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 lwkt_deschedule_self(td); 658 td->td_wmesg = "tsrace"; 659 lwkt_switch(); 660 kprintf("td %p %s: timeout race\n", td, td->td_comm); 661 } 662 if (td->td_flags & TDF_TIMEOUT) { 663 td->td_flags &= ~TDF_TIMEOUT; 664 error = EWOULDBLOCK; 665 } else { 666 /* does not block when on same cpu */ 667 callout_stop(&thandle); 668 } 669 } 670 td->td_flags &= ~TDF_TSLEEP_DESCHEDULED; 671 672 /* 673 * Make sure we have been removed from the sleepq. In most 674 * cases this will have been done for us already but it is 675 * possible for a scheduling IPI to be in-flight from a 676 * previous tsleep/tsleep_interlock() or due to a straight-out 677 * call to lwkt_schedule() (in the case of an interrupt thread), 678 * causing a spurious wakeup. 679 */ 680 _tsleep_remove(td); 681 td->td_wmesg = NULL; 682 683 /* 684 * Figure out the correct error return. If interrupted by a 685 * signal we want to return EINTR or ERESTART. 686 */ 687 resume: 688 if (lp) { 689 if (catch && error == 0) { 690 if (sig != 0 || (sig = CURSIG(lp))) { 691 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 692 error = EINTR; 693 else 694 error = ERESTART; 695 } 696 } 697 lp->lwp_flags &= ~LWP_SINTR; 698 lwkt_reltoken(&lp->lwp_token); 699 } 700 logtsleep1(tsleep_end); 701 crit_exit_quick(td); 702 return (error); 703 } 704 705 /* 706 * Interlocked spinlock sleep. An exclusively held spinlock must 707 * be passed to ssleep(). The function will atomically release the 708 * spinlock and tsleep on the ident, then reacquire the spinlock and 709 * return. 710 * 711 * This routine is fairly important along the critical path, so optimize it 712 * heavily. 713 */ 714 int 715 ssleep(const volatile void *ident, struct spinlock *spin, int flags, 716 const char *wmesg, int timo) 717 { 718 globaldata_t gd = mycpu; 719 int error; 720 721 _tsleep_interlock(gd, ident, flags); 722 spin_unlock_quick(gd, spin); 723 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 724 _spin_lock_quick(gd, spin, wmesg); 725 726 return (error); 727 } 728 729 int 730 lksleep(const volatile void *ident, struct lock *lock, int flags, 731 const char *wmesg, int timo) 732 { 733 globaldata_t gd = mycpu; 734 int error; 735 736 _tsleep_interlock(gd, ident, flags); 737 lockmgr(lock, LK_RELEASE); 738 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 739 lockmgr(lock, LK_EXCLUSIVE); 740 741 return (error); 742 } 743 744 /* 745 * Interlocked mutex sleep. An exclusively held mutex must be passed 746 * to mtxsleep(). The function will atomically release the mutex 747 * and tsleep on the ident, then reacquire the mutex and return. 748 */ 749 int 750 mtxsleep(const volatile void *ident, struct mtx *mtx, int flags, 751 const char *wmesg, int timo) 752 { 753 globaldata_t gd = mycpu; 754 int error; 755 756 _tsleep_interlock(gd, ident, flags); 757 mtx_unlock(mtx); 758 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 759 mtx_lock_ex_quick(mtx); 760 761 return (error); 762 } 763 764 /* 765 * Interlocked serializer sleep. An exclusively held serializer must 766 * be passed to zsleep(). The function will atomically release 767 * the serializer and tsleep on the ident, then reacquire the serializer 768 * and return. 