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 */ 41 42 #include "opt_ktrace.h" 43 44 #include <sys/param.h> 45 #include <sys/systm.h> 46 #include <sys/proc.h> 47 #include <sys/kernel.h> 48 #include <sys/signalvar.h> 49 #include <sys/resourcevar.h> 50 #include <sys/vmmeter.h> 51 #include <sys/sysctl.h> 52 #include <sys/lock.h> 53 #include <sys/uio.h> 54 #ifdef KTRACE 55 #include <sys/ktrace.h> 56 #endif 57 #include <sys/xwait.h> 58 #include <sys/ktr.h> 59 #include <sys/serialize.h> 60 61 #include <sys/signal2.h> 62 #include <sys/thread2.h> 63 #include <sys/spinlock2.h> 64 #include <sys/mutex2.h> 65 66 #include <machine/cpu.h> 67 #include <machine/smp.h> 68 69 TAILQ_HEAD(tslpque, thread); 70 71 static void sched_setup (void *dummy); 72 SYSINIT(sched_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, sched_setup, NULL) 73 74 int hogticks; 75 int lbolt; 76 void *lbolt_syncer; 77 int sched_quantum; /* Roundrobin scheduling quantum in ticks. */ 78 int ncpus; 79 int ncpus2, ncpus2_shift, ncpus2_mask; /* note: mask not cpumask_t */ 80 int ncpus_fit, ncpus_fit_mask; /* note: mask not cpumask_t */ 81 int safepri; 82 int tsleep_now_works; 83 int tsleep_crypto_dump = 0; 84 85 static struct callout loadav_callout; 86 static struct callout schedcpu_callout; 87 MALLOC_DEFINE(M_TSLEEP, "tslpque", "tsleep queues"); 88 89 #define __DEALL(ident) __DEQUALIFY(void *, ident) 90 91 #if !defined(KTR_TSLEEP) 92 #define KTR_TSLEEP KTR_ALL 93 #endif 94 KTR_INFO_MASTER(tsleep); 95 KTR_INFO(KTR_TSLEEP, tsleep, tsleep_beg, 0, "tsleep enter %p", const volatile void *ident); 96 KTR_INFO(KTR_TSLEEP, tsleep, tsleep_end, 1, "tsleep exit"); 97 KTR_INFO(KTR_TSLEEP, tsleep, wakeup_beg, 2, "wakeup enter %p", const volatile void *ident); 98 KTR_INFO(KTR_TSLEEP, tsleep, wakeup_end, 3, "wakeup exit"); 99 KTR_INFO(KTR_TSLEEP, tsleep, ilockfail, 4, "interlock failed %p", const volatile void *ident); 100 101 #define logtsleep1(name) KTR_LOG(tsleep_ ## name) 102 #define logtsleep2(name, val) KTR_LOG(tsleep_ ## name, val) 103 104 struct loadavg averunnable = 105 { {0, 0, 0}, FSCALE }; /* load average, of runnable procs */ 106 /* 107 * Constants for averages over 1, 5, and 15 minutes 108 * when sampling at 5 second intervals. 109 */ 110 static fixpt_t cexp[3] = { 111 0.9200444146293232 * FSCALE, /* exp(-1/12) */ 112 0.9834714538216174 * FSCALE, /* exp(-1/60) */ 113 0.9944598480048967 * FSCALE, /* exp(-1/180) */ 114 }; 115 116 static void endtsleep (void *); 117 static void loadav (void *arg); 118 static void schedcpu (void *arg); 119 120 /* 121 * Adjust the scheduler quantum. The quantum is specified in microseconds. 122 * Note that 'tick' is in microseconds per tick. 123 */ 124 static int 125 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS) 126 { 127 int error, new_val; 128 129 new_val = sched_quantum * ustick; 130 error = sysctl_handle_int(oidp, &new_val, 0, req); 131 if (error != 0 || req->newptr == NULL) 132 return (error); 133 if (new_val < ustick) 134 return (EINVAL); 135 sched_quantum = new_val / ustick; 136 hogticks = 2 * sched_quantum; 137 return (0); 138 } 139 140 SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW, 141 0, sizeof sched_quantum, sysctl_kern_quantum, "I", ""); 142 143 static int pctcpu_decay = 10; 144 SYSCTL_INT(_kern, OID_AUTO, pctcpu_decay, CTLFLAG_RW, &pctcpu_decay, 0, ""); 145 146 /* 147 * kernel uses `FSCALE', userland (SHOULD) use kern.fscale 148 */ 149 int fscale __unused = FSCALE; /* exported to systat */ 150 SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, ""); 151 152 /* 153 * Recompute process priorities, once a second. 154 * 155 * Since the userland schedulers are typically event oriented, if the 156 * estcpu calculation at wakeup() time is not sufficient to make a 157 * process runnable relative to other processes in the system we have 158 * a 1-second recalc to help out. 159 * 160 * This code also allows us to store sysclock_t data in the process structure 161 * without fear of an overrun, since sysclock_t are guarenteed to hold 162 * several seconds worth of count. 163 * 164 * WARNING! callouts can preempt normal threads. However, they will not 165 * preempt a thread holding a spinlock so we *can* safely use spinlocks. 