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 1.87 2007/08/11 18:18:30 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 65 #include <machine/cpu.h> 66 #include <machine/smp.h> 67 68 TAILQ_HEAD(tslpque, thread); 69 70 static void sched_setup (void *dummy); 71 SYSINIT(sched_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, sched_setup, NULL) 72 73 int hogticks; 74 int lbolt; 75 int lbolt_syncer; 76 int sched_quantum; /* Roundrobin scheduling quantum in ticks. */ 77 int ncpus; 78 int ncpus2, ncpus2_shift, ncpus2_mask; 79 int ncpus_fit, ncpus_fit_mask; 80 int safepri; 81 int tsleep_now_works; 82 83 static struct callout loadav_callout; 84 static struct callout schedcpu_callout; 85 MALLOC_DEFINE(M_TSLEEP, "tslpque", "tsleep queues"); 86 87 #if !defined(KTR_TSLEEP) 88 #define KTR_TSLEEP KTR_ALL 89 #endif 90 KTR_INFO_MASTER(tsleep); 91 KTR_INFO(KTR_TSLEEP, tsleep, tsleep_beg, 0, "tsleep enter %p", sizeof(void *)); 92 KTR_INFO(KTR_TSLEEP, tsleep, tsleep_end, 1, "tsleep exit", 0); 93 KTR_INFO(KTR_TSLEEP, tsleep, wakeup_beg, 2, "wakeup enter %p", sizeof(void *)); 94 KTR_INFO(KTR_TSLEEP, tsleep, wakeup_end, 3, "wakeup exit", 0); 95 96 #define logtsleep1(name) KTR_LOG(tsleep_ ## name) 97 #define logtsleep2(name, val) KTR_LOG(tsleep_ ## name, val) 98 99 struct loadavg averunnable = 100 { {0, 0, 0}, FSCALE }; /* load average, of runnable procs */ 101 /* 102 * Constants for averages over 1, 5, and 15 minutes 103 * when sampling at 5 second intervals. 104 */ 105 static fixpt_t cexp[3] = { 106 0.9200444146293232 * FSCALE, /* exp(-1/12) */ 107 0.9834714538216174 * FSCALE, /* exp(-1/60) */ 108 0.9944598480048967 * FSCALE, /* exp(-1/180) */ 109 }; 110 111 static void endtsleep (void *); 112 static void unsleep_and_wakeup_thread(struct thread *td); 113 static void loadav (void *arg); 114 static void schedcpu (void *arg); 115 116 /* 117 * Adjust the scheduler quantum. The quantum is specified in microseconds. 118 * Note that 'tick' is in microseconds per tick. 119 */ 120 static int 121 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS) 122 { 123 int error, new_val; 124 125 new_val = sched_quantum * tick; 126 error = sysctl_handle_int(oidp, &new_val, 0, req); 127 if (error != 0 || req->newptr == NULL) 128 return (error); 129 if (new_val < tick) 130 return (EINVAL); 131 sched_quantum = new_val / tick; 132 hogticks = 2 * sched_quantum; 133 return (0); 134 } 135 136 SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW, 137 0, sizeof sched_quantum, sysctl_kern_quantum, "I", ""); 138 139 /* 140 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the 141 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below 142 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT). 143 * 144 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used: 145 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits). 146 * 147 * If you don't want to bother with the faster/more-accurate formula, you 148 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate 149 * (more general) method of calculating the %age of CPU used by a process. 150 * 151 * decay 95% of `lwp_pctcpu' in 60 seconds; see CCPU_SHIFT before changing 152 */ 153 #define CCPU_SHIFT 11 154 155 static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */ 156 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 157 158 /* 159 * kernel uses `FSCALE', userland (SHOULD) use kern.fscale 160 */ 161 int fscale __unused = FSCALE; /* exported to systat */ 162 SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, ""); 163 164 /* 165 * Recompute process priorities, once a second. 