1 /*- 2 * Copyright (c) 1982, 1986, 1990, 1991, 1993 3 * The Regents of the University of California. All rights reserved. 4 * (c) UNIX System Laboratories, Inc. 5 * All or some portions of this file are derived from material licensed 6 * to the University of California by American Telephone and Telegraph 7 * Co. or Unix System Laboratories, Inc. and are reproduced herein with 8 * the permission of UNIX System Laboratories, Inc. 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions 12 * are met: 13 * 1. Redistributions of source code must retain the above copyright 14 * notice, this list of conditions and the following disclaimer. 15 * 2. Redistributions in binary form must reproduce the above copyright 16 * notice, this list of conditions and the following disclaimer in the 17 * documentation and/or other materials provided with the distribution. 18 * 3. All advertising materials mentioning features or use of this software 19 * must display the following acknowledgement: 20 * This product includes software developed by the University of 21 * California, Berkeley and its contributors. 22 * 4. Neither the name of the University nor the names of its contributors 23 * may be used to endorse or promote products derived from this software 24 * without specific prior written permission. 25 * 26 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 27 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 28 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 29 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 30 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 31 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 32 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 33 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 34 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 35 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 36 * SUCH DAMAGE. 37 * 38 * @(#)kern_synch.c 8.9 (Berkeley) 5/19/95 39 * $FreeBSD: src/sys/kern/kern_synch.c,v 1.87.2.6 2002/10/13 07:29:53 kbyanc Exp $ 40 * $DragonFly: src/sys/kern/kern_synch.c,v 1.91 2008/09/09 04:06:13 dillon Exp $ 41 */ 42 43 #include "opt_ktrace.h" 44 45 #include <sys/param.h> 46 #include <sys/systm.h> 47 #include <sys/proc.h> 48 #include <sys/kernel.h> 49 #include <sys/signalvar.h> 50 #include <sys/resourcevar.h> 51 #include <sys/vmmeter.h> 52 #include <sys/sysctl.h> 53 #include <sys/lock.h> 54 #include <sys/uio.h> 55 #ifdef KTRACE 56 #include <sys/ktrace.h> 57 #endif 58 #include <sys/xwait.h> 59 #include <sys/ktr.h> 60 #include <sys/serialize.h> 61 62 #include <sys/signal2.h> 63 #include <sys/thread2.h> 64 #include <sys/spinlock2.h> 65 #include <sys/mutex2.h> 66 67 #include <machine/cpu.h> 68 #include <machine/smp.h> 69 70 TAILQ_HEAD(tslpque, thread); 71 72 static void sched_setup (void *dummy); 73 SYSINIT(sched_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, sched_setup, NULL) 74 75 int hogticks; 76 int lbolt; 77 int lbolt_syncer; 78 int sched_quantum; /* Roundrobin scheduling quantum in ticks. */ 79 int ncpus; 80 int ncpus2, ncpus2_shift, ncpus2_mask; /* note: mask not cpumask_t */ 81 int ncpus_fit, ncpus_fit_mask; /* note: mask not cpumask_t */ 82 int safepri; 83 int tsleep_now_works; 84 int tsleep_crypto_dump = 0; 85 86 static struct callout loadav_callout; 87 static struct callout schedcpu_callout; 88 MALLOC_DEFINE(M_TSLEEP, "tslpque", "tsleep queues"); 89 90 #define __DEALL(ident) __DEQUALIFY(void *, ident) 91 92 #if !defined(KTR_TSLEEP) 93 #define KTR_TSLEEP KTR_ALL 94 #endif 95 KTR_INFO_MASTER(tsleep); 96 KTR_INFO(KTR_TSLEEP, tsleep, tsleep_beg, 0, "tsleep enter %p", sizeof(void *)); 97 KTR_INFO(KTR_TSLEEP, tsleep, tsleep_end, 1, "tsleep exit", 0); 98 KTR_INFO(KTR_TSLEEP, tsleep, wakeup_beg, 2, "wakeup enter %p", sizeof(void *)); 99 KTR_INFO(KTR_TSLEEP, tsleep, wakeup_end, 3, "wakeup exit", 0); 100 KTR_INFO(KTR_TSLEEP, tsleep, ilockfail, 4, "interlock failed %p", sizeof(void *)); 101 102 #define logtsleep1(name) KTR_LOG(tsleep_ ## name) 103 #define logtsleep2(name, val) KTR_LOG(tsleep_ ## name, val) 104 105 struct loadavg averunnable = 106 { {0, 0, 0}, FSCALE }; /* load average, of runnable procs */ 107 /* 108 * Constants for averages over 1, 5, and 15 minutes 109 * when sampling at 5 second intervals. 