thr_kern.c revision 118817
1/* 2 * Copyright (C) 2003 Daniel M. Eischen <deischen@freebsd.org> 3 * Copyright (C) 2002 Jonathon Mini <mini@freebsd.org> 4 * Copyright (c) 1995-1998 John Birrell <jb@cimlogic.com.au> 5 * All rights reserved. 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 1. Redistributions of source code must retain the above copyright 11 * notice, this list of conditions and the following disclaimer. 12 * 2. Redistributions in binary form must reproduce the above copyright 13 * notice, this list of conditions and the following disclaimer in the 14 * documentation and/or other materials provided with the distribution. 15 * 3. All advertising materials mentioning features or use of this software 16 * must display the following acknowledgement: 17 * This product includes software developed by John Birrell. 18 * 4. Neither the name of the author nor the names of any co-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 JOHN BIRRELL 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 AUTHOR 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 */ 35#include <sys/cdefs.h> 36__FBSDID("$FreeBSD: head/lib/libkse/thread/thr_kern.c 118817 2003-08-12 08:01:34Z davidxu $"); 37 38#include <sys/types.h> 39#include <sys/kse.h> 40#include <sys/signalvar.h> 41#include <sys/queue.h> 42#include <machine/atomic.h> 43#include <machine/sigframe.h> 44 45#include <assert.h> 46#include <errno.h> 47#include <signal.h> 48#include <stdlib.h> 49#include <string.h> 50#include <time.h> 51#include <ucontext.h> 52#include <unistd.h> 53 54#include "atomic_ops.h" 55#include "thr_private.h" 56#include "libc_private.h" 57 58/*#define DEBUG_THREAD_KERN */ 59#ifdef DEBUG_THREAD_KERN 60#define DBG_MSG stdout_debug 61#else 62#define DBG_MSG(x...) 63#endif 64 65/* 66 * Define a high water mark for the maximum number of threads that 67 * will be cached. Once this level is reached, any extra threads 68 * will be free()'d. 69 */ 70#define MAX_CACHED_THREADS 100 71/* 72 * Define high water marks for the maximum number of KSEs and KSE groups 73 * that will be cached. Because we support 1:1 threading, there could have 74 * same number of KSEs and KSE groups as threads. Once these levels are 75 * reached, any extra KSE and KSE groups will be free()'d. 76 */ 77#ifdef SYSTEM_SCOPE_ONLY 78#define MAX_CACHED_KSES 100 79#define MAX_CACHED_KSEGS 100 80#else 81#define MAX_CACHED_KSES 50 82#define MAX_CACHED_KSEGS 50 83#endif 84 85#define KSE_STACKSIZE 16384 86 87#define KSE_SET_MBOX(kse, thrd) \ 88 (kse)->k_kcb->kcb_kmbx.km_curthread = &(thrd)->tcb->tcb_tmbx 89 90#define KSE_SET_EXITED(kse) (kse)->k_flags |= KF_EXITED 91 92/* 93 * Macros for manipulating the run queues. The priority queue 94 * routines use the thread's pqe link and also handle the setting 95 * and clearing of the thread's THR_FLAGS_IN_RUNQ flag. 96 */ 97#define KSE_RUNQ_INSERT_HEAD(kse, thrd) \ 98 _pq_insert_head(&(kse)->k_schedq->sq_runq, thrd) 99#define KSE_RUNQ_INSERT_TAIL(kse, thrd) \ 100 _pq_insert_tail(&(kse)->k_schedq->sq_runq, thrd) 101#define KSE_RUNQ_REMOVE(kse, thrd) \ 102 _pq_remove(&(kse)->k_schedq->sq_runq, thrd) 103#define KSE_RUNQ_FIRST(kse) _pq_first(&(kse)->k_schedq->sq_runq) 104 105#define KSE_RUNQ_THREADS(kse) ((kse)->k_schedq->sq_runq.pq_threads) 106 107/* 108 * We've got to keep track of everything that is allocated, not only 109 * to have a speedy free list, but also so they can be deallocated 110 * after a fork(). 111 */ 112static TAILQ_HEAD(, kse) active_kseq; 113static TAILQ_HEAD(, kse) free_kseq; 114static TAILQ_HEAD(, kse_group) free_kse_groupq; 115static TAILQ_HEAD(, kse_group) active_kse_groupq; 116static TAILQ_HEAD(, kse_group) gc_ksegq; 117static struct lock kse_lock; /* also used for kseg queue */ 118static int free_kse_count = 0; 119static int free_kseg_count = 0; 120static TAILQ_HEAD(, pthread) free_threadq; 121static struct lock thread_lock; 122static int free_thread_count = 0; 123static int inited = 0; 124static int active_threads = 1; 125static int active_kse_count = 0; 126static int active_kseg_count = 0; 127static u_int64_t next_uniqueid = 1; 128 129LIST_HEAD(thread_hash_head, pthread); 130#define THREAD_HASH_QUEUES 127 131static struct thread_hash_head thr_hashtable[THREAD_HASH_QUEUES]; 132#define THREAD_HASH(thrd) ((unsigned long)thrd % THREAD_HASH_QUEUES) 133 134#ifdef DEBUG_THREAD_KERN 135static void dump_queues(struct kse *curkse); 136#endif 137static void kse_check_completed(struct kse *kse); 138static void kse_check_waitq(struct kse *kse); 139static void kse_fini(struct kse *curkse); 140static void kse_reinit(struct kse *kse, int sys_scope); 141static void kse_sched_multi(struct kse_mailbox *kmbx); 142static void kse_sched_single(struct kse_mailbox *kmbx); 143static void kse_switchout_thread(struct kse *kse, struct pthread *thread); 144static void kse_wait(struct kse *kse, struct pthread *td_wait, int sigseq); 145static void kse_free_unlocked(struct kse *kse); 146static void kse_destroy(struct kse *kse); 147static void kseg_free_unlocked(struct kse_group *kseg); 148static void kseg_init(struct kse_group *kseg); 149static void kseg_reinit(struct kse_group *kseg); 150static void kseg_destroy(struct kse_group *kseg); 151static void kse_waitq_insert(struct pthread *thread); 152static void kse_wakeup_multi(struct kse *curkse); 153static struct kse_mailbox *kse_wakeup_one(struct pthread *thread); 154static void thr_cleanup(struct kse *kse, struct pthread *curthread); 155static void thr_link(struct pthread *thread); 156static void thr_resume_wrapper(int sig, siginfo_t *, ucontext_t *); 157static void thr_resume_check(struct pthread *curthread, ucontext_t *ucp, 158 struct pthread_sigframe *psf); 159static int thr_timedout(struct pthread *thread, struct timespec *curtime); 160static void thr_unlink(struct pthread *thread); 161static void thread_gc(struct pthread *thread); 162static void kse_gc(struct pthread *thread); 163static void kseg_gc(struct pthread *thread); 164 165static void __inline 166thr_accounting(struct pthread *thread) 167{ 168 if ((thread->slice_usec != -1) && 169 (thread->slice_usec <= TIMESLICE_USEC) && 170 (thread->attr.sched_policy != SCHED_FIFO)) { 171 thread->slice_usec += (thread->tcb->tcb_tmbx.tm_uticks 172 + thread->tcb->tcb_tmbx.tm_sticks) * _clock_res_usec; 173 /* Check for time quantum exceeded: */ 174 if (thread->slice_usec > TIMESLICE_USEC) 175 thread->slice_usec = -1; 176 } 177 thread->tcb->tcb_tmbx.tm_uticks = 0; 178 thread->tcb->tcb_tmbx.tm_sticks = 0; 179} 180 181/* 182 * This is called after a fork(). 183 * No locks need to be taken here since we are guaranteed to be 184 * single threaded. 185 * 186 * XXX 187 * POSIX says for threaded process, fork() function is used 188 * only to run new programs, and the effects of calling functions 189 * that require certain resources between the call to fork() and 190 * the call to an exec function are undefined. 191 * 192 * Here it is not safe to reinitialize the library after fork(). 193 * Because memory management may be corrupted, further calling 194 * malloc()/free() may cause undefined behavior. 195 */ 196void 197_kse_single_thread(struct pthread *curthread) 198{ 199#ifdef NOTYET 200 struct kse *kse; 201 struct kse_group *kseg; 202 struct pthread *thread; 203 kse_critical_t crit; 204 int i; 205 206 if (__isthreaded) { 207 _thr_rtld_fini(); 208 _thr_signal_deinit(); 209 } 210 __isthreaded = 0; 211 /* 212 * Restore signal mask early, so any memory problems could 213 * dump core. 214 */ 215 sigprocmask(SIG_SETMASK, &curthread->sigmask, NULL); 216 active_threads = 1; 217 218 /* 219 * Enter a loop to remove and free all threads other than 220 * the running thread from the active thread list: 221 */ 222 while ((thread = TAILQ_FIRST(&_thread_list)) != NULL) { 223 THR_GCLIST_REMOVE(thread); 224 /* 225 * Remove this thread from the list (the current 226 * thread will be removed but re-added by libpthread 227 * initialization. 