166 * 167 * Since the userland schedulers are typically event oriented, if the 168 * estcpu calculation at wakeup() time is not sufficient to make a 169 * process runnable relative to other processes in the system we have 170 * a 1-second recalc to help out. 171 * 172 * This code also allows us to store sysclock_t data in the process structure 173 * without fear of an overrun, since sysclock_t are guarenteed to hold 174 * several seconds worth of count. 175 * 176 * WARNING! callouts can preempt normal threads. However, they will not 177 * preempt a thread holding a spinlock so we *can* safely use spinlocks. 178 */ 179 static int schedcpu_stats(struct proc *p, void *data __unused); 180 static int schedcpu_resource(struct proc *p, void *data __unused); 181 182 static void 183 schedcpu(void *arg) 184 { 185 allproc_scan(schedcpu_stats, NULL); 186 allproc_scan(schedcpu_resource, NULL); 187 wakeup((caddr_t)&lbolt); 188 wakeup((caddr_t)&lbolt_syncer); 189 callout_reset(&schedcpu_callout, hz, schedcpu, NULL); 190 } 191 192 /* 193 * General process statistics once a second 194 */ 195 static int 196 schedcpu_stats(struct proc *p, void *data __unused) 197 { 198 struct lwp *lp; 199 200 crit_enter(); 201 p->p_swtime++; 202 FOREACH_LWP_IN_PROC(lp, p) { 203 if (lp->lwp_stat == LSSLEEP) 204 lp->lwp_slptime++; 205 206 /* 207 * Only recalculate processes that are active or have slept 208 * less then 2 seconds. The schedulers understand this. 209 */ 210 if (lp->lwp_slptime <= 1) { 211 p->p_usched->recalculate(lp); 212 } else { 213 lp->lwp_pctcpu = (lp->lwp_pctcpu * ccpu) >> FSHIFT; 214 } 215 } 216 crit_exit(); 217 return(0); 218 } 219 220 /* 221 * Resource checks. XXX break out since ksignal/killproc can block, 222 * limiting us to one process killed per second. There is probably 223 * a better way. 224 */ 225 static int 226 schedcpu_resource(struct proc *p, void *data __unused) 227 { 228 u_int64_t ttime; 229 struct lwp *lp; 230 231 crit_enter(); 232 if (p->p_stat == SIDL || 233 p->p_stat == SZOMB || 234 p->p_limit == NULL 235 ) { 236 crit_exit(); 237 return(0); 238 } 239 240 ttime = 0; 241 FOREACH_LWP_IN_PROC(lp, p) { 242 /* 243 * We may have caught an lp in the middle of being 244 * created, lwp_thread can be NULL. 245 */ 246 if (lp->lwp_thread) { 247 ttime += lp->lwp_thread->td_sticks; 248 ttime += lp->lwp_thread->td_uticks; 249 } 250 } 251 252 switch(plimit_testcpulimit(p->p_limit, ttime)) { 253 case PLIMIT_TESTCPU_KILL: 254 killproc(p, "exceeded maximum CPU limit"); 255 break; 256 case PLIMIT_TESTCPU_XCPU: 257 if ((p->p_flag & P_XCPU) == 0) { 258 p->p_flag |= P_XCPU; 259 ksignal(p, SIGXCPU); 260 } 261 break; 262 default: 263 break; 264 } 265 crit_exit(); 266 return(0); 267 } 268 269 /* 270 * This is only used by ps. Generate a cpu percentage use over 271 * a period of one second. 272 * 273 * MPSAFE 274 */ 275 void 276 updatepcpu(struct lwp *lp, int cpticks, int ttlticks) 277 { 278 fixpt_t acc; 279 int remticks; 280 281 acc = (cpticks << FSHIFT) / ttlticks; 282 if (ttlticks >= ESTCPUFREQ) { 283 lp->lwp_pctcpu = acc; 284 } else { 285 remticks = ESTCPUFREQ - ttlticks; 286 lp->lwp_pctcpu = (acc * ttlticks + lp->lwp_pctcpu * remticks) / 287 ESTCPUFREQ; 288 } 289 } 290 291 /* 292 * tsleep/wakeup hash table parameters. Try to find the sweet spot for 293 * like addresses being slept on. 294 */ 295 #define TABLESIZE 1024 296 #define LOOKUP(x) (((intptr_t)(x) >> 6) & (TABLESIZE - 1)) 297 298 static cpumask_t slpque_cpumasks[TABLESIZE]; 299 300 /* 301 * General scheduler initialization. We force a reschedule 25 times 302 * a second by default. Note that cpu0 is initialized in early boot and 303 * cannot make any high level calls. 304 * 305 * Each cpu has its own sleep queue. 