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 #ifdef KTRACE 55 #include <sys/uio.h> 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 #include <sys/mplock2.h> 67 68 #include <machine/cpu.h> 69 #include <machine/smp.h> 70 71 TAILQ_HEAD(tslpque, thread); 72 73 static void sched_setup (void *dummy); 74 SYSINIT(sched_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, sched_setup, NULL) 75 76 int hogticks; 77 int lbolt; 78 int lbolt_syncer; 79 int sched_quantum; /* Roundrobin scheduling quantum in ticks. */ 80 int ncpus; 81 int ncpus2, ncpus2_shift, ncpus2_mask; 82 int ncpus_fit, ncpus_fit_mask; 83 int safepri; 84 int tsleep_now_works; 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 #ifdef SMP 121 static void tsleep_wakeup(struct thread *td); 122 #endif 123 124 /* 125 * Adjust the scheduler quantum. The quantum is specified in microseconds. 126 * Note that 'tick' is in microseconds per tick. 127 */ 128 static int 129 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS) 130 { 131 int error, new_val; 132 133 new_val = sched_quantum * ustick; 134 error = sysctl_handle_int(oidp, &new_val, 0, req); 135 if (error != 0 || req->newptr == NULL) 136 return (error); 137 if (new_val < ustick) 138 return (EINVAL); 139 sched_quantum = new_val / ustick; 140 hogticks = 2 * sched_quantum; 141 return (0); 142 } 143 144 SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW, 145 0, sizeof sched_quantum, sysctl_kern_quantum, "I", ""); 146 147 /* 148 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the 149 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below 150 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT). 151 * 152 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used: 153 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits). 154 * 155 * If you don't want to bother with the faster/more-accurate formula, you 156 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate 157 * (more general) method of calculating the %age of CPU used by a process. 158 * 159 * decay 95% of `lwp_pctcpu' in 60 seconds; see CCPU_SHIFT before changing 160 */ 161 #define CCPU_SHIFT 11 162 163 static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */ 164 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 165 166 /* 167 * kernel uses `FSCALE', userland (SHOULD) use kern.fscale 168 */ 169 int fscale __unused = FSCALE; /* exported to systat */ 170 SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, ""); 171 172 /* 173 * Recompute process priorities, once a second. 174 * 175 * Since the userland schedulers are typically event oriented, if the 176 * estcpu calculation at wakeup() time is not sufficient to make a 177 * process runnable relative to other processes in the system we have 178 * a 1-second recalc to help out. 179 * 180 * This code also allows us to store sysclock_t data in the process structure 181 * without fear of an overrun, since sysclock_t are guarenteed to hold 182 * several seconds worth of count. 183 * 184 * WARNING! callouts can preempt normal threads. However, they will not 185 * preempt a thread holding a spinlock so we *can* safely use spinlocks. 186 */ 187 static int schedcpu_stats(struct proc *p, void *data __unused); 188 static int schedcpu_resource(struct proc *p, void *data __unused); 189 190 static void 191 schedcpu(void *arg) 192 { 193 allproc_scan(schedcpu_stats, NULL); 194 allproc_scan(schedcpu_resource, NULL); 195 wakeup((caddr_t)&lbolt); 196 wakeup((caddr_t)&lbolt_syncer); 197 callout_reset(&schedcpu_callout, hz, schedcpu, NULL); 198 } 199 200 /* 201 * General process statistics once a second 202 */ 203 static int 204 schedcpu_stats(struct proc *p, void *data __unused) 205 { 206 struct lwp *lp; 207 208 crit_enter(); 209 p->p_swtime++; 210 FOREACH_LWP_IN_PROC(lp, p) { 211 if (lp->lwp_stat == LSSLEEP) 212 lp->lwp_slptime++; 213 214 /* 215 * Only recalculate processes that are active or have