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. Neither the name of the University nor the names of its contributors 19 * may be used to endorse or promote products derived from this software 20 * without specific prior written permission. 21 * 22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 * 34 * @(#)kern_synch.c 8.9 (Berkeley) 5/19/95 35 * $FreeBSD: src/sys/kern/kern_synch.c,v 1.87.2.6 2002/10/13 07:29:53 kbyanc Exp $ 36 */ 37 38 #include "opt_ktrace.h" 39 40 #include <sys/param.h> 41 #include <sys/systm.h> 42 #include <sys/proc.h> 43 #include <sys/kernel.h> 44 #include <sys/signalvar.h> 45 #include <sys/resourcevar.h> 46 #include <sys/vmmeter.h> 47 #include <sys/sysctl.h> 48 #include <sys/lock.h> 49 #include <sys/uio.h> 50 #ifdef KTRACE 51 #include <sys/ktrace.h> 52 #endif 53 #include <sys/ktr.h> 54 #include <sys/serialize.h> 55 56 #include <sys/signal2.h> 57 #include <sys/thread2.h> 58 #include <sys/spinlock2.h> 59 #include <sys/mutex2.h> 60 61 #include <machine/cpu.h> 62 #include <machine/smp.h> 63 64 TAILQ_HEAD(tslpque, thread); 65 66 static void sched_setup (void *dummy); 67 SYSINIT(sched_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, sched_setup, NULL); 68 69 int lbolt; 70 void *lbolt_syncer; 71 int sched_quantum; /* Roundrobin scheduling quantum in ticks. */ 72 int ncpus; 73 int ncpus2, ncpus2_shift, ncpus2_mask; /* note: mask not cpumask_t */ 74 int ncpus_fit, ncpus_fit_mask; /* note: mask not cpumask_t */ 75 int safepri; 76 int tsleep_now_works; 77 int tsleep_crypto_dump = 0; 78 79 static struct callout loadav_callout; 80 static struct callout schedcpu_callout; 81 MALLOC_DEFINE(M_TSLEEP, "tslpque", "tsleep queues"); 82 83 #define __DEALL(ident) __DEQUALIFY(void *, ident) 84 85 #if !defined(KTR_TSLEEP) 86 #define KTR_TSLEEP KTR_ALL 87 #endif 88 KTR_INFO_MASTER(tsleep); 89 KTR_INFO(KTR_TSLEEP, tsleep, tsleep_beg, 0, "tsleep enter %p", const volatile void *ident); 90 KTR_INFO(KTR_TSLEEP, tsleep, tsleep_end, 1, "tsleep exit"); 91 KTR_INFO(KTR_TSLEEP, tsleep, wakeup_beg, 2, "wakeup enter %p", const volatile void *ident); 92 KTR_INFO(KTR_TSLEEP, tsleep, wakeup_end, 3, "wakeup exit"); 93 KTR_INFO(KTR_TSLEEP, tsleep, ilockfail, 4, "interlock failed %p", const volatile void *ident); 94 95 #define logtsleep1(name) KTR_LOG(tsleep_ ## name) 96 #define logtsleep2(name, val) KTR_LOG(tsleep_ ## name, val) 97 98 struct loadavg averunnable = 99 { {0, 0, 0}, FSCALE }; /* load average, of runnable procs */ 100 /* 101 * Constants for averages over 1, 5, and 15 minutes 102 * when sampling at 5 second intervals. 103 */ 104 static fixpt_t cexp[3] = { 105 0.9200444146293232 * FSCALE, /* exp(-1/12) */ 106 0.9834714538216174 * FSCALE, /* exp(-1/60) */ 107 0.9944598480048967 * FSCALE, /* exp(-1/180) */ 108 }; 109 110 static void endtsleep (void *); 111 static void loadav (void *arg); 112 static void schedcpu (void *arg); 113 114 /* 115 * Adjust the scheduler quantum. The quantum is specified in microseconds. 116 * Note that 'tick' is in microseconds per tick. 117 */ 118 static int 119 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS) 120 { 121 int error, new_val; 122 123 new_val = sched_quantum * ustick; 124 error = sysctl_handle_int(oidp, &new_val, 0, req); 125 if (error != 0 || req->newptr == NULL) 126 return (error); 127 if (new_val < ustick) 128 return (EINVAL); 129 sched_quantum = new_val / ustick; 130 return (0); 131 } 132 133 SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW, 134 0, sizeof sched_quantum, sysctl_kern_quantum, "I", ""); 135 136 static int pctcpu_decay = 10; 137 SYSCTL_INT(_kern, OID_AUTO, pctcpu_decay, CTLFLAG_RW, &pctcpu_decay, 0, ""); 138 139 /* 140 * kernel uses `FSCALE', userland (SHOULD) use kern.fscale 141 */ 142 int fscale __unused = FSCALE; /* exported to systat */ 143 SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, ""); 144 145 /* 146 * Recompute process priorities, once a second. 147 * 148 * Since the userland schedulers are typically event oriented, if the 149 * estcpu calculation at wakeup() time is not sufficient to make a 150 * process runnable relative to other processes in the system we have 151 * a 1-second recalc to help out. 152 * 153 * This code also allows us to store sysclock_t data in the process structure 154 * without fear of an overrun, since sysclock_t are guarenteed to hold 155 * several seconds worth of count. 