228 */ 229 TAILQ_REMOVE(&_thread_list, thread, tle); 230 /* Make sure this isn't the running thread: */ 231 if (thread != curthread) { 232 _thr_stack_free(&thread->attr); 233 if (thread->specific != NULL) 234 free(thread->specific); 235 for (i = 0; i < MAX_THR_LOCKLEVEL; i++) { 236 _lockuser_destroy(&thread->lockusers[i]); 237 } 238 _lock_destroy(&thread->lock); 239 free(thread); 240 } 241 } 242 243 TAILQ_INIT(&curthread->mutexq); /* initialize mutex queue */ 244 curthread->joiner = NULL; /* no joining threads yet */ 245 curthread->refcount = 0; 246 SIGEMPTYSET(curthread->sigpend); /* clear pending signals */ 247 if (curthread->specific != NULL) { 248 free(curthread->specific); 249 curthread->specific = NULL; 250 curthread->specific_data_count = 0; 251 } 252 253 /* Free the free KSEs: */ 254 while ((kse = TAILQ_FIRST(&free_kseq)) != NULL) { 255 TAILQ_REMOVE(&free_kseq, kse, k_qe); 256 for (i = 0; i < MAX_KSE_LOCKLEVEL; i++) { 257 _lockuser_destroy(&kse->k_lockusers[i]); 258 } 259 _lock_destroy(&kse->k_lock); 260 _kcb_dtor(kse->k_kcb); 261 if (kse->k_stack.ss_sp != NULL) 262 free(kse->k_stack.ss_sp); 263 free(kse); 264 } 265 free_kse_count = 0; 266 267 /* Free the active KSEs: */ 268 while ((kse = TAILQ_FIRST(&active_kseq)) != NULL) { 269 TAILQ_REMOVE(&active_kseq, kse, k_qe); 270 for (i = 0; i < MAX_KSE_LOCKLEVEL; i++) { 271 _lockuser_destroy(&kse->k_lockusers[i]); 272 } 273 _lock_destroy(&kse->k_lock); 274 if (kse->k_stack.ss_sp != NULL) 275 free(kse->k_stack.ss_sp); 276 free(kse); 277 } 278 active_kse_count = 0; 279 280 /* Free the free KSEGs: */ 281 while ((kseg = TAILQ_FIRST(&free_kse_groupq)) != NULL) { 282 TAILQ_REMOVE(&free_kse_groupq, kseg, kg_qe); 283 _lock_destroy(&kseg->kg_lock); 284 _pq_free(&kseg->kg_schedq.sq_runq); 285 free(kseg); 286 } 287 free_kseg_count = 0; 288 289 /* Free the active KSEGs: */ 290 while ((kseg = TAILQ_FIRST(&active_kse_groupq)) != NULL) { 291 TAILQ_REMOVE(&active_kse_groupq, kseg, kg_qe); 292 _lock_destroy(&kseg->kg_lock); 293 _pq_free(&kseg->kg_schedq.sq_runq); 294 free(kseg); 295 } 296 active_kseg_count = 0; 297 298 /* Free the free threads. */ 299 while ((thread = TAILQ_FIRST(&free_threadq)) != NULL) { 300 TAILQ_REMOVE(&free_threadq, thread, tle); 301 if (thread->specific != NULL) 302 free(thread->specific); 303 for (i = 0; i < MAX_THR_LOCKLEVEL; i++) { 304 _lockuser_destroy(&thread->lockusers[i]); 305 } 306 _lock_destroy(&thread->lock); 307 free(thread); 308 } 309 free_thread_count = 0; 310 311 /* Free the to-be-gc'd threads. */ 312 while ((thread = TAILQ_FIRST(&_thread_gc_list)) != NULL) { 313 TAILQ_REMOVE(&_thread_gc_list, thread, gcle); 314 for (i = 0; i < MAX_THR_LOCKLEVEL; i++) { 315 _lockuser_destroy(&thread->lockusers[i]); 316 } 317 _lock_destroy(&thread->lock); 318 free(thread); 319 } 320 TAILQ_INIT(&gc_ksegq); 321 _gc_count = 0; 322 323 if (inited != 0) { 324 /* 325 * Destroy these locks; they'll be recreated to assure they 326 * are in the unlocked state. 327 */ 328 _lock_destroy(&kse_lock); 329 _lock_destroy(&thread_lock); 330 _lock_destroy(&_thread_list_lock); 331 inited = 0; 332 } 333 334 /* 335 * After a fork(), the leftover thread goes back to being 336 * scope process. 337 */ 338 curthread->attr.flags &= ~PTHREAD_SCOPE_SYSTEM; 339 curthread->attr.flags |= PTHREAD_SCOPE_PROCESS; 340 341 /* 342 * After a fork, we are still operating on the thread's original 343 * stack. Don't clear the THR_FLAGS_USER from the thread's 344 * attribute flags. 345 */ 346 347 /* Initialize the threads library. */ 348 curthread->kse = NULL; 349 curthread->kseg = NULL; 350 _kse_initial = NULL; 351 _libpthread_init(curthread); 352#else 353 if (__isthreaded) { 354 _thr_rtld_fini(); 355 _thr_signal_deinit(); 356 } 357 __isthreaded = 0; 358 /* 359 * Restore signal mask early, so any memory problems could 360 * dump core. 361 */ 362 sigprocmask(SIG_SETMASK, &curthread->sigmask, NULL); 363 curthread->kse->k_kcb->kcb_kmbx.km_curthread = NULL; 364 active_threads = 1; 365#endif 366} 367 368/* 369 * This is used to initialize housekeeping and to initialize the 370 * KSD for the KSE. 371 */ 372void 373_kse_init(void) 374{ 375 if (inited == 0) { 376 TAILQ_INIT(&active_kseq); 377 TAILQ_INIT(&active_kse_groupq); 378 TAILQ_INIT(&free_kseq); 379 TAILQ_INIT(&free_kse_groupq); 380 TAILQ_INIT(&free_threadq); 381 TAILQ_INIT(&gc_ksegq); 382 if (_lock_init(&kse_lock, LCK_ADAPTIVE, 383 _kse_lock_wait, _kse_lock_wakeup) != 0) 384 PANIC("Unable to initialize free KSE queue lock"); 385 if (_lock_init(&thread_lock, LCK_ADAPTIVE, 386 _kse_lock_wait, _kse_lock_wakeup) != 0) 387 PANIC("Unable to initialize free thread queue lock"); 388 if (_lock_init(&_thread_list_lock, LCK_ADAPTIVE, 389 _kse_lock_wait, _kse_lock_wakeup) != 0) 390 PANIC("Unable to initialize thread list lock"); 391 active_kse_count = 0; 392 active_kseg_count = 0; 393 _gc_count = 0; 394 inited = 1; 395 } 396} 397 398int 399_kse_isthreaded(void) 400{ 401 return (__isthreaded != 0); 402} 403 404/* 405 * This is called when the first thread (other than the initial 406 * thread) is created. 407 */ 408int 409_kse_setthreaded(int threaded) 410{ 411 sigset_t sigset; 412 413 if ((threaded != 0) && (__isthreaded == 0)) { 414 /* 415 * Tell the kernel to create a KSE for the initial thread 416 * and enable upcalls in it. 417 */ 418 _kse_initial->k_flags |= KF_STARTED; 419 420#ifdef SYSTEM_SCOPE_ONLY 421 /* 422 * For bound thread, kernel reads mailbox pointer once, 423 * we'd set it here before calling kse_create 424 */ 425 _tcb_set(_kse_initial->k_kcb, _thr_initial->tcb); 426 KSE_SET_MBOX(_kse_initial, _thr_initial); 427 _kse_initial->k_kcb->kcb_kmbx.km_flags |= KMF_BOUND; 428#endif 429 SIGFILLSET(sigset); 430 __sys_sigprocmask(SIG_SETMASK, &sigset, &_thr_initial->sigmask); 431 _thr_signal_init(); 432 433 /* 434 * Locking functions in libc are required when there are 435 * threads other than the initial thread. 436 */ 437 _thr_rtld_init(); 438 439 __isthreaded = 1; 440 if (kse_create(&_kse_initial->k_kcb->kcb_kmbx, 0) != 0) { 441 _kse_initial->k_flags &= ~KF_STARTED; 442 __isthreaded = 0; 443 PANIC("kse_create() failed\n"); 444 return (-1); 445 } 446 447#ifndef SYSTEM_SCOPE_ONLY 448 /* Set current thread to initial thread */ 449 _tcb_set(_kse_initial->k_kcb, _thr_initial->tcb); 450 KSE_SET_MBOX(_kse_initial, _thr_initial); 451 _thr_start_sig_daemon(); 452 _thr_setmaxconcurrency(); 453#else 454 __sys_sigprocmask(SIG_SETMASK, &_thr_initial->sigmask, NULL); 455#endif 456 } 457 return (0); 458} 459 460/* 461 * Lock wait and wakeup handlers for KSE locks. These are only used by 462 * KSEs, and should never be used by threads. KSE locks include the 463 * KSE group lock (used for locking the scheduling queue) and the 464 * kse_lock defined above. 465 * 466 * When a KSE lock attempt blocks, the entire KSE blocks allowing another 467 * KSE to run. For the most part, it doesn't make much sense to try and 468 * schedule another thread because you need to lock the scheduling queue 469 * in order to do that. And since the KSE lock is used to lock the scheduling 470 * queue, you would just end up blocking again. 471 */ 472void 473_kse_lock_wait(struct lock *lock, struct lockuser *lu) 474{ 475 struct kse *curkse = (struct kse *)_LCK_GET_PRIVATE(lu); 476 struct timespec ts; 477 int saved_flags; 478 479 if (curkse->k_kcb->kcb_kmbx.km_curthread != NULL) 480 PANIC("kse_lock_wait does not disable upcall.\n"); 481 /* 482 * Enter a loop to wait until we get the lock. 483 */ 484 ts.tv_sec = 0; 485 ts.tv_nsec = 1000000; /* 1 sec */ 486 while (!_LCK_GRANTED(lu)) { 487 /* 488 * Yield the kse and wait to be notified when the lock 489 * is granted. 490 */ 491 saved_flags = curkse->k_kcb->kcb_kmbx.km_flags; 492 curkse->k_kcb->kcb_kmbx.km_flags |= KMF_NOUPCALL | 493 KMF_NOCOMPLETED; 494 kse_release(&ts); 495 curkse->k_kcb->kcb_kmbx.km_flags = saved_flags; 496 } 497} 498 499void 500_kse_lock_wakeup(struct lock *lock, struct lockuser *lu) 501{ 502 struct kse *curkse; 503 struct kse *kse; 504 struct kse_mailbox *mbx; 505 506 curkse = _get_curkse(); 507 kse = (struct kse *)_LCK_GET_PRIVATE(lu); 508 509 if (kse == curkse) 510 PANIC("KSE trying to wake itself up in lock"); 511 else { 512 mbx = &kse->k_kcb->kcb_kmbx; 513 _lock_grant(lock, lu); 514 /* 515 * Notify the owning kse that it has the lock. 516 * It is safe to pass invalid address to kse_wakeup 517 * even if the mailbox is not in kernel at all, 518 * and waking up a wrong kse is also harmless. 519 */ 520 kse_wakeup(mbx); 521 } 522} 523 524/* 525 * Thread wait and wakeup handlers for thread locks. These are only used 526 * by threads, never by KSEs. Thread locks include the per-thread lock 527 * (defined in its structure), and condition variable and mutex locks. 528 */ 529void 530_thr_lock_wait(struct lock *lock, struct lockuser *lu) 531{ 532 struct pthread *curthread = (struct pthread *)lu->lu_private; 533 534 do { 535 THR_LOCK_SWITCH(curthread); 536 THR_SET_STATE(curthread, PS_LOCKWAIT); 537 _thr_sched_switch_unlocked(curthread); 538 } while (!_LCK_GRANTED(lu)); 539} 540 541void 542_thr_lock_wakeup(struct lock *lock, struct lockuser *lu) 543{ 544 struct pthread *thread; 545 struct pthread *curthread; 546 struct kse_mailbox *kmbx; 547 548 curthread = _get_curthread(); 549 thread = (struct pthread *)_LCK_GET_PRIVATE(lu); 550 551 THR_SCHED_LOCK(curthread, thread); 552 _lock_grant(lock, lu); 553 kmbx = _thr_setrunnable_unlocked(thread); 554 THR_SCHED_UNLOCK(curthread, thread); 555 if (kmbx != NULL) 556 kse_wakeup(kmbx); 557} 558 559kse_critical_t 560_kse_critical_enter(void) 561{ 562 kse_critical_t crit; 563 564 crit = (kse_critical_t)_kcb_critical_enter(); 565 return (crit); 566} 567 568void 569_kse_critical_leave(kse_critical_t crit) 570{ 571 struct pthread *curthread; 572 573 _kcb_critical_leave((struct kse_thr_mailbox *)crit); 574 if ((crit != NULL) && ((curthread = _get_curthread()) != NULL)) 575 THR_YIELD_CHECK(curthread); 576} 577 578int 579_kse_in_critical(void) 580{ 581 return (_kcb_in_critical()); 582} 583 584void 585_thr_critical_enter(struct pthread *thread) 586{ 587 thread->critical_count++; 588} 589 590void 591_thr_critical_leave(struct pthread *thread) 592{ 593 thread->critical_count--; 594 THR_YIELD_CHECK(thread); 595} 596 597void 598_thr_sched_switch(struct pthread *curthread) 599{ 600 struct kse *curkse; 601 602 (void)_kse_critical_enter(); 603 curkse = _get_curkse(); 604 KSE_SCHED_LOCK(curkse, curkse->k_kseg); 605 _thr_sched_switch_unlocked(curthread); 606} 607 608/* 609 * XXX - We may need to take the scheduling lock before calling 610 * this, or perhaps take the lock within here before 611 * doing anything else. 612 */ 613void 614_thr_sched_switch_unlocked(struct pthread *curthread) 615{ 616 struct pthread *td; 617 struct pthread_sigframe psf; 618 struct kse *curkse; 619 int ret; 620 volatile int uts_once; 621 volatile int resume_once = 0; 622 ucontext_t uc; 623 624 /* We're in the scheduler, 5 by 5: */ 625 curkse = _get_curkse(); 626 _tcb_set(curkse->k_kcb, NULL); 627 628 curthread->need_switchout = 1; /* The thread yielded on its own. */ 629 curthread->critical_yield = 0; /* No need to yield anymore. */ 630 thr_accounting(curthread); 631 632 633 /* Thread can unlock the scheduler lock. */ 634 curthread->lock_switch = 1; 635 636 /* 637 * The signal frame is allocated off the stack because 638 * a thread can be interrupted by other signals while 639 * it is running down pending signals. 