306 */ 307 void 308 sleep_gdinit(globaldata_t gd) 309 { 310 static struct tslpque slpque_cpu0[TABLESIZE]; 311 int i; 312 313 if (gd->gd_cpuid == 0) { 314 sched_quantum = (hz + 24) / 25; 315 hogticks = 2 * sched_quantum; 316 317 gd->gd_tsleep_hash = slpque_cpu0; 318 } else { 319 gd->gd_tsleep_hash = kmalloc(sizeof(slpque_cpu0), 320 M_TSLEEP, M_WAITOK | M_ZERO); 321 } 322 for (i = 0; i < TABLESIZE; ++i) 323 TAILQ_INIT(&gd->gd_tsleep_hash[i]); 324 } 325 326 /* 327 * General sleep call. Suspends the current process until a wakeup is 328 * performed on the specified identifier. The process will then be made 329 * runnable with the specified priority. Sleeps at most timo/hz seconds 330 * (0 means no timeout). If flags includes PCATCH flag, signals are checked 331 * before and after sleeping, else signals are not checked. Returns 0 if 332 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a 333 * signal needs to be delivered, ERESTART is returned if the current system 334 * call should be restarted if possible, and EINTR is returned if the system 335 * call should be interrupted by the signal (return EINTR). 336 * 337 * Note that if we are a process, we release_curproc() before messing with 338 * the LWKT scheduler. 339 * 340 * During autoconfiguration or after a panic, a sleep will simply 341 * lower the priority briefly to allow interrupts, then return. 342 */ 343 int 344 tsleep(void *ident, int flags, const char *wmesg, int timo) 345 { 346 struct thread *td = curthread; 347 struct lwp *lp = td->td_lwp; 348 struct proc *p = td->td_proc; /* may be NULL */ 349 globaldata_t gd; 350 int sig; 351 int catch; 352 int id; 353 int error; 354 int oldpri; 355 struct callout thandle; 356 357 /* 358 * NOTE: removed KTRPOINT, it could cause races due to blocking 359 * even in stable. Just scrap it for now. 360 */ 361 if (tsleep_now_works == 0 || panicstr) { 362 /* 363 * After a panic, or before we actually have an operational 364 * softclock, just give interrupts a chance, then just return; 365 * 366 * don't run any other procs or panic below, 367 * in case this is the idle process and already asleep. 368 */ 369 splz(); 370 oldpri = td->td_pri & TDPRI_MASK; 371 lwkt_setpri_self(safepri); 372 lwkt_switch(); 373 lwkt_setpri_self(oldpri); 374 return (0); 375 } 376 logtsleep2(tsleep_beg, ident); 377 gd = td->td_gd; 378 KKASSERT(td != &gd->gd_idlethread); /* you must be kidding! */ 379 380 /* 381 * NOTE: all of this occurs on the current cpu, including any 382 * callout-based wakeups, so a critical section is a sufficient 383 * interlock. 384 * 385 * The entire sequence through to where we actually sleep must 386 * run without breaking the critical section. 387 */ 388 id = LOOKUP(ident); 389 catch = flags & PCATCH; 390 error = 0; 391 sig = 0; 392 393 crit_enter_quick(td); 394 395 KASSERT(ident != NULL, ("tsleep: no ident")); 396 KASSERT(lp == NULL || 397 lp->lwp_stat == LSRUN || /* Obvious */ 398 lp->lwp_stat == LSSTOP, /* Set in tstop */ 399 ("tsleep %p %s %d", 400 ident, wmesg, lp->lwp_stat)); 401 402 /* 403 * Setup for the current process (if this is a process). 404 */ 405 if (lp) { 406 if (catch) { 407 /* 408 * Early termination if PCATCH was set and a 409 * signal is pending, interlocked with the 410 * critical section. 411 * 412 * Early termination only occurs when tsleep() is 413 * entered while in a normal LSRUN state. 414 */ 415 if ((sig = CURSIG(lp)) != 0) 416 goto resume; 417 418 /* 419 * Early termination if PCATCH was set and a 420 * mailbox signal was possibly delivered prior to 421 * the system call even being made, in order to 422 * allow the user to interlock without having to 423 * make additional system calls. 424 */ 425 if (p->p_flag & P_MAILBOX) 426 goto resume; 427 428 /* 429 * Causes ksignal to wake us up when. 430 */ 431 lp->lwp_flag |= LWP_SINTR; 432 } 433 434 /* 435 * Make sure the current process has been untangled from 436 * the userland scheduler and initialize slptime to start 437 * counting. 