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/signal2.h> 51 #include <sys/resourcevar.h> 52 #include <sys/vmmeter.h> 53 #include <sys/sysctl.h> 54 #include <sys/lock.h> 55 #ifdef KTRACE 56 #include <sys/uio.h> 57 #include <sys/ktrace.h> 58 #endif 59 #include <sys/xwait.h> 60 #include <sys/ktr.h> 61 62 #include <sys/thread2.h> 63 #include <sys/spinlock2.h> 64 #include <sys/serialize.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 int lbolt_syncer; 77 int sched_quantum; /* Roundrobin scheduling quantum in ticks. */ 78 int ncpus; 79 int ncpus2, ncpus2_shift, ncpus2_mask; 80 int ncpus_fit, ncpus_fit_mask; 81 int safepri; 82 int tsleep_now_works; 83 84 static struct callout loadav_callout; 85 static struct callout schedcpu_callout; 86 MALLOC_DEFINE(M_TSLEEP, "tslpque", "tsleep queues"); 87 88 #if !defined(KTR_TSLEEP) 89 #define KTR_TSLEEP KTR_ALL 90 #endif 91 KTR_INFO_MASTER(tsleep); 92 KTR_INFO(KTR_TSLEEP, tsleep, tsleep_beg, 0, "tsleep enter %p", sizeof(void *)); 93 KTR_INFO(KTR_TSLEEP, tsleep, tsleep_end, 1, "tsleep exit", 0); 94 KTR_INFO(KTR_TSLEEP, tsleep, wakeup_beg, 2, "wakeup enter %p", sizeof(void *)); 95 KTR_INFO(KTR_TSLEEP, tsleep, wakeup_end, 3, "wakeup exit", 0); 96 KTR_INFO(KTR_TSLEEP, tsleep, ilockfail, 4, "interlock failed %p", sizeof(void *)); 97 98 #define logtsleep1(name) KTR_LOG(tsleep_ ## name) 99 #define logtsleep2(name, val) KTR_LOG(tsleep_ ## name, val) 100 101 struct loadavg averunnable = 102 { {0, 0, 0}, FSCALE }; /* load average, of runnable procs */ 103 /* 104 * Constants for averages over 1, 5, and 15 minutes 105 * when sampling at 5 second intervals. 106 */ 107 static fixpt_t cexp[3] = { 108 0.9200444146293232 * FSCALE, /* exp(-1/12) */ 109 0.9834714538216174 * FSCALE, /* exp(-1/60) */ 110 0.9944598480048967 * FSCALE, /* exp(-1/180) */ 111 }; 112 113 static void endtsleep (void *); 114 static void tsleep_wakeup(struct thread *td); 115 static void loadav (void *arg); 116 static void schedcpu (void *arg); 117 118 /* 119 * Adjust the scheduler quantum. The quantum is specified in microseconds. 120 * Note that 'tick' is in microseconds per tick. 121 */ 122 static int 123 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS) 124 { 125 int error, new_val; 126 127 new_val = sched_quantum * tick; 128 error = sysctl_handle_int(oidp, &new_val, 0, req); 129 if (error != 0 || req->newptr == NULL) 130 return (error); 131 if (new_val < tick) 132 return (EINVAL); 133 sched_quantum = new_val / tick; 134 hogticks = 2 * sched_quantum; 135 return (0); 136 } 137 138 SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW, 139 0, sizeof sched_quantum, sysctl_kern_quantum, "I", ""); 140 141 /* 142 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the 143 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below 144 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT). 145 * 146 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used: 147 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits). 148 * 149 * If you don't want to bother with the faster/more-accurate formula, you 150 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate 151 * (more general) method of calculating the %age of CPU used by a process. 152 * 153 * decay 95% of `lwp_pctcpu' in 60 seconds; see CCPU_SHIFT before changing 154 */ 155 #define CCPU_SHIFT 11 156 157 static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */ 158 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 159 160 /* 161 * kernel uses `FSCALE', userland (SHOULD) use kern.fscale 162 */ 163 int fscale __unused = FSCALE; /* exported to systat */ 164 SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, ""); 165 166 /* 167 * Recompute process priorities, once a second. 