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 tsleep_wakeup(struct thread *td); 116 static void loadav (void *arg); 117 static void schedcpu (void *arg); 118 119 /* 120 * Adjust the scheduler quantum. The quantum is specified in microseconds. 121 * Note that 'tick' is in microseconds per tick. 122 */ 123 static int 124 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS) 125 { 126 int error, new_val; 127 128 new_val = sched_quantum * tick; 129 error = sysctl_handle_int(oidp, &new_val, 0, req); 130 if (error != 0 || req->newptr == NULL) 131 return (error); 132 if (new_val < tick) 133 return (EINVAL); 134 sched_quantum = new_val / tick; 135 hogticks = 2 * sched_quantum; 136 return (0); 137 } 138 139 SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW, 140 0, sizeof sched_quantum, sysctl_kern_quantum, "I", ""); 141 142 /* 143 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the 144 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below 145 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT). 146 * 147 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used: 148 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits). 149 * 150 * If you don't want to bother with the faster/more-accurate formula, you 151 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate 152 * (more general) method of calculating the %age of CPU used by a process. 153 * 154 * decay 95% of `lwp_pctcpu' in 60 seconds; see CCPU_SHIFT before changing 155 */ 156 #define CCPU_SHIFT 11 157 158 static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */ 159 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 160 161 /* 162 * kernel uses `FSCALE', userland (SHOULD) use kern.fscale 163 */ 164 int fscale __unused = FSCALE; /* exported to systat */ 165 SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, ""); 166 167 /* 168 * Recompute process priorities, once a second. 169 * 170 * Since the userland schedulers are typically event oriented, if the 171 * estcpu calculation at wakeup() time is not sufficient to make a 172 * process runnable relative to other processes in the system we have 173 * a 1-second recalc to help out. 174 * 175 * This code also allows us to store sysclock_t data in the process structure 176 * without fear of an overrun, since sysclock_t are guarenteed to hold 177 * several seconds worth of count. 178 * 179 * WARNING! callouts can preempt normal threads. However, they will not 180 * preempt a thread holding a spinlock so we *can* safely use spinlocks. 181 */ 182 static int schedcpu_stats(struct proc *p, void *data __unused); 183 static int schedcpu_resource(struct proc *p, void *data __unused); 184 185 static void 186 schedcpu(void *arg) 187 { 188 allproc_scan(schedcpu_stats, NULL); 189 allproc_scan(schedcpu_resource, NULL); 190 wakeup((caddr_t)&lbolt); 191 wakeup((caddr_t)&lbolt_syncer); 192 callout_reset(&schedcpu_callout, hz, schedcpu, NULL); 193 } 194 195 /* 196 * General process statistics once a second 197 */ 198 static int 199 schedcpu_stats(struct proc *p, void *data __unused) 200 { 201 struct lwp *lp; 202 203 crit_enter(); 204 p->p_swtime++; 205 FOREACH_LWP_IN_PROC(lp, p) { 206 if (lp->lwp_stat == LSSLEEP) 207 lp->lwp_slptime++; 208 209 /* 210 * Only recalculate processes that are active or have slept 211 * less then 2 seconds. The schedulers understand this. 212 */ 213 if (lp->lwp_slptime <= 1) { 214 p->p_usched->recalculate(lp); 215 } else { 216 lp->lwp_pctcpu = (lp->lwp_pctcpu * ccpu) >> FSHIFT; 217 } 218 } 219 crit_exit(); 220 return(0); 221 } 222 223 /* 224 * Resource checks. XXX break out since ksignal/killproc can block, 225 * limiting us to one process killed per second. There is probably 226 * a better way. 227 */ 228 static int 229 schedcpu_resource(struct proc *p, void *data __unused) 230 { 231 u_int64_t ttime; 232 struct lwp *lp; 233 234 crit_enter(); 235 if (p->p_stat == SIDL || 236 p->p_stat == SZOMB || 237 p->p_limit == NULL 238 ) { 239 crit_exit(); 240 return(0); 241 } 242 243 ttime = 0; 244 FOREACH_LWP_IN_PROC(lp, p) { 245 /* 246 * We may have caught an lp in the middle of being 247 * created, lwp_thread can be NULL. 248 */ 249 if (lp->lwp_thread) { 250 ttime += lp->lwp_thread->td_sticks; 251 ttime += lp->lwp_thread->td_uticks; 252 } 253 } 254 255 switch(plimit_testcpulimit(p->p_limit, ttime)) { 256 case PLIMIT_TESTCPU_KILL: 257 killproc(p, "exceeded maximum CPU limit"); 258 break; 259 case PLIMIT_TESTCPU_XCPU: 260 if ((p->p_flag & P_XCPU) == 0) { 261 p->p_flag |= P_XCPU; 262 ksignal(p, SIGXCPU); 263 } 264 break; 265 default: 266 break; 267 } 268 crit_exit(); 269 return(0); 270 } 271 272 /* 273 * This is only used by ps. Generate a cpu percentage use over 274 * a period of one second. 275 * 276 * MPSAFE 277 */ 278 void 279 updatepcpu(struct lwp *lp, int cpticks, int ttlticks) 280 { 281 fixpt_t acc; 282 int remticks; 283 284 acc = (cpticks << FSHIFT) / ttlticks; 285 if (ttlticks >= ESTCPUFREQ) { 286 lp->lwp_pctcpu = acc; 287 } else { 288 remticks = ESTCPUFREQ - ttlticks; 289 lp->lwp_pctcpu = (acc * ttlticks + lp->lwp_pctcpu * remticks) / 290 ESTCPUFREQ; 291 } 292 } 293 294 /* 295 * tsleep/wakeup hash table parameters. Try to find the sweet spot for 296 * like addresses being slept on. 297 */ 298 #define TABLESIZE 1024 299 #define LOOKUP(x) (((intptr_t)(x) >> 6) & (TABLESIZE - 1)) 300 301 static cpumask_t slpque_cpumasks[TABLESIZE]; 302 303 /* 304 * General scheduler initialization. We force a reschedule 25 times 305 * a second by default. Note that cpu0 is initialized in early boot and 306 * cannot make any high level calls. 307 * 308 * Each cpu has its own sleep queue. 309 */ 310 void 311 sleep_gdinit(globaldata_t gd) 312 { 313 static struct tslpque slpque_cpu0[TABLESIZE]; 314 int i; 315 316 if (gd->gd_cpuid == 0) { 317 sched_quantum = (hz + 24) / 25; 318 hogticks = 2 * sched_quantum; 319 320 gd->gd_tsleep_hash = slpque_cpu0; 321 } else { 322 gd->gd_tsleep_hash = kmalloc(sizeof(slpque_cpu0), 323 M_TSLEEP, M_WAITOK | M_ZERO); 324 } 325 for (i = 0; i < TABLESIZE; ++i) 326 TAILQ_INIT(&gd->gd_tsleep_hash[i]); 327 } 328 329 /* 330 * This is a dandy function that allows us to interlock tsleep/wakeup 331 * operations with unspecified upper level locks, such as lockmgr locks, 332 * simply by holding a critical section. The sequence is: 333 * 334 * (acquire upper level lock) 335 * tsleep_interlock(blah) 336 * (release upper level lock) 337 * tsleep(blah, ...) 338 * 339 * Basically this functions queues us on the tsleep queue without actually 340 * descheduling us. When tsleep() is later called with PINTERLOCK it 341 * assumes the thread was already queued, otherwise it queues it there. 342 * 343 * Thus it is possible to receive the wakeup prior to going to sleep and 344 * the race conditions are covered. 