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 lwkt_deschedule_self(td); 621 td->td_flags |= TDF_TSLEEP_DESCHEDULED; 622 td->td_wmesg = wmesg; 623 624 /* 625 * Setup the timeout, if any 626 */ 627 if (timo) { 628 callout_init(&thandle); 629 callout_reset(&thandle, timo, endtsleep, td); 630 } 631 632 /* 633 * Beddy bye bye. 634 */ 635 if (lp) { 636 /* 637 * Ok, we are sleeping. Place us in the SSLEEP state. 638 */ 639 KKASSERT((lp->lwp_flag & LWP_ONRUNQ) == 0); 640 /* 641 * tstop() sets LSSTOP, so don't fiddle with that. 642 */ 643 if (lp->lwp_stat != LSSTOP) 644 lp->lwp_stat = LSSLEEP; 645 lp->lwp_ru.ru_nvcsw++; 646 lwkt_switch(); 647 648 /* 649 * And when we are woken up, put us back in LSRUN. If we 650 * slept for over a second, recalculate our estcpu. 651 */ 652 lp->lwp_stat = LSRUN; 653 if (lp->lwp_slptime) 654 p->p_usched->recalculate(lp); 655 lp->lwp_slptime = 0; 656 } else { 657 lwkt_switch(); 658 } 659 660 /* 661 * Make sure we haven't switched cpus while we were asleep. It's 662 * not supposed to happen. Cleanup our temporary flags. 663 */ 664 KKASSERT(gd == td->td_gd); 665 666 /* 667 * Cleanup the timeout. 668 */ 669 if (timo) { 670 if (td->td_flags & TDF_TIMEOUT) { 671 td->td_flags &= ~TDF_TIMEOUT; 672 error = EWOULDBLOCK; 673 } else { 674 callout_stop(&thandle); 675 } 676 } 677 678 /* 679 * Make sure we have been removed from the sleepq. This should 680 * have been done for us already. 681 * 682 * However, it is possible for a scheduling IPI to be in flight 683 * from a previous tsleep/tsleep_interlock or due to a straight-out 684 * call to lwkt_schedule() (in the case of an interrupt thread). 685 * So don't complain if DESCHEDULED is still set. 686 */ 687 _tsleep_remove(td); 688 td->td_wmesg = NULL; 689 if (td->td_flags & TDF_TSLEEP_DESCHEDULED) { 690 td->td_flags &= ~TDF_TSLEEP_DESCHEDULED; 691 } 692 693 /* 694 * Figure out the correct error return. If interrupted by a 695 * signal we want to return EINTR or ERESTART. 696 * 697 * If P_MAILBOX is set no automatic system call restart occurs 698 * and we return EINTR. P_MAILBOX is meant to be used as an 699 * interlock, the user must poll it prior to any system call 700 * that it wishes to interlock a mailbox signal against since 701 * the flag is cleared on *any* system call that sleeps. 702 * 703 * p->p_token is held in the p && catch case. 704 */ 705 resume: 706 if (p) { 707 if (catch && error == 0) { 708 if ((p->p_flag & P_MAILBOX) && sig == 0) { 709 error = EINTR; 710 } else if (sig != 0 || (sig = CURSIG(lp))) { 711 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 712 error = EINTR; 713 else 714 error = ERESTART; 715 } 716 } 717 if (catch) 718 lwkt_reltoken(&p->p_token); 719 lp->lwp_flag &= ~(LWP_BREAKTSLEEP | LWP_SINTR); 720 p->p_flag &= ~P_MAILBOX; 721 } 722 logtsleep1(tsleep_end); 723 crit_exit_quick(td); 724 return (error); 725 } 726 727 /* 728 * Interlocked spinlock sleep. An exclusively held spinlock must 729 * be passed to ssleep(). The function will atomically release the 730 * spinlock and tsleep on the ident, then reacquire the spinlock and 731 * return. 732 * 733 * This routine is fairly important along the critical path, so optimize it 734 * heavily. 