slept 216 * less then 2 seconds. The schedulers understand this. 217 */ 218 if (lp->lwp_slptime <= 1) { 219 p->p_usched->recalculate(lp); 220 } else { 221 lp->lwp_pctcpu = (lp->lwp_pctcpu * ccpu) >> FSHIFT; 222 } 223 } 224 crit_exit(); 225 return(0); 226 } 227 228 /* 229 * Resource checks. XXX break out since ksignal/killproc can block, 230 * limiting us to one process killed per second. There is probably 231 * a better way. 232 */ 233 static int 234 schedcpu_resource(struct proc *p, void *data __unused) 235 { 236 u_int64_t ttime; 237 struct lwp *lp; 238 239 crit_enter(); 240 if (p->p_stat == SIDL || 241 p->p_stat == SZOMB || 242 p->p_limit == NULL 243 ) { 244 crit_exit(); 245 return(0); 246 } 247 248 ttime = 0; 249 FOREACH_LWP_IN_PROC(lp, p) { 250 /* 251 * We may have caught an lp in the middle of being 252 * created, lwp_thread can be NULL. 253 */ 254 if (lp->lwp_thread) { 255 ttime += lp->lwp_thread->td_sticks; 256 ttime += lp->lwp_thread->td_uticks; 257 } 258 } 259 260 switch(plimit_testcpulimit(p->p_limit, ttime)) { 261 case PLIMIT_TESTCPU_KILL: 262 killproc(p, "exceeded maximum CPU limit"); 263 break; 264 case PLIMIT_TESTCPU_XCPU: 265 if ((p->p_flag & P_XCPU) == 0) { 266 p->p_flag |= P_XCPU; 267 ksignal(p, SIGXCPU); 268 } 269 break; 270 default: 271 break; 272 } 273 crit_exit(); 274 return(0); 275 } 276 277 /* 278 * This is only used by ps. Generate a cpu percentage use over 279 * a period of one second. 280 * 281 * MPSAFE 282 */ 283 void 284 updatepcpu(struct lwp *lp, int cpticks, int ttlticks) 285 { 286 fixpt_t acc; 287 int remticks; 288 289 acc = (cpticks << FSHIFT) / ttlticks; 290 if (ttlticks >= ESTCPUFREQ) { 291 lp->lwp_pctcpu = acc; 292 } else { 293 remticks = ESTCPUFREQ - ttlticks; 294 lp->lwp_pctcpu = (acc * ttlticks + lp->lwp_pctcpu * remticks) / 295 ESTCPUFREQ; 296 } 297 } 298 299 /* 300 * tsleep/wakeup hash table parameters. Try to find the sweet spot for 301 * like addresses being slept on. 302 */ 303 #define TABLESIZE 1024 304 #define LOOKUP(x) (((intptr_t)(x) >> 6) & (TABLESIZE - 1)) 305 306 static cpumask_t slpque_cpumasks[TABLESIZE]; 307 308 /* 309 * General scheduler initialization. We force a reschedule 25 times 310 * a second by default. Note that cpu0 is initialized in early boot and 311 * cannot make any high level calls. 312 * 313 * Each cpu has its own sleep queue. 314 */ 315 void 316 sleep_gdinit(globaldata_t gd) 317 { 318 static struct tslpque slpque_cpu0[TABLESIZE]; 319 int i; 320 321 if (gd->gd_cpuid == 0) { 322 sched_quantum = (hz + 24) / 25; 323 hogticks = 2 * sched_quantum; 324 325 gd->gd_tsleep_hash = slpque_cpu0; 326 } else { 327 gd->gd_tsleep_hash = kmalloc(sizeof(slpque_cpu0), 328 M_TSLEEP, M_WAITOK | M_ZERO); 329 } 330 for (i = 0; i < TABLESIZE; ++i) 331 TAILQ_INIT(&gd->gd_tsleep_hash[i]); 332 } 333 334 /* 335 * This is a dandy function that allows us to interlock tsleep/wakeup 336 * operations with unspecified upper level locks, such as lockmgr locks, 337 * simply by holding a critical section. The sequence is: 338 * 339 * (acquire upper level lock) 340 * tsleep_interlock(blah) 341 * (release upper level lock) 342 * tsleep(blah, ...) 343 * 344 * Basically this functions queues us on the tsleep queue without actually 345 * descheduling us. When tsleep() is later called with PINTERLOCK it 346 * assumes the thread was already queued, otherwise it queues it there. 347 * 348 * Thus it is possible to receive the wakeup prior to going to sleep and 349 * the race conditions are covered. 