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/mutex2.h> 65 #include <sys/serialize.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; 81 int ncpus_fit, ncpus_fit_mask; 82 int safepri; 83 int tsleep_now_works; 84 85 static struct callout loadav_callout; 86 static struct callout schedcpu_callout; 87 MALLOC_DEFINE(M_TSLEEP, "tslpque", "tsleep queues"); 88 89 #if !defined(KTR_TSLEEP) 90 #define KTR_TSLEEP KTR_ALL 91 #endif 92 KTR_INFO_MASTER(tsleep); 93 KTR_INFO(KTR_TSLEEP, tsleep, tsleep_beg, 0, "tsleep enter %p", sizeof(void *)); 94 KTR_INFO(KTR_TSLEEP, tsleep, tsleep_end, 1, "tsleep exit", 0); 95 KTR_INFO(KTR_TSLEEP, tsleep, wakeup_beg, 2, "wakeup enter %p", sizeof(void *)); 96 KTR_INFO(KTR_TSLEEP, tsleep, wakeup_end, 3, "wakeup exit", 0); 97 KTR_INFO(KTR_TSLEEP, tsleep, ilockfail, 4, "interlock failed %p", sizeof(void *)); 98 99 #define logtsleep1(name) KTR_LOG(tsleep_ ## name) 100 #define logtsleep2(name, val) KTR_LOG(tsleep_ ## name, val) 101 102 struct loadavg averunnable = 103 { {0, 0, 0}, FSCALE }; /* load average, of runnable procs */ 104 /* 105 * Constants for averages over 1, 5, and 15 minutes 106 * when sampling at 5 second intervals. 107 */ 108 static fixpt_t cexp[3] = { 109 0.9200444146293232 * FSCALE, /* exp(-1/12) */ 110 0.9834714538216174 * FSCALE, /* exp(-1/60) */ 111 0.9944598480048967 * FSCALE, /* exp(-1/180) */ 112 }; 113 114 static void endtsleep (void *); 115 static void loadav (void *arg); 116 static void schedcpu (void *arg); 117 #ifdef SMP 118 static void tsleep_wakeup(struct thread *td); 119 #endif 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 * tick; 131 error = sysctl_handle_int(oidp, &new_val, 0, req); 132 if (error != 0 || req->newptr == NULL) 133 return (error); 134 if (new_val < tick) 135 return (EINVAL); 136 sched_quantum = new_val / tick; 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 crit_enter(); 206 p->p_swtime++; 207 FOREACH_LWP_IN_PROC(lp, p) { 208 if (lp->lwp_stat == LSSLEEP) 209 lp->lwp_slptime++; 210 211 /* 212 * Only recalculate processes that are active or have slept 213 * less then 2 seconds. The schedulers understand this. 214 */ 215 if (lp->lwp_slptime <= 1) { 216 p->p_usched->recalculate(lp); 217 } else { 218 lp->lwp_pctcpu = (lp->lwp_pctcpu * ccpu) >> FSHIFT; 219 } 220 } 221 crit_exit(); 222 return(0); 223 } 224 225 /* 226 * Resource checks. XXX break out since ksignal/killproc can block, 227 * limiting us to one process killed per second. There is probably 228 * a better way. 229 */ 230 static int 231 schedcpu_resource(struct proc *p, void *data __unused) 232 { 233 u_int64_t ttime; 234 struct lwp *lp; 235 236 crit_enter(); 237 if (p->p_stat == SIDL || 238 p->p_stat == SZOMB || 239 p->p_limit == NULL 240 ) { 241 crit_exit(); 242 return(0); 243 } 244 245 ttime = 0; 246 FOREACH_LWP_IN_PROC(lp, p) { 247 /* 248 * We may have caught an lp in the middle of being 249 * created, lwp_thread can be NULL. 250 */ 251 if (lp->lwp_thread) { 252 ttime += lp->lwp_thread->td_sticks; 253 ttime += lp->lwp_thread->td_uticks; 254 } 255 } 256 257 switch(plimit_testcpulimit(p->p_limit, ttime)) { 258 case PLIMIT_TESTCPU_KILL: 259 killproc(p, "exceeded maximum CPU limit"); 260 break; 261 case PLIMIT_TESTCPU_XCPU: 262 if ((p->p_flag & P_XCPU) == 0) { 263 p->p_flag |= P_XCPU; 264 ksignal(p, SIGXCPU); 265 } 266 break; 267 default: 268 break; 269 } 270 crit_exit(); 271 return(0); 272 } 273 274 /* 275 * This is only used by ps. Generate a cpu percentage use over 276 * a period of one second. 