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 /* 210 * Threads may not be completely set up if process in SIDL state. 211 */ 212 if (p->p_stat == SIDL) 213 return(0); 214 215 crit_enter(); 216 p->p_swtime++; 217 FOREACH_LWP_IN_PROC(lp, p) { 218 if (lp->lwp_stat == LSSLEEP) 219 lp->lwp_slptime++; 220 221 /* 222 * Only recalculate processes that are active or have slept 223 * less then 2 seconds. The schedulers understand this. 224 */ 225 if (lp->lwp_slptime <= 1) { 226 p->p_usched->recalculate(lp); 227 } else { 228 lp->lwp_pctcpu = (lp->lwp_pctcpu * ccpu) >> FSHIFT; 229 } 230 } 231 crit_exit(); 232 return(0); 233 } 234 235 /* 236 * Resource checks. XXX break out since ksignal/killproc can block, 237 * limiting us to one process killed per second. There is probably 238 * a better way. 239 */ 240 static int 241 schedcpu_resource(struct proc *p, void *data __unused) 242 { 243 u_int64_t ttime; 244 struct lwp *lp; 245 246 crit_enter(); 247 if (p->p_stat == SIDL || 248 p->p_stat == SZOMB || 249 p->p_limit == NULL 250 ) { 251 crit_exit(); 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_flag & P_XCPU) == 0) { 273 p->p_flag |= P_XCPU; 274 ksignal(p, SIGXCPU); 275 } 276 break; 277 default: 278 break; 279 } 280 crit_exit(); 281 return(0); 282 } 283 284 /* 285 * This is only used by ps. Generate a cpu percentage use over 286 * a period of one second. 287 * 288 * MPSAFE 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 1024 311 #define LOOKUP(x) (((intptr_t)(x) >> 6) & (TABLESIZE - 1)) 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], gd->gd_cpumask); 370 } else { 371 td->td_flags |= TDF_TSLEEPQ; 372 } 373 id = LOOKUP(ident); 374 TAILQ_INSERT_TAIL(&gd->gd_tsleep_hash[id], td, td_sleepq); 375 atomic_set_cpumask(&slpque_cpumasks[id], gd->gd_cpumask); 376 td->td_wchan = ident; 377 td->td_wdomain = flags & PDOMAIN_MASK; 378 crit_exit_quick(td); 379 } 380 381 void 382 tsleep_interlock(const volatile void *ident, int flags) 383 { 384 _tsleep_interlock(mycpu, ident, flags); 385 } 386 387 /* 388 * Remove thread from sleepq. Must be called with a critical section held. 389 */ 390 static __inline void 391 _tsleep_remove(thread_t td) 392 { 393 globaldata_t gd = mycpu; 394 int id; 395 396 KKASSERT(td->td_gd == gd); 397 if (td->td_flags & TDF_TSLEEPQ) { 398 td->td_flags &= ~TDF_TSLEEPQ; 399 id = LOOKUP(td->td_wchan); 400 TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_sleepq); 401 if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL) 402 atomic_clear_cpumask(&slpque_cpumasks[id], gd->gd_cpumask); 403 td->td_wchan = NULL; 404 td->td_wdomain = 0; 405 } 406 } 407 408 void 409 tsleep_remove(thread_t td) 410 { 411 _tsleep_remove(td); 412 } 413 414 /* 415 * This function removes a thread from the tsleep queue and schedules 416 * it. This function may act asynchronously. The target thread may be 417 * sleeping on a different cpu. 418 * 419 * This function mus be called while in a critical section but if the 420 * target thread is sleeping on a different cpu we cannot safely probe 421 * td_flags. 422 * 423 * This function is only called from a different cpu via setrunnable() 424 * when the thread is in a known sleep. However, multiple wakeups are 425 * possible and we must hold the td to prevent a race against the thread 426 * exiting. 