735 */ 736 int 737 ssleep(const volatile void *ident, struct spinlock *spin, int flags, 738 const char *wmesg, int timo) 739 { 740 globaldata_t gd = mycpu; 741 int error; 742 743 _tsleep_interlock(gd, ident, flags); 744 spin_unlock_quick(gd, spin); 745 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 746 spin_lock_quick(gd, spin); 747 748 return (error); 749 } 750 751 int 752 lksleep(const volatile void *ident, struct lock *lock, 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 lockmgr(lock, LK_RELEASE); 760 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 761 lockmgr(lock, LK_EXCLUSIVE); 762 763 return (error); 764 } 765 766 /* 767 * Interlocked mutex sleep. An exclusively held mutex must be passed 768 * to mtxsleep(). The function will atomically release the mutex 769 * and tsleep on the ident, then reacquire the mutex and return. 770 */ 771 int 772 mtxsleep(const volatile void *ident, struct mtx *mtx, int flags, 773 const char *wmesg, int timo) 774 { 775 globaldata_t gd = mycpu; 776 int error; 777 778 _tsleep_interlock(gd, ident, flags); 779 mtx_unlock(mtx); 780 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 781 mtx_lock_ex_quick(mtx, wmesg); 782 783 return (error); 784 } 785 786 /* 787 * Interlocked serializer sleep. An exclusively held serializer must 788 * be passed to zsleep(). The function will atomically release 789 * the serializer and tsleep on the ident, then reacquire the serializer 790 * and return. 791 */ 792 int 793 zsleep(const volatile void *ident, struct lwkt_serialize *slz, int flags, 794 const char *wmesg, int timo) 795 { 796 globaldata_t gd = mycpu; 797 int ret; 798 799 ASSERT_SERIALIZED(slz); 800 801 _tsleep_interlock(gd, ident, flags); 802 lwkt_serialize_exit(slz); 803 ret = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 804 lwkt_serialize_enter(slz); 805 806 return ret; 807 } 808 809 /* 810 * Directly block on the LWKT thread by descheduling it. This 811 * is much faster then tsleep(), but the only legal way to wake 812 * us up is to directly schedule the thread. 813 * 814 * Setting TDF_SINTR will cause new signals to directly schedule us. 815 * 816 * This routine must be called while in a critical section. 817 */ 818 int 819 lwkt_sleep(const char *wmesg, int flags) 820 { 821 thread_t td = curthread; 822 int sig; 823 824 if ((flags & PCATCH) == 0 || td->td_lwp == NULL) { 825 td->td_flags |= TDF_BLOCKED; 826 td->td_wmesg = wmesg; 827 lwkt_deschedule_self(td); 828 lwkt_switch(); 829 td->td_wmesg = NULL; 830 td->td_flags &= ~TDF_BLOCKED; 831 return(0); 832 } 833 if ((sig = CURSIG(td->td_lwp)) != 0) { 834 if (SIGISMEMBER(td->td_proc->p_sigacts->ps_sigintr, sig)) 835 return(EINTR); 836 else 837 return(ERESTART); 838 839 } 840 td->td_flags |= TDF_BLOCKED | TDF_SINTR; 841 td->td_wmesg = wmesg; 842 lwkt_deschedule_self(td); 843 lwkt_switch(); 844 td->td_flags &= ~(TDF_BLOCKED | TDF_SINTR); 845 td->td_wmesg = NULL; 846 return(0); 847 } 848 849 /* 850 * Implement the timeout for tsleep. 851 * 852 * We set LWP_BREAKTSLEEP to indicate that an event has occured, but 853 * we only call setrunnable if the process is not stopped. 