168 * 169 * Since the userland schedulers are typically event oriented, if the 170 * estcpu calculation at wakeup() time is not sufficient to make a 171 * process runnable relative to other processes in the system we have 172 * a 1-second recalc to help out. 173 * 174 * This code also allows us to store sysclock_t data in the process structure 175 * without fear of an overrun, since sysclock_t are guarenteed to hold 176 * several seconds worth of count. 177 * 178 * WARNING! callouts can preempt normal threads. However, they will not 179 * preempt a thread holding a spinlock so we *can* safely use spinlocks. 180 */ 181 static int schedcpu_stats(struct proc *p, void *data __unused); 182 static int schedcpu_resource(struct proc *p, void *data __unused); 183 184 static void 185 schedcpu(void *arg) 186 { 187 allproc_scan(schedcpu_stats, NULL); 188 allproc_scan(schedcpu_resource, NULL); 189 wakeup((caddr_t)&lbolt); 190 wakeup((caddr_t)&lbolt_syncer); 191 callout_reset(&schedcpu_callout, hz, schedcpu, NULL); 192 } 193 194 /* 195 * General process statistics once a second 196 */ 197 static int 198 schedcpu_stats(struct proc *p, void *data __unused) 199 { 200 struct lwp *lp; 201 202 crit_enter(); 203 p->p_swtime++; 204 FOREACH_LWP_IN_PROC(lp, p) { 205 if (lp->lwp_stat == LSSLEEP) 206 lp->lwp_slptime++; 207 208 /* 209 * Only recalculate processes that are active or have slept 210 * less then 2 seconds. The schedulers understand this. 211 */ 212 if (lp->lwp_slptime <= 1) { 213 p->p_usched->recalculate(lp); 214 } else { 215 lp->lwp_pctcpu = (lp->lwp_pctcpu * ccpu) >> FSHIFT; 216 } 217 } 218 crit_exit(); 219 return(0); 220 } 221 222 /* 223 * Resource checks. XXX break out since ksignal/killproc can block, 224 * limiting us to one process killed per second. There is probably 225 * a better way. 226 */ 227 static int 228 schedcpu_resource(struct proc *p, void *data __unused) 229 { 230 u_int64_t ttime; 231 struct lwp *lp; 232 233 crit_enter(); 234 if (p->p_stat == SIDL || 235 p->p_stat == SZOMB || 236 p->p_limit == NULL 237 ) { 238 crit_exit(); 239 return(0); 240 } 241 242 ttime = 0; 243 FOREACH_LWP_IN_PROC(lp, p) { 244 /* 245 * We may have caught an lp in the middle of being 246 * created, lwp_thread can be NULL. 247 */ 248 if (lp->lwp_thread) { 249 ttime += lp->lwp_thread->td_sticks; 250 ttime += lp->lwp_thread->td_uticks; 251 } 252 } 253 254 switch(plimit_testcpulimit(p->p_limit, ttime)) { 255 case PLIMIT_TESTCPU_KILL: 256 killproc(p, "exceeded maximum CPU limit"); 257 break; 258 case PLIMIT_TESTCPU_XCPU: 259 if ((p->p_flag & P_XCPU) == 0) { 260 p->p_flag |= P_XCPU; 261 ksignal(p, SIGXCPU); 262 } 263 break; 264 default: 265 break; 266 } 267 crit_exit(); 268 return(0); 269 } 270 271 /* 272 * This is only used by ps. Generate a cpu percentage use over 273 * a period of one second. 274 * 275 * MPSAFE 276 */ 277 void 278 updatepcpu(struct lwp *lp, int cpticks, int ttlticks) 279 { 280 fixpt_t acc; 281 int remticks; 282 283 acc = (cpticks << FSHIFT) / ttlticks; 284 if (ttlticks >= ESTCPUFREQ) { 285 lp->lwp_pctcpu = acc; 286 } else { 287 remticks = ESTCPUFREQ - ttlticks; 288 lp->lwp_pctcpu = (acc * ttlticks + lp->lwp_pctcpu * remticks) / 289 ESTCPUFREQ; 290 } 291 } 292 293 /* 294 * tsleep/wakeup hash table parameters. Try to find the sweet spot for 295 * like addresses being slept on. 296 */ 297 #define TABLESIZE 1024 298 #define LOOKUP(x) (((intptr_t)(x) >> 6) & (TABLESIZE - 1)) 299 300 static cpumask_t slpque_cpumasks[TABLESIZE]; 301 302 /* 303 * General scheduler initialization. We force a reschedule 25 times 304 * a second by default. Note that cpu0 is initialized in early boot and 305 * cannot make any high level calls. 306 * 307 * Each cpu has its own sleep queue. 