427 */ 428 static __inline 429 void 430 _tsleep_wakeup(struct thread *td) 431 { 432 #ifdef SMP 433 globaldata_t gd = mycpu; 434 435 if (td->td_gd != gd) { 436 lwkt_hold(td); 437 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)tsleep_wakeup_remote, td); 438 return; 439 } 440 #endif 441 _tsleep_remove(td); 442 if (td->td_flags & TDF_TSLEEP_DESCHEDULED) { 443 td->td_flags &= ~TDF_TSLEEP_DESCHEDULED; 444 lwkt_schedule(td); 445 } 446 } 447 448 #ifdef SMP 449 static 450 void 451 tsleep_wakeup_remote(struct thread *td) 452 { 453 _tsleep_wakeup(td); 454 lwkt_rele(td); 455 } 456 #endif 457 458 459 /* 460 * General sleep call. Suspends the current process until a wakeup is 461 * performed on the specified identifier. The process will then be made 462 * runnable with the specified priority. Sleeps at most timo/hz seconds 463 * (0 means no timeout). If flags includes PCATCH flag, signals are checked 464 * before and after sleeping, else signals are not checked. Returns 0 if 465 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a 466 * signal needs to be delivered, ERESTART is returned if the current system 467 * call should be restarted if possible, and EINTR is returned if the system 468 * call should be interrupted by the signal (return EINTR). 469 * 470 * Note that if we are a process, we release_curproc() before messing with 471 * the LWKT scheduler. 472 * 473 * During autoconfiguration or after a panic, a sleep will simply 474 * lower the priority briefly to allow interrupts, then return. 475 */ 476 int 477 tsleep(const volatile void *ident, int flags, const char *wmesg, int timo) 478 { 479 struct thread *td = curthread; 480 struct lwp *lp = td->td_lwp; 481 struct proc *p = td->td_proc; /* may be NULL */ 482 globaldata_t gd; 483 int sig; 484 int catch; 485 int id; 486 int error; 487 int oldpri; 488 struct callout thandle; 489 490 /* 491 * NOTE: removed KTRPOINT, it could cause races due to blocking 492 * even in stable. Just scrap it for now. 493 */ 494 if (!tsleep_crypto_dump && (tsleep_now_works == 0 || panicstr)) { 495 /* 496 * After a panic, or before we actually have an operational 497 * softclock, just give interrupts a chance, then just return; 498 * 499 * don't run any other procs or panic below, 500 * in case this is the idle process and already asleep. 501 */ 502 splz(); 503 oldpri = td->td_pri; 504 lwkt_setpri_self(safepri); 505 lwkt_switch(); 506 lwkt_setpri_self(oldpri); 507 return (0); 508 } 509 logtsleep2(tsleep_beg, ident); 510 gd = td->td_gd; 511 KKASSERT(td != &gd->gd_idlethread); /* you must be kidding! */ 512 513 /* 514 * NOTE: all of this occurs on the current cpu, including any 515 * callout-based wakeups, so a critical section is a sufficient 516 * interlock. 517 * 518 * The entire sequence through to where we actually sleep must 519 * run without breaking the critical section. 520 */ 521 catch = flags & PCATCH; 522 error = 0; 523 sig = 0; 524 525 crit_enter_quick(td); 526 527 KASSERT(ident != NULL, ("tsleep: no ident")); 528 KASSERT(lp == NULL || 529 lp->lwp_stat == LSRUN || /* Obvious */ 530 lp->lwp_stat == LSSTOP, /* Set in tstop */ 531 ("tsleep %p %s %d", 532 ident, wmesg, lp->lwp_stat)); 533 534 /* 535 * We interlock the sleep queue if the caller has not already done 536 * it for us. This must be done before we potentially acquire any 537 * tokens or we can loose the wakeup. 538 */ 539 if ((flags & PINTERLOCKED) == 0) { 540 id = LOOKUP(ident); 541 _tsleep_interlock(gd, ident, flags); 542 } 543 544 /* 545 * Setup for the current process (if this is a process). 546 * 547 * We hold the process token if lp && catch. The resume 548 * code will release it. 549 */ 550 if (lp) { 551 if (catch) { 552 /* 553 * Early termination if PCATCH was set and a 554 * signal is pending, interlocked with the 555 * critical section. 