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 67 #include <machine/cpu.h> 68 #include <machine/smp.h> 69 70 TAILQ_HEAD(tslpque, thread); 71 72 static void sched_setup (void *dummy); 73 SYSINIT(sched_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, sched_setup, NULL) 74 75 int hogticks; 76 int lbolt; 77 int lbolt_syncer; 78 int sched_quantum; /* Roundrobin scheduling quantum in ticks. */ 79 int ncpus; 80 int ncpus2, ncpus2_shift, ncpus2_mask; /* note: mask not cpumask_t */ 81 int ncpus_fit, ncpus_fit_mask; /* note: mask not cpumask_t */ 82 int safepri; 83 int tsleep_now_works; 84 int tsleep_crypto_dump = 0; 85 86 static struct callout loadav_callout; 87 static struct callout schedcpu_callout; 88 MALLOC_DEFINE(M_TSLEEP, "tslpque", "tsleep queues"); 89 90 #define __DEALL(ident) __DEQUALIFY(void *, ident) 91 92 #if !defined(KTR_TSLEEP) 93 #define KTR_TSLEEP KTR_ALL 94 #endif 95 KTR_INFO_MASTER(tsleep); 96 KTR_INFO(KTR_TSLEEP, tsleep, tsleep_beg, 0, "tsleep enter %p", sizeof(void *)); 97 KTR_INFO(KTR_TSLEEP, tsleep, tsleep_end, 1, "tsleep exit", 0); 98 KTR_INFO(KTR_TSLEEP, tsleep, wakeup_beg, 2, "wakeup enter %p", sizeof(void *)); 99 KTR_INFO(KTR_TSLEEP, tsleep, wakeup_end, 3, "wakeup exit", 0); 100 KTR_INFO(KTR_TSLEEP, tsleep, ilockfail, 4, "interlock failed %p", sizeof(void *)); 101 102 #define logtsleep1(name) KTR_LOG(tsleep_ ## name) 103 #define logtsleep2(name, val) KTR_LOG(tsleep_ ## name, val) 104 105 struct loadavg averunnable = 106 { {0, 0, 0}, FSCALE }; /* load average, of runnable procs */ 107 /* 108 * Constants for averages over 1, 5, and 15 minutes 109 * when sampling at 5 second intervals. 110 */ 111 static fixpt_t cexp[3] = { 112 0.9200444146293232 * FSCALE, /* exp(-1/12) */ 113 0.9834714538216174 * FSCALE, /* exp(-1/60) */ 114 0.9944598480048967 * FSCALE, /* exp(-1/180) */ 115 }; 116 117 static void endtsleep (void *); 118 static void loadav (void *arg); 119 static void schedcpu (void *arg); 120 #ifdef SMP 121 static void tsleep_wakeup_remote(struct thread *td); 122 #endif 123 124 /* 125 * Adjust the scheduler quantum. The quantum is specified in microseconds. 126 * Note that 'tick' is in microseconds per tick. 127 */ 128 static int 129 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS) 130 { 131 int error, new_val; 132 133 new_val = sched_quantum * ustick; 134 error = sysctl_handle_int(oidp, &new_val, 0, req); 135 if (error != 0 || req->newptr == NULL) 136 return (error); 137 if (new_val < ustick) 138 return (EINVAL); 139 sched_quantum = new_val / ustick; 140 hogticks = 2 * sched_quantum; 141 return (0); 142 } 143 144 SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW, 145 0, sizeof sched_quantum, sysctl_kern_quantum, "I", ""); 146 147 /* 148 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the 149 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below 150 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT). 151 * 152 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used: 153 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits). 154 * 155 * If you don't want to bother with the faster/more-accurate formula, you 156 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate 157 * (more general) method of calculating the %age of CPU used by a process. 158 * 159 * decay 95% of `lwp_pctcpu' in 60 seconds; see CCPU_SHIFT before changing 160 */ 161 #define CCPU_SHIFT 11 162 163 static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */ 164 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 165 166 /* 167 * kernel uses `FSCALE', userland (SHOULD) use kern.fscale 168 */ 169 int fscale __unused = FSCALE; /* exported to systat */ 170 SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, ""); 171 172 /* 173 * Recompute process priorities, once a second. 174 * 175 * Since the userland schedulers are typically event oriented, if the 176 * estcpu calculation at wakeup() time is not sufficient to make a 177 * process runnable relative to other processes in the system we have 178 * a 1-second recalc to help out. 179 * 180 * This code also allows us to store sysclock_t data in the process structure 181 * without fear of an overrun, since sysclock_t are guarenteed to hold 182 * several seconds worth of count. 