166 */ 167 static int schedcpu_stats(struct proc *p, void *data __unused); 168 static int schedcpu_resource(struct proc *p, void *data __unused); 169 170 static void 171 schedcpu(void *arg) 172 { 173 allproc_scan(schedcpu_stats, NULL); 174 allproc_scan(schedcpu_resource, NULL); 175 wakeup((caddr_t)&lbolt); 176 wakeup(lbolt_syncer); 177 callout_reset(&schedcpu_callout, hz, schedcpu, NULL); 178 } 179 180 /* 181 * General process statistics once a second 182 */ 183 static int 184 schedcpu_stats(struct proc *p, void *data __unused) 185 { 186 struct lwp *lp; 187 188 /* 189 * Threads may not be completely set up if process in SIDL state. 190 */ 191 if (p->p_stat == SIDL) 192 return(0); 193 194 PHOLD(p); 195 if (lwkt_trytoken(&p->p_token) == FALSE) { 196 PRELE(p); 197 return(0); 198 } 199 200 p->p_swtime++; 201 FOREACH_LWP_IN_PROC(lp, p) { 202 if (lp->lwp_stat == LSSLEEP) { 203 ++lp->lwp_slptime; 204 if (lp->lwp_slptime == 1) 205 p->p_usched->uload_update(lp); 206 } 207 208 /* 209 * Only recalculate processes that are active or have slept 210 * less then 2 seconds. The schedulers understand this. 211 * Otherwise decay by 50% per second. 212 */ 213 if (lp->lwp_slptime <= 1) { 214 p->p_usched->recalculate(lp); 215 } else { 216 int decay; 217 218 decay = pctcpu_decay; 219 cpu_ccfence(); 220 if (decay <= 1) 221 decay = 1; 222 if (decay > 100) 223 decay = 100; 224 lp->lwp_pctcpu = (lp->lwp_pctcpu * (decay - 1)) / decay; 225 } 226 } 227 lwkt_reltoken(&p->p_token); 228 lwkt_yield(); 229 PRELE(p); 230 return(0); 231 } 232 233 /* 234 * Resource checks. XXX break out since ksignal/killproc can block, 235 * limiting us to one process killed per second. There is probably 236 * a better way. 237 */ 238 static int 239 schedcpu_resource(struct proc *p, void *data __unused) 240 { 241 u_int64_t ttime; 242 struct lwp *lp; 243 244 if (p->p_stat == SIDL) 245 return(0); 246 247 PHOLD(p); 248 if (lwkt_trytoken(&p->p_token) == FALSE) { 249 PRELE(p); 250 return(0); 251 } 252 253 if (p->p_stat == SZOMB || p->p_limit == NULL) { 254 lwkt_reltoken(&p->p_token); 255 PRELE(p); 256 return(0); 257 } 258 259 ttime = 0; 260 FOREACH_LWP_IN_PROC(lp, p) { 261 /* 262 * We may have caught an lp in the middle of being 263 * created, lwp_thread can be NULL. 264 */ 265 if (lp->lwp_thread) { 266 ttime += lp->lwp_thread->td_sticks; 267 ttime += lp->lwp_thread->td_uticks; 268 } 269 } 270 271 switch(plimit_testcpulimit(p->p_limit, ttime)) { 272 case PLIMIT_TESTCPU_KILL: 273 killproc(p, "exceeded maximum CPU limit"); 274 break; 275 case PLIMIT_TESTCPU_XCPU: 276 if ((p->p_flags & P_XCPU) == 0) { 277 p->p_flags |= P_XCPU; 278 ksignal(p, SIGXCPU); 279 } 280 break; 281 default: 282 break; 283 } 284 lwkt_reltoken(&p->p_token); 285 lwkt_yield(); 286 PRELE(p); 287 return(0); 288 } 289 290 /* 291 * This is only used by ps. Generate a cpu percentage use over 292 * a period of one second. 293 */ 294 void 295 updatepcpu(struct lwp *lp, int cpticks, int ttlticks) 296 { 297 fixpt_t acc; 298 int remticks; 299 300 acc = (cpticks << FSHIFT) / ttlticks; 301 if (ttlticks >= ESTCPUFREQ) { 302 lp->lwp_pctcpu = acc; 303 } else { 304 remticks = ESTCPUFREQ - ttlticks; 305 lp->lwp_pctcpu = (acc * ttlticks + lp->lwp_pctcpu * remticks) / 306 ESTCPUFREQ; 307 } 308 } 309 310 /* 311 * tsleep/wakeup hash table parameters. Try to find the sweet spot for 312 * like addresses being slept on. 313 */ 314 #define TABLESIZE 4001 315 #define LOOKUP(x) (((u_int)(uintptr_t)(x)) % TABLESIZE) 316 317 static cpumask_t slpque_cpumasks[TABLESIZE]; 318 319 /* 320 * General scheduler initialization. We force a reschedule 25 times 321 * a second by default. Note that cpu0 is initialized in early boot and 322 * cannot make any high level calls. 323 * 324 * Each cpu has its own sleep queue. 