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 #include <sys/kcollect.h> 51 #ifdef KTRACE 52 #include <sys/ktrace.h> 53 #endif 54 #include <sys/ktr.h> 55 #include <sys/serialize.h> 56 57 #include <sys/signal2.h> 58 #include <sys/thread2.h> 59 #include <sys/spinlock2.h> 60 #include <sys/mutex2.h> 61 62 #include <machine/cpu.h> 63 #include <machine/smp.h> 64 65 TAILQ_HEAD(tslpque, thread); 66 67 static void sched_setup (void *dummy); 68 SYSINIT(sched_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, sched_setup, NULL); 69 70 int lbolt; 71 void *lbolt_syncer; 72 int sched_quantum; /* Roundrobin scheduling quantum in ticks. */ 73 int ncpus; 74 int ncpus2, ncpus2_shift, ncpus2_mask; /* note: mask not cpumask_t */ 75 int ncpus_fit, ncpus_fit_mask; /* note: mask not cpumask_t */ 76 int safepri; 77 int tsleep_now_works; 78 int tsleep_crypto_dump = 0; 79 80 static struct callout loadav_callout; 81 static struct callout schedcpu_callout; 82 MALLOC_DEFINE(M_TSLEEP, "tslpque", "tsleep queues"); 83 84 #define __DEALL(ident) __DEQUALIFY(void *, ident) 85 86 #if !defined(KTR_TSLEEP) 87 #define KTR_TSLEEP KTR_ALL 88 #endif 89 KTR_INFO_MASTER(tsleep); 90 KTR_INFO(KTR_TSLEEP, tsleep, tsleep_beg, 0, "tsleep enter %p", const volatile void *ident); 91 KTR_INFO(KTR_TSLEEP, tsleep, tsleep_end, 1, "tsleep exit"); 92 KTR_INFO(KTR_TSLEEP, tsleep, wakeup_beg, 2, "wakeup enter %p", const volatile void *ident); 93 KTR_INFO(KTR_TSLEEP, tsleep, wakeup_end, 3, "wakeup exit"); 94 KTR_INFO(KTR_TSLEEP, tsleep, ilockfail, 4, "interlock failed %p", const volatile void *ident); 95 96 #define logtsleep1(name) KTR_LOG(tsleep_ ## name) 97 #define logtsleep2(name, val) KTR_LOG(tsleep_ ## name, val) 98 99 struct loadavg averunnable = 100 { {0, 0, 0}, FSCALE }; /* load average, of runnable procs */ 101 /* 102 * Constants for averages over 1, 5, and 15 minutes 103 * when sampling at 5 second intervals. 104 */ 105 static fixpt_t cexp[3] = { 106 0.9200444146293232 * FSCALE, /* exp(-1/12) */ 107 0.9834714538216174 * FSCALE, /* exp(-1/60) */ 108 0.9944598480048967 * FSCALE, /* exp(-1/180) */ 109 }; 110 111 static void endtsleep (void *); 112 static void loadav (void *arg); 113 static void schedcpu (void *arg); 114 115 /* 116 * Adjust the scheduler quantum. The quantum is specified in microseconds. 117 * Note that 'tick' is in microseconds per tick. 118 */ 119 static int 120 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS) 121 { 122 int error, new_val; 123 124 new_val = sched_quantum * ustick; 125 error = sysctl_handle_int(oidp, &new_val, 0, req); 126 if (error != 0 || req->newptr == NULL) 127 return (error); 128 if (new_val < ustick) 129 return (EINVAL); 130 sched_quantum = new_val / ustick; 131 return (0); 132 } 133 134 SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW, 135 0, sizeof sched_quantum, sysctl_kern_quantum, "I", ""); 136 137 static int pctcpu_decay = 10; 138 SYSCTL_INT(_kern, OID_AUTO, pctcpu_decay, CTLFLAG_RW, &pctcpu_decay, 0, ""); 139 140 /* 141 * kernel uses `FSCALE', userland (SHOULD) use kern.fscale 142 */ 143 int fscale __unused = FSCALE; /* exported to systat */ 144 SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, ""); 145 146 /* 147 * Recompute process priorities, once a second. 148 * 149 * Since the userland schedulers are typically event oriented, if the 150 * estcpu calculation at wakeup() time is not sufficient to make a 151 * process runnable relative to other processes in the system we have 152 * a 1-second recalc to help out. 153 * 154 * This code also allows us to store sysclock_t data in the process structure 155 * without fear of an overrun, since sysclock_t are guarenteed to hold 156 * several seconds worth of count. 157 * 158 * WARNING! callouts can preempt normal threads. However, they will not 159 * preempt a thread holding a spinlock so we *can* safely use spinlocks. 160 */ 161 static int schedcpu_stats(struct proc *p, void *data __unused); 162 static int schedcpu_resource(struct proc *p, void *data __unused); 163 164 static void 165 schedcpu(void *arg) 166 { 167 allproc_scan(schedcpu_stats, NULL); 168 allproc_scan(schedcpu_resource, NULL); 169 wakeup((caddr_t)&lbolt); 170 wakeup(lbolt_syncer); 171 callout_reset(&schedcpu_callout, hz, schedcpu, NULL); 172 } 173 174 /* 175 * General process statistics once a second 176 */ 177 static int 178 schedcpu_stats(struct proc *p, void *data __unused) 179 { 180 struct lwp *lp; 181 182 /* 183 * Threads may not be completely set up if process in SIDL state. 