640 */ 641 psf.psf_valid = 0; 642 curthread->curframe = &psf; 643 644 /* 645 * Enter the scheduler if any one of the following is true: 646 * 647 * o The current thread is dead; it's stack needs to be 648 * cleaned up and it can't be done while operating on 649 * it. 650 * o The current thread has signals pending, should 651 * let scheduler install signal trampoline for us. 652 * o There are no runnable threads. 653 * o The next thread to run won't unlock the scheduler 654 * lock. A side note: the current thread may be run 655 * instead of the next thread in the run queue, but 656 * we don't bother checking for that. 657 */ 658 if (curthread->attr.flags & PTHREAD_SCOPE_SYSTEM) 659 kse_sched_single(&curkse->k_kcb->kcb_kmbx); 660 else if ((curthread->state == PS_DEAD) || 661 (((td = KSE_RUNQ_FIRST(curkse)) == NULL) && 662 (curthread->state != PS_RUNNING)) || 663 ((td != NULL) && (td->lock_switch == 0))) { 664 curkse->k_switch = 1; 665 _thread_enter_uts(curthread->tcb, curkse->k_kcb); 666 } 667 else { 668 uts_once = 0; 669 THR_GETCONTEXT(&curthread->tcb->tcb_tmbx.tm_context); 670 if (uts_once == 0) { 671 uts_once = 1; 672 673 /* Switchout the current thread. */ 674 kse_switchout_thread(curkse, curthread); 675 676 /* Choose another thread to run. */ 677 td = KSE_RUNQ_FIRST(curkse); 678 KSE_RUNQ_REMOVE(curkse, td); 679 curkse->k_curthread = td; 680 681 /* 682 * Make sure the current thread's kse points to 683 * this kse. 684 */ 685 td->kse = curkse; 686 687 /* 688 * Reset the time slice if this thread is running 689 * for the first time or running again after using 690 * its full time slice allocation. 691 */ 692 if (td->slice_usec == -1) 693 td->slice_usec = 0; 694 695 /* Mark the thread active. */ 696 td->active = 1; 697 698 /* Remove the frame reference. */ 699 td->curframe = NULL; 700 701 /* 702 * Continue the thread at its current frame: 703 */ 704 ret = _thread_switch(curkse->k_kcb, td->tcb, 0); 705 /* This point should not be reached. */ 706 if (ret != 0) 707 PANIC("Bad return from _thread_switch"); 708 PANIC("Thread has returned from _thread_switch"); 709 } 710 } 711 712 if (psf.psf_valid) { 713 /* 714 * It is ugly we must increase critical count, because we 715 * have a frame saved, we must backout state in psf 716 * before we can process signals. 717 */ 718 curthread->critical_count++; 719 } 720 721 if (curthread->lock_switch != 0) { 722 /* 723 * Unlock the scheduling queue and leave the 724 * critical region. 725 */ 726 /* Don't trust this after a switch! */ 727 curkse = _get_curkse(); 728 729 curthread->lock_switch = 0; 730 KSE_SCHED_UNLOCK(curkse, curkse->k_kseg); 731 _kse_critical_leave(&curthread->tcb->tcb_tmbx); 732 } 733 /* 734 * This thread is being resumed; check for cancellations. 735 */ 736 if ((psf.psf_valid || 737 (curthread->check_pending && !THR_IN_CRITICAL(curthread)))) { 738 resume_once = 0; 739 THR_GETCONTEXT(&uc); 740 if (resume_once == 0) { 741 resume_once = 1; 742 curthread->check_pending = 0; 743 thr_resume_check(curthread, &uc, &psf); 744 } 745 } 746 THR_ACTIVATE_LAST_LOCK(curthread); 747} 748 749/* 750 * This is the scheduler for a KSE which runs a scope system thread. 751 * The multi-thread KSE scheduler should also work for a single threaded 752 * KSE, but we use a separate scheduler so that it can be fine-tuned 753 * to be more efficient (and perhaps not need a separate stack for 754 * the KSE, allowing it to use the thread's stack). 755 */ 756 757static void 758kse_sched_single(struct kse_mailbox *kmbx) 759{ 760 struct kse *curkse; 761 struct pthread *curthread; 762 struct timespec ts; 763 sigset_t sigmask; 764 int i, sigseqno, level, first = 0; 765 766 curkse = (struct kse *)kmbx->km_udata; 767 curthread = curkse->k_curthread; 768 769 if ((curkse->k_flags & KF_INITIALIZED) == 0) { 770 /* Setup this KSEs specific data. */ 771 _kcb_set(curkse->k_kcb); 772 _tcb_set(curkse->k_kcb, curthread->tcb); 773 curkse->k_flags |= KF_INITIALIZED; 774 first = 1; 775 curthread->active = 1; 776 777 /* Setup kernel signal masks for new thread. */ 778 __sys_sigprocmask(SIG_SETMASK, &curthread->sigmask, NULL); 779 /* 780 * Enter critical region, this is meanless for bound thread, 781 * It is used to let other code work, those code want mailbox 782 * to be cleared. 783 */ 784 (void)_kse_critical_enter(); 785 } 786 787 curthread->critical_yield = 0; 788 curthread->need_switchout = 0; 789 790 /* 791 * Lock the scheduling queue. 792 * 793 * There is no scheduling queue for single threaded KSEs, 794 * but we need a lock for protection regardless. 795 */ 796 if (curthread->lock_switch == 0) 797 KSE_SCHED_LOCK(curkse, curkse->k_kseg); 798 799 /* 800 * This has to do the job of kse_switchout_thread(), only 801 * for a single threaded KSE/KSEG. 802 */ 803 804 switch (curthread->state) { 805 case PS_DEAD: 806 curthread->check_pending = 0; 807 /* Unlock the scheduling queue and exit the KSE and thread. */ 808 thr_cleanup(curkse, curthread); 809 KSE_SCHED_UNLOCK(curkse, curkse->k_kseg); 810 PANIC("bound thread shouldn't get here\n"); 811 break; 812 813 case PS_SIGWAIT: 814 PANIC("bound thread does not have SIGWAIT state\n"); 815 816 case PS_SLEEP_WAIT: 817 PANIC("bound thread does not have SLEEP_WAIT state\n"); 818 819 case PS_SIGSUSPEND: 820 PANIC("bound thread does not have SIGSUSPEND state\n"); 821 822 case PS_COND_WAIT: 823 break; 824 825 case PS_LOCKWAIT: 826 /* 827 * This state doesn't timeout. 828 */ 829 curthread->wakeup_time.tv_sec = -1; 830 curthread->wakeup_time.tv_nsec = -1; 831 level = curthread->locklevel - 1; 832 if (_LCK_GRANTED(&curthread->lockusers[level])) 833 THR_SET_STATE(curthread, PS_RUNNING); 834 break; 835 836 case PS_RUNNING: 837 if ((curthread->flags & THR_FLAGS_SUSPENDED) != 0) { 838 THR_SET_STATE(curthread, PS_SUSPENDED); 839 } 840 curthread->wakeup_time.tv_sec = -1; 841 curthread->wakeup_time.tv_nsec = -1; 842 break; 843 844 case PS_JOIN: 845 case PS_MUTEX_WAIT: 846 case PS_SUSPENDED: 847 case PS_DEADLOCK: 848 default: 849 /* 850 * These states don't timeout and don't need 851 * to be in the waiting queue. 852 */ 853 curthread->wakeup_time.tv_sec = -1; 854 curthread->wakeup_time.tv_nsec = -1; 855 break; 856 } 857 858 while (curthread->state != PS_RUNNING) { 859 sigseqno = curkse->k_sigseqno; 860 if (curthread->check_pending != 0) { 861 /* 862 * Install pending signals into the frame, possible 863 * cause mutex or condvar backout. 864 */ 865 curthread->check_pending = 0; 866 SIGFILLSET(sigmask); 867 868 /* 869 * Lock out kernel signal code when we are processing 870 * signals, and get a fresh copy of signal mask. 871 */ 872 __sys_sigprocmask(SIG_SETMASK, &sigmask, 873 &curthread->sigmask); 874 for (i = 1; i <= _SIG_MAXSIG; i++) { 875 if (SIGISMEMBER(curthread->sigmask, i)) 876 continue; 877 if (SIGISMEMBER(curthread->sigpend, i)) 878 (void)_thr_sig_add(curthread, i, 879 &curthread->siginfo[i-1]); 880 } 881 __sys_sigprocmask(SIG_SETMASK, &curthread->sigmask, 882 NULL); 883 /* The above code might make thread runnable */ 884 if (curthread->state == PS_RUNNING) 885 break; 886 } 887 THR_DEACTIVATE_LAST_LOCK(curthread); 888 kse_wait(curkse, curthread, sigseqno); 889 THR_ACTIVATE_LAST_LOCK(curthread); 890 KSE_GET_TOD(curkse, &ts); 891 if (thr_timedout(curthread, &ts)) { 892 /* Indicate the thread timedout: */ 893 curthread->timeout = 1; 894 /* Make the thread runnable. */ 895 THR_SET_STATE(curthread, PS_RUNNING); 896 } 897 } 898 899 /* Remove the frame reference. */ 900 curthread->curframe = NULL; 901 902 if (curthread->lock_switch == 0) { 903 /* Unlock the scheduling queue. */ 904 KSE_SCHED_UNLOCK(curkse, curkse->k_kseg); 905 } 906 907 DBG_MSG("Continuing bound thread %p\n", curthread); 908 if (first) { 909 _kse_critical_leave(&curthread->tcb->tcb_tmbx); 910 pthread_exit(curthread->start_routine(curthread->arg)); 911 } 912} 913 914#ifdef DEBUG_THREAD_KERN 915static void 916dump_queues(struct kse *curkse) 917{ 918 struct pthread *thread; 919 920 DBG_MSG("Threads in waiting queue:\n"); 921 TAILQ_FOREACH(thread, &curkse->k_kseg->kg_schedq.sq_waitq, pqe) { 922 DBG_MSG(" thread %p, state %d, blocked %d\n", 923 thread, thread->state, thread->blocked); 924 } 925} 926#endif 927 928/* 929 * This is the scheduler for a KSE which runs multiple threads. 