556 * 557 * Early termination only occurs when tsleep() is 558 * entered while in a normal LSRUN state. 559 */ 560 lwkt_gettoken(&p->p_token); 561 if ((sig = CURSIG(lp)) != 0) 562 goto resume; 563 564 /* 565 * Early termination if PCATCH was set and a 566 * mailbox signal was possibly delivered prior to 567 * the system call even being made, in order to 568 * allow the user to interlock without having to 569 * make additional system calls. 570 */ 571 if (p->p_flag & P_MAILBOX) 572 goto resume; 573 574 /* 575 * Causes ksignal to wake us up if a signal is 576 * received (interlocked with p->p_token). 577 */ 578 lp->lwp_flag |= LWP_SINTR; 579 } 580 } else { 581 KKASSERT(p == NULL); 582 } 583 584 /* 585 * Make sure the current process has been untangled from 586 * the userland scheduler and initialize slptime to start 587 * counting. 588 */ 589 if (lp) { 590 p->p_usched->release_curproc(lp); 591 lp->lwp_slptime = 0; 592 } 593 594 /* 595 * If the interlocked flag is set but our cpu bit in the slpqueue 596 * is no longer set, then a wakeup was processed inbetween the 597 * tsleep_interlock() (ours or the callers), and here. This can 598 * occur under numerous circumstances including when we release the 599 * current process. 600 * 601 * Extreme loads can cause the sending of an IPI (e.g. wakeup()'s) 602 * to process incoming IPIs, thus draining incoming wakeups. 603 */ 604 if ((td->td_flags & TDF_TSLEEPQ) == 0) { 605 logtsleep2(ilockfail, ident); 606 goto resume; 607 } 608 609 /* 610 * scheduling is blocked while in a critical section. Coincide 611 * the descheduled-by-tsleep flag with the descheduling of the 612 * lwkt. 613 */ 614 lwkt_deschedule_self(td); 615 td->td_flags |= TDF_TSLEEP_DESCHEDULED; 616 td->td_wmesg = wmesg; 617 618 /* 619 * Setup the timeout, if any 620 */ 621 if (timo) { 622 callout_init(&thandle); 623 callout_reset(&thandle, timo, endtsleep, td); 624 } 625 626 /* 627 * Beddy bye bye. 628 */ 629 if (lp) { 630 /* 631 * Ok, we are sleeping. Place us in the SSLEEP state. 632 */ 633 KKASSERT((lp->lwp_flag & LWP_ONRUNQ) == 0); 634 /* 635 * tstop() sets LSSTOP, so don't fiddle with that. 636 */ 637 if (lp->lwp_stat != LSSTOP) 638 lp->lwp_stat = LSSLEEP; 639 lp->lwp_ru.ru_nvcsw++; 640 lwkt_switch(); 641 642 /* 643 * And when we are woken up, put us back in LSRUN. If we 644 * slept for over a second, recalculate our estcpu. 645 */ 646 lp->lwp_stat = LSRUN; 647 if (lp->lwp_slptime) 648 p->p_usched->recalculate(lp); 649 lp->lwp_slptime = 0; 650 } else { 651 lwkt_switch(); 652 } 653 654 /* 655 * Make sure we haven't switched cpus while we were asleep. It's 656 * not supposed to happen. Cleanup our temporary flags. 657 */ 658 KKASSERT(gd == td->td_gd); 659 660 /* 661 * Cleanup the timeout. 662 */ 663 if (timo) { 664 if (td->td_flags & TDF_TIMEOUT) { 665 td->td_flags &= ~TDF_TIMEOUT; 666 error = EWOULDBLOCK; 667 } else { 668 callout_stop(&thandle); 669 } 670 } 671 672 /* 673 * Make sure we have been removed from the sleepq. This should 674 * have been done for us already. 675 * 676 * However, it is possible for a scheduling IPI to be in flight 677 * from a previous tsleep/tsleep_interlock or due to a straight-out 678 * call to lwkt_schedule() (in the case of an interrupt thread). 679 * So don't complain if DESCHEDULED is still set. 680 */ 681 _tsleep_remove(td); 682 td->td_wmesg = NULL; 683 if (td->td_flags & TDF_TSLEEP_DESCHEDULED) { 684 td->td_flags &= ~TDF_TSLEEP_DESCHEDULED; 685 } 686 687 /* 688 * Figure out the correct error return. If interrupted by a 689 * signal we want to return EINTR or ERESTART. 