184 */ 185 if (p->p_stat == SIDL) 186 return(0); 187 188 PHOLD(p); 189 if (lwkt_trytoken(&p->p_token) == FALSE) { 190 PRELE(p); 191 return(0); 192 } 193 194 p->p_swtime++; 195 FOREACH_LWP_IN_PROC(lp, p) { 196 if (lp->lwp_stat == LSSLEEP) { 197 ++lp->lwp_slptime; 198 if (lp->lwp_slptime == 1) 199 p->p_usched->uload_update(lp); 200 } 201 202 /* 203 * Only recalculate processes that are active or have slept 204 * less then 2 seconds. The schedulers understand this. 205 * Otherwise decay by 50% per second. 206 */ 207 if (lp->lwp_slptime <= 1) { 208 p->p_usched->recalculate(lp); 209 } else { 210 int decay; 211 212 decay = pctcpu_decay; 213 cpu_ccfence(); 214 if (decay <= 1) 215 decay = 1; 216 if (decay > 100) 217 decay = 100; 218 lp->lwp_pctcpu = (lp->lwp_pctcpu * (decay - 1)) / decay; 219 } 220 } 221 lwkt_reltoken(&p->p_token); 222 lwkt_yield(); 223 PRELE(p); 224 return(0); 225 } 226 227 /* 228 * Resource checks. XXX break out since ksignal/killproc can block, 229 * limiting us to one process killed per second. There is probably 230 * a better way. 231 */ 232 static int 233 schedcpu_resource(struct proc *p, void *data __unused) 234 { 235 u_int64_t ttime; 236 struct lwp *lp; 237 238 if (p->p_stat == SIDL) 239 return(0); 240 241 PHOLD(p); 242 if (lwkt_trytoken(&p->p_token) == FALSE) { 243 PRELE(p); 244 return(0); 245 } 246 247 if (p->p_stat == SZOMB || p->p_limit == NULL) { 248 lwkt_reltoken(&p->p_token); 249 PRELE(p); 250 return(0); 251 } 252 253 ttime = 0; 254 FOREACH_LWP_IN_PROC(lp, p) { 255 /* 256 * We may have caught an lp in the middle of being 257 * created, lwp_thread can be NULL. 258 */ 259 if (lp->lwp_thread) { 260 ttime += lp->lwp_thread->td_sticks; 261 ttime += lp->lwp_thread->td_uticks; 262 } 263 } 264 265 switch(plimit_testcpulimit(p->p_limit, ttime)) { 266 case PLIMIT_TESTCPU_KILL: 267 killproc(p, "exceeded maximum CPU limit"); 268 break; 269 case PLIMIT_TESTCPU_XCPU: 270 if ((p->p_flags & P_XCPU) == 0) { 271 p->p_flags |= P_XCPU; 272 ksignal(p, SIGXCPU); 273 } 274 break; 275 default: 276 break; 277 } 278 lwkt_reltoken(&p->p_token); 279 lwkt_yield(); 280 PRELE(p); 281 return(0); 282 } 283 284 /* 285 * This is only used by ps. Generate a cpu percentage use over 286 * a period of one second. 287 */ 288 void 289 updatepcpu(struct lwp *lp, int cpticks, int ttlticks) 290 { 291 fixpt_t acc; 292 int remticks; 293 294 acc = (cpticks << FSHIFT) / ttlticks; 295 if (ttlticks >= ESTCPUFREQ) { 296 lp->lwp_pctcpu = acc; 297 } else { 298 remticks = ESTCPUFREQ - ttlticks; 299 lp->lwp_pctcpu = (acc * ttlticks + lp->lwp_pctcpu * remticks) / 300 ESTCPUFREQ; 301 } 302 } 303 304 /* 305 * tsleep/wakeup hash table parameters. Try to find the sweet spot for 306 * like addresses being slept on. 307 */ 308 #define TABLESIZE 4001 309 #define LOOKUP(x) (((u_int)(uintptr_t)(x)) % TABLESIZE) 310 311 static cpumask_t slpque_cpumasks[TABLESIZE]; 312 313 /* 314 * General scheduler initialization. We force a reschedule 25 times 315 * a second by default. Note that cpu0 is initialized in early boot and 316 * cannot make any high level calls. 317 * 318 * Each cpu has its own sleep queue. 319 */ 320 void 321 sleep_gdinit(globaldata_t gd) 322 { 323 static struct tslpque slpque_cpu0[TABLESIZE]; 324 int i; 325 326 if (gd->gd_cpuid == 0) { 327 sched_quantum = (hz + 24) / 25; 328 gd->gd_tsleep_hash = slpque_cpu0; 329 } else { 330 gd->gd_tsleep_hash = kmalloc(sizeof(slpque_cpu0), 331 M_TSLEEP, M_WAITOK | M_ZERO); 332 } 333 for (i = 0; i < TABLESIZE; ++i) 334 TAILQ_INIT(&gd->gd_tsleep_hash[i]); 335 } 336 337 /* 338 * This is a dandy function that allows us to interlock tsleep/wakeup 339 * operations with unspecified upper level locks, such as lockmgr locks, 340 * simply by holding a critical section. The sequence is: 341 * 342 * (acquire upper level lock) 343 * tsleep_interlock(blah) 344 * (release upper level lock) 345 * tsleep(blah, ...) 346 * 347 * Basically this functions queues us on the tsleep queue without actually 348 * descheduling us. When tsleep() is later called with PINTERLOCK it 349 * assumes the thread was already queued, otherwise it queues it there. 350 * 351 * Thus it is possible to receive the wakeup prior to going to sleep and 352 * the race conditions are covered. 