110 */ 111 static fixpt_t cexp[3] = { 112 0.9200444146293232 * FSCALE, /* exp(-1/12) */ 113 0.9834714538216174 * FSCALE, /* exp(-1/60) */ 114 0.9944598480048967 * FSCALE, /* exp(-1/180) */ 115 }; 116 117 static void endtsleep (void *); 118 static void loadav (void *arg); 119 static void schedcpu (void *arg); 120 121 /* 122 * Adjust the scheduler quantum. The quantum is specified in microseconds. 123 * Note that 'tick' is in microseconds per tick. 124 */ 125 static int 126 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS) 127 { 128 int error, new_val; 129 130 new_val = sched_quantum * ustick; 131 error = sysctl_handle_int(oidp, &new_val, 0, req); 132 if (error != 0 || req->newptr == NULL) 133 return (error); 134 if (new_val < ustick) 135 return (EINVAL); 136 sched_quantum = new_val / ustick; 137 hogticks = 2 * sched_quantum; 138 return (0); 139 } 140 141 SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW, 142 0, sizeof sched_quantum, sysctl_kern_quantum, "I", ""); 143 144 /* 145 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the 146 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below 147 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT). 148 * 149 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used: 150 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits). 151 * 152 * If you don't want to bother with the faster/more-accurate formula, you 153 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate 154 * (more general) method of calculating the %age of CPU used by a process. 155 * 156 * decay 95% of `lwp_pctcpu' in 60 seconds; see CCPU_SHIFT before changing 157 */ 158 #define CCPU_SHIFT 11 159 160 static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */ 161 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 162 163 /* 164 * kernel uses `FSCALE', userland (SHOULD) use kern.fscale 165 */ 166 int fscale __unused = FSCALE; /* exported to systat */ 167 SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, ""); 168 169 /* 170 * Recompute process priorities, once a second. 171 * 172 * Since the userland schedulers are typically event oriented, if the 173 * estcpu calculation at wakeup() time is not sufficient to make a 174 * process runnable relative to other processes in the system we have 175 * a 1-second recalc to help out. 176 * 177 * This code also allows us to store sysclock_t data in the process structure 178 * without fear of an overrun, since sysclock_t are guarenteed to hold 179 * several seconds worth of count. 180 * 181 * WARNING! callouts can preempt normal threads. However, they will not 182 * preempt a thread holding a spinlock so we *can* safely use spinlocks. 183 */ 184 static int schedcpu_stats(struct proc *p, void *data __unused); 185 static int schedcpu_resource(struct proc *p, void *data __unused); 186 187 static void 188 schedcpu(void *arg) 189 { 190 allproc_scan(schedcpu_stats, NULL); 191 allproc_scan(schedcpu_resource, NULL); 192 wakeup((caddr_t)&lbolt); 193 wakeup((caddr_t)&lbolt_syncer); 194 callout_reset(&schedcpu_callout, hz, schedcpu, NULL); 195 } 196 197 /* 198 * General process statistics once a second 199 */ 200 static int 201 schedcpu_stats(struct proc *p, void *data __unused) 202 { 203 struct lwp *lp; 204 205 /* 206 * Threads may not be completely set up if process in SIDL state. 207 */ 208 if (p->p_stat == SIDL) 209 return(0); 210 211 PHOLD(p); 212 if (lwkt_trytoken(&p->p_token) == FALSE) { 213 PRELE(p); 214 return(0); 215 } 216 217 p->p_swtime++; 218 FOREACH_LWP_IN_PROC(lp, p) { 219 if (lp->lwp_stat == LSSLEEP) 220 lp->lwp_slptime++; 221 222 /* 223 * Only recalculate processes that are active or have slept 224 * less then 2 seconds. The schedulers understand this. 225 */ 226 if (lp->lwp_slptime <= 1) { 227 p->p_usched->recalculate(lp); 228 } else { 229 lp->lwp_pctcpu = (lp->lwp_pctcpu * ccpu) >> FSHIFT; 230 } 231 } 232 lwkt_reltoken(&p->p_token); 233 lwkt_yield(); 234 PRELE(p); 235 return(0); 236 } 237 238 /* 239 * Resource checks. XXX break out since ksignal/killproc can block, 240 * limiting us to one process killed per second. There is probably 241 * a better way. 242 */ 243 static int 244 schedcpu_resource(struct proc *p, void *data __unused) 245 { 246 u_int64_t ttime; 247 struct lwp *lp; 248 249 if (p->p_stat == SIDL) 250 return(0); 251 252 PHOLD(p); 253 if (lwkt_trytoken(&p->p_token) == FALSE) { 254 PRELE(p); 255 return(0); 