277 * 278 * MPSAFE 279 */ 280 void 281 updatepcpu(struct lwp *lp, int cpticks, int ttlticks) 282 { 283 fixpt_t acc; 284 int remticks; 285 286 acc = (cpticks << FSHIFT) / ttlticks; 287 if (ttlticks >= ESTCPUFREQ) { 288 lp->lwp_pctcpu = acc; 289 } else { 290 remticks = ESTCPUFREQ - ttlticks; 291 lp->lwp_pctcpu = (acc * ttlticks + lp->lwp_pctcpu * remticks) / 292 ESTCPUFREQ; 293 } 294 } 295 296 /* 297 * tsleep/wakeup hash table parameters. Try to find the sweet spot for 298 * like addresses being slept on. 299 */ 300 #define TABLESIZE 1024 301 #define LOOKUP(x) (((intptr_t)(x) >> 6) & (TABLESIZE - 1)) 302 303 static cpumask_t slpque_cpumasks[TABLESIZE]; 304 305 /* 306 * General scheduler initialization. We force a reschedule 25 times 307 * a second by default. Note that cpu0 is initialized in early boot and 308 * cannot make any high level calls. 309 * 310 * Each cpu has its own sleep queue. 311 */ 312 void 313 sleep_gdinit(globaldata_t gd) 314 { 315 static struct tslpque slpque_cpu0[TABLESIZE]; 316 int i; 317 318 if (gd->gd_cpuid == 0) { 319 sched_quantum = (hz + 24) / 25; 320 hogticks = 2 * sched_quantum; 321 322 gd->gd_tsleep_hash = slpque_cpu0; 323 } else { 324 gd->gd_tsleep_hash = kmalloc(sizeof(slpque_cpu0), 325 M_TSLEEP, M_WAITOK | M_ZERO); 326 } 327 for (i = 0; i < TABLESIZE; ++i) 328 TAILQ_INIT(&gd->gd_tsleep_hash[i]); 329 } 330 331 /* 332 * This is a dandy function that allows us to interlock tsleep/wakeup 333 * operations with unspecified upper level locks, such as lockmgr locks, 334 * simply by holding a critical section. The sequence is: 335 * 336 * (acquire upper level lock) 337 * tsleep_interlock(blah) 338 * (release upper level lock) 339 * tsleep(blah, ...) 340 * 341 * Basically this functions queues us on the tsleep queue without actually 342 * descheduling us. When tsleep() is later called with PINTERLOCK it 343 * assumes the thread was already queued, otherwise it queues it there. 344 * 345 * Thus it is possible to receive the wakeup prior to going to sleep and 346 * the race conditions are covered. 347 */ 348 static __inline void 349 _tsleep_interlock(globaldata_t gd, void *ident, int flags) 350 { 351 thread_t td = gd->gd_curthread; 352 int id; 353 354 crit_enter_quick(td); 355 if (td->td_flags & TDF_TSLEEPQ) { 356 id = LOOKUP(td->td_wchan); 357 TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_sleepq); 358 if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL) 359 atomic_clear_int(&slpque_cpumasks[id], gd->gd_cpumask); 360 } else { 361 td->td_flags |= TDF_TSLEEPQ; 362 } 363 id = LOOKUP(ident); 364 TAILQ_INSERT_TAIL(&gd->gd_tsleep_hash[id], td, td_sleepq); 365 atomic_set_int(&slpque_cpumasks[id], gd->gd_cpumask); 366 td->td_wchan = ident; 367 td->td_wdomain = flags & PDOMAIN_MASK; 368 crit_exit_quick(td); 369 } 370 371 void 372 tsleep_interlock(void *ident, int flags) 373 { 374 _tsleep_interlock(mycpu, ident, flags); 375 } 376 377 /* 378 * Remove thread from sleepq. Must be called with a critical section held. 379 */ 380 static __inline void 381 _tsleep_remove(thread_t td) 382 { 383 globaldata_t gd = mycpu; 384 int id; 385 386 KKASSERT(td->td_gd == gd); 387 if (td->td_flags & TDF_TSLEEPQ) { 388 td->td_flags &= ~TDF_TSLEEPQ; 389 id = LOOKUP(td->td_wchan); 390 TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_sleepq); 391 if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL) 392 atomic_clear_int(&slpque_cpumasks[id], gd->gd_cpumask); 393 td->td_wchan = NULL; 394 td->td_wdomain = 0; 395 } 396 } 397 398 void 399 tsleep_remove(thread_t td) 400 { 401 _tsleep_remove(td); 402 } 403 404 /* 405 * This function removes a thread from the tsleep queue and schedules 406 * it. This function may act asynchronously. The target thread may be 407 * sleeping on a different cpu. 408 * 409 * This function mus be called while in a critical section but if the 410 * target thread is sleeping on a different cpu we cannot safely probe 411 * td_flags. 