345 */ 346 static __inline void 347 _tsleep_interlock(globaldata_t gd, void *ident, int flags) 348 { 349 thread_t td = gd->gd_curthread; 350 int id; 351 352 crit_enter_quick(td); 353 if (td->td_flags & TDF_TSLEEPQ) { 354 id = LOOKUP(td->td_wchan); 355 TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_sleepq); 356 if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL) 357 atomic_clear_int(&slpque_cpumasks[id], gd->gd_cpumask); 358 } else { 359 td->td_flags |= TDF_TSLEEPQ; 360 } 361 id = LOOKUP(ident); 362 TAILQ_INSERT_TAIL(&gd->gd_tsleep_hash[id], td, td_sleepq); 363 atomic_set_int(&slpque_cpumasks[id], gd->gd_cpumask); 364 td->td_wchan = ident; 365 td->td_wdomain = flags & PDOMAIN_MASK; 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 #ifdef SMP 416 globaldata_t gd = mycpu; 417 418 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 /* 547 * We interlock the sleep queue if the caller has not already done 548 * it for us. 549 */ 550 if ((flags & PINTERLOCKED) == 0) { 551 id = LOOKUP(ident); 552 _tsleep_interlock(gd, ident, flags); 553 } 554 555 /* 556 * 557 * If no interlock was set we do an integrated interlock here. 558 * Make sure the current process has been untangled from 559 * the userland scheduler and initialize slptime to start 560 * counting. We must interlock the sleep queue before doing 561 * this to avoid wakeup/process-ipi races which can occur under 562 * heavy loads. 563 */ 564 if (lp) { 565 p->p_usched->release_curproc(lp); 566 lp->lwp_slptime = 0; 567 } 568 569 /* 570 * If the interlocked flag is set but our cpu bit in the slpqueue 571 * is no longer set, then a wakeup was processed inbetween the 572 * tsleep_interlock() (ours or the callers), and here. This can 573 * occur under numerous circumstances including when we release the 574 * current process. 575 * 576 * Extreme loads can cause the sending of an IPI (e.g. wakeup()'s) 577 * to process incoming IPIs, thus draining incoming wakeups. 578 */ 579 if ((td->td_flags & TDF_TSLEEPQ) == 0) { 580 logtsleep2(ilockfail, ident); 581 goto resume; 582 } 583 584 /* 585 * scheduling is blocked while in a critical section. Coincide 586 * the descheduled-by-tsleep flag with the descheduling of the 587 * lwkt. 588 */ 589 lwkt_deschedule_self(td); 590 td->td_flags |= TDF_TSLEEP_DESCHEDULED; 591 td->td_wmesg = wmesg; 592 593 /* 594 * Setup the timeout, if any 595 */ 596 if (timo) { 597 callout_init(&thandle); 598 callout_reset(&thandle, timo, endtsleep, td); 599 } 600 601 /* 602 * Beddy bye bye. 603 */ 604 if (lp) { 605 /* 606 * Ok, we are sleeping. Place us in the SSLEEP state. 607 */ 608 KKASSERT((lp->lwp_flag & LWP_ONRUNQ) == 0); 609 /* 610 * tstop() sets LSSTOP, so don't fiddle with that. 611 */ 612 if (lp->lwp_stat != LSSTOP) 613 lp->lwp_stat = LSSLEEP; 614 lp->lwp_ru.ru_nvcsw++; 615 lwkt_switch(); 616 617 /* 618 * And when we are woken up, put us back in LSRUN. If we 619 * slept for over a second, recalculate our estcpu. 620 */ 621 lp->lwp_stat = LSRUN; 622 if (lp->lwp_slptime) 623 p->p_usched->recalculate(lp); 624 lp->lwp_slptime = 0; 625 } else { 626 lwkt_switch(); 627 } 628 629 /* 630 * Make sure we haven't switched cpus while we were asleep. It's 631 * not supposed to happen. Cleanup our temporary flags. 632 */ 633 KKASSERT(gd == td->td_gd); 634 635 /* 636 * Cleanup the timeout. 637 */ 638 if (timo) { 639 if (td->td_flags & TDF_TIMEOUT) { 640 td->td_flags &= ~TDF_TIMEOUT; 641 error = EWOULDBLOCK; 642 } else { 643 callout_stop(&thandle); 644 } 645 } 646 647 /* 648 * Make sure we have been removed from the sleepq. This should 649 * have been done for us already. 650 */ 651 _tsleep_remove(td); 652 td->td_wmesg = NULL; 653 if (td->td_flags & TDF_TSLEEP_DESCHEDULED) { 654 td->td_flags &= ~TDF_TSLEEP_DESCHEDULED; 655 kprintf("td %p (%s) unexpectedly rescheduled\n", 656 td, td->td_comm); 657 } 658 659 /* 660 * Figure out the correct error return. If interrupted by a 661 * signal we want to return EINTR or ERESTART. 