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. Neither the name of the University nor the names of its contributors 19 * may be used to endorse or promote products derived from this software 20 * without specific prior written permission. 21 * 22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 * 34 * @(#)kern_synch.c 8.9 (Berkeley) 5/19/95 35 * $FreeBSD: src/sys/kern/kern_synch.c,v 1.87.2.6 2002/10/13 07:29:53 kbyanc Exp $ 36 */ 37 38 #include "opt_ktrace.h" 39 40 #include <sys/param.h> 41 #include <sys/systm.h> 42 #include <sys/proc.h> 43 #include <sys/kernel.h> 44 #include <sys/signalvar.h> 45 #include <sys/resourcevar.h> 46 #include <sys/vmmeter.h> 47 #include <sys/sysctl.h> 48 #include <sys/lock.h> 49 #include <sys/uio.h> 50 #include <sys/priv.h> 51 #include <sys/kcollect.h> 52 #ifdef KTRACE 53 #include <sys/ktrace.h> 54 #endif 55 #include <sys/ktr.h> 56 #include <sys/serialize.h> 57 58 #include <sys/signal2.h> 59 #include <sys/thread2.h> 60 #include <sys/spinlock2.h> 61 #include <sys/mutex2.h> 62 63 #include <machine/cpu.h> 64 #include <machine/smp.h> 65 66 #include <vm/vm_extern.h> 67 68 struct tslpque { 69 TAILQ_HEAD(, thread) queue; 70 const volatile void *ident0; 71 const volatile void *ident1; 72 const volatile void *ident2; 73 const volatile void *ident3; 74 }; 75 76 static void sched_setup (void *dummy); 77 SYSINIT(sched_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, sched_setup, NULL); 78 static void sched_dyninit (void *dummy); 79 SYSINIT(sched_dyninit, SI_BOOT1_DYNALLOC, SI_ORDER_FIRST, sched_dyninit, NULL); 80 81 int lbolt; 82 void *lbolt_syncer; 83 int ncpus; 84 int ncpus_fit, ncpus_fit_mask; /* note: mask not cpumask_t */ 85 int safepri; 86 int tsleep_now_works; 87 int tsleep_crypto_dump = 0; 88 89 MALLOC_DEFINE(M_TSLEEP, "tslpque", "tsleep queues"); 90 91 #define __DEALL(ident) __DEQUALIFY(void *, ident) 92 93 #if !defined(KTR_TSLEEP) 94 #define KTR_TSLEEP KTR_ALL 95 #endif 96 KTR_INFO_MASTER(tsleep); 97 KTR_INFO(KTR_TSLEEP, tsleep, tsleep_beg, 0, "tsleep enter %p", const volatile void *ident); 98 KTR_INFO(KTR_TSLEEP, tsleep, tsleep_end, 1, "tsleep exit"); 99 KTR_INFO(KTR_TSLEEP, tsleep, wakeup_beg, 2, "wakeup enter %p", const volatile void *ident); 100 KTR_INFO(KTR_TSLEEP, tsleep, wakeup_end, 3, "wakeup exit"); 101 KTR_INFO(KTR_TSLEEP, tsleep, ilockfail, 4, "interlock failed %p", const volatile void *ident); 102 103 #define logtsleep1(name) KTR_LOG(tsleep_ ## name) 104 #define logtsleep2(name, val) KTR_LOG(tsleep_ ## name, val) 105 106 struct loadavg averunnable = 107 { {0, 0, 0}, FSCALE }; /* load average, of runnable procs */ 108 /* 109 * Constants for averages over 1, 5, and 15 minutes 110 * when sampling at 5 second intervals. 111 */ 112 static fixpt_t cexp[3] = { 113 0.9200444146293232 * FSCALE, /* exp(-1/12) */ 114 0.9834714538216174 * FSCALE, /* exp(-1/60) */ 115 0.9944598480048967 * FSCALE, /* exp(-1/180) */ 116 }; 117 118 static void endtsleep (void *); 119 static void loadav (void *arg); 120 static void schedcpu (void *arg); 121 122 static int pctcpu_decay = 10; 123 SYSCTL_INT(_kern, OID_AUTO, pctcpu_decay, CTLFLAG_RW, 124 &pctcpu_decay, 0, ""); 125 126 /* 127 * kernel uses `FSCALE', userland (SHOULD) use kern.fscale 128 */ 129 int fscale __unused = FSCALE; /* exported to systat */ 130 SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, ""); 131 132 /* 133 * Issue a wakeup() from userland (debugging) 134 */ 135 static int 136 sysctl_wakeup(SYSCTL_HANDLER_ARGS) 137 { 138 uint64_t ident = 1; 139 int error = 0; 140 141 if (req->newptr != NULL) { 142 if (priv_check(curthread, PRIV_ROOT)) 143 return (EPERM); 144 error = SYSCTL_IN(req, &ident, sizeof(ident)); 145 if (error) 146 return error; 147 kprintf("issue wakeup %016jx\n", ident); 148 wakeup((void *)(intptr_t)ident); 149 } 150 if (req->oldptr != NULL) { 151 error = SYSCTL_OUT(req, &ident, sizeof(ident)); 152 } 153 return error; 154 } 155 156 static int 157 sysctl_wakeup_umtx(SYSCTL_HANDLER_ARGS) 158 { 159 uint64_t ident = 1; 160 int error = 0; 161 162 if (req->newptr != NULL) { 163 if (priv_check(curthread, PRIV_ROOT)) 164 return (EPERM); 165 error = SYSCTL_IN(req, &ident, sizeof(ident)); 166 if (error) 167 return error; 168 kprintf("issue wakeup %016jx, PDOMAIN_UMTX\n", ident); 169 wakeup_domain((void *)(intptr_t)ident, PDOMAIN_UMTX); 170 } 171 if (req->oldptr != NULL) { 172 error = SYSCTL_OUT(req, &ident, sizeof(ident)); 173 } 174 return error; 175 } 176 177 SYSCTL_PROC(_debug, OID_AUTO, wakeup, CTLTYPE_UQUAD|CTLFLAG_RW, 0, 0, 178 sysctl_wakeup, "Q", "issue wakeup(addr)"); 179 SYSCTL_PROC(_debug, OID_AUTO, wakeup_umtx, CTLTYPE_UQUAD|CTLFLAG_RW, 0, 0, 180 sysctl_wakeup_umtx, "Q", "issue wakeup(addr, PDOMAIN_UMTX)"); 181 182 /* 183 * Recompute process priorities, once a second. 184 * 185 * Since the userland schedulers are typically event oriented, if the 186 * estcpu calculation at wakeup() time is not sufficient to make a 187 * process runnable relative to other processes in the system we have 188 * a 1-second recalc to help out. 189 * 190 * This code also allows us to store sysclock_t data in the process structure 191 * without fear of an overrun, since sysclock_t are guarenteed to hold 192 * several seconds worth of count. 193 * 194 * WARNING! callouts can preempt normal threads. However, they will not 195 * preempt a thread holding a spinlock so we *can* safely use spinlocks. 