412 */ 413 static __inline 414 void 415 _tsleep_wakeup(struct thread *td) 416 { 417 #ifdef SMP 418 globaldata_t gd = mycpu; 419 420 if (td->td_gd != gd) { 421 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)tsleep_wakeup, td); 422 return; 423 } 424 #endif 425 _tsleep_remove(td); 426 if (td->td_flags & TDF_TSLEEP_DESCHEDULED) { 427 td->td_flags &= ~TDF_TSLEEP_DESCHEDULED; 428 lwkt_schedule(td); 429 } 430 } 431 432 #ifdef SMP 433 static 434 void 435 tsleep_wakeup(struct thread *td) 436 { 437 _tsleep_wakeup(td); 438 } 439 #endif 440 441 442 /* 443 * General sleep call. Suspends the current process until a wakeup is 444 * performed on the specified identifier. The process will then be made 445 * runnable with the specified priority. Sleeps at most timo/hz seconds 446 * (0 means no timeout). If flags includes PCATCH flag, signals are checked 447 * before and after sleeping, else signals are not checked. Returns 0 if 448 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a 449 * signal needs to be delivered, ERESTART is returned if the current system 450 * call should be restarted if possible, and EINTR is returned if the system 451 * call should be interrupted by the signal (return EINTR). 452 * 453 * Note that if we are a process, we release_curproc() before messing with 454 * the LWKT scheduler. 455 * 456 * During autoconfiguration or after a panic, a sleep will simply 457 * lower the priority briefly to allow interrupts, then return. 458 */ 459 int 460 tsleep(void *ident, int flags, const char *wmesg, int timo) 461 { 462 struct thread *td = curthread; 463 struct lwp *lp = td->td_lwp; 464 struct proc *p = td->td_proc; /* may be NULL */ 465 globaldata_t gd; 466 int sig; 467 int catch; 468 int id; 469 int error; 470 int oldpri; 471 struct callout thandle; 472 473 /* 474 * NOTE: removed KTRPOINT, it could cause races due to blocking 475 * even in stable. Just scrap it for now. 476 */ 477 if (tsleep_now_works == 0 || panicstr) { 478 /* 479 * After a panic, or before we actually have an operational 480 * softclock, just give interrupts a chance, then just return; 481 * 482 * don't run any other procs or panic below, 483 * in case this is the idle process and already asleep. 484 */ 485 splz(); 486 oldpri = td->td_pri & TDPRI_MASK; 487 lwkt_setpri_self(safepri); 488 lwkt_switch(); 489 lwkt_setpri_self(oldpri); 490 return (0); 491 } 492 logtsleep2(tsleep_beg, ident); 493 gd = td->td_gd; 494 KKASSERT(td != &gd->gd_idlethread); /* you must be kidding! */ 495 496 /* 497 * NOTE: all of this occurs on the current cpu, including any 498 * callout-based wakeups, so a critical section is a sufficient 499 * interlock. 500 * 501 * The entire sequence through to where we actually sleep must 502 * run without breaking the critical section. 503 */ 504 catch = flags & PCATCH; 505 error = 0; 506 sig = 0; 507 508 crit_enter_quick(td); 509 510 KASSERT(ident != NULL, ("tsleep: no ident")); 511 KASSERT(lp == NULL || 512 lp->lwp_stat == LSRUN || /* Obvious */ 513 lp->lwp_stat == LSSTOP, /* Set in tstop */ 514 ("tsleep %p %s %d", 515 ident, wmesg, lp->lwp_stat)); 516 517 /* 518 * Setup for the current process (if this is a process). 519 */ 520 if (lp) { 521 if (catch) { 522 /* 523 * Early termination if PCATCH was set and a 524 * signal is pending, interlocked with the 525 * critical section. 526 * 527 * Early termination only occurs when tsleep() is 528 * entered while in a normal LSRUN state. 