662 * 663 * If P_MAILBOX is set no automatic system call restart occurs 664 * and we return EINTR. P_MAILBOX is meant to be used as an 665 * interlock, the user must poll it prior to any system call 666 * that it wishes to interlock a mailbox signal against since 667 * the flag is cleared on *any* system call that sleeps. 668 */ 669 resume: 670 if (p) { 671 if (catch && error == 0) { 672 if ((p->p_flag & P_MAILBOX) && sig == 0) { 673 error = EINTR; 674 } else if (sig != 0 || (sig = CURSIG(lp))) { 675 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 676 error = EINTR; 677 else 678 error = ERESTART; 679 } 680 } 681 lp->lwp_flag &= ~(LWP_BREAKTSLEEP | LWP_SINTR); 682 p->p_flag &= ~P_MAILBOX; 683 } 684 logtsleep1(tsleep_end); 685 crit_exit_quick(td); 686 return (error); 687 } 688 689 /* 690 * Interlocked spinlock sleep. An exclusively held spinlock must 691 * be passed to ssleep(). The function will atomically release the 692 * spinlock and tsleep on the ident, then reacquire the spinlock and 693 * return. 694 * 695 * This routine is fairly important along the critical path, so optimize it 696 * heavily. 697 */ 698 int 699 ssleep(void *ident, struct spinlock *spin, int flags, 700 const char *wmesg, int timo) 701 { 702 globaldata_t gd = mycpu; 703 int error; 704 705 _tsleep_interlock(gd, ident, flags); 706 spin_unlock_wr_quick(gd, spin); 707 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 708 spin_lock_wr_quick(gd, spin); 709 710 return (error); 711 } 712 713 int 714 lksleep(void *ident, struct lock *lock, int flags, 715 const char *wmesg, int timo) 716 { 717 globaldata_t gd = mycpu; 718 int error; 719 720 _tsleep_interlock(gd, ident, flags); 721 lockmgr(lock, LK_RELEASE); 722 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 723 lockmgr(lock, LK_EXCLUSIVE); 724 725 return (error); 726 } 727 728 /* 729 * Interlocked mutex sleep. An exclusively held mutex must be passed 730 * to mtxsleep(). The function will atomically release the mutex 731 * and tsleep on the ident, then reacquire the mutex and return. 732 */ 733 int 734 mtxsleep(void *ident, struct mtx *mtx, int flags, 735 const char *wmesg, int timo) 736 { 737 globaldata_t gd = mycpu; 738 int error; 739 740 _tsleep_interlock(gd, ident, flags); 741 mtx_unlock(mtx); 742 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 743 mtx_lock_ex_quick(mtx, wmesg); 744 745 return (error); 746 } 747 748 /* 749 * Interlocked serializer sleep. An exclusively held serializer must 750 * be passed to zsleep(). The function will atomically release 751 * the serializer and tsleep on the ident, then reacquire the serializer 752 * and return. 753 */ 754 int 755 zsleep(void *ident, struct lwkt_serialize *slz, int flags, 756 const char *wmesg, int timo) 757 { 758 globaldata_t gd = mycpu; 759 int ret; 760 761 ASSERT_SERIALIZED(slz); 762 763 _tsleep_interlock(gd, ident, flags); 764 lwkt_serialize_exit(slz); 765 ret = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 766 lwkt_serialize_enter(slz); 767 768 return ret; 769 } 770 771 /* 772 * Directly block on the LWKT thread by descheduling it. This 773 * is much faster then tsleep(), but the only legal way to wake 774 * us up is to directly schedule the thread. 775 * 776 * Setting TDF_SINTR will cause new signals to directly schedule us. 777 * 778 * This routine must be called while in a critical section. 779 */ 780 int 781 lwkt_sleep(const char *wmesg, int flags) 782 { 783 thread_t td = curthread; 784 int sig; 785 786 if ((flags & PCATCH) == 0 || td->td_lwp == NULL) { 787 td->td_flags |= TDF_BLOCKED; 788 td->td_wmesg = wmesg; 789 lwkt_deschedule_self(td); 790 lwkt_switch(); 791 td->td_wmesg = NULL; 792 td->td_flags &= ~TDF_BLOCKED; 793 return(0); 794 } 795 if ((sig = CURSIG(td->td_lwp)) != 0) { 796 if (SIGISMEMBER(td->td_proc->p_sigacts->ps_sigintr, sig)) 797 return(EINTR); 798 else 799 return(ERESTART); 800 801 } 802 td->td_flags |= TDF_BLOCKED | TDF_SINTR; 803 td->td_wmesg = wmesg; 804 lwkt_deschedule_self(td); 805 lwkt_switch(); 806 td->td_flags &= ~(TDF_BLOCKED | TDF_SINTR); 807 td->td_wmesg = NULL; 808 return(0); 809 } 810 811 /* 812 * Implement the timeout for tsleep. 