156 * 157 * WARNING! callouts can preempt normal threads. However, they will not 158 * preempt a thread holding a spinlock so we *can* safely use spinlocks. 159 */ 160 static int schedcpu_stats(struct proc *p, void *data __unused); 161 static int schedcpu_resource(struct proc *p, void *data __unused); 162 163 static void 164 schedcpu(void *arg) 165 { 166 allproc_scan(schedcpu_stats, NULL); 167 allproc_scan(schedcpu_resource, NULL); 168 wakeup((caddr_t)&lbolt); 169 wakeup(lbolt_syncer); 170 callout_reset(&schedcpu_callout, hz, schedcpu, NULL); 171 } 172 173 /* 174 * General process statistics once a second 175 */ 176 static int 177 schedcpu_stats(struct proc *p, void *data __unused) 178 { 179 struct lwp *lp; 180 181 /* 182 * Threads may not be completely set up if process in SIDL state. 183 */ 184 if (p->p_stat == SIDL) 185 return(0); 186 187 PHOLD(p); 188 if (lwkt_trytoken(&p->p_token) == FALSE) { 189 PRELE(p); 190 return(0); 191 } 192 193 p->p_swtime++; 194 FOREACH_LWP_IN_PROC(lp, p) { 195 if (lp->lwp_stat == LSSLEEP) { 196 ++lp->lwp_slptime; 197 if (lp->lwp_slptime == 1) 198 p->p_usched->uload_update(lp); 199 } 200 201 /* 202 * Only recalculate processes that are active or have slept 203 * less then 2 seconds. The schedulers understand this. 204 * Otherwise decay by 50% per second. 205 */ 206 if (lp->lwp_slptime <= 1) { 207 p->p_usched->recalculate(lp); 208 } else { 209 int decay; 210 211 decay = pctcpu_decay; 212 cpu_ccfence(); 213 if (decay <= 1) 214 decay = 1; 215 if (decay > 100) 216 decay = 100; 217 lp->lwp_pctcpu = (lp->lwp_pctcpu * (decay - 1)) / decay; 218 } 219 } 220 lwkt_reltoken(&p->p_token); 221 lwkt_yield(); 222 PRELE(p); 223 return(0); 224 } 225 226 /* 227 * Resource checks. XXX break out since ksignal/killproc can block, 228 * limiting us to one process killed per second. There is probably 229 * a better way. 230 */ 231 static int 232 schedcpu_resource(struct proc *p, void *data __unused) 233 { 234 u_int64_t ttime; 235 struct lwp *lp; 236 237 if (p->p_stat == SIDL) 238 return(0); 239 240 PHOLD(p); 241 if (lwkt_trytoken(&p->p_token) == FALSE) { 242 PRELE(p); 243 return(0); 244 } 245 246 if (p->p_stat == SZOMB || p->p_limit == NULL) { 247 lwkt_reltoken(&p->p_token); 248 PRELE(p); 249 return(0); 250 } 251 252 ttime = 0; 253 FOREACH_LWP_IN_PROC(lp, p) { 254 /* 255 * We may have caught an lp in the middle of being 256 * created, lwp_thread can be NULL. 257 */ 258 if (lp->lwp_thread) { 259 ttime += lp->lwp_thread->td_sticks; 260 ttime += lp->lwp_thread->td_uticks; 261 } 262 } 263 264 switch(plimit_testcpulimit(p->p_limit, ttime)) { 265 case PLIMIT_TESTCPU_KILL: 266 killproc(p, "exceeded maximum CPU limit"); 267 break; 268 case PLIMIT_TESTCPU_XCPU: 269 if ((p->p_flags & P_XCPU) == 0) { 270 p->p_flags |= P_XCPU; 271 ksignal(p, SIGXCPU); 272 } 273 break; 274 default: 275 break; 276 } 277 lwkt_reltoken(&p->p_token); 278 lwkt_yield(); 279 PRELE(p); 280 return(0); 281 } 282 283 /* 284 * This is only used by ps. Generate a cpu percentage use over 285 * a period of one second. 286 */ 287 void 288 updatepcpu(struct lwp *lp, int cpticks, int ttlticks) 289 { 290 fixpt_t acc; 291 int remticks; 292 293 acc = (cpticks << FSHIFT) / ttlticks; 294 if (ttlticks >= ESTCPUFREQ) { 295 lp->lwp_pctcpu = acc; 296 } else { 297 remticks = ESTCPUFREQ - ttlticks; 298 lp->lwp_pctcpu = (acc * ttlticks + lp->lwp_pctcpu * remticks) / 299 ESTCPUFREQ; 300 } 301 } 302 303 /* 304 * tsleep/wakeup hash table parameters. Try to find the sweet spot for 305 * like addresses being slept on. 306 */ 307 #define TABLESIZE 4001 308 #define LOOKUP(x) (((u_int)(uintptr_t)(x)) % TABLESIZE) 309 310 static cpumask_t slpque_cpumasks[TABLESIZE]; 311 312 /* 313 * General scheduler initialization. We force a reschedule 25 times 314 * a second by default. Note that cpu0 is initialized in early boot and 315 * cannot make any high level calls. 316 * 317 * Each cpu has its own sleep queue. 318 */ 319 void 320 sleep_gdinit(globaldata_t gd) 321 { 322 static struct tslpque slpque_cpu0[TABLESIZE]; 323 int i; 324 325 if (gd->gd_cpuid == 0) { 326 sched_quantum = (hz + 24) / 25; 327 gd->gd_tsleep_hash = slpque_cpu0; 328 } else { 329 gd->gd_tsleep_hash = kmalloc(sizeof(slpque_cpu0), 330 M_TSLEEP, M_WAITOK | M_ZERO); 331 } 332 for (i = 0; i < TABLESIZE; ++i) 333 TAILQ_INIT(&gd->gd_tsleep_hash[i]); 334 } 335 336 /* 337 * This is a dandy function that allows us to interlock tsleep/wakeup 338 * operations with unspecified upper level locks, such as lockmgr locks, 339 * simply by holding a critical section. The sequence is: 340 * 341 * (acquire upper level lock) 342 * tsleep_interlock(blah) 343 * (release upper level lock) 344 * tsleep(blah, ...) 