196 */ 197 static int schedcpu_stats(struct proc *p, void *data __unused); 198 static int schedcpu_resource(struct proc *p, void *data __unused); 199 200 static void 201 schedcpu(void *arg) 202 { 203 allproc_scan(schedcpu_stats, NULL, 1); 204 allproc_scan(schedcpu_resource, NULL, 1); 205 if (mycpu->gd_cpuid == 0) { 206 wakeup((caddr_t)&lbolt); 207 wakeup(lbolt_syncer); 208 } 209 callout_reset(&mycpu->gd_schedcpu_callout, hz, schedcpu, NULL); 210 } 211 212 /* 213 * General process statistics once a second 214 */ 215 static int 216 schedcpu_stats(struct proc *p, void *data __unused) 217 { 218 struct lwp *lp; 219 220 /* 221 * Threads may not be completely set up if process in SIDL state. 222 */ 223 if (p->p_stat == SIDL) 224 return(0); 225 226 PHOLD(p); 227 if (lwkt_trytoken(&p->p_token) == FALSE) { 228 PRELE(p); 229 return(0); 230 } 231 232 p->p_swtime++; 233 FOREACH_LWP_IN_PROC(lp, p) { 234 if (lp->lwp_stat == LSSLEEP) { 235 ++lp->lwp_slptime; 236 if (lp->lwp_slptime == 1) 237 p->p_usched->uload_update(lp); 238 } 239 240 /* 241 * Only recalculate processes that are active or have slept 242 * less then 2 seconds. The schedulers understand this. 243 * Otherwise decay by 50% per second. 244 */ 245 if (lp->lwp_slptime <= 1) { 246 p->p_usched->recalculate(lp); 247 } else { 248 int decay; 249 250 decay = pctcpu_decay; 251 cpu_ccfence(); 252 if (decay <= 1) 253 decay = 1; 254 if (decay > 100) 255 decay = 100; 256 lp->lwp_pctcpu = (lp->lwp_pctcpu * (decay - 1)) / decay; 257 } 258 } 259 lwkt_reltoken(&p->p_token); 260 lwkt_yield(); 261 PRELE(p); 262 return(0); 263 } 264 265 /* 266 * Resource checks. XXX break out since ksignal/killproc can block, 267 * limiting us to one process killed per second. There is probably 268 * a better way. 269 */ 270 static int 271 schedcpu_resource(struct proc *p, void *data __unused) 272 { 273 u_int64_t ttime; 274 struct lwp *lp; 275 276 if (p->p_stat == SIDL) 277 return(0); 278 279 PHOLD(p); 280 if (lwkt_trytoken(&p->p_token) == FALSE) { 281 PRELE(p); 282 return(0); 283 } 284 285 if (p->p_stat == SZOMB || p->p_limit == NULL) { 286 lwkt_reltoken(&p->p_token); 287 PRELE(p); 288 return(0); 289 } 290 291 ttime = 0; 292 FOREACH_LWP_IN_PROC(lp, p) { 293 /* 294 * We may have caught an lp in the middle of being 295 * created, lwp_thread can be NULL. 296 */ 297 if (lp->lwp_thread) { 298 ttime += lp->lwp_thread->td_sticks; 299 ttime += lp->lwp_thread->td_uticks; 300 } 301 } 302 303 switch(plimit_testcpulimit(p, ttime)) { 304 case PLIMIT_TESTCPU_KILL: 305 killproc(p, "exceeded maximum CPU limit"); 306 break; 307 case PLIMIT_TESTCPU_XCPU: 308 if ((p->p_flags & P_XCPU) == 0) { 309 p->p_flags |= P_XCPU; 310 ksignal(p, SIGXCPU); 311 } 312 break; 313 default: 314 break; 315 } 316 lwkt_reltoken(&p->p_token); 317 lwkt_yield(); 318 PRELE(p); 319 return(0); 320 } 321 322 /* 323 * This is only used by ps. Generate a cpu percentage use over 324 * a period of one second. 325 */ 326 void 327 updatepcpu(struct lwp *lp, int cpticks, int ttlticks) 328 { 329 fixpt_t acc; 330 int remticks; 331 332 acc = (cpticks << FSHIFT) / ttlticks; 333 if (ttlticks >= ESTCPUFREQ) { 334 lp->lwp_pctcpu = acc; 335 } else { 336 remticks = ESTCPUFREQ - ttlticks; 337 lp->lwp_pctcpu = (acc * ttlticks + lp->lwp_pctcpu * remticks) / 338 ESTCPUFREQ; 339 } 340 } 341 342 /* 343 * Handy macros to calculate hash indices. LOOKUP() calculates the 344 * global cpumask hash index, TCHASHSHIFT() converts that into the 345 * pcpu hash index. 346 * 347 * By making the pcpu hash arrays smaller we save a significant amount 348 * of memory at very low cost. The real cost is in IPIs, which are handled 349 * by the much larger global cpumask hash table. 350 */ 351 #define LOOKUP_PRIME 66555444443333333ULL 352 #define LOOKUP(x) ((((uintptr_t)(x) + ((uintptr_t)(x) >> 18)) ^ \ 353 LOOKUP_PRIME) % slpque_tablesize) 354 #define TCHASHSHIFT(x) ((x) >> 4) 355 356 static uint32_t slpque_tablesize; 357 static cpumask_t *slpque_cpumasks; 358 359 SYSCTL_UINT(_kern, OID_AUTO, slpque_tablesize, CTLFLAG_RD, &slpque_tablesize, 360 0, ""); 361 362 /* 363 * This is a dandy function that allows us to interlock tsleep/wakeup 364 * operations with unspecified upper level locks, such as lockmgr locks, 365 * simply by holding a critical section. The sequence is: 366 * 367 * (acquire upper level lock) 368 * tsleep_interlock(blah) 369 * (release upper level lock) 370 * tsleep(blah, ...) 371 * 372 * Basically this functions queues us on the tsleep queue without actually 373 * descheduling us. When tsleep() is later called with PINTERLOCK it 374 * assumes the thread was already queued, otherwise it queues it there. 375 * 376 * Thus it is possible to receive the wakeup prior to going to sleep and 377 * the race conditions are covered. 378 */ 379 static __inline void 380 _tsleep_interlock(globaldata_t gd, const volatile void *ident, int flags) 381 { 382 thread_t td = gd->gd_curthread; 383 struct tslpque *qp; 384 uint32_t cid; 385 uint32_t gid; 386 387 if (ident == NULL) { 388 kprintf("tsleep_interlock: NULL ident %s\n", td->td_comm); 389 print_backtrace(5); 390 } 391 392 crit_enter_quick(td); 393 if (td->td_flags & TDF_TSLEEPQ) { 394 /* 395 * Shortcut if unchanged 396 */ 397 if (td->td_wchan == ident && 398 td->td_wdomain == (flags & PDOMAIN_MASK)) { 399 crit_exit_quick(td); 400 return; 401 } 402 403 /* 404 * Remove current sleepq 405 */ 406 cid = LOOKUP(td->td_wchan); 407 gid = TCHASHSHIFT(cid); 408 qp = &gd->gd_tsleep_hash[gid]; 409 TAILQ_REMOVE(&qp->queue, td, td_sleepq); 410 if (TAILQ_FIRST(&qp->queue) == NULL) { 411 qp->ident0 = NULL; 412 qp->ident1 = NULL; 413 qp->ident2 = NULL; 414 qp->ident3 = NULL; 415 ATOMIC_CPUMASK_NANDBIT(slpque_cpumasks[cid], 416 gd->gd_cpuid); 417 } 418 } else { 419 td->td_flags |= TDF_TSLEEPQ; 420 } 421 cid = LOOKUP(ident); 422 gid = TCHASHSHIFT(cid); 423 qp = &gd->gd_tsleep_hash[gid]; 424 TAILQ_INSERT_TAIL(&qp->queue, td, td_sleepq); 425 if (qp->ident0 != ident && qp->ident1 != ident && 426 qp->ident2 != ident && qp->ident3 != ident) { 427 if (qp->ident0 == NULL) 428 qp->ident0 = ident; 429 else if (qp->ident1 == NULL) 430 qp->ident1 = ident; 431 else if (qp->ident2 == NULL) 432 qp->ident2 = ident; 433 else if (qp->ident3 == NULL) 434 qp->ident3 = ident; 435 else 436 qp->ident0 = (void *)(intptr_t)-1; 437 } 438 ATOMIC_CPUMASK_ORBIT(slpque_cpumasks[cid], gd->gd_cpuid); 439 td->td_wchan = ident; 440 td->td_wdomain = flags & PDOMAIN_MASK; 441 crit_exit_quick(td); 442 } 443 444 void 445 tsleep_interlock(const volatile void *ident, int flags) 446 { 447 _tsleep_interlock(mycpu, ident, flags); 448 } 449 450 /* 451 * Remove thread from sleepq. Must be called with a critical section held. 452 * The thread must not be migrating. 