308 */ 309 void 310 sleep_gdinit(globaldata_t gd) 311 { 312 static struct tslpque slpque_cpu0[TABLESIZE]; 313 int i; 314 315 if (gd->gd_cpuid == 0) { 316 sched_quantum = (hz + 24) / 25; 317 hogticks = 2 * sched_quantum; 318 319 gd->gd_tsleep_hash = slpque_cpu0; 320 } else { 321 gd->gd_tsleep_hash = kmalloc(sizeof(slpque_cpu0), 322 M_TSLEEP, M_WAITOK | M_ZERO); 323 } 324 for (i = 0; i < TABLESIZE; ++i) 325 TAILQ_INIT(&gd->gd_tsleep_hash[i]); 326 } 327 328 /* 329 * This is a dandy function that allows us to interlock tsleep/wakeup 330 * operations with unspecified upper level locks, such as lockmgr locks, 331 * simply by holding a critical section. The sequence is: 332 * 333 * (acquire upper level lock) 334 * tsleep_interlock(blah) 335 * (release upper level lock) 336 * tsleep(blah, ...) 337 * 338 * Basically this functions queues us on the tsleep queue without actually 339 * descheduling us. When tsleep() is later called with PINTERLOCK it 340 * assumes the thread was already queued, otherwise it queues it there. 341 * 342 * Thus it is possible to receive the wakeup prior to going to sleep and 343 * the race conditions are covered. 344 */ 345 static __inline void 346 _tsleep_interlock(globaldata_t gd, void *ident, int flags) 347 { 348 thread_t td = gd->gd_curthread; 349 int id; 350 351 crit_enter_quick(td); 352 if (td->td_flags & TDF_TSLEEPQ) { 353 id = LOOKUP(td->td_wchan); 354 TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_sleepq); 355 if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL) 356 atomic_clear_int(&slpque_cpumasks[id], gd->gd_cpumask); 357 } else { 358 td->td_flags |= TDF_TSLEEPQ; 359 } 360 id = LOOKUP(ident); 361 TAILQ_INSERT_TAIL(&gd->gd_tsleep_hash[id], td, td_sleepq); 362 atomic_set_int(&slpque_cpumasks[id], gd->gd_cpumask); 363 td->td_wchan = ident; 364 td->td_wdomain = flags & PDOMAIN_MASK; 365 atomic_set_int(&slpque_cpumasks[id], gd->gd_cpumask); 366 crit_exit_quick(td); 367 } 368 369 void 370 tsleep_interlock(void *ident, int flags) 371 { 372 _tsleep_interlock(mycpu, ident, flags); 373 } 374 375 /* 376 * Remove thread from sleepq. Must be called with a critical section held. 377 */ 378 static __inline void 379 _tsleep_remove(thread_t td) 380 { 381 globaldata_t gd = mycpu; 382 int id; 383 384 KKASSERT(td->td_gd == gd); 385 if (td->td_flags & TDF_TSLEEPQ) { 386 td->td_flags &= ~TDF_TSLEEPQ; 387 id = LOOKUP(td->td_wchan); 388 TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_sleepq); 389 if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL) 390 atomic_clear_int(&slpque_cpumasks[id], gd->gd_cpumask); 391 td->td_wchan = NULL; 392 td->td_wdomain = 0; 393 } 394 } 395 396 void 397 tsleep_remove(thread_t td) 398 { 399 _tsleep_remove(td); 400 } 401 402 /* 403 * This function removes a thread from the tsleep queue and schedules 404 * it. This function may act asynchronously. The target thread may be 405 * sleeping on a different cpu. 406 * 407 * This function mus be called while in a critical section but if the 408 * target thread is sleeping on a different cpu we cannot safely probe 409 * td_flags. 410 */ 411 static __inline 412 void 413 _tsleep_wakeup(struct thread *td) 414 { 415 globaldata_t gd = mycpu; 416 417 #ifdef SMP 418 if (td->td_gd != gd) { 419 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)tsleep_wakeup, td); 420 return; 421 } 422 #endif 423 _tsleep_remove(td); 424 if (td->td_flags & TDF_TSLEEP_DESCHEDULED) { 425 td->td_flags &= ~TDF_TSLEEP_DESCHEDULED; 426 lwkt_schedule(td); 427 } 428 } 429 430 static 431 void 432 tsleep_wakeup(struct thread *td) 433 { 434 _tsleep_wakeup(td); 435 } 436 437 438 /* 439 * General sleep call. Suspends the current process until a wakeup is 440 * performed on the specified identifier. The process will then be made 441 * runnable with the specified priority. Sleeps at most timo/hz seconds 442 * (0 means no timeout). If flags includes PCATCH flag, signals are checked 443 * before and after sleeping, else signals are not checked. Returns 0 if 444 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a 445 * signal needs to be delivered, ERESTART is returned if the current system 446 * call should be restarted if possible, and EINTR is returned if the system 447 * call should be interrupted by the signal (return EINTR). 