325 */ 326 void 327 sleep_gdinit(globaldata_t gd) 328 { 329 static struct tslpque slpque_cpu0[TABLESIZE]; 330 int i; 331 332 if (gd->gd_cpuid == 0) { 333 sched_quantum = (hz + 24) / 25; 334 hogticks = 2 * sched_quantum; 335 336 gd->gd_tsleep_hash = slpque_cpu0; 337 } else { 338 gd->gd_tsleep_hash = kmalloc(sizeof(slpque_cpu0), 339 M_TSLEEP, M_WAITOK | M_ZERO); 340 } 341 for (i = 0; i < TABLESIZE; ++i) 342 TAILQ_INIT(&gd->gd_tsleep_hash[i]); 343 } 344 345 /* 346 * This is a dandy function that allows us to interlock tsleep/wakeup 347 * operations with unspecified upper level locks, such as lockmgr locks, 348 * simply by holding a critical section. The sequence is: 349 * 350 * (acquire upper level lock) 351 * tsleep_interlock(blah) 352 * (release upper level lock) 353 * tsleep(blah, ...) 354 * 355 * Basically this functions queues us on the tsleep queue without actually 356 * descheduling us. When tsleep() is later called with PINTERLOCK it 357 * assumes the thread was already queued, otherwise it queues it there. 358 * 359 * Thus it is possible to receive the wakeup prior to going to sleep and 360 * the race conditions are covered. 361 */ 362 static __inline void 363 _tsleep_interlock(globaldata_t gd, const volatile void *ident, int flags) 364 { 365 thread_t td = gd->gd_curthread; 366 int id; 367 368 crit_enter_quick(td); 369 if (td->td_flags & TDF_TSLEEPQ) { 370 id = LOOKUP(td->td_wchan); 371 TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_sleepq); 372 if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL) { 373 atomic_clear_cpumask(&slpque_cpumasks[id], 374 gd->gd_cpumask); 375 } 376 } else { 377 td->td_flags |= TDF_TSLEEPQ; 378 } 379 id = LOOKUP(ident); 380 TAILQ_INSERT_TAIL(&gd->gd_tsleep_hash[id], td, td_sleepq); 381 atomic_set_cpumask(&slpque_cpumasks[id], gd->gd_cpumask); 382 td->td_wchan = ident; 383 td->td_wdomain = flags & PDOMAIN_MASK; 384 crit_exit_quick(td); 385 } 386 387 void 388 tsleep_interlock(const volatile void *ident, int flags) 389 { 390 _tsleep_interlock(mycpu, ident, flags); 391 } 392 393 /* 394 * Remove thread from sleepq. Must be called with a critical section held. 395 * The thread must not be migrating. 396 */ 397 static __inline void 398 _tsleep_remove(thread_t td) 399 { 400 globaldata_t gd = mycpu; 401 int id; 402 403 KKASSERT(td->td_gd == gd && IN_CRITICAL_SECT(td)); 404 KKASSERT((td->td_flags & TDF_MIGRATING) == 0); 405 if (td->td_flags & TDF_TSLEEPQ) { 406 td->td_flags &= ~TDF_TSLEEPQ; 407 id = LOOKUP(td->td_wchan); 408 TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_sleepq); 409 if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL) 410 atomic_clear_cpumask(&slpque_cpumasks[id], gd->gd_cpumask); 411 td->td_wchan = NULL; 412 td->td_wdomain = 0; 413 } 414 } 415 416 void 417 tsleep_remove(thread_t td) 418 { 419 _tsleep_remove(td); 420 } 421 422 /* 423 * General sleep call. Suspends the current process until a wakeup is 424 * performed on the specified identifier. The process will then be made 425 * runnable with the specified priority. Sleeps at most timo/hz seconds 426 * (0 means no timeout). If flags includes PCATCH flag, signals are checked 427 * before and after sleeping, else signals are not checked. Returns 0 if 428 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a 429 * signal needs to be delivered, ERESTART is returned if the current system 430 * call should be restarted if possible, and EINTR is returned if the system 431 * call should be interrupted by the signal (return EINTR). 432 * 433 * Note that if we are a process, we release_curproc() before messing with 434 * the LWKT scheduler. 435 * 436 * During autoconfiguration or after a panic, a sleep will simply 437 * lower the priority briefly to allow interrupts, then return. 438 * 439 * WARNING! This code can't block (short of switching away), or bad things 440 * will happen. No getting tokens, no blocking locks, etc. 441 */ 442 int 443 tsleep(const volatile void *ident, int flags, const char *wmesg, int timo) 444 { 445 struct thread *td = curthread; 446 struct lwp *lp = td->td_lwp; 447 struct proc *p = td->td_proc; /* may be NULL */ 448 globaldata_t gd; 449 int sig; 450 int catch; 451 int error; 452 int oldpri; 453 struct callout thandle; 454 455 /* 456 * Currently a severe hack. Make sure any delayed wakeups 457 * are flushed before we sleep or we might deadlock on whatever 458 * event we are sleeping on. 459 */ 460 if (td->td_flags & TDF_DELAYED_WAKEUP) 461 wakeup_end_delayed(); 462 463 /* 464 * NOTE: removed KTRPOINT, it could cause races due to blocking 465 * even in stable. Just scrap it for now. 466 */ 467 if (!tsleep_crypto_dump && (tsleep_now_works == 0 || panicstr)) { 468 /* 469 * After a panic, or before we actually have an operational 470 * softclock, just give interrupts a chance, then just return; 471 * 472 * don't run any other procs or panic below, 473 * in case this is the idle process and already asleep. 