453 */ 454 static __inline void 455 _tsleep_remove(thread_t td) 456 { 457 globaldata_t gd = mycpu; 458 struct tslpque *qp; 459 uint32_t cid; 460 uint32_t gid; 461 462 KKASSERT(td->td_gd == gd && IN_CRITICAL_SECT(td)); 463 KKASSERT((td->td_flags & TDF_MIGRATING) == 0); 464 if (td->td_flags & TDF_TSLEEPQ) { 465 td->td_flags &= ~TDF_TSLEEPQ; 466 cid = LOOKUP(td->td_wchan); 467 gid = TCHASHSHIFT(cid); 468 qp = &gd->gd_tsleep_hash[gid]; 469 TAILQ_REMOVE(&qp->queue, td, td_sleepq); 470 if (TAILQ_FIRST(&qp->queue) == NULL) { 471 ATOMIC_CPUMASK_NANDBIT(slpque_cpumasks[cid], 472 gd->gd_cpuid); 473 } 474 td->td_wchan = NULL; 475 td->td_wdomain = 0; 476 } 477 } 478 479 void 480 tsleep_remove(thread_t td) 481 { 482 _tsleep_remove(td); 483 } 484 485 /* 486 * General sleep call. Suspends the current process until a wakeup is 487 * performed on the specified identifier. The process will then be made 488 * runnable with the specified priority. Sleeps at most timo/hz seconds 489 * (0 means no timeout). If flags includes PCATCH flag, signals are checked 490 * before and after sleeping, else signals are not checked. Returns 0 if 491 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a 492 * signal needs to be delivered, ERESTART is returned if the current system 493 * call should be restarted if possible, and EINTR is returned if the system 494 * call should be interrupted by the signal (return EINTR). 495 * 496 * Note that if we are a process, we release_curproc() before messing with 497 * the LWKT scheduler. 498 * 499 * During autoconfiguration or after a panic, a sleep will simply 500 * lower the priority briefly to allow interrupts, then return. 501 * 502 * WARNING! This code can't block (short of switching away), or bad things 503 * will happen. No getting tokens, no blocking locks, etc. 504 */ 505 int 506 tsleep(const volatile void *ident, int flags, const char *wmesg, int timo) 507 { 508 struct thread *td = curthread; 509 struct lwp *lp = td->td_lwp; 510 struct proc *p = td->td_proc; /* may be NULL */ 511 globaldata_t gd; 512 int sig; 513 int catch; 514 int error; 515 int oldpri; 516 struct callout thandle; 517 518 /* 519 * Currently a severe hack. Make sure any delayed wakeups 520 * are flushed before we sleep or we might deadlock on whatever 521 * event we are sleeping on. 522 */ 523 if (td->td_flags & TDF_DELAYED_WAKEUP) 524 wakeup_end_delayed(); 525 526 /* 527 * NOTE: removed KTRPOINT, it could cause races due to blocking 528 * even in stable. Just scrap it for now. 529 */ 530 if (!tsleep_crypto_dump && (tsleep_now_works == 0 || panicstr)) { 531 /* 532 * After a panic, or before we actually have an operational 533 * softclock, just give interrupts a chance, then just return; 534 * 535 * don't run any other procs or panic below, 536 * in case this is the idle process and already asleep. 537 */ 538 splz(); 539 oldpri = td->td_pri; 540 lwkt_setpri_self(safepri); 541 lwkt_switch(); 542 lwkt_setpri_self(oldpri); 543 return (0); 544 } 545 logtsleep2(tsleep_beg, ident); 546 gd = td->td_gd; 547 KKASSERT(td != &gd->gd_idlethread); /* you must be kidding! */ 548 td->td_wakefromcpu = -1; /* overwritten by _wakeup */ 549 550 /* 551 * NOTE: all of this occurs on the current cpu, including any 552 * callout-based wakeups, so a critical section is a sufficient 553 * interlock. 554 * 555 * The entire sequence through to where we actually sleep must 556 * run without breaking the critical section. 557 */ 558 catch = flags & PCATCH; 559 error = 0; 560 sig = 0; 561 562 crit_enter_quick(td); 563 564 KASSERT(ident != NULL, ("tsleep: no ident")); 565 KASSERT(lp == NULL || 566 lp->lwp_stat == LSRUN || /* Obvious */ 567 lp->lwp_stat == LSSTOP, /* Set in tstop */ 568 ("tsleep %p %s %d", 569 ident, wmesg, lp->lwp_stat)); 570 571 /* 572 * We interlock the sleep queue if the caller has not already done 573 * it for us. This must be done before we potentially acquire any 574 * tokens or we can loose the wakeup. 575 */ 576 if ((flags & PINTERLOCKED) == 0) { 577 _tsleep_interlock(gd, ident, flags); 578 } 579 580 /* 581 * Setup for the current process (if this is a process). We must 582 * interlock with lwp_token to avoid remote wakeup races via 583 * setrunnable() 584 */ 585 if (lp) { 586 lwkt_gettoken(&lp->lwp_token); 587 588 /* 589 * If the umbrella process is in the SCORE state then 590 * make sure that the thread is flagged going into a 591 * normal sleep to allow the core dump to proceed, otherwise 592 * the coredump can end up waiting forever. If the normal 593 * sleep is woken up, the thread will enter a stopped state 594 * upon return to userland. 595 * 596 * We do not want to interrupt or cause a thread exist at 597 * this juncture because that will mess-up the state the 598 * coredump is trying to save. 599 */ 600 if (p->p_stat == SCORE) { 601 lwkt_gettoken(&p->p_token); 602 if ((lp->lwp_mpflags & LWP_MP_WSTOP) == 0) { 603 atomic_set_int(&lp->lwp_mpflags, LWP_MP_WSTOP); 604 ++p->p_nstopped; 605 } 606 lwkt_reltoken(&p->p_token); 607 } 608 609 /* 610 * PCATCH requested. 