930 */ 931static void 932kse_sched_multi(struct kse_mailbox *kmbx) 933{ 934 struct kse *curkse; 935 struct pthread *curthread, *td_wait; 936 struct pthread_sigframe *curframe; 937 int ret; 938 939 curkse = (struct kse *)kmbx->km_udata; 940 THR_ASSERT(curkse->k_kcb->kcb_kmbx.km_curthread == NULL, 941 "Mailbox not null in kse_sched_multi"); 942 943 /* Check for first time initialization: */ 944 if ((curkse->k_flags & KF_INITIALIZED) == 0) { 945 /* Setup this KSEs specific data. */ 946 _kcb_set(curkse->k_kcb); 947 948 /* Set this before grabbing the context. */ 949 curkse->k_flags |= KF_INITIALIZED; 950 } 951 952 /* This may have returned from a kse_release(). */ 953 if (KSE_WAITING(curkse)) { 954 DBG_MSG("Entered upcall when KSE is waiting."); 955 KSE_CLEAR_WAIT(curkse); 956 } 957 958 /* If this is an upcall; take the scheduler lock. */ 959 if (curkse->k_switch == 0) { 960 /* Set fake kcb */ 961 _tcb_set(curkse->k_kcb, NULL); 962 KSE_SCHED_LOCK(curkse, curkse->k_kseg); 963 } 964 curkse->k_switch = 0; 965 966 /* 967 * Now that the scheduler lock is held, get the current 968 * thread. The KSE's current thread cannot be safely 969 * examined without the lock because it could have returned 970 * as completed on another KSE. See kse_check_completed(). 971 */ 972 curthread = curkse->k_curthread; 973 974 if (KSE_IS_IDLE(curkse)) { 975 KSE_CLEAR_IDLE(curkse); 976 curkse->k_kseg->kg_idle_kses--; 977 } 978 /* 979 * If the current thread was completed in another KSE, then 980 * it will be in the run queue. Don't mark it as being blocked. 981 */ 982 if ((curthread != NULL) && 983 ((curthread->flags & THR_FLAGS_IN_RUNQ) == 0) && 984 (curthread->need_switchout == 0)) { 985 /* 986 * Assume the current thread is blocked; when the 987 * completed threads are checked and if the current 988 * thread is among the completed, the blocked flag 989 * will be cleared. 990 */ 991 curthread->blocked = 1; 992 } 993 994 /* Check for any unblocked threads in the kernel. */ 995 kse_check_completed(curkse); 996 997 /* 998 * Check for threads that have timed-out. 999 */ 1000 kse_check_waitq(curkse); 1001 1002 /* 1003 * Switchout the current thread, if necessary, as the last step 1004 * so that it is inserted into the run queue (if it's runnable) 1005 * _after_ any other threads that were added to it above. 1006 */ 1007 if (curthread == NULL) 1008 ; /* Nothing to do here. */ 1009 else if ((curthread->need_switchout == 0) && 1010 (curthread->blocked == 0) && (THR_IN_CRITICAL(curthread))) { 1011 /* 1012 * Resume the thread and tell it to yield when 1013 * it leaves the critical region. 1014 */ 1015 curthread->critical_yield = 1; 1016 curthread->active = 1; 1017 if ((curthread->flags & THR_FLAGS_IN_RUNQ) != 0) 1018 KSE_RUNQ_REMOVE(curkse, curthread); 1019 curkse->k_curthread = curthread; 1020 curthread->kse = curkse; 1021 DBG_MSG("Continuing thread %p in critical region\n", 1022 curthread); 1023 kse_wakeup_multi(curkse); 1024 KSE_SCHED_UNLOCK(curkse, curkse->k_kseg); 1025 ret = _thread_switch(curkse->k_kcb, curthread->tcb, 1); 1026 if (ret != 0) 1027 PANIC("Can't resume thread in critical region\n"); 1028 } 1029 else if ((curthread->flags & THR_FLAGS_IN_RUNQ) == 0) 1030 kse_switchout_thread(curkse, curthread); 1031 curkse->k_curthread = NULL; 1032 1033 kse_wakeup_multi(curkse); 1034 1035#ifdef DEBUG_THREAD_KERN 1036 dump_queues(curkse); 1037#endif 1038 1039 /* Check if there are no threads ready to run: */ 1040 while (((curthread = KSE_RUNQ_FIRST(curkse)) == NULL) && 1041 (curkse->k_kseg->kg_threadcount != 0)) { 1042 /* 1043 * Wait for a thread to become active or until there are 1044 * no more threads. 1045 */ 1046 td_wait = KSE_WAITQ_FIRST(curkse); 1047 kse_wait(curkse, td_wait, 0); 1048 kse_check_completed(curkse); 1049 kse_check_waitq(curkse); 1050 } 1051 1052 /* Check for no more threads: */ 1053 if (curkse->k_kseg->kg_threadcount == 0) { 1054 /* 1055 * Normally this shouldn't return, but it will if there 1056 * are other KSEs running that create new threads that 1057 * are assigned to this KSE[G]. For instance, if a scope 1058 * system thread were to create a scope process thread 1059 * and this kse[g] is the initial kse[g], then that newly 1060 * created thread would be assigned to us (the initial 1061 * kse[g]). 1062 */ 1063 KSE_SCHED_UNLOCK(curkse, curkse->k_kseg); 1064 kse_fini(curkse); 1065 /* never returns */ 1066 } 1067 1068 THR_ASSERT(curthread != NULL, 1069 "Return from kse_wait/fini without thread."); 1070 THR_ASSERT(curthread->state != PS_DEAD, 1071 "Trying to resume dead thread!"); 1072 KSE_RUNQ_REMOVE(curkse, curthread); 1073 1074 /* 1075 * Make the selected thread the current thread. 1076 */ 1077 curkse->k_curthread = curthread; 1078 1079 /* 1080 * Make sure the current thread's kse points to this kse. 1081 */ 1082 curthread->kse = curkse; 1083 1084 /* 1085 * Reset the time slice if this thread is running for the first 1086 * time or running again after using its full time slice allocation. 1087 */ 1088 if (curthread->slice_usec == -1) 1089 curthread->slice_usec = 0; 1090 1091 /* Mark the thread active. */ 1092 curthread->active = 1; 1093 1094 /* Remove the frame reference. */ 1095 curframe = curthread->curframe; 1096 curthread->curframe = NULL; 1097 1098 kse_wakeup_multi(curkse); 1099 1100 /* 1101 * The thread's current signal frame will only be NULL if it 1102 * is being resumed after being blocked in the kernel. In 1103 * this case, and if the thread needs to run down pending 1104 * signals or needs a cancellation check, we need to add a 1105 * signal frame to the thread's context. 1106 */ 1107#ifdef NOT_YET 1108 if ((((curframe == NULL) && (curthread->check_pending != 0)) || 1109 (((curthread->cancelflags & THR_AT_CANCEL_POINT) == 0) && 1110 ((curthread->cancelflags & PTHREAD_CANCEL_ASYNCHRONOUS) != 0))) && 1111 !THR_IN_CRITICAL(curthread)) 1112 signalcontext(&curthread->tcb->tcb_tmbx.tm_context, 0, 1113 (__sighandler_t *)thr_resume_wrapper); 1114#else 1115 if ((curframe == NULL) && (curthread->state == PS_RUNNING) && 1116 (curthread->check_pending != 0) && !THR_IN_CRITICAL(curthread)) { 1117 curthread->check_pending = 0; 1118 signalcontext(&curthread->tcb->tcb_tmbx.tm_context, 0, 1119 (__sighandler_t *)thr_resume_wrapper); 1120 } 1121#endif 1122 /* 1123 * Continue the thread at its current frame: 1124 */ 1125 if (curthread->lock_switch != 0) { 1126 /* 1127 * This thread came from a scheduler switch; it will 1128 * unlock the scheduler lock and set the mailbox. 1129 */ 1130 ret = _thread_switch(curkse->k_kcb, curthread->tcb, 0); 1131 } else { 1132 /* This thread won't unlock the scheduler lock. */ 1133 KSE_SCHED_UNLOCK(curkse, curkse->k_kseg); 1134 ret = _thread_switch(curkse->k_kcb, curthread->tcb, 1); 1135 } 1136 if (ret != 0) 1137 PANIC("Thread has returned from _thread_switch"); 1138 1139 /* This point should not be reached. */ 1140 PANIC("Thread has returned from _thread_switch"); 1141} 1142 1143static void 1144thr_resume_wrapper(int sig, siginfo_t *siginfo, ucontext_t *ucp) 1145{ 1146 struct pthread *curthread = _get_curthread(); 1147 struct kse *curkse; 1148 int ret, err_save = curthread->error; 1149 1150 DBG_MSG(">>> sig wrapper\n"); 1151 if (curthread->lock_switch) 1152 PANIC("thr_resume_wrapper, lock_switch != 0\n"); 1153 thr_resume_check(curthread, ucp, NULL); 1154 _kse_critical_enter(); 1155 curkse = _get_curkse(); 1156 curthread->tcb->tcb_tmbx.tm_context = *ucp; 1157 curthread->error = err_save; 1158 ret = _thread_switch(curkse->k_kcb, curthread->tcb, 1); 1159 if (ret != 0) 1160 PANIC("thr_resume_wrapper: thread has returned " 1161 "from _thread_switch"); 1162 /* THR_SETCONTEXT(ucp); */ /* not work, why ? */ 1163} 1164 1165static void 1166thr_resume_check(struct pthread *curthread, ucontext_t *ucp, 1167 struct pthread_sigframe *psf) 1168{ 1169 _thr_sig_rundown(curthread, ucp, psf); 1170 1171#ifdef NOT_YET 1172 if (((curthread->cancelflags & THR_AT_CANCEL_POINT) == 0) && 1173 ((curthread->cancelflags & PTHREAD_CANCEL_ASYNCHRONOUS) != 0)) 1174 pthread_testcancel(); 1175#endif 1176} 1177 1178/* 1179 * Clean up a thread. This must be called with the thread's KSE 1180 * scheduling lock held. The thread must be a thread from the 1181 * KSE's group. 1182 */ 1183static void 1184thr_cleanup(struct kse *curkse, struct pthread *thread) 1185{ 1186 struct pthread *joiner; 1187 struct kse_mailbox *kmbx = NULL; 1188 int sys_scope; 1189 1190 if ((joiner = thread->joiner) != NULL) { 1191 /* Joinee scheduler lock held; joiner won't leave. */ 1192 if (joiner->kseg == curkse->k_kseg) { 1193 if (joiner->join_status.thread == thread) { 1194 joiner->join_status.thread = NULL; 1195 joiner->join_status.ret = thread->ret; 1196 (void)_thr_setrunnable_unlocked(joiner); 1197 } 1198 } else { 1199 KSE_SCHED_UNLOCK(curkse, curkse->k_kseg); 1200 /* The joiner may have removed itself and exited. */ 1201 if (_thr_ref_add(thread, joiner, 0) == 0) { 1202 KSE_SCHED_LOCK(curkse, joiner->kseg); 1203 if (joiner->join_status.thread == thread) { 1204 joiner->join_status.thread = NULL; 1205 joiner->join_status.ret = thread->ret; 1206 kmbx = _thr_setrunnable_unlocked(joiner); 1207 } 1208 KSE_SCHED_UNLOCK(curkse, joiner->kseg); 1209 _thr_ref_delete(thread, joiner); 1210 if (kmbx != NULL) 1211 kse_wakeup(kmbx); 1212 } 1213 KSE_SCHED_LOCK(curkse, curkse->k_kseg); 1214 } 1215 thread->attr.flags |= PTHREAD_DETACHED; 1216 } 1217 1218 if (!