438 */ 439 if (flags & PNORESCHED) 440 td->td_flags |= TDF_NORESCHED; 441 p->p_usched->release_curproc(lp); 442 lp->lwp_slptime = 0; 443 } 444 445 /* 446 * Move our thread to the correct queue and setup our wchan, etc. 447 */ 448 lwkt_deschedule_self(td); 449 td->td_flags |= TDF_TSLEEPQ; 450 TAILQ_INSERT_TAIL(&gd->gd_tsleep_hash[id], td, td_threadq); 451 atomic_set_int(&slpque_cpumasks[id], gd->gd_cpumask); 452 453 td->td_wchan = ident; 454 td->td_wmesg = wmesg; 455 td->td_wdomain = flags & PDOMAIN_MASK; 456 457 /* 458 * Setup the timeout, if any 459 */ 460 if (timo) { 461 callout_init(&thandle); 462 callout_reset(&thandle, timo, endtsleep, td); 463 } 464 465 /* 466 * Beddy bye bye. 467 */ 468 if (lp) { 469 /* 470 * Ok, we are sleeping. Place us in the SSLEEP state. 471 */ 472 KKASSERT((lp->lwp_flag & LWP_ONRUNQ) == 0); 473 /* 474 * tstop() sets LSSTOP, so don't fiddle with that. 475 */ 476 if (lp->lwp_stat != LSSTOP) 477 lp->lwp_stat = LSSLEEP; 478 lp->lwp_ru.ru_nvcsw++; 479 lwkt_switch(); 480 481 /* 482 * And when we are woken up, put us back in LSRUN. If we 483 * slept for over a second, recalculate our estcpu. 484 */ 485 lp->lwp_stat = LSRUN; 486 if (lp->lwp_slptime) 487 p->p_usched->recalculate(lp); 488 lp->lwp_slptime = 0; 489 } else { 490 lwkt_switch(); 491 } 492 493 /* 494 * Make sure we haven't switched cpus while we were asleep. It's 495 * not supposed to happen. Cleanup our temporary flags. 496 */ 497 KKASSERT(gd == td->td_gd); 498 td->td_flags &= ~TDF_NORESCHED; 499 500 /* 501 * Cleanup the timeout. 502 */ 503 if (timo) { 504 if (td->td_flags & TDF_TIMEOUT) { 505 td->td_flags &= ~TDF_TIMEOUT; 506 error = EWOULDBLOCK; 507 } else { 508 callout_stop(&thandle); 509 } 510 } 511 512 /* 513 * Since td_threadq is used both for our run queue AND for the 514 * tsleep hash queue, we can't still be on it at this point because 515 * we've gotten cpu back. 516 */ 517 KASSERT((td->td_flags & TDF_TSLEEPQ) == 0, ("tsleep: impossible thread flags %08x", td->td_flags)); 518 td->td_wchan = NULL; 519 td->td_wmesg = NULL; 520 td->td_wdomain = 0; 521 522 /* 523 * Figure out the correct error return. If interrupted by a 524 * signal we want to return EINTR or ERESTART. 525 * 526 * If P_MAILBOX is set no automatic system call restart occurs 527 * and we return EINTR. P_MAILBOX is meant to be used as an 528 * interlock, the user must poll it prior to any system call 529 * that it wishes to interlock a mailbox signal against since 530 * the flag is cleared on *any* system call that sleeps. 531 */ 532 resume: 533 if (p) { 534 if (catch && error == 0) { 535 if ((p->p_flag & P_MAILBOX) && sig == 0) { 536 error = EINTR; 537 } else if (sig != 0 || (sig = CURSIG(lp))) { 538 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 539 error = EINTR; 540 else 541 error = ERESTART; 542 } 543 } 544 lp->lwp_flag &= ~(LWP_BREAKTSLEEP | LWP_SINTR); 545 p->p_flag &= ~P_MAILBOX; 546 } 547 logtsleep1(tsleep_end); 548 crit_exit_quick(td); 549 return (error); 550 } 551 552 /* 553 * This is a dandy function that allows us to interlock tsleep/wakeup 554 * operations with unspecified upper level locks, such as lockmgr locks, 555 * simply by holding a critical section. The sequence is: 556 * 557 * (enter critical section) 558 * (acquire upper level lock) 559 * tsleep_interlock(blah) 560 * (release upper level lock) 561 * tsleep(blah, ...) 562 * (exit critical section) 563 * 564 * Basically this function sets our cpumask for the ident which informs 565 * other cpus that our cpu 'might' be waiting (or about to wait on) the 566 * hash index related to the ident. The critical section prevents another 567 * cpu's wakeup() from being processed on our cpu until we are actually 568 * able to enter the tsleep(). Thus, no race occurs between our attempt 569 * to release a resource and sleep, and another cpu's attempt to acquire 570 * a resource and call wakeup. 