529 */ 530 if ((sig = CURSIG(lp)) != 0) 531 goto resume; 532 533 /* 534 * Early termination if PCATCH was set and a 535 * mailbox signal was possibly delivered prior to 536 * the system call even being made, in order to 537 * allow the user to interlock without having to 538 * make additional system calls. 539 */ 540 if (p->p_flag & P_MAILBOX) 541 goto resume; 542 543 /* 544 * Causes ksignal to wake us up when. 545 */ 546 lp->lwp_flag |= LWP_SINTR; 547 } 548 } 549 550 /* 551 * We interlock the sleep queue if the caller has not already done 552 * it for us. 553 */ 554 if ((flags & PINTERLOCKED) == 0) { 555 id = LOOKUP(ident); 556 _tsleep_interlock(gd, ident, flags); 557 } 558 559 /* 560 * 561 * If no interlock was set we do an integrated interlock here. 562 * Make sure the current process has been untangled from 563 * the userland scheduler and initialize slptime to start 564 * counting. We must interlock the sleep queue before doing 565 * this to avoid wakeup/process-ipi races which can occur under 566 * heavy loads. 567 */ 568 if (lp) { 569 p->p_usched->release_curproc(lp); 570 lp->lwp_slptime = 0; 571 } 572 573 /* 574 * If the interlocked flag is set but our cpu bit in the slpqueue 575 * is no longer set, then a wakeup was processed inbetween the 576 * tsleep_interlock() (ours or the callers), and here. This can 577 * occur under numerous circumstances including when we release the 578 * current process. 579 * 580 * Extreme loads can cause the sending of an IPI (e.g. wakeup()'s) 581 * to process incoming IPIs, thus draining incoming wakeups. 582 */ 583 if ((td->td_flags & TDF_TSLEEPQ) == 0) { 584 logtsleep2(ilockfail, ident); 585 goto resume; 586 } 587 588 /* 589 * scheduling is blocked while in a critical section. Coincide 590 * the descheduled-by-tsleep flag with the descheduling of the 591 * lwkt. 592 */ 593 lwkt_deschedule_self(td); 594 td->td_flags |= TDF_TSLEEP_DESCHEDULED; 595 td->td_wmesg = wmesg; 596 597 /* 598 * Setup the timeout, if any 599 */ 600 if (timo) { 601 callout_init(&thandle); 602 callout_reset(&thandle, timo, endtsleep, td); 603 } 604 605 /* 606 * Beddy bye bye. 607 */ 608 if (lp) { 609 /* 610 * Ok, we are sleeping. Place us in the SSLEEP state. 611 */ 612 KKASSERT((lp->lwp_flag & LWP_ONRUNQ) == 0); 613 /* 614 * tstop() sets LSSTOP, so don't fiddle with that. 615 */ 616 if (lp->lwp_stat != LSSTOP) 617 lp->lwp_stat = LSSLEEP; 618 lp->lwp_ru.ru_nvcsw++; 619 lwkt_switch(); 620 621 /* 622 * And when we are woken up, put us back in LSRUN. If we 623 * slept for over a second, recalculate our estcpu. 624 */ 625 lp->lwp_stat = LSRUN; 626 if (lp->lwp_slptime) 627 p->p_usched->recalculate(lp); 628 lp->lwp_slptime = 0; 629 } else { 630 lwkt_switch(); 631 } 632 633 /* 634 * Make sure we haven't switched cpus while we were asleep. It's 635 * not supposed to happen. Cleanup our temporary flags. 636 */ 637 KKASSERT(gd == td->td_gd); 638 639 /* 640 * Cleanup the timeout. 641 */ 642 if (timo) { 643 if (td->td_flags & TDF_TIMEOUT) { 644 td->td_flags &= ~TDF_TIMEOUT; 645 error = EWOULDBLOCK; 646 } else { 647 callout_stop(&thandle); 648 } 649 } 650 651 /* 652 * Make sure we have been removed from the sleepq. This should 653 * have been done for us already. 654 */ 655 _tsleep_remove(td); 656 td->td_wmesg = NULL; 657 if (td->td_flags & TDF_TSLEEP_DESCHEDULED) { 658 td->td_flags &= ~TDF_TSLEEP_DESCHEDULED; 659 kprintf("td %p (%s) unexpectedly rescheduled\n", 660 td, td->td_comm); 661 } 662 663 /* 664 * Figure out the correct error return. If interrupted by a 665 * signal we want to return EINTR or ERESTART. 