769 */ 770 int 771 zsleep(const volatile void *ident, struct lwkt_serialize *slz, int flags, 772 const char *wmesg, int timo) 773 { 774 globaldata_t gd = mycpu; 775 int ret; 776 777 ASSERT_SERIALIZED(slz); 778 779 _tsleep_interlock(gd, ident, flags); 780 lwkt_serialize_exit(slz); 781 ret = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 782 lwkt_serialize_enter(slz); 783 784 return ret; 785 } 786 787 /* 788 * Directly block on the LWKT thread by descheduling it. This 789 * is much faster then tsleep(), but the only legal way to wake 790 * us up is to directly schedule the thread. 791 * 792 * Setting TDF_SINTR will cause new signals to directly schedule us. 793 * 794 * This routine must be called while in a critical section. 795 */ 796 int 797 lwkt_sleep(const char *wmesg, int flags) 798 { 799 thread_t td = curthread; 800 int sig; 801 802 if ((flags & PCATCH) == 0 || td->td_lwp == NULL) { 803 td->td_flags |= TDF_BLOCKED; 804 td->td_wmesg = wmesg; 805 lwkt_deschedule_self(td); 806 lwkt_switch(); 807 td->td_wmesg = NULL; 808 td->td_flags &= ~TDF_BLOCKED; 809 return(0); 810 } 811 if ((sig = CURSIG(td->td_lwp)) != 0) { 812 if (SIGISMEMBER(td->td_proc->p_sigacts->ps_sigintr, sig)) 813 return(EINTR); 814 else 815 return(ERESTART); 816 817 } 818 td->td_flags |= TDF_BLOCKED | TDF_SINTR; 819 td->td_wmesg = wmesg; 820 lwkt_deschedule_self(td); 821 lwkt_switch(); 822 td->td_flags &= ~(TDF_BLOCKED | TDF_SINTR); 823 td->td_wmesg = NULL; 824 return(0); 825 } 826 827 /* 828 * Implement the timeout for tsleep. 829 * 830 * This type of callout timeout is scheduled on the same cpu the process 831 * is sleeping on. Also, at the moment, the MP lock is held. 832 */ 833 static void 834 endtsleep(void *arg) 835 { 836 thread_t td = arg; 837 struct lwp *lp; 838 839 /* 840 * We are going to have to get the lwp_token, which means we might 841 * block. This can race a tsleep getting woken up by other means 842 * so set TDF_TIMEOUT_RUNNING to force the tsleep to wait for our 843 * processing to complete (sorry tsleep!). 844 * 845 * We can safely set td_flags because td MUST be on the same cpu 846 * as we are. 847 */ 848 KKASSERT(td->td_gd == mycpu); 849 crit_enter(); 850 td->td_flags |= TDF_TIMEOUT_RUNNING | TDF_TIMEOUT; 851 852 /* 853 * This can block but TDF_TIMEOUT_RUNNING will prevent the thread 854 * from exiting the tsleep on us. The flag is interlocked by virtue 855 * of lp being on the same cpu as we are. 856 */ 857 if ((lp = td->td_lwp) != NULL) 858 lwkt_gettoken(&lp->lwp_token); 859 860 KKASSERT(td->td_flags & TDF_TSLEEP_DESCHEDULED); 861 862 if (lp) { 863 /* 864 * callout timer should never be set in tstop() because 865 * it passes a timeout of 0. 866 */ 867 KKASSERT(lp->lwp_stat != LSSTOP); 868 setrunnable(lp); 869 lwkt_reltoken(&lp->lwp_token); 870 } else { 871 _tsleep_remove(td); 872 lwkt_schedule(td); 873 } 874 KKASSERT(td->td_gd == mycpu); 875 td->td_flags &= ~TDF_TIMEOUT_RUNNING; 876 crit_exit(); 877 } 878 879 /* 880 * Make all processes sleeping on the specified identifier runnable. 881 * count may be zero or one only. 882 * 883 * The domain encodes the sleep/wakeup domain, flags, plus the originating 884 * cpu. 885 * 886 * This call may run without the MP lock held. We can only manipulate thread 887 * state on the cpu owning the thread. We CANNOT manipulate process state 888 * at all. 889 * 890 * _wakeup() can be passed to an IPI so we can't use (const volatile 891 * void *ident). 