571 * 572 * There isn't much of a point to this function unless you call it while 573 * holding a critical section. 574 */ 575 static __inline void 576 _tsleep_interlock(globaldata_t gd, void *ident) 577 { 578 int id = LOOKUP(ident); 579 580 atomic_set_int(&slpque_cpumasks[id], gd->gd_cpumask); 581 } 582 583 void 584 tsleep_interlock(void *ident) 585 { 586 _tsleep_interlock(mycpu, ident); 587 } 588 589 /* 590 * Interlocked spinlock sleep. An exclusively held spinlock must 591 * be passed to msleep(). The function will atomically release the 592 * spinlock and tsleep on the ident, then reacquire the spinlock and 593 * return. 594 * 595 * This routine is fairly important along the critical path, so optimize it 596 * heavily. 597 */ 598 int 599 msleep(void *ident, struct spinlock *spin, int flags, 600 const char *wmesg, int timo) 601 { 602 globaldata_t gd = mycpu; 603 int error; 604 605 crit_enter_gd(gd); 606 _tsleep_interlock(gd, ident); 607 spin_unlock_wr_quick(gd, spin); 608 error = tsleep(ident, flags, wmesg, timo); 609 spin_lock_wr_quick(gd, spin); 610 crit_exit_gd(gd); 611 612 return (error); 613 } 614 615 /* 616 * Directly block on the LWKT thread by descheduling it. This 617 * is much faster then tsleep(), but the only legal way to wake 618 * us up is to directly schedule the thread. 619 * 620 * Setting TDF_SINTR will cause new signals to directly schedule us. 621 * 622 * This routine is typically called while in a critical section. 623 */ 624 int 625 lwkt_sleep(const char *wmesg, int flags) 626 { 627 thread_t td = curthread; 628 int sig; 629 630 if ((flags & PCATCH) == 0 || td->td_lwp == NULL) { 631 td->td_flags |= TDF_BLOCKED; 632 td->td_wmesg = wmesg; 633 lwkt_deschedule_self(td); 634 lwkt_switch(); 635 td->td_wmesg = NULL; 636 td->td_flags &= ~TDF_BLOCKED; 637 return(0); 638 } 639 if ((sig = CURSIG(td->td_lwp)) != 0) { 640 if (SIGISMEMBER(td->td_proc->p_sigacts->ps_sigintr, sig)) 641 return(EINTR); 642 else 643 return(ERESTART); 644 645 } 646 td->td_flags |= TDF_BLOCKED | TDF_SINTR; 647 td->td_wmesg = wmesg; 648 lwkt_deschedule_self(td); 649 lwkt_switch(); 650 td->td_flags &= ~(TDF_BLOCKED | TDF_SINTR); 651 td->td_wmesg = NULL; 652 return(0); 653 } 654 655 /* 656 * Implement the timeout for tsleep. 657 * 658 * We set LWP_BREAKTSLEEP to indicate that an event has occured, but 659 * we only call setrunnable if the process is not stopped. 660 * 661 * This type of callout timeout is scheduled on the same cpu the process 662 * is sleeping on. Also, at the moment, the MP lock is held. 663 */ 664 static void 665 endtsleep(void *arg) 666 { 667 thread_t td = arg; 668 struct lwp *lp; 669 670 ASSERT_MP_LOCK_HELD(curthread); 671 crit_enter(); 672 673 /* 674 * cpu interlock. Thread flags are only manipulated on 675 * the cpu owning the thread. proc flags are only manipulated 676 * by the older of the MP lock. We have both. 677 */ 678 if (td->td_flags & TDF_TSLEEPQ) { 679 td->td_flags |= TDF_TIMEOUT; 680 681 if ((lp = td->td_lwp) != NULL) { 682 lp->lwp_flag |= LWP_BREAKTSLEEP; 683 if (lp->lwp_proc->p_stat != SSTOP) 684 setrunnable(lp); 685 } else { 686 unsleep_and_wakeup_thread(td); 687 } 688 } 689 crit_exit(); 690 } 691 692 /* 693 * Unsleep and wakeup a thread. This function runs without the MP lock 694 * which means that it can only manipulate thread state on the owning cpu, 695 * and cannot touch the process state at all. 696 */ 697 static 698 void 699 unsleep_and_wakeup_thread(struct thread *td) 700 { 701 globaldata_t gd = mycpu; 702 int id; 703 704 #ifdef SMP 705 if (td->td_gd != gd) { 706 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)unsleep_and_wakeup_thread, td); 707 return; 708 } 709 #endif 710 crit_enter(); 711 if (td->td_flags & TDF_TSLEEPQ) { 712 td->td_flags &= ~TDF_TSLEEPQ; 713 id = LOOKUP(td->td_wchan); 714 TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_threadq); 715 if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL) 716 atomic_clear_int(&slpque_cpumasks[id], gd->gd_cpumask); 717 lwkt_schedule(td); 718 } 719 crit_exit(); 720 } 721 722 /* 723 * Make all processes sleeping on the specified identifier runnable. 