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 1.77 2007/02/19 01:14:23 corecode Exp $ 41 */ 42 43 #include "opt_ktrace.h" 44 45 #include <sys/param.h> 46 #include <sys/systm.h> 47 #include <sys/proc.h> 48 #include <sys/kernel.h> 49 #include <sys/signalvar.h> 50 #include <sys/resourcevar.h> 51 #include <sys/vmmeter.h> 52 #include <sys/sysctl.h> 53 #include <sys/lock.h> 54 #ifdef KTRACE 55 #include <sys/uio.h> 56 #include <sys/ktrace.h> 57 #endif 58 #include <sys/xwait.h> 59 #include <sys/ktr.h> 60 61 #include <sys/thread2.h> 62 #include <sys/spinlock2.h> 63 64 #include <machine/cpu.h> 65 #include <machine/smp.h> 66 67 TAILQ_HEAD(tslpque, thread); 68 69 static void sched_setup (void *dummy); 70 SYSINIT(sched_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, sched_setup, NULL) 71 72 int hogticks; 73 int lbolt; 74 int lbolt_syncer; 75 int sched_quantum; /* Roundrobin scheduling quantum in ticks. */ 76 int ncpus; 77 int ncpus2, ncpus2_shift, ncpus2_mask; 78 int safepri; 79 80 static struct callout loadav_callout; 81 static struct callout schedcpu_callout; 82 MALLOC_DEFINE(M_TSLEEP, "tslpque", "tsleep queues"); 83 84 #if !defined(KTR_TSLEEP) 85 #define KTR_TSLEEP KTR_ALL 86 #endif 87 KTR_INFO_MASTER(tsleep); 88 KTR_INFO(KTR_TSLEEP, tsleep, tsleep_beg, 0, "tsleep enter", 0); 89 KTR_INFO(KTR_TSLEEP, tsleep, tsleep_end, 0, "tsleep exit", 0); 90 KTR_INFO(KTR_TSLEEP, tsleep, wakeup_beg, 0, "wakeup enter", 0); 91 KTR_INFO(KTR_TSLEEP, tsleep, wakeup_end, 0, "wakeup exit", 0); 92 #define logtsleep(name) KTR_LOG(tsleep_ ## name) 93 94 struct loadavg averunnable = 95 { {0, 0, 0}, FSCALE }; /* load average, of runnable procs */ 96 /* 97 * Constants for averages over 1, 5, and 15 minutes 98 * when sampling at 5 second intervals. 99 */ 100 static fixpt_t cexp[3] = { 101 0.9200444146293232 * FSCALE, /* exp(-1/12) */ 102 0.9834714538216174 * FSCALE, /* exp(-1/60) */ 103 0.9944598480048967 * FSCALE, /* exp(-1/180) */ 104 }; 105 106 static void endtsleep (void *); 107 static void unsleep_and_wakeup_thread(struct thread *td); 108 static void loadav (void *arg); 109 static void schedcpu (void *arg); 110 111 /* 112 * Adjust the scheduler quantum. The quantum is specified in microseconds. 113 * Note that 'tick' is in microseconds per tick. 114 */ 115 static int 116 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS) 117 { 118 int error, new_val; 119 120 new_val = sched_quantum * tick; 121 error = sysctl_handle_int(oidp, &new_val, 0, req); 122 if (error != 0 || req->newptr == NULL) 123 return (error); 124 if (new_val < tick) 125 return (EINVAL); 126 sched_quantum = new_val / tick; 127 hogticks = 2 * sched_quantum; 128 return (0); 129 } 130 131 SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW, 132 0, sizeof sched_quantum, sysctl_kern_quantum, "I", ""); 133 134 /* 135 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the 136 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below 137 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT). 138 * 139 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used: 140 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits). 141 * 142 * If you don't want to bother with the faster/more-accurate formula, you 143 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate 144 * (more general) method of calculating the %age of CPU used by a process. 