448 * 449 * Note that if we are a process, we release_curproc() before messing with 450 * the LWKT scheduler. 451 * 452 * During autoconfiguration or after a panic, a sleep will simply 453 * lower the priority briefly to allow interrupts, then return. 454 */ 455 int 456 tsleep(void *ident, int flags, const char *wmesg, int timo) 457 { 458 struct thread *td = curthread; 459 struct lwp *lp = td->td_lwp; 460 struct proc *p = td->td_proc; /* may be NULL */ 461 globaldata_t gd; 462 int sig; 463 int catch; 464 int id; 465 int error; 466 int oldpri; 467 struct callout thandle; 468 469 /* 470 * NOTE: removed KTRPOINT, it could cause races due to blocking 471 * even in stable. Just scrap it for now. 472 */ 473 if (tsleep_now_works == 0 || panicstr) { 474 /* 475 * After a panic, or before we actually have an operational 476 * softclock, just give interrupts a chance, then just return; 477 * 478 * don't run any other procs or panic below, 479 * in case this is the idle process and already asleep. 480 */ 481 splz(); 482 oldpri = td->td_pri & TDPRI_MASK; 483 lwkt_setpri_self(safepri); 484 lwkt_switch(); 485 lwkt_setpri_self(oldpri); 486 return (0); 487 } 488 logtsleep2(tsleep_beg, ident); 489 gd = td->td_gd; 490 KKASSERT(td != &gd->gd_idlethread); /* you must be kidding! */ 491 492 /* 493 * NOTE: all of this occurs on the current cpu, including any 494 * callout-based wakeups, so a critical section is a sufficient 495 * interlock. 496 * 497 * The entire sequence through to where we actually sleep must 498 * run without breaking the critical section. 499 */ 500 catch = flags & PCATCH; 501 error = 0; 502 sig = 0; 503 504 crit_enter_quick(td); 505 506 KASSERT(ident != NULL, ("tsleep: no ident")); 507 KASSERT(lp == NULL || 508 lp->lwp_stat == LSRUN || /* Obvious */ 509 lp->lwp_stat == LSSTOP, /* Set in tstop */ 510 ("tsleep %p %s %d", 511 ident, wmesg, lp->lwp_stat)); 512 513 /* 514 * Setup for the current process (if this is a process). 515 */ 516 if (lp) { 517 if (catch) { 518 /* 519 * Early termination if PCATCH was set and a 520 * signal is pending, interlocked with the 521 * critical section. 522 * 523 * Early termination only occurs when tsleep() is 524 * entered while in a normal LSRUN state. 525 */ 526 if ((sig = CURSIG(lp)) != 0) 527 goto resume; 528 529 /* 530 * Early termination if PCATCH was set and a 531 * mailbox signal was possibly delivered prior to 532 * the system call even being made, in order to 533 * allow the user to interlock without having to 534 * make additional system calls. 535 */ 536 if (p->p_flag & P_MAILBOX) 537 goto resume; 538 539 /* 540 * Causes ksignal to wake us up when. 541 */ 542 lp->lwp_flag |= LWP_SINTR; 543 } 544 545 /* 546 * Make sure the current process has been untangled from 547 * the userland scheduler and initialize slptime to start 548 * counting. 549 */ 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() and here. This can occur under extreme loads 558 * if the IPIQ fills up and gets processed synchronously by, say, 559 * a wakeup() or other IPI sent inbetween the interlock and here. 560 * 561 * Even the usched->release function just above can muff it up. 562 */ 563 if (flags & PINTERLOCKED) { 564 if ((td->td_flags & TDF_TSLEEPQ) == 0) { 565 logtsleep2(ilockfail, ident); 566 goto resume; 567 } 568 } else { 569 id = LOOKUP(ident); 570 _tsleep_interlock(gd, ident, flags); 571 } 572 lwkt_deschedule_self(td); 573 td->td_flags |= TDF_TSLEEP_DESCHEDULED; 574 td->td_wmesg = wmesg; 575 576 /* 577 * Setup the timeout, if any 578 */ 579 if (timo) { 580 callout_init(&thandle); 581 callout_reset(&thandle, timo, endtsleep, td); 582 } 583 584 /* 585 * Beddy bye bye. 586 */ 587 if (lp) { 588 /* 589 * Ok, we are sleeping. Place us in the SSLEEP state. 