854 * 855 * This type of callout timeout is scheduled on the same cpu the process 856 * is sleeping on. Also, at the moment, the MP lock is held. 857 */ 858 static void 859 endtsleep(void *arg) 860 { 861 thread_t td = arg; 862 struct lwp *lp; 863 864 crit_enter(); 865 lp = td->td_lwp; 866 867 if (lp) 868 lwkt_gettoken(&lp->lwp_proc->p_token); 869 870 /* 871 * cpu interlock. Thread flags are only manipulated on 872 * the cpu owning the thread. proc flags are only manipulated 873 * by the holder of p->p_token. We have both. 874 */ 875 if (td->td_flags & TDF_TSLEEP_DESCHEDULED) { 876 td->td_flags |= TDF_TIMEOUT; 877 878 if (lp) { 879 lp->lwp_flag |= LWP_BREAKTSLEEP; 880 if (lp->lwp_proc->p_stat != SSTOP) 881 setrunnable(lp); 882 } else { 883 _tsleep_wakeup(td); 884 } 885 } 886 if (lp) 887 lwkt_reltoken(&lp->lwp_proc->p_token); 888 crit_exit(); 889 } 890 891 /* 892 * Make all processes sleeping on the specified identifier runnable. 893 * count may be zero or one only. 894 * 895 * The domain encodes the sleep/wakeup domain AND the first cpu to check 896 * (which is always the current cpu). As we iterate across cpus 897 * 898 * This call may run without the MP lock held. We can only manipulate thread 899 * state on the cpu owning the thread. We CANNOT manipulate process state 900 * at all. 901 * 902 * _wakeup() can be passed to an IPI so we can't use (const volatile 903 * void *ident). 904 */ 905 static void 906 _wakeup(void *ident, int domain) 907 { 908 struct tslpque *qp; 909 struct thread *td; 910 struct thread *ntd; 911 globaldata_t gd; 912 #ifdef SMP 913 cpumask_t mask; 914 #endif 915 int id; 916 917 crit_enter(); 918 logtsleep2(wakeup_beg, ident); 919 gd = mycpu; 920 id = LOOKUP(ident); 921 qp = &gd->gd_tsleep_hash[id]; 922 restart: 923 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) { 924 ntd = TAILQ_NEXT(td, td_sleepq); 925 if (td->td_wchan == ident && 926 td->td_wdomain == (domain & PDOMAIN_MASK) 927 ) { 928 KKASSERT(td->td_gd == gd); 929 _tsleep_remove(td); 930 if (td->td_flags & TDF_TSLEEP_DESCHEDULED) { 931 td->td_flags &= ~TDF_TSLEEP_DESCHEDULED; 932 lwkt_schedule(td); 933 if (domain & PWAKEUP_ONE) 934 goto done; 935 } 936 goto restart; 937 } 938 } 939 940 #ifdef SMP 941 /* 942 * We finished checking the current cpu but there still may be 943 * more work to do. Either wakeup_one was requested and no matching 944 * thread was found, or a normal wakeup was requested and we have 945 * to continue checking cpus. 946 * 947 * It should be noted that this scheme is actually less expensive then 948 * the old scheme when waking up multiple threads, since we send 949 * only one IPI message per target candidate which may then schedule 950 * multiple threads. Before we could have wound up sending an IPI 951 * message for each thread on the target cpu (!= current cpu) that 952 * needed to be woken up. 953 * 954 * NOTE: Wakeups occuring on remote cpus are asynchronous. This 955 * should be ok since we are passing idents in the IPI rather then 956 * thread pointers. 