353 */ 354 static __inline void 355 _tsleep_interlock(globaldata_t gd, const volatile void *ident, int flags) 356 { 357 thread_t td = gd->gd_curthread; 358 int id; 359 360 crit_enter_quick(td); 361 if (td->td_flags & TDF_TSLEEPQ) { 362 id = LOOKUP(td->td_wchan); 363 TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_sleepq); 364 if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL) { 365 ATOMIC_CPUMASK_NANDBIT(slpque_cpumasks[id], 366 gd->gd_cpuid); 367 } 368 } else { 369 td->td_flags |= TDF_TSLEEPQ; 370 } 371 id = LOOKUP(ident); 372 TAILQ_INSERT_TAIL(&gd->gd_tsleep_hash[id], td, td_sleepq); 373 ATOMIC_CPUMASK_ORBIT(slpque_cpumasks[id], gd->gd_cpuid); 374 td->td_wchan = ident; 375 td->td_wdomain = flags & PDOMAIN_MASK; 376 crit_exit_quick(td); 377 } 378 379 void 380 tsleep_interlock(const volatile void *ident, int flags) 381 { 382 _tsleep_interlock(mycpu, ident, flags); 383 } 384 385 /* 386 * Remove thread from sleepq. Must be called with a critical section held. 387 * The thread must not be migrating. 388 */ 389 static __inline void 390 _tsleep_remove(thread_t td) 391 { 392 globaldata_t gd = mycpu; 393 int id; 394 395 KKASSERT(td->td_gd == gd && IN_CRITICAL_SECT(td)); 396 KKASSERT((td->td_flags & TDF_MIGRATING) == 0); 397 if (td->td_flags & TDF_TSLEEPQ) { 398 td->td_flags &= ~TDF_TSLEEPQ; 399 id = LOOKUP(td->td_wchan); 400 TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_sleepq); 401 if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL) { 402 ATOMIC_CPUMASK_NANDBIT(slpque_cpumasks[id], 403 gd->gd_cpuid); 404 } 405 td->td_wchan = NULL; 406 td->td_wdomain = 0; 407 } 408 } 409 410 void 411 tsleep_remove(thread_t td) 412 { 413 _tsleep_remove(td); 414 } 415 416 /* 417 * General sleep call. Suspends the current process until a wakeup is 418 * performed on the specified identifier. The process will then be made 419 * runnable with the specified priority. Sleeps at most timo/hz seconds 420 * (0 means no timeout). If flags includes PCATCH flag, signals are checked 421 * before and after sleeping, else signals are not checked. Returns 0 if 422 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a 423 * signal needs to be delivered, ERESTART is returned if the current system 424 * call should be restarted if possible, and EINTR is returned if the system 425 * call should be interrupted by the signal (return EINTR). 426 * 427 * Note that if we are a process, we release_curproc() before messing with 428 * the LWKT scheduler. 429 * 430 * During autoconfiguration or after a panic, a sleep will simply 431 * lower the priority briefly to allow interrupts, then return. 432 * 433 * WARNING! This code can't block (short of switching away), or bad things 434 * will happen. No getting tokens, no blocking locks, etc. 435 */ 436 int 437 tsleep(const volatile void *ident, int flags, const char *wmesg, int timo) 438 { 439 struct thread *td = curthread; 440 struct lwp *lp = td->td_lwp; 441 struct proc *p = td->td_proc; /* may be NULL */ 442 globaldata_t gd; 443 int sig; 444 int catch; 445 int error; 446 int oldpri; 447 struct callout thandle; 448 449 /* 450 * Currently a severe hack. Make sure any delayed wakeups 451 * are flushed before we sleep or we might deadlock on whatever 452 * event we are sleeping on. 453 */ 454 if (td->td_flags & TDF_DELAYED_WAKEUP) 455 wakeup_end_delayed(); 456 457 /* 458 * NOTE: removed KTRPOINT, it could cause races due to blocking 459 * even in stable. Just scrap it for now. 460 */ 461 if (!tsleep_crypto_dump && (tsleep_now_works == 0 || panicstr)) { 462 /* 463 * After a panic, or before we actually have an operational 464 * softclock, just give interrupts a chance, then just return; 465 * 466 * don't run any other procs or panic below, 467 * in case this is the idle process and already asleep. 468 */ 469 splz(); 470 oldpri = td->td_pri; 471 lwkt_setpri_self(safepri); 472 lwkt_switch(); 473 lwkt_setpri_self(oldpri); 474 return (0); 475 } 476 logtsleep2(tsleep_beg, ident); 477 gd = td->td_gd; 478 KKASSERT(td != &gd->gd_idlethread); /* you must be kidding! */ 479 td->td_wakefromcpu = -1; /* overwritten by _wakeup */ 480 481 /* 482 * NOTE: all of this occurs on the current cpu, including any 483 * callout-based wakeups, so a critical section is a sufficient 484 * interlock. 485 * 486 * The entire sequence through to where we actually sleep must 487 * run without breaking the critical section. 488 */ 489 catch = flags & PCATCH; 490 error = 0; 491 sig = 0; 492 493 crit_enter_quick(td); 494 495 KASSERT(ident != NULL, ("tsleep: no ident")); 496 KASSERT(lp == NULL || 497 lp->lwp_stat == LSRUN || /* Obvious */ 498 lp->lwp_stat == LSSTOP, /* Set in tstop */ 499 ("tsleep %p %s %d", 500 ident, wmesg, lp->lwp_stat)); 501 502 /* 503 * We interlock the sleep queue if the caller has not already done 504 * it for us. This must be done before we potentially acquire any 505 * tokens or we can loose the wakeup. 