590 */ 591 KKASSERT((lp->lwp_flag & LWP_ONRUNQ) == 0); 592 /* 593 * tstop() sets LSSTOP, so don't fiddle with that. 594 */ 595 if (lp->lwp_stat != LSSTOP) 596 lp->lwp_stat = LSSLEEP; 597 lp->lwp_ru.ru_nvcsw++; 598 lwkt_switch(); 599 600 /* 601 * And when we are woken up, put us back in LSRUN. If we 602 * slept for over a second, recalculate our estcpu. 603 */ 604 lp->lwp_stat = LSRUN; 605 if (lp->lwp_slptime) 606 p->p_usched->recalculate(lp); 607 lp->lwp_slptime = 0; 608 } else { 609 lwkt_switch(); 610 } 611 612 /* 613 * Make sure we haven't switched cpus while we were asleep. It's 614 * not supposed to happen. Cleanup our temporary flags. 615 */ 616 KKASSERT(gd == td->td_gd); 617 618 /* 619 * Cleanup the timeout. 620 */ 621 if (timo) { 622 if (td->td_flags & TDF_TIMEOUT) { 623 td->td_flags &= ~TDF_TIMEOUT; 624 error = EWOULDBLOCK; 625 } else { 626 callout_stop(&thandle); 627 } 628 } 629 630 /* 631 * Make sure we have been removed from the sleepq. This should 632 * have been done for us already. 633 */ 634 _tsleep_remove(td); 635 td->td_wmesg = NULL; 636 if (td->td_flags & TDF_TSLEEP_DESCHEDULED) { 637 td->td_flags &= ~TDF_TSLEEP_DESCHEDULED; 638 kprintf("td %p (%s) unexpectedly rescheduled\n", 639 td, td->td_comm); 640 } 641 642 /* 643 * Figure out the correct error return. If interrupted by a 644 * signal we want to return EINTR or ERESTART. 645 * 646 * If P_MAILBOX is set no automatic system call restart occurs 647 * and we return EINTR. P_MAILBOX is meant to be used as an 648 * interlock, the user must poll it prior to any system call 649 * that it wishes to interlock a mailbox signal against since 650 * the flag is cleared on *any* system call that sleeps. 651 */ 652 resume: 653 if (p) { 654 if (catch && error == 0) { 655 if ((p->p_flag & P_MAILBOX) && sig == 0) { 656 error = EINTR; 657 } else if (sig != 0 || (sig = CURSIG(lp))) { 658 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 659 error = EINTR; 660 else 661 error = ERESTART; 662 } 663 } 664 lp->lwp_flag &= ~(LWP_BREAKTSLEEP | LWP_SINTR); 665 p->p_flag &= ~P_MAILBOX; 666 } 667 logtsleep1(tsleep_end); 668 crit_exit_quick(td); 669 return (error); 670 } 671 672 /* 673 * Interlocked spinlock sleep. An exclusively held spinlock must 674 * be passed to msleep(). The function will atomically release the 675 * spinlock and tsleep on the ident, then reacquire the spinlock and 676 * return. 677 * 678 * This routine is fairly important along the critical path, so optimize it 679 * heavily. 680 */ 681 int 682 msleep(void *ident, struct spinlock *spin, int flags, 683 const char *wmesg, int timo) 684 { 685 globaldata_t gd = mycpu; 686 int error; 687 688 _tsleep_interlock(gd, ident, flags); 689 spin_unlock_wr_quick(gd, spin); 690 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 691 spin_lock_wr_quick(gd, spin); 692 693 return (error); 694 } 695 696 /* 697 * Interlocked serializer sleep. An exclusively held serializer must 698 * be passed to serialize_sleep(). The function will atomically release 699 * the serializer and tsleep on the ident, then reacquire the serializer 700 * and return. 701 */ 702 int 703 serialize_sleep(void *ident, struct lwkt_serialize *slz, int flags, 704 const char *wmesg, int timo) 705 { 706 globaldata_t gd = mycpu; 707 int ret; 708 709 ASSERT_SERIALIZED(slz); 710 711 _tsleep_interlock(gd, ident, flags); 712 lwkt_serialize_exit(slz); 713 ret = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 714 lwkt_serialize_enter(slz); 715 716 return ret; 717 } 718 719 /* 720 * Directly block on the LWKT thread by descheduling it. This 721 * is much faster then tsleep(), but the only legal way to wake 722 * us up is to directly schedule the thread. 