256 } 257 258 if (p->p_stat == SZOMB || p->p_limit == NULL) { 259 lwkt_reltoken(&p->p_token); 260 PRELE(p); 261 return(0); 262 } 263 264 ttime = 0; 265 FOREACH_LWP_IN_PROC(lp, p) { 266 /* 267 * We may have caught an lp in the middle of being 268 * created, lwp_thread can be NULL. 269 */ 270 if (lp->lwp_thread) { 271 ttime += lp->lwp_thread->td_sticks; 272 ttime += lp->lwp_thread->td_uticks; 273 } 274 } 275 276 switch(plimit_testcpulimit(p->p_limit, ttime)) { 277 case PLIMIT_TESTCPU_KILL: 278 killproc(p, "exceeded maximum CPU limit"); 279 break; 280 case PLIMIT_TESTCPU_XCPU: 281 if ((p->p_flags & P_XCPU) == 0) { 282 p->p_flags |= P_XCPU; 283 ksignal(p, SIGXCPU); 284 } 285 break; 286 default: 287 break; 288 } 289 lwkt_reltoken(&p->p_token); 290 lwkt_yield(); 291 PRELE(p); 292 return(0); 293 } 294 295 /* 296 * This is only used by ps. Generate a cpu percentage use over 297 * a period of one second. 298 * 299 * MPSAFE 300 */ 301 void 302 updatepcpu(struct lwp *lp, int cpticks, int ttlticks) 303 { 304 fixpt_t acc; 305 int remticks; 306 307 acc = (cpticks << FSHIFT) / ttlticks; 308 if (ttlticks >= ESTCPUFREQ) { 309 lp->lwp_pctcpu = acc; 310 } else { 311 remticks = ESTCPUFREQ - ttlticks; 312 lp->lwp_pctcpu = (acc * ttlticks + lp->lwp_pctcpu * remticks) / 313 ESTCPUFREQ; 314 } 315 } 316 317 /* 318 * tsleep/wakeup hash table parameters. Try to find the sweet spot for 319 * like addresses being slept on. 320 */ 321 #define TABLESIZE 4001 322 #define LOOKUP(x) (((u_int)(uintptr_t)(x)) % TABLESIZE) 323 324 static cpumask_t slpque_cpumasks[TABLESIZE]; 325 326 /* 327 * General scheduler initialization. We force a reschedule 25 times 328 * a second by default. Note that cpu0 is initialized in early boot and 329 * cannot make any high level calls. 330 * 331 * Each cpu has its own sleep queue. 332 */ 333 void 334 sleep_gdinit(globaldata_t gd) 335 { 336 static struct tslpque slpque_cpu0[TABLESIZE]; 337 int i; 338 339 if (gd->gd_cpuid == 0) { 340 sched_quantum = (hz + 24) / 25; 341 hogticks = 2 * sched_quantum; 342 343 gd->gd_tsleep_hash = slpque_cpu0; 344 } else { 345 gd->gd_tsleep_hash = kmalloc(sizeof(slpque_cpu0), 346 M_TSLEEP, M_WAITOK | M_ZERO); 347 } 348 for (i = 0; i < TABLESIZE; ++i) 349 TAILQ_INIT(&gd->gd_tsleep_hash[i]); 350 } 351 352 /* 353 * This is a dandy function that allows us to interlock tsleep/wakeup 354 * operations with unspecified upper level locks, such as lockmgr locks, 355 * simply by holding a critical section. The sequence is: 356 * 357 * (acquire upper level lock) 358 * tsleep_interlock(blah) 359 * (release upper level lock) 360 * tsleep(blah, ...) 361 * 362 * Basically this functions queues us on the tsleep queue without actually 363 * descheduling us. When tsleep() is later called with PINTERLOCK it 364 * assumes the thread was already queued, otherwise it queues it there. 365 * 366 * Thus it is possible to receive the wakeup prior to going to sleep and 367 * the race conditions are covered. 368 */ 369 static __inline void 370 _tsleep_interlock(globaldata_t gd, const volatile void *ident, int flags) 371 { 372 thread_t td = gd->gd_curthread; 373 int id; 374 375 crit_enter_quick(td); 376 if (td->td_flags & TDF_TSLEEPQ) { 377 id = LOOKUP(td->td_wchan); 378 TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_sleepq); 379 if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL) { 380 atomic_clear_cpumask(&slpque_cpumasks[id], 381 gd->gd_cpumask); 382 } 383 } else { 384 td->td_flags |= TDF_TSLEEPQ; 385 } 386 id = LOOKUP(ident); 387 TAILQ_INSERT_TAIL(&gd->gd_tsleep_hash[id], td, td_sleepq); 388 atomic_set_cpumask(&slpque_cpumasks[id], gd->gd_cpumask); 389 td->td_wchan = ident; 390 td->td_wdomain = flags & PDOMAIN_MASK; 391 crit_exit_quick(td); 392 } 393 394 void 395 tsleep_interlock(const volatile void *ident, int flags) 396 { 397 _tsleep_interlock(mycpu, ident, flags); 398 } 399 400 /* 401 * Remove thread from sleepq. Must be called with a critical section held. 402 * The thread must not be migrating. 