611 */ 612 if (catch) { 613 /* 614 * Early termination if PCATCH was set and a 615 * signal is pending, interlocked with the 616 * critical section. 617 * 618 * Early termination only occurs when tsleep() is 619 * entered while in a normal LSRUN state. 620 */ 621 if ((sig = CURSIG(lp)) != 0) 622 goto resume; 623 624 /* 625 * Causes ksignal to wake us up if a signal is 626 * received (interlocked with lp->lwp_token). 627 */ 628 lp->lwp_flags |= LWP_SINTR; 629 } 630 } else { 631 KKASSERT(p == NULL); 632 } 633 634 /* 635 * Make sure the current process has been untangled from 636 * the userland scheduler and initialize slptime to start 637 * counting. 638 * 639 * NOTE: td->td_wakefromcpu is pre-set by the release function 640 * for the dfly scheduler, and then adjusted by _wakeup() 641 */ 642 if (lp) { 643 p->p_usched->release_curproc(lp); 644 lp->lwp_slptime = 0; 645 } 646 647 /* 648 * For PINTERLOCKED operation, TDF_TSLEEPQ might not be set if 649 * a wakeup() was processed before the thread could go to sleep. 650 * 651 * If TDF_TSLEEPQ is set, make sure the ident matches the recorded 652 * ident. If it does not then the thread slept inbetween the 653 * caller's initial tsleep_interlock() call and the caller's tsleep() 654 * call. 655 * 656 * Extreme loads can cause the sending of an IPI (e.g. wakeup()'s) 657 * to process incoming IPIs, thus draining incoming wakeups. 658 */ 659 if ((td->td_flags & TDF_TSLEEPQ) == 0) { 660 logtsleep2(ilockfail, ident); 661 goto resume; 662 } else if (td->td_wchan != ident || 663 td->td_wdomain != (flags & PDOMAIN_MASK)) { 664 logtsleep2(ilockfail, ident); 665 goto resume; 666 } 667 668 /* 669 * scheduling is blocked while in a critical section. Coincide 670 * the descheduled-by-tsleep flag with the descheduling of the 671 * lwkt. 672 * 673 * The timer callout is localized on our cpu and interlocked by 674 * our critical section. 675 */ 676 lwkt_deschedule_self(td); 677 td->td_flags |= TDF_TSLEEP_DESCHEDULED; 678 td->td_wmesg = wmesg; 679 680 /* 681 * Setup the timeout, if any. The timeout is only operable while 682 * the thread is flagged descheduled. 683 */ 684 KKASSERT((td->td_flags & TDF_TIMEOUT) == 0); 685 if (timo) { 686 callout_init_mp(&thandle); 687 callout_reset(&thandle, timo, endtsleep, td); 688 } 689 690 /* 691 * Beddy bye bye. 692 */ 693 if (lp) { 694 /* 695 * Ok, we are sleeping. Place us in the SSLEEP state. 696 */ 697 KKASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) == 0); 698 699 /* 700 * tstop() sets LSSTOP, so don't fiddle with that. 701 */ 702 if (lp->lwp_stat != LSSTOP) 703 lp->lwp_stat = LSSLEEP; 704 lp->lwp_ru.ru_nvcsw++; 705 p->p_usched->uload_update(lp); 706 lwkt_switch(); 707 708 /* 709 * And when we are woken up, put us back in LSRUN. If we 710 * slept for over a second, recalculate our estcpu. 711 */ 712 lp->lwp_stat = LSRUN; 713 if (lp->lwp_slptime) { 714 p->p_usched->uload_update(lp); 715 p->p_usched->recalculate(lp); 716 } 717 lp->lwp_slptime = 0; 718 } else { 719 lwkt_switch(); 720 } 721 722 /* 723 * Make sure we haven't switched cpus while we were asleep. It's 724 * not supposed to happen. Cleanup our temporary flags. 725 */ 726 KKASSERT(gd == td->td_gd); 727 728 /* 729 * Cleanup the timeout. If the timeout has already occured thandle 730 * has already been stopped, otherwise stop thandle. If the timeout 731 * is running (the callout thread must be blocked trying to get 732 * lwp_token) then wait for us to get scheduled. 733 */ 734 if (timo) { 735 while (td->td_flags & TDF_TIMEOUT_RUNNING) { 736 /* else we won't get rescheduled! */ 737 if (lp->lwp_stat != LSSTOP) 738 lp->lwp_stat = LSSLEEP; 739 lwkt_deschedule_self(td); 740 td->td_wmesg = "tsrace"; 741 lwkt_switch(); 742 kprintf("td %p %s: timeout race\n", td, td->td_comm); 743 } 744 if (td->td_flags & TDF_TIMEOUT) { 745 td->td_flags &= ~TDF_TIMEOUT; 746 error = EWOULDBLOCK; 747 } else { 748 /* does not block when on same cpu */ 749 callout_stop(&thandle); 750 } 751 } 752 td->td_flags &= ~TDF_TSLEEP_DESCHEDULED; 753 754 /* 755 * Make sure we have been removed from the sleepq. In most 756 * cases this will have been done for us already but it is 757 * possible for a scheduling IPI to be in-flight from a 758 * previous tsleep/tsleep_interlock() or due to a straight-out 759 * call to lwkt_schedule() (in the case of an interrupt thread), 760 * causing a spurious wakeup. 761 */ 762 _tsleep_remove(td); 763 td->td_wmesg = NULL; 764 765 /* 766 * Figure out the correct error return. If interrupted by a 767 * signal we want to return EINTR or ERESTART. 768 */ 769 resume: 770 if (lp) { 771 if (catch && error == 0) { 772 if (sig != 0 || (sig = CURSIG(lp))) { 773 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 774 error = EINTR; 775 else 776 error = ERESTART; 777 } 778 } 779 780 lp->lwp_flags &= ~LWP_SINTR; 781 782 /* 783 * Unconditionally set us to LSRUN on resume. lwp_stat could 784 * be in a weird state due to the goto resume, particularly 785 * when tsleep() is called from tstop(). 786 */ 787 lp->lwp_stat = LSRUN; 788 lwkt_reltoken(&lp->lwp_token); 789 } 790 logtsleep1(tsleep_end); 791 crit_exit_quick(td); 792 793 return (error); 794 } 795 796 /* 797 * Interlocked spinlock sleep. An exclusively held spinlock must 798 * be passed to ssleep(). The function will atomically release the 799 * spinlock and tsleep on the ident, then reacquire the spinlock and 800 * return. 801 * 802 * This routine is fairly important along the critical path, so optimize it 803 * heavily. 