350 */ 351 static __inline void 352 _tsleep_interlock(globaldata_t gd, const volatile void *ident, int flags) 353 { 354 thread_t td = gd->gd_curthread; 355 int id; 356 357 crit_enter_quick(td); 358 if (td->td_flags & TDF_TSLEEPQ) { 359 id = LOOKUP(td->td_wchan); 360 TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_sleepq); 361 if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL) 362 atomic_clear_int(&slpque_cpumasks[id], gd->gd_cpumask); 363 } else { 364 td->td_flags |= TDF_TSLEEPQ; 365 } 366 id = LOOKUP(ident); 367 TAILQ_INSERT_TAIL(&gd->gd_tsleep_hash[id], td, td_sleepq); 368 atomic_set_int(&slpque_cpumasks[id], gd->gd_cpumask); 369 td->td_wchan = ident; 370 td->td_wdomain = flags & PDOMAIN_MASK; 371 crit_exit_quick(td); 372 } 373 374 void 375 tsleep_interlock(const volatile void *ident, int flags) 376 { 377 _tsleep_interlock(mycpu, ident, flags); 378 } 379 380 /* 381 * Remove thread from sleepq. Must be called with a critical section held. 382 */ 383 static __inline void 384 _tsleep_remove(thread_t td) 385 { 386 globaldata_t gd = mycpu; 387 int id; 388 389 KKASSERT(td->td_gd == gd); 390 if (td->td_flags & TDF_TSLEEPQ) { 391 td->td_flags &= ~TDF_TSLEEPQ; 392 id = LOOKUP(td->td_wchan); 393 TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_sleepq); 394 if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL) 395 atomic_clear_int(&slpque_cpumasks[id], gd->gd_cpumask); 396 td->td_wchan = NULL; 397 td->td_wdomain = 0; 398 } 399 } 400 401 void 402 tsleep_remove(thread_t td) 403 { 404 _tsleep_remove(td); 405 } 406 407 /* 408 * This function removes a thread from the tsleep queue and schedules 409 * it. This function may act asynchronously. The target thread may be 410 * sleeping on a different cpu. 411 * 412 * This function mus be called while in a critical section but if the 413 * target thread is sleeping on a different cpu we cannot safely probe 414 * td_flags. 415 */ 416 static __inline 417 void 418 _tsleep_wakeup(struct thread *td) 419 { 420 #ifdef SMP 421 globaldata_t gd = mycpu; 422 423 if (td->td_gd != gd) { 424 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)tsleep_wakeup, td); 425 return; 426 } 427 #endif 428 _tsleep_remove(td); 429 if (td->td_flags & TDF_TSLEEP_DESCHEDULED) { 430 td->td_flags &= ~TDF_TSLEEP_DESCHEDULED; 431 lwkt_schedule(td); 432 } 433 } 434 435 #ifdef SMP 436 static 437 void 438 tsleep_wakeup(struct thread *td) 439 { 440 _tsleep_wakeup(td); 441 } 442 #endif 443 444 445 /* 446 * General sleep call. Suspends the current process until a wakeup is 447 * performed on the specified identifier. The process will then be made 448 * runnable with the specified priority. Sleeps at most timo/hz seconds 449 * (0 means no timeout). If flags includes PCATCH flag, signals are checked 450 * before and after sleeping, else signals are not checked. Returns 0 if 451 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a 452 * signal needs to be delivered, ERESTART is returned if the current system 453 * call should be restarted if possible, and EINTR is returned if the system 454 * call should be interrupted by the signal (return EINTR). 455 * 456 * Note that if we are a process, we release_curproc() before messing with 457 * the LWKT scheduler. 458 * 459 * During autoconfiguration or after a panic, a sleep will simply 460 * lower the priority briefly to allow interrupts, then return. 461 */ 462 int 463 tsleep(const volatile void *ident, int flags, const char *wmesg, int timo) 464 { 465 struct thread *td = curthread; 466 struct lwp *lp = td->td_lwp; 467 struct proc *p = td->td_proc; /* may be NULL */ 468 globaldata_t gd; 469 int sig; 470 int catch; 471 int id; 472 int error; 473 int oldpri; 474 struct callout thandle; 475 476 /* 477 * NOTE: removed KTRPOINT, it could cause races due to blocking 478 * even in stable. Just scrap it for now. 479 */ 480 if (tsleep_now_works == 0 || panicstr) { 481 /* 482 * After a panic, or before we actually have an operational 483 * softclock, just give interrupts a chance, then just return; 484 * 485 * don't run any other procs or panic below, 486 * in case this is the idle process and already asleep. 