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(struct thread *td); 123 #endif 124 125 /* 126 * Adjust the scheduler quantum. The quantum is specified in microseconds. 127 * Note that 'tick' is in microseconds per tick. 128 */ 129 static int 130 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS) 131 { 132 int error, new_val; 133 134 new_val = sched_quantum * ustick; 135 error = sysctl_handle_int(oidp, &new_val, 0, req); 136 if (error != 0 || req->newptr == NULL) 137 return (error); 138 if (new_val < ustick) 139 return (EINVAL); 140 sched_quantum = new_val / ustick; 141 hogticks = 2 * sched_quantum; 142 return (0); 143 } 144 145 SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW, 146 0, sizeof sched_quantum, sysctl_kern_quantum, "I", ""); 147 148 /* 149 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the 150 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below 151 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT). 152 * 153 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used: 154 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits). 155 * 156 * If you don't want to bother with the faster/more-accurate formula, you 157 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate 158 * (more general) method of calculating the %age of CPU used by a process. 159 * 160 * decay 95% of `lwp_pctcpu' in 60 seconds; see CCPU_SHIFT before changing 161 */ 162 #define CCPU_SHIFT 11 163 164 static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */ 165 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 166 167 /* 168 * kernel uses `FSCALE', userland (SHOULD) use kern.fscale 169 */ 170 int fscale __unused = FSCALE; /* exported to systat */ 171 SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, ""); 172 173 /* 174 * Recompute process priorities, once a second. 175 * 176 * Since the userland schedulers are typically event oriented, if the 177 * estcpu calculation at wakeup() time is not sufficient to make a 178 * process runnable relative to other processes in the system we have 179 * a 1-second recalc to help out. 180 * 181 * This code also allows us to store sysclock_t data in the process structure 182 * without fear of an overrun, since sysclock_t are guarenteed to hold 183 * several seconds worth of count. 184 * 185 * WARNING! callouts can preempt normal threads. However, they will not 186 * preempt a thread holding a spinlock so we *can* safely use spinlocks. 187 */ 188 static int schedcpu_stats(struct proc *p, void *data __unused); 189 static int schedcpu_resource(struct proc *p, void *data __unused); 190 191 static void 192 schedcpu(void *arg) 193 { 194 allproc_scan(schedcpu_stats, NULL); 195 allproc_scan(schedcpu_resource, NULL); 196 wakeup((caddr_t)&lbolt); 197 wakeup((caddr_t)&lbolt_syncer); 198 callout_reset(&schedcpu_callout, hz, schedcpu, NULL); 199 } 200 201 /* 202 * General process statistics once a second 203 */ 204 static int 205 schedcpu_stats(struct proc *p, void *data __unused) 206 { 207 struct lwp *lp; 208 209 crit_enter(); 210 p->p_swtime++; 211 FOREACH_LWP_IN_PROC(lp, p) { 212 if (lp->lwp_stat == LSSLEEP) 213 lp->lwp_slptime++; 214 215 /* 216 * Only recalculate processes that are active or have slept 217 * less then 2 seconds. The schedulers understand this. 218 */ 219 if (lp->lwp_slptime <= 1) { 220 p->p_usched->recalculate(lp); 221 } else { 222 lp->lwp_pctcpu = (lp->lwp_pctcpu * ccpu) >> FSHIFT; 223 } 224 } 225 crit_exit(); 226 return(0); 227 } 228 229 /* 230 * Resource checks. XXX break out since ksignal/killproc can block, 231 * limiting us to one process killed per second. There is probably 232 * a better way. 233 */ 234 static int 235 schedcpu_resource(struct proc *p, void *data __unused) 236 { 237 u_int64_t ttime; 238 struct lwp *lp; 239 240 crit_enter(); 241 if (p->p_stat == SIDL || 242 p->p_stat == SZOMB || 243 p->p_limit == NULL 244 ) { 245 crit_exit(); 246 return(0); 247 } 248 249 ttime = 0; 250 FOREACH_LWP_IN_PROC(lp, p) { 251 /* 252 * We may have caught an lp in the middle of being 253 * created, lwp_thread can be NULL. 254 */ 255 if (lp->lwp_thread) { 256 ttime += lp->lwp_thread->td_sticks; 257 ttime += lp->lwp_thread->td_uticks; 258 } 259 } 260 261 switch(plimit_testcpulimit(p->p_limit, ttime)) { 262 case PLIMIT_TESTCPU_KILL: 263 killproc(p, "exceeded maximum CPU limit"); 264 break; 265 case PLIMIT_TESTCPU_XCPU: 266 if ((p->p_flag & P_XCPU) == 0) { 267 p->p_flag |= P_XCPU; 268 ksignal(p, SIGXCPU); 269 } 270 break; 271 default: 272 break; 273 } 274 crit_exit(); 275 return(0); 276 } 277 278 /* 279 * This is only used by ps. Generate a cpu percentage use over 280 * a period of one second. 281 * 282 * MPSAFE 283 */ 284 void 285 updatepcpu(struct lwp *lp, int cpticks, int ttlticks) 286 { 287 fixpt_t acc; 288 int remticks; 289 290 acc = (cpticks << FSHIFT) / ttlticks; 291 if (ttlticks >= ESTCPUFREQ) { 292 lp->lwp_pctcpu = acc; 293 } else { 294 remticks = ESTCPUFREQ - ttlticks; 295 lp->lwp_pctcpu = (acc * ttlticks + lp->lwp_pctcpu * remticks) / 296 ESTCPUFREQ; 297 } 298 } 299 300 /* 301 * tsleep/wakeup hash table parameters. Try to find the sweet spot for 302 * like addresses being slept on. 303 */ 304 #define TABLESIZE 1024 305 #define LOOKUP(x) (((intptr_t)(x) >> 6) & (TABLESIZE - 1)) 306 307 static cpumask_t slpque_cpumasks[TABLESIZE]; 308 309 /* 310 * General scheduler initialization. We force a reschedule 25 times 311 * a second by default. Note that cpu0 is initialized in early boot and 312 * cannot make any high level calls. 313 * 314 * Each cpu has its own sleep queue. 315 */ 316 void 317 sleep_gdinit(globaldata_t gd) 318 { 319 static struct tslpque slpque_cpu0[TABLESIZE]; 320 int i; 321 322 if (gd->gd_cpuid == 0) { 323 sched_quantum = (hz + 24) / 25; 324 hogticks = 2 * sched_quantum; 325 326 gd->gd_tsleep_hash = slpque_cpu0; 327 } else { 328 gd->gd_tsleep_hash = kmalloc(sizeof(slpque_cpu0), 329 M_TSLEEP, M_WAITOK | M_ZERO); 330 } 331 for (i = 0; i < TABLESIZE; ++i) 332 TAILQ_INIT(&gd->gd_tsleep_hash[i]); 333 } 334 335 /* 336 * This is a dandy function that allows us to interlock tsleep/wakeup 337 * operations with unspecified upper level locks, such as lockmgr locks, 338 * simply by holding a critical section. The sequence is: 339 * 340 * (acquire upper level lock) 341 * tsleep_interlock(blah) 342 * (release upper level lock) 343 * tsleep(blah, ...) 344 * 345 * Basically this functions queues us on the tsleep queue without actually 346 * descheduling us. When tsleep() is later called with PINTERLOCK it 347 * assumes the thread was already queued, otherwise it queues it there. 348 * 349 * Thus it is possible to receive the wakeup prior to going to sleep and 350 * the race conditions are covered. 351 */ 352 static __inline void 353 _tsleep_interlock(globaldata_t gd, const volatile void *ident, int flags) 354 { 355 thread_t td = gd->gd_curthread; 356 int id; 357 358 crit_enter_quick(td); 359 if (td->td_flags & TDF_TSLEEPQ) { 360 id = LOOKUP(td->td_wchan); 361 TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_sleepq); 362 if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL) 363 atomic_clear_cpumask(&slpque_cpumasks[id], gd->gd_cpumask); 364 } else { 365 td->td_flags |= TDF_TSLEEPQ; 366 } 367 id = LOOKUP(ident); 368 TAILQ_INSERT_TAIL(&gd->gd_tsleep_hash[id], td, td_sleepq); 369 atomic_set_cpumask(&slpque_cpumasks[id], gd->gd_cpumask); 370 td->td_wchan = ident; 371 td->td_wdomain = flags & PDOMAIN_MASK; 372 crit_exit_quick(td); 