804 */ 805 int 806 ssleep(const volatile void *ident, struct spinlock *spin, int flags, 807 const char *wmesg, int timo) 808 { 809 globaldata_t gd = mycpu; 810 int error; 811 812 _tsleep_interlock(gd, ident, flags); 813 spin_unlock_quick(gd, spin); 814 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 815 KKASSERT(gd == mycpu); 816 _spin_lock_quick(gd, spin, wmesg); 817 818 return (error); 819 } 820 821 int 822 lksleep(const volatile void *ident, struct lock *lock, int flags, 823 const char *wmesg, int timo) 824 { 825 globaldata_t gd = mycpu; 826 int error; 827 828 _tsleep_interlock(gd, ident, flags); 829 lockmgr(lock, LK_RELEASE); 830 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 831 lockmgr(lock, LK_EXCLUSIVE); 832 833 return (error); 834 } 835 836 /* 837 * Interlocked mutex sleep. An exclusively held mutex must be passed 838 * to mtxsleep(). The function will atomically release the mutex 839 * and tsleep on the ident, then reacquire the mutex and return. 840 */ 841 int 842 mtxsleep(const volatile void *ident, struct mtx *mtx, int flags, 843 const char *wmesg, int timo) 844 { 845 globaldata_t gd = mycpu; 846 int error; 847 848 _tsleep_interlock(gd, ident, flags); 849 mtx_unlock(mtx); 850 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 851 mtx_lock_ex_quick(mtx); 852 853 return (error); 854 } 855 856 /* 857 * Interlocked serializer sleep. An exclusively held serializer must 858 * be passed to zsleep(). The function will atomically release 859 * the serializer and tsleep on the ident, then reacquire the serializer 860 * and return. 861 */ 862 int 863 zsleep(const volatile void *ident, struct lwkt_serialize *slz, int flags, 864 const char *wmesg, int timo) 865 { 866 globaldata_t gd = mycpu; 867 int ret; 868 869 ASSERT_SERIALIZED(slz); 870 871 _tsleep_interlock(gd, ident, flags); 872 lwkt_serialize_exit(slz); 873 ret = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 874 lwkt_serialize_enter(slz); 875 876 return ret; 877 } 878 879 /* 880 * Directly block on the LWKT thread by descheduling it. This 881 * is much faster then tsleep(), but the only legal way to wake 882 * us up is to directly schedule the thread. 883 * 884 * Setting TDF_SINTR will cause new signals to directly schedule us. 885 * 886 * This routine must be called while in a critical section. 887 */ 888 int 889 lwkt_sleep(const char *wmesg, int flags) 890 { 891 thread_t td = curthread; 892 int sig; 893 894 if ((flags & PCATCH) == 0 || td->td_lwp == NULL) { 895 td->td_flags |= TDF_BLOCKED; 896 td->td_wmesg = wmesg; 897 lwkt_deschedule_self(td); 898 lwkt_switch(); 899 td->td_wmesg = NULL; 900 td->td_flags &= ~TDF_BLOCKED; 901 return(0); 902 } 903 if ((sig = CURSIG(td->td_lwp)) != 0) { 904 if (SIGISMEMBER(td->td_proc->p_sigacts->ps_sigintr, sig)) 905 return(EINTR); 906 else 907 return(ERESTART); 908 909 } 910 td->td_flags |= TDF_BLOCKED | TDF_SINTR; 911 td->td_wmesg = wmesg; 912 lwkt_deschedule_self(td); 913 lwkt_switch(); 914 td->td_flags &= ~(TDF_BLOCKED | TDF_SINTR); 915 td->td_wmesg = NULL; 916 return(0); 917 } 918 919 /* 920 * Implement the timeout for tsleep. 921 * 922 * This type of callout timeout is scheduled on the same cpu the process 923 * is sleeping on. Also, at the moment, the MP lock is held. 924 */ 925 static void 926 endtsleep(void *arg) 927 { 928 thread_t td = arg; 929 struct lwp *lp; 930 931 /* 932 * We are going to have to get the lwp_token, which means we might 933 * block. This can race a tsleep getting woken up by other means 934 * so set TDF_TIMEOUT_RUNNING to force the tsleep to wait for our 935 * processing to complete (sorry tsleep!). 936 * 937 * We can safely set td_flags because td MUST be on the same cpu 938 * as we are. 939 */ 940 KKASSERT(td->td_gd == mycpu); 941 crit_enter(); 942 td->td_flags |= TDF_TIMEOUT_RUNNING | TDF_TIMEOUT; 943 944 /* 945 * This can block but TDF_TIMEOUT_RUNNING will prevent the thread 946 * from exiting the tsleep on us. The flag is interlocked by virtue 947 * of lp being on the same cpu as we are. 948 */ 949 if ((lp = td->td_lwp) != NULL) 950 lwkt_gettoken(&lp->lwp_token); 951 952 KKASSERT(td->td_flags & TDF_TSLEEP_DESCHEDULED); 953 954 if (lp) { 955 /* 956 * callout timer should normally never be set in tstop() 957 * because it passes a timeout of 0. However, there is a 958 * case during thread exit (which SSTOP's all the threads) 959 * for which tstop() must break out and can (properly) leave 960 * the thread in LSSTOP. 961 */ 962 KKASSERT(lp->lwp_stat != LSSTOP || 963 (lp->lwp_mpflags & LWP_MP_WEXIT)); 964 setrunnable(lp); 965 lwkt_reltoken(&lp->lwp_token); 966 } else { 967 _tsleep_remove(td); 968 lwkt_schedule(td); 969 } 970 KKASSERT(td->td_gd == mycpu); 971 td->td_flags &= ~TDF_TIMEOUT_RUNNING; 972 crit_exit(); 973 } 974 975 /* 976 * Make all processes sleeping on the specified identifier runnable. 977 * count may be zero or one only. 978 * 979 * The domain encodes the sleep/wakeup domain, flags, plus the originating 980 * cpu. 981 * 982 * This call may run without the MP lock held. We can only manipulate thread 983 * state on the cpu owning the thread. We CANNOT manipulate process state 984 * at all. 985 * 986 * _wakeup() can be passed to an IPI so we can't use (const volatile 987 * void *ident). 