183 * 184 * WARNING! callouts can preempt normal threads. However, they will not 185 * preempt a thread holding a spinlock so we *can* safely use spinlocks. 186 */ 187 static int schedcpu_stats(struct proc *p, void *data __unused); 188 static int schedcpu_resource(struct proc *p, void *data __unused); 189 190 static void 191 schedcpu(void *arg) 192 { 193 allproc_scan(schedcpu_stats, NULL); 194 allproc_scan(schedcpu_resource, NULL); 195 wakeup((caddr_t)&lbolt); 196 wakeup((caddr_t)&lbolt_syncer); 197 callout_reset(&schedcpu_callout, hz, schedcpu, NULL); 198 } 199 200 /* 201 * General process statistics once a second 202 */ 203 static int 204 schedcpu_stats(struct proc *p, void *data __unused) 205 { 206 struct lwp *lp; 207 208 /* 209 * Threads may not be completely set up if process in SIDL state. 210 */ 211 if (p->p_stat == SIDL) 212 return(0); 213 214 PHOLD(p); 215 lwkt_gettoken(&p->p_token); 216 217 p->p_swtime++; 218 FOREACH_LWP_IN_PROC(lp, p) { 219 if (lp->lwp_stat == LSSLEEP) 220 lp->lwp_slptime++; 221 222 /* 223 * Only recalculate processes that are active or have slept 224 * less then 2 seconds. The schedulers understand this. 225 */ 226 if (lp->lwp_slptime <= 1) { 227 p->p_usched->recalculate(lp); 228 } else { 229 lp->lwp_pctcpu = (lp->lwp_pctcpu * ccpu) >> FSHIFT; 230 } 231 } 232 lwkt_reltoken(&p->p_token); 233 PRELE(p); 234 return(0); 235 } 236 237 /* 238 * Resource checks. XXX break out since ksignal/killproc can block, 239 * limiting us to one process killed per second. There is probably 240 * a better way. 241 */ 242 static int 243 schedcpu_resource(struct proc *p, void *data __unused) 244 { 245 u_int64_t ttime; 246 struct lwp *lp; 247 248 if (p->p_stat == SIDL) 249 return(0); 250 251 PHOLD(p); 252 lwkt_gettoken(&p->p_token); 253 254 if (p->p_stat == SZOMB || p->p_limit == NULL) { 255 lwkt_reltoken(&p->p_token); 256 PRELE(p); 257 return(0); 258 } 259 260 ttime = 0; 261 FOREACH_LWP_IN_PROC(lp, p) { 262 /* 263 * We may have caught an lp in the middle of being 264 * created, lwp_thread can be NULL. 265 */ 266 if (lp->lwp_thread) { 267 ttime += lp->lwp_thread->td_sticks; 268 ttime += lp->lwp_thread->td_uticks; 269 } 270 } 271 272 switch(plimit_testcpulimit(p->p_limit, ttime)) { 273 case PLIMIT_TESTCPU_KILL: 274 killproc(p, "exceeded maximum CPU limit"); 275 break; 276 case PLIMIT_TESTCPU_XCPU: 277 if ((p->p_flag & P_XCPU) == 0) { 278 p->p_flag |= P_XCPU; 279 ksignal(p, SIGXCPU); 280 } 281 break; 282 default: 283 break; 284 } 285 lwkt_reltoken(&p->p_token); 286 PRELE(p); 287 return(0); 288 } 289 290 /* 291 * This is only used by ps. Generate a cpu percentage use over 292 * a period of one second. 293 * 294 * MPSAFE 295 */ 296 void 297 updatepcpu(struct lwp *lp, int cpticks, int ttlticks) 298 { 299 fixpt_t acc; 300 int remticks; 301 302 acc = (cpticks << FSHIFT) / ttlticks; 303 if (ttlticks >= ESTCPUFREQ) { 304 lp->lwp_pctcpu = acc; 305 } else { 306 remticks = ESTCPUFREQ - ttlticks; 307 lp->lwp_pctcpu = (acc * ttlticks + lp->lwp_pctcpu * remticks) / 308 ESTCPUFREQ; 309 } 310 } 311 312 /* 313 * tsleep/wakeup hash table parameters. Try to find the sweet spot for 314 * like addresses being slept on. 315 */ 316 #define TABLESIZE 1024 317 #define LOOKUP(x) (((intptr_t)(x) >> 6) & (TABLESIZE - 1)) 318 319 static cpumask_t slpque_cpumasks[TABLESIZE]; 320 321 /* 322 * General scheduler initialization. We force a reschedule 25 times 323 * a second by default. Note that cpu0 is initialized in early boot and 324 * cannot make any high level calls. 325 * 326 * Each cpu has its own sleep queue. 327 */ 328 void 329 sleep_gdinit(globaldata_t gd) 330 { 331 static struct tslpque slpque_cpu0[TABLESIZE]; 332 int i; 333 334 if (gd->gd_cpuid == 0) { 335 sched_quantum = (hz + 24) / 25; 336 hogticks = 2 * sched_quantum; 337 338 gd->gd_tsleep_hash = slpque_cpu0; 339 } else { 340 gd->gd_tsleep_hash = kmalloc(sizeof(slpque_cpu0), 341 M_TSLEEP, M_WAITOK | M_ZERO); 342 } 343 for (i = 0; i < TABLESIZE; ++i) 344 TAILQ_INIT(&gd->gd_tsleep_hash[i]); 345 } 346 347 /* 348 * This is a dandy function that allows us to interlock tsleep/wakeup 349 * operations with unspecified upper level locks, such as lockmgr locks, 350 * simply by holding a critical section. The sequence is: 351 * 352 * (acquire upper level lock) 353 * tsleep_interlock(blah) 354 * (release upper level lock) 355 * tsleep(blah, ...) 