345 * 346 * Basically this functions queues us on the tsleep queue without actually 347 * descheduling us. When tsleep() is later called with PINTERLOCK it 348 * assumes the thread was already queued, otherwise it queues it there. 349 * 350 * Thus it is possible to receive the wakeup prior to going to sleep and 351 * the race conditions are covered. 352 */ 353 static __inline void 354 _tsleep_interlock(globaldata_t gd, const volatile void *ident, int flags) 355 { 356 thread_t td = gd->gd_curthread; 357 int id; 358 359 crit_enter_quick(td); 360 if (td->td_flags & TDF_TSLEEPQ) { 361 id = LOOKUP(td->td_wchan); 362 TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_sleepq); 363 if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL) { 364 ATOMIC_CPUMASK_NANDBIT(slpque_cpumasks[id], 365 gd->gd_cpuid); 366 } 367 } else { 368 td->td_flags |= TDF_TSLEEPQ; 369 } 370 id = LOOKUP(ident); 371 TAILQ_INSERT_TAIL(&gd->gd_tsleep_hash[id], td, td_sleepq); 372 ATOMIC_CPUMASK_ORBIT(slpque_cpumasks[id], gd->gd_cpuid); 373 td->td_wchan = ident; 374 td->td_wdomain = flags & PDOMAIN_MASK; 375 crit_exit_quick(td); 376 } 377 378 void 379 tsleep_interlock(const volatile void *ident, int flags) 380 { 381 _tsleep_interlock(mycpu, ident, flags); 382 } 383 384 /* 385 * Remove thread from sleepq. Must be called with a critical section held. 386 * The thread must not be migrating. 387 */ 388 static __inline void 389 _tsleep_remove(thread_t td) 390 { 391 globaldata_t gd = mycpu; 392 int id; 393 394 KKASSERT(td->td_gd == gd && IN_CRITICAL_SECT(td)); 395 KKASSERT((td->td_flags & TDF_MIGRATING) == 0); 396 if (td->td_flags & TDF_TSLEEPQ) { 397 td->td_flags &= ~TDF_TSLEEPQ; 398 id = LOOKUP(td->td_wchan); 399 TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_sleepq); 400 if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL) { 401 ATOMIC_CPUMASK_NANDBIT(slpque_cpumasks[id], 402 gd->gd_cpuid); 403 } 404 td->td_wchan = NULL; 405 td->td_wdomain = 0; 406 } 407 } 408 409 void 410 tsleep_remove(thread_t td) 411 { 412 _tsleep_remove(td); 413 } 414 415 /* 416 * General sleep call. Suspends the current process until a wakeup is 417 * performed on the specified identifier. The process will then be made 418 * runnable with the specified priority. Sleeps at most timo/hz seconds 419 * (0 means no timeout). If flags includes PCATCH flag, signals are checked 420 * before and after sleeping, else signals are not checked. Returns 0 if 421 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a 422 * signal needs to be delivered, ERESTART is returned if the current system 423 * call should be restarted if possible, and EINTR is returned if the system 424 * call should be interrupted by the signal (return EINTR). 425 * 426 * Note that if we are a process, we release_curproc() before messing with 427 * the LWKT scheduler. 428 * 429 * During autoconfiguration or after a panic, a sleep will simply 430 * lower the priority briefly to allow interrupts, then return. 431 * 432 * WARNING! This code can't block (short of switching away), or bad things 433 * will happen. No getting tokens, no blocking locks, etc. 434 */ 435 int 436 tsleep(const volatile void *ident, int flags, const char *wmesg, int timo) 437 { 438 struct thread *td = curthread; 439 struct lwp *lp = td->td_lwp; 440 struct proc *p = td->td_proc; /* may be NULL */ 441 globaldata_t gd; 442 int sig; 443 int catch; 444 int error; 445 int oldpri; 446 struct callout thandle; 447 448 /* 449 * Currently a severe hack. Make sure any delayed wakeups 450 * are flushed before we sleep or we might deadlock on whatever 451 * event we are sleeping on. 452 */ 453 if (td->td_flags & TDF_DELAYED_WAKEUP) 454 wakeup_end_delayed(); 455 456 /* 457 * NOTE: removed KTRPOINT, it could cause races due to blocking 458 * even in stable. Just scrap it for now. 459 */ 460 if (!tsleep_crypto_dump && (tsleep_now_works == 0 || panicstr)) { 461 /* 462 * After a panic, or before we actually have an operational 463 * softclock, just give interrupts a chance, then just return; 464 * 465 * don't run any other procs or panic below, 466 * in case this is the idle process and already asleep. 