666 * 667 * If P_MAILBOX is set no automatic system call restart occurs 668 * and we return EINTR. P_MAILBOX is meant to be used as an 669 * interlock, the user must poll it prior to any system call 670 * that it wishes to interlock a mailbox signal against since 671 * the flag is cleared on *any* system call that sleeps. 672 */ 673 resume: 674 if (p) { 675 if (catch && error == 0) { 676 if ((p->p_flag & P_MAILBOX) && sig == 0) { 677 error = EINTR; 678 } else if (sig != 0 || (sig = CURSIG(lp))) { 679 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 680 error = EINTR; 681 else 682 error = ERESTART; 683 } 684 } 685 lp->lwp_flag &= ~(LWP_BREAKTSLEEP | LWP_SINTR); 686 p->p_flag &= ~P_MAILBOX; 687 } 688 logtsleep1(tsleep_end); 689 crit_exit_quick(td); 690 return (error); 691 } 692 693 /* 694 * Interlocked spinlock sleep. An exclusively held spinlock must 695 * be passed to ssleep(). The function will atomically release the 696 * spinlock and tsleep on the ident, then reacquire the spinlock and 697 * return. 698 * 699 * This routine is fairly important along the critical path, so optimize it 700 * heavily. 701 */ 702 int 703 ssleep(void *ident, struct spinlock *spin, int flags, 704 const char *wmesg, int timo) 705 { 706 globaldata_t gd = mycpu; 707 int error; 708 709 _tsleep_interlock(gd, ident, flags); 710 spin_unlock_wr_quick(gd, spin); 711 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 712 spin_lock_wr_quick(gd, spin); 713 714 return (error); 715 } 716 717 int 718 lksleep(void *ident, struct lock *lock, int flags, 719 const char *wmesg, int timo) 720 { 721 globaldata_t gd = mycpu; 722 int error; 723 724 _tsleep_interlock(gd, ident, flags); 725 lockmgr(lock, LK_RELEASE); 726 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 727 lockmgr(lock, LK_EXCLUSIVE); 728 729 return (error); 730 } 731 732 /* 733 * Interlocked mutex sleep. An exclusively held mutex must be passed 734 * to mtxsleep(). The function will atomically release the mutex 735 * and tsleep on the ident, then reacquire the mutex and return. 736 */ 737 int 738 mtxsleep(void *ident, struct mtx *mtx, int flags, 739 const char *wmesg, int timo) 740 { 741 globaldata_t gd = mycpu; 742 int error; 743 744 _tsleep_interlock(gd, ident, flags); 745 mtx_unlock(mtx); 746 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 747 mtx_lock_ex_quick(mtx, wmesg); 748 749 return (error); 750 } 751 752 /* 753 * Interlocked serializer sleep. An exclusively held serializer must 754 * be passed to zsleep(). The function will atomically release 755 * the serializer and tsleep on the ident, then reacquire the serializer 756 * and return. 757 */ 758 int 759 zsleep(void *ident, struct lwkt_serialize *slz, int flags, 760 const char *wmesg, int timo) 761 { 762 globaldata_t gd = mycpu; 763 int ret; 764 765 ASSERT_SERIALIZED(slz); 766 767 _tsleep_interlock(gd, ident, flags); 768 lwkt_serialize_exit(slz); 769 ret = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 770 lwkt_serialize_enter(slz); 771 772 return ret; 773 } 774 775 /* 776 * Directly block on the LWKT thread by descheduling it. This 777 * is much faster then tsleep(), but the only legal way to wake 778 * us up is to directly schedule the thread. 779 * 780 * Setting TDF_SINTR will cause new signals to directly schedule us. 781 * 782 * This routine must be called while in a critical section. 783 */ 784 int 785 lwkt_sleep(const char *wmesg, int flags) 786 { 787 thread_t td = curthread; 788 int sig; 789 790 if ((flags & PCATCH) == 0 || td->td_lwp == NULL) { 791 td->td_flags |= TDF_BLOCKED; 792 td->td_wmesg = wmesg; 793 lwkt_deschedule_self(td); 794 lwkt_switch(); 795 td->td_wmesg = NULL; 796 td->td_flags &= ~TDF_BLOCKED; 797 return(0); 798 } 799 if ((sig = CURSIG(td->td_lwp)) != 0) { 800 if (SIGISMEMBER(td->td_proc->p_sigacts->ps_sigintr, sig)) 801 return(EINTR); 802 else 803 return(ERESTART); 804 805 } 806 td->td_flags |= TDF_BLOCKED | TDF_SINTR; 807 td->td_wmesg = wmesg; 808 lwkt_deschedule_self(td); 809 lwkt_switch(); 810 td->td_flags &= ~(TDF_BLOCKED | TDF_SINTR); 811 td->td_wmesg = NULL; 812 return(0); 813 } 814 815 /* 816 * Implement the timeout for tsleep. 