892 */ 893 static void 894 _wakeup(void *ident, int domain) 895 { 896 struct tslpque *qp; 897 struct thread *td; 898 struct thread *ntd; 899 globaldata_t gd; 900 cpumask_t mask; 901 int id; 902 903 crit_enter(); 904 logtsleep2(wakeup_beg, ident); 905 gd = mycpu; 906 id = LOOKUP(ident); 907 qp = &gd->gd_tsleep_hash[id]; 908 restart: 909 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) { 910 ntd = TAILQ_NEXT(td, td_sleepq); 911 if (td->td_wchan == ident && 912 td->td_wdomain == (domain & PDOMAIN_MASK) 913 ) { 914 KKASSERT(td->td_gd == gd); 915 _tsleep_remove(td); 916 td->td_wakefromcpu = PWAKEUP_DECODE(domain); 917 if (td->td_flags & TDF_TSLEEP_DESCHEDULED) { 918 lwkt_schedule(td); 919 if (domain & PWAKEUP_ONE) 920 goto done; 921 } 922 goto restart; 923 } 924 } 925 926 /* 927 * We finished checking the current cpu but there still may be 928 * more work to do. Either wakeup_one was requested and no matching 929 * thread was found, or a normal wakeup was requested and we have 930 * to continue checking cpus. 931 * 932 * It should be noted that this scheme is actually less expensive then 933 * the old scheme when waking up multiple threads, since we send 934 * only one IPI message per target candidate which may then schedule 935 * multiple threads. Before we could have wound up sending an IPI 936 * message for each thread on the target cpu (!= current cpu) that 937 * needed to be woken up. 938 * 939 * NOTE: Wakeups occuring on remote cpus are asynchronous. This 940 * should be ok since we are passing idents in the IPI rather then 941 * thread pointers. 942 */ 943 if ((domain & PWAKEUP_MYCPU) == 0) { 944 mask = slpque_cpumasks[id]; 945 CPUMASK_ANDMASK(mask, gd->gd_other_cpus); 946 if (CPUMASK_TESTNZERO(mask)) { 947 lwkt_send_ipiq2_mask(mask, _wakeup, ident, 948 domain | PWAKEUP_MYCPU); 949 } 950 } 951 done: 952 logtsleep1(wakeup_end); 953 crit_exit(); 954 } 955 956 /* 957 * Wakeup all threads tsleep()ing on the specified ident, on all cpus 958 */ 959 void 960 wakeup(const volatile void *ident) 961 { 962 globaldata_t gd = mycpu; 963 thread_t td = gd->gd_curthread; 964 965 if (td && (td->td_flags & TDF_DELAYED_WAKEUP)) { 966 /* 967 * If we are in a delayed wakeup section, record up to two wakeups in 968 * a per-CPU queue and issue them when we block or exit the delayed 969 * wakeup section. 970 */ 971 if (atomic_cmpset_ptr(&gd->gd_delayed_wakeup[0], NULL, ident)) 972 return; 973 if (atomic_cmpset_ptr(&gd->gd_delayed_wakeup[1], NULL, ident)) 974 return; 975 976 ident = atomic_swap_ptr(__DEQUALIFY(volatile void **, &gd->gd_delayed_wakeup[1]), 977 __DEALL(ident)); 978 ident = atomic_swap_ptr(__DEQUALIFY(volatile void **, &gd->gd_delayed_wakeup[0]), 979 __DEALL(ident)); 980 } 981 982 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, gd->gd_cpuid)); 983 } 984 985 /* 986 * Wakeup one thread tsleep()ing on the specified ident, on any cpu. 987 */ 988 void 989 wakeup_one(const volatile void *ident) 990 { 991 /* XXX potentially round-robin the first responding cpu */ 992 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mycpu->gd_cpuid) | 993 PWAKEUP_ONE); 994 } 995 996 /* 997 * Wakeup threads tsleep()ing on the specified ident on the current cpu 998 * only. 999 */ 1000 void 1001 wakeup_mycpu(const volatile void *ident) 1002 { 1003 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mycpu->gd_cpuid) | 1004 PWAKEUP_MYCPU); 1005 } 1006 1007 /* 1008 * Wakeup one thread tsleep()ing on