356 * 357 * Basically this functions queues us on the tsleep queue without actually 358 * descheduling us. When tsleep() is later called with PINTERLOCK it 359 * assumes the thread was already queued, otherwise it queues it there. 360 * 361 * Thus it is possible to receive the wakeup prior to going to sleep and 362 * the race conditions are covered. 363 */ 364 static __inline void 365 _tsleep_interlock(globaldata_t gd, const volatile void *ident, int flags) 366 { 367 thread_t td = gd->gd_curthread; 368 int id; 369 370 crit_enter_quick(td); 371 if (td->td_flags & TDF_TSLEEPQ) { 372 id = LOOKUP(td->td_wchan); 373 TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_sleepq); 374 if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL) 375 atomic_clear_cpumask(&slpque_cpumasks[id], gd->gd_cpumask); 376 } else { 377 td->td_flags |= TDF_TSLEEPQ; 378 } 379 id = LOOKUP(ident); 380 TAILQ_INSERT_TAIL(&gd->gd_tsleep_hash[id], td, td_sleepq); 381 atomic_set_cpumask(&slpque_cpumasks[id], gd->gd_cpumask); 382 td->td_wchan = ident; 383 td->td_wdomain = flags & PDOMAIN_MASK; 384 crit_exit_quick(td); 385 } 386 387 void 388 tsleep_interlock(const volatile void *ident, int flags) 389 { 390 _tsleep_interlock(mycpu, ident, flags); 391 } 392 393 /* 394 * Remove thread from sleepq. Must be called with a critical section held. 395 */ 396 static __inline void 397 _tsleep_remove(thread_t td) 398 { 399 globaldata_t gd = mycpu; 400 int id; 401 402 KKASSERT(td->td_gd == gd); 403 if (td->td_flags & TDF_TSLEEPQ) { 404 td->td_flags &= ~TDF_TSLEEPQ; 405 id = LOOKUP(td->td_wchan); 406 TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_sleepq); 407 if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL) 408 atomic_clear_cpumask(&slpque_cpumasks[id], gd->gd_cpumask); 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 * This function removes a thread from the tsleep queue and schedules 422 * it. This function may act asynchronously. The target thread may be 423 * sleeping on a different cpu. 424 * 425 * This function mus be called while in a critical section but if the 426 * target thread is sleeping on a different cpu we cannot safely probe 427 * td_flags. 428 * 429 * This function is only called from a different cpu via setrunnable() 430 * when the thread is in a known sleep. However, multiple wakeups are 431 * possible and we must hold the td to prevent a race against the thread 432 * exiting. 433 */ 434 static __inline 435 void 436 _tsleep_wakeup(struct thread *td) 437 { 438 #ifdef SMP 439 globaldata_t gd = mycpu; 440 441 if (td->td_gd != gd) { 442 lwkt_hold(td); 443 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)tsleep_wakeup_remote, td); 444 return; 445 } 446 #endif 447 _tsleep_remove(td); 448 if (td->td_flags & TDF_TSLEEP_DESCHEDULED) { 449 td->td_flags &= ~TDF_TSLEEP_DESCHEDULED; 450 lwkt_schedule(td); 451 } 452 } 453 454 #ifdef SMP 455 static 456 void 457 tsleep_wakeup_remote(struct thread *td) 458 { 459 _tsleep_wakeup(td); 460 lwkt_rele(td); 461 } 462 #endif 463 464 465 /* 466 * General sleep call. Suspends the current process until a wakeup is 467 * performed on the specified identifier. The process will then be made 468 * runnable with the specified priority. Sleeps at most timo/hz seconds 469 * (0 means no timeout). If flags includes PCATCH flag, signals are checked 470 * before and after sleeping, else signals are not checked. Returns 0 if 471 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a 472 * signal needs to be delivered, ERESTART is returned if the current system 473 * call should be restarted if possible, and EINTR is returned if the system 474 * call should be interrupted by the signal (return EINTR). 475 * 476 * Note that if we are a process, we release_curproc() before messing with 477 * the LWKT scheduler. 478 * 479 * During autoconfiguration or after a panic, a sleep will simply 480 * lower the priority briefly to allow interrupts, then return. 