(sys_scope = (thread->attr.flags & PTHREAD_SCOPE_SYSTEM))) { 1219 /* 1220 * Remove the thread from the KSEG's list of threads. 1221 */ 1222 KSEG_THRQ_REMOVE(thread->kseg, thread); 1223 /* 1224 * Migrate the thread to the main KSE so that this 1225 * KSE and KSEG can be cleaned when their last thread 1226 * exits. 1227 */ 1228 thread->kseg = _kse_initial->k_kseg; 1229 thread->kse = _kse_initial; 1230 } 1231 thread->flags |= THR_FLAGS_GC_SAFE; 1232 1233 /* 1234 * We can't hold the thread list lock while holding the 1235 * scheduler lock. 1236 */ 1237 KSE_SCHED_UNLOCK(curkse, curkse->k_kseg); 1238 DBG_MSG("Adding thread %p to GC list\n", thread); 1239 KSE_LOCK_ACQUIRE(curkse, &_thread_list_lock); 1240 THR_GCLIST_ADD(thread); 1241 /* Use thread_list_lock */ 1242 active_threads--; 1243#ifdef SYSTEM_SCOPE_ONLY 1244 if (active_threads == 0) { 1245#else 1246 if (active_threads == 1) { 1247#endif 1248 KSE_LOCK_RELEASE(curkse, &_thread_list_lock); 1249 exit(0); 1250 } 1251 KSE_LOCK_RELEASE(curkse, &_thread_list_lock); 1252 if (sys_scope) { 1253 /* 1254 * System scope thread is single thread group, 1255 * when thread is exited, its kse and ksegrp should 1256 * be recycled as well. 1257 * kse upcall stack belongs to thread, clear it here. 1258 */ 1259 curkse->k_stack.ss_sp = 0; 1260 curkse->k_stack.ss_size = 0; 1261 kse_exit(); 1262 PANIC("kse_exit() failed for system scope thread"); 1263 } 1264 KSE_SCHED_LOCK(curkse, curkse->k_kseg); 1265} 1266 1267void 1268_thr_gc(struct pthread *curthread) 1269{ 1270 thread_gc(curthread); 1271 kse_gc(curthread); 1272 kseg_gc(curthread); 1273} 1274 1275static void 1276thread_gc(struct pthread *curthread) 1277{ 1278 struct pthread *td, *td_next; 1279 kse_critical_t crit; 1280 TAILQ_HEAD(, pthread) worklist; 1281 1282 TAILQ_INIT(&worklist); 1283 crit = _kse_critical_enter(); 1284 KSE_LOCK_ACQUIRE(curthread->kse, &_thread_list_lock); 1285 1286 /* Check the threads waiting for GC. */ 1287 for (td = TAILQ_FIRST(&_thread_gc_list); td != NULL; td = td_next) { 1288 td_next = TAILQ_NEXT(td, gcle); 1289 if ((td->flags & THR_FLAGS_GC_SAFE) == 0) 1290 continue; 1291 else if (((td->attr.flags & PTHREAD_SCOPE_SYSTEM) != 0) && 1292 ((td->kse->k_kcb->kcb_kmbx.km_flags & KMF_DONE) == 0)) { 1293 /* 1294 * The thread and KSE are operating on the same 1295 * stack. Wait for the KSE to exit before freeing 1296 * the thread's stack as well as everything else. 1297 */ 1298 continue; 1299 } 1300 /* 1301 * Remove the thread from the GC list. If the thread 1302 * isn't yet detached, it will get added back to the 1303 * GC list at a later time. 1304 */ 1305 THR_GCLIST_REMOVE(td); 1306 DBG_MSG("Freeing thread %p stack\n", td); 1307 /* 1308 * We can free the thread stack since it's no longer 1309 * in use. 1310 */ 1311 _thr_stack_free(&td->attr); 1312 if (((td->attr.flags & PTHREAD_DETACHED) != 0) && 1313 (td->refcount == 0)) { 1314 /* 1315 * The thread has detached and is no longer 1316 * referenced. It is safe to remove all 1317 * remnants of the thread. 1318 */ 1319 THR_LIST_REMOVE(td); 1320 TAILQ_INSERT_HEAD(&worklist, td, gcle); 1321 } 1322 } 1323 KSE_LOCK_RELEASE(curthread->kse, &_thread_list_lock); 1324 _kse_critical_leave(crit); 1325 1326 while ((td = TAILQ_FIRST(&worklist)) != NULL) { 1327 TAILQ_REMOVE(&worklist, td, gcle); 1328 1329 if ((td->attr.flags & PTHREAD_SCOPE_SYSTEM) != 0) { 1330 crit = _kse_critical_enter(); 1331 KSE_LOCK_ACQUIRE(curthread->kse, &kse_lock); 1332 kse_free_unlocked(td->kse); 1333 kseg_free_unlocked(td->kseg); 1334 KSE_LOCK_RELEASE(curthread->kse, &kse_lock); 1335 _kse_critical_leave(crit); 1336 } 1337 /* 1338 * XXX we don't free initial thread, because there might 1339 * have some code referencing initial thread. 1340 */ 1341 if (td != _thr_initial) { 1342 DBG_MSG("Freeing thread %p\n", td); 1343 _thr_free(curthread, td); 1344 } else 1345 DBG_MSG("Initial thread won't be freed\n"); 1346 } 1347} 1348 1349static void 1350kse_gc(struct pthread *curthread) 1351{ 1352 kse_critical_t crit; 1353 TAILQ_HEAD(, kse) worklist; 1354 struct kse *kse; 1355 1356 if (free_kse_count <= MAX_CACHED_KSES) 1357 return; 1358 TAILQ_INIT(&worklist); 1359 crit = _kse_critical_enter(); 1360 KSE_LOCK_ACQUIRE(curthread->kse, &kse_lock); 1361 while (free_kse_count > MAX_CACHED_KSES) { 1362 kse = TAILQ_FIRST(&free_kseq); 1363 TAILQ_REMOVE(&free_kseq, kse, k_qe); 1364 TAILQ_INSERT_HEAD(&worklist, kse, k_qe); 1365 free_kse_count--; 1366 } 1367 KSE_LOCK_RELEASE(curthread->kse, &kse_lock); 1368 _kse_critical_leave(crit); 1369 1370 while ((kse = TAILQ_FIRST(&worklist))) { 1371 TAILQ_REMOVE(&worklist, kse, k_qe); 1372 kse_destroy(kse); 1373 } 1374} 1375 1376static void 1377kseg_gc(struct pthread *curthread) 1378{ 1379 kse_critical_t crit; 1380 TAILQ_HEAD(, kse_group) worklist; 1381 struct kse_group *kseg; 1382 1383 if (free_kseg_count <= MAX_CACHED_KSEGS) 1384 return; 1385 crit = _kse_critical_enter(); 1386 KSE_LOCK_ACQUIRE(curthread->kse, &kse_lock); 1387 while (free_kseg_count > MAX_CACHED_KSEGS) { 1388 kseg = TAILQ_FIRST(&free_kse_groupq); 1389 TAILQ_REMOVE(&free_kse_groupq, kseg, kg_qe); 1390 free_kseg_count--; 1391 TAILQ_INSERT_HEAD(&worklist, kseg, kg_qe); 1392 } 1393 KSE_LOCK_RELEASE(curthread->kse, &kse_lock); 1394 _kse_critical_leave(crit); 1395 1396 while ((kseg = TAILQ_FIRST(&worklist))) { 1397 TAILQ_REMOVE(&worklist, kseg, kg_qe); 1398 kseg_destroy(kseg); 1399 } 1400} 1401 1402/* 1403 * Only new threads that are running or suspended may be scheduled. 1404 */ 1405int 1406_thr_schedule_add(struct pthread *curthread, struct pthread *newthread) 1407{ 1408 kse_critical_t crit; 1409 int ret; 1410 1411 /* Add the new thread. */ 1412 thr_link(newthread); 1413 1414 /* 1415 * If this is the first time creating a thread, make sure 1416 * the mailbox is set for the current thread. 1417 */ 1418 if ((newthread->attr.flags & PTHREAD_SCOPE_SYSTEM) != 0) { 1419 /* We use the thread's stack as the KSE's stack. */ 1420 newthread->kse->k_kcb->kcb_kmbx.km_stack.ss_sp = 1421 newthread->attr.stackaddr_attr; 1422 newthread->kse->k_kcb->kcb_kmbx.km_stack.ss_size = 1423 newthread->attr.stacksize_attr; 1424 1425 /* 1426 * No need to lock the scheduling queue since the 1427 * KSE/KSEG pair have not yet been started. 1428 */ 1429 KSEG_THRQ_ADD(newthread->kseg, newthread); 1430 /* this thread never gives up kse */ 1431 newthread->active = 1; 1432 newthread->kse->k_curthread = newthread; 1433 newthread->kse->k_kcb->kcb_kmbx.km_flags = KMF_BOUND; 1434 newthread->kse->k_kcb->kcb_kmbx.km_func = 1435 (kse_func_t *)kse_sched_single; 1436 newthread->kse->k_kcb->kcb_kmbx.km_quantum = 0; 1437 KSE_SET_MBOX(newthread->kse, newthread); 1438 /* 1439 * This thread needs a new KSE and KSEG. 1440 */ 1441 newthread->kse->k_flags &= ~KF_INITIALIZED; 1442 newthread->kse->k_flags |= KF_STARTED; 1443 /* Fire up! */ 1444 ret = kse_create(&newthread->kse->k_kcb->kcb_kmbx, 1); 1445 if (ret != 0) 1446 ret = errno; 1447 } 1448 else { 1449 /* 1450 * Lock the KSE and add the new thread to its list of 1451 * assigned threads. If the new thread is runnable, also 1452 * add it to the KSE's run queue. 1453 */ 1454 crit = _kse_critical_enter(); 1455 KSE_SCHED_LOCK(curthread->kse, newthread->kseg); 1456 KSEG_THRQ_ADD(newthread->kseg, newthread); 1457 if (newthread->state == PS_RUNNING) 1458 THR_RUNQ_INSERT_TAIL(newthread); 1459 if ((newthread->kse->k_flags & KF_STARTED) == 0) { 1460 /* 1461 * This KSE hasn't been started yet. Start it 1462 * outside of holding the lock. 1463 */ 1464 newthread->kse->k_flags |= KF_STARTED; 1465 newthread->kse->k_kcb->kcb_kmbx.km_func = 1466 (kse_func_t *)kse_sched_multi; 1467 newthread->kse->k_kcb->kcb_kmbx.km_flags = 0; 1468 kse_create(&newthread->kse->k_kcb->kcb_kmbx, 0); 1469 } else if ((newthread->state == PS_RUNNING) && 1470 KSE_IS_IDLE(newthread->kse)) { 1471 /* 1472 * The thread is being scheduled on another KSEG. 1473 */ 1474 kse_wakeup_one(newthread); 1475 } 1476 KSE_SCHED_UNLOCK(curthread->kse, newthread->kseg); 1477 _kse_critical_leave(crit); 1478 ret = 0; 1479 } 1480 if (ret != 0) 1481 thr_unlink(newthread); 1482 1483 return (ret); 1484} 1485 1486void 1487kse_waitq_insert(struct pthread *thread) 1488{ 1489 struct pthread *td; 1490 1491 if (thread->wakeup_time.tv_sec == -1) 1492 TAILQ_INSERT_TAIL(&thread->kse->k_schedq->sq_waitq, thread, 1493 pqe); 1494 else { 1495 td = TAILQ_FIRST(&thread->kse->k_schedq->sq_waitq); 1496 while ((td != NULL) && (td->wakeup_time.tv_sec != -1) && 1497 ((td->wakeup_time.tv_sec < thread->wakeup_time.tv_sec) || 1498 ((td->wakeup_time.tv_sec == thread->wakeup_time.tv_sec) && 1499 (td->wakeup_time.tv_nsec <= thread->wakeup_time.tv_nsec)))) 1500 td = TAILQ_NEXT(td, pqe); 1501 if (td == NULL) 1502 TAILQ_INSERT_TAIL(&thread->kse->k_schedq->sq_waitq, 1503 thread, pqe); 1504 else 1505 TAILQ_INSERT_BEFORE(td, thread, pqe); 1506 } 1507 thread->flags |= THR_FLAGS_IN_WAITQ; 1508} 1509 1510/* 1511 * This must be called with the scheduling lock held. 1512 */ 1513static void 1514kse_check_completed(struct kse *kse) 1515{ 1516 struct pthread *thread; 1517 struct kse_thr_mailbox *completed; 1518 int sig; 1519 1520 if ((completed = kse->k_kcb->kcb_kmbx.km_completed) != NULL) { 1521 kse->k_kcb->kcb_kmbx.km_completed = NULL; 1522 while (completed != NULL) { 1523 thread = completed->tm_udata; 1524 DBG_MSG("Found completed thread %p, name %s\n", 1525 thread, 1526 (thread->name == NULL) ? "none" : thread->name); 1527 thread->blocked = 0; 1528 if (thread != kse->k_curthread) { 1529 thr_accounting(thread); 1530 if ((thread->flags & THR_FLAGS_SUSPENDED) != 0) 1531 THR_SET_STATE(thread, PS_SUSPENDED); 1532 else 1533 KSE_RUNQ_INSERT_TAIL(kse, thread); 1534 if ((thread->kse != kse) && 1535 (thread->kse->k_curthread == thread)) { 1536 /* 1537 * Remove this thread from its 1538 * previous KSE so that it (the KSE) 1539 * doesn't think it is still active. 