723 * 724 * Setting TDF_SINTR will cause new signals to directly schedule us. 725 * 726 * This routine must be called while in a critical section. 727 */ 728 int 729 lwkt_sleep(const char *wmesg, int flags) 730 { 731 thread_t td = curthread; 732 int sig; 733 734 if ((flags & PCATCH) == 0 || td->td_lwp == NULL) { 735 td->td_flags |= TDF_BLOCKED; 736 td->td_wmesg = wmesg; 737 lwkt_deschedule_self(td); 738 lwkt_switch(); 739 td->td_wmesg = NULL; 740 td->td_flags &= ~TDF_BLOCKED; 741 return(0); 742 } 743 if ((sig = CURSIG(td->td_lwp)) != 0) { 744 if (SIGISMEMBER(td->td_proc->p_sigacts->ps_sigintr, sig)) 745 return(EINTR); 746 else 747 return(ERESTART); 748 749 } 750 td->td_flags |= TDF_BLOCKED | TDF_SINTR; 751 td->td_wmesg = wmesg; 752 lwkt_deschedule_self(td); 753 lwkt_switch(); 754 td->td_flags &= ~(TDF_BLOCKED | TDF_SINTR); 755 td->td_wmesg = NULL; 756 return(0); 757 } 758 759 /* 760 * Implement the timeout for tsleep. 761 * 762 * We set LWP_BREAKTSLEEP to indicate that an event has occured, but 763 * we only call setrunnable if the process is not stopped. 764 * 765 * This type of callout timeout is scheduled on the same cpu the process 766 * is sleeping on. Also, at the moment, the MP lock is held. 767 */ 768 static void 769 endtsleep(void *arg) 770 { 771 thread_t td = arg; 772 struct lwp *lp; 773 774 ASSERT_MP_LOCK_HELD(curthread); 775 crit_enter(); 776 777 /* 778 * cpu interlock. Thread flags are only manipulated on 779 * the cpu owning the thread. proc flags are only manipulated 780 * by the older of the MP lock. We have both. 781 */ 782 if (td->td_flags & TDF_TSLEEP_DESCHEDULED) { 783 td->td_flags |= TDF_TIMEOUT; 784 785 if ((lp = td->td_lwp) != NULL) { 786 lp->lwp_flag |= LWP_BREAKTSLEEP; 787 if (lp->lwp_proc->p_stat != SSTOP) 788 setrunnable(lp); 789 } else { 790 _tsleep_wakeup(td); 791 } 792 } 793 crit_exit(); 794 } 795 796 /* 797 * Make all processes sleeping on the specified identifier runnable. 798 * count may be zero or one only. 799 * 800 * The domain encodes the sleep/wakeup domain AND the first cpu to check 801 * (which is always the current cpu). As we iterate across cpus 802 * 803 * This call may run without the MP lock held. We can only manipulate thread 804 * state on the cpu owning the thread. We CANNOT manipulate process state 805 * at all. 806 */ 807 static void 808 _wakeup(void *ident, int domain) 809 { 810 struct tslpque *qp; 811 struct thread *td; 812 struct thread *ntd; 813 globaldata_t gd; 814 #ifdef SMP 815 cpumask_t mask; 816 #endif 817 int id; 818 819 crit_enter(); 820 logtsleep2(wakeup_beg, ident); 821 gd = mycpu; 822 id = LOOKUP(ident); 823 qp = &gd->gd_tsleep_hash[id]; 824 restart: 825 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) { 826 ntd = TAILQ_NEXT(td, td_sleepq); 827 if (td->td_wchan == ident && 828 td->td_wdomain == (domain & PDOMAIN_MASK) 829 ) { 830 KKASSERT(td->td_gd == gd); 831 _tsleep_remove(td); 832 if (td->td_flags & TDF_TSLEEP_DESCHEDULED) { 833 td->td_flags &= ~TDF_TSLEEP_DESCHEDULED; 834 lwkt_schedule(td); 835 if (domain & PWAKEUP_ONE) 836 goto done; 837 } 838 goto restart; 839 } 840 } 841 842 #ifdef SMP 843 /* 844 * We finished checking the current cpu but there still may be 845 * more work to do. Either wakeup_one was requested and no matching 846 * thread was found, or a normal wakeup was requested and we have 847 * to continue checking cpus. 848 * 849 * It should be noted that this scheme is actually less expensive then 850 * the old scheme when waking up multiple threads, since we send 851 * only one IPI message per target candidate which may then schedule 852 * multiple threads. Before we could have wound up sending an IPI 853 * message for each thread on the target cpu (!