145 * 146 * decay 95% of `lwp_pctcpu' in 60 seconds; see CCPU_SHIFT before changing 147 */ 148 #define CCPU_SHIFT 11 149 150 static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */ 151 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 152 153 /* 154 * kernel uses `FSCALE', userland (SHOULD) use kern.fscale 155 */ 156 int fscale __unused = FSCALE; /* exported to systat */ 157 SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, ""); 158 159 /* 160 * Recompute process priorities, once a second. 161 * 162 * Since the userland schedulers are typically event oriented, if the 163 * estcpu calculation at wakeup() time is not sufficient to make a 164 * process runnable relative to other processes in the system we have 165 * a 1-second recalc to help out. 166 * 167 * This code also allows us to store sysclock_t data in the process structure 168 * without fear of an overrun, since sysclock_t are guarenteed to hold 169 * several seconds worth of count. 170 * 171 * WARNING! callouts can preempt normal threads. However, they will not 172 * preempt a thread holding a spinlock so we *can* safely use spinlocks. 173 */ 174 static int schedcpu_stats(struct proc *p, void *data __unused); 175 static int schedcpu_resource(struct proc *p, void *data __unused); 176 177 static void 178 schedcpu(void *arg) 179 { 180 allproc_scan(schedcpu_stats, NULL); 181 allproc_scan(schedcpu_resource, NULL); 182 wakeup((caddr_t)&lbolt); 183 wakeup((caddr_t)&lbolt_syncer); 184 callout_reset(&schedcpu_callout, hz, schedcpu, NULL); 185 } 186 187 /* 188 * General process statistics once a second 189 */ 190 static int 191 schedcpu_stats(struct proc *p, void *data __unused) 192 { 193 struct lwp *lp; 194 195 crit_enter(); 196 p->p_swtime++; 197 FOREACH_LWP_IN_PROC(lp, p) { 198 if (lp->lwp_stat == LSSLEEP) 199 lp->lwp_slptime++; 200 201 /* 202 * Only recalculate processes that are active or have slept 203 * less then 2 seconds. The schedulers understand this. 204 */ 205 if (lp->lwp_slptime <= 1) { 206 p->p_usched->recalculate(lp); 207 } else { 208 lp->lwp_pctcpu = (lp->lwp_pctcpu * ccpu) >> FSHIFT; 209 } 210 } 211 crit_exit(); 212 return(0); 213 } 214 215 /* 216 * Resource checks. XXX break out since ksignal/killproc can block, 217 * limiting us to one process killed per second. There is probably 218 * a better way. 219 */ 220 static int 221 schedcpu_resource(struct proc *p, void *data __unused) 222 { 223 u_int64_t ttime; 224 struct lwp *lp; 225 226 crit_enter(); 227 if (p->p_stat == SIDL || 228 p->p_stat == SZOMB || 229 p->p_limit == NULL 230 ) { 231 crit_exit(); 232 return(0); 233 } 234 235 ttime = 0; 236 FOREACH_LWP_IN_PROC(lp, p) { 237 ttime += lp->lwp_thread->td_sticks; 238 ttime += lp->lwp_thread->td_uticks; 239 } 240 241 switch(plimit_testcpulimit(p->p_limit, ttime)) { 242 case PLIMIT_TESTCPU_KILL: 243 killproc(p, "exceeded maximum CPU limit"); 244 break; 245 case PLIMIT_TESTCPU_XCPU: 246 if ((p->p_flag & P_XCPU) == 0) { 247 p->p_flag |= P_XCPU; 248 ksignal(p, SIGXCPU); 249 } 250 break; 251 default: 252 break; 253 } 254 crit_exit(); 255 return(0); 256 } 257 258 /* 259 * This is only used by ps. Generate a cpu percentage use over 260 * a period of one second. 261 * 262 * MPSAFE 263 */ 264 void 265 updatepcpu(struct lwp *lp, int cpticks, int ttlticks) 266 { 267 fixpt_t acc; 268 int remticks; 269 270 acc = (cpticks << FSHIFT) / ttlticks; 271 if (ttlticks >= ESTCPUFREQ) { 272 lp->lwp_pctcpu = acc; 273 } else { 274 remticks = ESTCPUFREQ - ttlticks; 275 lp->lwp_pctcpu = (acc * ttlticks + lp->lwp_pctcpu * remticks) / 276 ESTCPUFREQ; 277 } 278 } 279 280 /* 281 * We're only looking at 7 bits of the address; everything is 282 * aligned to 4, lots of things are aligned to greater powers 283 * of 2. Shift right by 8, i.e. drop the bottom 256 worth. 