1540 */ 1541 thread->kse->k_curthread = NULL; 1542 thread->active = 0; 1543 } 1544 } 1545 if ((sig = thread->tcb->tcb_tmbx.tm_syncsig.si_signo) 1546 != 0) { 1547 if (SIGISMEMBER(thread->sigmask, sig)) 1548 SIGADDSET(thread->sigpend, sig); 1549 else 1550 (void)_thr_sig_add(thread, sig, 1551 &thread->tcb->tcb_tmbx.tm_syncsig); 1552 thread->tcb->tcb_tmbx.tm_syncsig.si_signo = 0; 1553 } 1554 completed = completed->tm_next; 1555 } 1556 } 1557} 1558 1559/* 1560 * This must be called with the scheduling lock held. 1561 */ 1562static void 1563kse_check_waitq(struct kse *kse) 1564{ 1565 struct pthread *pthread; 1566 struct timespec ts; 1567 1568 KSE_GET_TOD(kse, &ts); 1569 1570 /* 1571 * Wake up threads that have timedout. This has to be 1572 * done before adding the current thread to the run queue 1573 * so that a CPU intensive thread doesn't get preference 1574 * over waiting threads. 1575 */ 1576 while (((pthread = KSE_WAITQ_FIRST(kse)) != NULL) && 1577 thr_timedout(pthread, &ts)) { 1578 /* Remove the thread from the wait queue: */ 1579 KSE_WAITQ_REMOVE(kse, pthread); 1580 DBG_MSG("Found timedout thread %p in waitq\n", pthread); 1581 1582 /* Indicate the thread timedout: */ 1583 pthread->timeout = 1; 1584 1585 /* Add the thread to the priority queue: */ 1586 if ((pthread->flags & THR_FLAGS_SUSPENDED) != 0) 1587 THR_SET_STATE(pthread, PS_SUSPENDED); 1588 else { 1589 THR_SET_STATE(pthread, PS_RUNNING); 1590 KSE_RUNQ_INSERT_TAIL(kse, pthread); 1591 } 1592 } 1593} 1594 1595static int 1596thr_timedout(struct pthread *thread, struct timespec *curtime) 1597{ 1598 if (thread->wakeup_time.tv_sec < 0) 1599 return (0); 1600 else if (thread->wakeup_time.tv_sec > curtime->tv_sec) 1601 return (0); 1602 else if ((thread->wakeup_time.tv_sec == curtime->tv_sec) && 1603 (thread->wakeup_time.tv_nsec > curtime->tv_nsec)) 1604 return (0); 1605 else 1606 return (1); 1607} 1608 1609/* 1610 * This must be called with the scheduling lock held. 1611 * 1612 * Each thread has a time slice, a wakeup time (used when it wants 1613 * to wait for a specified amount of time), a run state, and an 1614 * active flag. 1615 * 1616 * When a thread gets run by the scheduler, the active flag is 1617 * set to non-zero (1). When a thread performs an explicit yield 1618 * or schedules a state change, it enters the scheduler and the 1619 * active flag is cleared. When the active flag is still seen 1620 * set in the scheduler, that means that the thread is blocked in 1621 * the kernel (because it is cleared before entering the scheduler 1622 * in all other instances). 1623 * 1624 * The wakeup time is only set for those states that can timeout. 1625 * It is set to (-1, -1) for all other instances. 1626 * 1627 * The thread's run state, aside from being useful when debugging, 1628 * is used to place the thread in an appropriate queue. There 1629 * are 2 basic queues: 1630 * 1631 * o run queue - queue ordered by priority for all threads 1632 * that are runnable 1633 * o waiting queue - queue sorted by wakeup time for all threads 1634 * that are not otherwise runnable (not blocked 1635 * in kernel, not waiting for locks) 1636 * 1637 * The thread's time slice is used for round-robin scheduling 1638 * (the default scheduling policy). While a SCHED_RR thread 1639 * is runnable it's time slice accumulates. When it reaches 1640 * the time slice interval, it gets reset and added to the end 1641 * of the queue of threads at its priority. When a thread no 1642 * longer becomes runnable (blocks in kernel, waits, etc), its 1643 * time slice is reset. 1644 * 1645 * The job of kse_switchout_thread() is to handle all of the above. 1646 */ 1647static void 1648kse_switchout_thread(struct kse *kse, struct pthread *thread) 1649{ 1650 int level; 1651 int i; 1652 int restart; 1653 siginfo_t siginfo; 1654 1655 /* 1656 * Place the currently running thread into the 1657 * appropriate queue(s). 1658 */ 1659 DBG_MSG("Switching out thread %p, state %d\n", thread, thread->state); 1660 1661 THR_DEACTIVATE_LAST_LOCK(thread); 1662 if (thread->blocked != 0) { 1663 thread->active = 0; 1664 thread->need_switchout = 0; 1665 /* This thread must have blocked in the kernel. */ 1666 /* 1667 * Check for pending signals for this thread to 1668 * see if we need to interrupt it in the kernel. 1669 */ 1670 if (thread->check_pending != 0) { 1671 for (i = 1; i <= _SIG_MAXSIG; ++i) { 1672 if (SIGISMEMBER(thread->sigpend, i) && 1673 !SIGISMEMBER(thread->sigmask, i)) { 1674 restart = _thread_sigact[1 - 1].sa_flags & SA_RESTART; 1675 kse_thr_interrupt(&thread->tcb->tcb_tmbx, 1676 restart ? KSE_INTR_RESTART : KSE_INTR_INTERRUPT, 0); 1677 break; 1678 } 1679 } 1680 } 1681 } 1682 else { 1683 switch (thread->state) { 1684 case PS_DEAD: 1685 /* 1686 * The scheduler is operating on a different 1687 * stack. It is safe to do garbage collecting 1688 * here. 1689 */ 1690 thread->active = 0; 1691 thread->need_switchout = 0; 1692 thread->lock_switch = 0; 1693 thr_cleanup(kse, thread); 1694 return; 1695 break; 1696 1697 case PS_RUNNING: 1698 if ((thread->flags & THR_FLAGS_SUSPENDED) != 0) 1699 THR_SET_STATE(thread, PS_SUSPENDED); 1700 break; 1701 1702 case PS_COND_WAIT: 1703 case PS_SLEEP_WAIT: 1704 /* Insert into the waiting queue: */ 1705 KSE_WAITQ_INSERT(kse, thread); 1706 break; 1707 1708 case PS_LOCKWAIT: 1709 /* 1710 * This state doesn't timeout. 1711 */ 1712 thread->wakeup_time.tv_sec = -1; 1713 thread->wakeup_time.tv_nsec = -1; 1714 level = thread->locklevel - 1; 1715 if (!_LCK_GRANTED(&thread->lockusers[level])) 1716 KSE_WAITQ_INSERT(kse, thread); 1717 else 1718 THR_SET_STATE(thread, PS_RUNNING); 1719 break; 1720 1721 case PS_SIGWAIT: 1722 KSE_WAITQ_INSERT(kse, thread); 1723 break; 1724 case PS_JOIN: 1725 case PS_MUTEX_WAIT: 1726 case PS_SIGSUSPEND: 1727 case PS_SUSPENDED: 1728 case PS_DEADLOCK: 1729 default: 1730 /* 1731 * These states don't timeout. 1732 */ 1733 thread->wakeup_time.tv_sec = -1; 1734 thread->wakeup_time.tv_nsec = -1; 1735 1736 /* Insert into the waiting queue: */ 1737 KSE_WAITQ_INSERT(kse, thread); 1738 break; 1739 } 1740 thr_accounting(thread); 1741 if (thread->state == PS_RUNNING) { 1742 if (thread->slice_usec == -1) { 1743 /* 1744 * The thread exceeded its time quantum or 1745 * it yielded the CPU; place it at the tail 1746 * of the queue for its priority. 1747 */ 1748 KSE_RUNQ_INSERT_TAIL(kse, thread); 1749 } else { 1750 /* 1751 * The thread hasn't exceeded its interval 1752 * Place it at the head of the queue for its 1753 * priority. 1754 */ 1755 KSE_RUNQ_INSERT_HEAD(kse, thread); 1756 } 1757 } 1758 } 1759 thread->active = 0; 1760 thread->need_switchout = 0; 1761 if (thread->check_pending != 0) { 1762 /* Install pending signals into the frame. */ 1763 thread->check_pending = 0; 1764 KSE_LOCK_ACQUIRE(kse, &_thread_signal_lock); 1765 for (i = 1; i <= _SIG_MAXSIG; i++) { 1766 if (SIGISMEMBER(thread->sigmask, i)) 1767 continue; 1768 if (SIGISMEMBER(thread->sigpend, i)) 1769 (void)_thr_sig_add(thread, i, 1770 &thread->siginfo[i-1]); 1771 else if (SIGISMEMBER(_thr_proc_sigpending, i) && 1772 _thr_getprocsig_unlocked(i, &siginfo)) { 1773 (void)_thr_sig_add(thread, i, &siginfo); 1774 } 1775 } 1776 KSE_LOCK_RELEASE(kse, &_thread_signal_lock); 1777 } 1778} 1779 1780/* 1781 * This function waits for the smallest timeout value of any waiting 1782 * thread, or until it receives a message from another KSE. 1783 * 1784 * This must be called with the scheduling lock held. 1785 */ 1786static void 1787kse_wait(struct kse *kse, struct pthread *td_wait, int sigseqno) 1788{ 1789 struct timespec ts, ts_sleep; 1790 int saved_flags; 1791 1792 KSE_GET_TOD(kse, &ts); 1793 1794 if ((td_wait == NULL) || (td_wait->wakeup_time.tv_sec < 0)) { 1795 /* Limit sleep to no more than 1 minute. */ 1796 ts_sleep.tv_sec = 60; 1797 ts_sleep.tv_nsec = 0; 1798 } else { 1799 TIMESPEC_SUB(&ts_sleep, &td_wait->wakeup_time, &ts); 1800 if (ts_sleep.tv_sec > 60) { 1801 ts_sleep.tv_sec = 60; 1802 ts_sleep.tv_nsec = 0; 1803 } 1804 } 1805 /* Don't sleep for negative times. */ 1806 if ((ts_sleep.tv_sec >= 0) && (ts_sleep.tv_nsec >= 0)) { 1807 KSE_SET_IDLE(kse); 1808 kse->k_kseg->kg_idle_kses++; 1809 KSE_SCHED_UNLOCK(kse, kse->k_kseg); 1810 if ((kse->k_kseg->kg_flags & KGF_SINGLE_THREAD) && 1811 (kse->k_sigseqno != sigseqno)) 1812 ; /* don't sleep */ 1813 else { 1814 saved_flags = kse->k_kcb->kcb_kmbx.km_flags; 1815 kse->k_kcb->kcb_kmbx.km_flags |= KMF_NOUPCALL; 1816 kse_release(&ts_sleep); 1817 kse->k_kcb->kcb_kmbx.km_flags = saved_flags; 1818 } 1819 KSE_SCHED_LOCK(kse, kse->k_kseg); 1820 if (KSE_IS_IDLE(kse)) { 1821 KSE_CLEAR_IDLE(kse); 1822 kse->k_kseg->kg_idle_kses--; 1823 } 1824 } 1825} 1826 1827/* 1828 * Avoid calling this kse_exit() so as not to confuse it with the 1829 * system call of the same name. 1830 */ 1831static void 1832kse_fini(struct kse *kse) 1833{ 1834 /* struct kse_group *free_kseg = NULL; */ 1835 struct timespec ts; 1836 1837 /* 1838 * Check to see if this is one of the main kses. 1839 */ 1840 if (kse->k_kseg != _kse_initial->k_kseg) { 1841 PANIC("shouldn't get here"); 1842 /* This is for supporting thread groups. */ 1843#ifdef NOT_YET 1844 /* Remove this KSE from the KSEG's list of KSEs. */ 1845 KSE_SCHED_LOCK(kse, kse->k_kseg); 1846 TAILQ_REMOVE(&kse->k_kseg->kg_kseq, kse, k_kgqe); 1847 kse->k_kseg->kg_ksecount--; 1848 if (TAILQ_EMPTY(&kse->k_kseg->kg_kseq)) 1849 free_kseg = kse->k_kseg; 1850 KSE_SCHED_UNLOCK(kse, kse->k_kseg); 1851 1852 /* 1853 * Add this KSE to the list of free KSEs along with 1854 * the KSEG if is now orphaned. 