403 */ 404 static __inline void 405 _tsleep_remove(thread_t td) 406 { 407 globaldata_t gd = mycpu; 408 int id; 409 410 KKASSERT(td->td_gd == gd && IN_CRITICAL_SECT(td)); 411 KKASSERT((td->td_flags & TDF_MIGRATING) == 0); 412 if (td->td_flags & TDF_TSLEEPQ) { 413 td->td_flags &= ~TDF_TSLEEPQ; 414 id = LOOKUP(td->td_wchan); 415 TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_sleepq); 416 if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL) 417 atomic_clear_cpumask(&slpque_cpumasks[id], gd->gd_cpumask); 418 td->td_wchan = NULL; 419 td->td_wdomain = 0; 420 } 421 } 422 423 void 424 tsleep_remove(thread_t td) 425 { 426 _tsleep_remove(td); 427 } 428 429 /* 430 * General sleep call. Suspends the current process until a wakeup is 431 * performed on the specified identifier. The process will then be made 432 * runnable with the specified priority. Sleeps at most timo/hz seconds 433 * (0 means no timeout). If flags includes PCATCH flag, signals are checked 434 * before and after sleeping, else signals are not checked. Returns 0 if 435 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a 436 * signal needs to be delivered, ERESTART is returned if the current system 437 * call should be restarted if possible, and EINTR is returned if the system 438 * call should be interrupted by the signal (return EINTR). 439 * 440 * Note that if we are a process, we release_curproc() before messing with 441 * the LWKT scheduler. 442 * 443 * During autoconfiguration or after a panic, a sleep will simply 444 * lower the priority briefly to allow interrupts, then return. 445 * 446 * WARNING! This code can't block (short of switching away), or bad things 447 * will happen. No getting tokens, no blocking locks, etc. 448 */ 449 int 450 tsleep(const volatile void *ident, int flags, const char *wmesg, int timo) 451 { 452 struct thread *td = curthread; 453 struct lwp *lp = td->td_lwp; 454 struct proc *p = td->td_proc; /* may be NULL */ 455 globaldata_t gd; 456 int sig; 457 int catch; 458 int error; 459 int oldpri; 460 struct callout thandle; 461 462 /* 463 * NOTE: removed KTRPOINT, it could cause races due to blocking 464 * even in stable. Just scrap it for now. 465 */ 466 if (!tsleep_crypto_dump && (tsleep_now_works == 0 || panicstr)) { 467 /* 468 * After a panic, or before we actually have an operational 469 * softclock, just give interrupts a chance, then just return; 470 * 471 * don't run any other procs or panic below, 472 * in case this is the idle process and already asleep. 473 */ 474 splz(); 475 oldpri = td->td_pri; 476 lwkt_setpri_self(safepri); 477 lwkt_switch(); 478 lwkt_setpri_self(oldpri); 479 return (0); 480 } 481 logtsleep2(tsleep_beg, ident); 482 gd = td->td_gd; 483 KKASSERT(td != &gd->gd_idlethread); /* you must be kidding! */ 484 485 /* 486 * NOTE: all of this occurs on the current cpu, including any 487 * callout-based wakeups, so a critical section is a sufficient 488 * interlock. 489 * 490 * The entire sequence through to where we actually sleep must 491 * run without breaking the critical section. 492 */ 493 catch = flags & PCATCH; 494 error = 0; 495 sig = 0; 496 497 crit_enter_quick(td); 498 499 KASSERT(ident != NULL, ("tsleep: no ident")); 500 KASSERT(lp == NULL || 501 lp->lwp_stat == LSRUN || /* Obvious */ 502 lp->lwp_stat == LSSTOP, /* Set in tstop */ 503 ("tsleep %p %s %d", 504 ident, wmesg, lp->lwp_stat)); 505 506 /* 507 * We interlock the sleep queue if the caller has not already done 508 * it for us. This must be done before we potentially acquire any 509 * tokens or we can loose the wakeup. 510 */ 511 if ((flags & PINTERLOCKED) == 0) { 512 _tsleep_interlock(gd, ident, flags); 513 } 514 515 /* 516 * Setup for the current process (if this is a process). We must 517 * interlock with lwp_token to avoid remote wakeup races via 518 * setrunnable() 519 */ 520 if (lp) { 521 lwkt_gettoken(&lp->lwp_token); 522 if (catch) { 523 /* 524 * Early termination if PCATCH was set and a 525 * signal is pending, interlocked with the 526 * critical section. 527 * 528 * Early termination only occurs when tsleep() is 529 * entered while in a normal LSRUN state. 530 */ 531 if ((sig = CURSIG(lp)) != 0) 532 goto resume; 533 534 /* 535 * Causes ksignal to wake us up if a signal is 536 * received (interlocked with p->p_token). 537 */ 538 lp->lwp_flags |= LWP_SINTR; 539 } 540 } else { 541 KKASSERT(p == NULL); 542 } 543 544 /* 545 * Make sure the current process has been untangled from 546 * the userland scheduler and initialize slptime to start 547 * counting. 