284 */ 285 #define TABLESIZE 128 286 #define LOOKUP(x) (((intptr_t)(x) >> 8) & (TABLESIZE - 1)) 287 288 static cpumask_t slpque_cpumasks[TABLESIZE]; 289 290 /* 291 * General scheduler initialization. We force a reschedule 25 times 292 * a second by default. Note that cpu0 is initialized in early boot and 293 * cannot make any high level calls. 294 * 295 * Each cpu has its own sleep queue. 296 */ 297 void 298 sleep_gdinit(globaldata_t gd) 299 { 300 static struct tslpque slpque_cpu0[TABLESIZE]; 301 int i; 302 303 if (gd->gd_cpuid == 0) { 304 sched_quantum = (hz + 24) / 25; 305 hogticks = 2 * sched_quantum; 306 307 gd->gd_tsleep_hash = slpque_cpu0; 308 } else { 309 gd->gd_tsleep_hash = kmalloc(sizeof(slpque_cpu0), 310 M_TSLEEP, M_WAITOK | M_ZERO); 311 } 312 for (i = 0; i < TABLESIZE; ++i) 313 TAILQ_INIT(&gd->gd_tsleep_hash[i]); 314 } 315 316 /* 317 * General sleep call. Suspends the current process until a wakeup is 318 * performed on the specified identifier. The process will then be made 319 * runnable with the specified priority. Sleeps at most timo/hz seconds 320 * (0 means no timeout). If flags includes PCATCH flag, signals are checked 321 * before and after sleeping, else signals are not checked. Returns 0 if 322 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a 323 * signal needs to be delivered, ERESTART is returned if the current system 324 * call should be restarted if possible, and EINTR is returned if the system 325 * call should be interrupted by the signal (return EINTR). 326 * 327 * Note that if we are a process, we release_curproc() before messing with 328 * the LWKT scheduler. 329 * 330 * During autoconfiguration or after a panic, a sleep will simply 331 * lower the priority briefly to allow interrupts, then return. 332 */ 333 int 334 tsleep(void *ident, int flags, const char *wmesg, int timo) 335 { 336 struct thread *td = curthread; 337 struct lwp *lp = td->td_lwp; 338 struct proc *p = td->td_proc; /* may be NULL */ 339 globaldata_t gd; 340 int sig; 341 int catch; 342 int id; 343 int error; 344 int oldpri; 345 struct callout thandle; 346 347 /* 348 * NOTE: removed KTRPOINT, it could cause races due to blocking 349 * even in stable. Just scrap it for now. 350 */ 351 if (cold || panicstr) { 352 /* 353 * After a panic, or during autoconfiguration, 354 * just give interrupts a chance, then just return; 355 * don't run any other procs or panic below, 356 * in case this is the idle process and already asleep. 357 */ 358 splz(); 359 oldpri = td->td_pri & TDPRI_MASK; 360 lwkt_setpri_self(safepri); 361 lwkt_switch(); 362 lwkt_setpri_self(oldpri); 363 return (0); 364 } 365 logtsleep(tsleep_beg); 366 gd = td->td_gd; 367 KKASSERT(td != &gd->gd_idlethread); /* you must be kidding! */ 368 369 /* 370 * NOTE: all of this occurs on the current cpu, including any 371 * callout-based wakeups, so a critical section is a sufficient 372 * interlock. 373 * 374 * The entire sequence through to where we actually sleep must 375 * run without breaking the critical section. 376 */ 377 id = LOOKUP(ident); 378 catch = flags & PCATCH; 379 error = 0; 380 sig = 0; 381 382 crit_enter_quick(td); 383 384 KASSERT(ident != NULL, ("tsleep: no ident")); 385 KASSERT(lp == NULL || lp->lwp_stat == LSRUN, ("tsleep %p %s %d", 386 ident, wmesg, lp->lwp_stat)); 387 388 /* 389 * Setup for the current process (if this is a process). 390 */ 391 if (lp) { 392 if (catch) { 393 /* 394 * Early termination if PCATCH was set and a 395 * signal is pending, interlocked with the 396 * critical section. 397 * 398 * Early termination only occurs when tsleep() is 399 * entered while in a normal LSRUN state. 