467 */ 468 splz(); 469 oldpri = td->td_pri; 470 lwkt_setpri_self(safepri); 471 lwkt_switch(); 472 lwkt_setpri_self(oldpri); 473 return (0); 474 } 475 logtsleep2(tsleep_beg, ident); 476 gd = td->td_gd; 477 KKASSERT(td != &gd->gd_idlethread); /* you must be kidding! */ 478 td->td_wakefromcpu = -1; /* overwritten by _wakeup */ 479 480 /* 481 * NOTE: all of this occurs on the current cpu, including any 482 * callout-based wakeups, so a critical section is a sufficient 483 * interlock. 484 * 485 * The entire sequence through to where we actually sleep must 486 * run without breaking the critical section. 487 */ 488 catch = flags & PCATCH; 489 error = 0; 490 sig = 0; 491 492 crit_enter_quick(td); 493 494 KASSERT(ident != NULL, ("tsleep: no ident")); 495 KASSERT(lp == NULL || 496 lp->lwp_stat == LSRUN || /* Obvious */ 497 lp->lwp_stat == LSSTOP, /* Set in tstop */ 498 ("tsleep %p %s %d", 499 ident, wmesg, lp->lwp_stat)); 500 501 /* 502 * We interlock the sleep queue if the caller has not already done 503 * it for us. This must be done before we potentially acquire any 504 * tokens or we can loose the wakeup. 505 */ 506 if ((flags & PINTERLOCKED) == 0) { 507 _tsleep_interlock(gd, ident, flags); 508 } 509 510 /* 511 * Setup for the current process (if this is a process). We must 512 * interlock with lwp_token to avoid remote wakeup races via 513 * setrunnable() 514 */ 515 if (lp) { 516 lwkt_gettoken(&lp->lwp_token); 517 518 /* 519 * If the umbrella process is in the SCORE state then 520 * make sure that the thread is flagged going into a 521 * normal sleep to allow the core dump to proceed, otherwise 522 * the coredump can end up waiting forever. If the normal 523 * sleep is woken up, the thread will enter a stopped state 524 * upon return to userland. 525 * 526 * We do not want to interrupt or cause a thread exist at 527 * this juncture because that will mess-up the state the 528 * coredump is trying to save. 529 */ 530 if (p->p_stat == SCORE && 531 (lp->lwp_mpflags & LWP_MP_WSTOP) == 0) { 532 atomic_set_int(&lp->lwp_mpflags, LWP_MP_WSTOP); 533 ++p->p_nstopped; 534 } 535 536 /* 537 * PCATCH requested. 538 */ 539 if (catch) { 540 /* 541 * Early termination if PCATCH was set and a 542 * signal is pending, interlocked with the 543 * critical section. 544 * 545 * Early termination only occurs when tsleep() is 546 * entered while in a normal LSRUN state. 547 */ 548 if ((sig = CURSIG(lp)) != 0) 549 goto resume; 550 551 /* 552 * Causes ksignal to wake us up if a signal is 553 * received (interlocked with p->p_token). 554 */ 555 lp->lwp_flags |= LWP_SINTR; 556 } 557 } else { 558 KKASSERT(p == NULL); 559 } 560 561 /* 562 * Make sure the current process has been untangled from 563 * the userland scheduler and initialize slptime to start 564 * counting. 565 * 566 * NOTE: td->td_wakefromcpu is pre-set by the release function 567 * for the dfly scheduler, and then adjusted by _wakeup() 568 */ 569 if (lp) { 570 p->p_usched->release_curproc(lp); 571 lp->lwp_slptime = 0; 572 } 573 574 /* 575 * If the interlocked flag is set but our cpu bit in the slpqueue 576 * is no longer set, then a wakeup was processed inbetween the 577 * tsleep_interlock() (ours or the callers), and here. This can 578 * occur under numerous circumstances including when we release the 579 * current process. 580 * 581 * Extreme loads can cause the sending of an IPI (e.g. wakeup()'s) 582 * to process incoming IPIs, thus draining incoming wakeups. 583 */ 584 if ((td->td_flags & TDF_TSLEEPQ) == 0) { 585 logtsleep2(ilockfail, ident); 586 goto resume; 587 } 588 589 /* 590 * scheduling is blocked while in a critical section. Coincide 591 * the descheduled-by-tsleep flag with the descheduling of the 592 * lwkt. 593 * 594 * The timer callout is localized on our cpu and interlocked by 595 * our critical section. 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. The timeout is only operable while 603 * the thread is flagged descheduled. 604 */ 605 KKASSERT((td->td_flags & TDF_TIMEOUT) == 0); 606 if (timo) { 607 callout_init_mp(&thandle); 608 callout_reset(&thandle, timo, endtsleep, td); 609 } 610 611 /* 612 * Beddy bye bye. 613 */ 614 if (lp) { 615 /* 616 * Ok, we are sleeping. Place us in the SSLEEP state. 