481 */ 482 int 483 tsleep(const volatile void *ident, int flags, const char *wmesg, int timo) 484 { 485 struct thread *td = curthread; 486 struct lwp *lp = td->td_lwp; 487 struct proc *p = td->td_proc; /* may be NULL */ 488 globaldata_t gd; 489 int sig; 490 int catch; 491 int id; 492 int error; 493 int oldpri; 494 struct callout thandle; 495 496 /* 497 * NOTE: removed KTRPOINT, it could cause races due to blocking 498 * even in stable. Just scrap it for now. 499 */ 500 if (!tsleep_crypto_dump && (tsleep_now_works == 0 || panicstr)) { 501 /* 502 * After a panic, or before we actually have an operational 503 * softclock, just give interrupts a chance, then just return; 504 * 505 * don't run any other procs or panic below, 506 * in case this is the idle process and already asleep. 507 */ 508 splz(); 509 oldpri = td->td_pri; 510 lwkt_setpri_self(safepri); 511 lwkt_switch(); 512 lwkt_setpri_self(oldpri); 513 return (0); 514 } 515 logtsleep2(tsleep_beg, ident); 516 gd = td->td_gd; 517 KKASSERT(td != &gd->gd_idlethread); /* you must be kidding! */ 518 519 /* 520 * NOTE: all of this occurs on the current cpu, including any 521 * callout-based wakeups, so a critical section is a sufficient 522 * interlock. 523 * 524 * The entire sequence through to where we actually sleep must 525 * run without breaking the critical section. 526 */ 527 catch = flags & PCATCH; 528 error = 0; 529 sig = 0; 530 531 crit_enter_quick(td); 532 533 KASSERT(ident != NULL, ("tsleep: no ident")); 534 KASSERT(lp == NULL || 535 lp->lwp_stat == LSRUN || /* Obvious */ 536 lp->lwp_stat == LSSTOP, /* Set in tstop */ 537 ("tsleep %p %s %d", 538 ident, wmesg, lp->lwp_stat)); 539 540 /* 541 * We interlock the sleep queue if the caller has not already done 542 * it for us. This must be done before we potentially acquire any 543 * tokens or we can loose the wakeup. 544 */ 545 if ((flags & PINTERLOCKED) == 0) { 546 id = LOOKUP(ident); 547 _tsleep_interlock(gd, ident, flags); 548 } 549 550 /* 551 * Setup for the current process (if this is a process). 552 * 553 * We hold the process token if lp && catch. The resume 554 * code will release it. 555 */ 556 if (lp) { 557 if (catch) { 558 /* 559 * Early termination if PCATCH was set and a 560 * signal is pending, interlocked with the 561 * critical section. 562 * 563 * Early termination only occurs when tsleep() is 564 * entered while in a normal LSRUN state. 565 */ 566 lwkt_gettoken(&p->p_token); 567 if ((sig = CURSIG(lp)) != 0) 568 goto resume; 569 570 /* 571 * Early termination if PCATCH was set and a 572 * mailbox signal was possibly delivered prior to 573 * the system call even being made, in order to 574 * allow the user to interlock without having to 575 * make additional system calls. 576 */ 577 if (p->p_flag & P_MAILBOX) 578 goto resume; 579 580 /* 581 * Causes ksignal to wake us up if a signal is 582 * received (interlocked with p->p_token). 583 */ 584 lp->lwp_flag |= LWP_SINTR; 585 } 586 } else { 587 KKASSERT(p == NULL); 588 } 589 590 /* 591 * Make sure the current process has been untangled from 592 * the userland scheduler and initialize slptime to start 593 * counting. 594 */ 595 if (lp) { 596 p->p_usched->release_curproc(lp); 597 lp->lwp_slptime = 0; 598 } 599 600 /* 601 * If the interlocked flag is set but our cpu bit in the slpqueue 602 * is no longer set, then a wakeup was processed inbetween the 603 * tsleep_interlock() (ours or the callers), and here. This can 604 * occur under numerous circumstances including when we release the 605 * current process. 606 * 607 * Extreme loads can cause the sending of an IPI (e.g. wakeup()'s) 608 * to process incoming IPIs, thus draining incoming wakeups. 609 */ 610 if ((td->td_flags & TDF_TSLEEPQ) == 0) { 611 logtsleep2(ilockfail, ident); 612 goto resume; 613 } 614 615 /* 616 * scheduling is blocked while in a critical section. Coincide 617 * the descheduled-by-tsleep flag with the descheduling of the 618 * lwkt. 619 * 620 * The timer callout is localized on our cpu and interlocked by 621 * our critical section. 