548 */ 549 if (lp) { 550 p->p_usched->release_curproc(lp); 551 lp->lwp_slptime = 0; 552 } 553 554 /* 555 * If the interlocked flag is set but our cpu bit in the slpqueue 556 * is no longer set, then a wakeup was processed inbetween the 557 * tsleep_interlock() (ours or the callers), and here. This can 558 * occur under numerous circumstances including when we release the 559 * current process. 560 * 561 * Extreme loads can cause the sending of an IPI (e.g. wakeup()'s) 562 * to process incoming IPIs, thus draining incoming wakeups. 563 */ 564 if ((td->td_flags & TDF_TSLEEPQ) == 0) { 565 logtsleep2(ilockfail, ident); 566 goto resume; 567 } 568 569 /* 570 * scheduling is blocked while in a critical section. Coincide 571 * the descheduled-by-tsleep flag with the descheduling of the 572 * lwkt. 573 * 574 * The timer callout is localized on our cpu and interlocked by 575 * our critical section. 576 */ 577 lwkt_deschedule_self(td); 578 td->td_flags |= TDF_TSLEEP_DESCHEDULED; 579 td->td_wmesg = wmesg; 580 581 /* 582 * Setup the timeout, if any. The timeout is only operable while 583 * the thread is flagged descheduled. 584 */ 585 KKASSERT((td->td_flags & TDF_TIMEOUT) == 0); 586 if (timo) { 587 callout_init_mp(&thandle); 588 callout_reset(&thandle, timo, endtsleep, td); 589 } 590 591 /* 592 * Beddy bye bye. 593 */ 594 if (lp) { 595 /* 596 * Ok, we are sleeping. Place us in the SSLEEP state. 597 */ 598 KKASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) == 0); 599 /* 600 * tstop() sets LSSTOP, so don't fiddle with that. 601 */ 602 if (lp->lwp_stat != LSSTOP) 603 lp->lwp_stat = LSSLEEP; 604 lp->lwp_ru.ru_nvcsw++; 605 lwkt_switch(); 606 607 /* 608 * And when we are woken up, put us back in LSRUN. If we 609 * slept for over a second, recalculate our estcpu. 610 */ 611 lp->lwp_stat = LSRUN; 612 if (lp->lwp_slptime) 613 p->p_usched->recalculate(lp); 614 lp->lwp_slptime = 0; 615 } else { 616 lwkt_switch(); 617 } 618 619 /* 620 * Make sure we haven't switched cpus while we were asleep. It's 621 * not supposed to happen. Cleanup our temporary flags. 622 */ 623 KKASSERT(gd == td->td_gd); 624 625 /* 626 * Cleanup the timeout. If the timeout has already occured thandle 627 * has already been stopped, otherwise stop thandle. If the timeout 628 * is running (the callout thread must be blocked trying to get 629 * lwp_token) then wait for us to get scheduled. 630 */ 631 if (timo) { 632 while (td->td_flags & TDF_TIMEOUT_RUNNING) { 633 lwkt_deschedule_self(td); 634 td->td_wmesg = "tsrace"; 635 lwkt_switch(); 636 kprintf("td %p %s: timeout race\n", td, td->td_comm); 637 } 638 if (td->td_flags & TDF_TIMEOUT) { 639 td->td_flags &= ~TDF_TIMEOUT; 640 error = EWOULDBLOCK; 641 } else { 642 /* does not block when on same cpu */ 643 callout_stop(&thandle); 644 } 645 } 646 td->td_flags &= ~TDF_TSLEEP_DESCHEDULED; 647 648 /* 649 * Make sure we have been removed from the sleepq. In most 650 * cases this will have been done for us already but it is 651 * possible for a scheduling IPI to be in-flight from a 652 * previous tsleep/tsleep_interlock() or due to a straight-out 653 * call to lwkt_schedule() (in the case of an interrupt thread), 654 * causing a spurious wakeup. 655 */ 656 _tsleep_remove(td); 657 td->td_wmesg = NULL; 658 659 /* 660 * Figure out the correct error return. If interrupted by a 661 * signal we want to return EINTR or ERESTART. 662 */ 663 resume: 664 if (lp) { 665 if (catch && error == 0) { 666 if (sig != 0 || (sig = CURSIG(lp))) { 667 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 668 error = EINTR; 669 else 670 error = ERESTART; 671 } 672 } 673 lp->lwp_flags &= ~LWP_SINTR; 674 lwkt_reltoken(&lp->lwp_token); 675 } 676 logtsleep1(tsleep_end); 677 crit_exit_quick(td); 678 return (error); 679 } 680 681 /* 682 * Interlocked spinlock sleep. An exclusively held spinlock must 683 * be passed to ssleep(). The function will atomically release the 684 * spinlock and tsleep on the ident, then reacquire the spinlock and 685 * return. 686 * 687 * This routine is fairly important along the critical path, so optimize it 688 * heavily. 