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 1.85 2007/05/24 20:51:16 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", 0); 92 KTR_INFO(KTR_TSLEEP, tsleep, tsleep_end, 0, "tsleep exit", 0); 93 KTR_INFO(KTR_TSLEEP, tsleep, wakeup_beg, 0, "wakeup enter", 0); 94 KTR_INFO(KTR_TSLEEP, tsleep, wakeup_end, 0, "wakeup exit", 0); 95 #define logtsleep(name) KTR_LOG(tsleep_ ## name) 96 97 struct loadavg averunnable = 98 { {0, 0, 0}, FSCALE }; /* load average, of runnable procs */ 99 /* 100 * Constants for averages over 1, 5, and 15 minutes 101 * when sampling at 5 second intervals. 102 */ 103 static fixpt_t cexp[3] = { 104 0.9200444146293232 * FSCALE, /* exp(-1/12) */ 105 0.9834714538216174 * FSCALE, /* exp(-1/60) */ 106 0.9944598480048967 * FSCALE, /* exp(-1/180) */ 107 }; 108 109 static void endtsleep (void *); 110 static void unsleep_and_wakeup_thread(struct thread *td); 111 static void loadav (void *arg); 112 static void schedcpu (void *arg); 113 114 /* 115 * Adjust the scheduler quantum. The quantum is specified in microseconds. 116 * Note that 'tick' is in microseconds per tick. 117 */ 118 static int 119 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS) 120 { 121 int error, new_val; 122 123 new_val = sched_quantum * tick; 124 error = sysctl_handle_int(oidp, &new_val, 0, req); 125 if (error != 0 || req->newptr == NULL) 126 return (error); 127 if (new_val < tick) 128 return (EINVAL); 129 sched_quantum = new_val / tick; 130 hogticks = 2 * sched_quantum; 131 return (0); 132 } 133 134 SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW, 135 0, sizeof sched_quantum, sysctl_kern_quantum, "I", ""); 136 137 /* 138 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the 139 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below 140 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT). 141 * 142 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used: 143 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits). 144 * 145 * If you don't want to bother with the faster/more-accurate formula, you 146 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate 147 * (more general) method of calculating the %age of CPU used by a process. 148 * 149 * decay 95% of `lwp_pctcpu' in 60 seconds; see CCPU_SHIFT before changing 150 */ 151 #define CCPU_SHIFT 11 152 153 static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */ 154 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 155 156 /* 157 * kernel uses `FSCALE', userland (SHOULD) use kern.fscale 158 */ 159 int fscale __unused = FSCALE; /* exported to systat */ 160 SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, ""); 161 162 /* 163 * Recompute process priorities, once a second. 164 * 165 * Since the userland schedulers are typically event oriented, if the 166 * estcpu calculation at wakeup() time is not sufficient to make a 167 * process runnable relative to other processes in the system we have 168 * a 1-second recalc to help out. 169 * 170 * This code also allows us to store sysclock_t data in the process structure 171 * without fear of an overrun, since sysclock_t are guarenteed to hold 172 * several seconds worth of count. 173 * 174 * WARNING! callouts can preempt normal threads. However, they will not 175 * preempt a thread holding a spinlock so we *can* safely use spinlocks. 