690 * 691 * If P_MAILBOX is set no automatic system call restart occurs 692 * and we return EINTR. P_MAILBOX is meant to be used as an 693 * interlock, the user must poll it prior to any system call 694 * that it wishes to interlock a mailbox signal against since 695 * the flag is cleared on *any* system call that sleeps. 696 * 697 * p->p_token is held in the p && catch case. 698 */ 699 resume: 700 if (p) { 701 if (catch && error == 0) { 702 if ((p->p_flag & P_MAILBOX) && sig == 0) { 703 error = EINTR; 704 } else if (sig != 0 || (sig = CURSIG(lp))) { 705 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 706 error = EINTR; 707 else 708 error = ERESTART; 709 } 710 } 711 if (catch) 712 lwkt_reltoken(&p->p_token); 713 lp->lwp_flag &= ~(LWP_BREAKTSLEEP | LWP_SINTR); 714 p->p_flag &= ~P_MAILBOX; 715 } 716 logtsleep1(tsleep_end); 717 crit_exit_quick(td); 718 return (error); 719 } 720 721 /* 722 * Interlocked spinlock sleep. An exclusively held spinlock must 723 * be passed to ssleep(). The function will atomically release the 724 * spinlock and tsleep on the ident, then reacquire the spinlock and 725 * return. 726 * 727 * This routine is fairly important along the critical path, so optimize it 728 * heavily. 729 */ 730 int 731 ssleep(const volatile void *ident, struct spinlock *spin, 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 spin_unlock_quick(gd, spin); 739 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 740 spin_lock_quick(gd, spin); 741 742 return (error); 743 } 744 745 int 746 lksleep(const volatile void *ident, struct lock *lock, int flags, 747 const char *wmesg, int timo) 748 { 749 globaldata_t gd = mycpu; 750 int error; 751 752 _tsleep_interlock(gd, ident, flags); 753 lockmgr(lock, LK_RELEASE); 754 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 755 lockmgr(lock, LK_EXCLUSIVE); 756 757 return (error); 758 } 759 760 /* 761 * Interlocked mutex sleep. An exclusively held mutex must be passed 762 * to mtxsleep(). The function will atomically release the mutex 763 * and tsleep on the ident, then reacquire the mutex and return. 764 */ 765 int 766 mtxsleep(const volatile void *ident, struct mtx *mtx, int flags, 767 const char *wmesg, int timo) 768 { 769 globaldata_t gd = mycpu; 770 int error; 771 772 _tsleep_interlock(gd, ident, flags); 773 mtx_unlock(mtx); 774 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 775 mtx_lock_ex_quick(mtx, wmesg); 776 777 return (error); 778 } 779 780 /* 781 * Interlocked serializer sleep. An exclusively held serializer must 782 * be passed to zsleep(). The function will atomically release 783 * the serializer and tsleep on the ident, then reacquire the serializer 784 * and return. 785 */ 786 int 787 zsleep(const volatile void *ident, struct lwkt_serialize *slz, int flags, 788 const char *wmesg, int timo) 789 { 790 globaldata_t gd = mycpu; 791 int ret; 792 793 ASSERT_SERIALIZED(slz); 794 795 _tsleep_interlock(gd, ident, flags); 796 lwkt_serialize_exit(slz); 797 ret = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 798 lwkt_serialize_enter(slz); 799 800 return ret; 801 } 802 803 /* 804 * Directly block on the LWKT thread by descheduling it. This 805 * is much faster then tsleep(), but the only legal way to wake 806 * us up is to directly schedule the thread. 807 * 808 * Setting TDF_SINTR will cause new signals to directly schedule us. 809 * 810 * This routine must be called while in a critical section. 