617 */ 618 KKASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) == 0); 619 620 /* 621 * tstop() sets LSSTOP, so don't fiddle with that. 622 */ 623 if (lp->lwp_stat != LSSTOP) 624 lp->lwp_stat = LSSLEEP; 625 lp->lwp_ru.ru_nvcsw++; 626 p->p_usched->uload_update(lp); 627 lwkt_switch(); 628 629 /* 630 * And when we are woken up, put us back in LSRUN. If we 631 * slept for over a second, recalculate our estcpu. 632 */ 633 lp->lwp_stat = LSRUN; 634 if (lp->lwp_slptime) { 635 p->p_usched->uload_update(lp); 636 p->p_usched->recalculate(lp); 637 } 638 lp->lwp_slptime = 0; 639 } else { 640 lwkt_switch(); 641 } 642 643 /* 644 * Make sure we haven't switched cpus while we were asleep. It's 645 * not supposed to happen. Cleanup our temporary flags. 646 */ 647 KKASSERT(gd == td->td_gd); 648 649 /* 650 * Cleanup the timeout. If the timeout has already occured thandle 651 * has already been stopped, otherwise stop thandle. If the timeout 652 * is running (the callout thread must be blocked trying to get 653 * lwp_token) then wait for us to get scheduled. 654 */ 655 if (timo) { 656 while (td->td_flags & TDF_TIMEOUT_RUNNING) { 657 /* else we won't get rescheduled! */ 658 if (lp->lwp_stat != LSSTOP) 659 lp->lwp_stat = LSSLEEP; 660 lwkt_deschedule_self(td); 661 td->td_wmesg = "tsrace"; 662 lwkt_switch(); 663 kprintf("td %p %s: timeout race\n", td, td->td_comm); 664 } 665 if (td->td_flags & TDF_TIMEOUT) { 666 td->td_flags &= ~TDF_TIMEOUT; 667 error = EWOULDBLOCK; 668 } else { 669 /* does not block when on same cpu */ 670 callout_stop(&thandle); 671 } 672 } 673 td->td_flags &= ~TDF_TSLEEP_DESCHEDULED; 674 675 /* 676 * Make sure we have been removed from the sleepq. In most 677 * cases this will have been done for us already but it is 678 * possible for a scheduling IPI to be in-flight from a 679 * previous tsleep/tsleep_interlock() or due to a straight-out 680 * call to lwkt_schedule() (in the case of an interrupt thread), 681 * causing a spurious wakeup. 682 */ 683 _tsleep_remove(td); 684 td->td_wmesg = NULL; 685 686 /* 687 * Figure out the correct error return. If interrupted by a 688 * signal we want to return EINTR or ERESTART. 689 */ 690 resume: 691 if (lp) { 692 if (catch && error == 0) { 693 if (sig != 0 || (sig = CURSIG(lp))) { 694 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 695 error = EINTR; 696 else 697 error = ERESTART; 698 } 699 } 700 701 lp->lwp_flags &= ~LWP_SINTR; 702 703 /* 704 * Unconditionally set us to LSRUN on resume. lwp_stat could 705 * be in a weird state due to the goto resume, particularly 706 * when tsleep() is called from tstop(). 707 */ 708 lp->lwp_stat = LSRUN; 709 lwkt_reltoken(&lp->lwp_token); 710 } 711 logtsleep1(tsleep_end); 712 crit_exit_quick(td); 713 return (error); 714 } 715 716 /* 717 * Interlocked spinlock sleep. An exclusively held spinlock must 718 * be passed to ssleep(). The function will atomically release the 719 * spinlock and tsleep on the ident, then reacquire the spinlock and 720 * return. 721 * 722 * This routine is fairly important along the critical path, so optimize it 723 * heavily. 724 */ 725 int 726 ssleep(const volatile void *ident, struct spinlock *spin, int flags, 727 const char *wmesg, int timo) 728 { 729 globaldata_t gd = mycpu; 730 int error; 731 732 _tsleep_interlock(gd, ident, flags); 733 spin_unlock_quick(gd, spin); 734 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 735 _spin_lock_quick(gd, spin, wmesg); 736 737 return (error); 738 } 739 740 int 741 lksleep(const volatile void *ident, struct lock *lock, int flags, 742 const char *wmesg, int timo) 743 { 744 globaldata_t gd = mycpu; 745 int error; 746 747 _tsleep_interlock(gd, ident, flags); 748 lockmgr(lock, LK_RELEASE); 749 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 750 lockmgr(lock, LK_EXCLUSIVE); 751 752 return (error); 753 } 754 755 /* 756 * Interlocked mutex sleep. An exclusively held mutex must be passed 757 * to mtxsleep(). The function will atomically release the mutex 758 * and tsleep on the ident, then reacquire the mutex and return. 759 */ 760 int 761 mtxsleep(const volatile void *ident, struct mtx *mtx, int flags, 762 const char *wmesg, int timo) 763 { 764 globaldata_t gd = mycpu; 765 int error; 766 767 _tsleep_interlock(gd, ident, flags); 768 mtx_unlock(mtx); 769 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 770 mtx_lock_ex_quick(mtx); 771 772 return (error); 773 } 774 775 /* 776 * Interlocked serializer sleep. An exclusively held serializer must 777 * be passed to zsleep(). The function will atomically release 778 * the serializer and tsleep on the ident, then reacquire the serializer 779 * and return. 