957 */ 958 if ((domain & PWAKEUP_MYCPU) == 0 && 959 (mask = slpque_cpumasks[id] & gd->gd_other_cpus) != 0) { 960 lwkt_send_ipiq2_mask(mask, _wakeup, ident, 961 domain | PWAKEUP_MYCPU); 962 } 963 #endif 964 done: 965 logtsleep1(wakeup_end); 966 crit_exit(); 967 } 968 969 /* 970 * Wakeup all threads tsleep()ing on the specified ident, on all cpus 971 */ 972 void 973 wakeup(const volatile void *ident) 974 { 975 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mycpu->gd_cpuid)); 976 } 977 978 /* 979 * Wakeup one thread tsleep()ing on the specified ident, on any cpu. 980 */ 981 void 982 wakeup_one(const volatile void *ident) 983 { 984 /* XXX potentially round-robin the first responding cpu */ 985 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mycpu->gd_cpuid) | PWAKEUP_ONE); 986 } 987 988 /* 989 * Wakeup threads tsleep()ing on the specified ident on the current cpu 990 * only. 991 */ 992 void 993 wakeup_mycpu(const volatile void *ident) 994 { 995 _wakeup(__DEALL(ident), PWAKEUP_MYCPU); 996 } 997 998 /* 999 * Wakeup one thread tsleep()ing on the specified ident on the current cpu 1000 * only. 1001 */ 1002 void 1003 wakeup_mycpu_one(const volatile void *ident) 1004 { 1005 /* XXX potentially round-robin the first responding cpu */ 1006 _wakeup(__DEALL(ident), PWAKEUP_MYCPU|PWAKEUP_ONE); 1007 } 1008 1009 /* 1010 * Wakeup all thread tsleep()ing on the specified ident on the specified cpu 1011 * only. 1012 */ 1013 void 1014 wakeup_oncpu(globaldata_t gd, const volatile void *ident) 1015 { 1016 #ifdef SMP 1017 if (gd == mycpu) { 1018 _wakeup(__DEALL(ident), PWAKEUP_MYCPU); 1019 } else { 1020 lwkt_send_ipiq2(gd, _wakeup, __DEALL(ident), PWAKEUP_MYCPU); 1021 } 1022 #else 1023 _wakeup(__DEALL(ident), PWAKEUP_MYCPU); 1024 #endif 1025 } 1026 1027 /* 1028 * Wakeup one thread tsleep()ing on the specified ident on the specified cpu 1029 * only. 1030 */ 1031 void 1032 wakeup_oncpu_one(globaldata_t gd, const volatile void *ident) 1033 { 1034 #ifdef SMP 1035 if (gd == mycpu) { 1036 _wakeup(__DEALL(ident), PWAKEUP_MYCPU | PWAKEUP_ONE); 1037 } else { 1038 lwkt_send_ipiq2(gd, _wakeup, __DEALL(ident), 1039 PWAKEUP_MYCPU | PWAKEUP_ONE); 1040 } 1041 #else 1042 _wakeup(__DEALL(ident), PWAKEUP_MYCPU | PWAKEUP_ONE); 1043 #endif 1044 } 1045 1046 /* 1047 * Wakeup all threads waiting on the specified ident that slept using 1048 * the specified domain, on all cpus. 1049 */ 1050 void 1051 wakeup_domain(const volatile void *ident, int domain) 1052 { 1053 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(domain, mycpu->gd_cpuid)); 1054 } 1055 1056 /* 1057 * Wakeup one thread waiting on the specified ident that slept using 1058 * the specified domain, on any cpu. 1059 */ 1060 void 1061 wakeup_domain_one(const volatile void *ident, int domain) 1062 { 1063 /* XXX potentially round-robin the first responding cpu */ 1064 _wakeup(__DEALL(ident), 1065 PWAKEUP_ENCODE(domain, mycpu->gd_cpuid) | PWAKEUP_ONE); 1066 } 1067 1068 /* 1069 * setrunnable() 1070 * 1071 * Make a process runnable. lp->lwp_proc->p_token must be held on call. 1072 * This only has an effect if we are in SSLEEP. We only break out of the 1073 * tsleep if LWP_BREAKTSLEEP is set, otherwise we just fix-up the state. 