176 */ 177 static int schedcpu_stats(struct proc *p, void *data __unused); 178 static int schedcpu_resource(struct proc *p, void *data __unused); 179 180 static void 181 schedcpu(void *arg) 182 { 183 allproc_scan(schedcpu_stats, NULL); 184 allproc_scan(schedcpu_resource, NULL); 185 wakeup((caddr_t)&lbolt); 186 wakeup((caddr_t)&lbolt_syncer); 187 callout_reset(&schedcpu_callout, hz, schedcpu, NULL); 188 } 189 190 /* 191 * General process statistics once a second 192 */ 193 static int 194 schedcpu_stats(struct proc *p, void *data __unused) 195 { 196 struct lwp *lp; 197 198 crit_enter(); 199 p->p_swtime++; 200 FOREACH_LWP_IN_PROC(lp, p) { 201 if (lp->lwp_stat == LSSLEEP) 202 lp->lwp_slptime++; 203 204 /* 205 * Only recalculate processes that are active or have slept 206 * less then 2 seconds. The schedulers understand this. 207 */ 208 if (lp->lwp_slptime <= 1) { 209 p->p_usched->recalculate(lp); 210 } else { 211 lp->lwp_pctcpu = (lp->lwp_pctcpu * ccpu) >> FSHIFT; 212 } 213 } 214 crit_exit(); 215 return(0); 216 } 217 218 /* 219 * Resource checks. XXX break out since ksignal/killproc can block, 220 * limiting us to one process killed per second. There is probably 221 * a better way. 222 */ 223 static int 224 schedcpu_resource(struct proc *p, void *data __unused) 225 { 226 u_int64_t ttime; 227 struct lwp *lp; 228 229 crit_enter(); 230 if (p->p_stat == SIDL || 231 p->p_stat == SZOMB || 232 p->p_limit == NULL 233 ) { 234 crit_exit(); 235 return(0); 236 } 237 238 ttime = 0; 239 FOREACH_LWP_IN_PROC(lp, p) { 240 ttime += lp->lwp_thread->td_sticks; 241 ttime += lp->lwp_thread->td_uticks; 242 } 243 244 switch(plimit_testcpulimit(p->p_limit, ttime)) { 245 case PLIMIT_TESTCPU_KILL: 246 killproc(p, "exceeded maximum CPU limit"); 247 break; 248 case PLIMIT_TESTCPU_XCPU: 249 if ((p->p_flag & P_XCPU) == 0) { 250 p->p_flag |= P_XCPU; 251 ksignal(p, SIGXCPU); 252 } 253 break; 254 default: 255 break; 256 } 257 crit_exit(); 258 return(0); 259 } 260 261 /* 262 * This is only used by ps. Generate a cpu percentage use over 263 * a period of one second. 264 * 265 * MPSAFE 266 */ 267 void 268 updatepcpu(struct lwp *lp, int cpticks, int ttlticks) 269 { 270 fixpt_t acc; 271 int remticks; 272 273 acc = (cpticks << FSHIFT) / ttlticks; 274 if (ttlticks >= ESTCPUFREQ) { 275 lp->lwp_pctcpu = acc; 276 } else { 277 remticks = ESTCPUFREQ - ttlticks; 278 lp->lwp_pctcpu = (acc * ttlticks + lp->lwp_pctcpu * remticks) / 279 ESTCPUFREQ; 280 } 281 } 282 283 /* 284 * We're only looking at 7 bits of the address; everything is 285 * aligned to 4, lots of things are aligned to greater powers 286 * of 2. Shift right by 8, i.e. drop the bottom 256 worth. 287 */ 288 #define TABLESIZE 128 289 #define LOOKUP(x) (((intptr_t)(x) >> 8) & (TABLESIZE - 1)) 290 291 static cpumask_t slpque_cpumasks[TABLESIZE]; 292 293 /* 294 * General scheduler initialization. We force a reschedule 25 times 295 * a second by default. Note that cpu0 is initialized in early boot and 296 * cannot make any high level calls. 297 * 298 * Each cpu has its own sleep queue. 299 */ 300 void 301 sleep_gdinit(globaldata_t gd) 302 { 303 static struct tslpque slpque_cpu0[TABLESIZE]; 304 int i; 305 306 if (gd->gd_cpuid == 0) { 307 sched_quantum = (hz + 24) / 25; 308 hogticks = 2 * sched_quantum; 309 310 gd->gd_tsleep_hash = slpque_cpu0; 311 } else { 312 gd->gd_tsleep_hash = kmalloc(sizeof(slpque_cpu0), 313 M_TSLEEP, M_WAITOK | M_ZERO); 314 } 315 for (i = 0; i < TABLESIZE; ++i) 316 TAILQ_INIT(&gd->gd_tsleep_hash[i]); 317 } 318 319 /* 320 * General sleep call. Suspends the current process until a wakeup is 321 * performed on the specified identifier. The process will then be made 322 * runnable with the specified priority. Sleeps at most timo/hz seconds 323 * (0 means no timeout). If flags includes PCATCH flag, signals are checked 324 * before and after sleeping, else signals are not checked. Returns 0 if 325 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a 326 * signal needs to be delivered, ERESTART is returned if the current system 327 * call should be restarted if possible, and EINTR is returned if the system 328 * call should be interrupted by the signal (return EINTR). 