780 */ 781 int 782 zsleep(const volatile void *ident, struct lwkt_serialize *slz, int flags, 783 const char *wmesg, int timo) 784 { 785 globaldata_t gd = mycpu; 786 int ret; 787 788 ASSERT_SERIALIZED(slz); 789 790 _tsleep_interlock(gd, ident, flags); 791 lwkt_serialize_exit(slz); 792 ret = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 793 lwkt_serialize_enter(slz); 794 795 return ret; 796 } 797 798 /* 799 * Directly block on the LWKT thread by descheduling it. This 800 * is much faster then tsleep(), but the only legal way to wake 801 * us up is to directly schedule the thread. 802 * 803 * Setting TDF_SINTR will cause new signals to directly schedule us. 804 * 805 * This routine must be called while in a critical section. 806 */ 807 int 808 lwkt_sleep(const char *wmesg, int flags) 809 { 810 thread_t td = curthread; 811 int sig; 812 813 if ((flags & PCATCH) == 0 || td->td_lwp == NULL) { 814 td->td_flags |= TDF_BLOCKED; 815 td->td_wmesg = wmesg; 816 lwkt_deschedule_self(td); 817 lwkt_switch(); 818 td->td_wmesg = NULL; 819 td->td_flags &= ~TDF_BLOCKED; 820 return(0); 821 } 822 if ((sig = CURSIG(td->td_lwp)) != 0) { 823 if (SIGISMEMBER(td->td_proc->p_sigacts->ps_sigintr, sig)) 824 return(EINTR); 825 else 826 return(ERESTART); 827 828 } 829 td->td_flags |= TDF_BLOCKED | TDF_SINTR; 830 td->td_wmesg = wmesg; 831 lwkt_deschedule_self(td); 832 lwkt_switch(); 833 td->td_flags &= ~(TDF_BLOCKED | TDF_SINTR); 834 td->td_wmesg = NULL; 835 return(0); 836 } 837 838 /* 839 * Implement the timeout for tsleep. 840 * 841 * This type of callout timeout is scheduled on the same cpu the process 842 * is sleeping on. Also, at the moment, the MP lock is held. 843 */ 844 static void 845 endtsleep(void *arg) 846 { 847 thread_t td = arg; 848 struct lwp *lp; 849 850 /* 851 * We are going to have to get the lwp_token, which means we might 852 * block. This can race a tsleep getting woken up by other means 853 * so set TDF_TIMEOUT_RUNNING to force the tsleep to wait for our 854 * processing to complete (sorry tsleep!). 855 * 856 * We can safely set td_flags because td MUST be on the same cpu 857 * as we are. 858 */ 859 KKASSERT(td->td_gd == mycpu); 860 crit_enter(); 861 td->td_flags |= TDF_TIMEOUT_RUNNING | TDF_TIMEOUT; 862 863 /* 864 * This can block but TDF_TIMEOUT_RUNNING will prevent the thread 865 * from exiting the tsleep on us. The flag is interlocked by virtue 866 * of lp being on the same cpu as we are. 867 */ 868 if ((lp = td->td_lwp) != NULL) 869 lwkt_gettoken(&lp->lwp_token); 870 871 KKASSERT(td->td_flags & TDF_TSLEEP_DESCHEDULED); 872 873 if (lp) { 874 /* 875 * callout timer should normally never be set in tstop() 876 * because it passes a timeout of 0. However, there is a 877 * case during thread exit (which SSTOP's all the threads) 878 * for which tstop() must break out and can (properly) leave 879 * the thread in LSSTOP. 880 */ 881 KKASSERT(lp->lwp_stat != LSSTOP || 882 (lp->lwp_mpflags & LWP_MP_WEXIT)); 883 setrunnable(lp); 884 lwkt_reltoken(&lp->lwp_token); 885 } else { 886 _tsleep_remove(td); 887 lwkt_schedule(td); 888 } 889 KKASSERT(td->td_gd == mycpu); 890 td->td_flags &= ~TDF_TIMEOUT_RUNNING; 891 crit_exit(); 892 } 893 894 /* 895 * Make all processes sleeping on the specified identifier runnable. 896 * count may be zero or one only. 897 * 898 * The domain encodes the sleep/wakeup domain, flags, plus the originating 899 * cpu. 900 * 901 * This call may run without the MP lock held. We can only manipulate thread 902 * state on the cpu owning the thread. We CANNOT manipulate process state 903 * at all. 904 * 905 * _wakeup() can be passed to an IPI so we can't use (const volatile 906 * void *ident). 907 */ 908 static void 909 _wakeup(void *ident, int domain) 910 { 911 struct tslpque *qp; 912 struct thread *td; 913 struct thread *ntd; 914 globaldata_t gd; 915 cpumask_t mask; 916 int id; 917 918 crit_enter(); 919 logtsleep2(wakeup_beg, ident); 920 gd = mycpu; 921 id = LOOKUP(ident); 922 qp = &gd->gd_tsleep_hash[id]; 923 restart: 924 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) { 925 ntd = TAILQ_NEXT(td, td_sleepq); 926 if (td->td_wchan == ident && 927 td->td_wdomain == (domain & PDOMAIN_MASK) 928 ) { 929 KKASSERT(td->td_gd == gd); 930 _tsleep_remove(td); 931 td->td_wakefromcpu = PWAKEUP_DECODE(domain); 932 if (td->td_flags & TDF_TSLEEP_DESCHEDULED) { 933 lwkt_schedule(td); 934 if (domain & PWAKEUP_ONE) 935 goto done; 936 } 937 goto restart; 938 } 939 } 940 941 /* 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]; 960 CPUMASK_ANDMASK(mask, gd->gd_other_cpus); 961 if (CPUMASK_TESTNZERO(mask)) { 962 lwkt_send_ipiq2_mask(mask, _wakeup, ident, 963 domain | PWAKEUP_MYCPU); 964 } 965 } 966 done: 967 logtsleep1(wakeup_end); 968 crit_exit(); 969 } 970 971 /* 972 * Wakeup all threads tsleep()ing on the specified ident, on all cpus 973 */ 974 void 975 wakeup(const volatile void *ident) 976 { 977 globaldata_t gd = mycpu; 978 thread_t td = gd->gd_curthread; 979 980 if (td && (td->td_flags & TDF_DELAYED_WAKEUP)) { 981 /* 982 * If we are in a delayed wakeup section, record up to two wakeups in 983 * a per-CPU queue and issue them when we block or exit the delayed 984 * wakeup section. 985 */ 986 if (atomic_cmpset_ptr(&gd->gd_delayed_wakeup[0], NULL, ident)) 987 return; 988 if (atomic_cmpset_ptr(&gd->gd_delayed_wakeup[1], NULL, ident)) 989 return; 990 991 ident = atomic_swap_ptr(__DEQUALIFY(volatile void **, &gd->gd_delayed_wakeup[1]), 992 __DEALL(ident)); 993 ident = atomic_swap_ptr(__DEQUALIFY(volatile void **, &gd->gd_delayed_wakeup[0]), 994 __DEALL(ident)); 995 } 996 997 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, gd->gd_cpuid)); 998 } 999 1000 /* 1001 * Wakeup one thread tsleep()ing on the specified ident, on any cpu. 1002 */ 1003 void 1004 wakeup_one(const volatile void *ident) 1005 { 1006 /* XXX potentially round-robin the first responding cpu */ 1007 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mycpu->gd_cpuid) | 1008 PWAKEUP_ONE); 1009 } 1010 1011 /* 1012 * Wakeup threads tsleep()ing on the specified ident on the current cpu 1013 * only. 1014 */ 1015 void 1016 wakeup_mycpu(const volatile void *ident) 1017 { 1018 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mycpu->gd_cpuid) | 1019 PWAKEUP_MYCPU); 1020 } 1021 1022 /* 1023 * Wakeup one thread tsleep()ing on the specified ident on the current cpu 1024 * only. 1025 */ 1026 void 1027 wakeup_mycpu_one(const volatile void *ident) 1028 { 1029 /* XXX potentially round-robin the first responding cpu */ 1030 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mycpu->gd_cpuid) | 1031 PWAKEUP_MYCPU | PWAKEUP_ONE); 1032 } 1033 1034 /* 1035 * Wakeup all thread tsleep()ing on the specified ident on the specified cpu 1036 * only. 1037 */ 1038 void 1039 wakeup_oncpu(globaldata_t gd, const volatile void *ident) 1040 { 1041 globaldata_t mygd = mycpu; 1042 if (gd == mycpu) { 1043 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mygd->gd_cpuid) | 1044 PWAKEUP_MYCPU); 1045 } else { 1046 lwkt_send_ipiq2(gd, _wakeup, __DEALL(ident), 1047 PWAKEUP_ENCODE(0, mygd->gd_cpuid) | 1048 PWAKEUP_MYCPU); 1049 } 1050 } 1051 1052 /* 1053 * Wakeup one thread tsleep()ing on the specified ident on the specified cpu 1054 * only. 1055 */ 1056 void 1057 wakeup_oncpu_one(globaldata_t gd, const volatile void *ident) 1058 { 1059 globaldata_t mygd = mycpu; 1060 if (gd == mygd) { 1061 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mygd->gd_cpuid) | 1062 PWAKEUP_MYCPU | PWAKEUP_ONE); 1063 } else { 1064 lwkt_send_ipiq2(gd, _wakeup, __DEALL(ident), 1065 PWAKEUP_ENCODE(0, mygd->gd_cpuid) | 1066 PWAKEUP_MYCPU | PWAKEUP_ONE); 1067 } 1068 } 1069 1070 /* 1071 * Wakeup all threads waiting on the specified ident that slept using 1072 * the specified domain, on all cpus. 1073 */ 1074 void 1075 wakeup_domain(const volatile void *ident, int domain) 1076 { 1077 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(domain, mycpu->gd_cpuid)); 1078 } 1079 1080 /* 1081 * Wakeup one thread waiting on the specified ident that slept using 1082 * the specified domain, on any cpu. 1083 */ 1084 void 1085 wakeup_domain_one(const volatile void *ident, int domain) 1086 { 1087 /* XXX potentially round-robin the first responding cpu */ 1088 _wakeup(__DEALL(ident), 1089 PWAKEUP_ENCODE(domain, mycpu->gd_cpuid) | PWAKEUP_ONE); 1090 } 1091 1092 void 1093 wakeup_start_delayed(void) 1094 { 1095 globaldata_t gd = mycpu; 1096 1097 crit_enter(); 1098 gd->gd_curthread->td_flags |= TDF_DELAYED_WAKEUP; 1099 crit_exit(); 1100 } 1101 1102 void 1103 wakeup_end_delayed(void) 1104 { 1105 globaldata_t gd = mycpu; 1106 1107 if (gd->gd_curthread->td_flags & TDF_DELAYED_WAKEUP) { 1108 crit_enter(); 1109 gd->gd_curthread->td_flags &= ~TDF_DELAYED_WAKEUP; 1110 if (gd->gd_delayed_wakeup[0] || gd->gd_delayed_wakeup[1]) { 1111 if (gd->gd_delayed_wakeup[0]) { 1112 wakeup(gd->gd_delayed_wakeup[0]); 1113 gd->gd_delayed_wakeup[0] = NULL; 1114 } 1115 if (gd->gd_delayed_wakeup[1]) { 1116 wakeup(gd->gd_delayed_wakeup[1]); 1117 gd->gd_delayed_wakeup[1] = NULL; 1118 } 1119 } 1120 crit_exit(); 1121 } 1122 } 1123 1124 /* 1125 * setrunnable() 1126 * 1127 * Make a process runnable. lp->lwp_token must be held on call and this 1128 * function must be called from the cpu owning lp. 1129 * 1130 * This only has an effect if we are in LSSTOP or LSSLEEP. 