487 */ 488 splz(); 489 oldpri = td->td_pri & TDPRI_MASK; 490 lwkt_setpri_self(safepri); 491 lwkt_switch(); 492 lwkt_setpri_self(oldpri); 493 return (0); 494 } 495 logtsleep2(tsleep_beg, ident); 496 gd = td->td_gd; 497 KKASSERT(td != &gd->gd_idlethread); /* you must be kidding! */ 498 499 /* 500 * NOTE: all of this occurs on the current cpu, including any 501 * callout-based wakeups, so a critical section is a sufficient 502 * interlock. 503 * 504 * The entire sequence through to where we actually sleep must 505 * run without breaking the critical section. 506 */ 507 catch = flags & PCATCH; 508 error = 0; 509 sig = 0; 510 511 crit_enter_quick(td); 512 513 KASSERT(ident != NULL, ("tsleep: no ident")); 514 KASSERT(lp == NULL || 515 lp->lwp_stat == LSRUN || /* Obvious */ 516 lp->lwp_stat == LSSTOP, /* Set in tstop */ 517 ("tsleep %p %s %d", 518 ident, wmesg, lp->lwp_stat)); 519 520 /* 521 * Setup for the current process (if this is a process). 522 */ 523 if (lp) { 524 if (catch) { 525 /* 526 * Early termination if PCATCH was set and a 527 * signal is pending, interlocked with the 528 * critical section. 529 * 530 * Early termination only occurs when tsleep() is 531 * entered while in a normal LSRUN state. 532 */ 533 if ((sig = CURSIG(lp)) != 0) 534 goto resume; 535 536 /* 537 * Early termination if PCATCH was set and a 538 * mailbox signal was possibly delivered prior to 539 * the system call even being made, in order to 540 * allow the user to interlock without having to 541 * make additional system calls. 542 */ 543 if (p->p_flag & P_MAILBOX) 544 goto resume; 545 546 /* 547 * Causes ksignal to wake us up when. 548 */ 549 lp->lwp_flag |= LWP_SINTR; 550 } 551 } 552 553 /* 554 * We interlock the sleep queue if the caller has not already done 555 * it for us. 556 */ 557 if ((flags & PINTERLOCKED) == 0) { 558 id = LOOKUP(ident); 559 _tsleep_interlock(gd, ident, flags); 560 } 561 562 /* 563 * 564 * If no interlock was set we do an integrated interlock here. 565 * Make sure the current process has been untangled from 566 * the userland scheduler and initialize slptime to start 567 * counting. We must interlock the sleep queue before doing 568 * this to avoid wakeup/process-ipi races which can occur under 569 * heavy loads. 570 */ 571 if (lp) { 572 p->p_usched->release_curproc(lp); 573 lp->lwp_slptime = 0; 574 } 575 576 /* 577 * If the interlocked flag is set but our cpu bit in the slpqueue 578 * is no longer set, then a wakeup was processed inbetween the 579 * tsleep_interlock() (ours or the callers), and here. This can 580 * occur under numerous circumstances including when we release the 581 * current process. 582 * 583 * Extreme loads can cause the sending of an IPI (e.g. wakeup()'s) 584 * to process incoming IPIs, thus draining incoming wakeups. 585 */ 586 if ((td->td_flags & TDF_TSLEEPQ) == 0) { 587 logtsleep2(ilockfail, ident); 588 goto resume; 589 } 590 591 /* 592 * scheduling is blocked while in a critical section. Coincide 593 * the descheduled-by-tsleep flag with the descheduling of the 594 * lwkt. 595 */ 596 lwkt_deschedule_self(td); 597 td->td_flags |= TDF_TSLEEP_DESCHEDULED; 598 td->td_wmesg = wmesg; 599 600 /* 601 * Setup the timeout, if any 602 */ 603 if (timo) { 604 callout_init(&thandle); 605 callout_reset(&thandle, timo, endtsleep, td); 606 } 607 608 /* 609 * Beddy bye bye. 610 */ 611 if (lp) { 612 /* 613 * Ok, we are sleeping. Place us in the SSLEEP state. 614 */ 615 KKASSERT((lp->lwp_flag & LWP_ONRUNQ) == 0); 616 /* 617 * tstop() sets LSSTOP, so don't fiddle with that. 618 */ 619 if (lp->lwp_stat != LSSTOP) 620 lp->lwp_stat = LSSLEEP; 621 lp->lwp_ru.ru_nvcsw++; 622 lwkt_switch(); 623 624 /* 625 * And when we are woken up, put us back in LSRUN. If we 626 * slept for over a second, recalculate our estcpu. 627 */ 628 lp->lwp_stat = LSRUN; 629 if (lp->lwp_slptime) 630 p->p_usched->recalculate(lp); 631 lp->lwp_slptime = 0; 632 } else { 633 lwkt_switch(); 634 } 635 636 /* 637 * Make sure we haven't switched cpus while we were asleep. It's 638 * not supposed to happen. Cleanup our temporary flags. 