= current cpu) that 854 * needed to be woken up. 855 * 856 * NOTE: Wakeups occuring on remote cpus are asynchronous. This 857 * should be ok since we are passing idents in the IPI rather then 858 * thread pointers. 859 */ 860 if ((domain & PWAKEUP_MYCPU) == 0 && 861 (mask = slpque_cpumasks[id] & gd->gd_other_cpus) != 0) { 862 lwkt_send_ipiq2_mask(mask, _wakeup, ident, 863 domain | PWAKEUP_MYCPU); 864 } 865 #endif 866 done: 867 logtsleep1(wakeup_end); 868 crit_exit(); 869 } 870 871 /* 872 * Wakeup all threads tsleep()ing on the specified ident, on all cpus 873 */ 874 void 875 wakeup(void *ident) 876 { 877 _wakeup(ident, PWAKEUP_ENCODE(0, mycpu->gd_cpuid)); 878 } 879 880 /* 881 * Wakeup one thread tsleep()ing on the specified ident, on any cpu. 882 */ 883 void 884 wakeup_one(void *ident) 885 { 886 /* XXX potentially round-robin the first responding cpu */ 887 _wakeup(ident, PWAKEUP_ENCODE(0, mycpu->gd_cpuid) | PWAKEUP_ONE); 888 } 889 890 /* 891 * Wakeup threads tsleep()ing on the specified ident on the current cpu 892 * only. 893 */ 894 void 895 wakeup_mycpu(void *ident) 896 { 897 _wakeup(ident, PWAKEUP_MYCPU); 898 } 899 900 /* 901 * Wakeup one thread tsleep()ing on the specified ident on the current cpu 902 * only. 903 */ 904 void 905 wakeup_mycpu_one(void *ident) 906 { 907 /* XXX potentially round-robin the first responding cpu */ 908 _wakeup(ident, PWAKEUP_MYCPU|PWAKEUP_ONE); 909 } 910 911 /* 912 * Wakeup all thread tsleep()ing on the specified ident on the specified cpu 913 * only. 914 */ 915 void 916 wakeup_oncpu(globaldata_t gd, void *ident) 917 { 918 #ifdef SMP 919 if (gd == mycpu) { 920 _wakeup(ident, PWAKEUP_MYCPU); 921 } else { 922 lwkt_send_ipiq2(gd, _wakeup, ident, PWAKEUP_MYCPU); 923 } 924 #else 925 _wakeup(ident, PWAKEUP_MYCPU); 926 #endif 927 } 928 929 /* 930 * Wakeup one thread tsleep()ing on the specified ident on the specified cpu 931 * only. 932 */ 933 void 934 wakeup_oncpu_one(globaldata_t gd, void *ident) 935 { 936 #ifdef SMP 937 if (gd == mycpu) { 938 _wakeup(ident, PWAKEUP_MYCPU | PWAKEUP_ONE); 939 } else { 940 lwkt_send_ipiq2(gd, _wakeup, ident, PWAKEUP_MYCPU | PWAKEUP_ONE); 941 } 942 #else 943 _wakeup(ident, PWAKEUP_MYCPU | PWAKEUP_ONE); 944 #endif 945 } 946 947 /* 948 * Wakeup all threads waiting on the specified ident that slept using 949 * the specified domain, on all cpus. 950 */ 951 void 952 wakeup_domain(void *ident, int domain) 953 { 954 _wakeup(ident, PWAKEUP_ENCODE(domain, mycpu->gd_cpuid)); 955 } 956 957 /* 958 * Wakeup one thread waiting on the specified ident that slept using 959 * the specified domain, on any cpu. 960 */ 961 void 962 wakeup_domain_one(void *ident, int domain) 963 { 964 /* XXX potentially round-robin the first responding cpu */ 965 _wakeup(ident, PWAKEUP_ENCODE(domain, mycpu->gd_cpuid) | PWAKEUP_ONE); 966 } 967 968 /* 969 * setrunnable() 970 * 971 * Make a process runnable. The MP lock must be held on call. This only 972 * has an effect if we are in SSLEEP. We only break out of the 973 * tsleep if LWP_BREAKTSLEEP is set, otherwise we just fix-up the state. 974 * 975 * NOTE: With the MP lock held we can only safely manipulate the process 976 * structure. We cannot safely manipulate the thread structure. 977 */ 978 void 979 setrunnable(struct lwp *lp) 980 { 981 crit_enter(); 982 ASSERT_MP_LOCK_HELD(curthread); 983 if (lp->lwp_stat == LSSTOP) 984 lp->lwp_stat = LSSLEEP; 985 if (lp->lwp_stat == LSSLEEP && (lp->lwp_flag & LWP_BREAKTSLEEP)) 986 _tsleep_wakeup(lp->lwp_thread); 987 crit_exit(); 988 } 989 990 /* 991 * The process is stopped due to some condition, usually because p_stat is 992 * set to SSTOP, but also possibly due to being traced. 993 * 994 * NOTE! If the caller sets SSTOP, the caller must also clear P_WAITED 995 * because the parent may check the child's status before the child actually 996 * gets to this routine. 