811 */ 812 int 813 lwkt_sleep(const char *wmesg, int flags) 814 { 815 thread_t td = curthread; 816 int sig; 817 818 if ((flags & PCATCH) == 0 || td->td_lwp == NULL) { 819 td->td_flags |= TDF_BLOCKED; 820 td->td_wmesg = wmesg; 821 lwkt_deschedule_self(td); 822 lwkt_switch(); 823 td->td_wmesg = NULL; 824 td->td_flags &= ~TDF_BLOCKED; 825 return(0); 826 } 827 if ((sig = CURSIG(td->td_lwp)) != 0) { 828 if (SIGISMEMBER(td->td_proc->p_sigacts->ps_sigintr, sig)) 829 return(EINTR); 830 else 831 return(ERESTART); 832 833 } 834 td->td_flags |= TDF_BLOCKED | TDF_SINTR; 835 td->td_wmesg = wmesg; 836 lwkt_deschedule_self(td); 837 lwkt_switch(); 838 td->td_flags &= ~(TDF_BLOCKED | TDF_SINTR); 839 td->td_wmesg = NULL; 840 return(0); 841 } 842 843 /* 844 * Implement the timeout for tsleep. 845 * 846 * We set LWP_BREAKTSLEEP to indicate that an event has occured, but 847 * we only call setrunnable if the process is not stopped. 848 * 849 * This type of callout timeout is scheduled on the same cpu the process 850 * is sleeping on. Also, at the moment, the MP lock is held. 851 */ 852 static void 853 endtsleep(void *arg) 854 { 855 thread_t td = arg; 856 struct lwp *lp; 857 858 crit_enter(); 859 lp = td->td_lwp; 860 861 if (lp) 862 lwkt_gettoken(&lp->lwp_proc->p_token); 863 864 /* 865 * cpu interlock. Thread flags are only manipulated on 866 * the cpu owning the thread. proc flags are only manipulated 867 * by the holder of p->p_token. We have both. 868 */ 869 if (td->td_flags & TDF_TSLEEP_DESCHEDULED) { 870 td->td_flags |= TDF_TIMEOUT; 871 872 if (lp) { 873 lp->lwp_flag |= LWP_BREAKTSLEEP; 874 if (lp->lwp_proc->p_stat != SSTOP) 875 setrunnable(lp); 876 } else { 877 _tsleep_wakeup(td); 878 } 879 } 880 if (lp) 881 lwkt_reltoken(&lp->lwp_proc->p_token); 882 crit_exit(); 883 } 884 885 /* 886 * Make all processes sleeping on the specified identifier runnable. 887 * count may be zero or one only. 888 * 889 * The domain encodes the sleep/wakeup domain AND the first cpu to check 890 * (which is always the current cpu). As we iterate across cpus 891 * 892 * This call may run without the MP lock held. We can only manipulate thread 893 * state on the cpu owning the thread. We CANNOT manipulate process state 894 * at all. 895 * 896 * _wakeup() can be passed to an IPI so we can't use (const volatile 897 * void *ident). 898 */ 899 static void 900 _wakeup(void *ident, int domain) 901 { 902 struct tslpque *qp; 903 struct thread *td; 904 struct thread *ntd; 905 globaldata_t gd; 906 #ifdef SMP 907 cpumask_t mask; 908 #endif 909 int id; 910 911 crit_enter(); 912 logtsleep2(wakeup_beg, ident); 913 gd = mycpu; 914 id = LOOKUP(ident); 915 qp = &gd->gd_tsleep_hash[id]; 916 restart: 917 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) { 918 ntd = TAILQ_NEXT(td, td_sleepq); 919 if (td->td_wchan == ident && 920 td->td_wdomain == (domain & PDOMAIN_MASK) 921 ) { 922 KKASSERT(td->td_gd == gd); 923 _tsleep_remove(td); 924 if (td->td_flags & TDF_TSLEEP_DESCHEDULED) { 925 td->td_flags &= ~TDF_TSLEEP_DESCHEDULED; 926 lwkt_schedule(td); 927 if (domain & PWAKEUP_ONE) 928 goto done; 929 } 930 goto restart; 931 } 932 } 933 934 #ifdef SMP 935 /* 936 * We finished checking the current cpu but there still may be 937 * more work to do. Either wakeup_one was requested and no matching 938 * thread was found, or a normal wakeup was requested and we have 939 * to continue checking cpus. 940 * 941 * It should be noted that this scheme is actually less expensive then 942 * the old scheme when waking up multiple threads, since we send 943 * only one IPI message per target candidate which may then schedule 944 * multiple threads. Before we could have wound up sending an IPI 945 * message for each thread on the target cpu (!