622 */ 623 lwkt_deschedule_self(td); 624 td->td_flags |= TDF_TSLEEP_DESCHEDULED; 625 td->td_wmesg = wmesg; 626 627 /* 628 * Setup the timeout, if any. The timeout is only operable while 629 * the thread is flagged descheduled. 630 */ 631 KKASSERT((td->td_flags & TDF_TIMEOUT) == 0); 632 if (timo) { 633 callout_init_mp(&thandle); 634 callout_reset(&thandle, timo, endtsleep, td); 635 } 636 637 /* 638 * Beddy bye bye. 639 */ 640 if (lp) { 641 /* 642 * Ok, we are sleeping. Place us in the SSLEEP state. 643 */ 644 KKASSERT((lp->lwp_flag & LWP_ONRUNQ) == 0); 645 /* 646 * tstop() sets LSSTOP, so don't fiddle with that. 647 */ 648 if (lp->lwp_stat != LSSTOP) 649 lp->lwp_stat = LSSLEEP; 650 lp->lwp_ru.ru_nvcsw++; 651 lwkt_switch(); 652 td->td_flags &= ~TDF_TSLEEP_DESCHEDULED; 653 654 /* 655 * And when we are woken up, put us back in LSRUN. If we 656 * slept for over a second, recalculate our estcpu. 657 */ 658 lp->lwp_stat = LSRUN; 659 if (lp->lwp_slptime) 660 p->p_usched->recalculate(lp); 661 lp->lwp_slptime = 0; 662 } else { 663 lwkt_switch(); 664 td->td_flags &= ~TDF_TSLEEP_DESCHEDULED; 665 } 666 667 /* 668 * Make sure we haven't switched cpus while we were asleep. It's 669 * not supposed to happen. Cleanup our temporary flags. 670 */ 671 KKASSERT(gd == td->td_gd); 672 673 /* 674 * Cleanup the timeout. If the timeout has already occured thandle 675 * has already been stopped, otherwise stop thandle. 676 */ 677 if (timo) { 678 if (td->td_flags & TDF_TIMEOUT) { 679 td->td_flags &= ~TDF_TIMEOUT; 680 error = EWOULDBLOCK; 681 } else { 682 /* does not block when on same cpu */ 683 callout_stop(&thandle); 684 } 685 } 686 687 /* 688 * Make sure we have been removed from the sleepq. In most 689 * cases this will have been done for us already but it is 690 * possible for a scheduling IPI to be in-flight from a 691 * previous tsleep/tsleep_interlock() or due to a straight-out 692 * call to lwkt_schedule() (in the case of an interrupt thread), 693 * causing a spurious wakeup. 694 */ 695 _tsleep_remove(td); 696 td->td_wmesg = NULL; 697 698 /* 699 * Figure out the correct error return. If interrupted by a 700 * signal we want to return EINTR or ERESTART. 701 * 702 * If P_MAILBOX is set no automatic system call restart occurs 703 * and we return EINTR. P_MAILBOX is meant to be used as an 704 * interlock, the user must poll it prior to any system call 705 * that it wishes to interlock a mailbox signal against since 706 * the flag is cleared on *any* system call that sleeps. 707 * 708 * p->p_token is held in the p && catch case. 709 */ 710 resume: 711 if (p) { 712 if (catch && error == 0) { 713 if ((p->p_flag & P_MAILBOX) && sig == 0) { 714 error = EINTR; 715 } else if (sig != 0 || (sig = CURSIG(lp))) { 716 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 717 error = EINTR; 718 else 719 error = ERESTART; 720 } 721 } 722 if (catch) 723 lwkt_reltoken(&p->p_token); 724 lp->lwp_flag &= ~(LWP_BREAKTSLEEP | LWP_SINTR); 725 p->p_flag &= ~P_MAILBOX; 726 } 727 logtsleep1(tsleep_end); 728 crit_exit_quick(td); 729 return (error); 730 } 731 732 /* 733 * Interlocked spinlock sleep. An exclusively held spinlock must 734 * be passed to ssleep(). The function will atomically release the 735 * spinlock and tsleep on the ident, then reacquire the spinlock and 736 * return. 737 * 738 * This routine is fairly important along the critical path, so optimize it 739 * heavily. 740 */ 741 int 742 ssleep(const volatile void *ident, struct spinlock *spin, int flags, 743 const char *wmesg, int timo) 744 { 745 globaldata_t gd = mycpu; 746 int error; 747 748 _tsleep_interlock(gd, ident, flags); 749 spin_unlock_quick(gd, spin); 750 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 751 spin_lock_quick(gd, spin); 752 753 return (error); 754 } 755 756 int 757 lksleep(const volatile void *ident, struct lock *lock, int flags, 758 const char *wmesg, int timo) 759 { 760 globaldata_t gd = mycpu; 761 int error; 762 763 _tsleep_interlock(gd, ident, flags); 764 lockmgr(lock, LK_RELEASE); 765 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 766 lockmgr(lock, LK_EXCLUSIVE); 767 768 return (error); 769 } 770 771 /* 772 * Interlocked mutex sleep. An exclusively held mutex must be passed 773 * to mtxsleep(). The function will atomically release the mutex 774 * and tsleep on the ident, then reacquire the mutex and return. 