817 * 818 * We set LWP_BREAKTSLEEP to indicate that an event has occured, but 819 * we only call setrunnable if the process is not stopped. 820 * 821 * This type of callout timeout is scheduled on the same cpu the process 822 * is sleeping on. Also, at the moment, the MP lock is held. 823 */ 824 static void 825 endtsleep(void *arg) 826 { 827 thread_t td = arg; 828 struct lwp *lp; 829 830 ASSERT_MP_LOCK_HELD(curthread); 831 crit_enter(); 832 833 /* 834 * cpu interlock. Thread flags are only manipulated on 835 * the cpu owning the thread. proc flags are only manipulated 836 * by the older of the MP lock. We have both. 837 */ 838 if (td->td_flags & TDF_TSLEEP_DESCHEDULED) { 839 td->td_flags |= TDF_TIMEOUT; 840 841 if ((lp = td->td_lwp) != NULL) { 842 lp->lwp_flag |= LWP_BREAKTSLEEP; 843 if (lp->lwp_proc->p_stat != SSTOP) 844 setrunnable(lp); 845 } else { 846 _tsleep_wakeup(td); 847 } 848 } 849 crit_exit(); 850 } 851 852 /* 853 * Make all processes sleeping on the specified identifier runnable. 854 * count may be zero or one only. 855 * 856 * The domain encodes the sleep/wakeup domain AND the first cpu to check 857 * (which is always the current cpu). As we iterate across cpus 858 * 859 * This call may run without the MP lock held. We can only manipulate thread 860 * state on the cpu owning the thread. We CANNOT manipulate process state 861 * at all. 862 */ 863 static void 864 _wakeup(void *ident, int domain) 865 { 866 struct tslpque *qp; 867 struct thread *td; 868 struct thread *ntd; 869 globaldata_t gd; 870 #ifdef SMP 871 cpumask_t mask; 872 #endif 873 int id; 874 875 crit_enter(); 876 logtsleep2(wakeup_beg, ident); 877 gd = mycpu; 878 id = LOOKUP(ident); 879 qp = &gd->gd_tsleep_hash[id]; 880 restart: 881 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) { 882 ntd = TAILQ_NEXT(td, td_sleepq); 883 if (td->td_wchan == ident && 884 td->td_wdomain == (domain & PDOMAIN_MASK) 885 ) { 886 KKASSERT(td->td_gd == gd); 887 _tsleep_remove(td); 888 if (td->td_flags & TDF_TSLEEP_DESCHEDULED) { 889 td->td_flags &= ~TDF_TSLEEP_DESCHEDULED; 890 lwkt_schedule(td); 891 if (domain & PWAKEUP_ONE) 892 goto done; 893 } 894 goto restart; 895 } 896 } 897 898 #ifdef SMP 899 /* 900 * We finished checking the current cpu but there still may be 901 * more work to do. Either wakeup_one was requested and no matching 902 * thread was found, or a normal wakeup was requested and we have 903 * to continue checking cpus. 904 * 905 * It should be noted that this scheme is actually less expensive then 906 * the old scheme when waking up multiple threads, since we send 907 * only one IPI message per target candidate which may then schedule 908 * multiple threads. Before we could have wound up sending an IPI 909 * message for each thread on the target cpu (!= current cpu) that 910 * needed to be woken up. 911 * 912 * NOTE: Wakeups occuring on remote cpus are asynchronous. This 913 * should be ok since we are passing idents in the IPI rather then 914 * thread pointers. 