988 */ 989 static void 990 _wakeup(void *ident, int domain) 991 { 992 struct tslpque *qp; 993 struct thread *td; 994 struct thread *ntd; 995 globaldata_t gd; 996 cpumask_t mask; 997 uint32_t cid; 998 uint32_t gid; 999 int wids = 0; 1000 1001 crit_enter(); 1002 logtsleep2(wakeup_beg, ident); 1003 gd = mycpu; 1004 cid = LOOKUP(ident); 1005 gid = TCHASHSHIFT(cid); 1006 qp = &gd->gd_tsleep_hash[gid]; 1007 restart: 1008 for (td = TAILQ_FIRST(&qp->queue); td != NULL; td = ntd) { 1009 ntd = TAILQ_NEXT(td, td_sleepq); 1010 if (td->td_wchan == ident && 1011 td->td_wdomain == (domain & PDOMAIN_MASK) 1012 ) { 1013 KKASSERT(td->td_gd == gd); 1014 _tsleep_remove(td); 1015 td->td_wakefromcpu = PWAKEUP_DECODE(domain); 1016 if (td->td_flags & TDF_TSLEEP_DESCHEDULED) { 1017 lwkt_schedule(td); 1018 if (domain & PWAKEUP_ONE) 1019 goto done; 1020 } 1021 goto restart; 1022 } 1023 if (td->td_wchan == qp->ident0) 1024 wids |= 1; 1025 else if (td->td_wchan == qp->ident1) 1026 wids |= 2; 1027 else if (td->td_wchan == qp->ident2) 1028 wids |= 4; 1029 else if (td->td_wchan == qp->ident3) 1030 wids |= 8; 1031 else 1032 wids |= 16; /* force ident0 to be retained (-1) */ 1033 } 1034 1035 /* 1036 * Because a bunch of cpumask array entries cover the same queue, it 1037 * is possible for our bit to remain set in some of them and cause 1038 * spurious wakeup IPIs later on. Make sure that the bit is cleared 1039 * when a spurious IPI occurs to prevent further spurious IPIs. 1040 */ 1041 if (TAILQ_FIRST(&qp->queue) == NULL) { 1042 ATOMIC_CPUMASK_NANDBIT(slpque_cpumasks[cid], gd->gd_cpuid); 1043 qp->ident0 = NULL; 1044 qp->ident1 = NULL; 1045 qp->ident2 = NULL; 1046 qp->ident3 = NULL; 1047 } else { 1048 if ((wids & 1) == 0) { 1049 if ((wids & 16) == 0) { 1050 qp->ident0 = NULL; 1051 } else { 1052 KKASSERT(qp->ident0 == (void *)(intptr_t)-1); 1053 } 1054 } 1055 if ((wids & 2) == 0) 1056 qp->ident1 = NULL; 1057 if ((wids & 4) == 0) 1058 qp->ident2 = NULL; 1059 if ((wids & 8) == 0) 1060 qp->ident3 = NULL; 1061 } 1062 1063 /* 1064 * We finished checking the current cpu but there still may be 1065 * more work to do. Either wakeup_one was requested and no matching 1066 * thread was found, or a normal wakeup was requested and we have 1067 * to continue checking cpus. 1068 * 1069 * It should be noted that this scheme is actually less expensive then 1070 * the old scheme when waking up multiple threads, since we send 1071 * only one IPI message per target candidate which may then schedule 1072 * multiple threads. Before we could have wound up sending an IPI 1073 * message for each thread on the target cpu (!= current cpu) that 1074 * needed to be woken up. 1075 * 1076 * NOTE: Wakeups occuring on remote cpus are asynchronous. This 1077 * should be ok since we are passing idents in the IPI rather 1078 * then thread pointers. 1079 * 1080 * NOTE: We MUST mfence (or use an atomic op) prior to reading 1081 * the cpumask, as another cpu may have written to it in 1082 * a fashion interlocked with whatever the caller did before 1083 * calling wakeup(). Otherwise we might miss the interaction 1084 * (kern_mutex.c can cause this problem). 1085 * 1086 * lfence is insufficient as it may allow a written state to 1087 * reorder around the cpumask load. 1088 */ 1089 if ((domain & PWAKEUP_MYCPU) == 0) { 1090 globaldata_t tgd; 1091 const volatile void *id0; 1092 int n; 1093 1094 cpu_mfence(); 1095 /* cpu_lfence(); */ 1096 mask = slpque_cpumasks[cid]; 1097 CPUMASK_ANDMASK(mask, gd->gd_other_cpus); 1098 while (CPUMASK_TESTNZERO(mask)) { 1099 n = BSRCPUMASK(mask); 1100 CPUMASK_NANDBIT(mask, n); 1101 tgd = globaldata_find(n); 1102 1103 /* 1104 * Both ident0 compares must from a single load 1105 * to avoid ident0 update races crossing the two 1106 * compares. 1107 */ 1108 qp = &tgd->gd_tsleep_hash[gid]; 1109 id0 = qp->ident0; 1110 cpu_ccfence(); 1111 if (id0 == (void *)(intptr_t)-1) { 1112 lwkt_send_ipiq2(tgd, _wakeup, ident, 1113 domain | PWAKEUP_MYCPU); 1114 ++tgd->gd_cnt.v_wakeup_colls; 1115 } else if (id0 == ident || 1116 qp->ident1 == ident || 1117 qp->ident2 == ident || 1118 qp->ident3 == ident) { 1119 lwkt_send_ipiq2(tgd, _wakeup, ident, 1120 domain | PWAKEUP_MYCPU); 1121 } 1122 } 1123 #if 0 1124 if (CPUMASK_TESTNZERO(mask)) { 1125 lwkt_send_ipiq2_mask(mask, _wakeup, ident, 1126 domain | PWAKEUP_MYCPU); 1127 } 1128 #endif 1129 } 1130 done: 1131 logtsleep1(wakeup_end); 1132 crit_exit(); 1133 } 1134 1135 /* 1136 * Wakeup all threads tsleep()ing on the specified ident, on all cpus 1137 */ 1138 void 1139 wakeup(const volatile void *ident) 1140 { 1141 globaldata_t gd = mycpu; 1142 thread_t td = gd->gd_curthread; 1143 1144 if (td && (td->td_flags & TDF_DELAYED_WAKEUP)) { 1145 /* 1146 * If we are in a delayed wakeup section, record up to two wakeups in 1147 * a per-CPU queue and issue them when we block or exit the delayed 1148 * wakeup section. 1149 */ 1150 if (atomic_cmpset_ptr(&gd->gd_delayed_wakeup[0], NULL, ident)) 1151 return; 1152 if (atomic_cmpset_ptr(&gd->gd_delayed_wakeup[1], NULL, ident)) 1153 return; 1154 1155 ident = atomic_swap_ptr(__DEQUALIFY(volatile void **, &gd->gd_delayed_wakeup[1]), 1156 __DEALL(ident)); 1157 ident = atomic_swap_ptr(__DEQUALIFY(volatile void **, &gd->gd_delayed_wakeup[0]), 1158 __DEALL(ident)); 1159 } 1160 1161 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, gd->gd_cpuid)); 1162 } 1163 1164 /* 1165 * Wakeup one thread tsleep()ing on the specified ident, on any cpu. 1166 */ 1167 void 1168 wakeup_one(const volatile void *ident) 1169 { 1170 /* XXX potentially round-robin the first responding cpu */ 1171 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mycpu->gd_cpuid) | 1172 PWAKEUP_ONE); 1173 } 1174 1175 /* 1176 * Wakeup threads tsleep()ing on the specified ident on the current cpu 1177 * only. 1178 */ 1179 void 1180 wakeup_mycpu(const volatile void *ident) 1181 { 1182 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mycpu->gd_cpuid) | 1183 PWAKEUP_MYCPU); 1184 } 1185 1186 /* 1187 * Wakeup one thread tsleep()ing on the specified ident on the current cpu 1188 * only. 1189 */ 1190 