1131 */ 1132 void 1133 setrunnable(struct lwp *lp) 1134 { 1135 thread_t td = lp->lwp_thread; 1136 1137 ASSERT_LWKT_TOKEN_HELD(&lp->lwp_token); 1138 KKASSERT(td->td_gd == mycpu); 1139 crit_enter(); 1140 if (lp->lwp_stat == LSSTOP) 1141 lp->lwp_stat = LSSLEEP; 1142 if (lp->lwp_stat == LSSLEEP) { 1143 _tsleep_remove(td); 1144 lwkt_schedule(td); 1145 } else if (td->td_flags & TDF_SINTR) { 1146 lwkt_schedule(td); 1147 } 1148 crit_exit(); 1149 } 1150 1151 /* 1152 * The process is stopped due to some condition, usually because p_stat is 1153 * set to SSTOP, but also possibly due to being traced. 1154 * 1155 * Caller must hold p->p_token 1156 * 1157 * NOTE! If the caller sets SSTOP, the caller must also clear P_WAITED 1158 * because the parent may check the child's status before the child actually 1159 * gets to this routine. 1160 * 1161 * This routine is called with the current lwp only, typically just 1162 * before returning to userland if the process state is detected as 1163 * possibly being in a stopped state. 1164 */ 1165 void 1166 tstop(void) 1167 { 1168 struct lwp *lp = curthread->td_lwp; 1169 struct proc *p = lp->lwp_proc; 1170 struct proc *q; 1171 1172 lwkt_gettoken(&lp->lwp_token); 1173 crit_enter(); 1174 1175 /* 1176 * If LWP_MP_WSTOP is set, we were sleeping 1177 * while our process was stopped. At this point 1178 * we were already counted as stopped. 1179 */ 1180 if ((lp->lwp_mpflags & LWP_MP_WSTOP) == 0) { 1181 /* 1182 * If we're the last thread to stop, signal 1183 * our parent. 1184 */ 1185 p->p_nstopped++; 1186 atomic_set_int(&lp->lwp_mpflags, LWP_MP_WSTOP); 1187 wakeup(&p->p_nstopped); 1188 if (p->p_nstopped == p->p_nthreads) { 1189 /* 1190 * Token required to interlock kern_wait() 1191 */ 1192 q = p->p_pptr; 1193 PHOLD(q); 1194 lwkt_gettoken(&q->p_token); 1195 p->p_flags &= ~P_WAITED; 1196 wakeup(p->p_pptr); 1197 if ((q->p_sigacts->ps_flag & PS_NOCLDSTOP) == 0) 1198 ksignal(q, SIGCHLD); 1199 lwkt_reltoken(&q->p_token); 1200 PRELE(q); 1201 } 1202 } 1203 1204 /* 1205 * Wait here while in a stopped state, interlocked with lwp_token. 1206 * We must break-out if the whole process is trying to exit. 1207 */ 1208 while (STOPLWP(p, lp)) { 1209 lp->lwp_stat = LSSTOP; 1210 tsleep(p, 0, "stop", 0); 1211 } 1212 p->p_nstopped--; 1213 atomic_clear_int(&lp->lwp_mpflags, LWP_MP_WSTOP); 1214 crit_exit(); 1215 lwkt_reltoken(&lp->lwp_token); 1216 } 1217 1218 /* 1219 * Compute a tenex style load average of a quantity on 1220 * 1, 5 and 15 minute intervals. 1221 */ 1222 static int loadav_count_runnable(struct lwp *p, void *data); 1223 1224 static void 1225 loadav(void *arg) 1226 { 1227 struct loadavg *avg; 1228 int i, nrun; 1229 1230 nrun = 0; 1231 alllwp_scan(loadav_count_runnable, &nrun); 1232 avg = &averunnable; 1233 for (i = 0; i < 3; i++) { 1234 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] + 1235 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT; 1236 } 1237 1238 /* 1239 * Schedule the next update to occur after 5 seconds, but add a 1240 * random variation to avoid synchronisation with processes that 1241 * run at regular intervals. 1242 */ 1243 callout_reset(&loadav_callout, hz * 4 + (int)(krandom() % (hz * 2 + 1)), 1244 loadav, NULL); 1245 } 1246 1247 static int 1248 loadav_count_runnable(struct lwp *lp, void *data) 1249 { 1250 int *nrunp = data; 1251 thread_t td; 1252 1253 switch (lp->lwp_stat) { 1254 case LSRUN: 1255 if ((td = lp->lwp_thread) == NULL) 1256 break; 1257 if (td->td_flags & TDF_BLOCKED) 1258 break; 1259 ++*nrunp; 1260 break; 1261 default: 1262 break; 1263 } 1264 lwkt_yield(); 1265 return(0); 1266 } 1267 1268 /* ARGSUSED */ 1269 static void 1270 sched_setup(void *dummy) 1271 { 1272 callout_init_mp(&loadav_callout); 1273 callout_init_mp(&schedcpu_callout); 1274 1275 /* Kick off timeout driven events by calling first time. */ 1276 schedcpu(NULL); 1277 loadav(NULL); 1278 } 1279 1280