915 */ 916 if ((domain & PWAKEUP_MYCPU) == 0 && 917 (mask = slpque_cpumasks[id] & gd->gd_other_cpus) != 0) { 918 lwkt_send_ipiq2_mask(mask, _wakeup, ident, 919 domain | PWAKEUP_MYCPU); 920 } 921 #endif 922 done: 923 logtsleep1(wakeup_end); 924 crit_exit(); 925 } 926 927 /* 928 * Wakeup all threads tsleep()ing on the specified ident, on all cpus 929 */ 930 void 931 wakeup(void *ident) 932 { 933 _wakeup(ident, PWAKEUP_ENCODE(0, mycpu->gd_cpuid)); 934 } 935 936 /* 937 * Wakeup one thread tsleep()ing on the specified ident, on any cpu. 938 */ 939 void 940 wakeup_one(void *ident) 941 { 942 /* XXX potentially round-robin the first responding cpu */ 943 _wakeup(ident, PWAKEUP_ENCODE(0, mycpu->gd_cpuid) | PWAKEUP_ONE); 944 } 945 946 /* 947 * Wakeup threads tsleep()ing on the specified ident on the current cpu 948 * only. 949 */ 950 void 951 wakeup_mycpu(void *ident) 952 { 953 _wakeup(ident, PWAKEUP_MYCPU); 954 } 955 956 /* 957 * Wakeup one thread tsleep()ing on the specified ident on the current cpu 958 * only. 959 */ 960 void 961 wakeup_mycpu_one(void *ident) 962 { 963 /* XXX potentially round-robin the first responding cpu */ 964 _wakeup(ident, PWAKEUP_MYCPU|PWAKEUP_ONE); 965 } 966 967 /* 968 * Wakeup all thread tsleep()ing on the specified ident on the specified cpu 969 * only. 970 */ 971 void 972 wakeup_oncpu(globaldata_t gd, void *ident) 973 { 974 #ifdef SMP 975 if (gd == mycpu) { 976 _wakeup(ident, PWAKEUP_MYCPU); 977 } else { 978 lwkt_send_ipiq2(gd, _wakeup, ident, PWAKEUP_MYCPU); 979 } 980 #else 981 _wakeup(ident, PWAKEUP_MYCPU); 982 #endif 983 } 984 985 /* 986 * Wakeup one thread tsleep()ing on the specified ident on the specified cpu 987 * only. 988 */ 989 void 990 wakeup_oncpu_one(globaldata_t gd, void *ident) 991 { 992 #ifdef SMP 993 if (gd == mycpu) { 994 _wakeup(ident, PWAKEUP_MYCPU | PWAKEUP_ONE); 995 } else { 996 lwkt_send_ipiq2(gd, _wakeup, ident, PWAKEUP_MYCPU | PWAKEUP_ONE); 997 } 998 #else 999 _wakeup(ident, PWAKEUP_MYCPU | PWAKEUP_ONE); 1000 #endif 1001 } 1002 1003 /* 1004 * Wakeup all threads waiting on the specified ident that slept using 1005 * the specified domain, on all cpus. 1006 */ 1007 void 1008 wakeup_domain(void *ident, int domain) 1009 { 1010 _wakeup(ident, PWAKEUP_ENCODE(domain, mycpu->gd_cpuid)); 1011 } 1012 1013 /* 1014 * Wakeup one thread waiting on the specified ident that slept using 1015 * the specified domain, on any cpu. 1016 */ 1017 void 1018 wakeup_domain_one(void *ident, int domain) 1019 { 1020 /* XXX potentially round-robin the first responding cpu */ 1021 _wakeup(ident, PWAKEUP_ENCODE(domain, mycpu->gd_cpuid) | PWAKEUP_ONE); 1022 } 1023 1024 /* 1025 * setrunnable() 1026 * 1027 * Make a process runnable. The MP lock must be held on call. This only 1028 * has an effect if we are in SSLEEP. We only break out of the 1029 * tsleep if LWP_BREAKTSLEEP is set, otherwise we just fix-up the state. 1030 * 1031 * NOTE: With the MP lock held we can only safely manipulate the process 1032 * structure. We cannot safely manipulate the thread structure. 1033 */ 1034 void 1035 setrunnable(struct lwp *lp) 1036 { 1037 crit_enter(); 1038 ASSERT_MP_LOCK_HELD(curthread); 1039 if (lp->lwp_stat == LSSTOP) 1040 lp->lwp_stat = LSSLEEP; 1041 if (lp->lwp_stat == LSSLEEP && (lp->lwp_flag & LWP_BREAKTSLEEP)) 1042 _tsleep_wakeup(lp->lwp_thread); 1043 crit_exit(); 1044 } 1045 1046 /* 1047 * The process is stopped due to some condition, usually because p_stat is 1048 * set to SSTOP, but also possibly due to being traced. 1049 * 1050 * NOTE! If the caller sets SSTOP, the caller must also clear P_WAITED 1051 * because the parent may check the child's status before the child actually 1052 * gets to this routine. 1053 * 1054 * This routine is called with the current lwp only, typically just 1055 * before returning to userland. 1056 * 1057 * Setting LWP_BREAKTSLEEP before entering the tsleep will cause a passive 1058 * SIGCONT to break out of the tsleep. 