void 1191 wakeup_mycpu_one(const volatile void *ident) 1192 { 1193 /* XXX potentially round-robin the first responding cpu */ 1194 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mycpu->gd_cpuid) | 1195 PWAKEUP_MYCPU | PWAKEUP_ONE); 1196 } 1197 1198 /* 1199 * Wakeup all thread tsleep()ing on the specified ident on the specified cpu 1200 * only. 1201 */ 1202 void 1203 wakeup_oncpu(globaldata_t gd, const volatile void *ident) 1204 { 1205 globaldata_t mygd = mycpu; 1206 if (gd == mycpu) { 1207 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mygd->gd_cpuid) | 1208 PWAKEUP_MYCPU); 1209 } else { 1210 lwkt_send_ipiq2(gd, _wakeup, __DEALL(ident), 1211 PWAKEUP_ENCODE(0, mygd->gd_cpuid) | 1212 PWAKEUP_MYCPU); 1213 } 1214 } 1215 1216 /* 1217 * Wakeup one thread tsleep()ing on the specified ident on the specified cpu 1218 * only. 1219 */ 1220 void 1221 wakeup_oncpu_one(globaldata_t gd, const volatile void *ident) 1222 { 1223 globaldata_t mygd = mycpu; 1224 if (gd == mygd) { 1225 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mygd->gd_cpuid) | 1226 PWAKEUP_MYCPU | PWAKEUP_ONE); 1227 } else { 1228 lwkt_send_ipiq2(gd, _wakeup, __DEALL(ident), 1229 PWAKEUP_ENCODE(0, mygd->gd_cpuid) | 1230 PWAKEUP_MYCPU | PWAKEUP_ONE); 1231 } 1232 } 1233 1234 /* 1235 * Wakeup all threads waiting on the specified ident that slept using 1236 * the specified domain, on all cpus. 1237 */ 1238 void 1239 wakeup_domain(const volatile void *ident, int domain) 1240 { 1241 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(domain, mycpu->gd_cpuid)); 1242 } 1243 1244 /* 1245 * Wakeup one thread waiting on the specified ident that slept using 1246 * the specified domain, on any cpu. 1247 */ 1248 void 1249 wakeup_domain_one(const volatile void *ident, int domain) 1250 { 1251 /* XXX potentially round-robin the first responding cpu */ 1252 _wakeup(__DEALL(ident), 1253 PWAKEUP_ENCODE(domain, mycpu->gd_cpuid) | PWAKEUP_ONE); 1254 } 1255 1256 void 1257 wakeup_start_delayed(void) 1258 { 1259 globaldata_t gd = mycpu; 1260 1261 crit_enter(); 1262 gd->gd_curthread->td_flags |= TDF_DELAYED_WAKEUP; 1263 crit_exit(); 1264 } 1265 1266 void 1267 wakeup_end_delayed(void) 1268 { 1269 globaldata_t gd = mycpu; 1270 1271 if (gd->gd_curthread->td_flags & TDF_DELAYED_WAKEUP) { 1272 crit_enter(); 1273 gd->gd_curthread->td_flags &= ~TDF_DELAYED_WAKEUP; 1274 if (gd->gd_delayed_wakeup[0] || gd->gd_delayed_wakeup[1]) { 1275 if (gd->gd_delayed_wakeup[0]) { 1276 wakeup(gd->gd_delayed_wakeup[0]); 1277 gd->gd_delayed_wakeup[0] = NULL; 1278 } 1279 if (gd->gd_delayed_wakeup[1]) { 1280 wakeup(gd->gd_delayed_wakeup[1]); 1281 gd->gd_delayed_wakeup[1] = NULL; 1282 } 1283 } 1284 crit_exit(); 1285 } 1286 } 1287 1288 /* 1289 * setrunnable() 1290 * 1291 * Make a process runnable. lp->lwp_token must be held on call and this 1292 * function must be called from the cpu owning lp. 1293 * 1294 * This only has an effect if we are in LSSTOP or LSSLEEP. 1295 */ 1296 void 1297 setrunnable(struct lwp *lp) 1298 { 1299 thread_t td = lp->lwp_thread; 1300 1301 ASSERT_LWKT_TOKEN_HELD(&lp->lwp_token); 1302 KKASSERT(td->td_gd == mycpu); 1303 crit_enter(); 1304 if (lp->lwp_stat == LSSTOP) 1305 lp->lwp_stat = LSSLEEP; 1306 if (lp->lwp_stat == LSSLEEP) { 1307 _tsleep_remove(td); 1308 lwkt_schedule(td); 1309 } else if (td->td_flags & TDF_SINTR) { 1310 lwkt_schedule(td); 1311 } 1312 crit_exit(); 1313 } 1314 1315 /* 1316 * The process is stopped due to some condition, usually because p_stat is 1317 * set to SSTOP, but also possibly due to being traced. 1318 * 1319 * Caller must hold p->p_token 1320 * 1321 * NOTE! If the caller sets SSTOP, the caller must also clear P_WAITED 1322 * because the parent may check the child's status before the child actually 1323 * gets to this routine. 1324 * 1325 * This routine is called with the current lwp only, typically just 1326 * before returning to userland if the process state is detected as 1327 * possibly being in a stopped state. 1328 */ 1329 void 1330 tstop(void) 1331 { 1332 struct lwp *lp = curthread->td_lwp; 1333 struct proc *p = lp->lwp_proc; 1334 struct proc *q; 1335 1336 lwkt_gettoken(&lp->lwp_token); 1337 crit_enter(); 1338 1339 /* 1340 * If LWP_MP_WSTOP is set, we were sleeping 1341 * while our process was stopped. At this point 1342 * we were already counted as stopped. 1343 */ 1344 if ((lp->lwp_mpflags & LWP_MP_WSTOP) == 0) { 1345 /* 1346 * If we're the last thread to stop, signal 1347 * our parent. 1348 */ 1349 p->p_nstopped++; 1350 atomic_set_int(&lp->lwp_mpflags, LWP_MP_WSTOP); 1351 wakeup(&p->p_nstopped); 1352 if (p->p_nstopped == p->p_nthreads) { 1353 /* 1354 * Token required to interlock kern_wait() 1355 */ 1356 q = p->p_pptr; 1357 PHOLD(q); 1358 lwkt_gettoken(&q->p_token); 1359 p->p_flags &= ~P_WAITED; 1360 wakeup(p->p_pptr); 1361 if ((q->p_sigacts->ps_flag & PS_NOCLDSTOP) == 0) 1362 ksignal(q, SIGCHLD); 1363 lwkt_reltoken(&q->p_token); 1364 PRELE(q); 1365 } 1366 } 1367 1368 /* 1369 * Wait here while in a stopped state, interlocked with lwp_token. 1370 * We must break-out if the whole process is trying to exit. 1371 */ 1372 while (STOPLWP(p, lp)) { 1373 lp->lwp_stat = LSSTOP; 1374 tsleep(p, 0, "stop", 0); 1375 } 1376 p->p_nstopped--; 1377 atomic_clear_int(&lp->lwp_mpflags, LWP_MP_WSTOP); 1378 crit_exit(); 1379 lwkt_reltoken(&lp->lwp_token); 1380 } 1381 1382 /* 1383 * Compute a tenex style load average of a quantity on 1384 * 1, 5 and 15 minute intervals. This is a pcpu callout. 