474 */ 475 splz(); 476 oldpri = td->td_pri; 477 lwkt_setpri_self(safepri); 478 lwkt_switch(); 479 lwkt_setpri_self(oldpri); 480 return (0); 481 } 482 logtsleep2(tsleep_beg, ident); 483 gd = td->td_gd; 484 KKASSERT(td != &gd->gd_idlethread); /* you must be kidding! */ 485 td->td_wakefromcpu = -1; /* overwritten by _wakeup */ 486 487 /* 488 * NOTE: all of this occurs on the current cpu, including any 489 * callout-based wakeups, so a critical section is a sufficient 490 * interlock. 491 * 492 * The entire sequence through to where we actually sleep must 493 * run without breaking the critical section. 494 */ 495 catch = flags & PCATCH; 496 error = 0; 497 sig = 0; 498 499 crit_enter_quick(td); 500 501 KASSERT(ident != NULL, ("tsleep: no ident")); 502 KASSERT(lp == NULL || 503 lp->lwp_stat == LSRUN || /* Obvious */ 504 lp->lwp_stat == LSSTOP, /* Set in tstop */ 505 ("tsleep %p %s %d", 506 ident, wmesg, lp->lwp_stat)); 507 508 /* 509 * We interlock the sleep queue if the caller has not already done 510 * it for us. This must be done before we potentially acquire any 511 * tokens or we can loose the wakeup. 512 */ 513 if ((flags & PINTERLOCKED) == 0) { 514 _tsleep_interlock(gd, ident, flags); 515 } 516 517 /* 518 * Setup for the current process (if this is a process). We must 519 * interlock with lwp_token to avoid remote wakeup races via 520 * setrunnable() 521 */ 522 if (lp) { 523 lwkt_gettoken(&lp->lwp_token); 524 if (catch) { 525 /* 526 * Early termination if PCATCH was set and a 527 * signal is pending, interlocked with the 528 * critical section. 529 * 530 * Early termination only occurs when tsleep() is 531 * entered while in a normal LSRUN state. 532 */ 533 if ((sig = CURSIG(lp)) != 0) 534 goto resume; 535 536 /* 537 * Causes ksignal to wake us up if a signal is 538 * received (interlocked with p->p_token). 539 */ 540 lp->lwp_flags |= LWP_SINTR; 541 } 542 } else { 543 KKASSERT(p == NULL); 544 } 545 546 /* 547 * Make sure the current process has been untangled from 548 * the userland scheduler and initialize slptime to start 549 * counting. 550 * 551 * NOTE: td->td_wakefromcpu is pre-set by the release function 552 * for the dfly scheduler, and then adjusted by _wakeup() 553 */ 554 if (lp) { 555 p->p_usched->release_curproc(lp); 556 lp->lwp_slptime = 0; 557 } 558 559 /* 560 * If the interlocked flag is set but our cpu bit in the slpqueue 561 * is no longer set, then a wakeup was processed inbetween the 562 * tsleep_interlock() (ours or the callers), and here. This can 563 * occur under numerous circumstances including when we release the 564 * current process. 565 * 566 * Extreme loads can cause the sending of an IPI (e.g. wakeup()'s) 567 * to process incoming IPIs, thus draining incoming wakeups. 568 */ 569 if ((td->td_flags & TDF_TSLEEPQ) == 0) { 570 logtsleep2(ilockfail, ident); 571 goto resume; 572 } 573 574 /* 575 * scheduling is blocked while in a critical section. Coincide 576 * the descheduled-by-tsleep flag with the descheduling of the 577 * lwkt. 578 * 579 * The timer callout is localized on our cpu and interlocked by 580 * our critical section. 581 */ 582 lwkt_deschedule_self(td); 583 td->td_flags |= TDF_TSLEEP_DESCHEDULED; 584 td->td_wmesg = wmesg; 585 586 /* 587 * Setup the timeout, if any. The timeout is only operable while 588 * the thread is flagged descheduled. 589 */ 590 KKASSERT((td->td_flags & TDF_TIMEOUT) == 0); 591 if (timo) { 592 callout_init_mp(&thandle); 593 callout_reset(&thandle, timo, endtsleep, td); 594 } 595 596 /* 597 * Beddy bye bye. 598 */ 599 if (lp) { 600 /* 601 * Ok, we are sleeping. Place us in the SSLEEP state. 