775 */ 776 int 777 mtxsleep(const volatile void *ident, struct mtx *mtx, int flags, 778 const char *wmesg, int timo) 779 { 780 globaldata_t gd = mycpu; 781 int error; 782 783 _tsleep_interlock(gd, ident, flags); 784 mtx_unlock(mtx); 785 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 786 mtx_lock_ex_quick(mtx, wmesg); 787 788 return (error); 789 } 790 791 /* 792 * Interlocked serializer sleep. An exclusively held serializer must 793 * be passed to zsleep(). The function will atomically release 794 * the serializer and tsleep on the ident, then reacquire the serializer 795 * and return. 796 */ 797 int 798 zsleep(const volatile void *ident, struct lwkt_serialize *slz, int flags, 799 const char *wmesg, int timo) 800 { 801 globaldata_t gd = mycpu; 802 int ret; 803 804 ASSERT_SERIALIZED(slz); 805 806 _tsleep_interlock(gd, ident, flags); 807 lwkt_serialize_exit(slz); 808 ret = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 809 lwkt_serialize_enter(slz); 810 811 return ret; 812 } 813 814 /* 815 * Directly block on the LWKT thread by descheduling it. This 816 * is much faster then tsleep(), but the only legal way to wake 817 * us up is to directly schedule the thread. 818 * 819 * Setting TDF_SINTR will cause new signals to directly schedule us. 820 * 821 * This routine must be called while in a critical section. 822 */ 823 int 824 lwkt_sleep(const char *wmesg, int flags) 825 { 826 thread_t td = curthread; 827 int sig; 828 829 if ((flags & PCATCH) == 0 || td->td_lwp == NULL) { 830 td->td_flags |= TDF_BLOCKED; 831 td->td_wmesg = wmesg; 832 lwkt_deschedule_self(td); 833 lwkt_switch(); 834 td->td_wmesg = NULL; 835 td->td_flags &= ~TDF_BLOCKED; 836 return(0); 837 } 838 if ((sig = CURSIG(td->td_lwp)) != 0) { 839 if (SIGISMEMBER(td->td_proc->p_sigacts->ps_sigintr, sig)) 840 return(EINTR); 841 else 842 return(ERESTART); 843 844 } 845 td->td_flags |= TDF_BLOCKED | TDF_SINTR; 846 td->td_wmesg = wmesg; 847 lwkt_deschedule_self(td); 848 lwkt_switch(); 849 td->td_flags &= ~(TDF_BLOCKED | TDF_SINTR); 850 td->td_wmesg = NULL; 851 return(0); 852 } 853 854 /* 855 * Implement the timeout for tsleep. 856 * 857 * We set LWP_BREAKTSLEEP to indicate that an event has occured, but 858 * we only call setrunnable if the process is not stopped. 859 * 860 * This type of callout timeout is scheduled on the same cpu the process 861 * is sleeping on. Also, at the moment, the MP lock is held. 862 */ 863 static void 864 endtsleep(void *arg) 865 { 866 thread_t td = arg; 867 struct lwp *lp; 868 869 crit_enter(); 870 871 /* 872 * Do this before we potentially block acquiring the token. Setting 873 * TDF_TIMEOUT tells tsleep that we have already stopped the callout. 874 */ 875 lwkt_hold(td); 876 td->td_flags |= TDF_TIMEOUT; 877 878 /* 879 * This can block 880 */ 881 if ((lp = td->td_lwp) != NULL) 882 lwkt_gettoken(&lp->lwp_proc->p_token); 883 884 /* 885 * Only do nominal wakeup processing if TDF_TIMEOUT and 886 * TDF_TSLEEP_DESCHEDULED are both still set. Otherwise 887 * we raced a wakeup or we began executed and raced due to 888 * blocking in the token above, and should do nothing. 889 */ 890 if ((td->td_flags & (TDF_TIMEOUT | TDF_TSLEEP_DESCHEDULED)) == 891 (TDF_TIMEOUT | TDF_TSLEEP_DESCHEDULED)) { 892 if (lp) { 893 lp->lwp_flag |= LWP_BREAKTSLEEP; 894 if (lp->lwp_proc->p_stat != SSTOP) 895 setrunnable(lp); 896 } else { 897 _tsleep_wakeup(td); 898 } 899 } 900 if (lp) 901 lwkt_reltoken(&lp->lwp_proc->p_token); 902 lwkt_rele(td); 903 crit_exit(); 904 } 905 906 /* 907 * Make all processes sleeping on the specified identifier runnable. 908 * count may be zero or one only. 909 * 910 * The domain encodes the sleep/wakeup domain AND the first cpu to check 911 * (which is always the current cpu). As we iterate across cpus 912 * 913 * This call may run without the MP lock held. We can only manipulate thread 914 * state on the cpu owning the thread. We CANNOT manipulate process state 915 * at all. 916 * 917 * _wakeup() can be passed to an IPI so we can't use (const volatile 918 * void *ident). 