639 */ 640 KKASSERT(gd == td->td_gd); 641 642 /* 643 * Cleanup the timeout. 644 */ 645 if (timo) { 646 if (td->td_flags & TDF_TIMEOUT) { 647 td->td_flags &= ~TDF_TIMEOUT; 648 error = EWOULDBLOCK; 649 } else { 650 callout_stop(&thandle); 651 } 652 } 653 654 /* 655 * Make sure we have been removed from the sleepq. This should 656 * have been done for us already. 657 */ 658 _tsleep_remove(td); 659 td->td_wmesg = NULL; 660 if (td->td_flags & TDF_TSLEEP_DESCHEDULED) { 661 td->td_flags &= ~TDF_TSLEEP_DESCHEDULED; 662 kprintf("td %p (%s) unexpectedly rescheduled\n", 663 td, td->td_comm); 664 } 665 666 /* 667 * Figure out the correct error return. If interrupted by a 668 * signal we want to return EINTR or ERESTART. 669 * 670 * If P_MAILBOX is set no automatic system call restart occurs 671 * and we return EINTR. P_MAILBOX is meant to be used as an 672 * interlock, the user must poll it prior to any system call 673 * that it wishes to interlock a mailbox signal against since 674 * the flag is cleared on *any* system call that sleeps. 675 */ 676 resume: 677 if (p) { 678 if (catch && error == 0) { 679 if ((p->p_flag & P_MAILBOX) && sig == 0) { 680 error = EINTR; 681 } else if (sig != 0 || (sig = CURSIG(lp))) { 682 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 683 error = EINTR; 684 else 685 error = ERESTART; 686 } 687 } 688 lp->lwp_flag &= ~(LWP_BREAKTSLEEP | LWP_SINTR); 689 p->p_flag &= ~P_MAILBOX; 690 } 691 logtsleep1(tsleep_end); 692 crit_exit_quick(td); 693 return (error); 694 } 695 696 /* 697 * Interlocked spinlock sleep. An exclusively held spinlock must 698 * be passed to ssleep(). The function will atomically release the 699 * spinlock and tsleep on the ident, then reacquire the spinlock and 700 * return. 701 * 702 * This routine is fairly important along the critical path, so optimize it 703 * heavily. 704 */ 705 int 706 ssleep(const volatile void *ident, struct spinlock *spin, 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 spin_unlock_wr_quick(gd, spin); 714 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 715 spin_lock_wr_quick(gd, spin); 716 717 return (error); 718 } 719 720 int 721 lksleep(const volatile void *ident, struct lock *lock, int flags, 722 const char *wmesg, int timo) 723 { 724 globaldata_t gd = mycpu; 725 int error; 726 727 _tsleep_interlock(gd, ident, flags); 728 lockmgr(lock, LK_RELEASE); 729 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 730 lockmgr(lock, LK_EXCLUSIVE); 731 732 return (error); 733 } 734 735 /* 736 * Interlocked mutex sleep. An exclusively held mutex must be passed 737 * to mtxsleep(). The function will atomically release the mutex 738 * and tsleep on the ident, then reacquire the mutex and return. 739 */ 740 int 741 mtxsleep(const volatile void *ident, struct mtx *mtx, int flags, 742 const char *wmesg, int timo) 743 { 744 globaldata_t gd = mycpu; 745 int error; 746 747 _tsleep_interlock(gd, ident, flags); 748 mtx_unlock(mtx); 749 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 750 mtx_lock_ex_quick(mtx, wmesg); 751 752 return (error); 753 } 754 755 /* 756 * Interlocked serializer sleep. An exclusively held serializer must 757 * be passed to zsleep(). The function will atomically release 758 * the serializer and tsleep on the ident, then reacquire the serializer 759 * and return. 760 */ 761 int 762 zsleep(const volatile void *ident, struct lwkt_serialize *slz, int flags, 763 const char *wmesg, int timo) 764 { 765 globaldata_t gd = mycpu; 766 int ret; 767 768 ASSERT_SERIALIZED(slz); 769 770 _tsleep_interlock(gd, ident, flags); 771 lwkt_serialize_exit(slz); 772 ret = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 773 lwkt_serialize_enter(slz); 774 775 return ret; 776 } 777 778 /* 779 * Directly block on the LWKT thread by descheduling it. This 780 * is much faster then tsleep(), but the only legal way to wake 781 * us up is to directly schedule the thread. 