813 * 814 * We set LWP_BREAKTSLEEP to indicate that an event has occured, but 815 * we only call setrunnable if the process is not stopped. 816 * 817 * This type of callout timeout is scheduled on the same cpu the process 818 * is sleeping on. Also, at the moment, the MP lock is held. 819 */ 820 static void 821 endtsleep(void *arg) 822 { 823 thread_t td = arg; 824 struct lwp *lp; 825 826 ASSERT_MP_LOCK_HELD(curthread); 827 crit_enter(); 828 829 /* 830 * cpu interlock. Thread flags are only manipulated on 831 * the cpu owning the thread. proc flags are only manipulated 832 * by the older of the MP lock. We have both. 833 */ 834 if (td->td_flags & TDF_TSLEEP_DESCHEDULED) { 835 td->td_flags |= TDF_TIMEOUT; 836 837 if ((lp = td->td_lwp) != NULL) { 838 lp->lwp_flag |= LWP_BREAKTSLEEP; 839 if (lp->lwp_proc->p_stat != SSTOP) 840 setrunnable(lp); 841 } else { 842 _tsleep_wakeup(td); 843 } 844 } 845 crit_exit(); 846 } 847 848 /* 849 * Make all processes sleeping on the specified identifier runnable. 850 * count may be zero or one only. 851 * 852 * The domain encodes the sleep/wakeup domain AND the first cpu to check 853 * (which is always the current cpu). As we iterate across cpus 854 * 855 * This call may run without the MP lock held. We can only manipulate thread 856 * state on the cpu owning the thread. We CANNOT manipulate process state 857 * at all. 858 */ 859 static void 860 _wakeup(void *ident, int domain) 861 { 862 struct tslpque *qp; 863 struct thread *td; 864 struct thread *ntd; 865 globaldata_t gd; 866 #ifdef SMP 867 cpumask_t mask; 868 #endif 869 int id; 870 871 crit_enter(); 872 logtsleep2(wakeup_beg, ident); 873 gd = mycpu; 874 id = LOOKUP(ident); 875 qp = &gd->gd_tsleep_hash[id]; 876 restart: 877 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) { 878 ntd = TAILQ_NEXT(td, td_sleepq); 879 if (td->td_wchan == ident && 880 td->td_wdomain == (domain & PDOMAIN_MASK) 881 ) { 882 KKASSERT(td->td_gd == gd); 883 _tsleep_remove(td); 884 if (td->td_flags & TDF_TSLEEP_DESCHEDULED) { 885 td->td_flags &= ~TDF_TSLEEP_DESCHEDULED; 886 lwkt_schedule(td); 887 if (domain & PWAKEUP_ONE) 888 goto done; 889 } 890 goto restart; 891 } 892 } 893 894 #ifdef SMP 895 /* 896 * We finished checking the current cpu but there still may be 897 * more work to do. Either wakeup_one was requested and no matching 898 * thread was found, or a normal wakeup was requested and we have 899 * to continue checking cpus. 900 * 901 * It should be noted that this scheme is actually less expensive then 902 * the old scheme when waking up multiple threads, since we send 903 * only one IPI message per target candidate which may then schedule 904 * multiple threads. Before we could have wound up sending an IPI 905 * message for each thread on the target cpu (!= current cpu) that 906 * needed to be woken up. 907 * 908 * NOTE: Wakeups occuring on remote cpus are asynchronous. This 909 * should be ok since we are passing idents in the IPI rather then 910 * thread pointers. 