919 */ 920 static void 921 _wakeup(void *ident, int domain) 922 { 923 struct tslpque *qp; 924 struct thread *td; 925 struct thread *ntd; 926 globaldata_t gd; 927 #ifdef SMP 928 cpumask_t mask; 929 #endif 930 int id; 931 932 crit_enter(); 933 logtsleep2(wakeup_beg, ident); 934 gd = mycpu; 935 id = LOOKUP(ident); 936 qp = &gd->gd_tsleep_hash[id]; 937 restart: 938 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) { 939 ntd = TAILQ_NEXT(td, td_sleepq); 940 if (td->td_wchan == ident && 941 td->td_wdomain == (domain & PDOMAIN_MASK) 942 ) { 943 KKASSERT(td->td_gd == gd); 944 _tsleep_remove(td); 945 if (td->td_flags & TDF_TSLEEP_DESCHEDULED) { 946 td->td_flags &= ~TDF_TSLEEP_DESCHEDULED; 947 lwkt_schedule(td); 948 if (domain & PWAKEUP_ONE) 949 goto done; 950 } 951 goto restart; 952 } 953 } 954 955 #ifdef SMP 956 /* 957 * We finished checking the current cpu but there still may be 958 * more work to do. Either wakeup_one was requested and no matching 959 * thread was found, or a normal wakeup was requested and we have 960 * to continue checking cpus. 961 * 962 * It should be noted that this scheme is actually less expensive then 963 * the old scheme when waking up multiple threads, since we send 964 * only one IPI message per target candidate which may then schedule 965 * multiple threads. Before we could have wound up sending an IPI 966 * message for each thread on the target cpu (!= current cpu) that 967 * needed to be woken up. 968 * 969 * NOTE: Wakeups occuring on remote cpus are asynchronous. This 970 * should be ok since we are passing idents in the IPI rather then 971 * thread pointers. 972 */ 973 if ((domain & PWAKEUP_MYCPU) == 0 && 974 (mask = slpque_cpumasks[id] & gd->gd_other_cpus) != 0) { 975 lwkt_send_ipiq2_mask(mask, _wakeup, ident, 976 domain | PWAKEUP_MYCPU); 977 } 978 #endif 979 done: 980 logtsleep1(wakeup_end); 981 crit_exit(); 982 } 983 984 /* 985 * Wakeup all threads tsleep()ing on the specified ident, on all cpus 986 */ 987 void 988 wakeup(const volatile void *ident) 989 { 990 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mycpu->gd_cpuid)); 991 } 992 993 /* 994 * Wakeup one thread tsleep()ing on the specified ident, on any cpu. 995 */ 996 void 997 wakeup_one(const volatile void *ident) 998 { 999 /* XXX potentially round-robin the first responding cpu */ 1000 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mycpu->gd_cpuid) | PWAKEUP_ONE); 1001 } 1002 1003 /* 1004 * Wakeup threads tsleep()ing on the specified ident on the current cpu 1005 * only. 1006 */ 1007 void 1008 wakeup_mycpu(const volatile void *ident) 1009 { 1010 _wakeup(__DEALL(ident), PWAKEUP_MYCPU); 1011 } 1012 1013 /* 1014 * Wakeup one thread tsleep()ing on the specified ident on the current cpu 1015 * only. 1016 */ 1017 void 1018 wakeup_mycpu_one(const volatile void *ident) 1019 { 1020 /* XXX potentially round-robin the first responding cpu */ 1021 _wakeup(__DEALL(ident), PWAKEUP_MYCPU|PWAKEUP_ONE); 1022 } 1023 1024 /* 1025 * Wakeup all thread tsleep()ing on the specified ident on the specified cpu 1026 * only. 1027 */ 1028 void 1029 wakeup_oncpu(globaldata_t gd, const volatile void *ident) 1030 { 1031 #ifdef SMP 1032 if (gd == mycpu) { 1033 _wakeup(__DEALL(ident), PWAKEUP_MYCPU); 1034 } else { 1035 lwkt_send_ipiq2(gd, _wakeup, __DEALL(ident), PWAKEUP_MYCPU); 1036 } 1037 #else 1038 _wakeup(__DEALL(ident), PWAKEUP_MYCPU); 1039 #endif 1040 } 1041 1042 /* 1043 * Wakeup one thread tsleep()ing on the specified ident on the specified cpu 1044 * only. 1045 */ 1046 void 1047 wakeup_oncpu_one(globaldata_t gd, const volatile void *ident) 1048 { 1049 #ifdef SMP 1050 if (gd == mycpu) { 1051 _wakeup(__DEALL(ident), PWAKEUP_MYCPU | PWAKEUP_ONE); 1052 } else { 1053 lwkt_send_ipiq2(gd, _wakeup, __DEALL(ident), 1054 PWAKEUP_MYCPU | PWAKEUP_ONE); 1055 } 1056 #else 1057 _wakeup(__DEALL(ident), PWAKEUP_MYCPU | PWAKEUP_ONE); 1058 #endif 1059 } 1060 1061 /* 1062 * Wakeup all threads waiting on the specified ident that slept using 1063 * the specified domain, on all cpus. 1064 */ 1065 void 1066 wakeup_domain(const volatile void *ident, int domain) 1067 { 1068 