329 * 330 * Note that if we are a process, we release_curproc() before messing with 331 * the LWKT scheduler. 332 * 333 * During autoconfiguration or after a panic, a sleep will simply 334 * lower the priority briefly to allow interrupts, then return. 335 */ 336 int 337 tsleep(void *ident, int flags, const char *wmesg, int timo) 338 { 339 struct thread *td = curthread; 340 struct lwp *lp = td->td_lwp; 341 struct proc *p = td->td_proc; /* may be NULL */ 342 globaldata_t gd; 343 int sig; 344 int catch; 345 int id; 346 int error; 347 int oldpri; 348 struct callout thandle; 349 350 /* 351 * NOTE: removed KTRPOINT, it could cause races due to blocking 352 * even in stable. Just scrap it for now. 353 */ 354 if (tsleep_now_works == 0 || panicstr) { 355 /* 356 * After a panic, or before we actually have an operational 357 * softclock, just give interrupts a chance, then just return; 358 * 359 * don't run any other procs or panic below, 360 * in case this is the idle process and already asleep. 361 */ 362 splz(); 363 oldpri = td->td_pri & TDPRI_MASK; 364 lwkt_setpri_self(safepri); 365 lwkt_switch(); 366 lwkt_setpri_self(oldpri); 367 return (0); 368 } 369 logtsleep(tsleep_beg); 370 gd = td->td_gd; 371 KKASSERT(td != &gd->gd_idlethread); /* you must be kidding! */ 372 373 /* 374 * NOTE: all of this occurs on the current cpu, including any 375 * callout-based wakeups, so a critical section is a sufficient 376 * interlock. 377 * 378 * The entire sequence through to where we actually sleep must 379 * run without breaking the critical section. 380 */ 381 id = LOOKUP(ident); 382 catch = flags & PCATCH; 383 error = 0; 384 sig = 0; 385 386 crit_enter_quick(td); 387 388 KASSERT(ident != NULL, ("tsleep: no ident")); 389 KASSERT(lp == NULL || 390 lp->lwp_stat == LSRUN || /* Obvious */ 391 lp->lwp_stat == LSSTOP, /* Set in tstop */ 392 ("tsleep %p %s %d", 393 ident, wmesg, lp->lwp_stat)); 394 395 /* 396 * Setup for the current process (if this is a process). 397 */ 398 if (lp) { 399 if (catch) { 400 /* 401 * Early termination if PCATCH was set and a 402 * signal is pending, interlocked with the 403 * critical section. 404 * 405 * Early termination only occurs when tsleep() is 406 * entered while in a normal LSRUN state. 407 */ 408 if ((sig = CURSIG(lp)) != 0) 409 goto resume; 410 411 /* 412 * Early termination if PCATCH was set and a 413 * mailbox signal was possibly delivered prior to 414 * the system call even being made, in order to 415 * allow the user to interlock without having to 416 * make additional system calls. 417 */ 418 if (p->p_flag & P_MAILBOX) 419 goto resume; 420 421 /* 422 * Causes ksignal to wake us up when. 423 */ 424 lp->lwp_flag |= LWP_SINTR; 425 } 426 427 /* 428 * Make sure the current process has been untangled from 429 * the userland scheduler and initialize slptime to start 430 * counting. 431 */ 432 if (flags & PNORESCHED) 433 td->td_flags |= TDF_NORESCHED; 434 p->p_usched->release_curproc(lp); 435 lp->lwp_slptime = 0; 436 } 437 438 /* 439 * Move our thread to the correct queue and setup our wchan, etc. 440 */ 441 lwkt_deschedule_self(td); 442 td->td_flags |= TDF_TSLEEPQ; 443 TAILQ_INSERT_TAIL(&gd->gd_tsleep_hash[id], td, td_threadq); 444 atomic_set_int(&slpque_cpumasks[id], gd->gd_cpumask); 445 446 td->td_wchan = ident; 447 td->td_wmesg = wmesg; 448 td->td_wdomain = flags & PDOMAIN_MASK; 449 450 /* 451 * Setup the timeout, if any 452 */ 453 if (timo) { 454 callout_init(&thandle); 455 callout_reset(&thandle, timo, endtsleep, td); 456 } 457 458 /* 459 * Beddy bye bye. 460 */ 461 if (lp) { 462 /* 463 * Ok, we are sleeping. Place us in the SSLEEP state. 