400 */ 401 if ((sig = CURSIG(lp)) != 0) 402 goto resume; 403 404 /* 405 * Early termination if PCATCH was set and a 406 * mailbox signal was possibly delivered prior to 407 * the system call even being made, in order to 408 * allow the user to interlock without having to 409 * make additional system calls. 410 */ 411 if (p->p_flag & P_MAILBOX) 412 goto resume; 413 414 /* 415 * Causes ksignal to wake us up when. 416 */ 417 lp->lwp_flag |= LWP_SINTR; 418 } 419 420 /* 421 * Make sure the current process has been untangled from 422 * the userland scheduler and initialize slptime to start 423 * counting. 424 */ 425 if (flags & PNORESCHED) 426 td->td_flags |= TDF_NORESCHED; 427 p->p_usched->release_curproc(lp); 428 lp->lwp_slptime = 0; 429 } 430 431 /* 432 * Move our thread to the correct queue and setup our wchan, etc. 433 */ 434 lwkt_deschedule_self(td); 435 td->td_flags |= TDF_TSLEEPQ; 436 TAILQ_INSERT_TAIL(&gd->gd_tsleep_hash[id], td, td_threadq); 437 atomic_set_int(&slpque_cpumasks[id], gd->gd_cpumask); 438 439 td->td_wchan = ident; 440 td->td_wmesg = wmesg; 441 td->td_wdomain = flags & PDOMAIN_MASK; 442 443 /* 444 * Setup the timeout, if any 445 */ 446 if (timo) { 447 callout_init(&thandle); 448 callout_reset(&thandle, timo, endtsleep, td); 449 } 450 451 /* 452 * Beddy bye bye. 453 */ 454 if (lp) { 455 /* 456 * Ok, we are sleeping. Place us in the SSLEEP state. 457 */ 458 KKASSERT((lp->lwp_flag & LWP_ONRUNQ) == 0); 459 lp->lwp_stat = LSSLEEP; 460 lp->lwp_ru.ru_nvcsw++; 461 lwkt_switch(); 462 463 /* 464 * And when we are woken up, put us back in LSRUN. If we 465 * slept for over a second, recalculate our estcpu. 466 */ 467 lp->lwp_stat = LSRUN; 468 if (lp->lwp_slptime) 469 p->p_usched->recalculate(lp); 470 lp->lwp_slptime = 0; 471 } else { 472 lwkt_switch(); 473 } 474 475 /* 476 * Make sure we haven't switched cpus while we were asleep. It's 477 * not supposed to happen. Cleanup our temporary flags. 478 */ 479 KKASSERT(gd == td->td_gd); 480 td->td_flags &= ~TDF_NORESCHED; 481 482 /* 483 * Cleanup the timeout. 484 */ 485 if (timo) { 486 if (td->td_flags & TDF_TIMEOUT) { 487 td->td_flags &= ~TDF_TIMEOUT; 488 if (sig == 0) 489 error = EWOULDBLOCK; 490 } else { 491 callout_stop(&thandle); 492 } 493 } 494 495 /* 496 * Since td_threadq is used both for our run queue AND for the 497 * tsleep hash queue, we can't still be on it at this point because 498 * we've gotten cpu back. 499 */ 500 KASSERT((td->td_flags & TDF_TSLEEPQ) == 0, ("tsleep: impossible thread flags %08x", td->td_flags)); 501 td->td_wchan = NULL; 502 td->td_wmesg = NULL; 503 td->td_wdomain = 0; 504 505 /* 506 * Figure out the correct error return. If interrupted by a 507 * signal we want to return EINTR or ERESTART. 508 * 509 * If P_MAILBOX is set no automatic system call restart occurs 510 * and we return EINTR. P_MAILBOX is meant to be used as an 511 * interlock, the user must poll it prior to any system call 512 * that it wishes to interlock a mailbox signal against since 513 * the flag is cleared on *any* system call that sleeps. 514 */ 515 resume: 516 if (p) { 517 if (catch && error == 0) { 518 if ((p->p_flag & P_MAILBOX) && sig == 0) { 519 error = EINTR; 520 } else if (sig != 0 || (sig = CURSIG(lp))) { 521 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 522 error = EINTR; 523 else 524 error = ERESTART; 525 } 526 } 527 lp->lwp_flag &= ~(LWP_BREAKTSLEEP | LWP_SINTR); 528 p->p_flag &= ~P_MAILBOX; 529 } 530 logtsleep(tsleep_end); 531 crit_exit_quick(td); 532 return (error); 533 } 534 535 /* 536 * This is a dandy function that allows us to interlock tsleep/wakeup 537 * operations with unspecified upper level locks, such as lockmgr locks, 538 * simply by holding a critical section. The sequence is: 539 * 540 * (enter critical section) 541 * (acquire upper level lock) 542 * tsleep_interlock(blah) 543 * (release upper level lock) 544 * tsleep(blah, ...) 