506 */ 507 if ((flags & PINTERLOCKED) == 0) { 508 _tsleep_interlock(gd, ident, flags); 509 } 510 511 /* 512 * Setup for the current process (if this is a process). We must 513 * interlock with lwp_token to avoid remote wakeup races via 514 * setrunnable() 515 */ 516 if (lp) { 517 lwkt_gettoken(&lp->lwp_token); 518 519 /* 520 * If the umbrella process is in the SCORE state then 521 * make sure that the thread is flagged going into a 522 * normal sleep to allow the core dump to proceed, otherwise 523 * the coredump can end up waiting forever. If the normal 524 * sleep is woken up, the thread will enter a stopped state 525 * upon return to userland. 526 * 527 * We do not want to interrupt or cause a thread exist at 528 * this juncture because that will mess-up the state the 529 * coredump is trying to save. 530 */ 531 if (p->p_stat == SCORE && 532 (lp->lwp_mpflags & LWP_MP_WSTOP) == 0) { 533 atomic_set_int(&lp->lwp_mpflags, LWP_MP_WSTOP); 534 ++p->p_nstopped; 535 } 536 537 /* 538 * PCATCH requested. 539 */ 540 if (catch) { 541 /* 542 * Early termination if PCATCH was set and a 543 * signal is pending, interlocked with the 544 * critical section. 545 * 546 * Early termination only occurs when tsleep() is 547 * entered while in a normal LSRUN state. 548 */ 549 if ((sig = CURSIG(lp)) != 0) 550 goto resume; 551 552 /* 553 * Causes ksignal to wake us up if a signal is 554 * received (interlocked with lp->lwp_token). 555 */ 556 lp->lwp_flags |= LWP_SINTR; 557 } 558 } else { 559 KKASSERT(p == NULL); 560 } 561 562 /* 563 * Make sure the current process has been untangled from 564 * the userland scheduler and initialize slptime to start 565 * counting. 566 * 567 * NOTE: td->td_wakefromcpu is pre-set by the release function 568 * for the dfly scheduler, and then adjusted by _wakeup() 569 */ 570 if (lp) { 571 p->p_usched->release_curproc(lp); 572 lp->lwp_slptime = 0; 573 } 574 575 /* 576 * If the interlocked flag is set but our cpu bit in the slpqueue 577 * is no longer set, then a wakeup was processed inbetween the 578 * tsleep_interlock() (ours or the callers), and here. This can 579 * occur under numerous circumstances including when we release the 580 * current process. 581 * 582 * Extreme loads can cause the sending of an IPI (e.g. wakeup()'s) 583 * to process incoming IPIs, thus draining incoming wakeups. 584 */ 585 if ((td->td_flags & TDF_TSLEEPQ) == 0) { 586 logtsleep2(ilockfail, ident); 587 goto resume; 588 } 589 590 /* 591 * scheduling is blocked while in a critical section. Coincide 592 * the descheduled-by-tsleep flag with the descheduling of the 593 * lwkt. 594 * 595 * The timer callout is localized on our cpu and interlocked by 596 * our critical section. 597 */ 598 lwkt_deschedule_self(td); 599 td->td_flags |= TDF_TSLEEP_DESCHEDULED; 600 td->td_wmesg = wmesg; 601 602 /* 603 * Setup the timeout, if any. The timeout is only operable while 604 * the thread is flagged descheduled. 605 */ 606 KKASSERT((td->td_flags & TDF_TIMEOUT) == 0); 607 if (timo) { 608 callout_init_mp(&thandle); 609 callout_reset(&thandle, timo, endtsleep, td); 610 } 611 612 /* 613 * Beddy bye bye. 614 */ 615 if (lp) { 616 /* 617 * Ok, we are sleeping. Place us in the SSLEEP state. 618 */ 619 KKASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) == 0); 620 621 /* 622 * tstop() sets LSSTOP, so don't fiddle with that. 623 */ 624 if (lp->lwp_stat != LSSTOP) 625 lp->lwp_stat = LSSLEEP; 626 lp->lwp_ru.ru_nvcsw++; 627 p->p_usched->uload_update(lp); 628 lwkt_switch(); 629 630 /* 631 * And when we are woken up, put us back in LSRUN. If we 632 * slept for over a second, recalculate our estcpu. 633 */ 634 lp->lwp_stat = LSRUN; 635 if (lp->lwp_slptime) { 636 p->p_usched->uload_update(lp); 637 p->p_usched->recalculate(lp); 638 } 639 lp->lwp_slptime = 0; 640 } else { 641 lwkt_switch(); 642 } 643 644 /* 645 * Make sure we haven't switched cpus while we were asleep. It's 646 * not supposed to happen. Cleanup our temporary flags. 647 */ 648 KKASSERT(gd == td->td_gd); 649 650 /* 651 * Cleanup the timeout. If the timeout has already occured thandle 652 * has already been stopped, otherwise stop thandle. If the timeout 653 * is running (the callout thread must be blocked trying to get 654 * lwp_token) then wait for us to get scheduled. 