464 */ 465 KKASSERT((lp->lwp_flag & LWP_ONRUNQ) == 0); 466 /* 467 * tstop() sets LSSTOP, so don't fiddle with that. 468 */ 469 if (lp->lwp_stat != LSSTOP) 470 lp->lwp_stat = LSSLEEP; 471 lp->lwp_ru.ru_nvcsw++; 472 lwkt_switch(); 473 474 /* 475 * And when we are woken up, put us back in LSRUN. If we 476 * slept for over a second, recalculate our estcpu. 477 */ 478 lp->lwp_stat = LSRUN; 479 if (lp->lwp_slptime) 480 p->p_usched->recalculate(lp); 481 lp->lwp_slptime = 0; 482 } else { 483 lwkt_switch(); 484 } 485 486 /* 487 * Make sure we haven't switched cpus while we were asleep. It's 488 * not supposed to happen. Cleanup our temporary flags. 489 */ 490 KKASSERT(gd == td->td_gd); 491 td->td_flags &= ~TDF_NORESCHED; 492 493 /* 494 * Cleanup the timeout. 495 */ 496 if (timo) { 497 if (td->td_flags & TDF_TIMEOUT) { 498 td->td_flags &= ~TDF_TIMEOUT; 499 error = EWOULDBLOCK; 500 } else { 501 callout_stop(&thandle); 502 } 503 } 504 505 /* 506 * Since td_threadq is used both for our run queue AND for the 507 * tsleep hash queue, we can't still be on it at this point because 508 * we've gotten cpu back. 509 */ 510 KASSERT((td->td_flags & TDF_TSLEEPQ) == 0, ("tsleep: impossible thread flags %08x", td->td_flags)); 511 td->td_wchan = NULL; 512 td->td_wmesg = NULL; 513 td->td_wdomain = 0; 514 515 /* 516 * Figure out the correct error return. If interrupted by a 517 * signal we want to return EINTR or ERESTART. 518 * 519 * If P_MAILBOX is set no automatic system call restart occurs 520 * and we return EINTR. P_MAILBOX is meant to be used as an 521 * interlock, the user must poll it prior to any system call 522 * that it wishes to interlock a mailbox signal against since 523 * the flag is cleared on *any* system call that sleeps. 524 */ 525 resume: 526 if (p) { 527 if (catch && error == 0) { 528 if ((p->p_flag & P_MAILBOX) && sig == 0) { 529 error = EINTR; 530 } else if (sig != 0 || (sig = CURSIG(lp))) { 531 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 532 error = EINTR; 533 else 534 error = ERESTART; 535 } 536 } 537 lp->lwp_flag &= ~(LWP_BREAKTSLEEP | LWP_SINTR); 538 p->p_flag &= ~P_MAILBOX; 539 } 540 logtsleep(tsleep_end); 541 crit_exit_quick(td); 542 return (error); 543 } 544 545 /* 546 * This is a dandy function that allows us to interlock tsleep/wakeup 547 * operations with unspecified upper level locks, such as lockmgr locks, 548 * simply by holding a critical section. The sequence is: 549 * 550 * (enter critical section) 551 * (acquire upper level lock) 552 * tsleep_interlock(blah) 553 * (release upper level lock) 554 * tsleep(blah, ...) 555 * (exit critical section) 556 * 557 * Basically this function sets our cpumask for the ident which informs 558 * other cpus that our cpu 'might' be waiting (or about to wait on) the 559 * hash index related to the ident. The critical section prevents another 560 * cpu's wakeup() from being processed on our cpu until we are actually 561 * able to enter the tsleep(). Thus, no race occurs between our attempt 562 * to release a resource and sleep, and another cpu's attempt to acquire 563 * a resource and call wakeup. 564 * 565 * There isn't much of a point to this function unless you call it while 566 * holding a critical section. 567 */ 568 static __inline void 569 _tsleep_interlock(globaldata_t gd, void *ident) 570 { 571 int id = LOOKUP(ident); 572 573 atomic_set_int(&slpque_cpumasks[id], gd->gd_cpumask); 574 } 575 576 void 577 tsleep_interlock(void *ident) 578 { 579 _tsleep_interlock(mycpu, ident); 580 } 581 582 /* 583 * Interlocked spinlock sleep. An exclusively held spinlock must 584 * be passed to msleep(). The function will atomically release the 585 * spinlock and tsleep on the ident, then reacquire the spinlock and 586 * return. 