373 } 374 375 void 376 tsleep_interlock(const volatile void *ident, int flags) 377 { 378 _tsleep_interlock(mycpu, ident, flags); 379 } 380 381 /* 382 * Remove thread from sleepq. Must be called with a critical section held. 383 */ 384 static __inline void 385 _tsleep_remove(thread_t td) 386 { 387 globaldata_t gd = mycpu; 388 int id; 389 390 KKASSERT(td->td_gd == gd); 391 if (td->td_flags & TDF_TSLEEPQ) { 392 td->td_flags &= ~TDF_TSLEEPQ; 393 id = LOOKUP(td->td_wchan); 394 TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_sleepq); 395 if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL) 396 atomic_clear_cpumask(&slpque_cpumasks[id], gd->gd_cpumask); 397 td->td_wchan = NULL; 398 td->td_wdomain = 0; 399 } 400 } 401 402 void 403 tsleep_remove(thread_t td) 404 { 405 _tsleep_remove(td); 406 } 407 408 /* 409 * This function removes a thread from the tsleep queue and schedules 410 * it. This function may act asynchronously. The target thread may be 411 * sleeping on a different cpu. 412 * 413 * This function mus be called while in a critical section but if the 414 * target thread is sleeping on a different cpu we cannot safely probe 415 * td_flags. 416 */ 417 static __inline 418 void 419 _tsleep_wakeup(struct thread *td) 420 { 421 #ifdef SMP 422 globaldata_t gd = mycpu; 423 424 if (td->td_gd != gd) { 425 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)tsleep_wakeup, td); 426 return; 427 } 428 #endif 429 _tsleep_remove(td); 430 if (td->td_flags & TDF_TSLEEP_DESCHEDULED) { 431 td->td_flags &= ~TDF_TSLEEP_DESCHEDULED; 432 lwkt_schedule(td); 433 } 434 } 435 436 #ifdef SMP 437 static 438 void 439 tsleep_wakeup(struct thread *td) 440 { 441 _tsleep_wakeup(td); 442 } 443 #endif 444 445 446 /* 447 * General sleep call. Suspends the current process until a wakeup is 448 * performed on the specified identifier. The process will then be made 449 * runnable with the specified priority. Sleeps at most timo/hz seconds 450 * (0 means no timeout). If flags includes PCATCH flag, signals are checked 451 * before and after sleeping, else signals are not checked. Returns 0 if 452 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a 453 * signal needs to be delivered, ERESTART is returned if the current system 454 * call should be restarted if possible, and EINTR is returned if the system 455 * call should be interrupted by the signal (return EINTR). 456 * 457 * Note that if we are a process, we release_curproc() before messing with 458 * the LWKT scheduler. 459 * 460 * During autoconfiguration or after a panic, a sleep will simply 461 * lower the priority briefly to allow interrupts, then return. 462 */ 463 int 464 tsleep(const volatile void *ident, int flags, const char *wmesg, int timo) 465 { 466 struct thread *td = curthread; 467 struct lwp *lp = td->td_lwp; 468 struct proc *p = td->td_proc; /* may be NULL */ 469 globaldata_t gd; 470 int sig; 471 int catch; 472 int id; 473 int error; 474 int oldpri; 475 struct callout thandle; 476 477 /* 478 * NOTE: removed KTRPOINT, it could cause races due to blocking 479 * even in stable. Just scrap it for now. 480 */ 481 if (!tsleep_crypto_dump && (tsleep_now_works == 0 || panicstr)) { 482 /* 483 * After a panic, or before we actually have an operational 484 * softclock, just give interrupts a chance, then just return; 485 * 486 * don't run any other procs or panic below, 487 * in case this is the idle process and already asleep. 488 */ 489 splz(); 490 oldpri = td->td_pri; 491 lwkt_setpri_self(safepri); 492 lwkt_switch(); 493 lwkt_setpri_self(oldpri); 494 return (0); 495 } 496 logtsleep2(tsleep_beg, ident); 497 gd = td->td_gd; 498 KKASSERT(td != &gd->gd_idlethread); /* you must be kidding! */ 499 500 /* 501 * NOTE: all of this occurs on the current cpu, including any 502 * callout-based wakeups, so a critical section is a sufficient 503 * interlock. 