782 * 783 * Setting TDF_SINTR will cause new signals to directly schedule us. 784 * 785 * This routine must be called while in a critical section. 786 */ 787 int 788 lwkt_sleep(const char *wmesg, int flags) 789 { 790 thread_t td = curthread; 791 int sig; 792 793 if ((flags & PCATCH) == 0 || td->td_lwp == NULL) { 794 td->td_flags |= TDF_BLOCKED; 795 td->td_wmesg = wmesg; 796 lwkt_deschedule_self(td); 797 lwkt_switch(); 798 td->td_wmesg = NULL; 799 td->td_flags &= ~TDF_BLOCKED; 800 return(0); 801 } 802 if ((sig = CURSIG(td->td_lwp)) != 0) { 803 if (SIGISMEMBER(td->td_proc->p_sigacts->ps_sigintr, sig)) 804 return(EINTR); 805 else 806 return(ERESTART); 807 808 } 809 td->td_flags |= TDF_BLOCKED | TDF_SINTR; 810 td->td_wmesg = wmesg; 811 lwkt_deschedule_self(td); 812 lwkt_switch(); 813 td->td_flags &= ~(TDF_BLOCKED | TDF_SINTR); 814 td->td_wmesg = NULL; 815 return(0); 816 } 817 818 /* 819 * Implement the timeout for tsleep. 820 * 821 * We set LWP_BREAKTSLEEP to indicate that an event has occured, but 822 * we only call setrunnable if the process is not stopped. 823 * 824 * This type of callout timeout is scheduled on the same cpu the process 825 * is sleeping on. Also, at the moment, the MP lock is held. 826 */ 827 static void 828 endtsleep(void *arg) 829 { 830 thread_t td = arg; 831 struct lwp *lp; 832 833 ASSERT_MP_LOCK_HELD(curthread); 834 crit_enter(); 835 836 /* 837 * cpu interlock. Thread flags are only manipulated on 838 * the cpu owning the thread. proc flags are only manipulated 839 * by the older of the MP lock. We have both. 840 */ 841 if (td->td_flags & TDF_TSLEEP_DESCHEDULED) { 842 td->td_flags |= TDF_TIMEOUT; 843 844 if ((lp = td->td_lwp) != NULL) { 845 lp->lwp_flag |= LWP_BREAKTSLEEP; 846 if (lp->lwp_proc->p_stat != SSTOP) 847 setrunnable(lp); 848 } else { 849 _tsleep_wakeup(td); 850 } 851 } 852 crit_exit(); 853 } 854 855 /* 856 * Make all processes sleeping on the specified identifier runnable. 857 * count may be zero or one only. 858 * 859 * The domain encodes the sleep/wakeup domain AND the first cpu to check 860 * (which is always the current cpu). As we iterate across cpus 861 * 862 * This call may run without the MP lock held. We can only manipulate thread 863 * state on the cpu owning the thread. We CANNOT manipulate process state 864 * at all. 865 * 866 * _wakeup() can be passed to an IPI so we can't use (const volatile 867 * void *ident). 868 */ 869 static void 870 _wakeup(void *ident, int domain) 871 { 872 struct tslpque *qp; 873 struct thread *td; 874 struct thread *ntd; 875 globaldata_t gd; 876 #ifdef SMP 877 cpumask_t mask; 878 #endif 879 int id; 880 881 crit_enter(); 882 logtsleep2(wakeup_beg, ident); 883 gd = mycpu; 884 id = LOOKUP(ident); 885 qp = &gd->gd_tsleep_hash[id]; 886 restart: 887 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) { 888 ntd = TAILQ_NEXT(td, td_sleepq); 889 if (td->td_wchan == ident && 890 td->td_wdomain == (domain & PDOMAIN_MASK) 891 ) { 892 KKASSERT(td->td_gd == gd); 893 _tsleep_remove(td); 894 if (td->td_flags & TDF_TSLEEP_DESCHEDULED) { 895 td->td_flags &= ~TDF_TSLEEP_DESCHEDULED; 896 lwkt_schedule(td); 897 if (domain & PWAKEUP_ONE) 898 goto done; 899 } 900 goto restart; 901 } 902 } 903 904 #ifdef SMP 905 /* 906 * We finished checking the current cpu but there still may be 907 * more work to do. Either wakeup_one was requested and no matching 908 * thread was found, or a normal wakeup was requested and we have 909 * to continue checking cpus. 910 * 911 * It should be noted that this scheme is actually less expensive then 912 * the old scheme when waking up multiple threads, since we send 913 * only one IPI message per target candidate which may then schedule 914 * multiple threads. Before we could have wound up sending an IPI 915 * message for each thread on the target cpu (!= current cpu) that 916 * needed to be woken up. 917 * 918 * NOTE: Wakeups occuring on remote cpus are asynchronous. This 919 * should be ok since we are passing idents in the IPI rather then 920 * thread pointers. 