587 * 588 * This routine is fairly important along the critical path, so optimize it 589 * heavily. 590 */ 591 int 592 msleep(void *ident, struct spinlock *spin, int flags, 593 const char *wmesg, int timo) 594 { 595 globaldata_t gd = mycpu; 596 int error; 597 598 crit_enter_gd(gd); 599 _tsleep_interlock(gd, ident); 600 spin_unlock_wr_quick(gd, spin); 601 error = tsleep(ident, flags, wmesg, timo); 602 spin_lock_wr_quick(gd, spin); 603 crit_exit_gd(gd); 604 605 return (error); 606 } 607 608 /* 609 * Directly block on the LWKT thread by descheduling it. This 610 * is much faster then tsleep(), but the only legal way to wake 611 * us up is to directly schedule the thread. 612 * 613 * Setting TDF_SINTR will cause new signals to directly schedule us. 614 * 615 * This routine is typically called while in a critical section. 616 */ 617 int 618 lwkt_sleep(const char *wmesg, int flags) 619 { 620 thread_t td = curthread; 621 int sig; 622 623 if ((flags & PCATCH) == 0 || td->td_lwp == NULL) { 624 td->td_flags |= TDF_BLOCKED; 625 td->td_wmesg = wmesg; 626 lwkt_deschedule_self(td); 627 lwkt_switch(); 628 td->td_wmesg = NULL; 629 td->td_flags &= ~TDF_BLOCKED; 630 return(0); 631 } 632 if ((sig = CURSIG(td->td_lwp)) != 0) { 633 if (SIGISMEMBER(td->td_proc->p_sigacts->ps_sigintr, sig)) 634 return(EINTR); 635 else 636 return(ERESTART); 637 638 } 639 td->td_flags |= TDF_BLOCKED | TDF_SINTR; 640 td->td_wmesg = wmesg; 641 lwkt_deschedule_self(td); 642 lwkt_switch(); 643 td->td_flags &= ~(TDF_BLOCKED | TDF_SINTR); 644 td->td_wmesg = NULL; 645 return(0); 646 } 647 648 /* 649 * Implement the timeout for tsleep. 650 * 651 * We set LWP_BREAKTSLEEP to indicate that an event has occured, but 652 * we only call setrunnable if the process is not stopped. 653 * 654 * This type of callout timeout is scheduled on the same cpu the process 655 * is sleeping on. Also, at the moment, the MP lock is held. 656 */ 657 static void 658 endtsleep(void *arg) 659 { 660 thread_t td = arg; 661 struct lwp *lp; 662 663 ASSERT_MP_LOCK_HELD(curthread); 664 crit_enter(); 665 666 /* 667 * cpu interlock. Thread flags are only manipulated on 668 * the cpu owning the thread. proc flags are only manipulated 669 * by the older of the MP lock. We have both. 670 */ 671 if (td->td_flags & TDF_TSLEEPQ) { 672 td->td_flags |= TDF_TIMEOUT; 673 674 if ((lp = td->td_lwp) != NULL) { 675 lp->lwp_flag |= LWP_BREAKTSLEEP; 676 if (lp->lwp_proc->p_stat != SSTOP) 677 setrunnable(lp); 678 } else { 679 unsleep_and_wakeup_thread(td); 680 } 681 } 682 crit_exit(); 683 } 684 685 /* 686 * Unsleep and wakeup a thread. This function runs without the MP lock 687 * which means that it can only manipulate thread state on the owning cpu, 688 * and cannot touch the process state at all. 689 */ 690 static 691 void 692 unsleep_and_wakeup_thread(struct thread *td) 693 { 694 globaldata_t gd = mycpu; 695 int id; 696 697 #ifdef SMP 698 if (td->td_gd != gd) { 699 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)unsleep_and_wakeup_thread, td); 700 return; 701 } 702 #endif 703 crit_enter(); 704 if (td->td_flags & TDF_TSLEEPQ) { 705 td->td_flags &= ~TDF_TSLEEPQ; 706 id = LOOKUP(td->td_wchan); 707 TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_threadq); 708 if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL) 709 atomic_clear_int(&slpque_cpumasks[id], gd->gd_cpumask); 710 lwkt_schedule(td); 711 } 712 crit_exit(); 713 } 714 715 /* 716 * Make all processes sleeping on the specified identifier runnable. 