655 */ 656 if (timo) { 657 while (td->td_flags & TDF_TIMEOUT_RUNNING) { 658 /* else we won't get rescheduled! */ 659 if (lp->lwp_stat != LSSTOP) 660 lp->lwp_stat = LSSLEEP; 661 lwkt_deschedule_self(td); 662 td->td_wmesg = "tsrace"; 663 lwkt_switch(); 664 kprintf("td %p %s: timeout race\n", td, td->td_comm); 665 } 666 if (td->td_flags & TDF_TIMEOUT) { 667 td->td_flags &= ~TDF_TIMEOUT; 668 error = EWOULDBLOCK; 669 } else { 670 /* does not block when on same cpu */ 671 callout_stop(&thandle); 672 } 673 } 674 td->td_flags &= ~TDF_TSLEEP_DESCHEDULED; 675 676 /* 677 * Make sure we have been removed from the sleepq. In most 678 * cases this will have been done for us already but it is 679 * possible for a scheduling IPI to be in-flight from a 680 * previous tsleep/tsleep_interlock() or due to a straight-out 681 * call to lwkt_schedule() (in the case of an interrupt thread), 682 * causing a spurious wakeup. 683 */ 684 _tsleep_remove(td); 685 td->td_wmesg = NULL; 686 687 /* 688 * Figure out the correct error return. If interrupted by a 689 * signal we want to return EINTR or ERESTART. 690 */ 691 resume: 692 if (lp) { 693 if (catch && error == 0) { 694 if (sig != 0 || (sig = CURSIG(lp))) { 695 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 696 error = EINTR; 697 else 698 error = ERESTART; 699 } 700 } 701 702 lp->lwp_flags &= ~LWP_SINTR; 703 704 /* 705 * Unconditionally set us to LSRUN on resume. lwp_stat could 706 * be in a weird state due to the goto resume, particularly 707 * when tsleep() is called from tstop(). 708 */ 709 lp->lwp_stat = LSRUN; 710 lwkt_reltoken(&lp->lwp_token); 711 } 712 logtsleep1(tsleep_end); 713 crit_exit_quick(td); 714 return (error); 715 } 716 717 /* 718 * Interlocked spinlock sleep. An exclusively held spinlock must 719 * be passed to ssleep(). The function will atomically release the 720 * spinlock and tsleep on the ident, then reacquire the spinlock and 721 * return. 722 * 723 * This routine is fairly important along the critical path, so optimize it 724 * heavily. 725 */ 726 int 727 ssleep(const volatile void *ident, struct spinlock *spin, int flags, 728 const char *wmesg, int timo) 729 { 730 globaldata_t gd = mycpu; 731 int error; 732 733 _tsleep_interlock(gd, ident, flags); 734 spin_unlock_quick(gd, spin); 735 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 736 _spin_lock_quick(gd, spin, wmesg); 737 738 return (error); 739 } 740 741 int 742 lksleep(const volatile void *ident, struct lock *lock, int flags, 743 const char *wmesg, int timo) 744 { 745 globaldata_t gd = mycpu; 746 int error; 747 748 _tsleep_interlock(gd, ident, flags); 749 lockmgr(lock, LK_RELEASE); 750 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 751 lockmgr(lock, LK_EXCLUSIVE); 752 753 return (error); 754 } 755 756 /* 757 * Interlocked mutex sleep. An exclusively held mutex must be passed 758 * to mtxsleep(). The function will atomically release the mutex 759 * and tsleep on the ident, then reacquire the mutex and return. 760 */ 761 int 762 mtxsleep(const volatile void *ident, struct mtx *mtx, int flags, 763 const char *wmesg, int timo) 764 { 765 globaldata_t gd = mycpu; 766 int error; 767 768 _tsleep_interlock(gd, ident, flags); 769 mtx_unlock(mtx); 770 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 771 mtx_lock_ex_quick(mtx); 772 773 return (error); 774 } 775 776 /* 777 * Interlocked serializer sleep. An exclusively held serializer must 778 * be passed to zsleep(). The function will atomically release 779 * the serializer and tsleep on the ident, then reacquire the serializer 780 * and return. 781 */ 782 int 783 zsleep(const volatile void *ident, struct lwkt_serialize *slz, int flags, 784 const char *wmesg, int timo) 785 { 786 globaldata_t gd = mycpu; 787 int ret; 788 789 ASSERT_SERIALIZED(slz); 790 791 _tsleep_interlock(gd, ident, flags); 792 lwkt_serialize_exit(slz); 793 ret = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 794 lwkt_serialize_enter(slz); 795 796 return ret; 797 } 798 799 /* 800 * Directly block on the LWKT thread by descheduling it. This 801 * is much faster then tsleep(), but the only legal way to wake 802 * us up is to directly schedule the thread. 803 * 804 * Setting TDF_SINTR will cause new signals to directly schedule us. 805 * 806 * This routine must be called while in a critical section. 