921 */ 922 if ((domain & PWAKEUP_MYCPU) == 0 && 923 (mask = slpque_cpumasks[id] & gd->gd_other_cpus) != 0) { 924 lwkt_send_ipiq2_mask(mask, _wakeup, ident, 925 domain | PWAKEUP_MYCPU); 926 } 927 #endif 928 done: 929 logtsleep1(wakeup_end); 930 crit_exit(); 931 } 932 933 /* 934 * Wakeup all threads tsleep()ing on the specified ident, on all cpus 935 */ 936 void 937 wakeup(const volatile void *ident) 938 { 939 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mycpu->gd_cpuid)); 940 } 941 942 /* 943 * Wakeup one thread tsleep()ing on the specified ident, on any cpu. 944 */ 945 void 946 wakeup_one(const volatile void *ident) 947 { 948 /* XXX potentially round-robin the first responding cpu */ 949 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mycpu->gd_cpuid) | PWAKEUP_ONE); 950 } 951 952 /* 953 * Wakeup threads tsleep()ing on the specified ident on the current cpu 954 * only. 955 */ 956 void 957 wakeup_mycpu(const volatile void *ident) 958 { 959 _wakeup(__DEALL(ident), PWAKEUP_MYCPU); 960 } 961 962 /* 963 * Wakeup one thread tsleep()ing on the specified ident on the current cpu 964 * only. 965 */ 966 void 967 wakeup_mycpu_one(const volatile void *ident) 968 { 969 /* XXX potentially round-robin the first responding cpu */ 970 _wakeup(__DEALL(ident), PWAKEUP_MYCPU|PWAKEUP_ONE); 971 } 972 973 /* 974 * Wakeup all thread tsleep()ing on the specified ident on the specified cpu 975 * only. 976 */ 977 void 978 wakeup_oncpu(globaldata_t gd, const volatile void *ident) 979 { 980 #ifdef SMP 981 if (gd == mycpu) { 982 _wakeup(__DEALL(ident), PWAKEUP_MYCPU); 983 } else { 984 lwkt_send_ipiq2(gd, _wakeup, __DEALL(ident), PWAKEUP_MYCPU); 985 } 986 #else 987 _wakeup(__DEALL(ident), PWAKEUP_MYCPU); 988 #endif 989 } 990 991 /* 992 * Wakeup one thread tsleep()ing on the specified ident on the specified cpu 993 * only. 994 */ 995 void 996 wakeup_oncpu_one(globaldata_t gd, const volatile void *ident) 997 { 998 #ifdef SMP 999 if (gd == mycpu) { 1000 _wakeup(__DEALL(ident), PWAKEUP_MYCPU | PWAKEUP_ONE); 1001 } else { 1002 lwkt_send_ipiq2(gd, _wakeup, __DEALL(ident), 1003 PWAKEUP_MYCPU | PWAKEUP_ONE); 1004 } 1005 #else 1006 _wakeup(__DEALL(ident), PWAKEUP_MYCPU | PWAKEUP_ONE); 1007 #endif 1008 } 1009 1010 /* 1011 * Wakeup all threads waiting on the specified ident that slept using 1012 * the specified domain, on all cpus. 1013 */ 1014 void 1015 wakeup_domain(const volatile void *ident, int domain) 1016 { 1017 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(domain, mycpu->gd_cpuid)); 1018 } 1019 1020 /* 1021 * Wakeup one thread waiting on the specified ident that slept using 1022 * the specified domain, on any cpu. 1023 */ 1024 void 1025 wakeup_domain_one(const volatile void *ident, int domain) 1026 { 1027 /* XXX potentially round-robin the first responding cpu */ 1028 _wakeup(__DEALL(ident), 1029 PWAKEUP_ENCODE(domain, mycpu->gd_cpuid) | PWAKEUP_ONE); 1030 } 1031 1032 /* 1033 * setrunnable() 1034 * 1035 * Make a process runnable. The MP lock must be held on call. This only 1036 * has an effect if we are in SSLEEP. We only break out of the 1037 * tsleep if LWP_BREAKTSLEEP is set, otherwise we just fix-up the state. 1038 * 1039 * NOTE: With the MP lock held we can only safely manipulate the process 1040 * structure. We cannot safely manipulate the thread structure. 1041 */ 1042 void 1043 setrunnable(struct lwp *lp) 1044 { 1045 crit_enter(); 1046 ASSERT_MP_LOCK_HELD(curthread); 1047 if (lp->lwp_stat == LSSTOP) 1048 lp->lwp_stat = LSSLEEP; 1049 if (lp->lwp_stat == LSSLEEP && (lp->lwp_flag & LWP_BREAKTSLEEP)) 1050 _tsleep_wakeup(lp->lwp_thread); 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 * NOTE! If the caller sets SSTOP, the caller must also clear P_WAITED 1059 * because the parent may check the child's status before the child actually 1060 * gets to this routine. 