997 * 998 * This routine is called with the current lwp only, typically just 999 * before returning to userland. 1000 * 1001 * Setting LWP_BREAKTSLEEP before entering the tsleep will cause a passive 1002 * SIGCONT to break out of the tsleep. 1003 */ 1004 void 1005 tstop(void) 1006 { 1007 struct lwp *lp = curthread->td_lwp; 1008 struct proc *p = lp->lwp_proc; 1009 1010 crit_enter(); 1011 /* 1012 * If LWP_WSTOP is set, we were sleeping 1013 * while our process was stopped. At this point 1014 * we were already counted as stopped. 1015 */ 1016 if ((lp->lwp_flag & LWP_WSTOP) == 0) { 1017 /* 1018 * If we're the last thread to stop, signal 1019 * our parent. 1020 */ 1021 p->p_nstopped++; 1022 lp->lwp_flag |= LWP_WSTOP; 1023 wakeup(&p->p_nstopped); 1024 if (p->p_nstopped == p->p_nthreads) { 1025 p->p_flag &= ~P_WAITED; 1026 wakeup(p->p_pptr); 1027 if ((p->p_pptr->p_sigacts->ps_flag & PS_NOCLDSTOP) == 0) 1028 ksignal(p->p_pptr, SIGCHLD); 1029 } 1030 } 1031 while (p->p_stat == SSTOP) { 1032 lp->lwp_flag |= LWP_BREAKTSLEEP; 1033 lp->lwp_stat = LSSTOP; 1034 tsleep(p, 0, "stop", 0); 1035 } 1036 p->p_nstopped--; 1037 lp->lwp_flag &= ~LWP_WSTOP; 1038 crit_exit(); 1039 } 1040 1041 /* 1042 * Yield / synchronous reschedule. This is a bit tricky because the trap 1043 * code might have set a lazy release on the switch function. Setting 1044 * P_PASSIVE_ACQ will ensure that the lazy release executes when we call 1045 * switch, and that we are given a greater chance of affinity with our 1046 * current cpu. 1047 * 1048 * We call lwkt_setpri_self() to rotate our thread to the end of the lwkt 1049 * run queue. lwkt_switch() will also execute any assigned passive release 1050 * (which usually calls release_curproc()), allowing a same/higher priority 1051 * process to be designated as the current process. 1052 * 1053 * While it is possible for a lower priority process to be designated, 1054 * it's call to lwkt_maybe_switch() in acquire_curproc() will likely 1055 * round-robin back to us and we will be able to re-acquire the current 1056 * process designation. 1057 */ 1058 void 1059 uio_yield(void) 1060 { 1061 struct thread *td = curthread; 1062 struct proc *p = td->td_proc; 1063 1064 lwkt_setpri_self(td->td_pri & TDPRI_MASK); 1065 if (p) { 1066 p->p_flag |= P_PASSIVE_ACQ; 1067 lwkt_switch(); 1068 p->p_flag &= ~P_PASSIVE_ACQ; 1069 } else { 1070 lwkt_switch(); 1071 } 1072 } 1073 1074 /* 1075 * Compute a tenex style load average of a quantity on 1076 * 1, 5 and 15 minute intervals. 1077 */ 1078 static int loadav_count_runnable(struct lwp *p, void *data); 1079 1080 static void 1081 loadav(void *arg) 1082 { 1083 struct loadavg *avg; 1084 int i, nrun; 1085 1086 nrun = 0; 1087 alllwp_scan(loadav_count_runnable, &nrun); 1088 avg = &averunnable; 1089 for (i = 0; i < 3; i++) { 1090 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] + 1091 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT; 1092 } 1093 1094 /* 1095 * Schedule the next update to occur after 5 seconds, but add a 1096 * random variation to avoid synchronisation with processes that 1097 * run at regular intervals. 1098 */ 1099 callout_reset(&loadav_callout, hz * 4 + (int)(krandom() % (hz * 2 + 1)), 1100 loadav, NULL); 1101 } 1102 1103 static int 1104 loadav_count_runnable(struct lwp *lp, void *data) 1105 { 1106 int *nrunp = data; 1107 thread_t td; 1108 1109 switch (lp->lwp_stat) { 1110 case LSRUN: 1111 if ((td = lp->lwp_thread) == NULL) 1112 break; 1113 if (td->td_flags & TDF_BLOCKED) 1114 break; 1115 ++*nrunp; 1116 break; 1117 default: 1118 break; 1119 } 1120 return(0); 1121 } 1122 1123 /* ARGSUSED */ 1124 static void 1125 sched_setup(void *dummy) 1126 { 1127 callout_init(&loadav_callout); 1128 callout_init(&schedcpu_callout); 1129 1130 /* Kick off timeout driven events by calling first time. */ 1131 schedcpu(NULL); 1132 loadav(NULL); 1133 } 1134 1135