the specified ident on the current cpu 1009 * only. 1010 */ 1011 void 1012 wakeup_mycpu_one(const volatile void *ident) 1013 { 1014 /* XXX potentially round-robin the first responding cpu */ 1015 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mycpu->gd_cpuid) | 1016 PWAKEUP_MYCPU | PWAKEUP_ONE); 1017 } 1018 1019 /* 1020 * Wakeup all thread tsleep()ing on the specified ident on the specified cpu 1021 * only. 1022 */ 1023 void 1024 wakeup_oncpu(globaldata_t gd, const volatile void *ident) 1025 { 1026 globaldata_t mygd = mycpu; 1027 if (gd == mycpu) { 1028 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mygd->gd_cpuid) | 1029 PWAKEUP_MYCPU); 1030 } else { 1031 lwkt_send_ipiq2(gd, _wakeup, __DEALL(ident), 1032 PWAKEUP_ENCODE(0, mygd->gd_cpuid) | 1033 PWAKEUP_MYCPU); 1034 } 1035 } 1036 1037 /* 1038 * Wakeup one thread tsleep()ing on the specified ident on the specified cpu 1039 * only. 1040 */ 1041 void 1042 wakeup_oncpu_one(globaldata_t gd, const volatile void *ident) 1043 { 1044 globaldata_t mygd = mycpu; 1045 if (gd == mygd) { 1046 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mygd->gd_cpuid) | 1047 PWAKEUP_MYCPU | PWAKEUP_ONE); 1048 } else { 1049 lwkt_send_ipiq2(gd, _wakeup, __DEALL(ident), 1050 PWAKEUP_ENCODE(0, mygd->gd_cpuid) | 1051 PWAKEUP_MYCPU | PWAKEUP_ONE); 1052 } 1053 } 1054 1055 /* 1056 * Wakeup all threads waiting on the specified ident that slept using 1057 * the specified domain, on all cpus. 1058 */ 1059 void 1060 wakeup_domain(const volatile void *ident, int domain) 1061 { 1062 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(domain, mycpu->gd_cpuid)); 1063 } 1064 1065 /* 1066 * Wakeup one thread waiting on the specified ident that slept using 1067 * the specified domain, on any cpu. 1068 */ 1069 void 1070 wakeup_domain_one(const volatile void *ident, int domain) 1071 { 1072 /* XXX potentially round-robin the first responding cpu */ 1073 _wakeup(__DEALL(ident), 1074 PWAKEUP_ENCODE(domain, mycpu->gd_cpuid) | PWAKEUP_ONE); 1075 } 1076 1077 void 1078 wakeup_start_delayed(void) 1079 { 1080 globaldata_t gd = mycpu; 1081 1082 crit_enter(); 1083 gd->gd_curthread->td_flags |= TDF_DELAYED_WAKEUP; 1084 crit_exit(); 1085 } 1086 1087 void 1088 wakeup_end_delayed(void) 1089 { 1090 globaldata_t gd = mycpu; 1091 1092 if (gd->gd_curthread->td_flags & TDF_DELAYED_WAKEUP) { 1093 crit_enter(); 1094 gd->gd_curthread->td_flags &= ~TDF_DELAYED_WAKEUP; 1095 if (gd->gd_delayed_wakeup[0] || gd->gd_delayed_wakeup[1]) { 1096 if (gd->gd_delayed_wakeup[0]) { 1097 wakeup(gd->gd_delayed_wakeup[0]); 1098 gd->gd_delayed_wakeup[0] = NULL; 1099 } 1100 if (gd->gd_delayed_wakeup[1]) { 1101 wakeup(gd->gd_delayed_wakeup[1]); 1102 gd->gd_delayed_wakeup[1] = NULL; 1103 } 1104 } 1105 crit_exit(); 1106 } 1107 } 1108 1109 /* 1110 * setrunnable() 1111 * 1112 * Make a process runnable. lp->lwp_token must be held on call and this 1113 * function must be called from the cpu owning lp. 1114 * 1115 * This only has an effect if we are in LSSTOP or LSSLEEP. 1116 */ 1117 void 1118 setrunnable(struct lwp *lp) 1119 { 1120 thread_t td = lp->lwp_thread; 1121 1122 ASSERT_LWKT_TOKEN_HELD(&lp->lwp_token); 1123 KKASSERT(td->td_gd == mycpu); 1124 crit_enter(); 1125 if (lp->lwp_stat == LSSTOP) 1126 lp->lwp_stat = LSSLEEP; 1127 if (lp->lwp_stat == LSSLEEP) { 1128 _tsleep_remove(td); 1129 lwkt_schedule(td); 1130 } else if (td->td_flags & TDF_SINTR) { 1131 lwkt_schedule(td); 1132 } 1133 crit_exit(); 1134 } 1135 1136 /* 1137 * The process is stopped due to some condition, usually because p_stat is 1138 * set to SSTOP, but also possibly due to being traced. 