= current cpu) that 946 * needed to be woken up. 947 * 948 * NOTE: Wakeups occuring on remote cpus are asynchronous. This 949 * should be ok since we are passing idents in the IPI rather then 950 * thread pointers. 951 */ 952 if ((domain & PWAKEUP_MYCPU) == 0 && 953 (mask = slpque_cpumasks[id] & gd->gd_other_cpus) != 0) { 954 lwkt_send_ipiq2_mask(mask, _wakeup, ident, 955 domain | PWAKEUP_MYCPU); 956 } 957 #endif 958 done: 959 logtsleep1(wakeup_end); 960 crit_exit(); 961 } 962 963 /* 964 * Wakeup all threads tsleep()ing on the specified ident, on all cpus 965 */ 966 void 967 wakeup(const volatile void *ident) 968 { 969 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mycpu->gd_cpuid)); 970 } 971 972 /* 973 * Wakeup one thread tsleep()ing on the specified ident, on any cpu. 974 */ 975 void 976 wakeup_one(const volatile void *ident) 977 { 978 /* XXX potentially round-robin the first responding cpu */ 979 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mycpu->gd_cpuid) | PWAKEUP_ONE); 980 } 981 982 /* 983 * Wakeup threads tsleep()ing on the specified ident on the current cpu 984 * only. 985 */ 986 void 987 wakeup_mycpu(const volatile void *ident) 988 { 989 _wakeup(__DEALL(ident), PWAKEUP_MYCPU); 990 } 991 992 /* 993 * Wakeup one thread tsleep()ing on the specified ident on the current cpu 994 * only. 995 */ 996 void 997 wakeup_mycpu_one(const volatile void *ident) 998 { 999 /* XXX potentially round-robin the first responding cpu */ 1000 _wakeup(__DEALL(ident), PWAKEUP_MYCPU|PWAKEUP_ONE); 1001 } 1002 1003 /* 1004 * Wakeup all thread tsleep()ing on the specified ident on the specified cpu 1005 * only. 1006 */ 1007 void 1008 wakeup_oncpu(globaldata_t gd, const volatile void *ident) 1009 { 1010 #ifdef SMP 1011 if (gd == mycpu) { 1012 _wakeup(__DEALL(ident), PWAKEUP_MYCPU); 1013 } else { 1014 lwkt_send_ipiq2(gd, _wakeup, __DEALL(ident), PWAKEUP_MYCPU); 1015 } 1016 #else 1017 _wakeup(__DEALL(ident), PWAKEUP_MYCPU); 1018 #endif 1019 } 1020 1021 /* 1022 * Wakeup one thread tsleep()ing on the specified ident on the specified cpu 1023 * only. 1024 */ 1025 void 1026 wakeup_oncpu_one(globaldata_t gd, const volatile void *ident) 1027 { 1028 #ifdef SMP 1029 if (gd == mycpu) { 1030 _wakeup(__DEALL(ident), PWAKEUP_MYCPU | PWAKEUP_ONE); 1031 } else { 1032 lwkt_send_ipiq2(gd, _wakeup, __DEALL(ident), 1033 PWAKEUP_MYCPU | PWAKEUP_ONE); 1034 } 1035 #else 1036 _wakeup(__DEALL(ident), PWAKEUP_MYCPU | PWAKEUP_ONE); 1037 #endif 1038 } 1039 1040 /* 1041 * Wakeup all threads waiting on the specified ident that slept using 1042 * the specified domain, on all cpus. 1043 */ 1044 void 1045 wakeup_domain(const volatile void *ident, int domain) 1046 { 1047 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(domain, mycpu->gd_cpuid)); 1048 } 1049 1050 /* 1051 * Wakeup one thread waiting on the specified ident that slept using 1052 * the specified domain, on any cpu. 1053 */ 1054 void 1055 wakeup_domain_one(const volatile void *ident, int domain) 1056 { 1057 /* XXX potentially round-robin the first responding cpu */ 1058 _wakeup(__DEALL(ident), 1059 PWAKEUP_ENCODE(domain, mycpu->gd_cpuid) | PWAKEUP_ONE); 1060 } 1061 1062 /* 1063 * setrunnable() 1064 * 1065 * Make a process runnable. lp->lwp_proc->p_token must be held on call. 1066 * This only has an effect if we are in SSLEEP. We only break out of the 1067 * tsleep if LWP_BREAKTSLEEP is set, otherwise we just fix-up the state. 