1855 */ 1856 KSE_LOCK_ACQUIRE(kse, &kse_lock); 1857 if (free_kseg != NULL) 1858 kseg_free_unlocked(free_kseg); 1859 kse_free_unlocked(kse); 1860 KSE_LOCK_RELEASE(kse, &kse_lock); 1861 kse_exit(); 1862 /* Never returns. */ 1863 PANIC("kse_exit()"); 1864#endif 1865 } else { 1866#ifdef NOT_YET 1867 /* 1868 * In future, we might allow program to kill 1869 * kse in initial group. 1870 */ 1871 if (kse != _kse_initial) { 1872 KSE_SCHED_LOCK(kse, kse->k_kseg); 1873 TAILQ_REMOVE(&kse->k_kseg->kg_kseq, kse, k_kgqe); 1874 kse->k_kseg->kg_ksecount--; 1875 KSE_SCHED_UNLOCK(kse, kse->k_kseg); 1876 KSE_LOCK_ACQUIRE(kse, &kse_lock); 1877 kse_free_unlocked(kse); 1878 KSE_LOCK_RELEASE(kse, &kse_lock); 1879 kse_exit(); 1880 /* Never returns. */ 1881 PANIC("kse_exit() failed for initial kseg"); 1882 } 1883#endif 1884 KSE_SCHED_LOCK(kse, kse->k_kseg); 1885 KSE_SET_IDLE(kse); 1886 kse->k_kseg->kg_idle_kses++; 1887 KSE_SCHED_UNLOCK(kse, kse->k_kseg); 1888 ts.tv_sec = 120; 1889 ts.tv_nsec = 0; 1890 kse->k_kcb->kcb_kmbx.km_flags = 0; 1891 kse_release(&ts); 1892 /* Never reach */ 1893 } 1894} 1895 1896void 1897_thr_set_timeout(const struct timespec *timeout) 1898{ 1899 struct pthread *curthread = _get_curthread(); 1900 struct timespec ts; 1901 1902 /* Reset the timeout flag for the running thread: */ 1903 curthread->timeout = 0; 1904 1905 /* Check if the thread is to wait forever: */ 1906 if (timeout == NULL) { 1907 /* 1908 * Set the wakeup time to something that can be recognised as 1909 * different to an actual time of day: 1910 */ 1911 curthread->wakeup_time.tv_sec = -1; 1912 curthread->wakeup_time.tv_nsec = -1; 1913 } 1914 /* Check if no waiting is required: */ 1915 else if ((timeout->tv_sec == 0) && (timeout->tv_nsec == 0)) { 1916 /* Set the wake up time to 'immediately': */ 1917 curthread->wakeup_time.tv_sec = 0; 1918 curthread->wakeup_time.tv_nsec = 0; 1919 } else { 1920 /* Calculate the time for the current thread to wakeup: */ 1921 KSE_GET_TOD(curthread->kse, &ts); 1922 TIMESPEC_ADD(&curthread->wakeup_time, &ts, timeout); 1923 } 1924} 1925 1926void 1927_thr_panic_exit(char *file, int line, char *msg) 1928{ 1929 char buf[256]; 1930 1931 snprintf(buf, sizeof(buf), "(%s:%d) %s\n", file, line, msg); 1932 __sys_write(2, buf, strlen(buf)); 1933 abort(); 1934} 1935 1936void 1937_thr_setrunnable(struct pthread *curthread, struct pthread *thread) 1938{ 1939 kse_critical_t crit; 1940 struct kse_mailbox *kmbx; 1941 1942 crit = _kse_critical_enter(); 1943 KSE_SCHED_LOCK(curthread->kse, thread->kseg); 1944 kmbx = _thr_setrunnable_unlocked(thread); 1945 KSE_SCHED_UNLOCK(curthread->kse, thread->kseg); 1946 _kse_critical_leave(crit); 1947 if (kmbx != NULL) 1948 kse_wakeup(kmbx); 1949} 1950 1951struct kse_mailbox * 1952_thr_setrunnable_unlocked(struct pthread *thread) 1953{ 1954 struct kse_mailbox *kmbx = NULL; 1955 1956 if ((thread->kseg->kg_flags & KGF_SINGLE_THREAD) != 0) { 1957 /* No silly queues for these threads. */ 1958 if ((thread->flags & THR_FLAGS_SUSPENDED) != 0) 1959 THR_SET_STATE(thread, PS_SUSPENDED); 1960 else { 1961 THR_SET_STATE(thread, PS_RUNNING); 1962 kmbx = kse_wakeup_one(thread); 1963 } 1964 1965 } else if (thread->state != PS_RUNNING) { 1966 if ((thread->flags & THR_FLAGS_IN_WAITQ) != 0) 1967 KSE_WAITQ_REMOVE(thread->kse, thread); 1968 if ((thread->flags & THR_FLAGS_SUSPENDED) != 0) 1969 THR_SET_STATE(thread, PS_SUSPENDED); 1970 else { 1971 THR_SET_STATE(thread, PS_RUNNING); 1972 if ((thread->blocked == 0) && (thread->active == 0) && 1973 (thread->flags & THR_FLAGS_IN_RUNQ) == 0) 1974 THR_RUNQ_INSERT_TAIL(thread); 1975 /* 1976 * XXX - Threads are not yet assigned to specific 1977 * KSEs; they are assigned to the KSEG. So 1978 * the fact that a thread's KSE is waiting 1979 * doesn't necessarily mean that it will be 1980 * the KSE that runs the thread after the 1981 * lock is granted. But we don't know if the 1982 * other KSEs within the same KSEG are also 1983 * in a waiting state or not so we err on the 1984 * side of caution and wakeup the thread's 1985 * last known KSE. We ensure that the 1986 * threads KSE doesn't change while it's 1987 * scheduling lock is held so it is safe to 1988 * reference it (the KSE). If the KSE wakes 1989 * up and doesn't find any more work it will 1990 * again go back to waiting so no harm is 1991 * done. 1992 */ 1993 kmbx = kse_wakeup_one(thread); 1994 } 1995 } 1996 return (kmbx); 1997} 1998 1999static struct kse_mailbox * 2000kse_wakeup_one(struct pthread *thread) 2001{ 2002 struct kse *ke; 2003 2004 if (KSE_IS_IDLE(thread->kse)) { 2005 KSE_CLEAR_IDLE(thread->kse); 2006 thread->kseg->kg_idle_kses--; 2007 return (&thread->kse->k_kcb->kcb_kmbx); 2008 } else { 2009 TAILQ_FOREACH(ke, &thread->kseg->kg_kseq, k_kgqe) { 2010 if (KSE_IS_IDLE(ke)) { 2011 KSE_CLEAR_IDLE(ke); 2012 ke->k_kseg->kg_idle_kses--; 2013 return (&ke->k_kcb->kcb_kmbx); 2014 } 2015 } 2016 } 2017 return (NULL); 2018} 2019 2020static void 2021kse_wakeup_multi(struct kse *curkse) 2022{ 2023 struct kse *ke; 2024 int tmp; 2025 2026 if ((tmp = KSE_RUNQ_THREADS(curkse)) && curkse->k_kseg->kg_idle_kses) { 2027 TAILQ_FOREACH(ke, &curkse->k_kseg->kg_kseq, k_kgqe) { 2028 if (KSE_IS_IDLE(ke)) { 2029 KSE_CLEAR_IDLE(ke); 2030 ke->k_kseg->kg_idle_kses--; 2031 KSE_WAKEUP(ke); 2032 if (--tmp == 0) 2033 break; 2034 } 2035 } 2036 } 2037} 2038 2039/* 2040 * Allocate a new KSEG. 2041 * 2042 * We allow the current thread to be NULL in the case that this 2043 * is the first time a KSEG is being created (library initialization). 2044 * In this case, we don't need to (and can't) take any locks. 2045 */ 2046struct kse_group * 2047_kseg_alloc(struct pthread *curthread) 2048{ 2049 struct kse_group *kseg = NULL; 2050 kse_critical_t crit; 2051 2052 if ((curthread != NULL) && (free_kseg_count > 0)) { 2053 /* Use the kse lock for the kseg queue. */ 2054 crit = _kse_critical_enter(); 2055 KSE_LOCK_ACQUIRE(curthread->kse, &kse_lock); 2056 if ((kseg = TAILQ_FIRST(&free_kse_groupq)) != NULL) { 2057 TAILQ_REMOVE(&free_kse_groupq, kseg, kg_qe); 2058 free_kseg_count--; 2059 active_kseg_count++; 2060 TAILQ_INSERT_TAIL(&active_kse_groupq, kseg, kg_qe); 2061 } 2062 KSE_LOCK_RELEASE(curthread->kse, &kse_lock); 2063 _kse_critical_leave(crit); 2064 if (kseg) 2065 kseg_reinit(kseg); 2066 } 2067 2068 /* 2069 * If requested, attempt to allocate a new KSE group only if the 2070 * KSE allocation was successful and a KSE group wasn't found in 2071 * the free list. 2072 */ 2073 if ((kseg == NULL) && 2074 ((kseg = (struct kse_group *)malloc(sizeof(*kseg))) != NULL)) { 2075 if (_pq_alloc(&kseg->kg_schedq.sq_runq, 2076 THR_MIN_PRIORITY, THR_LAST_PRIORITY) != 0) { 2077 free(kseg); 2078 kseg = NULL; 2079 } else { 2080 kseg_init(kseg); 2081 /* Add the KSEG to the list of active KSEGs. */ 2082 if (curthread != NULL) { 2083 crit = _kse_critical_enter(); 2084 KSE_LOCK_ACQUIRE(curthread->kse, &kse_lock); 2085 active_kseg_count++; 2086 TAILQ_INSERT_TAIL(&active_kse_groupq, 2087 kseg, kg_qe); 2088 KSE_LOCK_RELEASE(curthread->kse, &kse_lock); 2089 _kse_critical_leave(crit); 2090 } else { 2091 active_kseg_count++; 2092 TAILQ_INSERT_TAIL(&active_kse_groupq, 2093 kseg, kg_qe); 2094 } 2095 } 2096 } 2097 return (kseg); 2098} 2099 2100static void 2101kseg_init(struct kse_group *kseg) 2102{ 2103 kseg_reinit(kseg); 2104 _lock_init(&kseg->kg_lock, LCK_ADAPTIVE, _kse_lock_wait, 2105 _kse_lock_wakeup); 2106} 2107 2108static void 2109kseg_reinit(struct kse_group *kseg) 2110{ 2111 TAILQ_INIT(&kseg->kg_kseq); 2112 TAILQ_INIT(&kseg->kg_threadq); 2113 TAILQ_INIT(&kseg->kg_schedq.sq_waitq); 2114 kseg->kg_threadcount = 0; 2115 kseg->kg_ksecount = 0; 2116 kseg->kg_idle_kses = 0; 2117 kseg->kg_flags = 0; 2118} 2119 2120/* 2121 * This must be called with the kse lock held and when there are 2122 * no more threads that reference it. 2123 */ 2124static void 2125kseg_free_unlocked(struct kse_group *kseg) 2126{ 2127 TAILQ_REMOVE(&active_kse_groupq, kseg, kg_qe); 2128 TAILQ_INSERT_HEAD(&free_kse_groupq, kseg, kg_qe); 2129 free_kseg_count++; 2130 active_kseg_count--; 2131} 2132 2133void 2134_kseg_free(struct kse_group *kseg) 2135{ 2136 struct kse *curkse; 2137 kse_critical_t crit; 2138 2139 crit = _kse_critical_enter(); 2140 curkse = _get_curkse(); 2141 KSE_LOCK_ACQUIRE(curkse, &kse_lock); 2142 kseg_free_unlocked(kseg); 2143 KSE_LOCK_RELEASE(curkse, &kse_lock); 2144 _kse_critical_leave(crit); 2145} 2146 2147static void 2148kseg_destroy(struct kse_group *kseg) 2149{ 2150 _lock_destroy(&kseg->kg_lock); 2151 _pq_free(&kseg->kg_schedq.sq_runq); 2152 free(kseg); 2153} 2154 2155/* 2156 * Allocate a new KSE. 2157 * 2158 * We allow the current thread to be NULL in the case that this 2159 * is the first time a KSE is being created (library initialization). 