807 */ 808 int 809 lwkt_sleep(const char *wmesg, int flags) 810 { 811 thread_t td = curthread; 812 int sig; 813 814 if ((flags & PCATCH) == 0 || td->td_lwp == NULL) { 815 td->td_flags |= TDF_BLOCKED; 816 td->td_wmesg = wmesg; 817 lwkt_deschedule_self(td); 818 lwkt_switch(); 819 td->td_wmesg = NULL; 820 td->td_flags &= ~TDF_BLOCKED; 821 return(0); 822 } 823 if ((sig = CURSIG(td->td_lwp)) != 0) { 824 if (SIGISMEMBER(td->td_proc->p_sigacts->ps_sigintr, sig)) 825 return(EINTR); 826 else 827 return(ERESTART); 828 829 } 830 td->td_flags |= TDF_BLOCKED | TDF_SINTR; 831 td->td_wmesg = wmesg; 832 lwkt_deschedule_self(td); 833 lwkt_switch(); 834 td->td_flags &= ~(TDF_BLOCKED | TDF_SINTR); 835 td->td_wmesg = NULL; 836 return(0); 837 } 838 839 /* 840 * Implement the timeout for tsleep. 841 * 842 * This type of callout timeout is scheduled on the same cpu the process 843 * is sleeping on. Also, at the moment, the MP lock is held. 844 */ 845 static void 846 endtsleep(void *arg) 847 { 848 thread_t td = arg; 849 struct lwp *lp; 850 851 /* 852 * We are going to have to get the lwp_token, which means we might 853 * block. This can race a tsleep getting woken up by other means 854 * so set TDF_TIMEOUT_RUNNING to force the tsleep to wait for our 855 * processing to complete (sorry tsleep!). 856 * 857 * We can safely set td_flags because td MUST be on the same cpu 858 * as we are. 859 */ 860 KKASSERT(td->td_gd == mycpu); 861 crit_enter(); 862 td->td_flags |= TDF_TIMEOUT_RUNNING | TDF_TIMEOUT; 863 864 /* 865 * This can block but TDF_TIMEOUT_RUNNING will prevent the thread 866 * from exiting the tsleep on us. The flag is interlocked by virtue 867 * of lp being on the same cpu as we are. 868 */ 869 if ((lp = td->td_lwp) != NULL) 870 lwkt_gettoken(&lp->lwp_token); 871 872 KKASSERT(td->td_flags & TDF_TSLEEP_DESCHEDULED); 873 874 if (lp) { 875 /* 876 * callout timer should normally never be set in tstop() 877 * because it passes a timeout of 0. However, there is a 878 * case during thread exit (which SSTOP's all the threads) 879 * for which tstop() must break out and can (properly) leave 880 * the thread in LSSTOP. 881 */ 882 KKASSERT(lp->lwp_stat != LSSTOP || 883 (lp->lwp_mpflags & LWP_MP_WEXIT)); 884 setrunnable(lp); 885 lwkt_reltoken(&lp->lwp_token); 886 } else { 887 _tsleep_remove(td); 888 lwkt_schedule(td); 889 } 890 KKASSERT(td->td_gd == mycpu); 891 td->td_flags &= ~TDF_TIMEOUT_RUNNING; 892 crit_exit(); 893 } 894 895 /* 896 * Make all processes sleeping on the specified identifier runnable. 897 * count may be zero or one only. 898 * 899 * The domain encodes the sleep/wakeup domain, flags, plus the originating 900 * cpu. 901 * 902 * This call may run without the MP lock held. We can only manipulate thread 903 * state on the cpu owning the thread. We CANNOT manipulate process state 904 * at all. 905 * 906 * _wakeup() can be passed to an IPI so we can't use (const volatile 907 * void *ident). 908 */ 909 static void 910 _wakeup(void *ident, int domain) 911 { 912 struct tslpque *qp; 913 struct thread *td; 914 struct thread *ntd; 915 globaldata_t gd; 916 cpumask_t mask; 917 int id; 918 919 crit_enter(); 920 logtsleep2(wakeup_beg, ident); 921 gd = mycpu; 922 id = LOOKUP(ident); 923 qp = &gd->gd_tsleep_hash[id]; 924 restart: 925 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) { 926 ntd = TAILQ_NEXT(td, td_sleepq); 927 if (td->td_wchan == ident && 928 td->td_wdomain == (domain & PDOMAIN_MASK) 929 ) { 930 KKASSERT(td->td_gd == gd); 931 _tsleep_remove(td); 932 td->td_wakefromcpu = PWAKEUP_DECODE(domain); 933 if (td->td_flags & TDF_TSLEEP_DESCHEDULED) { 934 lwkt_schedule(td); 935 if (domain & PWAKEUP_ONE) 936 goto done; 937 } 938 goto restart; 939 } 940 } 941 942 /* 943 * We finished checking the current cpu but there still may be 944 * more work to do. Either wakeup_one was requested and no matching 945 * thread was found, or a normal wakeup was requested and we have 946 * to continue checking cpus. 947 * 948 * It should be noted that this scheme is actually less expensive then 949 * the old scheme when waking up multiple threads, since we send 950 * only one IPI message per target candidate which may then schedule 951 * multiple threads. Before we could have wound up sending an IPI 952 * message for each thread on the target cpu (!= current cpu) that 953 * needed to be woken up. 954 * 955 * NOTE: Wakeups occuring on remote cpus are asynchronous. This 956 * should be ok since we are passing idents in the IPI rather then 957 * thread pointers. 