504 * 505 * The entire sequence through to where we actually sleep must 506 * run without breaking the critical section. 507 */ 508 catch = flags & PCATCH; 509 error = 0; 510 sig = 0; 511 512 crit_enter_quick(td); 513 514 KASSERT(ident != NULL, ("tsleep: no ident")); 515 KASSERT(lp == NULL || 516 lp->lwp_stat == LSRUN || /* Obvious */ 517 lp->lwp_stat == LSSTOP, /* Set in tstop */ 518 ("tsleep %p %s %d", 519 ident, wmesg, lp->lwp_stat)); 520 521 /* 522 * Setup for the current process (if this is a process). 523 */ 524 if (lp) { 525 if (catch) { 526 /* 527 * Early termination if PCATCH was set and a 528 * signal is pending, interlocked with the 529 * critical section. 530 * 531 * Early termination only occurs when tsleep() is 532 * entered while in a normal LSRUN state. 533 */ 534 if ((sig = CURSIG(lp)) != 0) 535 goto resume; 536 537 /* 538 * Early termination if PCATCH was set and a 539 * mailbox signal was possibly delivered prior to 540 * the system call even being made, in order to 541 * allow the user to interlock without having to 542 * make additional system calls. 543 */ 544 if (p->p_flag & P_MAILBOX) 545 goto resume; 546 547 /* 548 * Causes ksignal to wake us up when. 549 */ 550 lp->lwp_flag |= LWP_SINTR; 551 } 552 } 553 554 /* 555 * We interlock the sleep queue if the caller has not already done 556 * it for us. 557 */ 558 if ((flags & PINTERLOCKED) == 0) { 559 id = LOOKUP(ident); 560 _tsleep_interlock(gd, ident, flags); 561 } 562 563 /* 564 * 565 * If no interlock was set we do an integrated interlock here. 566 * Make sure the current process has been untangled from 567 * the userland scheduler and initialize slptime to start 568 * counting. We must interlock the sleep queue before doing 569 * this to avoid wakeup/process-ipi races which can occur under 570 * heavy loads. 571 */ 572 if (lp) { 573 p->p_usched->release_curproc(lp); 574 lp->lwp_slptime = 0; 575 } 576 577 /* 578 * If the interlocked flag is set but our cpu bit in the slpqueue 579 * is no longer set, then a wakeup was processed inbetween the 580 * tsleep_interlock() (ours or the callers), and here. This can 581 * occur under numerous circumstances including when we release the 582 * current process. 583 * 584 * Extreme loads can cause the sending of an IPI (e.g. wakeup()'s) 585 * to process incoming IPIs, thus draining incoming wakeups. 586 */ 587 if ((td->td_flags & TDF_TSLEEPQ) == 0) { 588 logtsleep2(ilockfail, ident); 589 goto resume; 590 } 591 592 /* 593 * scheduling is blocked while in a critical section. Coincide 594 * the descheduled-by-tsleep flag with the descheduling of the 595 * lwkt. 596 */ 597 lwkt_deschedule_self(td); 598 td->td_flags |= TDF_TSLEEP_DESCHEDULED; 599 td->td_wmesg = wmesg; 600 601 /* 602 * Setup the timeout, if any 603 */ 604 if (timo) { 605 callout_init(&thandle); 606 callout_reset(&thandle, timo, endtsleep, td); 607 } 608 609 /* 610 * Beddy bye bye. 611 */ 612 if (lp) { 613 /* 614 * Ok, we are sleeping. Place us in the SSLEEP state. 615 */ 616 KKASSERT((lp->lwp_flag & LWP_ONRUNQ) == 0); 617 /* 618 * tstop() sets LSSTOP, so don't fiddle with that. 619 */ 620 if (lp->lwp_stat != LSSTOP) 621 lp->lwp_stat = LSSLEEP; 622 lp->lwp_ru.ru_nvcsw++; 623 lwkt_switch(); 624 625 /* 626 * And when we are woken up, put us back in LSRUN. If we 627 * slept for over a second, recalculate our estcpu. 628 */ 629 lp->lwp_stat = LSRUN; 630 if (lp->lwp_slptime) 631 p->p_usched->recalculate(lp); 632 lp->lwp_slptime = 0; 633 } else { 634 lwkt_switch(); 635 } 636 637 /* 638 * Make sure we haven't switched cpus while we were asleep. It's 639 * not supposed to happen. Cleanup our temporary flags. 