911 */ 912 if ((domain & PWAKEUP_MYCPU) == 0 && 913 (mask = slpque_cpumasks[id] & gd->gd_other_cpus) != 0) { 914 lwkt_send_ipiq2_mask(mask, _wakeup, ident, 915 domain | PWAKEUP_MYCPU); 916 } 917 #endif 918 done: 919 logtsleep1(wakeup_end); 920 crit_exit(); 921 } 922 923 /* 924 * Wakeup all threads tsleep()ing on the specified ident, on all cpus 925 */ 926 void 927 wakeup(void *ident) 928 { 929 _wakeup(ident, PWAKEUP_ENCODE(0, mycpu->gd_cpuid)); 930 } 931 932 /* 933 * Wakeup one thread tsleep()ing on the specified ident, on any cpu. 934 */ 935 void 936 wakeup_one(void *ident) 937 { 938 /* XXX potentially round-robin the first responding cpu */ 939 _wakeup(ident, PWAKEUP_ENCODE(0, mycpu->gd_cpuid) | PWAKEUP_ONE); 940 } 941 942 /* 943 * Wakeup threads tsleep()ing on the specified ident on the current cpu 944 * only. 945 */ 946 void 947 wakeup_mycpu(void *ident) 948 { 949 _wakeup(ident, PWAKEUP_MYCPU); 950 } 951 952 /* 953 * Wakeup one thread tsleep()ing on the specified ident on the current cpu 954 * only. 955 */ 956 void 957 wakeup_mycpu_one(void *ident) 958 { 959 /* XXX potentially round-robin the first responding cpu */ 960 _wakeup(ident, PWAKEUP_MYCPU|PWAKEUP_ONE); 961 } 962 963 /* 964 * Wakeup all thread tsleep()ing on the specified ident on the specified cpu 965 * only. 966 */ 967 void 968 wakeup_oncpu(globaldata_t gd, void *ident) 969 { 970 #ifdef SMP 971 if (gd == mycpu) { 972 _wakeup(ident, PWAKEUP_MYCPU); 973 } else { 974 lwkt_send_ipiq2(gd, _wakeup, ident, PWAKEUP_MYCPU); 975 } 976 #else 977 _wakeup(ident, PWAKEUP_MYCPU); 978 #endif 979 } 980 981 /* 982 * Wakeup one thread tsleep()ing on the specified ident on the specified cpu 983 * only. 984 */ 985 void 986 wakeup_oncpu_one(globaldata_t gd, void *ident) 987 { 988 #ifdef SMP 989 if (gd == mycpu) { 990 _wakeup(ident, PWAKEUP_MYCPU | PWAKEUP_ONE); 991 } else { 992 lwkt_send_ipiq2(gd, _wakeup, ident, PWAKEUP_MYCPU | PWAKEUP_ONE); 993 } 994 #else 995 _wakeup(ident, PWAKEUP_MYCPU | PWAKEUP_ONE); 996 #endif 997 } 998 999 /* 1000 * Wakeup all threads waiting on the specified ident that slept using 1001 * the specified domain, on all cpus. 1002 */ 1003 void 1004 wakeup_domain(void *ident, int domain) 1005 { 1006 _wakeup(ident, PWAKEUP_ENCODE(domain, mycpu->gd_cpuid)); 1007 } 1008 1009 /* 1010 * Wakeup one thread waiting on the specified ident that slept using 1011 * the specified domain, on any cpu. 1012 */ 1013 void 1014 wakeup_domain_one(void *ident, int domain) 1015 { 1016 /* XXX potentially round-robin the first responding cpu */ 1017 _wakeup(ident, PWAKEUP_ENCODE(domain, mycpu->gd_cpuid) | PWAKEUP_ONE); 1018 } 1019 1020 /* 1021 * setrunnable() 1022 * 1023 * Make a process runnable. The MP lock must be held on call. This only 1024 * has an effect if we are in SSLEEP. We only break out of the 1025 * tsleep if LWP_BREAKTSLEEP is set, otherwise we just fix-up the state. 1026 * 1027 * NOTE: With the MP lock held we can only safely manipulate the process 1028 * structure. We cannot safely manipulate the thread structure. 1029 */ 1030 void 1031 setrunnable(struct lwp *lp) 1032 { 1033 crit_enter(); 1034 ASSERT_MP_LOCK_HELD(curthread); 1035 if (lp->lwp_stat == LSSTOP) 1036 lp->lwp_stat = LSSLEEP; 1037 if (lp->lwp_stat == LSSLEEP && (lp->lwp_flag & LWP_BREAKTSLEEP)) 1038 _tsleep_wakeup(lp->lwp_thread); 1039 crit_exit(); 1040 } 1041 1042 /* 1043 * The process is stopped due to some condition, usually because p_stat is 1044 * set to SSTOP, but also possibly due to being traced. 1045 * 1046 * NOTE! If the caller sets SSTOP, the caller must also clear P_WAITED 1047 * because the parent may check the child's status before the child actually 1048 * gets to this routine. 1049 * 1050 * This routine is called with the current lwp only, typically just 1051 * before returning to userland. 1052 * 1053 * Setting LWP_BREAKTSLEEP before entering the tsleep will cause a passive 1054 * SIGCONT to break out of the tsleep. 