1059 */ 1060 void 1061 tstop(void) 1062 { 1063 struct lwp *lp = curthread->td_lwp; 1064 struct proc *p = lp->lwp_proc; 1065 1066 crit_enter(); 1067 /* 1068 * If LWP_WSTOP is set, we were sleeping 1069 * while our process was stopped. At this point 1070 * we were already counted as stopped. 1071 */ 1072 if ((lp->lwp_flag & LWP_WSTOP) == 0) { 1073 /* 1074 * If we're the last thread to stop, signal 1075 * our parent. 1076 */ 1077 p->p_nstopped++; 1078 lp->lwp_flag |= LWP_WSTOP; 1079 wakeup(&p->p_nstopped); 1080 if (p->p_nstopped == p->p_nthreads) { 1081 p->p_flag &= ~P_WAITED; 1082 wakeup(p->p_pptr); 1083 if ((p->p_pptr->p_sigacts->ps_flag & PS_NOCLDSTOP) == 0) 1084 ksignal(p->p_pptr, SIGCHLD); 1085 } 1086 } 1087 while (p->p_stat == SSTOP) { 1088 lp->lwp_flag |= LWP_BREAKTSLEEP; 1089 lp->lwp_stat = LSSTOP; 1090 tsleep(p, 0, "stop", 0); 1091 } 1092 p->p_nstopped--; 1093 lp->lwp_flag &= ~LWP_WSTOP; 1094 crit_exit(); 1095 } 1096 1097 /* 1098 * Yield / synchronous reschedule. This is a bit tricky because the trap 1099 * code might have set a lazy release on the switch function. Setting 1100 * P_PASSIVE_ACQ will ensure that the lazy release executes when we call 1101 * switch, and that we are given a greater chance of affinity with our 1102 * current cpu. 1103 * 1104 * We call lwkt_setpri_self() to rotate our thread to the end of the lwkt 1105 * run queue. lwkt_switch() will also execute any assigned passive release 1106 * (which usually calls release_curproc()), allowing a same/higher priority 1107 * process to be designated as the current process. 1108 * 1109 * While it is possible for a lower priority process to be designated, 1110 * it's call to lwkt_maybe_switch() in acquire_curproc() will likely 1111 * round-robin back to us and we will be able to re-acquire the current 1112 * process designation. 1113 * 1114 * MPSAFE 1115 */ 1116 void 1117 uio_yield(void) 1118 { 1119 struct thread *td = curthread; 1120 struct proc *p = td->td_proc; 1121 1122 lwkt_setpri_self(td->td_pri & TDPRI_MASK); 1123 if (p) { 1124 p->p_flag |= P_PASSIVE_ACQ; 1125 lwkt_switch(); 1126 p->p_flag &= ~P_PASSIVE_ACQ; 1127 } else { 1128 lwkt_switch(); 1129 } 1130 } 1131 1132 /* 1133 * Compute a tenex style load average of a quantity on 1134 * 1, 5 and 15 minute intervals. 1135 */ 1136 static int loadav_count_runnable(struct lwp *p, void *data); 1137 1138 static void 1139 loadav(void *arg) 1140 { 1141 struct loadavg *avg; 1142 int i, nrun; 1143 1144 nrun = 0; 1145 alllwp_scan(loadav_count_runnable, &nrun); 1146 avg = &averunnable; 1147 for (i = 0; i < 3; i++) { 1148 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] + 1149 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT; 1150 } 1151 1152 /* 1153 * Schedule the next update to occur after 5 seconds, but add a 1154 * random variation to avoid synchronisation with processes that 1155 * run at regular intervals. 1156 */ 1157 callout_reset(&loadav_callout, hz * 4 + (int)(krandom() % (hz * 2 + 1)), 1158 loadav, NULL); 1159 } 1160 1161 static int 1162 loadav_count_runnable(struct lwp *lp, void *data) 1163 { 1164 int *nrunp = data; 1165 thread_t td; 1166 1167 switch (lp->lwp_stat) { 1168 case LSRUN: 1169 if ((td = lp->lwp_thread) == NULL) 1170 break; 1171 if (td->td_flags & TDF_BLOCKED) 1172 break; 1173 ++*nrunp; 1174 break; 1175 default: 1176 break; 1177 } 1178 return(0); 1179 } 1180 1181 /* ARGSUSED */ 1182 static void 1183 sched_setup(void *dummy) 1184 { 1185 callout_init(&loadav_callout); 1186 callout_init(&schedcpu_callout); 1187 1188 /* Kick off timeout driven events by calling first time. */ 1189 schedcpu(NULL); 1190 loadav(NULL); 1191 } 1192 1193