602 */ 603 KKASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) == 0); 604 605 /* 606 * tstop() sets LSSTOP, so don't fiddle with that. 607 */ 608 if (lp->lwp_stat != LSSTOP) 609 lp->lwp_stat = LSSLEEP; 610 lp->lwp_ru.ru_nvcsw++; 611 p->p_usched->uload_update(lp); 612 lwkt_switch(); 613 614 /* 615 * And when we are woken up, put us back in LSRUN. If we 616 * slept for over a second, recalculate our estcpu. 617 */ 618 lp->lwp_stat = LSRUN; 619 if (lp->lwp_slptime) { 620 p->p_usched->uload_update(lp); 621 p->p_usched->recalculate(lp); 622 } 623 lp->lwp_slptime = 0; 624 } else { 625 lwkt_switch(); 626 } 627 628 /* 629 * Make sure we haven't switched cpus while we were asleep. It's 630 * not supposed to happen. Cleanup our temporary flags. 631 */ 632 KKASSERT(gd == td->td_gd); 633 634 /* 635 * Cleanup the timeout. If the timeout has already occured thandle 636 * has already been stopped, otherwise stop thandle. If the timeout 637 * is running (the callout thread must be blocked trying to get 638 * lwp_token) then wait for us to get scheduled. 639 */ 640 if (timo) { 641 while (td->td_flags & TDF_TIMEOUT_RUNNING) { 642 lwkt_deschedule_self(td); 643 td->td_wmesg = "tsrace"; 644 lwkt_switch(); 645 kprintf("td %p %s: timeout race\n", td, td->td_comm); 646 } 647 if (td->td_flags & TDF_TIMEOUT) { 648 td->td_flags &= ~TDF_TIMEOUT; 649 error = EWOULDBLOCK; 650 } else { 651 /* does not block when on same cpu */ 652 callout_stop(&thandle); 653 } 654 } 655 td->td_flags &= ~TDF_TSLEEP_DESCHEDULED; 656 657 /* 658 * Make sure we have been removed from the sleepq. In most 659 * cases this will have been done for us already but it is 660 * possible for a scheduling IPI to be in-flight from a 661 * previous tsleep/tsleep_interlock() or due to a straight-out 662 * call to lwkt_schedule() (in the case of an interrupt thread), 663 * causing a spurious wakeup. 664 */ 665 _tsleep_remove(td); 666 td->td_wmesg = NULL; 667 668 /* 669 * Figure out the correct error return. If interrupted by a 670 * signal we want to return EINTR or ERESTART. 671 */ 672 resume: 673 if (lp) { 674 if (catch && error == 0) { 675 if (sig != 0 || (sig = CURSIG(lp))) { 676 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 677 error = EINTR; 678 else 679 error = ERESTART; 680 } 681 } 682 lp->lwp_flags &= ~LWP_SINTR; 683 lwkt_reltoken(&lp->lwp_token); 684 } 685 logtsleep1(tsleep_end); 686 crit_exit_quick(td); 687 return (error); 688 } 689 690 /* 691 * Interlocked spinlock sleep. An exclusively held spinlock must 692 * be passed to ssleep(). The function will atomically release the 693 * spinlock and tsleep on the ident, then reacquire the spinlock and 694 * return. 695 * 696 * This routine is fairly important along the critical path, so optimize it 697 * heavily. 698 */ 699 int 700 ssleep(const volatile void *ident, struct spinlock *spin, int flags, 701 const char *wmesg, int timo) 702 { 703 globaldata_t gd = mycpu; 704 int error; 705 706 _tsleep_interlock(gd, ident, flags); 707 spin_unlock_quick(gd, spin); 708 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 709 spin_lock_quick(gd, spin); 710 711 return (error); 712 } 713 714 int 715 lksleep(const volatile void *ident, struct lock *lock, 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 lockmgr(lock, LK_RELEASE); 723 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 724 lockmgr(lock, LK_EXCLUSIVE); 725 726 return (error); 727 } 728 729 /* 730 * Interlocked mutex sleep. An exclusively held mutex must be passed 731 * to mtxsleep(). The function will atomically release the mutex 732 * and tsleep on the ident, then reacquire the mutex and return. 733 */ 734 int 735 mtxsleep(const volatile void *ident, struct mtx *mtx, int flags, 736 const char *wmesg, int timo) 737 { 738 globaldata_t gd = mycpu; 739 int error; 740 741 _tsleep_interlock(gd, ident, flags); 742 mtx_unlock(mtx); 743 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 744 mtx_lock_ex_quick(mtx, wmesg); 745 746 return (error); 747 } 748 749 /* 750 * Interlocked serializer sleep. An exclusively held serializer must 751 * be passed to zsleep(). The function will atomically release 752 * the serializer and tsleep on the ident, then reacquire the serializer 753 * and return. 