1055 */ 1056 void 1057 tstop(void) 1058 { 1059 struct lwp *lp = curthread->td_lwp; 1060 struct proc *p = lp->lwp_proc; 1061 1062 crit_enter(); 1063 /* 1064 * If LWP_WSTOP is set, we were sleeping 1065 * while our process was stopped. At this point 1066 * we were already counted as stopped. 1067 */ 1068 if ((lp->lwp_flag & LWP_WSTOP) == 0) { 1069 /* 1070 * If we're the last thread to stop, signal 1071 * our parent. 1072 */ 1073 p->p_nstopped++; 1074 lp->lwp_flag |= LWP_WSTOP; 1075 wakeup(&p->p_nstopped); 1076 if (p->p_nstopped == p->p_nthreads) { 1077 p->p_flag &= ~P_WAITED; 1078 wakeup(p->p_pptr); 1079 if ((p->p_pptr->p_sigacts->ps_flag & PS_NOCLDSTOP) == 0) 1080 ksignal(p->p_pptr, SIGCHLD); 1081 } 1082 } 1083 while (p->p_stat == SSTOP) { 1084 lp->lwp_flag |= LWP_BREAKTSLEEP; 1085 lp->lwp_stat = LSSTOP; 1086 tsleep(p, 0, "stop", 0); 1087 } 1088 p->p_nstopped--; 1089 lp->lwp_flag &= ~LWP_WSTOP; 1090 crit_exit(); 1091 } 1092 1093 /* 1094 * Yield / synchronous reschedule. This is a bit tricky because the trap 1095 * code might have set a lazy release on the switch function. Setting 1096 * P_PASSIVE_ACQ will ensure that the lazy release executes when we call 1097 * switch, and that we are given a greater chance of affinity with our 1098 * current cpu. 1099 * 1100 * We call lwkt_setpri_self() to rotate our thread to the end of the lwkt 1101 * run queue. lwkt_switch() will also execute any assigned passive release 1102 * (which usually calls release_curproc()), allowing a same/higher priority 1103 * process to be designated as the current process. 1104 * 1105 * While it is possible for a lower priority process to be designated, 1106 * it's call to lwkt_maybe_switch() in acquire_curproc() will likely 1107 * round-robin back to us and we will be able to re-acquire the current 1108 * process designation. 1109 */ 1110 void 1111 uio_yield(void) 1112 { 1113 struct thread *td = curthread; 1114 struct proc *p = td->td_proc; 1115 1116 lwkt_setpri_self(td->td_pri & TDPRI_MASK); 1117 if (p) { 1118 p->p_flag |= P_PASSIVE_ACQ; 1119 lwkt_switch(); 1120 p->p_flag &= ~P_PASSIVE_ACQ; 1121 } else { 1122 lwkt_switch(); 1123 } 1124 } 1125 1126 /* 1127 * Compute a tenex style load average of a quantity on 1128 * 1, 5 and 15 minute intervals. 1129 */ 1130 static int loadav_count_runnable(struct lwp *p, void *data); 1131 1132 static void 1133 loadav(void *arg) 1134 { 1135 struct loadavg *avg; 1136 int i, nrun; 1137 1138 nrun = 0; 1139 alllwp_scan(loadav_count_runnable, &nrun); 1140 avg = &averunnable; 1141 for (i = 0; i < 3; i++) { 1142 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] + 1143 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT; 1144 } 1145 1146 /* 1147 * Schedule the next update to occur after 5 seconds, but add a 1148 * random variation to avoid synchronisation with processes that 1149 * run at regular intervals. 1150 */ 1151 callout_reset(&loadav_callout, hz * 4 + (int)(krandom() % (hz * 2 + 1)), 1152 loadav, NULL); 1153 } 1154 1155 static int 1156 loadav_count_runnable(struct lwp *lp, void *data) 1157 { 1158 int *nrunp = data; 1159 thread_t td; 1160 1161 switch (lp->lwp_stat) { 1162 case LSRUN: 1163 if ((td = lp->lwp_thread) == NULL) 1164 break; 1165 if (td->td_flags & TDF_BLOCKED) 1166 break; 1167 ++*nrunp; 1168 break; 1169 default: 1170 break; 1171 } 1172 return(0); 1173 } 1174 1175 /* ARGSUSED */ 1176 static void 1177 sched_setup(void *dummy) 1178 { 1179 callout_init(&loadav_callout); 1180 callout_init(&schedcpu_callout); 1181 1182 /* Kick off timeout driven events by calling first time. */ 1183 schedcpu(NULL); 1184 loadav(NULL); 1185 } 1186 1187