724 * count may be zero or one only. 725 * 726 * The domain encodes the sleep/wakeup domain AND the first cpu to check 727 * (which is always the current cpu). As we iterate across cpus 728 * 729 * This call may run without the MP lock held. We can only manipulate thread 730 * state on the cpu owning the thread. We CANNOT manipulate process state 731 * at all. 732 */ 733 static void 734 _wakeup(void *ident, int domain) 735 { 736 struct tslpque *qp; 737 struct thread *td; 738 struct thread *ntd; 739 globaldata_t gd; 740 #ifdef SMP 741 cpumask_t mask; 742 cpumask_t tmask; 743 int startcpu; 744 int nextcpu; 745 #endif 746 int id; 747 748 crit_enter(); 749 logtsleep2(wakeup_beg, ident); 750 gd = mycpu; 751 id = LOOKUP(ident); 752 qp = &gd->gd_tsleep_hash[id]; 753 restart: 754 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) { 755 ntd = TAILQ_NEXT(td, td_threadq); 756 if (td->td_wchan == ident && 757 td->td_wdomain == (domain & PDOMAIN_MASK) 758 ) { 759 KKASSERT(td->td_flags & TDF_TSLEEPQ); 760 td->td_flags &= ~TDF_TSLEEPQ; 761 TAILQ_REMOVE(qp, td, td_threadq); 762 if (TAILQ_FIRST(qp) == NULL) { 763 atomic_clear_int(&slpque_cpumasks[id], 764 gd->gd_cpumask); 765 } 766 lwkt_schedule(td); 767 if (domain & PWAKEUP_ONE) 768 goto done; 769 goto restart; 770 } 771 } 772 773 #ifdef SMP 774 /* 775 * We finished checking the current cpu but there still may be 776 * more work to do. Either wakeup_one was requested and no matching 777 * thread was found, or a normal wakeup was requested and we have 778 * to continue checking cpus. 779 * 780 * The cpu that started the wakeup sequence is encoded in the domain. 781 * We use this information to determine which cpus still need to be 782 * checked, locate a candidate cpu, and chain the wakeup 783 * asynchronously with an IPI message. 784 * 785 * It should be noted that this scheme is actually less expensive then 786 * the old scheme when waking up multiple threads, since we send 787 * only one IPI message per target candidate which may then schedule 788 * multiple threads. Before we could have wound up sending an IPI 789 * message for each thread on the target cpu (!= current cpu) that 790 * needed to be woken up. 791 * 792 * NOTE: Wakeups occuring on remote cpus are asynchronous. This 793 * should be ok since we are passing idents in the IPI rather then 794 * thread pointers. 795 */ 796 if ((domain & PWAKEUP_MYCPU) == 0 && 797 (mask = slpque_cpumasks[id]) != 0 798 ) { 799 /* 800 * Look for a cpu that might have work to do. Mask out cpus 801 * which have already been processed. 802 * 803 * 31xxxxxxxxxxxxxxxxxxxxxxxxxxxxx0 804 * ^ ^ ^ 805 * start currentcpu start 806 * case2 case1 807 * * * * 808 * 11111111111111110000000000000111 case1 809 * 00000000111111110000000000000000 case2 810 * 811 * case1: We started at start_case1 and processed through 812 * to the current cpu. We have to check any bits 813 * after the current cpu, then check bits before 814 * the starting cpu. 815 * 816 * case2: We have already checked all the bits from 817 * start_case2 to the end, and from 0 to the current 818 * cpu. We just have the bits from the current cpu 819 * to start_case2 left to check. 