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(domain, mycpu->gd_cpuid)); 1069 } 1070 1071 /* 1072 * Wakeup one thread waiting on the specified ident that slept using 1073 * the specified domain, on any cpu. 1074 */ 1075 void 1076 wakeup_domain_one(const volatile void *ident, int domain) 1077 { 1078 /* XXX potentially round-robin the first responding cpu */ 1079 _wakeup(__DEALL(ident), 1080 PWAKEUP_ENCODE(domain, mycpu->gd_cpuid) | PWAKEUP_ONE); 1081 } 1082 1083 /* 1084 * setrunnable() 1085 * 1086 * Make a process runnable. lp->lwp_proc->p_token must be held on call. 1087 * This only has an effect if we are in SSLEEP. We only break out of the 1088 * tsleep if LWP_BREAKTSLEEP is set, otherwise we just fix-up the state. 1089 * 1090 * NOTE: With p_token held we can only safely manipulate the process 1091 * structure and the lp's lwp_stat. 1092 */ 1093 void 1094 setrunnable(struct lwp *lp) 1095 { 1096 ASSERT_LWKT_TOKEN_HELD(&lp->lwp_proc->p_token); 1097 crit_enter(); 1098 if (lp->lwp_stat == LSSTOP) 1099 lp->lwp_stat = LSSLEEP; 1100 if (lp->lwp_stat == LSSLEEP && (lp->lwp_flag & LWP_BREAKTSLEEP)) 1101 _tsleep_wakeup(lp->lwp_thread); 1102 crit_exit(); 1103 } 1104 1105 /* 1106 * The process is stopped due to some condition, usually because p_stat is 1107 * set to SSTOP, but also possibly due to being traced. 1108 * 1109 * NOTE! If the caller sets SSTOP, the caller must also clear P_WAITED 1110 * because the parent may check the child's status before the child actually 1111 * gets to this routine. 1112 * 1113 * This routine is called with the current lwp only, typically just 1114 * before returning to userland. 1115 * 1116 * Setting LWP_BREAKTSLEEP before entering the tsleep will cause a passive 1117 * SIGCONT to break out of the tsleep. 1118 */ 1119 void 1120 tstop(void) 1121 { 1122 struct lwp *lp = curthread->td_lwp; 1123 struct proc *p = lp->lwp_proc; 1124 1125 crit_enter(); 1126 /* 1127 * If LWP_WSTOP is set, we were sleeping 1128 * while our process was stopped. At this point 1129 * we were already counted as stopped. 1130 */ 1131 if ((lp->lwp_flag & LWP_WSTOP) == 0) { 1132 /* 1133 * If we're the last thread to stop, signal 1134 * our parent. 1135 */ 1136 p->p_nstopped++; 1137 lp->lwp_flag |= LWP_WSTOP; 1138 wakeup(&p->p_nstopped); 1139 if (p->p_nstopped == p->p_nthreads) { 1140 p->p_flag &= ~P_WAITED; 1141 wakeup(p->p_pptr); 1142 if ((p->p_pptr->p_sigacts->ps_flag & PS_NOCLDSTOP) == 0) 1143 ksignal(p->p_pptr, SIGCHLD); 1144 } 1145 } 1146 while (p->p_stat == SSTOP) { 1147 lp->lwp_flag |= LWP_BREAKTSLEEP; 1148 lp->lwp_stat = LSSTOP; 1149 tsleep(p, 0, "stop", 0); 1150 } 1151 p->p_nstopped--; 1152 lp->lwp_flag &= ~LWP_WSTOP; 1153 crit_exit(); 1154 } 1155 1156 /* 1157 * Compute a tenex style load average of a quantity on 1158 * 1, 5 and 15 minute intervals. 1159 */ 1160 static int loadav_count_runnable(struct lwp *p, void *data); 1161 1162 static void 1163 loadav(void *arg) 1164 { 1165 struct loadavg *avg; 1166 int i, nrun; 1167 1168 nrun = 0; 1169 alllwp_scan(loadav_count_runnable, &nrun); 1170 avg = &averunnable; 1171 for (i = 0; i < 3; i++) { 1172 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] + 1173 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT; 1174 } 1175 1176 /* 1177 * Schedule the next update to occur after 5 seconds, but add a 1178 * random variation to avoid synchronisation with processes that 1179 * run at regular intervals. 1180 */ 1181 callout_reset(&loadav_callout, hz * 4 + (int)(krandom() % (hz * 2 + 1)), 1182 loadav, NULL); 1183 } 1184 1185 static int 1186 loadav_count_runnable(struct lwp *lp, void *data) 1187 { 1188 int *nrunp = data; 1189 thread_t td; 1190 1191 switch (lp->lwp_stat) { 1192 case LSRUN: 1193 if ((td = lp->lwp_thread) == NULL) 1194 break; 1195 if (td->td_flags & TDF_BLOCKED) 1196 break; 1197 ++*nrunp; 1198 break; 1199 default: 1200 break; 1201 } 1202 return(0); 1203 } 1204 1205 /* ARGSUSED */ 1206 static void 1207 sched_setup(void *dummy) 1208 { 1209 callout_init_mp(&loadav_callout); 1210 callout_init_mp(&schedcpu_callout); 1211 1212 /* Kick off timeout driven events by calling first time. */ 1213 schedcpu(NULL); 1214 loadav(NULL); 1215 } 1216 1217