545 * (exit critical section) 546 * 547 * Basically this function sets our cpumask for the ident which informs 548 * other cpus that our cpu 'might' be waiting (or about to wait on) the 549 * hash index related to the ident. The critical section prevents another 550 * cpu's wakeup() from being processed on our cpu until we are actually 551 * able to enter the tsleep(). Thus, no race occurs between our attempt 552 * to release a resource and sleep, and another cpu's attempt to acquire 553 * a resource and call wakeup. 554 * 555 * There isn't much of a point to this function unless you call it while 556 * holding a critical section. 557 */ 558 static __inline void 559 _tsleep_interlock(globaldata_t gd, void *ident) 560 { 561 int id = LOOKUP(ident); 562 563 atomic_set_int(&slpque_cpumasks[id], gd->gd_cpumask); 564 } 565 566 void 567 tsleep_interlock(void *ident) 568 { 569 _tsleep_interlock(mycpu, ident); 570 } 571 572 /* 573 * Interlocked spinlock sleep. An exclusively held spinlock must 574 * be passed to msleep(). The function will atomically release the 575 * spinlock and tsleep on the ident, then reacquire the spinlock and 576 * return. 577 * 578 * This routine is fairly important along the critical path, so optimize it 579 * heavily. 580 */ 581 int 582 msleep(void *ident, struct spinlock *spin, int flags, 583 const char *wmesg, int timo) 584 { 585 globaldata_t gd = mycpu; 586 int error; 587 588 crit_enter_gd(gd); 589 _tsleep_interlock(gd, ident); 590 spin_unlock_wr_quick(gd, spin); 591 error = tsleep(ident, flags, wmesg, timo); 592 spin_lock_wr_quick(gd, spin); 593 crit_exit_gd(gd); 594 595 return (error); 596 } 597 598 /* 599 * Implement the timeout for tsleep. 600 * 601 * We set LWP_BREAKTSLEEP to indicate that an event has occured, but 602 * we only call setrunnable if the process is not stopped. 603 * 604 * This type of callout timeout is scheduled on the same cpu the process 605 * is sleeping on. Also, at the moment, the MP lock is held. 606 */ 607 static void 608 endtsleep(void *arg) 609 { 610 thread_t td = arg; 611 struct lwp *lp; 612 613 ASSERT_MP_LOCK_HELD(curthread); 614 crit_enter(); 615 616 /* 617 * cpu interlock. Thread flags are only manipulated on 618 * the cpu owning the thread. proc flags are only manipulated 619 * by the older of the MP lock. We have both. 620 */ 621 if (td->td_flags & TDF_TSLEEPQ) { 622 td->td_flags |= TDF_TIMEOUT; 623 624 if ((lp = td->td_lwp) != NULL) { 625 lp->lwp_flag |= LWP_BREAKTSLEEP; 626 if (lp->lwp_proc->p_stat != SSTOP) 627 setrunnable(lp); 628 } else { 629 unsleep_and_wakeup_thread(td); 630 } 631 } 632 crit_exit(); 633 } 634 635 /* 636 * Unsleep and wakeup a thread. This function runs without the MP lock 637 * which means that it can only manipulate thread state on the owning cpu, 638 * and cannot touch the process state at all. 639 */ 640 static 641 void 642 unsleep_and_wakeup_thread(struct thread *td) 643 { 644 globaldata_t gd = mycpu; 645 int id; 646 647 #ifdef SMP 648 if (td->td_gd != gd) { 649 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)unsleep_and_wakeup_thread, td); 650 return; 651 } 652 #endif 653 crit_enter(); 654 if (td->td_flags & TDF_TSLEEPQ) { 655 td->td_flags &= ~TDF_TSLEEPQ; 656 id = LOOKUP(td->td_wchan); 657 TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_threadq); 658 if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL) 659 atomic_clear_int(&slpque_cpumasks[id], gd->gd_cpumask); 660 lwkt_schedule(td); 661 } 662 crit_exit(); 663 } 664 665 /* 666 * Make all processes sleeping on the specified identifier runnable. 