1068 * 1069 * NOTE: With p_token held we can only safely manipulate the process 1070 * structure and the lp's lwp_stat. 1071 */ 1072 void 1073 setrunnable(struct lwp *lp) 1074 { 1075 ASSERT_LWKT_TOKEN_HELD(&lp->lwp_proc->p_token); 1076 crit_enter(); 1077 if (lp->lwp_stat == LSSTOP) 1078 lp->lwp_stat = LSSLEEP; 1079 if (lp->lwp_stat == LSSLEEP && (lp->lwp_flag & LWP_BREAKTSLEEP)) 1080 _tsleep_wakeup(lp->lwp_thread); 1081 crit_exit(); 1082 } 1083 1084 /* 1085 * The process is stopped due to some condition, usually because p_stat is 1086 * set to SSTOP, but also possibly due to being traced. 1087 * 1088 * NOTE! If the caller sets SSTOP, the caller must also clear P_WAITED 1089 * because the parent may check the child's status before the child actually 1090 * gets to this routine. 1091 * 1092 * This routine is called with the current lwp only, typically just 1093 * before returning to userland. 1094 * 1095 * Setting LWP_BREAKTSLEEP before entering the tsleep will cause a passive 1096 * SIGCONT to break out of the tsleep. 1097 */ 1098 void 1099 tstop(void) 1100 { 1101 struct lwp *lp = curthread->td_lwp; 1102 struct proc *p = lp->lwp_proc; 1103 1104 crit_enter(); 1105 /* 1106 * If LWP_WSTOP is set, we were sleeping 1107 * while our process was stopped. At this point 1108 * we were already counted as stopped. 1109 */ 1110 if ((lp->lwp_flag & LWP_WSTOP) == 0) { 1111 /* 1112 * If we're the last thread to stop, signal 1113 * our parent. 1114 */ 1115 p->p_nstopped++; 1116 lp->lwp_flag |= LWP_WSTOP; 1117 wakeup(&p->p_nstopped); 1118 if (p->p_nstopped == p->p_nthreads) { 1119 p->p_flag &= ~P_WAITED; 1120 wakeup(p->p_pptr); 1121 if ((p->p_pptr->p_sigacts->ps_flag & PS_NOCLDSTOP) == 0) 1122 ksignal(p->p_pptr, SIGCHLD); 1123 } 1124 } 1125 while (p->p_stat == SSTOP) { 1126 lp->lwp_flag |= LWP_BREAKTSLEEP; 1127 lp->lwp_stat = LSSTOP; 1128 tsleep(p, 0, "stop", 0); 1129 } 1130 p->p_nstopped--; 1131 lp->lwp_flag &= ~LWP_WSTOP; 1132 crit_exit(); 1133 } 1134 1135 /* 1136 * Compute a tenex style load average of a quantity on 1137 * 1, 5 and 15 minute intervals. 1138 */ 1139 static int loadav_count_runnable(struct lwp *p, void *data); 1140 1141 static void 1142 loadav(void *arg) 1143 { 1144 struct loadavg *avg; 1145 int i, nrun; 1146 1147 nrun = 0; 1148 alllwp_scan(loadav_count_runnable, &nrun); 1149 avg = &averunnable; 1150 for (i = 0; i < 3; i++) { 1151 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] + 1152 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT; 1153 } 1154 1155 /* 1156 * Schedule the next update to occur after 5 seconds, but add a 1157 * random variation to avoid synchronisation with processes that 1158 * run at regular intervals. 1159 */ 1160 callout_reset(&loadav_callout, hz * 4 + (int)(krandom() % (hz * 2 + 1)), 1161 loadav, NULL); 1162 } 1163 1164 static int 1165 loadav_count_runnable(struct lwp *lp, void *data) 1166 { 1167 int *nrunp = data; 1168 thread_t td; 1169 1170 switch (lp->lwp_stat) { 1171 case LSRUN: 1172 if ((td = lp->lwp_thread) == NULL) 1173 break; 1174 if (td->td_flags & TDF_BLOCKED) 1175 break; 1176 ++*nrunp; 1177 break; 1178 default: 1179 break; 1180 } 1181 return(0); 1182 } 1183 1184 /* ARGSUSED */ 1185 static void 1186 sched_setup(void *dummy) 1187 { 1188 callout_init(&loadav_callout); 1189 callout_init(&schedcpu_callout); 1190 1191 /* Kick off timeout driven events by calling first time. */ 1192 schedcpu(NULL); 1193 loadav(NULL); 1194 } 1195 1196