958 */ 959 if ((domain & PWAKEUP_MYCPU) == 0) { 960 mask = slpque_cpumasks[id]; 961 CPUMASK_ANDMASK(mask, gd->gd_other_cpus); 962 if (CPUMASK_TESTNZERO(mask)) { 963 lwkt_send_ipiq2_mask(mask, _wakeup, ident, 964 domain | PWAKEUP_MYCPU); 965 } 966 } 967 done: 968 logtsleep1(wakeup_end); 969 crit_exit(); 970 } 971 972 /* 973 * Wakeup all threads tsleep()ing on the specified ident, on all cpus 974 */ 975 void 976 wakeup(const volatile void *ident) 977 { 978 globaldata_t gd = mycpu; 979 thread_t td = gd->gd_curthread; 980 981 if (td && (td->td_flags & TDF_DELAYED_WAKEUP)) { 982 /* 983 * If we are in a delayed wakeup section, record up to two wakeups in 984 * a per-CPU queue and issue them when we block or exit the delayed 985 * wakeup section. 986 */ 987 if (atomic_cmpset_ptr(&gd->gd_delayed_wakeup[0], NULL, ident)) 988 return; 989 if (atomic_cmpset_ptr(&gd->gd_delayed_wakeup[1], NULL, ident)) 990 return; 991 992 ident = atomic_swap_ptr(__DEQUALIFY(volatile void **, &gd->gd_delayed_wakeup[1]), 993 __DEALL(ident)); 994 ident = atomic_swap_ptr(__DEQUALIFY(volatile void **, &gd->gd_delayed_wakeup[0]), 995 __DEALL(ident)); 996 } 997 998 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, gd->gd_cpuid)); 999 } 1000 1001 /* 1002 * Wakeup one thread tsleep()ing on the specified ident, on any cpu. 1003 */ 1004 void 1005 wakeup_one(const volatile void *ident) 1006 { 1007 /* XXX potentially round-robin the first responding cpu */ 1008 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mycpu->gd_cpuid) | 1009 PWAKEUP_ONE); 1010 } 1011 1012 /* 1013 * Wakeup threads tsleep()ing on the specified ident on the current cpu 1014 * only. 1015 */ 1016 void 1017 wakeup_mycpu(const volatile void *ident) 1018 { 1019 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mycpu->gd_cpuid) | 1020 PWAKEUP_MYCPU); 1021 } 1022 1023 /* 1024 * Wakeup one thread tsleep()ing on the specified ident on the current cpu 1025 * only. 1026 */ 1027 void 1028 wakeup_mycpu_one(const volatile void *ident) 1029 { 1030 /* XXX potentially round-robin the first responding cpu */ 1031 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mycpu->gd_cpuid) | 1032 PWAKEUP_MYCPU | PWAKEUP_ONE); 1033 } 1034 1035 /* 1036 * Wakeup all thread tsleep()ing on the specified ident on the specified cpu 1037 * only. 1038 */ 1039 void 1040 wakeup_oncpu(globaldata_t gd, const volatile void *ident) 1041 { 1042 globaldata_t mygd = mycpu; 1043 if (gd == mycpu) { 1044 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mygd->gd_cpuid) | 1045 PWAKEUP_MYCPU); 1046 } else { 1047 lwkt_send_ipiq2(gd, _wakeup, __DEALL(ident), 1048 PWAKEUP_ENCODE(0, mygd->gd_cpuid) | 1049 PWAKEUP_MYCPU); 1050 } 1051 } 1052 1053 /* 1054 * Wakeup one thread tsleep()ing on the specified ident on the specified cpu 1055 * only. 1056 */ 1057 void 1058 wakeup_oncpu_one(globaldata_t gd, const volatile void *ident) 1059 { 1060 globaldata_t mygd = mycpu; 1061 if (gd == mygd) { 1062 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mygd->gd_cpuid) | 1063 PWAKEUP_MYCPU | PWAKEUP_ONE); 1064 } else { 1065 lwkt_send_ipiq2(gd, _wakeup, __DEALL(ident), 1066 PWAKEUP_ENCODE(0, mygd->gd_cpuid) | 1067 PWAKEUP_MYCPU | PWAKEUP_ONE); 1068 } 1069 } 1070 1071 /* 1072 * Wakeup all threads waiting on the specified ident that slept using 1073 * the specified domain, on all cpus. 1074 */ 1075 void 1076 wakeup_domain(const volatile void *ident, int domain) 1077 { 1078 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(domain, mycpu->gd_cpuid)); 1079 } 1080 1081 /* 1082 * Wakeup one thread waiting on the specified ident that slept using 1083 * the specified domain, on any cpu. 1084 */ 1085 void 1086 wakeup_domain_one(const volatile void *ident, int domain) 1087 { 1088 /* XXX potentially round-robin the first responding cpu */ 1089 _wakeup(__DEALL(ident), 1090 PWAKEUP_ENCODE(domain, mycpu->gd_cpuid) | PWAKEUP_ONE); 1091 } 1092 1093 void 1094 wakeup_start_delayed(void) 1095 { 1096 globaldata_t gd = mycpu; 1097 1098 crit_enter(); 1099 gd->gd_curthread->td_flags |= TDF_DELAYED_WAKEUP; 1100 crit_exit(); 1101 } 1102 1103 void 1104 wakeup_end_delayed(void) 1105 { 1106 globaldata_t gd = mycpu; 1107 1108 if (gd->gd_curthread->td_flags & TDF_DELAYED_WAKEUP) { 1109 crit_enter(); 1110 gd->gd_curthread->td_flags &= ~TDF_DELAYED_WAKEUP; 1111 if (gd->gd_delayed_wakeup[0] || gd->gd_delayed_wakeup[1]) { 1112 if (gd->gd_delayed_wakeup[0]) { 1113 wakeup(gd->gd_delayed_wakeup[0]); 1114 gd->gd_delayed_wakeup[0] = NULL; 1115 } 1116 if (gd->gd_delayed_wakeup[1]) { 1117 wakeup(gd->gd_delayed_wakeup[1]); 1118 gd->gd_delayed_wakeup[1] = NULL; 1119 } 1120 } 1121 crit_exit(); 1122 } 1123 } 1124 1125 /* 1126 * setrunnable() 1127 * 1128 * Make a process runnable. lp->lwp_token must be held on call and this 1129 * function must be called from the cpu owning lp. 1130 * 1131 * This only has an effect if we are in LSSTOP or LSSLEEP. 