689 */ 690 int 691 ssleep(const volatile void *ident, struct spinlock *spin, int flags, 692 const char *wmesg, int timo) 693 { 694 globaldata_t gd = mycpu; 695 int error; 696 697 _tsleep_interlock(gd, ident, flags); 698 spin_unlock_quick(gd, spin); 699 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 700 spin_lock_quick(gd, spin); 701 702 return (error); 703 } 704 705 int 706 lksleep(const volatile void *ident, struct lock *lock, int flags, 707 const char *wmesg, int timo) 708 { 709 globaldata_t gd = mycpu; 710 int error; 711 712 _tsleep_interlock(gd, ident, flags); 713 lockmgr(lock, LK_RELEASE); 714 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 715 lockmgr(lock, LK_EXCLUSIVE); 716 717 return (error); 718 } 719 720 /* 721 * Interlocked mutex sleep. An exclusively held mutex must be passed 722 * to mtxsleep(). The function will atomically release the mutex 723 * and tsleep on the ident, then reacquire the mutex and return. 724 */ 725 int 726 mtxsleep(const volatile void *ident, struct mtx *mtx, int flags, 727 const char *wmesg, int timo) 728 { 729 globaldata_t gd = mycpu; 730 int error; 731 732 _tsleep_interlock(gd, ident, flags); 733 mtx_unlock(mtx); 734 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 735 mtx_lock_ex_quick(mtx, wmesg); 736 737 return (error); 738 } 739 740 /* 741 * Interlocked serializer sleep. An exclusively held serializer must 742 * be passed to zsleep(). The function will atomically release 743 * the serializer and tsleep on the ident, then reacquire the serializer 744 * and return. 745 */ 746 int 747 zsleep(const volatile void *ident, struct lwkt_serialize *slz, int flags, 748 const char *wmesg, int timo) 749 { 750 globaldata_t gd = mycpu; 751 int ret; 752 753 ASSERT_SERIALIZED(slz); 754 755 _tsleep_interlock(gd, ident, flags); 756 lwkt_serialize_exit(slz); 757 ret = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 758 lwkt_serialize_enter(slz); 759 760 return ret; 761 } 762 763 /* 764 * Directly block on the LWKT thread by descheduling it. This 765 * is much faster then tsleep(), but the only legal way to wake 766 * us up is to directly schedule the thread. 767 * 768 * Setting TDF_SINTR will cause new signals to directly schedule us. 769 * 770 * This routine must be called while in a critical section. 771 */ 772 int 773 lwkt_sleep(const char *wmesg, int flags) 774 { 775 thread_t td = curthread; 776 int sig; 777 778 if ((flags & PCATCH) == 0 || td->td_lwp == NULL) { 779 td->td_flags |= TDF_BLOCKED; 780 td->td_wmesg = wmesg; 781 lwkt_deschedule_self(td); 782 lwkt_switch(); 783 td->td_wmesg = NULL; 784 td->td_flags &= ~TDF_BLOCKED; 785 return(0); 786 } 787 if ((sig = CURSIG(td->td_lwp)) != 0) { 788 if (SIGISMEMBER(td->td_proc->p_sigacts->ps_sigintr, sig)) 789 return(EINTR); 790 else 791 return(ERESTART); 792 793 } 794 td->td_flags |= TDF_BLOCKED | TDF_SINTR; 795 td->td_wmesg = wmesg; 796 lwkt_deschedule_self(td); 797 lwkt_switch(); 798 td->td_flags &= ~(TDF_BLOCKED | TDF_SINTR); 799 td->td_wmesg = NULL; 800 return(0); 801 } 802 803 /* 804 * Implement the timeout for tsleep. 805 * 806 * This type of callout timeout is scheduled on the same cpu the process 807 * is sleeping on. Also, at the moment, the MP lock is held. 808 */ 809 static void 810 endtsleep(void *arg) 811 { 812 thread_t td = arg; 813 struct lwp *lp; 814 815 /* 816 * We are going to have to get the lwp_token, which means we might 817 * block. This can race a tsleep getting woken up by other means 818 * so set TDF_TIMEOUT_RUNNING to force the tsleep to wait for our 819 * processing to complete (sorry tsleep!). 820 * 821 * We can safely set td_flags because td MUST be on the same cpu 822 * as we are. 823 */ 824 KKASSERT(td->td_gd == mycpu); 825 crit_enter(); 826 td->td_flags |= TDF_TIMEOUT_RUNNING | TDF_TIMEOUT; 827 828 /* 829 * This can block but TDF_TIMEOUT_RUNNING will prevent the thread 830 * from exiting the tsleep on us. The flag is interlocked by virtue 831 * of lp being on the same cpu as we are. 832 */ 833 if ((lp = td->td_lwp) != NULL) 834 lwkt_gettoken(&lp->lwp_token); 835 836 KKASSERT(td->td_flags & TDF_TSLEEP_DESCHEDULED); 837 838 if (lp) { 839 if (lp->lwp_proc->p_stat != SSTOP) 840 setrunnable(lp); 841 lwkt_reltoken(&lp->lwp_token); 842 } else { 843 _tsleep_remove(td); 844 lwkt_schedule(td); 845 } 846 KKASSERT(td->td_gd == mycpu); 847 td->td_flags &= ~TDF_TIMEOUT_RUNNING; 848 crit_exit(); 849 } 850 851 /* 852 * Make all processes sleeping on the specified identifier runnable. 853 * count may be zero or one only. 854 * 855 * The domain encodes the sleep/wakeup domain AND the first cpu to check 856 * (which is always the current cpu). As we iterate across cpus 857 * 858 * This call may run without the MP lock held. We can only manipulate thread 859 * state on the cpu owning the thread. We CANNOT manipulate process state 860 * at all. 861 * 862 * _wakeup() can be passed to an IPI so we can't use (const volatile 863 * void *ident). 