754 */ 755 int 756 zsleep(const volatile void *ident, struct lwkt_serialize *slz, int flags, 757 const char *wmesg, int timo) 758 { 759 globaldata_t gd = mycpu; 760 int ret; 761 762 ASSERT_SERIALIZED(slz); 763 764 _tsleep_interlock(gd, ident, flags); 765 lwkt_serialize_exit(slz); 766 ret = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 767 lwkt_serialize_enter(slz); 768 769 return ret; 770 } 771 772 /* 773 * Directly block on the LWKT thread by descheduling it. This 774 * is much faster then tsleep(), but the only legal way to wake 775 * us up is to directly schedule the thread. 776 * 777 * Setting TDF_SINTR will cause new signals to directly schedule us. 778 * 779 * This routine must be called while in a critical section. 780 */ 781 int 782 lwkt_sleep(const char *wmesg, int flags) 783 { 784 thread_t td = curthread; 785 int sig; 786 787 if ((flags & PCATCH) == 0 || td->td_lwp == NULL) { 788 td->td_flags |= TDF_BLOCKED; 789 td->td_wmesg = wmesg; 790 lwkt_deschedule_self(td); 791 lwkt_switch(); 792 td->td_wmesg = NULL; 793 td->td_flags &= ~TDF_BLOCKED; 794 return(0); 795 } 796 if ((sig = CURSIG(td->td_lwp)) != 0) { 797 if (SIGISMEMBER(td->td_proc->p_sigacts->ps_sigintr, sig)) 798 return(EINTR); 799 else 800 return(ERESTART); 801 802 } 803 td->td_flags |= TDF_BLOCKED | TDF_SINTR; 804 td->td_wmesg = wmesg; 805 lwkt_deschedule_self(td); 806 lwkt_switch(); 807 td->td_flags &= ~(TDF_BLOCKED | TDF_SINTR); 808 td->td_wmesg = NULL; 809 return(0); 810 } 811 812 /* 813 * Implement the timeout for tsleep. 814 * 815 * This type of callout timeout is scheduled on the same cpu the process 816 * is sleeping on. Also, at the moment, the MP lock is held. 817 */ 818 static void 819 endtsleep(void *arg) 820 { 821 thread_t td = arg; 822 struct lwp *lp; 823 824 /* 825 * We are going to have to get the lwp_token, which means we might 826 * block. This can race a tsleep getting woken up by other means 827 * so set TDF_TIMEOUT_RUNNING to force the tsleep to wait for our 828 * processing to complete (sorry tsleep!). 829 * 830 * We can safely set td_flags because td MUST be on the same cpu 831 * as we are. 832 */ 833 KKASSERT(td->td_gd == mycpu); 834 crit_enter(); 835 td->td_flags |= TDF_TIMEOUT_RUNNING | TDF_TIMEOUT; 836 837 /* 838 * This can block but TDF_TIMEOUT_RUNNING will prevent the thread 839 * from exiting the tsleep on us. The flag is interlocked by virtue 840 * of lp being on the same cpu as we are. 841 */ 842 if ((lp = td->td_lwp) != NULL) 843 lwkt_gettoken(&lp->lwp_token); 844 845 KKASSERT(td->td_flags & TDF_TSLEEP_DESCHEDULED); 846 847 if (lp) { 848 if (lp->lwp_proc->p_stat != SSTOP) 849 setrunnable(lp); 850 lwkt_reltoken(&lp->lwp_token); 851 } else { 852 _tsleep_remove(td); 853 lwkt_schedule(td); 854 } 855 KKASSERT(td->td_gd == mycpu); 856 td->td_flags &= ~TDF_TIMEOUT_RUNNING; 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, flags, plus the originating 865 * cpu. 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 cpumask_t mask; 882 int id; 883 884 crit_enter(); 885 logtsleep2(wakeup_beg, ident); 886 gd = mycpu; 887 id = LOOKUP(ident); 888 qp = &gd->gd_tsleep_hash[id]; 889 restart: 890 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) { 891 ntd = TAILQ_NEXT(td, td_sleepq); 892 if (td->td_wchan == ident && 893 td->td_wdomain == (domain & PDOMAIN_MASK) 894 ) { 895 KKASSERT(td->td_gd == gd); 896 _tsleep_remove(td); 897 td->td_wakefromcpu = PWAKEUP_DECODE(domain); 898 if (td->td_flags & TDF_TSLEEP_DESCHEDULED) { 899 lwkt_schedule(td); 900 if (domain & PWAKEUP_ONE) 901 goto done; 902 } 903 goto restart; 904 } 905 } 906 907 /* 908 * We finished checking the current cpu but there still may be 909 * more work to do. Either wakeup_one was requested and no matching 910 * thread was found, or a normal wakeup was requested and we have 911 * to continue checking cpus. 912 * 913 * It should be noted that this scheme is actually less expensive then 914 * the old scheme when waking up multiple threads, since we send 915 * only one IPI message per target candidate which may then schedule 916 * multiple threads. Before we could have wound up sending an IPI 917 * message for each thread on the target cpu (!