640 */ 641 KKASSERT(gd == td->td_gd); 642 643 /* 644 * Cleanup the timeout. 645 */ 646 if (timo) { 647 if (td->td_flags & TDF_TIMEOUT) { 648 td->td_flags &= ~TDF_TIMEOUT; 649 error = EWOULDBLOCK; 650 } else { 651 callout_stop(&thandle); 652 } 653 } 654 655 /* 656 * Make sure we have been removed from the sleepq. This should 657 * have been done for us already. 658 * 659 * However, it is possible for a scheduling IPI to be in flight 660 * from a previous tsleep/tsleep_interlock or due to a straight-out 661 * call to lwkt_schedule() (in the case of an interrupt thread). 662 * So don't complain if DESCHEDULED is still set. 663 */ 664 _tsleep_remove(td); 665 td->td_wmesg = NULL; 666 if (td->td_flags & TDF_TSLEEP_DESCHEDULED) { 667 td->td_flags &= ~TDF_TSLEEP_DESCHEDULED; 668 } 669 670 /* 671 * Figure out the correct error return. If interrupted by a 672 * signal we want to return EINTR or ERESTART. 673 * 674 * If P_MAILBOX is set no automatic system call restart occurs 675 * and we return EINTR. P_MAILBOX is meant to be used as an 676 * interlock, the user must poll it prior to any system call 677 * that it wishes to interlock a mailbox signal against since 678 * the flag is cleared on *any* system call that sleeps. 679 */ 680 resume: 681 if (p) { 682 if (catch && error == 0) { 683 if ((p->p_flag & P_MAILBOX) && sig == 0) { 684 error = EINTR; 685 } else if (sig != 0 || (sig = CURSIG(lp))) { 686 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 687 error = EINTR; 688 else 689 error = ERESTART; 690 } 691 } 692 lp->lwp_flag &= ~(LWP_BREAKTSLEEP | LWP_SINTR); 693 p->p_flag &= ~P_MAILBOX; 694 } 695 logtsleep1(tsleep_end); 696 crit_exit_quick(td); 697 return (error); 698 } 699 700 /* 701 * Interlocked spinlock sleep. An exclusively held spinlock must 702 * be passed to ssleep(). The function will atomically release the 703 * spinlock and tsleep on the ident, then reacquire the spinlock and 704 * return. 705 * 706 * This routine is fairly important along the critical path, so optimize it 707 * heavily. 708 */ 709 int 710 ssleep(const volatile void *ident, struct spinlock *spin, int flags, 711 const char *wmesg, int timo) 712 { 713 globaldata_t gd = mycpu; 714 int error; 715 716 _tsleep_interlock(gd, ident, flags); 717 spin_unlock_quick(gd, spin); 718 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 719 spin_lock_quick(gd, spin); 720 721 return (error); 722 } 723 724 int 725 lksleep(const volatile void *ident, struct lock *lock, int flags, 726 const char *wmesg, int timo) 727 { 728 globaldata_t gd = mycpu; 729 int error; 730 731 _tsleep_interlock(gd, ident, flags); 732 lockmgr(lock, LK_RELEASE); 733 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 734 lockmgr(lock, LK_EXCLUSIVE); 735 736 return (error); 737 } 738 739 /* 740 * Interlocked mutex sleep. An exclusively held mutex must be passed 741 * to mtxsleep(). The function will atomically release the mutex 742 * and tsleep on the ident, then reacquire the mutex and return. 743 */ 744 int 745 mtxsleep(const volatile void *ident, struct mtx *mtx, int flags, 746 const char *wmesg, int timo) 747 { 748 globaldata_t gd = mycpu; 749 int error; 750 751 _tsleep_interlock(gd, ident, flags); 752 mtx_unlock(mtx); 753 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 754 mtx_lock_ex_quick(mtx, wmesg); 755 756 return (error); 757 } 758 759 /* 760 * Interlocked serializer sleep. An exclusively held serializer must 761 * be passed to zsleep(). The function will atomically release 762 * the serializer and tsleep on the ident, then reacquire the serializer 763 * and return. 