864 */ 865 static void 866 _wakeup(void *ident, int domain) 867 { 868 struct tslpque *qp; 869 struct thread *td; 870 struct thread *ntd; 871 globaldata_t gd; 872 #ifdef SMP 873 cpumask_t mask; 874 #endif 875 int id; 876 877 crit_enter(); 878 logtsleep2(wakeup_beg, ident); 879 gd = mycpu; 880 id = LOOKUP(ident); 881 qp = &gd->gd_tsleep_hash[id]; 882 restart: 883 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) { 884 ntd = TAILQ_NEXT(td, td_sleepq); 885 if (td->td_wchan == ident && 886 td->td_wdomain == (domain & PDOMAIN_MASK) 887 ) { 888 KKASSERT(td->td_gd == gd); 889 _tsleep_remove(td); 890 if (td->td_flags & TDF_TSLEEP_DESCHEDULED) { 891 lwkt_schedule(td); 892 if (domain & PWAKEUP_ONE) 893 goto done; 894 } 895 goto restart; 896 } 897 } 898 899 #ifdef SMP 900 /* 901 * We finished checking the current cpu but there still may be 902 * more work to do. Either wakeup_one was requested and no matching 903 * thread was found, or a normal wakeup was requested and we have 904 * to continue checking cpus. 905 * 906 * It should be noted that this scheme is actually less expensive then 907 * the old scheme when waking up multiple threads, since we send 908 * only one IPI message per target candidate which may then schedule 909 * multiple threads. Before we could have wound up sending an IPI 910 * message for each thread on the target cpu (!= current cpu) that 911 * needed to be woken up. 912 * 913 * NOTE: Wakeups occuring on remote cpus are asynchronous. This 914 * should be ok since we are passing idents in the IPI rather then 915 * thread pointers. 916 */ 917 if ((domain & PWAKEUP_MYCPU) == 0 && 918 (mask = slpque_cpumasks[id] & gd->gd_other_cpus) != 0) { 919 lwkt_send_ipiq2_mask(mask, _wakeup, ident, 920 domain | PWAKEUP_MYCPU); 921 } 922 #endif 923 done: 924 logtsleep1(wakeup_end); 925 crit_exit(); 926 } 927 928 /* 929 * Wakeup all threads tsleep()ing on the specified ident, on all cpus 930 */ 931 void 932 wakeup(const volatile void *ident) 933 { 934 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mycpu->gd_cpuid)); 935 } 936 937 /* 938 * Wakeup one thread tsleep()ing on the specified ident, on any cpu. 939 */ 940 void 941 wakeup_one(const volatile void *ident) 942 { 943 /* XXX potentially round-robin the first responding cpu */ 944 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mycpu->gd_cpuid) | PWAKEUP_ONE); 945 } 946 947 /* 948 * Wakeup threads tsleep()ing on the specified ident on the current cpu 949 * only. 950 */ 951 void 952 wakeup_mycpu(const volatile void *ident) 953 { 954 _wakeup(__DEALL(ident), PWAKEUP_MYCPU); 955 } 956 957 /* 958 * Wakeup one thread tsleep()ing on the specified ident on the current cpu 959 * only. 960 */ 961 void 962 wakeup_mycpu_one(const volatile void *ident) 963 { 964 /* XXX potentially round-robin the first responding cpu */ 965 _wakeup(__DEALL(ident), PWAKEUP_MYCPU|PWAKEUP_ONE); 966 } 967 968 /* 969 * Wakeup all thread tsleep()ing on the specified ident on the specified cpu 970 * only. 971 */ 972 void 973 wakeup_oncpu(globaldata_t gd, const volatile void *ident) 974 { 975 #ifdef SMP 976 if (gd == mycpu) { 977 _wakeup(__DEALL(ident), PWAKEUP_MYCPU); 978 } else { 979 lwkt_send_ipiq2(gd, _wakeup, __DEALL(ident), PWAKEUP_MYCPU); 980 } 981 #else 982 _wakeup(__DEALL(ident), PWAKEUP_MYCPU); 983 #endif 984 } 985 986 /* 987 * Wakeup one thread tsleep()ing on the specified ident on the specified cpu 988 * only. 989 */ 990 void 991 wakeup_oncpu_one(globaldata_t gd, const volatile void *ident) 992 { 993 #ifdef SMP 994 if (gd == mycpu) { 995 _wakeup(__DEALL(ident), PWAKEUP_MYCPU | PWAKEUP_ONE); 996 } else { 997 lwkt_send_ipiq2(gd, _wakeup, __DEALL(ident), 998 PWAKEUP_MYCPU | PWAKEUP_ONE); 999 } 1000 #else 1001 _wakeup(__DEALL(ident), PWAKEUP_MYCPU | PWAKEUP_ONE); 1002 #endif 1003 } 1004 1005 /* 1006 * Wakeup all threads waiting on the specified ident that slept using 1007 * the specified domain, on all cpus. 1008 */ 1009 void 1010 wakeup_domain(const volatile void *ident, int domain) 1011 { 1012 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(domain, mycpu->gd_cpuid)); 1013 } 1014 1015 /* 1016 * Wakeup one thread waiting on the specified ident that slept using 1017 * the specified domain, on any cpu. 1018 */ 1019 void 1020 wakeup_domain_one(const volatile void *ident, int domain) 1021 { 1022 /* XXX potentially round-robin the first responding cpu */ 1023 _wakeup(__DEALL(ident), 1024 PWAKEUP_ENCODE(domain, mycpu->gd_cpuid) | PWAKEUP_ONE); 1025 } 1026 1027 /* 1028 * setrunnable() 1029 * 1030 * Make a process runnable. lp->lwp_token must be held on call and this 1031 * function must be called from the cpu owning lp. 1032 * 1033 * This only has an effect if we are in LSSTOP or LSSLEEP. 