= current cpu) that 918 * needed to be woken up. 919 * 920 * NOTE: Wakeups occuring on remote cpus are asynchronous. This 921 * should be ok since we are passing idents in the IPI rather then 922 * thread pointers. 923 */ 924 if ((domain & PWAKEUP_MYCPU) == 0 && 925 (mask = slpque_cpumasks[id] & gd->gd_other_cpus) != 0) { 926 lwkt_send_ipiq2_mask(mask, _wakeup, ident, 927 domain | PWAKEUP_MYCPU); 928 } 929 done: 930 logtsleep1(wakeup_end); 931 crit_exit(); 932 } 933 934 /* 935 * Wakeup all threads tsleep()ing on the specified ident, on all cpus 936 */ 937 void 938 wakeup(const volatile void *ident) 939 { 940 globaldata_t gd = mycpu; 941 thread_t td = gd->gd_curthread; 942 943 if (td && (td->td_flags & TDF_DELAYED_WAKEUP)) { 944 if (!atomic_cmpset_ptr(&gd->gd_delayed_wakeup[0], NULL, ident)) { 945 if (!atomic_cmpset_ptr(&gd->gd_delayed_wakeup[1], NULL, ident)) 946 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, gd->gd_cpuid)); 947 } 948 return; 949 } 950 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, gd->gd_cpuid)); 951 } 952 953 /* 954 * Wakeup one thread tsleep()ing on the specified ident, on any cpu. 955 */ 956 void 957 wakeup_one(const volatile void *ident) 958 { 959 /* XXX potentially round-robin the first responding cpu */ 960 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mycpu->gd_cpuid) | 961 PWAKEUP_ONE); 962 } 963 964 /* 965 * Wakeup threads tsleep()ing on the specified ident on the current cpu 966 * only. 967 */ 968 void 969 wakeup_mycpu(const volatile void *ident) 970 { 971 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mycpu->gd_cpuid) | 972 PWAKEUP_MYCPU); 973 } 974 975 /* 976 * Wakeup one thread tsleep()ing on the specified ident on the current cpu 977 * only. 978 */ 979 void 980 wakeup_mycpu_one(const volatile void *ident) 981 { 982 /* XXX potentially round-robin the first responding cpu */ 983 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mycpu->gd_cpuid) | 984 PWAKEUP_MYCPU | PWAKEUP_ONE); 985 } 986 987 /* 988 * Wakeup all thread tsleep()ing on the specified ident on the specified cpu 989 * only. 990 */ 991 void 992 wakeup_oncpu(globaldata_t gd, const volatile void *ident) 993 { 994 globaldata_t mygd = mycpu; 995 if (gd == mycpu) { 996 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mygd->gd_cpuid) | 997 PWAKEUP_MYCPU); 998 } else { 999 lwkt_send_ipiq2(gd, _wakeup, __DEALL(ident), 1000 PWAKEUP_ENCODE(0, mygd->gd_cpuid) | 1001 PWAKEUP_MYCPU); 1002 } 1003 } 1004 1005 /* 1006 * Wakeup one thread tsleep()ing on the specified ident on the specified cpu 1007 * only. 1008 */ 1009 void 1010 wakeup_oncpu_one(globaldata_t gd, const volatile void *ident) 1011 { 1012 globaldata_t mygd = mycpu; 1013 if (gd == mygd) { 1014 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mygd->gd_cpuid) | 1015 PWAKEUP_MYCPU | PWAKEUP_ONE); 1016 } else { 1017 lwkt_send_ipiq2(gd, _wakeup, __DEALL(ident), 1018 PWAKEUP_ENCODE(0, mygd->gd_cpuid) | 1019 PWAKEUP_MYCPU | PWAKEUP_ONE); 1020 } 1021 } 1022 1023 /* 1024 * Wakeup all threads waiting on the specified ident that slept using 1025 * the specified domain, on all cpus. 1026 */ 1027 void 1028 wakeup_domain(const volatile void *ident, int domain) 1029 { 1030 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(domain, mycpu->gd_cpuid)); 1031 } 1032 1033 /* 1034 * Wakeup one thread waiting on the specified ident that slept using 1035 * the specified domain, on any cpu. 1036 */ 1037 void 1038 wakeup_domain_one(const volatile void *ident, int domain) 1039 { 1040 /* XXX potentially round-robin the first responding cpu */ 1041 _wakeup(__DEALL(ident), 1042 PWAKEUP_ENCODE(domain, mycpu->gd_cpuid) | PWAKEUP_ONE); 1043 } 1044 1045 void 1046 wakeup_start_delayed(void) 1047 { 1048 globaldata_t gd = mycpu; 1049 1050 crit_enter(); 1051 gd->gd_curthread->td_flags |= TDF_DELAYED_WAKEUP; 1052 crit_exit(); 1053 } 1054 1055 void 1056 wakeup_end_delayed(void) 1057 { 1058 globaldata_t gd = mycpu; 1059 1060 if (gd->gd_curthread->td_flags & TDF_DELAYED_WAKEUP) { 1061 crit_enter(); 1062 gd->gd_curthread->td_flags &= ~TDF_DELAYED_WAKEUP; 1063 if (gd->gd_delayed_wakeup[0] || gd->gd_delayed_wakeup[1]) { 1064 if (gd->gd_delayed_wakeup[0]) { 1065 wakeup(gd->gd_delayed_wakeup[0]); 1066 gd->gd_delayed_wakeup[0] = NULL; 1067 } 1068 if (gd->gd_delayed_wakeup[1]) { 1069 wakeup(gd->gd_delayed_wakeup[1]); 1070 gd->gd_delayed_wakeup[1] = NULL; 1071 } 1072 } 1073 crit_exit(); 1074 } 1075 } 1076 1077 /* 1078 * setrunnable() 1079 * 1080 * Make a process runnable. lp->lwp_token must be held on call and this 1081 * function must be called from the cpu owning lp. 1082 * 1083 * This only has an effect if we are in LSSTOP or LSSLEEP. 