717 * count may be zero or one only. 718 * 719 * The domain encodes the sleep/wakeup domain AND the first cpu to check 720 * (which is always the current cpu). As we iterate across cpus 721 * 722 * This call may run without the MP lock held. We can only manipulate thread 723 * state on the cpu owning the thread. We CANNOT manipulate process state 724 * at all. 725 */ 726 static void 727 _wakeup(void *ident, int domain) 728 { 729 struct tslpque *qp; 730 struct thread *td; 731 struct thread *ntd; 732 globaldata_t gd; 733 #ifdef SMP 734 cpumask_t mask; 735 cpumask_t tmask; 736 int startcpu; 737 int nextcpu; 738 #endif 739 int id; 740 741 crit_enter(); 742 logtsleep(wakeup_beg); 743 gd = mycpu; 744 id = LOOKUP(ident); 745 qp = &gd->gd_tsleep_hash[id]; 746 restart: 747 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) { 748 ntd = TAILQ_NEXT(td, td_threadq); 749 if (td->td_wchan == ident && 750 td->td_wdomain == (domain & PDOMAIN_MASK) 751 ) { 752 KKASSERT(td->td_flags & TDF_TSLEEPQ); 753 td->td_flags &= ~TDF_TSLEEPQ; 754 TAILQ_REMOVE(qp, td, td_threadq); 755 if (TAILQ_FIRST(qp) == NULL) { 756 atomic_clear_int(&slpque_cpumasks[id], 757 gd->gd_cpumask); 758 } 759 lwkt_schedule(td); 760 if (domain & PWAKEUP_ONE) 761 goto done; 762 goto restart; 763 } 764 } 765 766 #ifdef SMP 767 /* 768 * We finished checking the current cpu but there still may be 769 * more work to do. Either wakeup_one was requested and no matching 770 * thread was found, or a normal wakeup was requested and we have 771 * to continue checking cpus. 772 * 773 * The cpu that started the wakeup sequence is encoded in the domain. 774 * We use this information to determine which cpus still need to be 775 * checked, locate a candidate cpu, and chain the wakeup 776 * asynchronously with an IPI message. 777 * 778 * It should be noted that this scheme is actually less expensive then 779 * the old scheme when waking up multiple threads, since we send 780 * only one IPI message per target candidate which may then schedule 781 * multiple threads. Before we could have wound up sending an IPI 782 * message for each thread on the target cpu (!= current cpu) that 783 * needed to be woken up. 784 * 785 * NOTE: Wakeups occuring on remote cpus are asynchronous. This 786 * should be ok since we are passing idents in the IPI rather then 787 * thread pointers. 788 */ 789 if ((domain & PWAKEUP_MYCPU) == 0 && 790 (mask = slpque_cpumasks[id]) != 0 791 ) { 792 /* 793 * Look for a cpu that might have work to do. Mask out cpus 794 * which have already been processed. 795 * 796 * 31xxxxxxxxxxxxxxxxxxxxxxxxxxxxx0 797 * ^ ^ ^ 798 * start currentcpu start 799 * case2 case1 800 * * * * 801 * 11111111111111110000000000000111 case1 802 * 00000000111111110000000000000000 case2 803 * 804 * case1: We started at start_case1 and processed through 805 * to the current cpu. We have to check any bits 806 * after the current cpu, then check bits before 807 * the starting cpu. 808 * 809 * case2: We have already checked all the bits from 810 * start_case2 to the end, and from 0 to the current 811 * cpu. We just have the bits from the current cpu 812 * to start_case2 left to check. 813 */ 814 startcpu = PWAKEUP_DECODE(domain); 815 if (gd->gd_cpuid >= startcpu) { 816 /* 817 * CASE1 818 */ 819 tmask = mask & ~((gd->gd_cpumask << 1) - 1); 820 if (mask & tmask) { 821 nextcpu = bsfl(mask & tmask); 822 lwkt_send_ipiq2(globaldata_find(nextcpu), 823 _wakeup, ident, domain); 824 } else { 825 tmask = (1 << startcpu) - 1; 826 if (mask & tmask) { 827 nextcpu = bsfl(mask & tmask); 828 lwkt_send_ipiq2( 829 globaldata_find(nextcpu), 830 _wakeup, ident, domain); 831 } 832 } 833 } else { 834 /* 835 * CASE2 836 */ 837 tmask = ~((gd->gd_cpumask << 1) - 1) & 838 ((1 << startcpu) - 1); 839 if (mask & tmask) { 840 nextcpu = bsfl(mask & tmask); 841 lwkt_send_ipiq2(globaldata_find(nextcpu), 842 _wakeup, ident, domain); 843 } 844 } 845 } 846 #endif 847 done: 848 logtsleep(wakeup_end); 849 crit_exit(); 850 } 851 852 /* 853 * Wakeup all threads tsleep()ing on the specified ident, on all cpus 854 */ 855 void 856 wakeup(void *ident) 857 { 858 _wakeup(ident, PWAKEUP_ENCODE(0, mycpu->gd_cpuid)); 859 } 860 861 /* 862 * Wakeup one thread tsleep()ing on the specified ident, on any cpu. 863 */ 864 void 865 wakeup_one(void *ident) 866 { 867 /* XXX potentially round-robin the first responding cpu */ 868 _wakeup(ident, PWAKEUP_ENCODE(0, mycpu->gd_cpuid) | PWAKEUP_ONE); 869 } 870 871 /* 872 * Wakeup threads tsleep()ing on the specified ident on the current cpu 873 * only. 874 */ 875 void 876 wakeup_mycpu(void *ident) 877 { 878 _wakeup(ident, PWAKEUP_MYCPU); 879 } 880 881 /* 882 * Wakeup one thread tsleep()ing on the specified ident on the current cpu 883 * only. 884 */ 885 void 886 wakeup_mycpu_one(void *ident) 887 { 888 /* XXX potentially round-robin the first responding cpu */ 889 _wakeup(ident, PWAKEUP_MYCPU|PWAKEUP_ONE); 890 } 891 892 /* 893 * Wakeup all thread tsleep()ing on the specified ident on the specified cpu 894 * only. 895 */ 896 void 897 wakeup_oncpu(globaldata_t gd, void *ident) 898 { 899 #ifdef SMP 900 if (gd == mycpu) { 901 _wakeup(ident, PWAKEUP_MYCPU); 902 } else { 903 lwkt_send_ipiq2(gd, _wakeup, ident, PWAKEUP_MYCPU); 904 } 905 #else 906 _wakeup(ident, PWAKEUP_MYCPU); 907 #endif 908 } 909 910 /* 911 * Wakeup one thread tsleep()ing on the specified ident on the specified cpu 912 * only. 913 */ 914 void 915 wakeup_oncpu_one(globaldata_t gd, void *ident) 916 { 917 #ifdef SMP 918 if (gd == mycpu) { 919 _wakeup(ident, PWAKEUP_MYCPU | PWAKEUP_ONE); 920 } else { 921 lwkt_send_ipiq2(gd, _wakeup, ident, PWAKEUP_MYCPU | PWAKEUP_ONE); 922 } 923 #else 924 _wakeup(ident, PWAKEUP_MYCPU | PWAKEUP_ONE); 925 #endif 926 } 927 928 /* 929 * Wakeup all threads waiting on the specified ident that slept using 930 * the specified domain, on all cpus. 931 */ 932 void 933 wakeup_domain(void *ident, int domain) 934 { 935 _wakeup(ident, PWAKEUP_ENCODE(domain, mycpu->gd_cpuid)); 936 } 937 938 /* 939 * Wakeup one thread waiting on the specified ident that slept using 940 * the specified domain, on any cpu. 941 */ 942 void 943 wakeup_domain_one(void *ident, int domain) 944 { 945 /* XXX potentially round-robin the first responding cpu */ 946 _wakeup(ident, PWAKEUP_ENCODE(domain, mycpu->gd_cpuid) | PWAKEUP_ONE); 947 } 948 949 /* 950 * setrunnable() 951 * 952 * Make a process runnable. The MP lock must be held on call. This only 953 * has an effect if we are in SSLEEP. We only break out of the 954 * tsleep if LWP_BREAKTSLEEP is set, otherwise we just fix-up the state. 955 * 956 * NOTE: With the MP lock held we can only safely manipulate the process 957 * structure. We cannot safely manipulate the thread structure. 958 */ 959 void 960 setrunnable(struct lwp *lp) 961 { 962 crit_enter(); 963 ASSERT_MP_LOCK_HELD(curthread); 964 if (lp->lwp_stat == LSSTOP) 965 lp->lwp_stat = LSSLEEP; 966 if (lp->lwp_stat == LSSLEEP && (lp->lwp_flag & LWP_BREAKTSLEEP)) 967 unsleep_and_wakeup_thread(lp->lwp_thread); 968 crit_exit(); 969 } 970 971 /* 972 * The process is stopped due to some condition, usually because p_stat is 973 * set to SSTOP, but also possibly due to being traced. 974 * 975 * NOTE! If the caller sets SSTOP, the caller must also clear P_WAITED 976 * because the parent may check the child's status before the child actually 977 * gets to this routine. 978 * 979 * This routine is called with the current lwp only, typically just 980 * before returning to userland. 