2160 * In this case, we don't need to (and can't) take any locks. 2161 */ 2162struct kse * 2163_kse_alloc(struct pthread *curthread, int sys_scope) 2164{ 2165 struct kse *kse = NULL; 2166 char *stack; 2167 kse_critical_t crit; 2168 int i; 2169 2170 if ((curthread != NULL) && (free_kse_count > 0)) { 2171 crit = _kse_critical_enter(); 2172 KSE_LOCK_ACQUIRE(curthread->kse, &kse_lock); 2173 /* Search for a finished KSE. */ 2174 kse = TAILQ_FIRST(&free_kseq); 2175 while ((kse != NULL) && 2176 ((kse->k_kcb->kcb_kmbx.km_flags & KMF_DONE) == 0)) { 2177 kse = TAILQ_NEXT(kse, k_qe); 2178 } 2179 if (kse != NULL) { 2180 DBG_MSG("found an unused kse.\n"); 2181 TAILQ_REMOVE(&free_kseq, kse, k_qe); 2182 free_kse_count--; 2183 TAILQ_INSERT_TAIL(&active_kseq, kse, k_qe); 2184 active_kse_count++; 2185 } 2186 KSE_LOCK_RELEASE(curthread->kse, &kse_lock); 2187 _kse_critical_leave(crit); 2188 if (kse != NULL) 2189 kse_reinit(kse, sys_scope); 2190 } 2191 if ((kse == NULL) && 2192 ((kse = (struct kse *)malloc(sizeof(*kse))) != NULL)) { 2193 if (sys_scope != 0) 2194 stack = NULL; 2195 else if ((stack = malloc(KSE_STACKSIZE)) == NULL) { 2196 free(kse); 2197 return (NULL); 2198 } 2199 bzero(kse, sizeof(*kse)); 2200 2201 /* Initialize KCB without the lock. */ 2202 if ((kse->k_kcb = _kcb_ctor(kse)) == NULL) { 2203 if (stack != NULL) 2204 free(stack); 2205 free(kse); 2206 return (NULL); 2207 } 2208 2209 /* Initialize the lockusers. */ 2210 for (i = 0; i < MAX_KSE_LOCKLEVEL; i++) { 2211 _lockuser_init(&kse->k_lockusers[i], (void *)kse); 2212 _LCK_SET_PRIVATE2(&kse->k_lockusers[i], NULL); 2213 } 2214 /* _lock_init(kse->k_lock, ...) */ 2215 2216 if (curthread != NULL) { 2217 crit = _kse_critical_enter(); 2218 KSE_LOCK_ACQUIRE(curthread->kse, &kse_lock); 2219 } 2220 kse->k_flags = 0; 2221 TAILQ_INSERT_TAIL(&active_kseq, kse, k_qe); 2222 active_kse_count++; 2223 if (curthread != NULL) { 2224 KSE_LOCK_RELEASE(curthread->kse, &kse_lock); 2225 _kse_critical_leave(crit); 2226 } 2227 /* 2228 * Create the KSE context. 2229 * Scope system threads (one thread per KSE) are not required 2230 * to have a stack for an unneeded kse upcall. 2231 */ 2232 if (!sys_scope) { 2233 kse->k_kcb->kcb_kmbx.km_func = (kse_func_t *)kse_sched_multi; 2234 kse->k_stack.ss_sp = stack; 2235 kse->k_stack.ss_size = KSE_STACKSIZE; 2236 } else { 2237 kse->k_kcb->kcb_kmbx.km_func = (kse_func_t *)kse_sched_single; 2238 kse->k_stack.ss_sp = NULL; 2239 kse->k_stack.ss_size = 0; 2240 } 2241 kse->k_kcb->kcb_kmbx.km_udata = (void *)kse; 2242 kse->k_kcb->kcb_kmbx.km_quantum = 20000; 2243 /* 2244 * We need to keep a copy of the stack in case it 2245 * doesn't get used; a KSE running a scope system 2246 * thread will use that thread's stack. 2247 */ 2248 kse->k_kcb->kcb_kmbx.km_stack = kse->k_stack; 2249 } 2250 return (kse); 2251} 2252 2253static void 2254kse_reinit(struct kse *kse, int sys_scope) 2255{ 2256 if (!sys_scope) { 2257 kse->k_kcb->kcb_kmbx.km_func = (kse_func_t *)kse_sched_multi; 2258 if (kse->k_stack.ss_sp == NULL) { 2259 /* XXX check allocation failure */ 2260 kse->k_stack.ss_sp = (char *) malloc(KSE_STACKSIZE); 2261 kse->k_stack.ss_size = KSE_STACKSIZE; 2262 } 2263 kse->k_kcb->kcb_kmbx.km_quantum = 20000; 2264 } else { 2265 kse->k_kcb->kcb_kmbx.km_func = (kse_func_t *)kse_sched_single; 2266 if (kse->k_stack.ss_sp) 2267 free(kse->k_stack.ss_sp); 2268 kse->k_stack.ss_sp = NULL; 2269 kse->k_stack.ss_size = 0; 2270 kse->k_kcb->kcb_kmbx.km_quantum = 0; 2271 } 2272 kse->k_kcb->kcb_kmbx.km_stack = kse->k_stack; 2273 kse->k_kcb->kcb_kmbx.km_udata = (void *)kse; 2274 kse->k_kcb->kcb_kmbx.km_curthread = NULL; 2275 kse->k_kcb->kcb_kmbx.km_flags = 0; 2276 kse->k_curthread = NULL; 2277 kse->k_kseg = 0; 2278 kse->k_schedq = 0; 2279 kse->k_locklevel = 0; 2280 SIGEMPTYSET(kse->k_sigmask); 2281 bzero(&kse->k_sigq, sizeof(kse->k_sigq)); 2282 kse->k_check_sigq = 0; 2283 kse->k_flags = 0; 2284 kse->k_waiting = 0; 2285 kse->k_idle = 0; 2286 kse->k_error = 0; 2287 kse->k_cpu = 0; 2288 kse->k_done = 0; 2289 kse->k_switch = 0; 2290 kse->k_sigseqno = 0; 2291} 2292 2293void 2294kse_free_unlocked(struct kse *kse) 2295{ 2296 TAILQ_REMOVE(&active_kseq, kse, k_qe); 2297 active_kse_count--; 2298 kse->k_kseg = NULL; 2299 kse->k_kcb->kcb_kmbx.km_quantum = 20000; 2300 kse->k_flags = 0; 2301 TAILQ_INSERT_HEAD(&free_kseq, kse, k_qe); 2302 free_kse_count++; 2303} 2304 2305void 2306_kse_free(struct pthread *curthread, struct kse *kse) 2307{ 2308 kse_critical_t crit; 2309 2310 if (curthread == NULL) 2311 kse_free_unlocked(kse); 2312 else { 2313 crit = _kse_critical_enter(); 2314 KSE_LOCK_ACQUIRE(curthread->kse, &kse_lock); 2315 kse_free_unlocked(kse); 2316 KSE_LOCK_RELEASE(curthread->kse, &kse_lock); 2317 _kse_critical_leave(crit); 2318 } 2319} 2320 2321static void 2322kse_destroy(struct kse *kse) 2323{ 2324 int i; 2325 2326 if (kse->k_stack.ss_sp != NULL) 2327 free(kse->k_stack.ss_sp); 2328 _kcb_dtor(kse->k_kcb); 2329 for (i = 0; i < MAX_KSE_LOCKLEVEL; ++i) 2330 _lockuser_destroy(&kse->k_lockusers[i]); 2331 _lock_destroy(&kse->k_lock); 2332 free(kse); 2333} 2334 2335struct pthread * 2336_thr_alloc(struct pthread *curthread) 2337{ 2338 kse_critical_t crit; 2339 struct pthread *thread = NULL; 2340 2341 if (curthread != NULL) { 2342 if (GC_NEEDED()) 2343 _thr_gc(curthread); 2344 if (free_thread_count > 0) { 2345 crit = _kse_critical_enter(); 2346 KSE_LOCK_ACQUIRE(curthread->kse, &thread_lock); 2347 if ((thread = TAILQ_FIRST(&free_threadq)) != NULL) { 2348 TAILQ_REMOVE(&free_threadq, thread, tle); 2349 free_thread_count--; 2350 } 2351 KSE_LOCK_RELEASE(curthread->kse, &thread_lock); 2352 _kse_critical_leave(crit); 2353 } 2354 } 2355 if ((thread == NULL) && 2356 ((thread = malloc(sizeof(struct pthread))) != NULL)) { 2357 bzero(thread, sizeof(struct pthread)); 2358 if ((thread->tcb = _tcb_ctor(thread)) == NULL) { 2359 free(thread); 2360 thread = NULL; 2361 } 2362 } 2363 return (thread); 2364} 2365 2366void 2367_thr_free(struct pthread *curthread, struct pthread *thread) 2368{ 2369 kse_critical_t crit; 2370 int i; 2371 2372 DBG_MSG("Freeing thread %p\n", thread); 2373 if ((curthread == NULL) || (free_thread_count >= MAX_CACHED_THREADS)) { 2374 for (i = 0; i < MAX_THR_LOCKLEVEL; i++) { 2375 _lockuser_destroy(&thread->lockusers[i]); 2376 } 2377 _lock_destroy(&thread->lock); 2378 _tcb_dtor(thread->tcb); 2379 free(thread); 2380 } 2381 else { 2382 /* Reinitialize any important fields here. */ 2383 thread->lock_switch = 0; 2384 sigemptyset(&thread->sigpend); 2385 thread->check_pending = 0; 2386 2387 /* Add the thread to the free thread list. */ 2388 crit = _kse_critical_enter(); 2389 KSE_LOCK_ACQUIRE(curthread->kse, &thread_lock); 2390 TAILQ_INSERT_TAIL(&free_threadq, thread, tle); 2391 free_thread_count++; 2392 KSE_LOCK_RELEASE(curthread->kse, &thread_lock); 2393 _kse_critical_leave(crit); 2394 } 2395} 2396 2397/* 2398 * Add an active thread: 2399 * 2400 * o Assign the thread a unique id (which GDB uses to track 2401 * threads. 2402 * o Add the thread to the list of all threads and increment 2403 * number of active threads. 2404 */ 2405static void 2406thr_link(struct pthread *thread) 2407{ 2408 kse_critical_t crit; 2409 struct kse *curkse; 2410 struct pthread *curthread; 2411 2412 crit = _kse_critical_enter(); 2413 curkse = _get_curkse(); 2414 curthread = _get_curthread(); 2415 thread->sigmask = curthread->sigmask; 2416 KSE_LOCK_ACQUIRE(curkse, &_thread_list_lock); 2417 /* 2418 * Initialize the unique id (which GDB uses to track 2419 * threads), add the thread to the list of all threads, 2420 * and 2421 */ 2422 thread->uniqueid = next_uniqueid++; 2423 THR_LIST_ADD(thread); 2424 active_threads++; 2425 KSE_LOCK_RELEASE(curkse, &_thread_list_lock); 2426 _kse_critical_leave(crit); 2427} 2428 2429/* 2430 * Remove an active thread. 2431 */ 2432static void 2433thr_unlink(struct pthread *thread) 2434{ 2435 kse_critical_t crit; 2436 struct kse *curkse; 2437 2438 crit = _kse_critical_enter(); 2439 curkse = _get_curkse(); 2440 KSE_LOCK_ACQUIRE(curkse, &_thread_list_lock); 2441 THR_LIST_REMOVE(thread); 2442 active_threads--; 2443 KSE_LOCK_RELEASE(curkse, &_thread_list_lock); 2444 _kse_critical_leave(crit); 2445} 2446 2447void 2448_thr_hash_add(struct pthread *thread) 2449{ 2450 struct thread_hash_head *head; 2451 2452 head = &thr_hashtable[THREAD_HASH(thread)]; 2453 LIST_INSERT_HEAD(head, thread, hle); 2454} 2455 2456void 2457_thr_hash_remove(struct pthread *thread) 2458{ 2459 LIST_REMOVE(thread, hle); 2460} 2461 2462struct pthread * 2463_thr_hash_find(struct pthread *thread) 2464{ 2465 struct pthread *td; 2466 struct thread_hash_head *head; 2467 2468 head = &thr_hashtable[THREAD_HASH(thread)]; 2469 LIST_FOREACH(td, head, hle) { 2470 if (td == thread) 2471 return (thread); 2472 } 2473 return (NULL); 2474} 2475 2476