820 */ 821 startcpu = PWAKEUP_DECODE(domain); 822 if (gd->gd_cpuid >= startcpu) { 823 /* 824 * CASE1 825 */ 826 tmask = mask & ~((gd->gd_cpumask << 1) - 1); 827 if (mask & tmask) { 828 nextcpu = bsfl(mask & tmask); 829 lwkt_send_ipiq2(globaldata_find(nextcpu), 830 _wakeup, ident, domain); 831 } else { 832 tmask = (1 << startcpu) - 1; 833 if (mask & tmask) { 834 nextcpu = bsfl(mask & tmask); 835 lwkt_send_ipiq2( 836 globaldata_find(nextcpu), 837 _wakeup, ident, domain); 838 } 839 } 840 } else { 841 /* 842 * CASE2 843 */ 844 tmask = ~((gd->gd_cpumask << 1) - 1) & 845 ((1 << startcpu) - 1); 846 if (mask & tmask) { 847 nextcpu = bsfl(mask & tmask); 848 lwkt_send_ipiq2(globaldata_find(nextcpu), 849 _wakeup, ident, domain); 850 } 851 } 852 } 853 #endif 854 done: 855 logtsleep1(wakeup_end); 856 crit_exit(); 857 } 858 859 /* 860 * Wakeup all threads tsleep()ing on the specified ident, on all cpus 861 */ 862 void 863 wakeup(void *ident) 864 { 865 _wakeup(ident, PWAKEUP_ENCODE(0, mycpu->gd_cpuid)); 866 } 867 868 /* 869 * Wakeup one thread tsleep()ing on the specified ident, on any cpu. 870 */ 871 void 872 wakeup_one(void *ident) 873 { 874 /* XXX potentially round-robin the first responding cpu */ 875 _wakeup(ident, PWAKEUP_ENCODE(0, mycpu->gd_cpuid) | PWAKEUP_ONE); 876 } 877 878 /* 879 * Wakeup threads tsleep()ing on the specified ident on the current cpu 880 * only. 881 */ 882 void 883 wakeup_mycpu(void *ident) 884 { 885 _wakeup(ident, PWAKEUP_MYCPU); 886 } 887 888 /* 889 * Wakeup one thread tsleep()ing on the specified ident on the current cpu 890 * only. 891 */ 892 void 893 wakeup_mycpu_one(void *ident) 894 { 895 /* XXX potentially round-robin the first responding cpu */ 896 _wakeup(ident, PWAKEUP_MYCPU|PWAKEUP_ONE); 897 } 898 899 /* 900 * Wakeup all thread tsleep()ing on the specified ident on the specified cpu 901 * only. 902 */ 903 void 904 wakeup_oncpu(globaldata_t gd, void *ident) 905 { 906 #ifdef SMP 907 if (gd == mycpu) { 908 _wakeup(ident, PWAKEUP_MYCPU); 909 } else { 910 lwkt_send_ipiq2(gd, _wakeup, ident, PWAKEUP_MYCPU); 911 } 912 #else 913 _wakeup(ident, PWAKEUP_MYCPU); 914 #endif 915 } 916 917 /* 918 * Wakeup one thread tsleep()ing on the specified ident on the specified cpu 919 * only. 920 */ 921 void 922 wakeup_oncpu_one(globaldata_t gd, void *ident) 923 { 924 #ifdef SMP 925 if (gd == mycpu) { 926 _wakeup(ident, PWAKEUP_MYCPU | PWAKEUP_ONE); 927 } else { 928 lwkt_send_ipiq2(gd, _wakeup, ident, PWAKEUP_MYCPU | PWAKEUP_ONE); 929 } 930 #else 931 _wakeup(ident, PWAKEUP_MYCPU | PWAKEUP_ONE); 932 #endif 933 } 934 935 /* 936 * Wakeup all threads waiting on the specified ident that slept using 937 * the specified domain, on all cpus. 938 */ 939 void 940 wakeup_domain(void *ident, int domain) 941 { 942 _wakeup(ident, PWAKEUP_ENCODE(domain, mycpu->gd_cpuid)); 943 } 944 945 /* 946 * Wakeup one thread waiting on the specified ident that slept using 947 * the specified domain, on any cpu. 948 */ 949 void 950 wakeup_domain_one(void *ident, int domain) 951 { 952 /* XXX potentially round-robin the first responding cpu */ 953 _wakeup(ident, PWAKEUP_ENCODE(domain, mycpu->gd_cpuid) | PWAKEUP_ONE); 954 } 955 956 /* 957 * setrunnable() 958 * 959 * Make a process runnable. The MP lock must be held on call. This only 960 * has an effect if we are in SSLEEP. We only break out of the 961 * tsleep if LWP_BREAKTSLEEP is set, otherwise we just fix-up the state. 962 * 963 * NOTE: With the MP lock held we can only safely manipulate the process 964 * structure. We cannot safely manipulate the thread structure. 965 */ 966 void 967 setrunnable(struct lwp *lp) 968 { 969 crit_enter(); 970 ASSERT_MP_LOCK_HELD(curthread); 971 if (lp->lwp_stat == LSSTOP) 972 lp->lwp_stat = LSSLEEP; 973 if (lp->lwp_stat == LSSLEEP && (lp->lwp_flag & LWP_BREAKTSLEEP)) 974 unsleep_and_wakeup_thread(lp->lwp_thread); 975 crit_exit(); 976 } 977 978 /* 979 * The process is stopped due to some condition, usually because p_stat is 980 * set to SSTOP, but also possibly due to being traced. 981 * 982 * NOTE! If the caller sets SSTOP, the caller must also clear P_WAITED 983 * because the parent may check the child's status before the child actually 984 * gets to this routine. 