1034 */ 1035 void 1036 setrunnable(struct lwp *lp) 1037 { 1038 thread_t td = lp->lwp_thread; 1039 1040 ASSERT_LWKT_TOKEN_HELD(&lp->lwp_token); 1041 KKASSERT(td->td_gd == mycpu); 1042 crit_enter(); 1043 if (lp->lwp_stat == LSSTOP) 1044 lp->lwp_stat = LSSLEEP; 1045 if (lp->lwp_stat == LSSLEEP) { 1046 _tsleep_remove(td); 1047 lwkt_schedule(td); 1048 } else if (td->td_flags & TDF_SINTR) { 1049 lwkt_schedule(td); 1050 } 1051 crit_exit(); 1052 } 1053 1054 /* 1055 * The process is stopped due to some condition, usually because p_stat is 1056 * set to SSTOP, but also possibly due to being traced. 1057 * 1058 * Caller must hold p->p_token 1059 * 1060 * NOTE! If the caller sets SSTOP, the caller must also clear P_WAITED 1061 * because the parent may check the child's status before the child actually 1062 * gets to this routine. 1063 * 1064 * This routine is called with the current lwp only, typically just 1065 * before returning to userland if the process state is detected as 1066 * possibly being in a stopped state. 1067 */ 1068 void 1069 tstop(void) 1070 { 1071 struct lwp *lp = curthread->td_lwp; 1072 struct proc *p = lp->lwp_proc; 1073 struct proc *q; 1074 1075 lwkt_gettoken(&lp->lwp_token); 1076 crit_enter(); 1077 1078 /* 1079 * If LWP_MP_WSTOP is set, we were sleeping 1080 * while our process was stopped. At this point 1081 * we were already counted as stopped. 1082 */ 1083 if ((lp->lwp_mpflags & LWP_MP_WSTOP) == 0) { 1084 /* 1085 * If we're the last thread to stop, signal 1086 * our parent. 1087 */ 1088 p->p_nstopped++; 1089 atomic_set_int(&lp->lwp_mpflags, LWP_MP_WSTOP); 1090 wakeup(&p->p_nstopped); 1091 if (p->p_nstopped == p->p_nthreads) { 1092 /* 1093 * Token required to interlock kern_wait() 1094 */ 1095 q = p->p_pptr; 1096 PHOLD(q); 1097 lwkt_gettoken(&q->p_token); 1098 p->p_flags &= ~P_WAITED; 1099 wakeup(p->p_pptr); 1100 if ((q->p_sigacts->ps_flag & PS_NOCLDSTOP) == 0) 1101 ksignal(q, SIGCHLD); 1102 lwkt_reltoken(&q->p_token); 1103 PRELE(q); 1104 } 1105 } 1106 while (p->p_stat == SSTOP) { 1107 lp->lwp_stat = LSSTOP; 1108 tsleep(p, 0, "stop", 0); 1109 } 1110 p->p_nstopped--; 1111 atomic_clear_int(&lp->lwp_mpflags, LWP_MP_WSTOP); 1112 crit_exit(); 1113 lwkt_reltoken(&lp->lwp_token); 1114 } 1115 1116 /* 1117 * Compute a tenex style load average of a quantity on 1118 * 1, 5 and 15 minute intervals. 1119 */ 1120 static int loadav_count_runnable(struct lwp *p, void *data); 1121 1122 static void 1123 loadav(void *arg) 1124 { 1125 struct loadavg *avg; 1126 int i, nrun; 1127 1128 nrun = 0; 1129 alllwp_scan(loadav_count_runnable, &nrun); 1130 avg = &averunnable; 1131 for (i = 0; i < 3; i++) { 1132 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] + 1133 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT; 1134 } 1135 1136 /* 1137 * Schedule the next update to occur after 5 seconds, but add a 1138 * random variation to avoid synchronisation with processes that 1139 * run at regular intervals. 1140 */ 1141 callout_reset(&loadav_callout, hz * 4 + (int)(krandom() % (hz * 2 + 1)), 1142 loadav, NULL); 1143 } 1144 1145 static int 1146 loadav_count_runnable(struct lwp *lp, void *data) 1147 { 1148 int *nrunp = data; 1149 thread_t td; 1150 1151 switch (lp->lwp_stat) { 1152 case LSRUN: 1153 if ((td = lp->lwp_thread) == NULL) 1154 break; 1155 if (td->td_flags & TDF_BLOCKED) 1156 break; 1157 ++*nrunp; 1158 break; 1159 default: 1160 break; 1161 } 1162 lwkt_yield(); 1163 return(0); 1164 } 1165 1166 /* ARGSUSED */ 1167 static void 1168 sched_setup(void *dummy) 1169 { 1170 callout_init_mp(&loadav_callout); 1171 callout_init_mp(&schedcpu_callout); 1172 1173 /* Kick off timeout driven events by calling first time. */ 1174 schedcpu(NULL); 1175 loadav(NULL); 1176 } 1177 1178