1139 * 1140 * Caller must hold p->p_token 1141 * 1142 * NOTE! If the caller sets SSTOP, the caller must also clear P_WAITED 1143 * because the parent may check the child's status before the child actually 1144 * gets to this routine. 1145 * 1146 * This routine is called with the current lwp only, typically just 1147 * before returning to userland if the process state is detected as 1148 * possibly being in a stopped state. 1149 */ 1150 void 1151 tstop(void) 1152 { 1153 struct lwp *lp = curthread->td_lwp; 1154 struct proc *p = lp->lwp_proc; 1155 struct proc *q; 1156 1157 lwkt_gettoken(&lp->lwp_token); 1158 crit_enter(); 1159 1160 /* 1161 * If LWP_MP_WSTOP is set, we were sleeping 1162 * while our process was stopped. At this point 1163 * we were already counted as stopped. 1164 */ 1165 if ((lp->lwp_mpflags & LWP_MP_WSTOP) == 0) { 1166 /* 1167 * If we're the last thread to stop, signal 1168 * our parent. 1169 */ 1170 p->p_nstopped++; 1171 atomic_set_int(&lp->lwp_mpflags, LWP_MP_WSTOP); 1172 wakeup(&p->p_nstopped); 1173 if (p->p_nstopped == p->p_nthreads) { 1174 /* 1175 * Token required to interlock kern_wait() 1176 */ 1177 q = p->p_pptr; 1178 PHOLD(q); 1179 lwkt_gettoken(&q->p_token); 1180 p->p_flags &= ~P_WAITED; 1181 wakeup(p->p_pptr); 1182 if ((q->p_sigacts->ps_flag & PS_NOCLDSTOP) == 0) 1183 ksignal(q, SIGCHLD); 1184 lwkt_reltoken(&q->p_token); 1185 PRELE(q); 1186 } 1187 } 1188 while (p->p_stat == SSTOP || p->p_stat == SCORE) { 1189 lp->lwp_stat = LSSTOP; 1190 tsleep(p, 0, "stop", 0); 1191 } 1192 p->p_nstopped--; 1193 atomic_clear_int(&lp->lwp_mpflags, LWP_MP_WSTOP); 1194 crit_exit(); 1195 lwkt_reltoken(&lp->lwp_token); 1196 } 1197 1198 /* 1199 * Compute a tenex style load average of a quantity on 1200 * 1, 5 and 15 minute intervals. 1201 */ 1202 static int loadav_count_runnable(struct lwp *p, void *data); 1203 1204 static void 1205 loadav(void *arg) 1206 { 1207 struct loadavg *avg; 1208 int i, nrun; 1209 1210 nrun = 0; 1211 alllwp_scan(loadav_count_runnable, &nrun); 1212 avg = &averunnable; 1213 for (i = 0; i < 3; i++) { 1214 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] + 1215 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT; 1216 } 1217 1218 /* 1219 * Schedule the next update to occur after 5 seconds, but add a 1220 * random variation to avoid synchronisation with processes that 1221 * run at regular intervals. 1222 */ 1223 callout_reset(&loadav_callout, hz * 4 + (int)(krandom() % (hz * 2 + 1)), 1224 loadav, NULL); 1225 } 1226 1227 static int 1228 loadav_count_runnable(struct lwp *lp, void *data) 1229 { 1230 int *nrunp = data; 1231 thread_t td; 1232 1233 switch (lp->lwp_stat) { 1234 case LSRUN: 1235 if ((td = lp->lwp_thread) == NULL) 1236 break; 1237 if (td->td_flags & TDF_BLOCKED) 1238 break; 1239 ++*nrunp; 1240 break; 1241 default: 1242 break; 1243 } 1244 lwkt_yield(); 1245 return(0); 1246 } 1247 1248 /* ARGSUSED */ 1249 static void 1250 sched_setup(void *dummy) 1251 { 1252 callout_init_mp(&loadav_callout); 1253 callout_init_mp(&schedcpu_callout); 1254 1255 /* Kick off timeout driven events by calling first time. */ 1256 schedcpu(NULL); 1257 loadav(NULL); 1258 } 1259 1260