1385 * 1386 * We segment the lwp scan on a pcpu basis. This does NOT 1387 * mean the associated lwps are on this cpu, it is done 1388 * just to break the work up. 1389 * 1390 * The callout on cpu0 rolls up the stats from the other 1391 * cpus. 1392 */ 1393 static int loadav_count_runnable(struct lwp *p, void *data); 1394 1395 static void 1396 loadav(void *arg) 1397 { 1398 globaldata_t gd = mycpu; 1399 struct loadavg *avg; 1400 int i, nrun; 1401 1402 nrun = 0; 1403 alllwp_scan(loadav_count_runnable, &nrun, 1); 1404 gd->gd_loadav_nrunnable = nrun; 1405 if (gd->gd_cpuid == 0) { 1406 avg = &averunnable; 1407 nrun = 0; 1408 for (i = 0; i < ncpus; ++i) 1409 nrun += globaldata_find(i)->gd_loadav_nrunnable; 1410 for (i = 0; i < 3; i++) { 1411 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] + 1412 (long)nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT; 1413 } 1414 } 1415 1416 /* 1417 * Schedule the next update to occur after 5 seconds, but add a 1418 * random variation to avoid synchronisation with processes that 1419 * run at regular intervals. 1420 */ 1421 callout_reset(&gd->gd_loadav_callout, 1422 hz * 4 + (int)(krandom() % (hz * 2 + 1)), 1423 loadav, NULL); 1424 } 1425 1426 static int 1427 loadav_count_runnable(struct lwp *lp, void *data) 1428 { 1429 int *nrunp = data; 1430 thread_t td; 1431 1432 switch (lp->lwp_stat) { 1433 case LSRUN: 1434 if ((td = lp->lwp_thread) == NULL) 1435 break; 1436 if (td->td_flags & TDF_BLOCKED) 1437 break; 1438 ++*nrunp; 1439 break; 1440 default: 1441 break; 1442 } 1443 lwkt_yield(); 1444 return(0); 1445 } 1446 1447 /* 1448 * Regular data collection 1449 */ 1450 static uint64_t 1451 collect_load_callback(int n) 1452 { 1453 int fscale = averunnable.fscale; 1454 1455 return ((averunnable.ldavg[0] * 100 + (fscale >> 1)) / fscale); 1456 } 1457 1458 static void 1459 sched_setup(void *dummy __unused) 1460 { 1461 globaldata_t save_gd = mycpu; 1462 globaldata_t gd; 1463 int n; 1464 1465 kcollect_register(KCOLLECT_LOAD, "load", collect_load_callback, 1466 KCOLLECT_SCALE(KCOLLECT_LOAD_FORMAT, 0)); 1467 1468 /* 1469 * Kick off timeout driven events by calling first time. We 1470 * split the work across available cpus to help scale it, 1471 * it can eat a lot of cpu when there are a lot of processes 1472 * on the system. 1473 */ 1474 for (n = 0; n < ncpus; ++n) { 1475 gd = globaldata_find(n); 1476 lwkt_setcpu_self(gd); 1477 callout_init_mp(&gd->gd_loadav_callout); 1478 callout_init_mp(&gd->gd_schedcpu_callout); 1479 schedcpu(NULL); 1480 loadav(NULL); 1481 } 1482 lwkt_setcpu_self(save_gd); 1483 } 1484 1485 /* 1486 * Extremely early initialization, dummy-up the tables so we don't have 1487 * to conditionalize for NULL in _wakeup() and tsleep_interlock(). Even 1488 * though the system isn't blocking this early, these functions still 1489 * try to access the hash table. 1490 * 1491 * This setup will be overridden once sched_dyninit() -> sleep_gdinit() 1492 * is called. 1493 */ 1494 void 1495 sleep_early_gdinit(globaldata_t gd) 1496 { 1497 static struct tslpque dummy_slpque; 1498 static cpumask_t dummy_cpumasks; 1499 1500 slpque_tablesize = 1; 1501 gd->gd_tsleep_hash = &dummy_slpque; 1502 slpque_cpumasks = &dummy_cpumasks; 1503 TAILQ_INIT(&dummy_slpque.queue); 1504 } 1505 1506 /* 1507 * PCPU initialization. Called after KMALLOC is operational, by 1508 * sched_dyninit() for cpu 0, and by mi_gdinit() for other cpus later. 1509 * 1510 * WARNING! The pcpu hash table is smaller than the global cpumask 1511 * hash table, which can save us a lot of memory when maxproc 1512 * is set high. 1513 */ 1514 void 1515 sleep_gdinit(globaldata_t gd) 1516 { 1517 struct thread *td; 1518 size_t hash_size; 1519 uint32_t n; 1520 uint32_t i; 1521 1522 /* 1523 * This shouldn't happen, that is there shouldn't be any threads 1524 * waiting on the dummy tsleep queue this early in the boot. 1525 */ 1526 if (gd->gd_cpuid == 0) { 1527 struct tslpque *qp = &gd->gd_tsleep_hash[0]; 1528 TAILQ_FOREACH(td, &qp->queue, td_sleepq) { 1529 kprintf("SLEEP_GDINIT SWITCH %s\n", td->td_comm); 1530 } 1531 } 1532 1533 /* 1534 * Note that we have to allocate one extra slot because we are 1535 * shifting a modulo value. TCHASHSHIFT(slpque_tablesize - 1) can 1536 * return the same value as TCHASHSHIFT(slpque_tablesize). 1537 */ 1538 n = TCHASHSHIFT(slpque_tablesize) + 1; 1539 1540 hash_size = sizeof(struct tslpque) * n; 1541 gd->gd_tsleep_hash = (void *)kmem_alloc3(&kernel_map, hash_size, 1542 VM_SUBSYS_GD, 1543 KM_CPU(gd->gd_cpuid)); 1544 memset(gd->gd_tsleep_hash, 0, hash_size); 1545 for (i = 0; i < n; ++i) 1546 TAILQ_INIT(&gd->gd_tsleep_hash[i].queue); 1547 } 1548 1549 /* 1550 * Dynamic initialization after the memory system is operational. 1551 */ 1552 static void 1553 sched_dyninit(void *dummy __unused) 1554 { 1555 int tblsize; 1556 int tblsize2; 1557 int n; 1558 1559 /* 1560 * Calculate table size for slpque hash. We want a prime number 1561 * large enough to avoid overloading slpque_cpumasks when the 1562 * system has a large number of sleeping processes, which will 1563 * spam IPIs on wakeup(). 1564 * 1565 * While it is true this is really a per-lwp factor, generally 1566 * speaking the maxproc limit is a good metric to go by. 1567 */ 1568 for (tblsize = maxproc | 1; ; tblsize += 2) { 1569 if (tblsize % 3 == 0) 1570 continue; 1571 if (tblsize % 5 == 0) 1572 continue; 1573 tblsize2 = (tblsize / 2) | 1; 1574 for (n = 7; n < tblsize2; n += 2) { 1575 if (tblsize % n == 0) 1576 break; 1577 } 1578 if (n == tblsize2) 1579 break; 1580 } 1581 1582 /* 1583 * PIDs are currently limited to 6 digits. Cap the table size 1584 * at double this. 1585 */ 1586 if (tblsize > 2000003) 1587 tblsize = 2000003; 1588 1589 slpque_tablesize = tblsize; 1590 slpque_cpumasks = kmalloc(sizeof(*slpque_cpumasks) * slpque_tablesize, 1591 M_TSLEEP, M_WAITOK | M_ZERO); 1592 sleep_gdinit(mycpu); 1593 } 1594