1132 */ 1133 void 1134 setrunnable(struct lwp *lp) 1135 { 1136 thread_t td = lp->lwp_thread; 1137 1138 ASSERT_LWKT_TOKEN_HELD(&lp->lwp_token); 1139 KKASSERT(td->td_gd == mycpu); 1140 crit_enter(); 1141 if (lp->lwp_stat == LSSTOP) 1142 lp->lwp_stat = LSSLEEP; 1143 if (lp->lwp_stat == LSSLEEP) { 1144 _tsleep_remove(td); 1145 lwkt_schedule(td); 1146 } else if (td->td_flags & TDF_SINTR) { 1147 lwkt_schedule(td); 1148 } 1149 crit_exit(); 1150 } 1151 1152 /* 1153 * The process is stopped due to some condition, usually because p_stat is 1154 * set to SSTOP, but also possibly due to being traced. 1155 * 1156 * Caller must hold p->p_token 1157 * 1158 * NOTE! If the caller sets SSTOP, the caller must also clear P_WAITED 1159 * because the parent may check the child's status before the child actually 1160 * gets to this routine. 1161 * 1162 * This routine is called with the current lwp only, typically just 1163 * before returning to userland if the process state is detected as 1164 * possibly being in a stopped state. 1165 */ 1166 void 1167 tstop(void) 1168 { 1169 struct lwp *lp = curthread->td_lwp; 1170 struct proc *p = lp->lwp_proc; 1171 struct proc *q; 1172 1173 lwkt_gettoken(&lp->lwp_token); 1174 crit_enter(); 1175 1176 /* 1177 * If LWP_MP_WSTOP is set, we were sleeping 1178 * while our process was stopped. At this point 1179 * we were already counted as stopped. 1180 */ 1181 if ((lp->lwp_mpflags & LWP_MP_WSTOP) == 0) { 1182 /* 1183 * If we're the last thread to stop, signal 1184 * our parent. 1185 */ 1186 p->p_nstopped++; 1187 atomic_set_int(&lp->lwp_mpflags, LWP_MP_WSTOP); 1188 wakeup(&p->p_nstopped); 1189 if (p->p_nstopped == p->p_nthreads) { 1190 /* 1191 * Token required to interlock kern_wait() 1192 */ 1193 q = p->p_pptr; 1194 PHOLD(q); 1195 lwkt_gettoken(&q->p_token); 1196 p->p_flags &= ~P_WAITED; 1197 wakeup(p->p_pptr); 1198 if ((q->p_sigacts->ps_flag & PS_NOCLDSTOP) == 0) 1199 ksignal(q, SIGCHLD); 1200 lwkt_reltoken(&q->p_token); 1201 PRELE(q); 1202 } 1203 } 1204 1205 /* 1206 * Wait here while in a stopped state, interlocked with lwp_token. 1207 * We must break-out if the whole process is trying to exit. 1208 */ 1209 while (STOPLWP(p, lp)) { 1210 lp->lwp_stat = LSSTOP; 1211 tsleep(p, 0, "stop", 0); 1212 } 1213 p->p_nstopped--; 1214 atomic_clear_int(&lp->lwp_mpflags, LWP_MP_WSTOP); 1215 crit_exit(); 1216 lwkt_reltoken(&lp->lwp_token); 1217 } 1218 1219 /* 1220 * Compute a tenex style load average of a quantity on 1221 * 1, 5 and 15 minute intervals. 1222 */ 1223 static int loadav_count_runnable(struct lwp *p, void *data); 1224 1225 static void 1226 loadav(void *arg) 1227 { 1228 struct loadavg *avg; 1229 int i, nrun; 1230 1231 nrun = 0; 1232 alllwp_scan(loadav_count_runnable, &nrun); 1233 avg = &averunnable; 1234 for (i = 0; i < 3; i++) { 1235 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] + 1236 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT; 1237 } 1238 1239 /* 1240 * Schedule the next update to occur after 5 seconds, but add a 1241 * random variation to avoid synchronisation with processes that 1242 * run at regular intervals. 1243 */ 1244 callout_reset(&loadav_callout, hz * 4 + (int)(krandom() % (hz * 2 + 1)), 1245 loadav, NULL); 1246 } 1247 1248 static int 1249 loadav_count_runnable(struct lwp *lp, void *data) 1250 { 1251 int *nrunp = data; 1252 thread_t td; 1253 1254 switch (lp->lwp_stat) { 1255 case LSRUN: 1256 if ((td = lp->lwp_thread) == NULL) 1257 break; 1258 if (td->td_flags & TDF_BLOCKED) 1259 break; 1260 ++*nrunp; 1261 break; 1262 default: 1263 break; 1264 } 1265 lwkt_yield(); 1266 return(0); 1267 } 1268 1269 /* 1270 * Regular data collection 1271 */ 1272 static uint64_t 1273 collect_load_callback(int n) 1274 { 1275 int fscale = averunnable.fscale; 1276 1277 return ((averunnable.ldavg[0] * 100 + (fscale >> 1)) / fscale); 1278 } 1279 1280 /* ARGSUSED */ 1281 static void 1282 sched_setup(void *dummy) 1283 { 1284 callout_init_mp(&loadav_callout); 1285 callout_init_mp(&schedcpu_callout); 1286 kcollect_register(KCOLLECT_LOAD, "load", collect_load_callback, 1287 KCOLLECT_SCALE(KCOLLECT_LOAD_FORMAT, 0)); 1288 /* Kick off timeout driven events by calling first time. */ 1289 schedcpu(NULL); 1290 loadav(NULL); 1291 } 1292 1293