1084 */ 1085 void 1086 setrunnable(struct lwp *lp) 1087 { 1088 thread_t td = lp->lwp_thread; 1089 1090 ASSERT_LWKT_TOKEN_HELD(&lp->lwp_token); 1091 KKASSERT(td->td_gd == mycpu); 1092 crit_enter(); 1093 if (lp->lwp_stat == LSSTOP) 1094 lp->lwp_stat = LSSLEEP; 1095 if (lp->lwp_stat == LSSLEEP) { 1096 _tsleep_remove(td); 1097 lwkt_schedule(td); 1098 } else if (td->td_flags & TDF_SINTR) { 1099 lwkt_schedule(td); 1100 } 1101 crit_exit(); 1102 } 1103 1104 /* 1105 * The process is stopped due to some condition, usually because p_stat is 1106 * set to SSTOP, but also possibly due to being traced. 1107 * 1108 * Caller must hold p->p_token 1109 * 1110 * NOTE! If the caller sets SSTOP, the caller must also clear P_WAITED 1111 * because the parent may check the child's status before the child actually 1112 * gets to this routine. 1113 * 1114 * This routine is called with the current lwp only, typically just 1115 * before returning to userland if the process state is detected as 1116 * possibly being in a stopped state. 1117 */ 1118 void 1119 tstop(void) 1120 { 1121 struct lwp *lp = curthread->td_lwp; 1122 struct proc *p = lp->lwp_proc; 1123 struct proc *q; 1124 1125 lwkt_gettoken(&lp->lwp_token); 1126 crit_enter(); 1127 1128 /* 1129 * If LWP_MP_WSTOP is set, we were sleeping 1130 * while our process was stopped. At this point 1131 * we were already counted as stopped. 1132 */ 1133 if ((lp->lwp_mpflags & LWP_MP_WSTOP) == 0) { 1134 /* 1135 * If we're the last thread to stop, signal 1136 * our parent. 1137 */ 1138 p->p_nstopped++; 1139 atomic_set_int(&lp->lwp_mpflags, LWP_MP_WSTOP); 1140 wakeup(&p->p_nstopped); 1141 if (p->p_nstopped == p->p_nthreads) { 1142 /* 1143 * Token required to interlock kern_wait() 1144 */ 1145 q = p->p_pptr; 1146 PHOLD(q); 1147 lwkt_gettoken(&q->p_token); 1148 p->p_flags &= ~P_WAITED; 1149 wakeup(p->p_pptr); 1150 if ((q->p_sigacts->ps_flag & PS_NOCLDSTOP) == 0) 1151 ksignal(q, SIGCHLD); 1152 lwkt_reltoken(&q->p_token); 1153 PRELE(q); 1154 } 1155 } 1156 while (p->p_stat == SSTOP) { 1157 lp->lwp_stat = LSSTOP; 1158 tsleep(p, 0, "stop", 0); 1159 } 1160 p->p_nstopped--; 1161 atomic_clear_int(&lp->lwp_mpflags, LWP_MP_WSTOP); 1162 crit_exit(); 1163 lwkt_reltoken(&lp->lwp_token); 1164 } 1165 1166 /* 1167 * Compute a tenex style load average of a quantity on 1168 * 1, 5 and 15 minute intervals. 1169 */ 1170 static int loadav_count_runnable(struct lwp *p, void *data); 1171 1172 static void 1173 loadav(void *arg) 1174 { 1175 struct loadavg *avg; 1176 int i, nrun; 1177 1178 nrun = 0; 1179 alllwp_scan(loadav_count_runnable, &nrun); 1180 avg = &averunnable; 1181 for (i = 0; i < 3; i++) { 1182 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] + 1183 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT; 1184 } 1185 1186 /* 1187 * Schedule the next update to occur after 5 seconds, but add a 1188 * random variation to avoid synchronisation with processes that 1189 * run at regular intervals. 1190 */ 1191 callout_reset(&loadav_callout, hz * 4 + (int)(krandom() % (hz * 2 + 1)), 1192 loadav, NULL); 1193 } 1194 1195 static int 1196 loadav_count_runnable(struct lwp *lp, void *data) 1197 { 1198 int *nrunp = data; 1199 thread_t td; 1200 1201 switch (lp->lwp_stat) { 1202 case LSRUN: 1203 if ((td = lp->lwp_thread) == NULL) 1204 break; 1205 if (td->td_flags & TDF_BLOCKED) 1206 break; 1207 ++*nrunp; 1208 break; 1209 default: 1210 break; 1211 } 1212 lwkt_yield(); 1213 return(0); 1214 } 1215 1216 /* ARGSUSED */ 1217 static void 1218 sched_setup(void *dummy) 1219 { 1220 callout_init_mp(&loadav_callout); 1221 callout_init_mp(&schedcpu_callout); 1222 1223 /* Kick off timeout driven events by calling first time. */ 1224 schedcpu(NULL); 1225 loadav(NULL); 1226 } 1227 1228