764 */ 765 int 766 zsleep(const volatile void *ident, struct lwkt_serialize *slz, int flags, 767 const char *wmesg, int timo) 768 { 769 globaldata_t gd = mycpu; 770 int ret; 771 772 ASSERT_SERIALIZED(slz); 773 774 _tsleep_interlock(gd, ident, flags); 775 lwkt_serialize_exit(slz); 776 ret = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 777 lwkt_serialize_enter(slz); 778 779 return ret; 780 } 781 782 /* 783 * Directly block on the LWKT thread by descheduling it. This 784 * is much faster then tsleep(), but the only legal way to wake 785 * us up is to directly schedule the thread. 786 * 787 * Setting TDF_SINTR will cause new signals to directly schedule us. 788 * 789 * This routine must be called while in a critical section. 790 */ 791 int 792 lwkt_sleep(const char *wmesg, int flags) 793 { 794 thread_t td = curthread; 795 int sig; 796 797 if ((flags & PCATCH) == 0 || td->td_lwp == NULL) { 798 td->td_flags |= TDF_BLOCKED; 799 td->td_wmesg = wmesg; 800 lwkt_deschedule_self(td); 801 lwkt_switch(); 802 td->td_wmesg = NULL; 803 td->td_flags &= ~TDF_BLOCKED; 804 return(0); 805 } 806 if ((sig = CURSIG(td->td_lwp)) != 0) { 807 if (SIGISMEMBER(td->td_proc->p_sigacts->ps_sigintr, sig)) 808 return(EINTR); 809 else 810 return(ERESTART); 811 812 } 813 td->td_flags |= TDF_BLOCKED | TDF_SINTR; 814 td->td_wmesg = wmesg; 815 lwkt_deschedule_self(td); 816 lwkt_switch(); 817 td->td_flags &= ~(TDF_BLOCKED | TDF_SINTR); 818 td->td_wmesg = NULL; 819 return(0); 820 } 821 822 /* 823 * Implement the timeout for tsleep. 824 * 825 * We set LWP_BREAKTSLEEP to indicate that an event has occured, but 826 * we only call setrunnable if the process is not stopped. 827 * 828 * This type of callout timeout is scheduled on the same cpu the process 829 * is sleeping on. Also, at the moment, the MP lock is held. 830 */ 831 static void 832 endtsleep(void *arg) 833 { 834 thread_t td = arg; 835 struct lwp *lp; 836 837 crit_enter(); 838 lwkt_gettoken(&proc_token); 839 840 /* 841 * cpu interlock. Thread flags are only manipulated on 842 * the cpu owning the thread. proc flags are only manipulated 843 * by the older of the MP lock. We have both. 844 */ 845 if (td->td_flags & TDF_TSLEEP_DESCHEDULED) { 846 td->td_flags |= TDF_TIMEOUT; 847 848 if ((lp = td->td_lwp) != NULL) { 849 lp->lwp_flag |= LWP_BREAKTSLEEP; 850 if (lp->lwp_proc->p_stat != SSTOP) 851 setrunnable(lp); 852 } else { 853 _tsleep_wakeup(td); 854 } 855 } 856 lwkt_reltoken(&proc_token); 857 crit_exit(); 858 } 859 860 /* 861 * Make all processes sleeping on the specified identifier runnable. 862 * count may be zero or one only. 863 * 864 * The domain encodes the sleep/wakeup domain AND the first cpu to check 865 * (which is always the current cpu). As we iterate across cpus 866 * 867 * This call may run without the MP lock held. We can only manipulate thread 868 * state on the cpu owning the thread. We CANNOT manipulate process state 869 * at all. 870 * 871 * _wakeup() can be passed to an IPI so we can't use (const volatile 872 * void *ident). 873 */ 874 static void 875 _wakeup(void *ident, int domain) 876 { 877 struct tslpque *qp; 878 struct thread *td; 879 struct thread *ntd; 880 globaldata_t gd; 881 #ifdef SMP 882 cpumask_t mask; 883 #endif 884 int id; 885 886 crit_enter(); 887 logtsleep2(wakeup_beg, ident); 888 gd = mycpu; 889 id = LOOKUP(ident); 890 qp = &gd->gd_tsleep_hash[id]; 891 restart: 892 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) { 893 ntd = TAILQ_NEXT(td, td_sleepq); 894 if (td->td_wchan == ident && 895 td->td_wdomain == (domain & PDOMAIN_MASK) 896 ) { 897 KKASSERT(td->td_gd == gd); 898 _tsleep_remove(td); 899 if (td->td_flags & TDF_TSLEEP_DESCHEDULED) { 900 td->td_flags &= ~TDF_TSLEEP_DESCHEDULED; 901 lwkt_schedule(td); 902 if (domain & PWAKEUP_ONE) 903 goto done; 904 } 905 goto restart; 906 } 907 } 908 909 #ifdef SMP 910 /* 911 * We finished checking the current cpu but there still may be 912 * more work to do. Either wakeup_one was requested and no matching 913 * thread was found, or a normal wakeup was requested and we have 914 * to continue checking cpus. 915 * 916 * It should be noted that this scheme is actually less expensive then 917 * the old scheme when waking up multiple threads, since we send 918 * only one IPI message per target candidate which may then schedule 919 * multiple threads. Before we could have wound up sending an IPI 920 * message for each thread on the target cpu (!= current cpu) that 921 * needed to be woken up. 922 * 923 * NOTE: Wakeups occuring on remote cpus are asynchronous. This 924 * should be ok since we are passing idents in the IPI rather then 925 * thread pointers. 926 */ 927 if ((domain & PWAKEUP_MYCPU) == 0 && 928 (mask = slpque_cpumasks[id] & gd->gd_other_cpus) != 0) { 929 lwkt_send_ipiq2_mask(mask, _wakeup, ident, 930 domain | PWAKEUP_MYCPU); 931 } 932 #endif 933 done: 934 logtsleep1(wakeup_end); 935 crit_exit(); 936 } 937 938 /* 939 * Wakeup all threads tsleep()ing on the specified ident, on all cpus 940 */ 941 void 942 wakeup(const volatile void *ident) 943 { 944 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mycpu->gd_cpuid)); 945 } 946 947 /* 948 * Wakeup one thread tsleep()ing on the specified ident, on any cpu. 949 */ 950 void 951 wakeup_one(const volatile void *ident) 952 { 953 /* XXX potentially round-robin the first responding cpu */ 954 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mycpu->gd_cpuid) | PWAKEUP_ONE); 955 } 956 957 /* 958 * Wakeup threads tsleep()ing on the specified ident on the current cpu 959 * only. 960 */ 961 void 962 wakeup_mycpu(const volatile void *ident) 963 { 964 _wakeup(__DEALL(ident), PWAKEUP_MYCPU); 965 } 966 967 /* 968 * Wakeup one thread tsleep()ing on the specified ident on the current cpu 969 * only. 970 */ 971 void 972 wakeup_mycpu_one(const volatile void *ident) 973 { 974 /* XXX potentially round-robin the first responding cpu */ 975 _wakeup(__DEALL(ident), PWAKEUP_MYCPU|PWAKEUP_ONE); 976 } 977 978 /* 979 * Wakeup all thread tsleep()ing on the specified ident on the specified cpu 980 * only. 981 */ 982 void 983 wakeup_oncpu(globaldata_t gd, const volatile void *ident) 984 { 985 #ifdef SMP 986 if (gd == mycpu) { 987 _wakeup(__DEALL(ident), PWAKEUP_MYCPU); 988 } else { 989 lwkt_send_ipiq2(gd, _wakeup, __DEALL(ident), PWAKEUP_MYCPU); 990 } 991 #else 992 _wakeup(__DEALL(ident), PWAKEUP_MYCPU); 993 #endif 994 } 995 996 /* 997 * Wakeup one thread tsleep()ing on the specified ident on the specified cpu 998 * only. 999 */ 1000 void 1001 wakeup_oncpu_one(globaldata_t gd, const volatile void *ident) 1002 { 1003 #ifdef SMP 1004 if (gd == mycpu) { 1005 _wakeup(__DEALL(ident), PWAKEUP_MYCPU | PWAKEUP_ONE); 1006 } else { 1007 lwkt_send_ipiq2(gd, _wakeup, __DEALL(ident), 1008 PWAKEUP_MYCPU | PWAKEUP_ONE); 1009 } 1010 #else 1011 _wakeup(__DEALL(ident), PWAKEUP_MYCPU | PWAKEUP_ONE); 1012 #endif 1013 } 1014 1015 /* 1016 * Wakeup all threads waiting on the specified ident that slept using 1017 * the specified domain, on all cpus. 1018 */ 1019 void 1020 wakeup_domain(const volatile void *ident, int domain) 1021 { 1022 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(domain, mycpu->gd_cpuid)); 1023 } 1024 1025 /* 1026 * Wakeup one thread waiting on the specified ident that slept using 1027 * the specified domain, on any cpu. 1028 */ 1029 void 1030 wakeup_domain_one(const volatile void *ident, int domain) 1031 { 1032 /* XXX potentially round-robin the first responding cpu */ 1033 _wakeup(__DEALL(ident), 1034 PWAKEUP_ENCODE(domain, mycpu->gd_cpuid) | PWAKEUP_ONE); 1035 } 1036 1037 /* 1038 * setrunnable() 1039 * 1040 * Make a process runnable. The proc_token must be held on call. This only 1041 * has an effect if we are in SSLEEP. We only break out of the 1042 * tsleep if LWP_BREAKTSLEEP is set, otherwise we just fix-up the state. 