667 * count may be zero or one only. 668 * 669 * The domain encodes the sleep/wakeup domain AND the first cpu to check 670 * (which is always the current cpu). As we iterate across cpus 671 * 672 * This call may run without the MP lock held. We can only manipulate thread 673 * state on the cpu owning the thread. We CANNOT manipulate process state 674 * at all. 675 */ 676 static void 677 _wakeup(void *ident, int domain) 678 { 679 struct tslpque *qp; 680 struct thread *td; 681 struct thread *ntd; 682 globaldata_t gd; 683 #ifdef SMP 684 cpumask_t mask; 685 cpumask_t tmask; 686 int startcpu; 687 int nextcpu; 688 #endif 689 int id; 690 691 crit_enter(); 692 logtsleep(wakeup_beg); 693 gd = mycpu; 694 id = LOOKUP(ident); 695 qp = &gd->gd_tsleep_hash[id]; 696 restart: 697 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) { 698 ntd = TAILQ_NEXT(td, td_threadq); 699 if (td->td_wchan == ident && 700 td->td_wdomain == (domain & PDOMAIN_MASK) 701 ) { 702 KKASSERT(td->td_flags & TDF_TSLEEPQ); 703 td->td_flags &= ~TDF_TSLEEPQ; 704 TAILQ_REMOVE(qp, td, td_threadq); 705 if (TAILQ_FIRST(qp) == NULL) { 706 atomic_clear_int(&slpque_cpumasks[id], 707 gd->gd_cpumask); 708 } 709 lwkt_schedule(td); 710 if (domain & PWAKEUP_ONE) 711 goto done; 712 goto restart; 713 } 714 } 715 716 #ifdef SMP 717 /* 718 * We finished checking the current cpu but there still may be 719 * more work to do. Either wakeup_one was requested and no matching 720 * thread was found, or a normal wakeup was requested and we have 721 * to continue checking cpus. 722 * 723 * The cpu that started the wakeup sequence is encoded in the domain. 724 * We use this information to determine which cpus still need to be 725 * checked, locate a candidate cpu, and chain the wakeup 726 * asynchronously with an IPI message. 727 * 728 * It should be noted that this scheme is actually less expensive then 729 * the old scheme when waking up multiple threads, since we send 730 * only one IPI message per target candidate which may then schedule 731 * multiple threads. Before we could have wound up sending an IPI 732 * message for each thread on the target cpu (!= current cpu) that 733 * needed to be woken up. 734 * 735 * NOTE: Wakeups occuring on remote cpus are asynchronous. This 736 * should be ok since we are passing idents in the IPI rather then 737 * thread pointers. 738 */ 739 if ((domain & PWAKEUP_MYCPU) == 0 && 740 (mask = slpque_cpumasks[id]) != 0 741 ) { 742 /* 743 * Look for a cpu that might have work to do. Mask out cpus 744 * which have already been processed. 745 * 746 * 31xxxxxxxxxxxxxxxxxxxxxxxxxxxxx0 747 * ^ ^ ^ 748 * start currentcpu start 749 * case2 case1 750 * * * * 751 * 11111111111111110000000000000111 case1 752 * 00000000111111110000000000000000 case2 753 * 754 * case1: We started at start_case1 and processed through 755 * to the current cpu. We have to check any bits 756 * after the current cpu, then check bits before 757 * the starting cpu. 758 * 759 * case2: We have already checked all the bits from 760 * start_case2 to the end, and from 0 to the current 761 * cpu. We just have the bits from the current cpu 762 * to start_case2 left to check. 