985 * 986 * This routine is called with the current lwp only, typically just 987 * before returning to userland. 988 * 989 * Setting LWP_BREAKTSLEEP before entering the tsleep will cause a passive 990 * SIGCONT to break out of the tsleep. 991 */ 992 void 993 tstop(void) 994 { 995 struct lwp *lp = curthread->td_lwp; 996 struct proc *p = lp->lwp_proc; 997 998 lp->lwp_flag |= LWP_BREAKTSLEEP; 999 lp->lwp_stat = LSSTOP; 1000 crit_enter(); 1001 /* 1002 * If LWP_WSTOP is set, we were sleeping 1003 * while our process was stopped. At this point 1004 * we were already counted as stopped. 1005 */ 1006 if ((lp->lwp_flag & LWP_WSTOP) == 0) { 1007 /* 1008 * If we're the last thread to stop, signal 1009 * our parent. 1010 */ 1011 p->p_nstopped++; 1012 lp->lwp_flag |= LWP_WSTOP; 1013 if (p->p_nstopped == p->p_nthreads) { 1014 p->p_flag &= ~P_WAITED; 1015 wakeup(p->p_pptr); 1016 if ((p->p_pptr->p_sigacts->ps_flag & PS_NOCLDSTOP) == 0) 1017 ksignal(p->p_pptr, SIGCHLD); 1018 } 1019 } 1020 tsleep(lp->lwp_proc, 0, "stop", 0); 1021 p->p_nstopped--; 1022 crit_exit(); 1023 } 1024 1025 /* 1026 * Yield / synchronous reschedule. This is a bit tricky because the trap 1027 * code might have set a lazy release on the switch function. Setting 1028 * P_PASSIVE_ACQ will ensure that the lazy release executes when we call 1029 * switch, and that we are given a greater chance of affinity with our 1030 * current cpu. 1031 * 1032 * We call lwkt_setpri_self() to rotate our thread to the end of the lwkt 1033 * run queue. lwkt_switch() will also execute any assigned passive release 1034 * (which usually calls release_curproc()), allowing a same/higher priority 1035 * process to be designated as the current process. 1036 * 1037 * While it is possible for a lower priority process to be designated, 1038 * it's call to lwkt_maybe_switch() in acquire_curproc() will likely 1039 * round-robin back to us and we will be able to re-acquire the current 1040 * process designation. 1041 */ 1042 void 1043 uio_yield(void) 1044 { 1045 struct thread *td = curthread; 1046 struct proc *p = td->td_proc; 1047 1048 lwkt_setpri_self(td->td_pri & TDPRI_MASK); 1049 if (p) { 1050 p->p_flag |= P_PASSIVE_ACQ; 1051 lwkt_switch(); 1052 p->p_flag &= ~P_PASSIVE_ACQ; 1053 } else { 1054 lwkt_switch(); 1055 } 1056 } 1057 1058 /* 1059 * Compute a tenex style load average of a quantity on 1060 * 1, 5 and 15 minute intervals. 1061 */ 1062 static int loadav_count_runnable(struct lwp *p, void *data); 1063 1064 static void 1065 loadav(void *arg) 1066 { 1067 struct loadavg *avg; 1068 int i, nrun; 1069 1070 nrun = 0; 1071 alllwp_scan(loadav_count_runnable, &nrun); 1072 avg = &averunnable; 1073 for (i = 0; i < 3; i++) { 1074 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] + 1075 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT; 1076 } 1077 1078 /* 1079 * Schedule the next update to occur after 5 seconds, but add a 1080 * random variation to avoid synchronisation with processes that 1081 * run at regular intervals. 1082 */ 1083 callout_reset(&loadav_callout, hz * 4 + (int)(krandom() % (hz * 2 + 1)), 1084 loadav, NULL); 1085 } 1086 1087 static int 1088 loadav_count_runnable(struct lwp *lp, void *data) 1089 { 1090 int *nrunp = data; 1091 thread_t td; 1092 1093 switch (lp->lwp_stat) { 1094 case LSRUN: 1095 if ((td = lp->lwp_thread) == NULL) 1096 break; 1097 if (td->td_flags & TDF_BLOCKED) 1098 break; 1099 ++*nrunp; 1100 break; 1101 default: 1102 break; 1103 } 1104 return(0); 1105 } 1106 1107 /* ARGSUSED */ 1108 static void 1109 sched_setup(void *dummy) 1110 { 1111 callout_init(&loadav_callout); 1112 callout_init(&schedcpu_callout); 1113 1114 /* Kick off timeout driven events by calling first time. */ 1115 schedcpu(NULL); 1116 loadav(NULL); 1117 } 1118 1119