1074 * 1075 * NOTE: With p_token held we can only safely manipulate the process 1076 * structure and the lp's lwp_stat. 1077 */ 1078 void 1079 setrunnable(struct lwp *lp) 1080 { 1081 ASSERT_LWKT_TOKEN_HELD(&lp->lwp_proc->p_token); 1082 crit_enter(); 1083 if (lp->lwp_stat == LSSTOP) 1084 lp->lwp_stat = LSSLEEP; 1085 if (lp->lwp_stat == LSSLEEP && (lp->lwp_flag & LWP_BREAKTSLEEP)) 1086 _tsleep_wakeup(lp->lwp_thread); 1087 crit_exit(); 1088 } 1089 1090 /* 1091 * The process is stopped due to some condition, usually because p_stat is 1092 * set to SSTOP, but also possibly due to being traced. 1093 * 1094 * NOTE! If the caller sets SSTOP, the caller must also clear P_WAITED 1095 * because the parent may check the child's status before the child actually 1096 * gets to this routine. 1097 * 1098 * This routine is called with the current lwp only, typically just 1099 * before returning to userland. 1100 * 1101 * Setting LWP_BREAKTSLEEP before entering the tsleep will cause a passive 1102 * SIGCONT to break out of the tsleep. 1103 */ 1104 void 1105 tstop(void) 1106 { 1107 struct lwp *lp = curthread->td_lwp; 1108 struct proc *p = lp->lwp_proc; 1109 1110 crit_enter(); 1111 /* 1112 * If LWP_WSTOP is set, we were sleeping 1113 * while our process was stopped. At this point 1114 * we were already counted as stopped. 1115 */ 1116 if ((lp->lwp_flag & LWP_WSTOP) == 0) { 1117 /* 1118 * If we're the last thread to stop, signal 1119 * our parent. 1120 */ 1121 p->p_nstopped++; 1122 lp->lwp_flag |= LWP_WSTOP; 1123 wakeup(&p->p_nstopped); 1124 if (p->p_nstopped == p->p_nthreads) { 1125 p->p_flag &= ~P_WAITED; 1126 wakeup(p->p_pptr); 1127 if ((p->p_pptr->p_sigacts->ps_flag & PS_NOCLDSTOP) == 0) 1128 ksignal(p->p_pptr, SIGCHLD); 1129 } 1130 } 1131 while (p->p_stat == SSTOP) { 1132 lp->lwp_flag |= LWP_BREAKTSLEEP; 1133 lp->lwp_stat = LSSTOP; 1134 tsleep(p, 0, "stop", 0); 1135 } 1136 p->p_nstopped--; 1137 lp->lwp_flag &= ~LWP_WSTOP; 1138 crit_exit(); 1139 } 1140 1141 /* 1142 * Compute a tenex style load average of a quantity on 1143 * 1, 5 and 15 minute intervals. 1144 */ 1145 static int loadav_count_runnable(struct lwp *p, void *data); 1146 1147 static void 1148 loadav(void *arg) 1149 { 1150 struct loadavg *avg; 1151 int i, nrun; 1152 1153 nrun = 0; 1154 alllwp_scan(loadav_count_runnable, &nrun); 1155 avg = &averunnable; 1156 for (i = 0; i < 3; i++) { 1157 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] + 1158 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT; 1159 } 1160 1161 /* 1162 * Schedule the next update to occur after 5 seconds, but add a 1163 * random variation to avoid synchronisation with processes that 1164 * run at regular intervals. 1165 */ 1166 callout_reset(&loadav_callout, hz * 4 + (int)(krandom() % (hz * 2 + 1)), 1167 loadav, NULL); 1168 } 1169 1170 static int 1171 loadav_count_runnable(struct lwp *lp, void *data) 1172 { 1173 int *nrunp = data; 1174 thread_t td; 1175 1176 switch (lp->lwp_stat) { 1177 case LSRUN: 1178 if ((td = lp->lwp_thread) == NULL) 1179 break; 1180 if (td->td_flags & TDF_BLOCKED) 1181 break; 1182 ++*nrunp; 1183 break; 1184 default: 1185 break; 1186 } 1187 return(0); 1188 } 1189 1190 /* ARGSUSED */ 1191 static void 1192 sched_setup(void *dummy) 1193 { 1194 callout_init(&loadav_callout); 1195 callout_init(&schedcpu_callout); 1196 1197 /* Kick off timeout driven events by calling first time. */ 1198 schedcpu(NULL); 1199 loadav(NULL); 1200 } 1201 1202