763 */ 764 startcpu = PWAKEUP_DECODE(domain); 765 if (gd->gd_cpuid >= startcpu) { 766 /* 767 * CASE1 768 */ 769 tmask = mask & ~((gd->gd_cpumask << 1) - 1); 770 if (mask & tmask) { 771 nextcpu = bsfl(mask & tmask); 772 lwkt_send_ipiq2(globaldata_find(nextcpu), 773 _wakeup, ident, domain); 774 } else { 775 tmask = (1 << startcpu) - 1; 776 if (mask & tmask) { 777 nextcpu = bsfl(mask & tmask); 778 lwkt_send_ipiq2( 779 globaldata_find(nextcpu), 780 _wakeup, ident, domain); 781 } 782 } 783 } else { 784 /* 785 * CASE2 786 */ 787 tmask = ~((gd->gd_cpumask << 1) - 1) & 788 ((1 << startcpu) - 1); 789 if (mask & tmask) { 790 nextcpu = bsfl(mask & tmask); 791 lwkt_send_ipiq2(globaldata_find(nextcpu), 792 _wakeup, ident, domain); 793 } 794 } 795 } 796 #endif 797 done: 798 logtsleep(wakeup_end); 799 crit_exit(); 800 } 801 802 /* 803 * Wakeup all threads tsleep()ing on the specified ident, on all cpus 804 */ 805 void 806 wakeup(void *ident) 807 { 808 _wakeup(ident, PWAKEUP_ENCODE(0, mycpu->gd_cpuid)); 809 } 810 811 /* 812 * Wakeup one thread tsleep()ing on the specified ident, on any cpu. 813 */ 814 void 815 wakeup_one(void *ident) 816 { 817 /* XXX potentially round-robin the first responding cpu */ 818 _wakeup(ident, PWAKEUP_ENCODE(0, mycpu->gd_cpuid) | PWAKEUP_ONE); 819 } 820 821 /* 822 * Wakeup threads tsleep()ing on the specified ident on the current cpu 823 * only. 824 */ 825 void 826 wakeup_mycpu(void *ident) 827 { 828 _wakeup(ident, PWAKEUP_MYCPU); 829 } 830 831 /* 832 * Wakeup one thread tsleep()ing on the specified ident on the current cpu 833 * only. 834 */ 835 void 836 wakeup_mycpu_one(void *ident) 837 { 838 /* XXX potentially round-robin the first responding cpu */ 839 _wakeup(ident, PWAKEUP_MYCPU|PWAKEUP_ONE); 840 } 841 842 /* 843 * Wakeup all thread tsleep()ing on the specified ident on the specified cpu 844 * only. 845 */ 846 void 847 wakeup_oncpu(globaldata_t gd, void *ident) 848 { 849 #ifdef SMP 850 if (gd == mycpu) { 851 _wakeup(ident, PWAKEUP_MYCPU); 852 } else { 853 lwkt_send_ipiq2(gd, _wakeup, ident, PWAKEUP_MYCPU); 854 } 855 #else 856 _wakeup(ident, PWAKEUP_MYCPU); 857 #endif 858 } 859 860 /* 861 * Wakeup one thread tsleep()ing on the specified ident on the specified cpu 862 * only. 863 */ 864 void 865 wakeup_oncpu_one(globaldata_t gd, void *ident) 866 { 867 #ifdef SMP 868 if (gd == mycpu) { 869 _wakeup(ident, PWAKEUP_MYCPU | PWAKEUP_ONE); 870 } else { 871 lwkt_send_ipiq2(gd, _wakeup, ident, PWAKEUP_MYCPU | PWAKEUP_ONE); 872 } 873 #else 874 _wakeup(ident, PWAKEUP_MYCPU | PWAKEUP_ONE); 875 #endif 876 } 877 878 /* 879 * Wakeup all threads waiting on the specified ident that slept using 880 * the specified domain, on all cpus. 881 */ 882 void 883 wakeup_domain(void *ident, int domain) 884 { 885 _wakeup(ident, PWAKEUP_ENCODE(domain, mycpu->gd_cpuid)); 886 } 887 888 /* 889 * Wakeup one thread waiting on the specified ident that slept using 890 * the specified domain, on any cpu. 891 */ 892 void 893 wakeup_domain_one(void *ident, int domain) 894 { 895 /* XXX potentially round-robin the first responding cpu */ 896 _wakeup(ident, PWAKEUP_ENCODE(domain, mycpu->gd_cpuid) | PWAKEUP_ONE); 897 } 898 899 /* 900 * setrunnable() 901 * 902 * Make a process runnable. The MP lock must be held on call. This only 903 * has an effect if we are in SSLEEP. We only break out of the 904 * tsleep if LWP_BREAKTSLEEP is set, otherwise we just fix-up the state. 905 * 906 * NOTE: With the MP lock held we can only safely manipulate the process 907 * structure. We cannot safely manipulate the thread structure. 