1061 * 1062 * This routine is called with the current lwp only, typically just 1063 * before returning to userland. 1064 * 1065 * Setting LWP_BREAKTSLEEP before entering the tsleep will cause a passive 1066 * SIGCONT to break out of the tsleep. 1067 */ 1068 void 1069 tstop(void) 1070 { 1071 struct lwp *lp = curthread->td_lwp; 1072 struct proc *p = lp->lwp_proc; 1073 1074 crit_enter(); 1075 /* 1076 * If LWP_WSTOP is set, we were sleeping 1077 * while our process was stopped. At this point 1078 * we were already counted as stopped. 1079 */ 1080 if ((lp->lwp_flag & LWP_WSTOP) == 0) { 1081 /* 1082 * If we're the last thread to stop, signal 1083 * our parent. 1084 */ 1085 p->p_nstopped++; 1086 lp->lwp_flag |= LWP_WSTOP; 1087 wakeup(&p->p_nstopped); 1088 if (p->p_nstopped == p->p_nthreads) { 1089 p->p_flag &= ~P_WAITED; 1090 wakeup(p->p_pptr); 1091 if ((p->p_pptr->p_sigacts->ps_flag & PS_NOCLDSTOP) == 0) 1092 ksignal(p->p_pptr, SIGCHLD); 1093 } 1094 } 1095 while (p->p_stat == SSTOP) { 1096 lp->lwp_flag |= LWP_BREAKTSLEEP; 1097 lp->lwp_stat = LSSTOP; 1098 tsleep(p, 0, "stop", 0); 1099 } 1100 p->p_nstopped--; 1101 lp->lwp_flag &= ~LWP_WSTOP; 1102 crit_exit(); 1103 } 1104 1105 /* 1106 * Yield / synchronous reschedule. This is a bit tricky because the trap 1107 * code might have set a lazy release on the switch function. Setting 1108 * P_PASSIVE_ACQ will ensure that the lazy release executes when we call 1109 * switch, and that we are given a greater chance of affinity with our 1110 * current cpu. 1111 * 1112 * We call lwkt_setpri_self() to rotate our thread to the end of the lwkt 1113 * run queue. lwkt_switch() will also execute any assigned passive release 1114 * (which usually calls release_curproc()), allowing a same/higher priority 1115 * process to be designated as the current process. 1116 * 1117 * While it is possible for a lower priority process to be designated, 1118 * it's call to lwkt_maybe_switch() in acquire_curproc() will likely 1119 * round-robin back to us and we will be able to re-acquire the current 1120 * process designation. 1121 * 1122 * MPSAFE 1123 */ 1124 void 1125 uio_yield(void) 1126 { 1127 struct thread *td = curthread; 1128 struct proc *p = td->td_proc; 1129 1130 lwkt_setpri_self(td->td_pri & TDPRI_MASK); 1131 if (p) { 1132 p->p_flag |= P_PASSIVE_ACQ; 1133 lwkt_switch(); 1134 p->p_flag &= ~P_PASSIVE_ACQ; 1135 } else { 1136 lwkt_switch(); 1137 } 1138 } 1139 1140 /* 1141 * Compute a tenex style load average of a quantity on 1142 * 1, 5 and 15 minute intervals. 1143 */ 1144 static int loadav_count_runnable(struct lwp *p, void *data); 1145 1146 static void 1147 loadav(void *arg) 1148 { 1149 struct loadavg *avg; 1150 int i, nrun; 1151 1152 nrun = 0; 1153 alllwp_scan(loadav_count_runnable, &nrun); 1154 avg = &averunnable; 1155 for (i = 0; i < 3; i++) { 1156 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] + 1157 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT; 1158 } 1159 1160 /* 1161 * Schedule the next update to occur after 5 seconds, but add a 1162 * random variation to avoid synchronisation with processes that 1163 * run at regular intervals. 1164 */ 1165 callout_reset(&loadav_callout, hz * 4 + (int)(krandom() % (hz * 2 + 1)), 1166 loadav, NULL); 1167 } 1168 1169 static int 1170 loadav_count_runnable(struct lwp *lp, void *data) 1171 { 1172 int *nrunp = data; 1173 thread_t td; 1174 1175 switch (lp->lwp_stat) { 1176 case LSRUN: 1177 if ((td = lp->lwp_thread) == NULL) 1178 break; 1179 if (td->td_flags & TDF_BLOCKED) 1180 break; 1181 ++*nrunp; 1182 break; 1183 default: 1184 break; 1185 } 1186 return(0); 1187 } 1188 1189 /* ARGSUSED */ 1190 static void 1191 sched_setup(void *dummy) 1192 { 1193 callout_init(&loadav_callout); 1194 callout_init(&schedcpu_callout); 1195 1196 /* Kick off timeout driven events by calling first time. */ 1197 schedcpu(NULL); 1198 loadav(NULL); 1199 } 1200 1201