981 * 982 * Setting LWP_BREAKTSLEEP before entering the tsleep will cause a passive 983 * SIGCONT to break out of the tsleep. 984 */ 985 void 986 tstop(void) 987 { 988 struct lwp *lp = curthread->td_lwp; 989 struct proc *p = lp->lwp_proc; 990 991 lp->lwp_flag |= LWP_BREAKTSLEEP; 992 lp->lwp_stat = LSSTOP; 993 crit_enter(); 994 /* 995 * If LWP_WSTOP is set, we were sleeping 996 * while our process was stopped. At this point 997 * we were already counted as stopped. 998 */ 999 if ((lp->lwp_flag & LWP_WSTOP) == 0) { 1000 /* 1001 * If we're the last thread to stop, signal 1002 * our parent. 1003 */ 1004 p->p_nstopped++; 1005 lp->lwp_flag |= LWP_WSTOP; 1006 if (p->p_nstopped == p->p_nthreads) { 1007 p->p_flag &= ~P_WAITED; 1008 wakeup(p->p_pptr); 1009 if ((p->p_pptr->p_sigacts->ps_flag & PS_NOCLDSTOP) == 0) 1010 ksignal(p->p_pptr, SIGCHLD); 1011 } 1012 } 1013 tsleep(lp->lwp_proc, 0, "stop", 0); 1014 p->p_nstopped--; 1015 crit_exit(); 1016 } 1017 1018 /* 1019 * Yield / synchronous reschedule. This is a bit tricky because the trap 1020 * code might have set a lazy release on the switch function. Setting 1021 * P_PASSIVE_ACQ will ensure that the lazy release executes when we call 1022 * switch, and that we are given a greater chance of affinity with our 1023 * current cpu. 1024 * 1025 * We call lwkt_setpri_self() to rotate our thread to the end of the lwkt 1026 * run queue. lwkt_switch() will also execute any assigned passive release 1027 * (which usually calls release_curproc()), allowing a same/higher priority 1028 * process to be designated as the current process. 1029 * 1030 * While it is possible for a lower priority process to be designated, 1031 * it's call to lwkt_maybe_switch() in acquire_curproc() will likely 1032 * round-robin back to us and we will be able to re-acquire the current 1033 * process designation. 1034 */ 1035 void 1036 uio_yield(void) 1037 { 1038 struct thread *td = curthread; 1039 struct proc *p = td->td_proc; 1040 1041 lwkt_setpri_self(td->td_pri & TDPRI_MASK); 1042 if (p) { 1043 p->p_flag |= P_PASSIVE_ACQ; 1044 lwkt_switch(); 1045 p->p_flag &= ~P_PASSIVE_ACQ; 1046 } else { 1047 lwkt_switch(); 1048 } 1049 } 1050 1051 /* 1052 * Compute a tenex style load average of a quantity on 1053 * 1, 5 and 15 minute intervals. 1054 */ 1055 static int loadav_count_runnable(struct lwp *p, void *data); 1056 1057 static void 1058 loadav(void *arg) 1059 { 1060 struct loadavg *avg; 1061 int i, nrun; 1062 1063 nrun = 0; 1064 alllwp_scan(loadav_count_runnable, &nrun); 1065 avg = &averunnable; 1066 for (i = 0; i < 3; i++) { 1067 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] + 1068 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT; 1069 } 1070 1071 /* 1072 * Schedule the next update to occur after 5 seconds, but add a 1073 * random variation to avoid synchronisation with processes that 1074 * run at regular intervals. 1075 */ 1076 callout_reset(&loadav_callout, hz * 4 + (int)(krandom() % (hz * 2 + 1)), 1077 loadav, NULL); 1078 } 1079 1080 static int 1081 loadav_count_runnable(struct lwp *lp, void *data) 1082 { 1083 int *nrunp = data; 1084 thread_t td; 1085 1086 switch (lp->lwp_stat) { 1087 case LSRUN: 1088 if ((td = lp->lwp_thread) == NULL) 1089 break; 1090 if (td->td_flags & TDF_BLOCKED) 1091 break; 1092 ++*nrunp; 1093 break; 1094 default: 1095 break; 1096 } 1097 return(0); 1098 } 1099 1100 /* ARGSUSED */ 1101 static void 1102 sched_setup(void *dummy) 1103 { 1104 callout_init(&loadav_callout); 1105 callout_init(&schedcpu_callout); 1106 1107 /* Kick off timeout driven events by calling first time. */ 1108 schedcpu(NULL); 1109 loadav(NULL); 1110 } 1111 1112