908 */ 909 void 910 setrunnable(struct lwp *lp) 911 { 912 enum lwpstat stat; 913 914 crit_enter(); 915 ASSERT_MP_LOCK_HELD(curthread); 916 stat = lp->lwp_stat; 917 lp->lwp_stat = LSRUN; 918 if (stat == LSSLEEP && (lp->lwp_flag & LWP_BREAKTSLEEP)) 919 unsleep_and_wakeup_thread(lp->lwp_thread); 920 crit_exit(); 921 } 922 923 /* 924 * The process is stopped due to some condition, usually because p_stat is 925 * set to SSTOP, but also possibly due to being traced. 926 * 927 * NOTE! If the caller sets SSTOP, the caller must also clear P_WAITED 928 * because the parent may check the child's status before the child actually 929 * gets to this routine. 930 * 931 * This routine is called with the current lwp only, typically just 932 * before returning to userland. 933 * 934 * Setting LWP_BREAKTSLEEP before entering the tsleep will cause a passive 935 * SIGCONT to break out of the tsleep. 936 */ 937 void 938 tstop(void) 939 { 940 struct lwp *lp = curthread->td_lwp; 941 942 lp->lwp_flag |= LWP_BREAKTSLEEP; 943 tsleep(lp->lwp_proc, 0, "stop", 0); 944 } 945 946 /* 947 * Yield / synchronous reschedule. This is a bit tricky because the trap 948 * code might have set a lazy release on the switch function. Setting 949 * P_PASSIVE_ACQ will ensure that the lazy release executes when we call 950 * switch, and that we are given a greater chance of affinity with our 951 * current cpu. 952 * 953 * We call lwkt_setpri_self() to rotate our thread to the end of the lwkt 954 * run queue. lwkt_switch() will also execute any assigned passive release 955 * (which usually calls release_curproc()), allowing a same/higher priority 956 * process to be designated as the current process. 957 * 958 * While it is possible for a lower priority process to be designated, 959 * it's call to lwkt_maybe_switch() in acquire_curproc() will likely 960 * round-robin back to us and we will be able to re-acquire the current 961 * process designation. 962 */ 963 void 964 uio_yield(void) 965 { 966 struct thread *td = curthread; 967 struct proc *p = td->td_proc; 968 969 lwkt_setpri_self(td->td_pri & TDPRI_MASK); 970 if (p) { 971 p->p_flag |= P_PASSIVE_ACQ; 972 lwkt_switch(); 973 p->p_flag &= ~P_PASSIVE_ACQ; 974 } else { 975 lwkt_switch(); 976 } 977 } 978 979 /* 980 * Compute a tenex style load average of a quantity on 981 * 1, 5 and 15 minute intervals. 982 */ 983 static int loadav_count_runnable(struct lwp *p, void *data); 984 985 static void 986 loadav(void *arg) 987 { 988 struct loadavg *avg; 989 int i, nrun; 990 991 nrun = 0; 992 alllwp_scan(loadav_count_runnable, &nrun); 993 avg = &averunnable; 994 for (i = 0; i < 3; i++) { 995 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] + 996 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT; 997 } 998 999 /* 1000 * Schedule the next update to occur after 5 seconds, but add a 1001 * random variation to avoid synchronisation with processes that 1002 * run at regular intervals. 1003 */ 1004 callout_reset(&loadav_callout, hz * 4 + (int)(krandom() % (hz * 2 + 1)), 1005 loadav, NULL); 1006 } 1007 1008 static int 1009 loadav_count_runnable(struct lwp *lp, void *data) 1010 { 1011 int *nrunp = data; 1012 thread_t td; 1013 1014 switch (lp->lwp_stat) { 1015 case LSRUN: 1016 if ((td = lp->lwp_thread) == NULL) 1017 break; 1018 if (td->td_flags & TDF_BLOCKED) 1019 break; 1020 ++*nrunp; 1021 break; 1022 default: 1023 break; 1024 } 1025 return(0); 1026 } 1027 1028 /* ARGSUSED */ 1029 static void 1030 sched_setup(void *dummy) 1031 { 1032 callout_init(&loadav_callout); 1033 callout_init(&schedcpu_callout); 1034 1035 /* Kick off timeout driven events by calling first time. */ 1036 schedcpu(NULL); 1037 loadav(NULL); 1038 } 1039 1040