1043 * 1044 * NOTE: With the MP lock held we can only safely manipulate the process 1045 * structure. We cannot safely manipulate the thread structure. 1046 */ 1047 void 1048 setrunnable(struct lwp *lp) 1049 { 1050 ASSERT_LWKT_TOKEN_HELD(&proc_token); 1051 crit_enter(); 1052 if (lp->lwp_stat == LSSTOP) 1053 lp->lwp_stat = LSSLEEP; 1054 if (lp->lwp_stat == LSSLEEP && (lp->lwp_flag & LWP_BREAKTSLEEP)) 1055 _tsleep_wakeup(lp->lwp_thread); 1056 crit_exit(); 1057 } 1058 1059 /* 1060 * The process is stopped due to some condition, usually because p_stat is 1061 * set to SSTOP, but also possibly due to being traced. 1062 * 1063 * NOTE! If the caller sets SSTOP, the caller must also clear P_WAITED 1064 * because the parent may check the child's status before the child actually 1065 * gets to this routine. 1066 * 1067 * This routine is called with the current lwp only, typically just 1068 * before returning to userland. 1069 * 1070 * Setting LWP_BREAKTSLEEP before entering the tsleep will cause a passive 1071 * SIGCONT to break out of the tsleep. 1072 */ 1073 void 1074 tstop(void) 1075 { 1076 struct lwp *lp = curthread->td_lwp; 1077 struct proc *p = lp->lwp_proc; 1078 1079 crit_enter(); 1080 /* 1081 * If LWP_WSTOP is set, we were sleeping 1082 * while our process was stopped. At this point 1083 * we were already counted as stopped. 1084 */ 1085 if ((lp->lwp_flag & LWP_WSTOP) == 0) { 1086 /* 1087 * If we're the last thread to stop, signal 1088 * our parent. 1089 */ 1090 p->p_nstopped++; 1091 lp->lwp_flag |= LWP_WSTOP; 1092 wakeup(&p->p_nstopped); 1093 if (p->p_nstopped == p->p_nthreads) { 1094 p->p_flag &= ~P_WAITED; 1095 wakeup(p->p_pptr); 1096 if ((p->p_pptr->p_sigacts->ps_flag & PS_NOCLDSTOP) == 0) 1097 ksignal(p->p_pptr, SIGCHLD); 1098 } 1099 } 1100 while (p->p_stat == SSTOP) { 1101 lp->lwp_flag |= LWP_BREAKTSLEEP; 1102 lp->lwp_stat = LSSTOP; 1103 tsleep(p, 0, "stop", 0); 1104 } 1105 p->p_nstopped--; 1106 lp->lwp_flag &= ~LWP_WSTOP; 1107 crit_exit(); 1108 } 1109 1110 /* 1111 * Compute a tenex style load average of a quantity on 1112 * 1, 5 and 15 minute intervals. 1113 */ 1114 static int loadav_count_runnable(struct lwp *p, void *data); 1115 1116 static void 1117 loadav(void *arg) 1118 { 1119 struct loadavg *avg; 1120 int i, nrun; 1121 1122 nrun = 0; 1123 alllwp_scan(loadav_count_runnable, &nrun); 1124 avg = &averunnable; 1125 for (i = 0; i < 3; i++) { 1126 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] + 1127 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT; 1128 } 1129 1130 /* 1131 * Schedule the next update to occur after 5 seconds, but add a 1132 * random variation to avoid synchronisation with processes that 1133 * run at regular intervals. 1134 */ 1135 callout_reset(&loadav_callout, hz * 4 + (int)(krandom() % (hz * 2 + 1)), 1136 loadav, NULL); 1137 } 1138 1139 static int 1140 loadav_count_runnable(struct lwp *lp, void *data) 1141 { 1142 int *nrunp = data; 1143 thread_t td; 1144 1145 switch (lp->lwp_stat) { 1146 case LSRUN: 1147 if ((td = lp->lwp_thread) == NULL) 1148 break; 1149 if (td->td_flags & TDF_BLOCKED) 1150 break; 1151 ++*nrunp; 1152 break; 1153 default: 1154 break; 1155 } 1156 return(0); 1157 } 1158 1159 /* ARGSUSED */ 1160 static void 1161 sched_setup(void *dummy) 1162 { 1163 callout_init(&loadav_callout); 1164 callout_init(&schedcpu_callout); 1165 1166 /* Kick off timeout driven events by calling first time. */ 1167 schedcpu(NULL); 1168 loadav(NULL); 1169 } 1170 1171