 1103 if ((p->p_pptr->p_sigacts->ps_flag & PS_NOCLDSTOP) == 0) 1104 ksignal(p->p_pptr, SIGCHLD); 1105 } 1106 } 1107 while (p->p_stat == SSTOP) { 1108 lp->lwp_flag |= LWP_BREAKTSLEEP; 1109 lp->lwp_stat = LSSTOP; 1110 tsleep(p, 0, "stop", 0); 1111 } 1112 p->p_nstopped--; 1113 lp->lwp_flag &= ~LWP_WSTOP; 1114 crit_exit(); 1115 } 1116 1117 /* 1118 * Compute a tenex style load average of a quantity on 1119 * 1, 5 and 15 minute intervals. 1120 */ 1121 static int loadav_count_runnable(struct lwp *p, void *data); 1122 1123 static void 1124 loadav(void *arg) 1125 { 1126 struct loadavg *avg; 1127 int i, nrun; 1128 1129 nrun = 0; 1130 alllwp_scan(loadav_count_runnable, &nrun); 1131 avg = &averunnable; 1132 for (i = 0; i < 3; i++) { 1133 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] + 1134 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT; 1135 } 1136 1137 /* 1138 * Schedule the next update to occur after 5 seconds, but add a 1139 * random variation to avoid synchronisation with processes that 1140 * run at regular intervals. 1141 */ 1142 callout_reset(&loadav_callout, hz * 4 + (int)(krandom() % (hz * 2 + 1)), 1143 loadav, NULL); 1144 } 1145 1146 static int 1147 loadav_count_runnable(struct lwp *lp, void *data) 1148 { 1149 int *nrunp = data; 1150 thread_t td; 1151 1152 switch (lp->lwp_stat) { 1153 case LSRUN: 1154 if ((td = lp->lwp_thread) == NULL) 1155 break; 1156 if (td->td_flags & TDF_BLOCKED) 1157 break; 1158 ++*nrunp; 1159 break; 1160 default: 1161 break; 1162 } 1163 return(0); 1164 } 1165 1166 /* ARGSUSED */ 1167 static void 1168 sched_setup(void *dummy) 1169 { 1170 callout_init(&loadav_callout); 1171 callout_init(&schedcpu_callout); 1172 1173 /* Kick off timeout driven events by calling first time. */ 1174 schedcpu(NULL); 1175 loadav(NULL); 1176 } 1177 1178